maxjhandsome
/
llvm-project
forked from OSchip/llvm-project
1
0
Fork 0
Code Issues Pull Requests Packages Releases Wiki Activity
llvm-project/llvm/test/CodeGen/X86/avx512bwvl-intrinsics-upgra...

3829 lines
228 KiB
LLVM
Raw Normal View History

[X86][AVX512] Add support for UNPCK masked shuffle comments llvm-svn: 274464
2016-07-03 22:26:21 +08:00
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
[X86][SKX][KNL] Updated regression tests to use -mattr instead of -mcpu flag.NFC. NFC. Updated 8 regression tests to use -mattr instead of -mcpu flag as follows: -mcpu=knl --> -mattr=+avx512f -mcpu=skx --> -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq The updates are as part of the preparation of a large commit to add all instruction scheduling for the SKX target. Reviewers: delena, zvi, RKSimon Differential Revision: https://reviews.llvm.org/D38222 Change-Id: I2381c9b5bb75ecacfca017243c22d054f6eddd14 llvm-svn: 314306
2017-09-27 22:44:15 +08:00
; RUN: llc < %s -mtriple=x86_64-apple-darwin -mattr=+avx512f,+avx512bw,+avx512vl --show-mc-encoding| FileCheck %s
[AVX512] Remove masked store intrinsics. Clang now emits generic masked store intrinsics instead. The intrinsics will be autoupgraded to the same generic masked stores. llvm-svn: 271245
2016-05-31 09:50:02 +08:00
[x86] Lowering Mask Set1 intrinsics to LLVM IR This patch, together with a matching clang patch (https://reviews.llvm.org/D37668), implements the lowering of X86 mask set1 intrinsics to IR. Differential Revision: https://reviews.llvm.org/D37669 llvm-svn: 313625
2017-09-19 19:03:06 +08:00
declare <16 x i8> @llvm.x86.avx512.mask.pbroadcast.b.gpr.128(i8, <16 x i8>, i16)
define <16 x i8>@test_int_x86_avx512_mask_pbroadcast_b_gpr_128(i8 %x0, <16 x i8> %x1, i16 %mask) {
; CHECK-LABEL: test_int_x86_avx512_mask_pbroadcast_b_gpr_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[x86] Lowering Mask Set1 intrinsics to LLVM IR This patch, together with a matching clang patch (https://reviews.llvm.org/D37668), implements the lowering of X86 mask set1 intrinsics to IR. Differential Revision: https://reviews.llvm.org/D37669 llvm-svn: 313625
2017-09-19 19:03:06 +08:00
; CHECK-NEXT: vpbroadcastb %edi, %xmm1 ## encoding: [0x62,0xf2,0x7d,0x08,0x7a,0xcf]
; CHECK-NEXT: kmovd %esi, %k1 ## encoding: [0xc5,0xfb,0x92,0xce]
; CHECK-NEXT: vpbroadcastb %edi, %xmm0 {%k1} ## encoding: [0x62,0xf2,0x7d,0x09,0x7a,0xc7]
[X86][SKX][KNL] Updated regression tests to use -mattr instead of -mcpu flag.NFC. NFC. Updated 8 regression tests to use -mattr instead of -mcpu flag as follows: -mcpu=knl --> -mattr=+avx512f -mcpu=skx --> -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq The updates are as part of the preparation of a large commit to add all instruction scheduling for the SKX target. Reviewers: delena, zvi, RKSimon Differential Revision: https://reviews.llvm.org/D38222 Change-Id: I2381c9b5bb75ecacfca017243c22d054f6eddd14 llvm-svn: 314306
2017-09-27 22:44:15 +08:00
; CHECK-NEXT: vpbroadcastb %edi, %xmm2 {%k1} {z} ## encoding: [0x62,0xf2,0x7d,0x89,0x7a,0xd7]
; CHECK-NEXT: vpaddb %xmm2, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xfc,0xc2]
[x86] Lowering Mask Set1 intrinsics to LLVM IR This patch, together with a matching clang patch (https://reviews.llvm.org/D37668), implements the lowering of X86 mask set1 intrinsics to IR. Differential Revision: https://reviews.llvm.org/D37669 llvm-svn: 313625
2017-09-19 19:03:06 +08:00
; CHECK-NEXT: vpaddb %xmm0, %xmm1, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf1,0xfc,0xc0]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <16 x i8> @llvm.x86.avx512.mask.pbroadcast.b.gpr.128(i8 %x0, <16 x i8> %x1, i16 -1)
%res1 = call <16 x i8> @llvm.x86.avx512.mask.pbroadcast.b.gpr.128(i8 %x0, <16 x i8> %x1, i16 %mask)
%res2 = call <16 x i8> @llvm.x86.avx512.mask.pbroadcast.b.gpr.128(i8 %x0, <16 x i8> zeroinitializer, i16 %mask)
%res3 = add <16 x i8> %res, %res1
%res4 = add <16 x i8> %res2, %res3
ret <16 x i8> %res4
}
declare <8 x i16> @llvm.x86.avx512.mask.pbroadcast.w.gpr.128(i16, <8 x i16>, i8)
define <8 x i16>@test_int_x86_avx512_mask_pbroadcast_w_gpr_128(i16 %x0, <8 x i16> %x1, i8 %mask) {
; CHECK-LABEL: test_int_x86_avx512_mask_pbroadcast_w_gpr_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[x86] Lowering Mask Set1 intrinsics to LLVM IR This patch, together with a matching clang patch (https://reviews.llvm.org/D37668), implements the lowering of X86 mask set1 intrinsics to IR. Differential Revision: https://reviews.llvm.org/D37669 llvm-svn: 313625
2017-09-19 19:03:06 +08:00
; CHECK-NEXT: vpbroadcastw %edi, %xmm1 ## encoding: [0x62,0xf2,0x7d,0x08,0x7b,0xcf]
; CHECK-NEXT: kmovd %esi, %k1 ## encoding: [0xc5,0xfb,0x92,0xce]
; CHECK-NEXT: vpbroadcastw %edi, %xmm0 {%k1} ## encoding: [0x62,0xf2,0x7d,0x09,0x7b,0xc7]
[X86][SKX][KNL] Updated regression tests to use -mattr instead of -mcpu flag.NFC. NFC. Updated 8 regression tests to use -mattr instead of -mcpu flag as follows: -mcpu=knl --> -mattr=+avx512f -mcpu=skx --> -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq The updates are as part of the preparation of a large commit to add all instruction scheduling for the SKX target. Reviewers: delena, zvi, RKSimon Differential Revision: https://reviews.llvm.org/D38222 Change-Id: I2381c9b5bb75ecacfca017243c22d054f6eddd14 llvm-svn: 314306
2017-09-27 22:44:15 +08:00
; CHECK-NEXT: vpbroadcastw %edi, %xmm2 {%k1} {z} ## encoding: [0x62,0xf2,0x7d,0x89,0x7b,0xd7]
; CHECK-NEXT: vpaddw %xmm2, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xfd,0xc2]
[x86] Lowering Mask Set1 intrinsics to LLVM IR This patch, together with a matching clang patch (https://reviews.llvm.org/D37668), implements the lowering of X86 mask set1 intrinsics to IR. Differential Revision: https://reviews.llvm.org/D37669 llvm-svn: 313625
2017-09-19 19:03:06 +08:00
; CHECK-NEXT: vpaddw %xmm0, %xmm1, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf1,0xfd,0xc0]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <8 x i16> @llvm.x86.avx512.mask.pbroadcast.w.gpr.128(i16 %x0, <8 x i16> %x1, i8 -1)
%res1 = call <8 x i16> @llvm.x86.avx512.mask.pbroadcast.w.gpr.128(i16 %x0, <8 x i16> %x1, i8 %mask)
%res2 = call <8 x i16> @llvm.x86.avx512.mask.pbroadcast.w.gpr.128(i16 %x0, <8 x i16> zeroinitializer, i8 %mask)
%res3 = add <8 x i16> %res, %res1
%res4 = add <8 x i16> %res2, %res3
ret <8 x i16> %res4
}
declare <32 x i8> @llvm.x86.avx512.mask.pbroadcast.b.gpr.256(i8, <32 x i8>, i32)
define <32 x i8>@test_int_x86_avx512_mask_pbroadcast_b_gpr_256(i8 %x0, <32 x i8> %x1, i32 %mask) {
; CHECK-LABEL: test_int_x86_avx512_mask_pbroadcast_b_gpr_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[x86] Lowering Mask Set1 intrinsics to LLVM IR This patch, together with a matching clang patch (https://reviews.llvm.org/D37668), implements the lowering of X86 mask set1 intrinsics to IR. Differential Revision: https://reviews.llvm.org/D37669 llvm-svn: 313625
2017-09-19 19:03:06 +08:00
; CHECK-NEXT: vpbroadcastb %edi, %ymm1 ## encoding: [0x62,0xf2,0x7d,0x28,0x7a,0xcf]
; CHECK-NEXT: kmovd %esi, %k1 ## encoding: [0xc5,0xfb,0x92,0xce]
; CHECK-NEXT: vpbroadcastb %edi, %ymm0 {%k1} ## encoding: [0x62,0xf2,0x7d,0x29,0x7a,0xc7]
[X86][SKX][KNL] Updated regression tests to use -mattr instead of -mcpu flag.NFC. NFC. Updated 8 regression tests to use -mattr instead of -mcpu flag as follows: -mcpu=knl --> -mattr=+avx512f -mcpu=skx --> -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq The updates are as part of the preparation of a large commit to add all instruction scheduling for the SKX target. Reviewers: delena, zvi, RKSimon Differential Revision: https://reviews.llvm.org/D38222 Change-Id: I2381c9b5bb75ecacfca017243c22d054f6eddd14 llvm-svn: 314306
2017-09-27 22:44:15 +08:00
; CHECK-NEXT: vpbroadcastb %edi, %ymm2 {%k1} {z} ## encoding: [0x62,0xf2,0x7d,0xa9,0x7a,0xd7]
; CHECK-NEXT: vpaddb %ymm2, %ymm0, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xfd,0xfc,0xc2]
[x86] Lowering Mask Set1 intrinsics to LLVM IR This patch, together with a matching clang patch (https://reviews.llvm.org/D37668), implements the lowering of X86 mask set1 intrinsics to IR. Differential Revision: https://reviews.llvm.org/D37669 llvm-svn: 313625
2017-09-19 19:03:06 +08:00
; CHECK-NEXT: vpaddb %ymm0, %ymm1, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf5,0xfc,0xc0]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <32 x i8> @llvm.x86.avx512.mask.pbroadcast.b.gpr.256(i8 %x0, <32 x i8> %x1, i32 -1)
%res1 = call <32 x i8> @llvm.x86.avx512.mask.pbroadcast.b.gpr.256(i8 %x0, <32 x i8> %x1, i32 %mask)
%res2 = call <32 x i8> @llvm.x86.avx512.mask.pbroadcast.b.gpr.256(i8 %x0, <32 x i8> zeroinitializer, i32 %mask)
%res3 = add <32 x i8> %res, %res1
%res4 = add <32 x i8> %res2, %res3
ret <32 x i8> %res4
}
declare <16 x i16> @llvm.x86.avx512.mask.pbroadcast.w.gpr.256(i16, <16 x i16>, i16)
define <16 x i16>@test_int_x86_avx512_mask_pbroadcast_w_gpr_256(i16 %x0, <16 x i16> %x1, i16 %mask) {
; CHECK-LABEL: test_int_x86_avx512_mask_pbroadcast_w_gpr_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[x86] Lowering Mask Set1 intrinsics to LLVM IR This patch, together with a matching clang patch (https://reviews.llvm.org/D37668), implements the lowering of X86 mask set1 intrinsics to IR. Differential Revision: https://reviews.llvm.org/D37669 llvm-svn: 313625
2017-09-19 19:03:06 +08:00
; CHECK-NEXT: vpbroadcastw %edi, %ymm1 ## encoding: [0x62,0xf2,0x7d,0x28,0x7b,0xcf]
; CHECK-NEXT: kmovd %esi, %k1 ## encoding: [0xc5,0xfb,0x92,0xce]
; CHECK-NEXT: vpbroadcastw %edi, %ymm0 {%k1} ## encoding: [0x62,0xf2,0x7d,0x29,0x7b,0xc7]
[X86][SKX][KNL] Updated regression tests to use -mattr instead of -mcpu flag.NFC. NFC. Updated 8 regression tests to use -mattr instead of -mcpu flag as follows: -mcpu=knl --> -mattr=+avx512f -mcpu=skx --> -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq The updates are as part of the preparation of a large commit to add all instruction scheduling for the SKX target. Reviewers: delena, zvi, RKSimon Differential Revision: https://reviews.llvm.org/D38222 Change-Id: I2381c9b5bb75ecacfca017243c22d054f6eddd14 llvm-svn: 314306
2017-09-27 22:44:15 +08:00
; CHECK-NEXT: vpbroadcastw %edi, %ymm2 {%k1} {z} ## encoding: [0x62,0xf2,0x7d,0xa9,0x7b,0xd7]
; CHECK-NEXT: vpaddw %ymm2, %ymm0, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xfd,0xfd,0xc2]
[x86] Lowering Mask Set1 intrinsics to LLVM IR This patch, together with a matching clang patch (https://reviews.llvm.org/D37668), implements the lowering of X86 mask set1 intrinsics to IR. Differential Revision: https://reviews.llvm.org/D37669 llvm-svn: 313625
2017-09-19 19:03:06 +08:00
; CHECK-NEXT: vpaddw %ymm0, %ymm1, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf5,0xfd,0xc0]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <16 x i16> @llvm.x86.avx512.mask.pbroadcast.w.gpr.256(i16 %x0, <16 x i16> %x1, i16 -1)
%res1 = call <16 x i16> @llvm.x86.avx512.mask.pbroadcast.w.gpr.256(i16 %x0, <16 x i16> %x1, i16 %mask)
%res2 = call <16 x i16> @llvm.x86.avx512.mask.pbroadcast.w.gpr.256(i16 %x0, <16 x i16> zeroinitializer, i16 %mask)
%res3 = add <16 x i16> %res, %res1
%res4 = add <16 x i16> %res2, %res3
ret <16 x i16> %res4
}
[X86][AVX512] Autoupgrade the BROADCAST intrinsics llvm-svn: 274550
2016-07-05 21:58:47 +08:00
declare <32 x i8> @llvm.x86.avx512.pbroadcastb.256(<16 x i8>, <32 x i8>, i32)
define <32 x i8>@test_int_x86_avx512_pbroadcastb_256(<16 x i8> %x0, <32 x i8> %x1, i32 %mask) {
; CHECK-LABEL: test_int_x86_avx512_pbroadcastb_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpbroadcastb %xmm0, %ymm2 ## EVEX TO VEX Compression encoding: [0xc4,0xe2,0x7d,0x78,0xd0]
[X86][AVX512] Autoupgrade the BROADCAST intrinsics llvm-svn: 274550
2016-07-05 21:58:47 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
; CHECK-NEXT: vpbroadcastb %xmm0, %ymm1 {%k1} ## encoding: [0x62,0xf2,0x7d,0x29,0x78,0xc8]
Add LiveRangeShrink pass to shrink live range within BB. Summary: LiveRangeShrink pass moves instruction right after the definition with the same BB if the instruction and its operands all have more than one use. This pass is inexpensive and guarantees optimal live-range within BB. Reviewers: davidxl, wmi, hfinkel, MatzeB, andreadb Reviewed By: MatzeB, andreadb Subscribers: hiraditya, jyknight, sanjoy, skatkov, gberry, jholewinski, qcolombet, javed.absar, krytarowski, atrick, spatel, RKSimon, andreadb, MatzeB, mehdi_amini, mgorny, efriedma, davide, dberlin, llvm-commits Differential Revision: https://reviews.llvm.org/D32563 llvm-svn: 304371
2017-06-01 07:25:25 +08:00
; CHECK-NEXT: vpbroadcastb %xmm0, %ymm0 {%k1} {z} ## encoding: [0x62,0xf2,0x7d,0xa9,0x78,0xc0]
[X86][SKX][KNL] Updated regression tests to use -mattr instead of -mcpu flag.NFC. NFC. Updated 8 regression tests to use -mattr instead of -mcpu flag as follows: -mcpu=knl --> -mattr=+avx512f -mcpu=skx --> -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq The updates are as part of the preparation of a large commit to add all instruction scheduling for the SKX target. Reviewers: delena, zvi, RKSimon Differential Revision: https://reviews.llvm.org/D38222 Change-Id: I2381c9b5bb75ecacfca017243c22d054f6eddd14 llvm-svn: 314306
2017-09-27 22:44:15 +08:00
; CHECK-NEXT: vpaddb %ymm0, %ymm1, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf5,0xfc,0xc0]
; CHECK-NEXT: vpaddb %ymm0, %ymm2, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xed,0xfc,0xc0]
[X86][AVX512] Autoupgrade the BROADCAST intrinsics llvm-svn: 274550
2016-07-05 21:58:47 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <32 x i8> @llvm.x86.avx512.pbroadcastb.256(<16 x i8> %x0, <32 x i8> %x1, i32 -1)
%res1 = call <32 x i8> @llvm.x86.avx512.pbroadcastb.256(<16 x i8> %x0, <32 x i8> %x1, i32 %mask)
%res2 = call <32 x i8> @llvm.x86.avx512.pbroadcastb.256(<16 x i8> %x0, <32 x i8> zeroinitializer, i32 %mask)
%res3 = add <32 x i8> %res, %res1
%res4 = add <32 x i8> %res2, %res3
ret <32 x i8> %res4
}
declare <16 x i8> @llvm.x86.avx512.pbroadcastb.128(<16 x i8>, <16 x i8>, i16)
define <16 x i8>@test_int_x86_avx512_pbroadcastb_128(<16 x i8> %x0, <16 x i8> %x1, i16 %mask) {
; CHECK-LABEL: test_int_x86_avx512_pbroadcastb_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpbroadcastb %xmm0, %xmm2 ## EVEX TO VEX Compression encoding: [0xc4,0xe2,0x79,0x78,0xd0]
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
[X86][AVX512] Autoupgrade the BROADCAST intrinsics llvm-svn: 274550
2016-07-05 21:58:47 +08:00
; CHECK-NEXT: vpbroadcastb %xmm0, %xmm1 {%k1} ## encoding: [0x62,0xf2,0x7d,0x09,0x78,0xc8]
Add LiveRangeShrink pass to shrink live range within BB. Summary: LiveRangeShrink pass moves instruction right after the definition with the same BB if the instruction and its operands all have more than one use. This pass is inexpensive and guarantees optimal live-range within BB. Reviewers: davidxl, wmi, hfinkel, MatzeB, andreadb Reviewed By: MatzeB, andreadb Subscribers: hiraditya, jyknight, sanjoy, skatkov, gberry, jholewinski, qcolombet, javed.absar, krytarowski, atrick, spatel, RKSimon, andreadb, MatzeB, mehdi_amini, mgorny, efriedma, davide, dberlin, llvm-commits Differential Revision: https://reviews.llvm.org/D32563 llvm-svn: 304371
2017-06-01 07:25:25 +08:00
; CHECK-NEXT: vpbroadcastb %xmm0, %xmm0 {%k1} {z} ## encoding: [0x62,0xf2,0x7d,0x89,0x78,0xc0]
[X86][SKX][KNL] Updated regression tests to use -mattr instead of -mcpu flag.NFC. NFC. Updated 8 regression tests to use -mattr instead of -mcpu flag as follows: -mcpu=knl --> -mattr=+avx512f -mcpu=skx --> -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq The updates are as part of the preparation of a large commit to add all instruction scheduling for the SKX target. Reviewers: delena, zvi, RKSimon Differential Revision: https://reviews.llvm.org/D38222 Change-Id: I2381c9b5bb75ecacfca017243c22d054f6eddd14 llvm-svn: 314306
2017-09-27 22:44:15 +08:00
; CHECK-NEXT: vpaddb %xmm0, %xmm1, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf1,0xfc,0xc0]
; CHECK-NEXT: vpaddb %xmm0, %xmm2, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xe9,0xfc,0xc0]
[X86][AVX512] Autoupgrade the BROADCAST intrinsics llvm-svn: 274550
2016-07-05 21:58:47 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <16 x i8> @llvm.x86.avx512.pbroadcastb.128(<16 x i8> %x0, <16 x i8> %x1, i16 -1)
%res1 = call <16 x i8> @llvm.x86.avx512.pbroadcastb.128(<16 x i8> %x0, <16 x i8> %x1, i16 %mask)
%res2 = call <16 x i8> @llvm.x86.avx512.pbroadcastb.128(<16 x i8> %x0, <16 x i8> zeroinitializer, i16 %mask)
%res3 = add <16 x i8> %res, %res1
%res4 = add <16 x i8> %res2, %res3
ret <16 x i8> %res4
}
declare <16 x i16> @llvm.x86.avx512.pbroadcastw.256(<8 x i16>, <16 x i16>, i16)
define <16 x i16>@test_int_x86_avx512_pbroadcastw_256(<8 x i16> %x0, <16 x i16> %x1, i16 %mask) {
; CHECK-LABEL: test_int_x86_avx512_pbroadcastw_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpbroadcastw %xmm0, %ymm2 ## EVEX TO VEX Compression encoding: [0xc4,0xe2,0x7d,0x79,0xd0]
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
[X86][AVX512] Autoupgrade the BROADCAST intrinsics llvm-svn: 274550
2016-07-05 21:58:47 +08:00
; CHECK-NEXT: vpbroadcastw %xmm0, %ymm1 {%k1} ## encoding: [0x62,0xf2,0x7d,0x29,0x79,0xc8]
Add LiveRangeShrink pass to shrink live range within BB. Summary: LiveRangeShrink pass moves instruction right after the definition with the same BB if the instruction and its operands all have more than one use. This pass is inexpensive and guarantees optimal live-range within BB. Reviewers: davidxl, wmi, hfinkel, MatzeB, andreadb Reviewed By: MatzeB, andreadb Subscribers: hiraditya, jyknight, sanjoy, skatkov, gberry, jholewinski, qcolombet, javed.absar, krytarowski, atrick, spatel, RKSimon, andreadb, MatzeB, mehdi_amini, mgorny, efriedma, davide, dberlin, llvm-commits Differential Revision: https://reviews.llvm.org/D32563 llvm-svn: 304371
2017-06-01 07:25:25 +08:00
; CHECK-NEXT: vpbroadcastw %xmm0, %ymm0 {%k1} {z} ## encoding: [0x62,0xf2,0x7d,0xa9,0x79,0xc0]
[X86][SKX][KNL] Updated regression tests to use -mattr instead of -mcpu flag.NFC. NFC. Updated 8 regression tests to use -mattr instead of -mcpu flag as follows: -mcpu=knl --> -mattr=+avx512f -mcpu=skx --> -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq The updates are as part of the preparation of a large commit to add all instruction scheduling for the SKX target. Reviewers: delena, zvi, RKSimon Differential Revision: https://reviews.llvm.org/D38222 Change-Id: I2381c9b5bb75ecacfca017243c22d054f6eddd14 llvm-svn: 314306
2017-09-27 22:44:15 +08:00
; CHECK-NEXT: vpaddw %ymm0, %ymm1, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf5,0xfd,0xc0]
; CHECK-NEXT: vpaddw %ymm0, %ymm2, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xed,0xfd,0xc0]
[X86][AVX512] Autoupgrade the BROADCAST intrinsics llvm-svn: 274550
2016-07-05 21:58:47 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <16 x i16> @llvm.x86.avx512.pbroadcastw.256(<8 x i16> %x0, <16 x i16> %x1, i16 -1)
%res1 = call <16 x i16> @llvm.x86.avx512.pbroadcastw.256(<8 x i16> %x0, <16 x i16> %x1, i16 %mask)
%res2 = call <16 x i16> @llvm.x86.avx512.pbroadcastw.256(<8 x i16> %x0, <16 x i16> zeroinitializer, i16 %mask)
%res3 = add <16 x i16> %res, %res1
%res4 = add <16 x i16> %res2, %res3
ret <16 x i16> %res4
}
declare <8 x i16> @llvm.x86.avx512.pbroadcastw.128(<8 x i16>, <8 x i16>, i8)
define <8 x i16>@test_int_x86_avx512_pbroadcastw_128(<8 x i16> %x0, <8 x i16> %x1, i8 %mask) {
; CHECK-LABEL: test_int_x86_avx512_pbroadcastw_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpbroadcastw %xmm0, %xmm2 ## EVEX TO VEX Compression encoding: [0xc4,0xe2,0x79,0x79,0xd0]
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
[X86][AVX512] Autoupgrade the BROADCAST intrinsics llvm-svn: 274550
2016-07-05 21:58:47 +08:00
; CHECK-NEXT: vpbroadcastw %xmm0, %xmm1 {%k1} ## encoding: [0x62,0xf2,0x7d,0x09,0x79,0xc8]
Add LiveRangeShrink pass to shrink live range within BB. Summary: LiveRangeShrink pass moves instruction right after the definition with the same BB if the instruction and its operands all have more than one use. This pass is inexpensive and guarantees optimal live-range within BB. Reviewers: davidxl, wmi, hfinkel, MatzeB, andreadb Reviewed By: MatzeB, andreadb Subscribers: hiraditya, jyknight, sanjoy, skatkov, gberry, jholewinski, qcolombet, javed.absar, krytarowski, atrick, spatel, RKSimon, andreadb, MatzeB, mehdi_amini, mgorny, efriedma, davide, dberlin, llvm-commits Differential Revision: https://reviews.llvm.org/D32563 llvm-svn: 304371
2017-06-01 07:25:25 +08:00
; CHECK-NEXT: vpbroadcastw %xmm0, %xmm0 {%k1} {z} ## encoding: [0x62,0xf2,0x7d,0x89,0x79,0xc0]
[X86][SKX][KNL] Updated regression tests to use -mattr instead of -mcpu flag.NFC. NFC. Updated 8 regression tests to use -mattr instead of -mcpu flag as follows: -mcpu=knl --> -mattr=+avx512f -mcpu=skx --> -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq The updates are as part of the preparation of a large commit to add all instruction scheduling for the SKX target. Reviewers: delena, zvi, RKSimon Differential Revision: https://reviews.llvm.org/D38222 Change-Id: I2381c9b5bb75ecacfca017243c22d054f6eddd14 llvm-svn: 314306
2017-09-27 22:44:15 +08:00
; CHECK-NEXT: vpaddw %xmm0, %xmm1, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf1,0xfd,0xc0]
; CHECK-NEXT: vpaddw %xmm0, %xmm2, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xe9,0xfd,0xc0]
[X86][AVX512] Autoupgrade the BROADCAST intrinsics llvm-svn: 274550
2016-07-05 21:58:47 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <8 x i16> @llvm.x86.avx512.pbroadcastw.128(<8 x i16> %x0, <8 x i16> %x1, i8 -1)
%res1 = call <8 x i16> @llvm.x86.avx512.pbroadcastw.128(<8 x i16> %x0, <8 x i16> %x1, i8 %mask)
%res2 = call <8 x i16> @llvm.x86.avx512.pbroadcastw.128(<8 x i16> %x0, <8 x i16> zeroinitializer, i8 %mask)
%res3 = add <8 x i16> %res, %res1
%res4 = add <8 x i16> %res2, %res3
ret <8 x i16> %res4
}
declare <64 x i8> @llvm.x86.avx512.pbroadcastb.512(<16 x i8>, <64 x i8>, i64)
define <64 x i8>@test_int_x86_avx512_pbroadcastb_512(<16 x i8> %x0, <64 x i8> %x1, i64 %mask) {
; CHECK-LABEL: test_int_x86_avx512_pbroadcastb_512:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[X86][AVX512] Autoupgrade the BROADCAST intrinsics llvm-svn: 274550
2016-07-05 21:58:47 +08:00
; CHECK-NEXT: vpbroadcastb %xmm0, %zmm2 ## encoding: [0x62,0xf2,0x7d,0x48,0x78,0xd0]
; CHECK-NEXT: kmovq %rdi, %k1 ## encoding: [0xc4,0xe1,0xfb,0x92,0xcf]
; CHECK-NEXT: vpbroadcastb %xmm0, %zmm1 {%k1} ## encoding: [0x62,0xf2,0x7d,0x49,0x78,0xc8]
Add LiveRangeShrink pass to shrink live range within BB. Summary: LiveRangeShrink pass moves instruction right after the definition with the same BB if the instruction and its operands all have more than one use. This pass is inexpensive and guarantees optimal live-range within BB. Reviewers: davidxl, wmi, hfinkel, MatzeB, andreadb Reviewed By: MatzeB, andreadb Subscribers: hiraditya, jyknight, sanjoy, skatkov, gberry, jholewinski, qcolombet, javed.absar, krytarowski, atrick, spatel, RKSimon, andreadb, MatzeB, mehdi_amini, mgorny, efriedma, davide, dberlin, llvm-commits Differential Revision: https://reviews.llvm.org/D32563 llvm-svn: 304371
2017-06-01 07:25:25 +08:00
; CHECK-NEXT: vpbroadcastb %xmm0, %zmm0 {%k1} {z} ## encoding: [0x62,0xf2,0x7d,0xc9,0x78,0xc0]
[X86][SKX][KNL] Updated regression tests to use -mattr instead of -mcpu flag.NFC. NFC. Updated 8 regression tests to use -mattr instead of -mcpu flag as follows: -mcpu=knl --> -mattr=+avx512f -mcpu=skx --> -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq The updates are as part of the preparation of a large commit to add all instruction scheduling for the SKX target. Reviewers: delena, zvi, RKSimon Differential Revision: https://reviews.llvm.org/D38222 Change-Id: I2381c9b5bb75ecacfca017243c22d054f6eddd14 llvm-svn: 314306
2017-09-27 22:44:15 +08:00
; CHECK-NEXT: vpaddb %zmm0, %zmm1, %zmm0 ## encoding: [0x62,0xf1,0x75,0x48,0xfc,0xc0]
; CHECK-NEXT: vpaddb %zmm0, %zmm2, %zmm0 ## encoding: [0x62,0xf1,0x6d,0x48,0xfc,0xc0]
[X86][AVX512] Autoupgrade the BROADCAST intrinsics llvm-svn: 274550
2016-07-05 21:58:47 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <64 x i8> @llvm.x86.avx512.pbroadcastb.512(<16 x i8> %x0, <64 x i8> %x1, i64 -1)
%res1 = call <64 x i8> @llvm.x86.avx512.pbroadcastb.512(<16 x i8> %x0, <64 x i8> %x1, i64 %mask)
%res2 = call <64 x i8> @llvm.x86.avx512.pbroadcastb.512(<16 x i8> %x0, <64 x i8> zeroinitializer, i64 %mask)
%res3 = add <64 x i8> %res, %res1
%res4 = add <64 x i8> %res2, %res3
ret <64 x i8> %res4
}
declare <32 x i16> @llvm.x86.avx512.pbroadcastw.512(<8 x i16>, <32 x i16>, i32)
define <32 x i16>@test_int_x86_avx512_pbroadcastw_512(<8 x i16> %x0, <32 x i16> %x1, i32 %mask) {
; CHECK-LABEL: test_int_x86_avx512_pbroadcastw_512:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[X86][AVX512] Autoupgrade the BROADCAST intrinsics llvm-svn: 274550
2016-07-05 21:58:47 +08:00
; CHECK-NEXT: vpbroadcastw %xmm0, %zmm2 ## encoding: [0x62,0xf2,0x7d,0x48,0x79,0xd0]
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
; CHECK-NEXT: vpbroadcastw %xmm0, %zmm1 {%k1} ## encoding: [0x62,0xf2,0x7d,0x49,0x79,0xc8]
Add LiveRangeShrink pass to shrink live range within BB. Summary: LiveRangeShrink pass moves instruction right after the definition with the same BB if the instruction and its operands all have more than one use. This pass is inexpensive and guarantees optimal live-range within BB. Reviewers: davidxl, wmi, hfinkel, MatzeB, andreadb Reviewed By: MatzeB, andreadb Subscribers: hiraditya, jyknight, sanjoy, skatkov, gberry, jholewinski, qcolombet, javed.absar, krytarowski, atrick, spatel, RKSimon, andreadb, MatzeB, mehdi_amini, mgorny, efriedma, davide, dberlin, llvm-commits Differential Revision: https://reviews.llvm.org/D32563 llvm-svn: 304371
2017-06-01 07:25:25 +08:00
; CHECK-NEXT: vpbroadcastw %xmm0, %zmm0 {%k1} {z} ## encoding: [0x62,0xf2,0x7d,0xc9,0x79,0xc0]
[X86][SKX][KNL] Updated regression tests to use -mattr instead of -mcpu flag.NFC. NFC. Updated 8 regression tests to use -mattr instead of -mcpu flag as follows: -mcpu=knl --> -mattr=+avx512f -mcpu=skx --> -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq The updates are as part of the preparation of a large commit to add all instruction scheduling for the SKX target. Reviewers: delena, zvi, RKSimon Differential Revision: https://reviews.llvm.org/D38222 Change-Id: I2381c9b5bb75ecacfca017243c22d054f6eddd14 llvm-svn: 314306
2017-09-27 22:44:15 +08:00
; CHECK-NEXT: vpaddw %zmm0, %zmm1, %zmm0 ## encoding: [0x62,0xf1,0x75,0x48,0xfd,0xc0]
; CHECK-NEXT: vpaddw %zmm0, %zmm2, %zmm0 ## encoding: [0x62,0xf1,0x6d,0x48,0xfd,0xc0]
[X86][AVX512] Autoupgrade the BROADCAST intrinsics llvm-svn: 274550
2016-07-05 21:58:47 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <32 x i16> @llvm.x86.avx512.pbroadcastw.512(<8 x i16> %x0, <32 x i16> %x1, i32 -1)
%res1 = call <32 x i16> @llvm.x86.avx512.pbroadcastw.512(<8 x i16> %x0, <32 x i16> %x1, i32 %mask)
%res2 = call <32 x i16> @llvm.x86.avx512.pbroadcastw.512(<8 x i16> %x0, <32 x i16> zeroinitializer, i32 %mask)
%res3 = add <32 x i16> %res, %res1
%res4 = add <32 x i16> %res2, %res3
ret <32 x i16> %res4
}
[AVX512] Remove masked store intrinsics. Clang now emits generic masked store intrinsics instead. The intrinsics will be autoupgraded to the same generic masked stores. llvm-svn: 271245
2016-05-31 09:50:02 +08:00
declare void @llvm.x86.avx512.mask.storeu.b.128(i8*, <16 x i8>, i16)
define void@test_int_x86_avx512_mask_storeu_b_128(i8* %ptr1, i8* %ptr2, <16 x i8> %x1, i16 %x2) {
; CHECK-LABEL: test_int_x86_avx512_mask_storeu_b_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edx, %k1 ## encoding: [0xc5,0xfb,0x92,0xca]
[AVX512] Remove masked store intrinsics. Clang now emits generic masked store intrinsics instead. The intrinsics will be autoupgraded to the same generic masked stores. llvm-svn: 271245
2016-05-31 09:50:02 +08:00
; CHECK-NEXT: vmovdqu8 %xmm0, (%rdi) {%k1} ## encoding: [0x62,0xf1,0x7f,0x09,0x7f,0x07]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vmovdqu %xmm0, (%rsi) ## EVEX TO VEX Compression encoding: [0xc5,0xfa,0x7f,0x06]
[AVX512] Remove masked store intrinsics. Clang now emits generic masked store intrinsics instead. The intrinsics will be autoupgraded to the same generic masked stores. llvm-svn: 271245
2016-05-31 09:50:02 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
call void @llvm.x86.avx512.mask.storeu.b.128(i8* %ptr1, <16 x i8> %x1, i16 %x2)
call void @llvm.x86.avx512.mask.storeu.b.128(i8* %ptr2, <16 x i8> %x1, i16 -1)
ret void
}
declare void @llvm.x86.avx512.mask.storeu.b.256(i8*, <32 x i8>, i32)
define void@test_int_x86_avx512_mask_storeu_b_256(i8* %ptr1, i8* %ptr2, <32 x i8> %x1, i32 %x2) {
; CHECK-LABEL: test_int_x86_avx512_mask_storeu_b_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX512] Remove masked store intrinsics. Clang now emits generic masked store intrinsics instead. The intrinsics will be autoupgraded to the same generic masked stores. llvm-svn: 271245
2016-05-31 09:50:02 +08:00
; CHECK-NEXT: kmovd %edx, %k1 ## encoding: [0xc5,0xfb,0x92,0xca]
; CHECK-NEXT: vmovdqu8 %ymm0, (%rdi) {%k1} ## encoding: [0x62,0xf1,0x7f,0x29,0x7f,0x07]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vmovdqu %ymm0, (%rsi) ## EVEX TO VEX Compression encoding: [0xc5,0xfe,0x7f,0x06]
[X86][SKX][KNL] Updated regression tests to use -mattr instead of -mcpu flag.NFC. NFC. Updated 8 regression tests to use -mattr instead of -mcpu flag as follows: -mcpu=knl --> -mattr=+avx512f -mcpu=skx --> -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq The updates are as part of the preparation of a large commit to add all instruction scheduling for the SKX target. Reviewers: delena, zvi, RKSimon Differential Revision: https://reviews.llvm.org/D38222 Change-Id: I2381c9b5bb75ecacfca017243c22d054f6eddd14 llvm-svn: 314306
2017-09-27 22:44:15 +08:00
; CHECK-NEXT: vzeroupper ## encoding: [0xc5,0xf8,0x77]
[AVX512] Remove masked store intrinsics. Clang now emits generic masked store intrinsics instead. The intrinsics will be autoupgraded to the same generic masked stores. llvm-svn: 271245
2016-05-31 09:50:02 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
call void @llvm.x86.avx512.mask.storeu.b.256(i8* %ptr1, <32 x i8> %x1, i32 %x2)
call void @llvm.x86.avx512.mask.storeu.b.256(i8* %ptr2, <32 x i8> %x1, i32 -1)
ret void
}
declare void @llvm.x86.avx512.mask.storeu.w.128(i8*, <8 x i16>, i8)
define void@test_int_x86_avx512_mask_storeu_w_128(i8* %ptr1, i8* %ptr2, <8 x i16> %x1, i8 %x2) {
; CHECK-LABEL: test_int_x86_avx512_mask_storeu_w_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edx, %k1 ## encoding: [0xc5,0xfb,0x92,0xca]
[AVX512] Remove masked store intrinsics. Clang now emits generic masked store intrinsics instead. The intrinsics will be autoupgraded to the same generic masked stores. llvm-svn: 271245
2016-05-31 09:50:02 +08:00
; CHECK-NEXT: vmovdqu16 %xmm0, (%rdi) {%k1} ## encoding: [0x62,0xf1,0xff,0x09,0x7f,0x07]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vmovdqu %xmm0, (%rsi) ## EVEX TO VEX Compression encoding: [0xc5,0xfa,0x7f,0x06]
[AVX512] Remove masked store intrinsics. Clang now emits generic masked store intrinsics instead. The intrinsics will be autoupgraded to the same generic masked stores. llvm-svn: 271245
2016-05-31 09:50:02 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
call void @llvm.x86.avx512.mask.storeu.w.128(i8* %ptr1, <8 x i16> %x1, i8 %x2)
call void @llvm.x86.avx512.mask.storeu.w.128(i8* %ptr2, <8 x i16> %x1, i8 -1)
ret void
}
declare void @llvm.x86.avx512.mask.storeu.w.256(i8*, <16 x i16>, i16)
define void@test_int_x86_avx512_mask_storeu_w_256(i8* %ptr1, i8* %ptr2, <16 x i16> %x1, i16 %x2) {
; CHECK-LABEL: test_int_x86_avx512_mask_storeu_w_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edx, %k1 ## encoding: [0xc5,0xfb,0x92,0xca]
[AVX512] Remove masked store intrinsics. Clang now emits generic masked store intrinsics instead. The intrinsics will be autoupgraded to the same generic masked stores. llvm-svn: 271245
2016-05-31 09:50:02 +08:00
; CHECK-NEXT: vmovdqu16 %ymm0, (%rdi) {%k1} ## encoding: [0x62,0xf1,0xff,0x29,0x7f,0x07]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vmovdqu %ymm0, (%rsi) ## EVEX TO VEX Compression encoding: [0xc5,0xfe,0x7f,0x06]
[X86][SKX][KNL] Updated regression tests to use -mattr instead of -mcpu flag.NFC. NFC. Updated 8 regression tests to use -mattr instead of -mcpu flag as follows: -mcpu=knl --> -mattr=+avx512f -mcpu=skx --> -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq The updates are as part of the preparation of a large commit to add all instruction scheduling for the SKX target. Reviewers: delena, zvi, RKSimon Differential Revision: https://reviews.llvm.org/D38222 Change-Id: I2381c9b5bb75ecacfca017243c22d054f6eddd14 llvm-svn: 314306
2017-09-27 22:44:15 +08:00
; CHECK-NEXT: vzeroupper ## encoding: [0xc5,0xf8,0x77]
[AVX512] Remove masked store intrinsics. Clang now emits generic masked store intrinsics instead. The intrinsics will be autoupgraded to the same generic masked stores. llvm-svn: 271245
2016-05-31 09:50:02 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
call void @llvm.x86.avx512.mask.storeu.w.256(i8* %ptr1, <16 x i16> %x1, i16 %x2)
call void @llvm.x86.avx512.mask.storeu.w.256(i8* %ptr2, <16 x i16> %x1, i16 -1)
ret void
}
[AVX512] Remove masked load intrinsics. Clang now emits generic masked load intrinsics instead. The intrinsics will be autoupgraded to the same generic masked loads. llvm-svn: 271478
2016-06-02 12:19:36 +08:00
declare <8 x i16> @llvm.x86.avx512.mask.loadu.w.128(i8*, <8 x i16>, i8)
define <8 x i16>@test_int_x86_avx512_mask_loadu_w_128(i8* %ptr, i8* %ptr2, <8 x i16> %x1, i8 %mask) {
; CHECK-LABEL: test_int_x86_avx512_mask_loadu_w_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vmovdqu (%rdi), %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xfa,0x6f,0x07]
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edx, %k1 ## encoding: [0xc5,0xfb,0x92,0xca]
[AVX512] Remove masked load intrinsics. Clang now emits generic masked load intrinsics instead. The intrinsics will be autoupgraded to the same generic masked loads. llvm-svn: 271478
2016-06-02 12:19:36 +08:00
; CHECK-NEXT: vmovdqu16 (%rsi), %xmm0 {%k1} ## encoding: [0x62,0xf1,0xff,0x09,0x6f,0x06]
; CHECK-NEXT: vmovdqu16 (%rdi), %xmm1 {%k1} {z} ## encoding: [0x62,0xf1,0xff,0x89,0x6f,0x0f]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpaddw %xmm1, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xfd,0xc1]
[AVX512] Remove masked load intrinsics. Clang now emits generic masked load intrinsics instead. The intrinsics will be autoupgraded to the same generic masked loads. llvm-svn: 271478
2016-06-02 12:19:36 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res0 = call <8 x i16> @llvm.x86.avx512.mask.loadu.w.128(i8* %ptr, <8 x i16> %x1, i8 -1)
%res = call <8 x i16> @llvm.x86.avx512.mask.loadu.w.128(i8* %ptr2, <8 x i16> %res0, i8 %mask)
%res1 = call <8 x i16> @llvm.x86.avx512.mask.loadu.w.128(i8* %ptr, <8 x i16> zeroinitializer, i8 %mask)
%res2 = add <8 x i16> %res, %res1
ret <8 x i16> %res2
}
declare <16 x i16> @llvm.x86.avx512.mask.loadu.w.256(i8*, <16 x i16>, i16)
define <16 x i16>@test_int_x86_avx512_mask_loadu_w_256(i8* %ptr, i8* %ptr2, <16 x i16> %x1, i16 %mask) {
; CHECK-LABEL: test_int_x86_avx512_mask_loadu_w_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vmovdqu (%rdi), %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xfe,0x6f,0x07]
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edx, %k1 ## encoding: [0xc5,0xfb,0x92,0xca]
[AVX512] Remove masked load intrinsics. Clang now emits generic masked load intrinsics instead. The intrinsics will be autoupgraded to the same generic masked loads. llvm-svn: 271478
2016-06-02 12:19:36 +08:00
; CHECK-NEXT: vmovdqu16 (%rsi), %ymm0 {%k1} ## encoding: [0x62,0xf1,0xff,0x29,0x6f,0x06]
; CHECK-NEXT: vmovdqu16 (%rdi), %ymm1 {%k1} {z} ## encoding: [0x62,0xf1,0xff,0xa9,0x6f,0x0f]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpaddw %ymm1, %ymm0, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xfd,0xfd,0xc1]
[AVX512] Remove masked load intrinsics. Clang now emits generic masked load intrinsics instead. The intrinsics will be autoupgraded to the same generic masked loads. llvm-svn: 271478
2016-06-02 12:19:36 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res0 = call <16 x i16> @llvm.x86.avx512.mask.loadu.w.256(i8* %ptr, <16 x i16> %x1, i16 -1)
%res = call <16 x i16> @llvm.x86.avx512.mask.loadu.w.256(i8* %ptr2, <16 x i16> %res0, i16 %mask)
%res1 = call <16 x i16> @llvm.x86.avx512.mask.loadu.w.256(i8* %ptr, <16 x i16> zeroinitializer, i16 %mask)
%res2 = add <16 x i16> %res, %res1
ret <16 x i16> %res2
}
declare <16 x i8> @llvm.x86.avx512.mask.loadu.b.128(i8*, <16 x i8>, i16)
define <16 x i8>@test_int_x86_avx512_mask_loadu_b_128(i8* %ptr, i8* %ptr2, <16 x i8> %x1, i16 %mask) {
; CHECK-LABEL: test_int_x86_avx512_mask_loadu_b_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vmovdqu (%rdi), %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xfa,0x6f,0x07]
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edx, %k1 ## encoding: [0xc5,0xfb,0x92,0xca]
[AVX512] Remove masked load intrinsics. Clang now emits generic masked load intrinsics instead. The intrinsics will be autoupgraded to the same generic masked loads. llvm-svn: 271478
2016-06-02 12:19:36 +08:00
; CHECK-NEXT: vmovdqu8 (%rsi), %xmm0 {%k1} ## encoding: [0x62,0xf1,0x7f,0x09,0x6f,0x06]
; CHECK-NEXT: vmovdqu8 (%rdi), %xmm1 {%k1} {z} ## encoding: [0x62,0xf1,0x7f,0x89,0x6f,0x0f]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpaddb %xmm1, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xfc,0xc1]
[AVX512] Remove masked load intrinsics. Clang now emits generic masked load intrinsics instead. The intrinsics will be autoupgraded to the same generic masked loads. llvm-svn: 271478
2016-06-02 12:19:36 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res0 = call <16 x i8> @llvm.x86.avx512.mask.loadu.b.128(i8* %ptr, <16 x i8> %x1, i16 -1)
%res = call <16 x i8> @llvm.x86.avx512.mask.loadu.b.128(i8* %ptr2, <16 x i8> %res0, i16 %mask)
%res1 = call <16 x i8> @llvm.x86.avx512.mask.loadu.b.128(i8* %ptr, <16 x i8> zeroinitializer, i16 %mask)
%res2 = add <16 x i8> %res, %res1
ret <16 x i8> %res2
}
declare <32 x i8> @llvm.x86.avx512.mask.loadu.b.256(i8*, <32 x i8>, i32)
define <32 x i8>@test_int_x86_avx512_mask_loadu_b_256(i8* %ptr, i8* %ptr2, <32 x i8> %x1, i32 %mask) {
; CHECK-LABEL: test_int_x86_avx512_mask_loadu_b_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vmovdqu (%rdi), %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xfe,0x6f,0x07]
[AVX512] Remove masked load intrinsics. Clang now emits generic masked load intrinsics instead. The intrinsics will be autoupgraded to the same generic masked loads. llvm-svn: 271478
2016-06-02 12:19:36 +08:00
; CHECK-NEXT: kmovd %edx, %k1 ## encoding: [0xc5,0xfb,0x92,0xca]
; CHECK-NEXT: vmovdqu8 (%rsi), %ymm0 {%k1} ## encoding: [0x62,0xf1,0x7f,0x29,0x6f,0x06]
; CHECK-NEXT: vmovdqu8 (%rdi), %ymm1 {%k1} {z} ## encoding: [0x62,0xf1,0x7f,0xa9,0x6f,0x0f]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpaddb %ymm1, %ymm0, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xfd,0xfc,0xc1]
[AVX512] Remove masked load intrinsics. Clang now emits generic masked load intrinsics instead. The intrinsics will be autoupgraded to the same generic masked loads. llvm-svn: 271478
2016-06-02 12:19:36 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res0 = call <32 x i8> @llvm.x86.avx512.mask.loadu.b.256(i8* %ptr, <32 x i8> %x1, i32 -1)
%res = call <32 x i8> @llvm.x86.avx512.mask.loadu.b.256(i8* %ptr2, <32 x i8> %res0, i32 %mask)
%res1 = call <32 x i8> @llvm.x86.avx512.mask.loadu.b.256(i8* %ptr, <32 x i8> zeroinitializer, i32 %mask)
%res2 = add <32 x i8> %res, %res1
ret <32 x i8> %res2
}
[AVX512] Remove masked palignr intrinsics and auto-upgrade them to native IR of vector shuffle and select. llvm-svn: 271872
2016-06-06 14:12:54 +08:00
declare <16 x i8> @llvm.x86.avx512.mask.palignr.128(<16 x i8>, <16 x i8>, i32, <16 x i8>, i16)
define <16 x i8>@test_int_x86_avx512_mask_palignr_128(<16 x i8> %x0, <16 x i8> %x1, <16 x i8> %x3, i16 %x4) {
; CHECK-LABEL: test_int_x86_avx512_mask_palignr_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpalignr $2, %xmm1, %xmm0, %xmm3 ## EVEX TO VEX Compression encoding: [0xc4,0xe3,0x79,0x0f,0xd9,0x02]
[X86] Fix a test I failed to re-generate in r272249. llvm-svn: 272250
2016-06-09 15:10:34 +08:00
; CHECK-NEXT: ## xmm3 = xmm1[2,3,4,5,6,7,8,9,10,11,12,13,14,15],xmm0[0,1]
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
[AVX512] Remove masked palignr intrinsics and auto-upgrade them to native IR of vector shuffle and select. llvm-svn: 271872
2016-06-06 14:12:54 +08:00
; CHECK-NEXT: vpalignr $2, %xmm1, %xmm0, %xmm2 {%k1} ## encoding: [0x62,0xf3,0x7d,0x09,0x0f,0xd1,0x02]
[X86][AVX512] Add support for VPALIGNR/PSHUFD/PSHUFHW/PSHUFLW masked shuffle comments llvm-svn: 274466
2016-07-03 23:00:51 +08:00
; CHECK-NEXT: ## xmm2 {%k1} = xmm1[2,3,4,5,6,7,8,9,10,11,12,13,14,15],xmm0[0,1]
[AVX512] Remove masked palignr intrinsics and auto-upgrade them to native IR of vector shuffle and select. llvm-svn: 271872
2016-06-06 14:12:54 +08:00
; CHECK-NEXT: vpalignr $2, %xmm1, %xmm0, %xmm0 {%k1} {z} ## encoding: [0x62,0xf3,0x7d,0x89,0x0f,0xc1,0x02]
[X86][AVX512] Add support for VPALIGNR/PSHUFD/PSHUFHW/PSHUFLW masked shuffle comments llvm-svn: 274466
2016-07-03 23:00:51 +08:00
; CHECK-NEXT: ## xmm0 {%k1} {z} = xmm1[2,3,4,5,6,7,8,9,10,11,12,13,14,15],xmm0[0,1]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpaddb %xmm3, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xfc,0xc3]
[X86][SKX][KNL] Updated regression tests to use -mattr instead of -mcpu flag.NFC. NFC. Updated 8 regression tests to use -mattr instead of -mcpu flag as follows: -mcpu=knl --> -mattr=+avx512f -mcpu=skx --> -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq The updates are as part of the preparation of a large commit to add all instruction scheduling for the SKX target. Reviewers: delena, zvi, RKSimon Differential Revision: https://reviews.llvm.org/D38222 Change-Id: I2381c9b5bb75ecacfca017243c22d054f6eddd14 llvm-svn: 314306
2017-09-27 22:44:15 +08:00
; CHECK-NEXT: vpaddb %xmm0, %xmm2, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xe9,0xfc,0xc0]
[AVX512] Remove masked palignr intrinsics and auto-upgrade them to native IR of vector shuffle and select. llvm-svn: 271872
2016-06-06 14:12:54 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <16 x i8> @llvm.x86.avx512.mask.palignr.128(<16 x i8> %x0, <16 x i8> %x1, i32 2, <16 x i8> %x3, i16 %x4)
%res1 = call <16 x i8> @llvm.x86.avx512.mask.palignr.128(<16 x i8> %x0, <16 x i8> %x1, i32 2, <16 x i8> zeroinitializer, i16 %x4)
%res2 = call <16 x i8> @llvm.x86.avx512.mask.palignr.128(<16 x i8> %x0, <16 x i8> %x1, i32 2, <16 x i8> %x3, i16 -1)
%res3 = add <16 x i8> %res, %res1
%res4 = add <16 x i8> %res3, %res2
ret <16 x i8> %res4
}
declare <32 x i8> @llvm.x86.avx512.mask.palignr.256(<32 x i8>, <32 x i8>, i32, <32 x i8>, i32)
define <32 x i8>@test_int_x86_avx512_mask_palignr_256(<32 x i8> %x0, <32 x i8> %x1, <32 x i8> %x3, i32 %x4) {
; CHECK-LABEL: test_int_x86_avx512_mask_palignr_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpalignr $2, %ymm1, %ymm0, %ymm3 ## EVEX TO VEX Compression encoding: [0xc4,0xe3,0x7d,0x0f,0xd9,0x02]
[X86] Fix a test I failed to re-generate in r272249. llvm-svn: 272250
2016-06-09 15:10:34 +08:00
; CHECK-NEXT: ## ymm3 = ymm1[2,3,4,5,6,7,8,9,10,11,12,13,14,15],ymm0[0,1],ymm1[18,19,20,21,22,23,24,25,26,27,28,29,30,31],ymm0[16,17]
[AVX512] Remove masked palignr intrinsics and auto-upgrade them to native IR of vector shuffle and select. llvm-svn: 271872
2016-06-06 14:12:54 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
; CHECK-NEXT: vpalignr $2, %ymm1, %ymm0, %ymm2 {%k1} ## encoding: [0x62,0xf3,0x7d,0x29,0x0f,0xd1,0x02]
[X86][AVX512] Add support for VPALIGNR/PSHUFD/PSHUFHW/PSHUFLW masked shuffle comments llvm-svn: 274466
2016-07-03 23:00:51 +08:00
; CHECK-NEXT: ## ymm2 {%k1} = ymm1[2,3,4,5,6,7,8,9,10,11,12,13,14,15],ymm0[0,1],ymm1[18,19,20,21,22,23,24,25,26,27,28,29,30,31],ymm0[16,17]
[AVX512] Remove masked palignr intrinsics and auto-upgrade them to native IR of vector shuffle and select. llvm-svn: 271872
2016-06-06 14:12:54 +08:00
; CHECK-NEXT: vpalignr $2, %ymm1, %ymm0, %ymm0 {%k1} {z} ## encoding: [0x62,0xf3,0x7d,0xa9,0x0f,0xc1,0x02]
[X86][AVX512] Add support for VPALIGNR/PSHUFD/PSHUFHW/PSHUFLW masked shuffle comments llvm-svn: 274466
2016-07-03 23:00:51 +08:00
; CHECK-NEXT: ## ymm0 {%k1} {z} = ymm1[2,3,4,5,6,7,8,9,10,11,12,13,14,15],ymm0[0,1],ymm1[18,19,20,21,22,23,24,25,26,27,28,29,30,31],ymm0[16,17]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpaddb %ymm3, %ymm0, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xfd,0xfc,0xc3]
[X86][SKX][KNL] Updated regression tests to use -mattr instead of -mcpu flag.NFC. NFC. Updated 8 regression tests to use -mattr instead of -mcpu flag as follows: -mcpu=knl --> -mattr=+avx512f -mcpu=skx --> -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq The updates are as part of the preparation of a large commit to add all instruction scheduling for the SKX target. Reviewers: delena, zvi, RKSimon Differential Revision: https://reviews.llvm.org/D38222 Change-Id: I2381c9b5bb75ecacfca017243c22d054f6eddd14 llvm-svn: 314306
2017-09-27 22:44:15 +08:00
; CHECK-NEXT: vpaddb %ymm0, %ymm2, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xed,0xfc,0xc0]
[AVX512] Remove masked palignr intrinsics and auto-upgrade them to native IR of vector shuffle and select. llvm-svn: 271872
2016-06-06 14:12:54 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <32 x i8> @llvm.x86.avx512.mask.palignr.256(<32 x i8> %x0, <32 x i8> %x1, i32 2, <32 x i8> %x3, i32 %x4)
%res1 = call <32 x i8> @llvm.x86.avx512.mask.palignr.256(<32 x i8> %x0, <32 x i8> %x1, i32 2, <32 x i8> zeroinitializer, i32 %x4)
%res2 = call <32 x i8> @llvm.x86.avx512.mask.palignr.256(<32 x i8> %x0, <32 x i8> %x1, i32 2, <32 x i8> %x3, i32 -1)
%res3 = add <32 x i8> %res, %res1
%res4 = add <32 x i8> %res3, %res2
ret <32 x i8> %res4
}
[AVX512] Remove maksed pshufd, pshuflw, and phufhw intrinsics and autoupgrade them to selects and shufflevector. llvm-svn: 272527
2016-06-13 10:36:48 +08:00
declare <8 x i16> @llvm.x86.avx512.mask.pshufh.w.128(<8 x i16>, i32, <8 x i16>, i8)
define <8 x i16>@test_int_x86_avx512_mask_pshufh_w_128(<8 x i16> %x0, i32 %x1, <8 x i16> %x2, i8 %x3) {
; CHECK-LABEL: test_int_x86_avx512_mask_pshufh_w_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpshufhw $3, %xmm0, %xmm2 ## EVEX TO VEX Compression encoding: [0xc5,0xfa,0x70,0xd0,0x03]
[AVX512] Remove maksed pshufd, pshuflw, and phufhw intrinsics and autoupgrade them to selects and shufflevector. llvm-svn: 272527
2016-06-13 10:36:48 +08:00
; CHECK-NEXT: ## xmm2 = xmm0[0,1,2,3,7,4,4,4]
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %esi, %k1 ## encoding: [0xc5,0xfb,0x92,0xce]
[AVX512] Remove maksed pshufd, pshuflw, and phufhw intrinsics and autoupgrade them to selects and shufflevector. llvm-svn: 272527
2016-06-13 10:36:48 +08:00
; CHECK-NEXT: vpshufhw $3, %xmm0, %xmm1 {%k1} ## encoding: [0x62,0xf1,0x7e,0x09,0x70,0xc8,0x03]
[X86][AVX512] Add support for VPALIGNR/PSHUFD/PSHUFHW/PSHUFLW masked shuffle comments llvm-svn: 274466
2016-07-03 23:00:51 +08:00
; CHECK-NEXT: ## xmm1 {%k1} = xmm0[0,1,2,3,7,4,4,4]
[AVX512] Remove maksed pshufd, pshuflw, and phufhw intrinsics and autoupgrade them to selects and shufflevector. llvm-svn: 272527
2016-06-13 10:36:48 +08:00
; CHECK-NEXT: vpshufhw $3, %xmm0, %xmm0 {%k1} {z} ## encoding: [0x62,0xf1,0x7e,0x89,0x70,0xc0,0x03]
[X86][AVX512] Add support for VPALIGNR/PSHUFD/PSHUFHW/PSHUFLW masked shuffle comments llvm-svn: 274466
2016-07-03 23:00:51 +08:00
; CHECK-NEXT: ## xmm0 {%k1} {z} = xmm0[0,1,2,3,7,4,4,4]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpaddw %xmm2, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xfd,0xc2]
[X86][SKX][KNL] Updated regression tests to use -mattr instead of -mcpu flag.NFC. NFC. Updated 8 regression tests to use -mattr instead of -mcpu flag as follows: -mcpu=knl --> -mattr=+avx512f -mcpu=skx --> -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq The updates are as part of the preparation of a large commit to add all instruction scheduling for the SKX target. Reviewers: delena, zvi, RKSimon Differential Revision: https://reviews.llvm.org/D38222 Change-Id: I2381c9b5bb75ecacfca017243c22d054f6eddd14 llvm-svn: 314306
2017-09-27 22:44:15 +08:00
; CHECK-NEXT: vpaddw %xmm0, %xmm1, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf1,0xfd,0xc0]
[AVX512] Remove maksed pshufd, pshuflw, and phufhw intrinsics and autoupgrade them to selects and shufflevector. llvm-svn: 272527
2016-06-13 10:36:48 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <8 x i16> @llvm.x86.avx512.mask.pshufh.w.128(<8 x i16> %x0, i32 3, <8 x i16> %x2, i8 %x3)
%res1 = call <8 x i16> @llvm.x86.avx512.mask.pshufh.w.128(<8 x i16> %x0, i32 3, <8 x i16> zeroinitializer, i8 %x3)
%res2 = call <8 x i16> @llvm.x86.avx512.mask.pshufh.w.128(<8 x i16> %x0, i32 3, <8 x i16> %x2, i8 -1)
%res3 = add <8 x i16> %res, %res1
%res4 = add <8 x i16> %res3, %res2
ret <8 x i16> %res4
}
declare <16 x i16> @llvm.x86.avx512.mask.pshufh.w.256(<16 x i16>, i32, <16 x i16>, i16)
define <16 x i16>@test_int_x86_avx512_mask_pshufh_w_256(<16 x i16> %x0, i32 %x1, <16 x i16> %x2, i16 %x3) {
; CHECK-LABEL: test_int_x86_avx512_mask_pshufh_w_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpshufhw $3, %ymm0, %ymm2 ## EVEX TO VEX Compression encoding: [0xc5,0xfe,0x70,0xd0,0x03]
[AVX512] Remove maksed pshufd, pshuflw, and phufhw intrinsics and autoupgrade them to selects and shufflevector. llvm-svn: 272527
2016-06-13 10:36:48 +08:00
; CHECK-NEXT: ## ymm2 = ymm0[0,1,2,3,7,4,4,4,8,9,10,11,15,12,12,12]
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %esi, %k1 ## encoding: [0xc5,0xfb,0x92,0xce]
[AVX512] Remove maksed pshufd, pshuflw, and phufhw intrinsics and autoupgrade them to selects and shufflevector. llvm-svn: 272527
2016-06-13 10:36:48 +08:00
; CHECK-NEXT: vpshufhw $3, %ymm0, %ymm1 {%k1} ## encoding: [0x62,0xf1,0x7e,0x29,0x70,0xc8,0x03]
[X86][AVX512] Add support for VPALIGNR/PSHUFD/PSHUFHW/PSHUFLW masked shuffle comments llvm-svn: 274466
2016-07-03 23:00:51 +08:00
; CHECK-NEXT: ## ymm1 {%k1} = ymm0[0,1,2,3,7,4,4,4,8,9,10,11,15,12,12,12]
[AVX512] Remove maksed pshufd, pshuflw, and phufhw intrinsics and autoupgrade them to selects and shufflevector. llvm-svn: 272527
2016-06-13 10:36:48 +08:00
; CHECK-NEXT: vpshufhw $3, %ymm0, %ymm0 {%k1} {z} ## encoding: [0x62,0xf1,0x7e,0xa9,0x70,0xc0,0x03]
[X86][AVX512] Add support for VPALIGNR/PSHUFD/PSHUFHW/PSHUFLW masked shuffle comments llvm-svn: 274466
2016-07-03 23:00:51 +08:00
; CHECK-NEXT: ## ymm0 {%k1} {z} = ymm0[0,1,2,3,7,4,4,4,8,9,10,11,15,12,12,12]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpaddw %ymm2, %ymm0, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xfd,0xfd,0xc2]
[X86][SKX][KNL] Updated regression tests to use -mattr instead of -mcpu flag.NFC. NFC. Updated 8 regression tests to use -mattr instead of -mcpu flag as follows: -mcpu=knl --> -mattr=+avx512f -mcpu=skx --> -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq The updates are as part of the preparation of a large commit to add all instruction scheduling for the SKX target. Reviewers: delena, zvi, RKSimon Differential Revision: https://reviews.llvm.org/D38222 Change-Id: I2381c9b5bb75ecacfca017243c22d054f6eddd14 llvm-svn: 314306
2017-09-27 22:44:15 +08:00
; CHECK-NEXT: vpaddw %ymm0, %ymm1, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf5,0xfd,0xc0]
[AVX512] Remove maksed pshufd, pshuflw, and phufhw intrinsics and autoupgrade them to selects and shufflevector. llvm-svn: 272527
2016-06-13 10:36:48 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <16 x i16> @llvm.x86.avx512.mask.pshufh.w.256(<16 x i16> %x0, i32 3, <16 x i16> %x2, i16 %x3)
%res1 = call <16 x i16> @llvm.x86.avx512.mask.pshufh.w.256(<16 x i16> %x0, i32 3, <16 x i16> zeroinitializer, i16 %x3)
%res2 = call <16 x i16> @llvm.x86.avx512.mask.pshufh.w.256(<16 x i16> %x0, i32 3, <16 x i16> %x2, i16 -1)
%res3 = add <16 x i16> %res, %res1
%res4 = add <16 x i16> %res3, %res2
ret <16 x i16> %res4
}
declare <8 x i16> @llvm.x86.avx512.mask.pshufl.w.128(<8 x i16>, i32, <8 x i16>, i8)
define <8 x i16>@test_int_x86_avx512_mask_pshufl_w_128(<8 x i16> %x0, i32 %x1, <8 x i16> %x2, i8 %x3) {
; CHECK-LABEL: test_int_x86_avx512_mask_pshufl_w_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpshuflw $3, %xmm0, %xmm2 ## EVEX TO VEX Compression encoding: [0xc5,0xfb,0x70,0xd0,0x03]
[AVX512] Remove maksed pshufd, pshuflw, and phufhw intrinsics and autoupgrade them to selects and shufflevector. llvm-svn: 272527
2016-06-13 10:36:48 +08:00
; CHECK-NEXT: ## xmm2 = xmm0[3,0,0,0,4,5,6,7]
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %esi, %k1 ## encoding: [0xc5,0xfb,0x92,0xce]
[AVX512] Remove maksed pshufd, pshuflw, and phufhw intrinsics and autoupgrade them to selects and shufflevector. llvm-svn: 272527
2016-06-13 10:36:48 +08:00
; CHECK-NEXT: vpshuflw $3, %xmm0, %xmm1 {%k1} ## encoding: [0x62,0xf1,0x7f,0x09,0x70,0xc8,0x03]
[X86][AVX512] Add support for VPALIGNR/PSHUFD/PSHUFHW/PSHUFLW masked shuffle comments llvm-svn: 274466
2016-07-03 23:00:51 +08:00
; CHECK-NEXT: ## xmm1 {%k1} = xmm0[3,0,0,0,4,5,6,7]
[AVX512] Remove maksed pshufd, pshuflw, and phufhw intrinsics and autoupgrade them to selects and shufflevector. llvm-svn: 272527
2016-06-13 10:36:48 +08:00
; CHECK-NEXT: vpshuflw $3, %xmm0, %xmm0 {%k1} {z} ## encoding: [0x62,0xf1,0x7f,0x89,0x70,0xc0,0x03]
[X86][AVX512] Add support for VPALIGNR/PSHUFD/PSHUFHW/PSHUFLW masked shuffle comments llvm-svn: 274466
2016-07-03 23:00:51 +08:00
; CHECK-NEXT: ## xmm0 {%k1} {z} = xmm0[3,0,0,0,4,5,6,7]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpaddw %xmm2, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xfd,0xc2]
[X86][SKX][KNL] Updated regression tests to use -mattr instead of -mcpu flag.NFC. NFC. Updated 8 regression tests to use -mattr instead of -mcpu flag as follows: -mcpu=knl --> -mattr=+avx512f -mcpu=skx --> -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq The updates are as part of the preparation of a large commit to add all instruction scheduling for the SKX target. Reviewers: delena, zvi, RKSimon Differential Revision: https://reviews.llvm.org/D38222 Change-Id: I2381c9b5bb75ecacfca017243c22d054f6eddd14 llvm-svn: 314306
2017-09-27 22:44:15 +08:00
; CHECK-NEXT: vpaddw %xmm0, %xmm1, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf1,0xfd,0xc0]
[AVX512] Remove maksed pshufd, pshuflw, and phufhw intrinsics and autoupgrade them to selects and shufflevector. llvm-svn: 272527
2016-06-13 10:36:48 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <8 x i16> @llvm.x86.avx512.mask.pshufl.w.128(<8 x i16> %x0, i32 3, <8 x i16> %x2, i8 %x3)
%res1 = call <8 x i16> @llvm.x86.avx512.mask.pshufl.w.128(<8 x i16> %x0, i32 3, <8 x i16> zeroinitializer, i8 %x3)
%res2 = call <8 x i16> @llvm.x86.avx512.mask.pshufl.w.128(<8 x i16> %x0, i32 3, <8 x i16> %x2, i8 -1)
%res3 = add <8 x i16> %res, %res1
%res4 = add <8 x i16> %res3, %res2
ret <8 x i16> %res4
}
declare <16 x i16> @llvm.x86.avx512.mask.pshufl.w.256(<16 x i16>, i32, <16 x i16>, i16)
define <16 x i16>@test_int_x86_avx512_mask_pshufl_w_256(<16 x i16> %x0, i32 %x1, <16 x i16> %x2, i16 %x3) {
; CHECK-LABEL: test_int_x86_avx512_mask_pshufl_w_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpshuflw $3, %ymm0, %ymm2 ## EVEX TO VEX Compression encoding: [0xc5,0xff,0x70,0xd0,0x03]
[AVX512] Remove maksed pshufd, pshuflw, and phufhw intrinsics and autoupgrade them to selects and shufflevector. llvm-svn: 272527
2016-06-13 10:36:48 +08:00
; CHECK-NEXT: ## ymm2 = ymm0[3,0,0,0,4,5,6,7,11,8,8,8,12,13,14,15]
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %esi, %k1 ## encoding: [0xc5,0xfb,0x92,0xce]
[AVX512] Remove maksed pshufd, pshuflw, and phufhw intrinsics and autoupgrade them to selects and shufflevector. llvm-svn: 272527
2016-06-13 10:36:48 +08:00
; CHECK-NEXT: vpshuflw $3, %ymm0, %ymm1 {%k1} ## encoding: [0x62,0xf1,0x7f,0x29,0x70,0xc8,0x03]
[X86][AVX512] Add support for VPALIGNR/PSHUFD/PSHUFHW/PSHUFLW masked shuffle comments llvm-svn: 274466
2016-07-03 23:00:51 +08:00
; CHECK-NEXT: ## ymm1 {%k1} = ymm0[3,0,0,0,4,5,6,7,11,8,8,8,12,13,14,15]
[AVX512] Remove maksed pshufd, pshuflw, and phufhw intrinsics and autoupgrade them to selects and shufflevector. llvm-svn: 272527
2016-06-13 10:36:48 +08:00
; CHECK-NEXT: vpshuflw $3, %ymm0, %ymm0 {%k1} {z} ## encoding: [0x62,0xf1,0x7f,0xa9,0x70,0xc0,0x03]
[X86][AVX512] Add support for VPALIGNR/PSHUFD/PSHUFHW/PSHUFLW masked shuffle comments llvm-svn: 274466
2016-07-03 23:00:51 +08:00
; CHECK-NEXT: ## ymm0 {%k1} {z} = ymm0[3,0,0,0,4,5,6,7,11,8,8,8,12,13,14,15]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpaddw %ymm2, %ymm0, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xfd,0xfd,0xc2]
[X86][SKX][KNL] Updated regression tests to use -mattr instead of -mcpu flag.NFC. NFC. Updated 8 regression tests to use -mattr instead of -mcpu flag as follows: -mcpu=knl --> -mattr=+avx512f -mcpu=skx --> -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq The updates are as part of the preparation of a large commit to add all instruction scheduling for the SKX target. Reviewers: delena, zvi, RKSimon Differential Revision: https://reviews.llvm.org/D38222 Change-Id: I2381c9b5bb75ecacfca017243c22d054f6eddd14 llvm-svn: 314306
2017-09-27 22:44:15 +08:00
; CHECK-NEXT: vpaddw %ymm0, %ymm1, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf5,0xfd,0xc0]
[AVX512] Remove maksed pshufd, pshuflw, and phufhw intrinsics and autoupgrade them to selects and shufflevector. llvm-svn: 272527
2016-06-13 10:36:48 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <16 x i16> @llvm.x86.avx512.mask.pshufl.w.256(<16 x i16> %x0, i32 3, <16 x i16> %x2, i16 %x3)
%res1 = call <16 x i16> @llvm.x86.avx512.mask.pshufl.w.256(<16 x i16> %x0, i32 3, <16 x i16> zeroinitializer, i16 %x3)
%res2 = call <16 x i16> @llvm.x86.avx512.mask.pshufl.w.256(<16 x i16> %x0, i32 3, <16 x i16> %x2, i16 -1)
%res3 = add <16 x i16> %res, %res1
%res4 = add <16 x i16> %res3, %res2
ret <16 x i16> %res4
}
[AVX512] Remove the masked vpcmpeq/vcmpgt intrinsics and autoupgrade them to native icmps. llvm-svn: 273240
2016-06-21 11:53:24 +08:00
define i32 @test_pcmpeq_b_256(<32 x i8> %a, <32 x i8> %b) {
; CHECK-LABEL: test_pcmpeq_b_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX512] Remove the masked vpcmpeq/vcmpgt intrinsics and autoupgrade them to native icmps. llvm-svn: 273240
2016-06-21 11:53:24 +08:00
; CHECK-NEXT: vpcmpeqb %ymm1, %ymm0, %k0 ## encoding: [0x62,0xf1,0x7d,0x28,0x74,0xc1]
; CHECK-NEXT: kmovd %k0, %eax ## encoding: [0xc5,0xfb,0x93,0xc0]
[X86][SKX][KNL] Updated regression tests to use -mattr instead of -mcpu flag.NFC. NFC. Updated 8 regression tests to use -mattr instead of -mcpu flag as follows: -mcpu=knl --> -mattr=+avx512f -mcpu=skx --> -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq The updates are as part of the preparation of a large commit to add all instruction scheduling for the SKX target. Reviewers: delena, zvi, RKSimon Differential Revision: https://reviews.llvm.org/D38222 Change-Id: I2381c9b5bb75ecacfca017243c22d054f6eddd14 llvm-svn: 314306
2017-09-27 22:44:15 +08:00
; CHECK-NEXT: vzeroupper ## encoding: [0xc5,0xf8,0x77]
[AVX512] Remove the masked vpcmpeq/vcmpgt intrinsics and autoupgrade them to native icmps. llvm-svn: 273240
2016-06-21 11:53:24 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call i32 @llvm.x86.avx512.mask.pcmpeq.b.256(<32 x i8> %a, <32 x i8> %b, i32 -1)
ret i32 %res
}
define i32 @test_mask_pcmpeq_b_256(<32 x i8> %a, <32 x i8> %b, i32 %mask) {
; CHECK-LABEL: test_mask_pcmpeq_b_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX512] Remove the masked vpcmpeq/vcmpgt intrinsics and autoupgrade them to native icmps. llvm-svn: 273240
2016-06-21 11:53:24 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
; CHECK-NEXT: vpcmpeqb %ymm1, %ymm0, %k0 {%k1} ## encoding: [0x62,0xf1,0x7d,0x29,0x74,0xc1]
; CHECK-NEXT: kmovd %k0, %eax ## encoding: [0xc5,0xfb,0x93,0xc0]
[X86][SKX][KNL] Updated regression tests to use -mattr instead of -mcpu flag.NFC. NFC. Updated 8 regression tests to use -mattr instead of -mcpu flag as follows: -mcpu=knl --> -mattr=+avx512f -mcpu=skx --> -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq The updates are as part of the preparation of a large commit to add all instruction scheduling for the SKX target. Reviewers: delena, zvi, RKSimon Differential Revision: https://reviews.llvm.org/D38222 Change-Id: I2381c9b5bb75ecacfca017243c22d054f6eddd14 llvm-svn: 314306
2017-09-27 22:44:15 +08:00
; CHECK-NEXT: vzeroupper ## encoding: [0xc5,0xf8,0x77]
[AVX512] Remove the masked vpcmpeq/vcmpgt intrinsics and autoupgrade them to native icmps. llvm-svn: 273240
2016-06-21 11:53:24 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call i32 @llvm.x86.avx512.mask.pcmpeq.b.256(<32 x i8> %a, <32 x i8> %b, i32 %mask)
ret i32 %res
}
declare i32 @llvm.x86.avx512.mask.pcmpeq.b.256(<32 x i8>, <32 x i8>, i32)
define i16 @test_pcmpeq_w_256(<16 x i16> %a, <16 x i16> %b) {
; CHECK-LABEL: test_pcmpeq_w_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX512] Remove the masked vpcmpeq/vcmpgt intrinsics and autoupgrade them to native icmps. llvm-svn: 273240
2016-06-21 11:53:24 +08:00
; CHECK-NEXT: vpcmpeqw %ymm1, %ymm0, %k0 ## encoding: [0x62,0xf1,0x7d,0x28,0x75,0xc1]
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %k0, %eax ## encoding: [0xc5,0xfb,0x93,0xc0]
[CodeGen] Use MachineOperand::print in the MIRPrinter for MO_Register. Work towards the unification of MIR and debug output by refactoring the interfaces. For MachineOperand::print, keep a simple version that can be easily called from `dump()`, and a more complex one which will be called from both the MIRPrinter and MachineInstr::print. Add extra checks inside MachineOperand for detached operands (operands with getParent() == nullptr). https://reviews.llvm.org/D40836 * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/kill: ([^ ]+) ([^ ]+)<def> ([^ ]+)/kill: \1 def \2 \3/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/kill: ([^ ]+) ([^ ]+) ([^ ]+)<def>/kill: \1 \2 def \3/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/kill: def ([^ ]+) ([^ ]+) ([^ ]+)<def>/kill: def \1 \2 def \3/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/<def>//g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<kill>/killed \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-use,kill>/implicit killed \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<dead>/dead \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<def[ ]*,[ ]*dead>/dead \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-def[ ]*,[ ]*dead>/implicit-def dead \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-def>/implicit-def \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-use>/implicit \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<internal>/internal \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<undef>/undef \1/g' llvm-svn: 320022
2017-12-07 18:40:31 +08:00
; CHECK-NEXT: ## kill: def %ax killed %ax killed %eax
[X86][SKX][KNL] Updated regression tests to use -mattr instead of -mcpu flag.NFC. NFC. Updated 8 regression tests to use -mattr instead of -mcpu flag as follows: -mcpu=knl --> -mattr=+avx512f -mcpu=skx --> -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq The updates are as part of the preparation of a large commit to add all instruction scheduling for the SKX target. Reviewers: delena, zvi, RKSimon Differential Revision: https://reviews.llvm.org/D38222 Change-Id: I2381c9b5bb75ecacfca017243c22d054f6eddd14 llvm-svn: 314306
2017-09-27 22:44:15 +08:00
; CHECK-NEXT: vzeroupper ## encoding: [0xc5,0xf8,0x77]
[AVX512] Remove the masked vpcmpeq/vcmpgt intrinsics and autoupgrade them to native icmps. llvm-svn: 273240
2016-06-21 11:53:24 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call i16 @llvm.x86.avx512.mask.pcmpeq.w.256(<16 x i16> %a, <16 x i16> %b, i16 -1)
ret i16 %res
}
define i16 @test_mask_pcmpeq_w_256(<16 x i16> %a, <16 x i16> %b, i16 %mask) {
; CHECK-LABEL: test_mask_pcmpeq_w_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
[AVX512] Remove the masked vpcmpeq/vcmpgt intrinsics and autoupgrade them to native icmps. llvm-svn: 273240
2016-06-21 11:53:24 +08:00
; CHECK-NEXT: vpcmpeqw %ymm1, %ymm0, %k0 {%k1} ## encoding: [0x62,0xf1,0x7d,0x29,0x75,0xc1]
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %k0, %eax ## encoding: [0xc5,0xfb,0x93,0xc0]
[CodeGen] Use MachineOperand::print in the MIRPrinter for MO_Register. Work towards the unification of MIR and debug output by refactoring the interfaces. For MachineOperand::print, keep a simple version that can be easily called from `dump()`, and a more complex one which will be called from both the MIRPrinter and MachineInstr::print. Add extra checks inside MachineOperand for detached operands (operands with getParent() == nullptr). https://reviews.llvm.org/D40836 * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/kill: ([^ ]+) ([^ ]+)<def> ([^ ]+)/kill: \1 def \2 \3/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/kill: ([^ ]+) ([^ ]+) ([^ ]+)<def>/kill: \1 \2 def \3/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/kill: def ([^ ]+) ([^ ]+) ([^ ]+)<def>/kill: def \1 \2 def \3/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/<def>//g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<kill>/killed \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-use,kill>/implicit killed \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<dead>/dead \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<def[ ]*,[ ]*dead>/dead \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-def[ ]*,[ ]*dead>/implicit-def dead \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-def>/implicit-def \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-use>/implicit \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<internal>/internal \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<undef>/undef \1/g' llvm-svn: 320022
2017-12-07 18:40:31 +08:00
; CHECK-NEXT: ## kill: def %ax killed %ax killed %eax
[X86][SKX][KNL] Updated regression tests to use -mattr instead of -mcpu flag.NFC. NFC. Updated 8 regression tests to use -mattr instead of -mcpu flag as follows: -mcpu=knl --> -mattr=+avx512f -mcpu=skx --> -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq The updates are as part of the preparation of a large commit to add all instruction scheduling for the SKX target. Reviewers: delena, zvi, RKSimon Differential Revision: https://reviews.llvm.org/D38222 Change-Id: I2381c9b5bb75ecacfca017243c22d054f6eddd14 llvm-svn: 314306
2017-09-27 22:44:15 +08:00
; CHECK-NEXT: vzeroupper ## encoding: [0xc5,0xf8,0x77]
[AVX512] Remove the masked vpcmpeq/vcmpgt intrinsics and autoupgrade them to native icmps. llvm-svn: 273240
2016-06-21 11:53:24 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call i16 @llvm.x86.avx512.mask.pcmpeq.w.256(<16 x i16> %a, <16 x i16> %b, i16 %mask)
ret i16 %res
}
declare i16 @llvm.x86.avx512.mask.pcmpeq.w.256(<16 x i16>, <16 x i16>, i16)
define i32 @test_pcmpgt_b_256(<32 x i8> %a, <32 x i8> %b) {
; CHECK-LABEL: test_pcmpgt_b_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX512] Remove the masked vpcmpeq/vcmpgt intrinsics and autoupgrade them to native icmps. llvm-svn: 273240
2016-06-21 11:53:24 +08:00
; CHECK-NEXT: vpcmpgtb %ymm1, %ymm0, %k0 ## encoding: [0x62,0xf1,0x7d,0x28,0x64,0xc1]
; CHECK-NEXT: kmovd %k0, %eax ## encoding: [0xc5,0xfb,0x93,0xc0]
[X86][SKX][KNL] Updated regression tests to use -mattr instead of -mcpu flag.NFC. NFC. Updated 8 regression tests to use -mattr instead of -mcpu flag as follows: -mcpu=knl --> -mattr=+avx512f -mcpu=skx --> -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq The updates are as part of the preparation of a large commit to add all instruction scheduling for the SKX target. Reviewers: delena, zvi, RKSimon Differential Revision: https://reviews.llvm.org/D38222 Change-Id: I2381c9b5bb75ecacfca017243c22d054f6eddd14 llvm-svn: 314306
2017-09-27 22:44:15 +08:00
; CHECK-NEXT: vzeroupper ## encoding: [0xc5,0xf8,0x77]
[AVX512] Remove the masked vpcmpeq/vcmpgt intrinsics and autoupgrade them to native icmps. llvm-svn: 273240
2016-06-21 11:53:24 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call i32 @llvm.x86.avx512.mask.pcmpgt.b.256(<32 x i8> %a, <32 x i8> %b, i32 -1)
ret i32 %res
}
define i32 @test_mask_pcmpgt_b_256(<32 x i8> %a, <32 x i8> %b, i32 %mask) {
; CHECK-LABEL: test_mask_pcmpgt_b_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX512] Remove the masked vpcmpeq/vcmpgt intrinsics and autoupgrade them to native icmps. llvm-svn: 273240
2016-06-21 11:53:24 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
; CHECK-NEXT: vpcmpgtb %ymm1, %ymm0, %k0 {%k1} ## encoding: [0x62,0xf1,0x7d,0x29,0x64,0xc1]
; CHECK-NEXT: kmovd %k0, %eax ## encoding: [0xc5,0xfb,0x93,0xc0]
[X86][SKX][KNL] Updated regression tests to use -mattr instead of -mcpu flag.NFC. NFC. Updated 8 regression tests to use -mattr instead of -mcpu flag as follows: -mcpu=knl --> -mattr=+avx512f -mcpu=skx --> -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq The updates are as part of the preparation of a large commit to add all instruction scheduling for the SKX target. Reviewers: delena, zvi, RKSimon Differential Revision: https://reviews.llvm.org/D38222 Change-Id: I2381c9b5bb75ecacfca017243c22d054f6eddd14 llvm-svn: 314306
2017-09-27 22:44:15 +08:00
; CHECK-NEXT: vzeroupper ## encoding: [0xc5,0xf8,0x77]
[AVX512] Remove the masked vpcmpeq/vcmpgt intrinsics and autoupgrade them to native icmps. llvm-svn: 273240
2016-06-21 11:53:24 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call i32 @llvm.x86.avx512.mask.pcmpgt.b.256(<32 x i8> %a, <32 x i8> %b, i32 %mask)
ret i32 %res
}
declare i32 @llvm.x86.avx512.mask.pcmpgt.b.256(<32 x i8>, <32 x i8>, i32)
define i16 @test_pcmpgt_w_256(<16 x i16> %a, <16 x i16> %b) {
; CHECK-LABEL: test_pcmpgt_w_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX512] Remove the masked vpcmpeq/vcmpgt intrinsics and autoupgrade them to native icmps. llvm-svn: 273240
2016-06-21 11:53:24 +08:00
; CHECK-NEXT: vpcmpgtw %ymm1, %ymm0, %k0 ## encoding: [0x62,0xf1,0x7d,0x28,0x65,0xc1]
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %k0, %eax ## encoding: [0xc5,0xfb,0x93,0xc0]
[CodeGen] Use MachineOperand::print in the MIRPrinter for MO_Register. Work towards the unification of MIR and debug output by refactoring the interfaces. For MachineOperand::print, keep a simple version that can be easily called from `dump()`, and a more complex one which will be called from both the MIRPrinter and MachineInstr::print. Add extra checks inside MachineOperand for detached operands (operands with getParent() == nullptr). https://reviews.llvm.org/D40836 * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/kill: ([^ ]+) ([^ ]+)<def> ([^ ]+)/kill: \1 def \2 \3/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/kill: ([^ ]+) ([^ ]+) ([^ ]+)<def>/kill: \1 \2 def \3/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/kill: def ([^ ]+) ([^ ]+) ([^ ]+)<def>/kill: def \1 \2 def \3/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/<def>//g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<kill>/killed \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-use,kill>/implicit killed \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<dead>/dead \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<def[ ]*,[ ]*dead>/dead \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-def[ ]*,[ ]*dead>/implicit-def dead \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-def>/implicit-def \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-use>/implicit \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<internal>/internal \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<undef>/undef \1/g' llvm-svn: 320022
2017-12-07 18:40:31 +08:00
; CHECK-NEXT: ## kill: def %ax killed %ax killed %eax
[X86][SKX][KNL] Updated regression tests to use -mattr instead of -mcpu flag.NFC. NFC. Updated 8 regression tests to use -mattr instead of -mcpu flag as follows: -mcpu=knl --> -mattr=+avx512f -mcpu=skx --> -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq The updates are as part of the preparation of a large commit to add all instruction scheduling for the SKX target. Reviewers: delena, zvi, RKSimon Differential Revision: https://reviews.llvm.org/D38222 Change-Id: I2381c9b5bb75ecacfca017243c22d054f6eddd14 llvm-svn: 314306
2017-09-27 22:44:15 +08:00
; CHECK-NEXT: vzeroupper ## encoding: [0xc5,0xf8,0x77]
[AVX512] Remove the masked vpcmpeq/vcmpgt intrinsics and autoupgrade them to native icmps. llvm-svn: 273240
2016-06-21 11:53:24 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call i16 @llvm.x86.avx512.mask.pcmpgt.w.256(<16 x i16> %a, <16 x i16> %b, i16 -1)
ret i16 %res
}
define i16 @test_mask_pcmpgt_w_256(<16 x i16> %a, <16 x i16> %b, i16 %mask) {
; CHECK-LABEL: test_mask_pcmpgt_w_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
[AVX512] Remove the masked vpcmpeq/vcmpgt intrinsics and autoupgrade them to native icmps. llvm-svn: 273240
2016-06-21 11:53:24 +08:00
; CHECK-NEXT: vpcmpgtw %ymm1, %ymm0, %k0 {%k1} ## encoding: [0x62,0xf1,0x7d,0x29,0x65,0xc1]
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %k0, %eax ## encoding: [0xc5,0xfb,0x93,0xc0]
[CodeGen] Use MachineOperand::print in the MIRPrinter for MO_Register. Work towards the unification of MIR and debug output by refactoring the interfaces. For MachineOperand::print, keep a simple version that can be easily called from `dump()`, and a more complex one which will be called from both the MIRPrinter and MachineInstr::print. Add extra checks inside MachineOperand for detached operands (operands with getParent() == nullptr). https://reviews.llvm.org/D40836 * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/kill: ([^ ]+) ([^ ]+)<def> ([^ ]+)/kill: \1 def \2 \3/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/kill: ([^ ]+) ([^ ]+) ([^ ]+)<def>/kill: \1 \2 def \3/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/kill: def ([^ ]+) ([^ ]+) ([^ ]+)<def>/kill: def \1 \2 def \3/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/<def>//g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<kill>/killed \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-use,kill>/implicit killed \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<dead>/dead \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<def[ ]*,[ ]*dead>/dead \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-def[ ]*,[ ]*dead>/implicit-def dead \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-def>/implicit-def \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-use>/implicit \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<internal>/internal \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<undef>/undef \1/g' llvm-svn: 320022
2017-12-07 18:40:31 +08:00
; CHECK-NEXT: ## kill: def %ax killed %ax killed %eax
[X86][SKX][KNL] Updated regression tests to use -mattr instead of -mcpu flag.NFC. NFC. Updated 8 regression tests to use -mattr instead of -mcpu flag as follows: -mcpu=knl --> -mattr=+avx512f -mcpu=skx --> -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq The updates are as part of the preparation of a large commit to add all instruction scheduling for the SKX target. Reviewers: delena, zvi, RKSimon Differential Revision: https://reviews.llvm.org/D38222 Change-Id: I2381c9b5bb75ecacfca017243c22d054f6eddd14 llvm-svn: 314306
2017-09-27 22:44:15 +08:00
; CHECK-NEXT: vzeroupper ## encoding: [0xc5,0xf8,0x77]
[AVX512] Remove the masked vpcmpeq/vcmpgt intrinsics and autoupgrade them to native icmps. llvm-svn: 273240
2016-06-21 11:53:24 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call i16 @llvm.x86.avx512.mask.pcmpgt.w.256(<16 x i16> %a, <16 x i16> %b, i16 %mask)
ret i16 %res
}
declare i16 @llvm.x86.avx512.mask.pcmpgt.w.256(<16 x i16>, <16 x i16>, i16)
[AVX-512] Remove and autoupgrade the masked integer compare intrinsics Summary: These intrinsics aren't used by clang and haven't been for a while. There's some really terrible codegen in the 32-bit target for avx512bw due to i64 not being legal. But as I said these intrinsics aren't used by clang even before this patch so this codegen reflects our clang behavior today. Reviewers: spatel, RKSimon, zvi, igorb Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D34389 llvm-svn: 306047
2017-06-23 04:11:01 +08:00
define i16 @test_pcmpeq_b_128(<16 x i8> %a, <16 x i8> %b) {
; CHECK-LABEL: test_pcmpeq_b_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove and autoupgrade the masked integer compare intrinsics Summary: These intrinsics aren't used by clang and haven't been for a while. There's some really terrible codegen in the 32-bit target for avx512bw due to i64 not being legal. But as I said these intrinsics aren't used by clang even before this patch so this codegen reflects our clang behavior today. Reviewers: spatel, RKSimon, zvi, igorb Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D34389 llvm-svn: 306047
2017-06-23 04:11:01 +08:00
; CHECK-NEXT: vpcmpeqb %xmm1, %xmm0, %k0 ## encoding: [0x62,0xf1,0x7d,0x08,0x74,0xc1]
; CHECK-NEXT: kmovd %k0, %eax ## encoding: [0xc5,0xfb,0x93,0xc0]
[CodeGen] Use MachineOperand::print in the MIRPrinter for MO_Register. Work towards the unification of MIR and debug output by refactoring the interfaces. For MachineOperand::print, keep a simple version that can be easily called from `dump()`, and a more complex one which will be called from both the MIRPrinter and MachineInstr::print. Add extra checks inside MachineOperand for detached operands (operands with getParent() == nullptr). https://reviews.llvm.org/D40836 * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/kill: ([^ ]+) ([^ ]+)<def> ([^ ]+)/kill: \1 def \2 \3/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/kill: ([^ ]+) ([^ ]+) ([^ ]+)<def>/kill: \1 \2 def \3/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/kill: def ([^ ]+) ([^ ]+) ([^ ]+)<def>/kill: def \1 \2 def \3/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/<def>//g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<kill>/killed \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-use,kill>/implicit killed \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<dead>/dead \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<def[ ]*,[ ]*dead>/dead \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-def[ ]*,[ ]*dead>/implicit-def dead \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-def>/implicit-def \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-use>/implicit \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<internal>/internal \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<undef>/undef \1/g' llvm-svn: 320022
2017-12-07 18:40:31 +08:00
; CHECK-NEXT: ## kill: def %ax killed %ax killed %eax
[AVX-512] Remove and autoupgrade the masked integer compare intrinsics Summary: These intrinsics aren't used by clang and haven't been for a while. There's some really terrible codegen in the 32-bit target for avx512bw due to i64 not being legal. But as I said these intrinsics aren't used by clang even before this patch so this codegen reflects our clang behavior today. Reviewers: spatel, RKSimon, zvi, igorb Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D34389 llvm-svn: 306047
2017-06-23 04:11:01 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call i16 @llvm.x86.avx512.mask.pcmpeq.b.128(<16 x i8> %a, <16 x i8> %b, i16 -1)
ret i16 %res
}
define i16 @test_mask_pcmpeq_b_128(<16 x i8> %a, <16 x i8> %b, i16 %mask) {
; CHECK-LABEL: test_mask_pcmpeq_b_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove and autoupgrade the masked integer compare intrinsics Summary: These intrinsics aren't used by clang and haven't been for a while. There's some really terrible codegen in the 32-bit target for avx512bw due to i64 not being legal. But as I said these intrinsics aren't used by clang even before this patch so this codegen reflects our clang behavior today. Reviewers: spatel, RKSimon, zvi, igorb Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D34389 llvm-svn: 306047
2017-06-23 04:11:01 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
; CHECK-NEXT: vpcmpeqb %xmm1, %xmm0, %k0 {%k1} ## encoding: [0x62,0xf1,0x7d,0x09,0x74,0xc1]
; CHECK-NEXT: kmovd %k0, %eax ## encoding: [0xc5,0xfb,0x93,0xc0]
[CodeGen] Use MachineOperand::print in the MIRPrinter for MO_Register. Work towards the unification of MIR and debug output by refactoring the interfaces. For MachineOperand::print, keep a simple version that can be easily called from `dump()`, and a more complex one which will be called from both the MIRPrinter and MachineInstr::print. Add extra checks inside MachineOperand for detached operands (operands with getParent() == nullptr). https://reviews.llvm.org/D40836 * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/kill: ([^ ]+) ([^ ]+)<def> ([^ ]+)/kill: \1 def \2 \3/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/kill: ([^ ]+) ([^ ]+) ([^ ]+)<def>/kill: \1 \2 def \3/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/kill: def ([^ ]+) ([^ ]+) ([^ ]+)<def>/kill: def \1 \2 def \3/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/<def>//g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<kill>/killed \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-use,kill>/implicit killed \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<dead>/dead \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<def[ ]*,[ ]*dead>/dead \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-def[ ]*,[ ]*dead>/implicit-def dead \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-def>/implicit-def \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-use>/implicit \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<internal>/internal \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<undef>/undef \1/g' llvm-svn: 320022
2017-12-07 18:40:31 +08:00
; CHECK-NEXT: ## kill: def %ax killed %ax killed %eax
[AVX-512] Remove and autoupgrade the masked integer compare intrinsics Summary: These intrinsics aren't used by clang and haven't been for a while. There's some really terrible codegen in the 32-bit target for avx512bw due to i64 not being legal. But as I said these intrinsics aren't used by clang even before this patch so this codegen reflects our clang behavior today. Reviewers: spatel, RKSimon, zvi, igorb Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D34389 llvm-svn: 306047
2017-06-23 04:11:01 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call i16 @llvm.x86.avx512.mask.pcmpeq.b.128(<16 x i8> %a, <16 x i8> %b, i16 %mask)
ret i16 %res
}
declare i16 @llvm.x86.avx512.mask.pcmpeq.b.128(<16 x i8>, <16 x i8>, i16)
define i8 @test_pcmpeq_w_128(<8 x i16> %a, <8 x i16> %b) {
; CHECK-LABEL: test_pcmpeq_w_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove and autoupgrade the masked integer compare intrinsics Summary: These intrinsics aren't used by clang and haven't been for a while. There's some really terrible codegen in the 32-bit target for avx512bw due to i64 not being legal. But as I said these intrinsics aren't used by clang even before this patch so this codegen reflects our clang behavior today. Reviewers: spatel, RKSimon, zvi, igorb Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D34389 llvm-svn: 306047
2017-06-23 04:11:01 +08:00
; CHECK-NEXT: vpcmpeqw %xmm1, %xmm0, %k0 ## encoding: [0x62,0xf1,0x7d,0x08,0x75,0xc1]
; CHECK-NEXT: kmovd %k0, %eax ## encoding: [0xc5,0xfb,0x93,0xc0]
[CodeGen] Use MachineOperand::print in the MIRPrinter for MO_Register. Work towards the unification of MIR and debug output by refactoring the interfaces. For MachineOperand::print, keep a simple version that can be easily called from `dump()`, and a more complex one which will be called from both the MIRPrinter and MachineInstr::print. Add extra checks inside MachineOperand for detached operands (operands with getParent() == nullptr). https://reviews.llvm.org/D40836 * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/kill: ([^ ]+) ([^ ]+)<def> ([^ ]+)/kill: \1 def \2 \3/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/kill: ([^ ]+) ([^ ]+) ([^ ]+)<def>/kill: \1 \2 def \3/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/kill: def ([^ ]+) ([^ ]+) ([^ ]+)<def>/kill: def \1 \2 def \3/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/<def>//g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<kill>/killed \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-use,kill>/implicit killed \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<dead>/dead \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<def[ ]*,[ ]*dead>/dead \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-def[ ]*,[ ]*dead>/implicit-def dead \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-def>/implicit-def \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-use>/implicit \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<internal>/internal \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<undef>/undef \1/g' llvm-svn: 320022
2017-12-07 18:40:31 +08:00
; CHECK-NEXT: ## kill: def %al killed %al killed %eax
[AVX-512] Remove and autoupgrade the masked integer compare intrinsics Summary: These intrinsics aren't used by clang and haven't been for a while. There's some really terrible codegen in the 32-bit target for avx512bw due to i64 not being legal. But as I said these intrinsics aren't used by clang even before this patch so this codegen reflects our clang behavior today. Reviewers: spatel, RKSimon, zvi, igorb Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D34389 llvm-svn: 306047
2017-06-23 04:11:01 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call i8 @llvm.x86.avx512.mask.pcmpeq.w.128(<8 x i16> %a, <8 x i16> %b, i8 -1)
ret i8 %res
}
define i8 @test_mask_pcmpeq_w_128(<8 x i16> %a, <8 x i16> %b, i8 %mask) {
; CHECK-LABEL: test_mask_pcmpeq_w_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove and autoupgrade the masked integer compare intrinsics Summary: These intrinsics aren't used by clang and haven't been for a while. There's some really terrible codegen in the 32-bit target for avx512bw due to i64 not being legal. But as I said these intrinsics aren't used by clang even before this patch so this codegen reflects our clang behavior today. Reviewers: spatel, RKSimon, zvi, igorb Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D34389 llvm-svn: 306047
2017-06-23 04:11:01 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
; CHECK-NEXT: vpcmpeqw %xmm1, %xmm0, %k0 {%k1} ## encoding: [0x62,0xf1,0x7d,0x09,0x75,0xc1]
; CHECK-NEXT: kmovd %k0, %eax ## encoding: [0xc5,0xfb,0x93,0xc0]
[CodeGen] Use MachineOperand::print in the MIRPrinter for MO_Register. Work towards the unification of MIR and debug output by refactoring the interfaces. For MachineOperand::print, keep a simple version that can be easily called from `dump()`, and a more complex one which will be called from both the MIRPrinter and MachineInstr::print. Add extra checks inside MachineOperand for detached operands (operands with getParent() == nullptr). https://reviews.llvm.org/D40836 * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/kill: ([^ ]+) ([^ ]+)<def> ([^ ]+)/kill: \1 def \2 \3/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/kill: ([^ ]+) ([^ ]+) ([^ ]+)<def>/kill: \1 \2 def \3/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/kill: def ([^ ]+) ([^ ]+) ([^ ]+)<def>/kill: def \1 \2 def \3/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/<def>//g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<kill>/killed \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-use,kill>/implicit killed \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<dead>/dead \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<def[ ]*,[ ]*dead>/dead \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-def[ ]*,[ ]*dead>/implicit-def dead \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-def>/implicit-def \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-use>/implicit \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<internal>/internal \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<undef>/undef \1/g' llvm-svn: 320022
2017-12-07 18:40:31 +08:00
; CHECK-NEXT: ## kill: def %al killed %al killed %eax
[AVX-512] Remove and autoupgrade the masked integer compare intrinsics Summary: These intrinsics aren't used by clang and haven't been for a while. There's some really terrible codegen in the 32-bit target for avx512bw due to i64 not being legal. But as I said these intrinsics aren't used by clang even before this patch so this codegen reflects our clang behavior today. Reviewers: spatel, RKSimon, zvi, igorb Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D34389 llvm-svn: 306047
2017-06-23 04:11:01 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call i8 @llvm.x86.avx512.mask.pcmpeq.w.128(<8 x i16> %a, <8 x i16> %b, i8 %mask)
ret i8 %res
}
declare i8 @llvm.x86.avx512.mask.pcmpeq.w.128(<8 x i16>, <8 x i16>, i8)
define i16 @test_pcmpgt_b_128(<16 x i8> %a, <16 x i8> %b) {
; CHECK-LABEL: test_pcmpgt_b_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove and autoupgrade the masked integer compare intrinsics Summary: These intrinsics aren't used by clang and haven't been for a while. There's some really terrible codegen in the 32-bit target for avx512bw due to i64 not being legal. But as I said these intrinsics aren't used by clang even before this patch so this codegen reflects our clang behavior today. Reviewers: spatel, RKSimon, zvi, igorb Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D34389 llvm-svn: 306047
2017-06-23 04:11:01 +08:00
; CHECK-NEXT: vpcmpgtb %xmm1, %xmm0, %k0 ## encoding: [0x62,0xf1,0x7d,0x08,0x64,0xc1]
; CHECK-NEXT: kmovd %k0, %eax ## encoding: [0xc5,0xfb,0x93,0xc0]
[CodeGen] Use MachineOperand::print in the MIRPrinter for MO_Register. Work towards the unification of MIR and debug output by refactoring the interfaces. For MachineOperand::print, keep a simple version that can be easily called from `dump()`, and a more complex one which will be called from both the MIRPrinter and MachineInstr::print. Add extra checks inside MachineOperand for detached operands (operands with getParent() == nullptr). https://reviews.llvm.org/D40836 * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/kill: ([^ ]+) ([^ ]+)<def> ([^ ]+)/kill: \1 def \2 \3/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/kill: ([^ ]+) ([^ ]+) ([^ ]+)<def>/kill: \1 \2 def \3/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/kill: def ([^ ]+) ([^ ]+) ([^ ]+)<def>/kill: def \1 \2 def \3/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/<def>//g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<kill>/killed \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-use,kill>/implicit killed \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<dead>/dead \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<def[ ]*,[ ]*dead>/dead \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-def[ ]*,[ ]*dead>/implicit-def dead \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-def>/implicit-def \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-use>/implicit \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<internal>/internal \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<undef>/undef \1/g' llvm-svn: 320022
2017-12-07 18:40:31 +08:00
; CHECK-NEXT: ## kill: def %ax killed %ax killed %eax
[AVX-512] Remove and autoupgrade the masked integer compare intrinsics Summary: These intrinsics aren't used by clang and haven't been for a while. There's some really terrible codegen in the 32-bit target for avx512bw due to i64 not being legal. But as I said these intrinsics aren't used by clang even before this patch so this codegen reflects our clang behavior today. Reviewers: spatel, RKSimon, zvi, igorb Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D34389 llvm-svn: 306047
2017-06-23 04:11:01 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call i16 @llvm.x86.avx512.mask.pcmpgt.b.128(<16 x i8> %a, <16 x i8> %b, i16 -1)
ret i16 %res
}
define i16 @test_mask_pcmpgt_b_128(<16 x i8> %a, <16 x i8> %b, i16 %mask) {
; CHECK-LABEL: test_mask_pcmpgt_b_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove and autoupgrade the masked integer compare intrinsics Summary: These intrinsics aren't used by clang and haven't been for a while. There's some really terrible codegen in the 32-bit target for avx512bw due to i64 not being legal. But as I said these intrinsics aren't used by clang even before this patch so this codegen reflects our clang behavior today. Reviewers: spatel, RKSimon, zvi, igorb Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D34389 llvm-svn: 306047
2017-06-23 04:11:01 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
; CHECK-NEXT: vpcmpgtb %xmm1, %xmm0, %k0 {%k1} ## encoding: [0x62,0xf1,0x7d,0x09,0x64,0xc1]
; CHECK-NEXT: kmovd %k0, %eax ## encoding: [0xc5,0xfb,0x93,0xc0]
[CodeGen] Use MachineOperand::print in the MIRPrinter for MO_Register. Work towards the unification of MIR and debug output by refactoring the interfaces. For MachineOperand::print, keep a simple version that can be easily called from `dump()`, and a more complex one which will be called from both the MIRPrinter and MachineInstr::print. Add extra checks inside MachineOperand for detached operands (operands with getParent() == nullptr). https://reviews.llvm.org/D40836 * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/kill: ([^ ]+) ([^ ]+)<def> ([^ ]+)/kill: \1 def \2 \3/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/kill: ([^ ]+) ([^ ]+) ([^ ]+)<def>/kill: \1 \2 def \3/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/kill: def ([^ ]+) ([^ ]+) ([^ ]+)<def>/kill: def \1 \2 def \3/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/<def>//g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<kill>/killed \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-use,kill>/implicit killed \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<dead>/dead \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<def[ ]*,[ ]*dead>/dead \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-def[ ]*,[ ]*dead>/implicit-def dead \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-def>/implicit-def \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-use>/implicit \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<internal>/internal \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<undef>/undef \1/g' llvm-svn: 320022
2017-12-07 18:40:31 +08:00
; CHECK-NEXT: ## kill: def %ax killed %ax killed %eax
[AVX-512] Remove and autoupgrade the masked integer compare intrinsics Summary: These intrinsics aren't used by clang and haven't been for a while. There's some really terrible codegen in the 32-bit target for avx512bw due to i64 not being legal. But as I said these intrinsics aren't used by clang even before this patch so this codegen reflects our clang behavior today. Reviewers: spatel, RKSimon, zvi, igorb Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D34389 llvm-svn: 306047
2017-06-23 04:11:01 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call i16 @llvm.x86.avx512.mask.pcmpgt.b.128(<16 x i8> %a, <16 x i8> %b, i16 %mask)
ret i16 %res
}
declare i16 @llvm.x86.avx512.mask.pcmpgt.b.128(<16 x i8>, <16 x i8>, i16)
define i8 @test_pcmpgt_w_128(<8 x i16> %a, <8 x i16> %b) {
; CHECK-LABEL: test_pcmpgt_w_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove and autoupgrade the masked integer compare intrinsics Summary: These intrinsics aren't used by clang and haven't been for a while. There's some really terrible codegen in the 32-bit target for avx512bw due to i64 not being legal. But as I said these intrinsics aren't used by clang even before this patch so this codegen reflects our clang behavior today. Reviewers: spatel, RKSimon, zvi, igorb Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D34389 llvm-svn: 306047
2017-06-23 04:11:01 +08:00
; CHECK-NEXT: vpcmpgtw %xmm1, %xmm0, %k0 ## encoding: [0x62,0xf1,0x7d,0x08,0x65,0xc1]
; CHECK-NEXT: kmovd %k0, %eax ## encoding: [0xc5,0xfb,0x93,0xc0]
[CodeGen] Use MachineOperand::print in the MIRPrinter for MO_Register. Work towards the unification of MIR and debug output by refactoring the interfaces. For MachineOperand::print, keep a simple version that can be easily called from `dump()`, and a more complex one which will be called from both the MIRPrinter and MachineInstr::print. Add extra checks inside MachineOperand for detached operands (operands with getParent() == nullptr). https://reviews.llvm.org/D40836 * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/kill: ([^ ]+) ([^ ]+)<def> ([^ ]+)/kill: \1 def \2 \3/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/kill: ([^ ]+) ([^ ]+) ([^ ]+)<def>/kill: \1 \2 def \3/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/kill: def ([^ ]+) ([^ ]+) ([^ ]+)<def>/kill: def \1 \2 def \3/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/<def>//g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<kill>/killed \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-use,kill>/implicit killed \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<dead>/dead \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<def[ ]*,[ ]*dead>/dead \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-def[ ]*,[ ]*dead>/implicit-def dead \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-def>/implicit-def \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-use>/implicit \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<internal>/internal \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<undef>/undef \1/g' llvm-svn: 320022
2017-12-07 18:40:31 +08:00
; CHECK-NEXT: ## kill: def %al killed %al killed %eax
[AVX-512] Remove and autoupgrade the masked integer compare intrinsics Summary: These intrinsics aren't used by clang and haven't been for a while. There's some really terrible codegen in the 32-bit target for avx512bw due to i64 not being legal. But as I said these intrinsics aren't used by clang even before this patch so this codegen reflects our clang behavior today. Reviewers: spatel, RKSimon, zvi, igorb Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D34389 llvm-svn: 306047
2017-06-23 04:11:01 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call i8 @llvm.x86.avx512.mask.pcmpgt.w.128(<8 x i16> %a, <8 x i16> %b, i8 -1)
ret i8 %res
}
define i8 @test_mask_pcmpgt_w_128(<8 x i16> %a, <8 x i16> %b, i8 %mask) {
; CHECK-LABEL: test_mask_pcmpgt_w_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove and autoupgrade the masked integer compare intrinsics Summary: These intrinsics aren't used by clang and haven't been for a while. There's some really terrible codegen in the 32-bit target for avx512bw due to i64 not being legal. But as I said these intrinsics aren't used by clang even before this patch so this codegen reflects our clang behavior today. Reviewers: spatel, RKSimon, zvi, igorb Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D34389 llvm-svn: 306047
2017-06-23 04:11:01 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
; CHECK-NEXT: vpcmpgtw %xmm1, %xmm0, %k0 {%k1} ## encoding: [0x62,0xf1,0x7d,0x09,0x65,0xc1]
; CHECK-NEXT: kmovd %k0, %eax ## encoding: [0xc5,0xfb,0x93,0xc0]
[CodeGen] Use MachineOperand::print in the MIRPrinter for MO_Register. Work towards the unification of MIR and debug output by refactoring the interfaces. For MachineOperand::print, keep a simple version that can be easily called from `dump()`, and a more complex one which will be called from both the MIRPrinter and MachineInstr::print. Add extra checks inside MachineOperand for detached operands (operands with getParent() == nullptr). https://reviews.llvm.org/D40836 * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/kill: ([^ ]+) ([^ ]+)<def> ([^ ]+)/kill: \1 def \2 \3/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/kill: ([^ ]+) ([^ ]+) ([^ ]+)<def>/kill: \1 \2 def \3/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/kill: def ([^ ]+) ([^ ]+) ([^ ]+)<def>/kill: def \1 \2 def \3/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/<def>//g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<kill>/killed \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-use,kill>/implicit killed \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<dead>/dead \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<def[ ]*,[ ]*dead>/dead \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-def[ ]*,[ ]*dead>/implicit-def dead \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-def>/implicit-def \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-use>/implicit \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<internal>/internal \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<undef>/undef \1/g' llvm-svn: 320022
2017-12-07 18:40:31 +08:00
; CHECK-NEXT: ## kill: def %al killed %al killed %eax
[AVX-512] Remove and autoupgrade the masked integer compare intrinsics Summary: These intrinsics aren't used by clang and haven't been for a while. There's some really terrible codegen in the 32-bit target for avx512bw due to i64 not being legal. But as I said these intrinsics aren't used by clang even before this patch so this codegen reflects our clang behavior today. Reviewers: spatel, RKSimon, zvi, igorb Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D34389 llvm-svn: 306047
2017-06-23 04:11:01 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call i8 @llvm.x86.avx512.mask.pcmpgt.w.128(<8 x i16> %a, <8 x i16> %b, i8 %mask)
ret i8 %res
}
declare i8 @llvm.x86.avx512.mask.pcmpgt.w.128(<8 x i16>, <8 x i16>, i8)
[AVX512] Remove masked unpack intrinsics and autoupgrade to vectorshuffle and selects. llvm-svn: 273543
2016-06-23 15:37:33 +08:00
declare <16 x i8> @llvm.x86.avx512.mask.punpckhb.w.128(<16 x i8>, <16 x i8>, <16 x i8>, i16)
define <16 x i8>@test_int_x86_avx512_mask_punpckhb_w_128(<16 x i8> %x0, <16 x i8> %x1, <16 x i8> %x2, i16 %x3) {
; CHECK-LABEL: test_int_x86_avx512_mask_punpckhb_w_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpunpckhbw %xmm1, %xmm0, %xmm3 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0x68,0xd9]
[AVX512] Remove masked unpack intrinsics and autoupgrade to vectorshuffle and selects. llvm-svn: 273543
2016-06-23 15:37:33 +08:00
; CHECK-NEXT: ## xmm3 = xmm0[8],xmm1[8],xmm0[9],xmm1[9],xmm0[10],xmm1[10],xmm0[11],xmm1[11],xmm0[12],xmm1[12],xmm0[13],xmm1[13],xmm0[14],xmm1[14],xmm0[15],xmm1[15]
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
[AVX512] Remove masked unpack intrinsics and autoupgrade to vectorshuffle and selects. llvm-svn: 273543
2016-06-23 15:37:33 +08:00
; CHECK-NEXT: vpunpckhbw %xmm1, %xmm0, %xmm2 {%k1} ## encoding: [0x62,0xf1,0x7d,0x09,0x68,0xd1]
[X86][AVX512] Add support for UNPCK masked shuffle comments llvm-svn: 274464
2016-07-03 22:26:21 +08:00
; CHECK-NEXT: ## xmm2 {%k1} = xmm0[8],xmm1[8],xmm0[9],xmm1[9],xmm0[10],xmm1[10],xmm0[11],xmm1[11],xmm0[12],xmm1[12],xmm0[13],xmm1[13],xmm0[14],xmm1[14],xmm0[15],xmm1[15]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpaddb %xmm3, %xmm2, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xe9,0xfc,0xc3]
[AVX512] Remove masked unpack intrinsics and autoupgrade to vectorshuffle and selects. llvm-svn: 273543
2016-06-23 15:37:33 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <16 x i8> @llvm.x86.avx512.mask.punpckhb.w.128(<16 x i8> %x0, <16 x i8> %x1, <16 x i8> %x2, i16 %x3)
%res1 = call <16 x i8> @llvm.x86.avx512.mask.punpckhb.w.128(<16 x i8> %x0, <16 x i8> %x1, <16 x i8> %x2, i16 -1)
%res2 = add <16 x i8> %res, %res1
ret <16 x i8> %res2
}
declare <16 x i8> @llvm.x86.avx512.mask.punpcklb.w.128(<16 x i8>, <16 x i8>, <16 x i8>, i16)
define <16 x i8>@test_int_x86_avx512_mask_punpcklb_w_128(<16 x i8> %x0, <16 x i8> %x1, <16 x i8> %x2, i16 %x3) {
; CHECK-LABEL: test_int_x86_avx512_mask_punpcklb_w_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpunpcklbw %xmm1, %xmm0, %xmm3 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0x60,0xd9]
[AVX512] Remove masked unpack intrinsics and autoupgrade to vectorshuffle and selects. llvm-svn: 273543
2016-06-23 15:37:33 +08:00
; CHECK-NEXT: ## xmm3 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3],xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
[AVX512] Remove masked unpack intrinsics and autoupgrade to vectorshuffle and selects. llvm-svn: 273543
2016-06-23 15:37:33 +08:00
; CHECK-NEXT: vpunpcklbw %xmm1, %xmm0, %xmm2 {%k1} ## encoding: [0x62,0xf1,0x7d,0x09,0x60,0xd1]
[X86][AVX512] Add support for UNPCK masked shuffle comments llvm-svn: 274464
2016-07-03 22:26:21 +08:00
; CHECK-NEXT: ## xmm2 {%k1} = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3],xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpaddb %xmm3, %xmm2, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xe9,0xfc,0xc3]
[AVX512] Remove masked unpack intrinsics and autoupgrade to vectorshuffle and selects. llvm-svn: 273543
2016-06-23 15:37:33 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <16 x i8> @llvm.x86.avx512.mask.punpcklb.w.128(<16 x i8> %x0, <16 x i8> %x1, <16 x i8> %x2, i16 %x3)
%res1 = call <16 x i8> @llvm.x86.avx512.mask.punpcklb.w.128(<16 x i8> %x0, <16 x i8> %x1, <16 x i8> %x2, i16 -1)
%res2 = add <16 x i8> %res, %res1
ret <16 x i8> %res2
}
declare <32 x i8> @llvm.x86.avx512.mask.punpckhb.w.256(<32 x i8>, <32 x i8>, <32 x i8>, i32)
define <32 x i8>@test_int_x86_avx512_mask_punpckhb_w_256(<32 x i8> %x0, <32 x i8> %x1, <32 x i8> %x2, i32 %x3) {
; CHECK-LABEL: test_int_x86_avx512_mask_punpckhb_w_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpunpckhbw %ymm1, %ymm0, %ymm3 ## EVEX TO VEX Compression encoding: [0xc5,0xfd,0x68,0xd9]
[AVX512] Remove masked unpack intrinsics and autoupgrade to vectorshuffle and selects. llvm-svn: 273543
2016-06-23 15:37:33 +08:00
; CHECK-NEXT: ## ymm3 = ymm0[8],ymm1[8],ymm0[9],ymm1[9],ymm0[10],ymm1[10],ymm0[11],ymm1[11],ymm0[12],ymm1[12],ymm0[13],ymm1[13],ymm0[14],ymm1[14],ymm0[15],ymm1[15],ymm0[24],ymm1[24],ymm0[25],ymm1[25],ymm0[26],ymm1[26],ymm0[27],ymm1[27],ymm0[28],ymm1[28],ymm0[29],ymm1[29],ymm0[30],ymm1[30],ymm0[31],ymm1[31]
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
; CHECK-NEXT: vpunpckhbw %ymm1, %ymm0, %ymm2 {%k1} ## encoding: [0x62,0xf1,0x7d,0x29,0x68,0xd1]
[X86][AVX512] Add support for UNPCK masked shuffle comments llvm-svn: 274464
2016-07-03 22:26:21 +08:00
; CHECK-NEXT: ## ymm2 {%k1} = ymm0[8],ymm1[8],ymm0[9],ymm1[9],ymm0[10],ymm1[10],ymm0[11],ymm1[11],ymm0[12],ymm1[12],ymm0[13],ymm1[13],ymm0[14],ymm1[14],ymm0[15],ymm1[15],ymm0[24],ymm1[24],ymm0[25],ymm1[25],ymm0[26],ymm1[26],ymm0[27],ymm1[27],ymm0[28],ymm1[28],ymm0[29],ymm1[29],ymm0[30],ymm1[30],ymm0[31],ymm1[31]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpaddb %ymm3, %ymm2, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xed,0xfc,0xc3]
[AVX512] Remove masked unpack intrinsics and autoupgrade to vectorshuffle and selects. llvm-svn: 273543
2016-06-23 15:37:33 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <32 x i8> @llvm.x86.avx512.mask.punpckhb.w.256(<32 x i8> %x0, <32 x i8> %x1, <32 x i8> %x2, i32 %x3)
%res1 = call <32 x i8> @llvm.x86.avx512.mask.punpckhb.w.256(<32 x i8> %x0, <32 x i8> %x1, <32 x i8> %x2, i32 -1)
%res2 = add <32 x i8> %res, %res1
ret <32 x i8> %res2
}
declare <32 x i8> @llvm.x86.avx512.mask.punpcklb.w.256(<32 x i8>, <32 x i8>, <32 x i8>, i32)
define <32 x i8>@test_int_x86_avx512_mask_punpcklb_w_256(<32 x i8> %x0, <32 x i8> %x1, <32 x i8> %x2, i32 %x3) {
; CHECK-LABEL: test_int_x86_avx512_mask_punpcklb_w_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpunpcklbw %ymm1, %ymm0, %ymm3 ## EVEX TO VEX Compression encoding: [0xc5,0xfd,0x60,0xd9]
[AVX512] Remove masked unpack intrinsics and autoupgrade to vectorshuffle and selects. llvm-svn: 273543
2016-06-23 15:37:33 +08:00
; CHECK-NEXT: ## ymm3 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[4],ymm1[4],ymm0[5],ymm1[5],ymm0[6],ymm1[6],ymm0[7],ymm1[7],ymm0[16],ymm1[16],ymm0[17],ymm1[17],ymm0[18],ymm1[18],ymm0[19],ymm1[19],ymm0[20],ymm1[20],ymm0[21],ymm1[21],ymm0[22],ymm1[22],ymm0[23],ymm1[23]
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
; CHECK-NEXT: vpunpcklbw %ymm1, %ymm0, %ymm2 {%k1} ## encoding: [0x62,0xf1,0x7d,0x29,0x60,0xd1]
[X86][AVX512] Add support for UNPCK masked shuffle comments llvm-svn: 274464
2016-07-03 22:26:21 +08:00
; CHECK-NEXT: ## ymm2 {%k1} = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[4],ymm1[4],ymm0[5],ymm1[5],ymm0[6],ymm1[6],ymm0[7],ymm1[7],ymm0[16],ymm1[16],ymm0[17],ymm1[17],ymm0[18],ymm1[18],ymm0[19],ymm1[19],ymm0[20],ymm1[20],ymm0[21],ymm1[21],ymm0[22],ymm1[22],ymm0[23],ymm1[23]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpaddb %ymm3, %ymm2, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xed,0xfc,0xc3]
[AVX512] Remove masked unpack intrinsics and autoupgrade to vectorshuffle and selects. llvm-svn: 273543
2016-06-23 15:37:33 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <32 x i8> @llvm.x86.avx512.mask.punpcklb.w.256(<32 x i8> %x0, <32 x i8> %x1, <32 x i8> %x2, i32 %x3)
%res1 = call <32 x i8> @llvm.x86.avx512.mask.punpcklb.w.256(<32 x i8> %x0, <32 x i8> %x1, <32 x i8> %x2, i32 -1)
%res2 = add <32 x i8> %res, %res1
ret <32 x i8> %res2
}
declare <8 x i16> @llvm.x86.avx512.mask.punpcklw.d.128(<8 x i16>, <8 x i16>, <8 x i16>, i8)
define <8 x i16>@test_int_x86_avx512_mask_punpcklw_d_128(<8 x i16> %x0, <8 x i16> %x1, <8 x i16> %x2, i8 %x3) {
; CHECK-LABEL: test_int_x86_avx512_mask_punpcklw_d_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpunpcklwd %xmm1, %xmm0, %xmm3 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0x61,0xd9]
[AVX512] Remove masked unpack intrinsics and autoupgrade to vectorshuffle and selects. llvm-svn: 273543
2016-06-23 15:37:33 +08:00
; CHECK-NEXT: ## xmm3 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
[AVX512] Remove masked unpack intrinsics and autoupgrade to vectorshuffle and selects. llvm-svn: 273543
2016-06-23 15:37:33 +08:00
; CHECK-NEXT: vpunpcklwd %xmm1, %xmm0, %xmm2 {%k1} ## encoding: [0x62,0xf1,0x7d,0x09,0x61,0xd1]
[X86][AVX512] Add support for UNPCK masked shuffle comments llvm-svn: 274464
2016-07-03 22:26:21 +08:00
; CHECK-NEXT: ## xmm2 {%k1} = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpaddw %xmm3, %xmm2, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xe9,0xfd,0xc3]
[AVX512] Remove masked unpack intrinsics and autoupgrade to vectorshuffle and selects. llvm-svn: 273543
2016-06-23 15:37:33 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <8 x i16> @llvm.x86.avx512.mask.punpcklw.d.128(<8 x i16> %x0, <8 x i16> %x1, <8 x i16> %x2, i8 %x3)
%res1 = call <8 x i16> @llvm.x86.avx512.mask.punpcklw.d.128(<8 x i16> %x0, <8 x i16> %x1, <8 x i16> %x2, i8 -1)
%res2 = add <8 x i16> %res, %res1
ret <8 x i16> %res2
}
declare <8 x i16> @llvm.x86.avx512.mask.punpckhw.d.128(<8 x i16>, <8 x i16>, <8 x i16>, i8)
define <8 x i16>@test_int_x86_avx512_mask_punpckhw_d_128(<8 x i16> %x0, <8 x i16> %x1, <8 x i16> %x2, i8 %x3) {
; CHECK-LABEL: test_int_x86_avx512_mask_punpckhw_d_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpunpckhwd %xmm1, %xmm0, %xmm3 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0x69,0xd9]
[AVX512] Remove masked unpack intrinsics and autoupgrade to vectorshuffle and selects. llvm-svn: 273543
2016-06-23 15:37:33 +08:00
; CHECK-NEXT: ## xmm3 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
[AVX512] Remove masked unpack intrinsics and autoupgrade to vectorshuffle and selects. llvm-svn: 273543
2016-06-23 15:37:33 +08:00
; CHECK-NEXT: vpunpckhwd %xmm1, %xmm0, %xmm2 {%k1} ## encoding: [0x62,0xf1,0x7d,0x09,0x69,0xd1]
[X86][AVX512] Add support for UNPCK masked shuffle comments llvm-svn: 274464
2016-07-03 22:26:21 +08:00
; CHECK-NEXT: ## xmm2 {%k1} = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpaddw %xmm3, %xmm2, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xe9,0xfd,0xc3]
[AVX512] Remove masked unpack intrinsics and autoupgrade to vectorshuffle and selects. llvm-svn: 273543
2016-06-23 15:37:33 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <8 x i16> @llvm.x86.avx512.mask.punpckhw.d.128(<8 x i16> %x0, <8 x i16> %x1, <8 x i16> %x2, i8 %x3)
%res1 = call <8 x i16> @llvm.x86.avx512.mask.punpckhw.d.128(<8 x i16> %x0, <8 x i16> %x1, <8 x i16> %x2, i8 -1)
%res2 = add <8 x i16> %res, %res1
ret <8 x i16> %res2
}
declare <16 x i16> @llvm.x86.avx512.mask.punpcklw.d.256(<16 x i16>, <16 x i16>, <16 x i16>, i16)
define <16 x i16>@test_int_x86_avx512_mask_punpcklw_d_256(<16 x i16> %x0, <16 x i16> %x1, <16 x i16> %x2, i16 %x3) {
; CHECK-LABEL: test_int_x86_avx512_mask_punpcklw_d_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpunpcklwd %ymm1, %ymm0, %ymm3 ## EVEX TO VEX Compression encoding: [0xc5,0xfd,0x61,0xd9]
[AVX512] Remove masked unpack intrinsics and autoupgrade to vectorshuffle and selects. llvm-svn: 273543
2016-06-23 15:37:33 +08:00
; CHECK-NEXT: ## ymm3 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[8],ymm1[8],ymm0[9],ymm1[9],ymm0[10],ymm1[10],ymm0[11],ymm1[11]
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
[AVX512] Remove masked unpack intrinsics and autoupgrade to vectorshuffle and selects. llvm-svn: 273543
2016-06-23 15:37:33 +08:00
; CHECK-NEXT: vpunpcklwd %ymm1, %ymm0, %ymm2 {%k1} ## encoding: [0x62,0xf1,0x7d,0x29,0x61,0xd1]
[X86][AVX512] Add support for UNPCK masked shuffle comments llvm-svn: 274464
2016-07-03 22:26:21 +08:00
; CHECK-NEXT: ## ymm2 {%k1} = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[8],ymm1[8],ymm0[9],ymm1[9],ymm0[10],ymm1[10],ymm0[11],ymm1[11]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpaddw %ymm3, %ymm2, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xed,0xfd,0xc3]
[AVX512] Remove masked unpack intrinsics and autoupgrade to vectorshuffle and selects. llvm-svn: 273543
2016-06-23 15:37:33 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <16 x i16> @llvm.x86.avx512.mask.punpcklw.d.256(<16 x i16> %x0, <16 x i16> %x1, <16 x i16> %x2, i16 %x3)
%res1 = call <16 x i16> @llvm.x86.avx512.mask.punpcklw.d.256(<16 x i16> %x0, <16 x i16> %x1, <16 x i16> %x2, i16 -1)
%res2 = add <16 x i16> %res, %res1
ret <16 x i16> %res2
}
declare <16 x i16> @llvm.x86.avx512.mask.punpckhw.d.256(<16 x i16>, <16 x i16>, <16 x i16>, i16)
define <16 x i16>@test_int_x86_avx512_mask_punpckhw_d_256(<16 x i16> %x0, <16 x i16> %x1, <16 x i16> %x2, i16 %x3) {
; CHECK-LABEL: test_int_x86_avx512_mask_punpckhw_d_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpunpckhwd %ymm1, %ymm0, %ymm3 ## EVEX TO VEX Compression encoding: [0xc5,0xfd,0x69,0xd9]
[AVX512] Remove masked unpack intrinsics and autoupgrade to vectorshuffle and selects. llvm-svn: 273543
2016-06-23 15:37:33 +08:00
; CHECK-NEXT: ## ymm3 = ymm0[4],ymm1[4],ymm0[5],ymm1[5],ymm0[6],ymm1[6],ymm0[7],ymm1[7],ymm0[12],ymm1[12],ymm0[13],ymm1[13],ymm0[14],ymm1[14],ymm0[15],ymm1[15]
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
[AVX512] Remove masked unpack intrinsics and autoupgrade to vectorshuffle and selects. llvm-svn: 273543
2016-06-23 15:37:33 +08:00
; CHECK-NEXT: vpunpckhwd %ymm1, %ymm0, %ymm2 {%k1} ## encoding: [0x62,0xf1,0x7d,0x29,0x69,0xd1]
[X86][AVX512] Add support for UNPCK masked shuffle comments llvm-svn: 274464
2016-07-03 22:26:21 +08:00
; CHECK-NEXT: ## ymm2 {%k1} = ymm0[4],ymm1[4],ymm0[5],ymm1[5],ymm0[6],ymm1[6],ymm0[7],ymm1[7],ymm0[12],ymm1[12],ymm0[13],ymm1[13],ymm0[14],ymm1[14],ymm0[15],ymm1[15]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpaddw %ymm3, %ymm2, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xed,0xfd,0xc3]
[AVX512] Remove masked unpack intrinsics and autoupgrade to vectorshuffle and selects. llvm-svn: 273543
2016-06-23 15:37:33 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <16 x i16> @llvm.x86.avx512.mask.punpckhw.d.256(<16 x i16> %x0, <16 x i16> %x1, <16 x i16> %x2, i16 %x3)
%res1 = call <16 x i16> @llvm.x86.avx512.mask.punpckhw.d.256(<16 x i16> %x0, <16 x i16> %x1, <16 x i16> %x2, i16 -1)
%res2 = add <16 x i16> %res, %res1
ret <16 x i16> %res2
}
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
define <8 x i16> @test_mask_add_epi16_rr_128(<8 x i16> %a, <8 x i16> %b) {
; CHECK-LABEL: test_mask_add_epi16_rr_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpaddw %xmm1, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xfd,0xc1]
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <8 x i16> @llvm.x86.avx512.mask.padd.w.128(<8 x i16> %a, <8 x i16> %b, <8 x i16> zeroinitializer, i8 -1)
ret <8 x i16> %res
}
define <8 x i16> @test_mask_add_epi16_rrk_128(<8 x i16> %a, <8 x i16> %b, <8 x i16> %passThru, i8 %mask) {
; CHECK-LABEL: test_mask_add_epi16_rrk_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: vpaddw %xmm1, %xmm0, %xmm2 {%k1} ## encoding: [0x62,0xf1,0x7d,0x09,0xfd,0xd1]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vmovdqa %xmm2, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6f,0xc2]
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <8 x i16> @llvm.x86.avx512.mask.padd.w.128(<8 x i16> %a, <8 x i16> %b, <8 x i16> %passThru, i8 %mask)
ret <8 x i16> %res
}
define <8 x i16> @test_mask_add_epi16_rrkz_128(<8 x i16> %a, <8 x i16> %b, i8 %mask) {
; CHECK-LABEL: test_mask_add_epi16_rrkz_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: vpaddw %xmm1, %xmm0, %xmm0 {%k1} {z} ## encoding: [0x62,0xf1,0x7d,0x89,0xfd,0xc1]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <8 x i16> @llvm.x86.avx512.mask.padd.w.128(<8 x i16> %a, <8 x i16> %b, <8 x i16> zeroinitializer, i8 %mask)
ret <8 x i16> %res
}
define <8 x i16> @test_mask_add_epi16_rm_128(<8 x i16> %a, <8 x i16>* %ptr_b) {
; CHECK-LABEL: test_mask_add_epi16_rm_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpaddw (%rdi), %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xfd,0x07]
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%b = load <8 x i16>, <8 x i16>* %ptr_b
%res = call <8 x i16> @llvm.x86.avx512.mask.padd.w.128(<8 x i16> %a, <8 x i16> %b, <8 x i16> zeroinitializer, i8 -1)
ret <8 x i16> %res
}
define <8 x i16> @test_mask_add_epi16_rmk_128(<8 x i16> %a, <8 x i16>* %ptr_b, <8 x i16> %passThru, i8 %mask) {
; CHECK-LABEL: test_mask_add_epi16_rmk_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %esi, %k1 ## encoding: [0xc5,0xfb,0x92,0xce]
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: vpaddw (%rdi), %xmm0, %xmm1 {%k1} ## encoding: [0x62,0xf1,0x7d,0x09,0xfd,0x0f]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vmovdqa %xmm1, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6f,0xc1]
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%b = load <8 x i16>, <8 x i16>* %ptr_b
%res = call <8 x i16> @llvm.x86.avx512.mask.padd.w.128(<8 x i16> %a, <8 x i16> %b, <8 x i16> %passThru, i8 %mask)
ret <8 x i16> %res
}
define <8 x i16> @test_mask_add_epi16_rmkz_128(<8 x i16> %a, <8 x i16>* %ptr_b, i8 %mask) {
; CHECK-LABEL: test_mask_add_epi16_rmkz_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %esi, %k1 ## encoding: [0xc5,0xfb,0x92,0xce]
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: vpaddw (%rdi), %xmm0, %xmm0 {%k1} {z} ## encoding: [0x62,0xf1,0x7d,0x89,0xfd,0x07]
; CHECK-NEXT: retq ## encoding: [0xc3]
%b = load <8 x i16>, <8 x i16>* %ptr_b
%res = call <8 x i16> @llvm.x86.avx512.mask.padd.w.128(<8 x i16> %a, <8 x i16> %b, <8 x i16> zeroinitializer, i8 %mask)
ret <8 x i16> %res
}
declare <8 x i16> @llvm.x86.avx512.mask.padd.w.128(<8 x i16>, <8 x i16>, <8 x i16>, i8)
define <16 x i16> @test_mask_add_epi16_rr_256(<16 x i16> %a, <16 x i16> %b) {
; CHECK-LABEL: test_mask_add_epi16_rr_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpaddw %ymm1, %ymm0, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xfd,0xfd,0xc1]
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <16 x i16> @llvm.x86.avx512.mask.padd.w.256(<16 x i16> %a, <16 x i16> %b, <16 x i16> zeroinitializer, i16 -1)
ret <16 x i16> %res
}
define <16 x i16> @test_mask_add_epi16_rrk_256(<16 x i16> %a, <16 x i16> %b, <16 x i16> %passThru, i16 %mask) {
; CHECK-LABEL: test_mask_add_epi16_rrk_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: vpaddw %ymm1, %ymm0, %ymm2 {%k1} ## encoding: [0x62,0xf1,0x7d,0x29,0xfd,0xd1]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vmovdqa %ymm2, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xfd,0x6f,0xc2]
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <16 x i16> @llvm.x86.avx512.mask.padd.w.256(<16 x i16> %a, <16 x i16> %b, <16 x i16> %passThru, i16 %mask)
ret <16 x i16> %res
}
define <16 x i16> @test_mask_add_epi16_rrkz_256(<16 x i16> %a, <16 x i16> %b, i16 %mask) {
; CHECK-LABEL: test_mask_add_epi16_rrkz_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: vpaddw %ymm1, %ymm0, %ymm0 {%k1} {z} ## encoding: [0x62,0xf1,0x7d,0xa9,0xfd,0xc1]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <16 x i16> @llvm.x86.avx512.mask.padd.w.256(<16 x i16> %a, <16 x i16> %b, <16 x i16> zeroinitializer, i16 %mask)
ret <16 x i16> %res
}
define <16 x i16> @test_mask_add_epi16_rm_256(<16 x i16> %a, <16 x i16>* %ptr_b) {
; CHECK-LABEL: test_mask_add_epi16_rm_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpaddw (%rdi), %ymm0, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xfd,0xfd,0x07]
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%b = load <16 x i16>, <16 x i16>* %ptr_b
%res = call <16 x i16> @llvm.x86.avx512.mask.padd.w.256(<16 x i16> %a, <16 x i16> %b, <16 x i16> zeroinitializer, i16 -1)
ret <16 x i16> %res
}
define <16 x i16> @test_mask_add_epi16_rmk_256(<16 x i16> %a, <16 x i16>* %ptr_b, <16 x i16> %passThru, i16 %mask) {
; CHECK-LABEL: test_mask_add_epi16_rmk_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %esi, %k1 ## encoding: [0xc5,0xfb,0x92,0xce]
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: vpaddw (%rdi), %ymm0, %ymm1 {%k1} ## encoding: [0x62,0xf1,0x7d,0x29,0xfd,0x0f]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vmovdqa %ymm1, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xfd,0x6f,0xc1]
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%b = load <16 x i16>, <16 x i16>* %ptr_b
%res = call <16 x i16> @llvm.x86.avx512.mask.padd.w.256(<16 x i16> %a, <16 x i16> %b, <16 x i16> %passThru, i16 %mask)
ret <16 x i16> %res
}
define <16 x i16> @test_mask_add_epi16_rmkz_256(<16 x i16> %a, <16 x i16>* %ptr_b, i16 %mask) {
; CHECK-LABEL: test_mask_add_epi16_rmkz_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %esi, %k1 ## encoding: [0xc5,0xfb,0x92,0xce]
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: vpaddw (%rdi), %ymm0, %ymm0 {%k1} {z} ## encoding: [0x62,0xf1,0x7d,0xa9,0xfd,0x07]
; CHECK-NEXT: retq ## encoding: [0xc3]
%b = load <16 x i16>, <16 x i16>* %ptr_b
%res = call <16 x i16> @llvm.x86.avx512.mask.padd.w.256(<16 x i16> %a, <16 x i16> %b, <16 x i16> zeroinitializer, i16 %mask)
ret <16 x i16> %res
}
declare <16 x i16> @llvm.x86.avx512.mask.padd.w.256(<16 x i16>, <16 x i16>, <16 x i16>, i16)
define <8 x i16> @test_mask_sub_epi16_rr_128(<8 x i16> %a, <8 x i16> %b) {
; CHECK-LABEL: test_mask_sub_epi16_rr_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpsubw %xmm1, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xf9,0xc1]
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <8 x i16> @llvm.x86.avx512.mask.psub.w.128(<8 x i16> %a, <8 x i16> %b, <8 x i16> zeroinitializer, i8 -1)
ret <8 x i16> %res
}
define <8 x i16> @test_mask_sub_epi16_rrk_128(<8 x i16> %a, <8 x i16> %b, <8 x i16> %passThru, i8 %mask) {
; CHECK-LABEL: test_mask_sub_epi16_rrk_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: vpsubw %xmm1, %xmm0, %xmm2 {%k1} ## encoding: [0x62,0xf1,0x7d,0x09,0xf9,0xd1]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vmovdqa %xmm2, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6f,0xc2]
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <8 x i16> @llvm.x86.avx512.mask.psub.w.128(<8 x i16> %a, <8 x i16> %b, <8 x i16> %passThru, i8 %mask)
ret <8 x i16> %res
}
define <8 x i16> @test_mask_sub_epi16_rrkz_128(<8 x i16> %a, <8 x i16> %b, i8 %mask) {
; CHECK-LABEL: test_mask_sub_epi16_rrkz_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: vpsubw %xmm1, %xmm0, %xmm0 {%k1} {z} ## encoding: [0x62,0xf1,0x7d,0x89,0xf9,0xc1]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <8 x i16> @llvm.x86.avx512.mask.psub.w.128(<8 x i16> %a, <8 x i16> %b, <8 x i16> zeroinitializer, i8 %mask)
ret <8 x i16> %res
}
define <8 x i16> @test_mask_sub_epi16_rm_128(<8 x i16> %a, <8 x i16>* %ptr_b) {
; CHECK-LABEL: test_mask_sub_epi16_rm_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpsubw (%rdi), %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xf9,0x07]
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%b = load <8 x i16>, <8 x i16>* %ptr_b
%res = call <8 x i16> @llvm.x86.avx512.mask.psub.w.128(<8 x i16> %a, <8 x i16> %b, <8 x i16> zeroinitializer, i8 -1)
ret <8 x i16> %res
}
define <8 x i16> @test_mask_sub_epi16_rmk_128(<8 x i16> %a, <8 x i16>* %ptr_b, <8 x i16> %passThru, i8 %mask) {
; CHECK-LABEL: test_mask_sub_epi16_rmk_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %esi, %k1 ## encoding: [0xc5,0xfb,0x92,0xce]
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: vpsubw (%rdi), %xmm0, %xmm1 {%k1} ## encoding: [0x62,0xf1,0x7d,0x09,0xf9,0x0f]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vmovdqa %xmm1, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6f,0xc1]
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%b = load <8 x i16>, <8 x i16>* %ptr_b
%res = call <8 x i16> @llvm.x86.avx512.mask.psub.w.128(<8 x i16> %a, <8 x i16> %b, <8 x i16> %passThru, i8 %mask)
ret <8 x i16> %res
}
define <8 x i16> @test_mask_sub_epi16_rmkz_128(<8 x i16> %a, <8 x i16>* %ptr_b, i8 %mask) {
; CHECK-LABEL: test_mask_sub_epi16_rmkz_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %esi, %k1 ## encoding: [0xc5,0xfb,0x92,0xce]
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: vpsubw (%rdi), %xmm0, %xmm0 {%k1} {z} ## encoding: [0x62,0xf1,0x7d,0x89,0xf9,0x07]
; CHECK-NEXT: retq ## encoding: [0xc3]
%b = load <8 x i16>, <8 x i16>* %ptr_b
%res = call <8 x i16> @llvm.x86.avx512.mask.psub.w.128(<8 x i16> %a, <8 x i16> %b, <8 x i16> zeroinitializer, i8 %mask)
ret <8 x i16> %res
}
declare <8 x i16> @llvm.x86.avx512.mask.psub.w.128(<8 x i16>, <8 x i16>, <8 x i16>, i8)
define <16 x i16> @test_mask_sub_epi16_rr_256(<16 x i16> %a, <16 x i16> %b) {
; CHECK-LABEL: test_mask_sub_epi16_rr_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpsubw %ymm1, %ymm0, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xfd,0xf9,0xc1]
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <16 x i16> @llvm.x86.avx512.mask.psub.w.256(<16 x i16> %a, <16 x i16> %b, <16 x i16> zeroinitializer, i16 -1)
ret <16 x i16> %res
}
define <16 x i16> @test_mask_sub_epi16_rrk_256(<16 x i16> %a, <16 x i16> %b, <16 x i16> %passThru, i16 %mask) {
; CHECK-LABEL: test_mask_sub_epi16_rrk_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: vpsubw %ymm1, %ymm0, %ymm2 {%k1} ## encoding: [0x62,0xf1,0x7d,0x29,0xf9,0xd1]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vmovdqa %ymm2, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xfd,0x6f,0xc2]
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <16 x i16> @llvm.x86.avx512.mask.psub.w.256(<16 x i16> %a, <16 x i16> %b, <16 x i16> %passThru, i16 %mask)
ret <16 x i16> %res
}
define <16 x i16> @test_mask_sub_epi16_rrkz_256(<16 x i16> %a, <16 x i16> %b, i16 %mask) {
; CHECK-LABEL: test_mask_sub_epi16_rrkz_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: vpsubw %ymm1, %ymm0, %ymm0 {%k1} {z} ## encoding: [0x62,0xf1,0x7d,0xa9,0xf9,0xc1]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <16 x i16> @llvm.x86.avx512.mask.psub.w.256(<16 x i16> %a, <16 x i16> %b, <16 x i16> zeroinitializer, i16 %mask)
ret <16 x i16> %res
}
define <16 x i16> @test_mask_sub_epi16_rm_256(<16 x i16> %a, <16 x i16>* %ptr_b) {
; CHECK-LABEL: test_mask_sub_epi16_rm_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpsubw (%rdi), %ymm0, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xfd,0xf9,0x07]
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%b = load <16 x i16>, <16 x i16>* %ptr_b
%res = call <16 x i16> @llvm.x86.avx512.mask.psub.w.256(<16 x i16> %a, <16 x i16> %b, <16 x i16> zeroinitializer, i16 -1)
ret <16 x i16> %res
}
define <16 x i16> @test_mask_sub_epi16_rmk_256(<16 x i16> %a, <16 x i16>* %ptr_b, <16 x i16> %passThru, i16 %mask) {
; CHECK-LABEL: test_mask_sub_epi16_rmk_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %esi, %k1 ## encoding: [0xc5,0xfb,0x92,0xce]
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: vpsubw (%rdi), %ymm0, %ymm1 {%k1} ## encoding: [0x62,0xf1,0x7d,0x29,0xf9,0x0f]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vmovdqa %ymm1, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xfd,0x6f,0xc1]
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%b = load <16 x i16>, <16 x i16>* %ptr_b
%res = call <16 x i16> @llvm.x86.avx512.mask.psub.w.256(<16 x i16> %a, <16 x i16> %b, <16 x i16> %passThru, i16 %mask)
ret <16 x i16> %res
}
define <16 x i16> @test_mask_sub_epi16_rmkz_256(<16 x i16> %a, <16 x i16>* %ptr_b, i16 %mask) {
; CHECK-LABEL: test_mask_sub_epi16_rmkz_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %esi, %k1 ## encoding: [0xc5,0xfb,0x92,0xce]
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: vpsubw (%rdi), %ymm0, %ymm0 {%k1} {z} ## encoding: [0x62,0xf1,0x7d,0xa9,0xf9,0x07]
; CHECK-NEXT: retq ## encoding: [0xc3]
%b = load <16 x i16>, <16 x i16>* %ptr_b
%res = call <16 x i16> @llvm.x86.avx512.mask.psub.w.256(<16 x i16> %a, <16 x i16> %b, <16 x i16> zeroinitializer, i16 %mask)
ret <16 x i16> %res
}
declare <16 x i16> @llvm.x86.avx512.mask.psub.w.256(<16 x i16>, <16 x i16>, <16 x i16>, i16)
define <32 x i16> @test_mask_add_epi16_rr_512(<32 x i16> %a, <32 x i16> %b) {
; CHECK-LABEL: test_mask_add_epi16_rr_512:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: vpaddw %zmm1, %zmm0, %zmm0 ## encoding: [0x62,0xf1,0x7d,0x48,0xfd,0xc1]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <32 x i16> @llvm.x86.avx512.mask.padd.w.512(<32 x i16> %a, <32 x i16> %b, <32 x i16> zeroinitializer, i32 -1)
ret <32 x i16> %res
}
define <32 x i16> @test_mask_add_epi16_rrk_512(<32 x i16> %a, <32 x i16> %b, <32 x i16> %passThru, i32 %mask) {
; CHECK-LABEL: test_mask_add_epi16_rrk_512:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
; CHECK-NEXT: vpaddw %zmm1, %zmm0, %zmm2 {%k1} ## encoding: [0x62,0xf1,0x7d,0x49,0xfd,0xd1]
; CHECK-NEXT: vmovdqa64 %zmm2, %zmm0 ## encoding: [0x62,0xf1,0xfd,0x48,0x6f,0xc2]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <32 x i16> @llvm.x86.avx512.mask.padd.w.512(<32 x i16> %a, <32 x i16> %b, <32 x i16> %passThru, i32 %mask)
ret <32 x i16> %res
}
define <32 x i16> @test_mask_add_epi16_rrkz_512(<32 x i16> %a, <32 x i16> %b, i32 %mask) {
; CHECK-LABEL: test_mask_add_epi16_rrkz_512:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
; CHECK-NEXT: vpaddw %zmm1, %zmm0, %zmm0 {%k1} {z} ## encoding: [0x62,0xf1,0x7d,0xc9,0xfd,0xc1]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <32 x i16> @llvm.x86.avx512.mask.padd.w.512(<32 x i16> %a, <32 x i16> %b, <32 x i16> zeroinitializer, i32 %mask)
ret <32 x i16> %res
}
define <32 x i16> @test_mask_add_epi16_rm_512(<32 x i16> %a, <32 x i16>* %ptr_b) {
; CHECK-LABEL: test_mask_add_epi16_rm_512:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: vpaddw (%rdi), %zmm0, %zmm0 ## encoding: [0x62,0xf1,0x7d,0x48,0xfd,0x07]
; CHECK-NEXT: retq ## encoding: [0xc3]
%b = load <32 x i16>, <32 x i16>* %ptr_b
%res = call <32 x i16> @llvm.x86.avx512.mask.padd.w.512(<32 x i16> %a, <32 x i16> %b, <32 x i16> zeroinitializer, i32 -1)
ret <32 x i16> %res
}
define <32 x i16> @test_mask_add_epi16_rmk_512(<32 x i16> %a, <32 x i16>* %ptr_b, <32 x i16> %passThru, i32 %mask) {
; CHECK-LABEL: test_mask_add_epi16_rmk_512:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: kmovd %esi, %k1 ## encoding: [0xc5,0xfb,0x92,0xce]
; CHECK-NEXT: vpaddw (%rdi), %zmm0, %zmm1 {%k1} ## encoding: [0x62,0xf1,0x7d,0x49,0xfd,0x0f]
; CHECK-NEXT: vmovdqa64 %zmm1, %zmm0 ## encoding: [0x62,0xf1,0xfd,0x48,0x6f,0xc1]
; CHECK-NEXT: retq ## encoding: [0xc3]
%b = load <32 x i16>, <32 x i16>* %ptr_b
%res = call <32 x i16> @llvm.x86.avx512.mask.padd.w.512(<32 x i16> %a, <32 x i16> %b, <32 x i16> %passThru, i32 %mask)
ret <32 x i16> %res
}
define <32 x i16> @test_mask_add_epi16_rmkz_512(<32 x i16> %a, <32 x i16>* %ptr_b, i32 %mask) {
; CHECK-LABEL: test_mask_add_epi16_rmkz_512:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: kmovd %esi, %k1 ## encoding: [0xc5,0xfb,0x92,0xce]
; CHECK-NEXT: vpaddw (%rdi), %zmm0, %zmm0 {%k1} {z} ## encoding: [0x62,0xf1,0x7d,0xc9,0xfd,0x07]
; CHECK-NEXT: retq ## encoding: [0xc3]
%b = load <32 x i16>, <32 x i16>* %ptr_b
%res = call <32 x i16> @llvm.x86.avx512.mask.padd.w.512(<32 x i16> %a, <32 x i16> %b, <32 x i16> zeroinitializer, i32 %mask)
ret <32 x i16> %res
}
declare <32 x i16> @llvm.x86.avx512.mask.padd.w.512(<32 x i16>, <32 x i16>, <32 x i16>, i32)
define <32 x i16> @test_mask_sub_epi16_rr_512(<32 x i16> %a, <32 x i16> %b) {
; CHECK-LABEL: test_mask_sub_epi16_rr_512:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: vpsubw %zmm1, %zmm0, %zmm0 ## encoding: [0x62,0xf1,0x7d,0x48,0xf9,0xc1]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <32 x i16> @llvm.x86.avx512.mask.psub.w.512(<32 x i16> %a, <32 x i16> %b, <32 x i16> zeroinitializer, i32 -1)
ret <32 x i16> %res
}
define <32 x i16> @test_mask_sub_epi16_rrk_512(<32 x i16> %a, <32 x i16> %b, <32 x i16> %passThru, i32 %mask) {
; CHECK-LABEL: test_mask_sub_epi16_rrk_512:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
; CHECK-NEXT: vpsubw %zmm1, %zmm0, %zmm2 {%k1} ## encoding: [0x62,0xf1,0x7d,0x49,0xf9,0xd1]
; CHECK-NEXT: vmovdqa64 %zmm2, %zmm0 ## encoding: [0x62,0xf1,0xfd,0x48,0x6f,0xc2]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <32 x i16> @llvm.x86.avx512.mask.psub.w.512(<32 x i16> %a, <32 x i16> %b, <32 x i16> %passThru, i32 %mask)
ret <32 x i16> %res
}
define <32 x i16> @test_mask_sub_epi16_rrkz_512(<32 x i16> %a, <32 x i16> %b, i32 %mask) {
; CHECK-LABEL: test_mask_sub_epi16_rrkz_512:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
; CHECK-NEXT: vpsubw %zmm1, %zmm0, %zmm0 {%k1} {z} ## encoding: [0x62,0xf1,0x7d,0xc9,0xf9,0xc1]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <32 x i16> @llvm.x86.avx512.mask.psub.w.512(<32 x i16> %a, <32 x i16> %b, <32 x i16> zeroinitializer, i32 %mask)
ret <32 x i16> %res
}
define <32 x i16> @test_mask_sub_epi16_rm_512(<32 x i16> %a, <32 x i16>* %ptr_b) {
; CHECK-LABEL: test_mask_sub_epi16_rm_512:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: vpsubw (%rdi), %zmm0, %zmm0 ## encoding: [0x62,0xf1,0x7d,0x48,0xf9,0x07]
; CHECK-NEXT: retq ## encoding: [0xc3]
%b = load <32 x i16>, <32 x i16>* %ptr_b
%res = call <32 x i16> @llvm.x86.avx512.mask.psub.w.512(<32 x i16> %a, <32 x i16> %b, <32 x i16> zeroinitializer, i32 -1)
ret <32 x i16> %res
}
define <32 x i16> @test_mask_sub_epi16_rmk_512(<32 x i16> %a, <32 x i16>* %ptr_b, <32 x i16> %passThru, i32 %mask) {
; CHECK-LABEL: test_mask_sub_epi16_rmk_512:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: kmovd %esi, %k1 ## encoding: [0xc5,0xfb,0x92,0xce]
; CHECK-NEXT: vpsubw (%rdi), %zmm0, %zmm1 {%k1} ## encoding: [0x62,0xf1,0x7d,0x49,0xf9,0x0f]
; CHECK-NEXT: vmovdqa64 %zmm1, %zmm0 ## encoding: [0x62,0xf1,0xfd,0x48,0x6f,0xc1]
; CHECK-NEXT: retq ## encoding: [0xc3]
%b = load <32 x i16>, <32 x i16>* %ptr_b
%res = call <32 x i16> @llvm.x86.avx512.mask.psub.w.512(<32 x i16> %a, <32 x i16> %b, <32 x i16> %passThru, i32 %mask)
ret <32 x i16> %res
}
define <32 x i16> @test_mask_sub_epi16_rmkz_512(<32 x i16> %a, <32 x i16>* %ptr_b, i32 %mask) {
; CHECK-LABEL: test_mask_sub_epi16_rmkz_512:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: kmovd %esi, %k1 ## encoding: [0xc5,0xfb,0x92,0xce]
; CHECK-NEXT: vpsubw (%rdi), %zmm0, %zmm0 {%k1} {z} ## encoding: [0x62,0xf1,0x7d,0xc9,0xf9,0x07]
; CHECK-NEXT: retq ## encoding: [0xc3]
%b = load <32 x i16>, <32 x i16>* %ptr_b
%res = call <32 x i16> @llvm.x86.avx512.mask.psub.w.512(<32 x i16> %a, <32 x i16> %b, <32 x i16> zeroinitializer, i32 %mask)
ret <32 x i16> %res
}
declare <32 x i16> @llvm.x86.avx512.mask.psub.w.512(<32 x i16>, <32 x i16>, <32 x i16>, i32)
define <32 x i16> @test_mask_mullo_epi16_rr_512(<32 x i16> %a, <32 x i16> %b) {
; CHECK-LABEL: test_mask_mullo_epi16_rr_512:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: vpmullw %zmm1, %zmm0, %zmm0 ## encoding: [0x62,0xf1,0x7d,0x48,0xd5,0xc1]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <32 x i16> @llvm.x86.avx512.mask.pmull.w.512(<32 x i16> %a, <32 x i16> %b, <32 x i16> zeroinitializer, i32 -1)
ret <32 x i16> %res
}
define <32 x i16> @test_mask_mullo_epi16_rrk_512(<32 x i16> %a, <32 x i16> %b, <32 x i16> %passThru, i32 %mask) {
; CHECK-LABEL: test_mask_mullo_epi16_rrk_512:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
; CHECK-NEXT: vpmullw %zmm1, %zmm0, %zmm2 {%k1} ## encoding: [0x62,0xf1,0x7d,0x49,0xd5,0xd1]
; CHECK-NEXT: vmovdqa64 %zmm2, %zmm0 ## encoding: [0x62,0xf1,0xfd,0x48,0x6f,0xc2]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <32 x i16> @llvm.x86.avx512.mask.pmull.w.512(<32 x i16> %a, <32 x i16> %b, <32 x i16> %passThru, i32 %mask)
ret <32 x i16> %res
}
define <32 x i16> @test_mask_mullo_epi16_rrkz_512(<32 x i16> %a, <32 x i16> %b, i32 %mask) {
; CHECK-LABEL: test_mask_mullo_epi16_rrkz_512:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
; CHECK-NEXT: vpmullw %zmm1, %zmm0, %zmm0 {%k1} {z} ## encoding: [0x62,0xf1,0x7d,0xc9,0xd5,0xc1]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <32 x i16> @llvm.x86.avx512.mask.pmull.w.512(<32 x i16> %a, <32 x i16> %b, <32 x i16> zeroinitializer, i32 %mask)
ret <32 x i16> %res
}
define <32 x i16> @test_mask_mullo_epi16_rm_512(<32 x i16> %a, <32 x i16>* %ptr_b) {
; CHECK-LABEL: test_mask_mullo_epi16_rm_512:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: vpmullw (%rdi), %zmm0, %zmm0 ## encoding: [0x62,0xf1,0x7d,0x48,0xd5,0x07]
; CHECK-NEXT: retq ## encoding: [0xc3]
%b = load <32 x i16>, <32 x i16>* %ptr_b
%res = call <32 x i16> @llvm.x86.avx512.mask.pmull.w.512(<32 x i16> %a, <32 x i16> %b, <32 x i16> zeroinitializer, i32 -1)
ret <32 x i16> %res
}
define <32 x i16> @test_mask_mullo_epi16_rmk_512(<32 x i16> %a, <32 x i16>* %ptr_b, <32 x i16> %passThru, i32 %mask) {
; CHECK-LABEL: test_mask_mullo_epi16_rmk_512:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: kmovd %esi, %k1 ## encoding: [0xc5,0xfb,0x92,0xce]
; CHECK-NEXT: vpmullw (%rdi), %zmm0, %zmm1 {%k1} ## encoding: [0x62,0xf1,0x7d,0x49,0xd5,0x0f]
; CHECK-NEXT: vmovdqa64 %zmm1, %zmm0 ## encoding: [0x62,0xf1,0xfd,0x48,0x6f,0xc1]
; CHECK-NEXT: retq ## encoding: [0xc3]
%b = load <32 x i16>, <32 x i16>* %ptr_b
%res = call <32 x i16> @llvm.x86.avx512.mask.pmull.w.512(<32 x i16> %a, <32 x i16> %b, <32 x i16> %passThru, i32 %mask)
ret <32 x i16> %res
}
define <32 x i16> @test_mask_mullo_epi16_rmkz_512(<32 x i16> %a, <32 x i16>* %ptr_b, i32 %mask) {
; CHECK-LABEL: test_mask_mullo_epi16_rmkz_512:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: kmovd %esi, %k1 ## encoding: [0xc5,0xfb,0x92,0xce]
; CHECK-NEXT: vpmullw (%rdi), %zmm0, %zmm0 {%k1} {z} ## encoding: [0x62,0xf1,0x7d,0xc9,0xd5,0x07]
; CHECK-NEXT: retq ## encoding: [0xc3]
%b = load <32 x i16>, <32 x i16>* %ptr_b
%res = call <32 x i16> @llvm.x86.avx512.mask.pmull.w.512(<32 x i16> %a, <32 x i16> %b, <32 x i16> zeroinitializer, i32 %mask)
ret <32 x i16> %res
}
declare <32 x i16> @llvm.x86.avx512.mask.pmull.w.512(<32 x i16>, <32 x i16>, <32 x i16>, i32)
define <8 x i16> @test_mask_mullo_epi16_rr_128(<8 x i16> %a, <8 x i16> %b) {
; CHECK-LABEL: test_mask_mullo_epi16_rr_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpmullw %xmm1, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xd5,0xc1]
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <8 x i16> @llvm.x86.avx512.mask.pmull.w.128(<8 x i16> %a, <8 x i16> %b, <8 x i16> zeroinitializer, i8 -1)
ret <8 x i16> %res
}
define <8 x i16> @test_mask_mullo_epi16_rrk_128(<8 x i16> %a, <8 x i16> %b, <8 x i16> %passThru, i8 %mask) {
; CHECK-LABEL: test_mask_mullo_epi16_rrk_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: vpmullw %xmm1, %xmm0, %xmm2 {%k1} ## encoding: [0x62,0xf1,0x7d,0x09,0xd5,0xd1]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vmovdqa %xmm2, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6f,0xc2]
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <8 x i16> @llvm.x86.avx512.mask.pmull.w.128(<8 x i16> %a, <8 x i16> %b, <8 x i16> %passThru, i8 %mask)
ret <8 x i16> %res
}
define <8 x i16> @test_mask_mullo_epi16_rrkz_128(<8 x i16> %a, <8 x i16> %b, i8 %mask) {
; CHECK-LABEL: test_mask_mullo_epi16_rrkz_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: vpmullw %xmm1, %xmm0, %xmm0 {%k1} {z} ## encoding: [0x62,0xf1,0x7d,0x89,0xd5,0xc1]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <8 x i16> @llvm.x86.avx512.mask.pmull.w.128(<8 x i16> %a, <8 x i16> %b, <8 x i16> zeroinitializer, i8 %mask)
ret <8 x i16> %res
}
define <8 x i16> @test_mask_mullo_epi16_rm_128(<8 x i16> %a, <8 x i16>* %ptr_b) {
; CHECK-LABEL: test_mask_mullo_epi16_rm_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpmullw (%rdi), %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xd5,0x07]
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%b = load <8 x i16>, <8 x i16>* %ptr_b
%res = call <8 x i16> @llvm.x86.avx512.mask.pmull.w.128(<8 x i16> %a, <8 x i16> %b, <8 x i16> zeroinitializer, i8 -1)
ret <8 x i16> %res
}
define <8 x i16> @test_mask_mullo_epi16_rmk_128(<8 x i16> %a, <8 x i16>* %ptr_b, <8 x i16> %passThru, i8 %mask) {
; CHECK-LABEL: test_mask_mullo_epi16_rmk_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %esi, %k1 ## encoding: [0xc5,0xfb,0x92,0xce]
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: vpmullw (%rdi), %xmm0, %xmm1 {%k1} ## encoding: [0x62,0xf1,0x7d,0x09,0xd5,0x0f]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vmovdqa %xmm1, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6f,0xc1]
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%b = load <8 x i16>, <8 x i16>* %ptr_b
%res = call <8 x i16> @llvm.x86.avx512.mask.pmull.w.128(<8 x i16> %a, <8 x i16> %b, <8 x i16> %passThru, i8 %mask)
ret <8 x i16> %res
}
define <8 x i16> @test_mask_mullo_epi16_rmkz_128(<8 x i16> %a, <8 x i16>* %ptr_b, i8 %mask) {
; CHECK-LABEL: test_mask_mullo_epi16_rmkz_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %esi, %k1 ## encoding: [0xc5,0xfb,0x92,0xce]
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: vpmullw (%rdi), %xmm0, %xmm0 {%k1} {z} ## encoding: [0x62,0xf1,0x7d,0x89,0xd5,0x07]
; CHECK-NEXT: retq ## encoding: [0xc3]
%b = load <8 x i16>, <8 x i16>* %ptr_b
%res = call <8 x i16> @llvm.x86.avx512.mask.pmull.w.128(<8 x i16> %a, <8 x i16> %b, <8 x i16> zeroinitializer, i8 %mask)
ret <8 x i16> %res
}
declare <8 x i16> @llvm.x86.avx512.mask.pmull.w.128(<8 x i16>, <8 x i16>, <8 x i16>, i8)
define <16 x i16> @test_mask_mullo_epi16_rr_256(<16 x i16> %a, <16 x i16> %b) {
; CHECK-LABEL: test_mask_mullo_epi16_rr_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpmullw %ymm1, %ymm0, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xfd,0xd5,0xc1]
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <16 x i16> @llvm.x86.avx512.mask.pmull.w.256(<16 x i16> %a, <16 x i16> %b, <16 x i16> zeroinitializer, i16 -1)
ret <16 x i16> %res
}
define <16 x i16> @test_mask_mullo_epi16_rrk_256(<16 x i16> %a, <16 x i16> %b, <16 x i16> %passThru, i16 %mask) {
; CHECK-LABEL: test_mask_mullo_epi16_rrk_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: vpmullw %ymm1, %ymm0, %ymm2 {%k1} ## encoding: [0x62,0xf1,0x7d,0x29,0xd5,0xd1]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vmovdqa %ymm2, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xfd,0x6f,0xc2]
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <16 x i16> @llvm.x86.avx512.mask.pmull.w.256(<16 x i16> %a, <16 x i16> %b, <16 x i16> %passThru, i16 %mask)
ret <16 x i16> %res
}
define <16 x i16> @test_mask_mullo_epi16_rrkz_256(<16 x i16> %a, <16 x i16> %b, i16 %mask) {
; CHECK-LABEL: test_mask_mullo_epi16_rrkz_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: vpmullw %ymm1, %ymm0, %ymm0 {%k1} {z} ## encoding: [0x62,0xf1,0x7d,0xa9,0xd5,0xc1]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <16 x i16> @llvm.x86.avx512.mask.pmull.w.256(<16 x i16> %a, <16 x i16> %b, <16 x i16> zeroinitializer, i16 %mask)
ret <16 x i16> %res
}
define <16 x i16> @test_mask_mullo_epi16_rm_256(<16 x i16> %a, <16 x i16>* %ptr_b) {
; CHECK-LABEL: test_mask_mullo_epi16_rm_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpmullw (%rdi), %ymm0, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xfd,0xd5,0x07]
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%b = load <16 x i16>, <16 x i16>* %ptr_b
%res = call <16 x i16> @llvm.x86.avx512.mask.pmull.w.256(<16 x i16> %a, <16 x i16> %b, <16 x i16> zeroinitializer, i16 -1)
ret <16 x i16> %res
}
define <16 x i16> @test_mask_mullo_epi16_rmk_256(<16 x i16> %a, <16 x i16>* %ptr_b, <16 x i16> %passThru, i16 %mask) {
; CHECK-LABEL: test_mask_mullo_epi16_rmk_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %esi, %k1 ## encoding: [0xc5,0xfb,0x92,0xce]
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: vpmullw (%rdi), %ymm0, %ymm1 {%k1} ## encoding: [0x62,0xf1,0x7d,0x29,0xd5,0x0f]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vmovdqa %ymm1, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xfd,0x6f,0xc1]
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%b = load <16 x i16>, <16 x i16>* %ptr_b
%res = call <16 x i16> @llvm.x86.avx512.mask.pmull.w.256(<16 x i16> %a, <16 x i16> %b, <16 x i16> %passThru, i16 %mask)
ret <16 x i16> %res
}
define <16 x i16> @test_mask_mullo_epi16_rmkz_256(<16 x i16> %a, <16 x i16>* %ptr_b, i16 %mask) {
; CHECK-LABEL: test_mask_mullo_epi16_rmkz_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %esi, %k1 ## encoding: [0xc5,0xfb,0x92,0xce]
[AVX-512] Remove masked integer add/sub/mull intrinsics and upgrade to native IR. llvm-svn: 280611
2016-09-04 10:09:53 +08:00
; CHECK-NEXT: vpmullw (%rdi), %ymm0, %ymm0 {%k1} {z} ## encoding: [0x62,0xf1,0x7d,0xa9,0xd5,0x07]
; CHECK-NEXT: retq ## encoding: [0xc3]
%b = load <16 x i16>, <16 x i16>* %ptr_b
%res = call <16 x i16> @llvm.x86.avx512.mask.pmull.w.256(<16 x i16> %a, <16 x i16> %b, <16 x i16> zeroinitializer, i16 %mask)
ret <16 x i16> %res
}
declare <16 x i16> @llvm.x86.avx512.mask.pmull.w.256(<16 x i16>, <16 x i16>, <16 x i16>, i16)
[AVX-512] Remove masked pmin/pmax intrinsics and autoupgrade to native IR. Clang patch to replace 512-bit vector and 64-bit element versions with native IR will follow. llvm-svn: 284955
2016-10-24 12:04:16 +08:00
declare <16 x i8> @llvm.x86.avx512.mask.pmaxs.b.128(<16 x i8>, <16 x i8>, <16 x i8>, i16)
define <16 x i8>@test_int_x86_avx512_mask_pmaxs_b_128(<16 x i8> %x0, <16 x i8> %x1, <16 x i8> %x2, i16 %mask) {
; CHECK-LABEL: test_int_x86_avx512_mask_pmaxs_b_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
[AVX-512] Remove masked pmin/pmax intrinsics and autoupgrade to native IR. Clang patch to replace 512-bit vector and 64-bit element versions with native IR will follow. llvm-svn: 284955
2016-10-24 12:04:16 +08:00
; CHECK-NEXT: vpmaxsb %xmm1, %xmm0, %xmm2 {%k1} ## encoding: [0x62,0xf2,0x7d,0x09,0x3c,0xd1]
; CHECK-NEXT: vpmaxsb %xmm1, %xmm0, %xmm0 {%k1} {z} ## encoding: [0x62,0xf2,0x7d,0x89,0x3c,0xc1]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpaddb %xmm0, %xmm2, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xe9,0xfc,0xc0]
[AVX-512] Remove masked pmin/pmax intrinsics and autoupgrade to native IR. Clang patch to replace 512-bit vector and 64-bit element versions with native IR will follow. llvm-svn: 284955
2016-10-24 12:04:16 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <16 x i8> @llvm.x86.avx512.mask.pmaxs.b.128(<16 x i8> %x0, <16 x i8> %x1, <16 x i8> %x2 ,i16 %mask)
%res1 = call <16 x i8> @llvm.x86.avx512.mask.pmaxs.b.128(<16 x i8> %x0, <16 x i8> %x1, <16 x i8> zeroinitializer, i16 %mask)
%res2 = add <16 x i8> %res, %res1
ret <16 x i8> %res2
}
declare <32 x i8> @llvm.x86.avx512.mask.pmaxs.b.256(<32 x i8>, <32 x i8>, <32 x i8>, i32)
define <32 x i8>@test_int_x86_avx512_mask_pmaxs_b_256(<32 x i8> %x0, <32 x i8> %x1, <32 x i8> %x2, i32 %x3) {
; CHECK-LABEL: test_int_x86_avx512_mask_pmaxs_b_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpmaxsb %ymm1, %ymm0, %ymm3 ## EVEX TO VEX Compression encoding: [0xc4,0xe2,0x7d,0x3c,0xd9]
[AVX-512] Remove masked pmin/pmax intrinsics and autoupgrade to native IR. Clang patch to replace 512-bit vector and 64-bit element versions with native IR will follow. llvm-svn: 284955
2016-10-24 12:04:16 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
; CHECK-NEXT: vpmaxsb %ymm1, %ymm0, %ymm2 {%k1} ## encoding: [0x62,0xf2,0x7d,0x29,0x3c,0xd1]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpaddb %ymm3, %ymm2, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xed,0xfc,0xc3]
[AVX-512] Remove masked pmin/pmax intrinsics and autoupgrade to native IR. Clang patch to replace 512-bit vector and 64-bit element versions with native IR will follow. llvm-svn: 284955
2016-10-24 12:04:16 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <32 x i8> @llvm.x86.avx512.mask.pmaxs.b.256(<32 x i8> %x0, <32 x i8> %x1, <32 x i8> %x2, i32 %x3)
%res1 = call <32 x i8> @llvm.x86.avx512.mask.pmaxs.b.256(<32 x i8> %x0, <32 x i8> %x1, <32 x i8> %x2, i32 -1)
%res2 = add <32 x i8> %res, %res1
ret <32 x i8> %res2
}
declare <8 x i16> @llvm.x86.avx512.mask.pmaxs.w.128(<8 x i16>, <8 x i16>, <8 x i16>, i8)
define <8 x i16>@test_int_x86_avx512_mask_pmaxs_w_128(<8 x i16> %x0, <8 x i16> %x1, <8 x i16> %x2, i8 %x3) {
; CHECK-LABEL: test_int_x86_avx512_mask_pmaxs_w_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpmaxsw %xmm1, %xmm0, %xmm3 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xee,0xd9]
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
[AVX-512] Remove masked pmin/pmax intrinsics and autoupgrade to native IR. Clang patch to replace 512-bit vector and 64-bit element versions with native IR will follow. llvm-svn: 284955
2016-10-24 12:04:16 +08:00
; CHECK-NEXT: vpmaxsw %xmm1, %xmm0, %xmm2 {%k1} ## encoding: [0x62,0xf1,0x7d,0x09,0xee,0xd1]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpaddw %xmm3, %xmm2, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xe9,0xfd,0xc3]
[AVX-512] Remove masked pmin/pmax intrinsics and autoupgrade to native IR. Clang patch to replace 512-bit vector and 64-bit element versions with native IR will follow. llvm-svn: 284955
2016-10-24 12:04:16 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <8 x i16> @llvm.x86.avx512.mask.pmaxs.w.128(<8 x i16> %x0, <8 x i16> %x1, <8 x i16> %x2, i8 %x3)
%res1 = call <8 x i16> @llvm.x86.avx512.mask.pmaxs.w.128(<8 x i16> %x0, <8 x i16> %x1, <8 x i16> %x2, i8 -1)
%res2 = add <8 x i16> %res, %res1
ret <8 x i16> %res2
}
declare <16 x i16> @llvm.x86.avx512.mask.pmaxs.w.256(<16 x i16>, <16 x i16>, <16 x i16>, i16)
define <16 x i16>@test_int_x86_avx512_mask_pmaxs_w_256(<16 x i16> %x0, <16 x i16> %x1, <16 x i16> %x2, i16 %mask) {
; CHECK-LABEL: test_int_x86_avx512_mask_pmaxs_w_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
[AVX-512] Remove masked pmin/pmax intrinsics and autoupgrade to native IR. Clang patch to replace 512-bit vector and 64-bit element versions with native IR will follow. llvm-svn: 284955
2016-10-24 12:04:16 +08:00
; CHECK-NEXT: vpmaxsw %ymm1, %ymm0, %ymm2 {%k1} ## encoding: [0x62,0xf1,0x7d,0x29,0xee,0xd1]
; CHECK-NEXT: vpmaxsw %ymm1, %ymm0, %ymm0 {%k1} {z} ## encoding: [0x62,0xf1,0x7d,0xa9,0xee,0xc1]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpaddw %ymm0, %ymm2, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xed,0xfd,0xc0]
[AVX-512] Remove masked pmin/pmax intrinsics and autoupgrade to native IR. Clang patch to replace 512-bit vector and 64-bit element versions with native IR will follow. llvm-svn: 284955
2016-10-24 12:04:16 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <16 x i16> @llvm.x86.avx512.mask.pmaxs.w.256(<16 x i16> %x0, <16 x i16> %x1, <16 x i16> %x2, i16 %mask)
%res1 = call <16 x i16> @llvm.x86.avx512.mask.pmaxs.w.256(<16 x i16> %x0, <16 x i16> %x1, <16 x i16> zeroinitializer, i16 %mask)
%res2 = add <16 x i16> %res, %res1
ret <16 x i16> %res2
}
declare <16 x i8> @llvm.x86.avx512.mask.pmaxu.b.128(<16 x i8>, <16 x i8>, <16 x i8>, i16)
define <16 x i8>@test_int_x86_avx512_mask_pmaxu_b_128(<16 x i8> %x0, <16 x i8> %x1, <16 x i8> %x2,i16 %mask) {
; CHECK-LABEL: test_int_x86_avx512_mask_pmaxu_b_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
[AVX-512] Remove masked pmin/pmax intrinsics and autoupgrade to native IR. Clang patch to replace 512-bit vector and 64-bit element versions with native IR will follow. llvm-svn: 284955
2016-10-24 12:04:16 +08:00
; CHECK-NEXT: vpmaxub %xmm1, %xmm0, %xmm2 {%k1} ## encoding: [0x62,0xf1,0x7d,0x09,0xde,0xd1]
; CHECK-NEXT: vpmaxub %xmm1, %xmm0, %xmm0 {%k1} {z} ## encoding: [0x62,0xf1,0x7d,0x89,0xde,0xc1]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpaddb %xmm0, %xmm2, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xe9,0xfc,0xc0]
[AVX-512] Remove masked pmin/pmax intrinsics and autoupgrade to native IR. Clang patch to replace 512-bit vector and 64-bit element versions with native IR will follow. llvm-svn: 284955
2016-10-24 12:04:16 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <16 x i8> @llvm.x86.avx512.mask.pmaxu.b.128(<16 x i8> %x0, <16 x i8> %x1, <16 x i8> %x2, i16 %mask)
%res1 = call <16 x i8> @llvm.x86.avx512.mask.pmaxu.b.128(<16 x i8> %x0, <16 x i8> %x1, <16 x i8> zeroinitializer, i16 %mask)
%res2 = add <16 x i8> %res, %res1
ret <16 x i8> %res2
}
declare <32 x i8> @llvm.x86.avx512.mask.pmaxu.b.256(<32 x i8>, <32 x i8>, <32 x i8>, i32)
define <32 x i8>@test_int_x86_avx512_mask_pmaxu_b_256(<32 x i8> %x0, <32 x i8> %x1, <32 x i8> %x2, i32 %x3) {
; CHECK-LABEL: test_int_x86_avx512_mask_pmaxu_b_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpmaxub %ymm1, %ymm0, %ymm3 ## EVEX TO VEX Compression encoding: [0xc5,0xfd,0xde,0xd9]
[AVX-512] Remove masked pmin/pmax intrinsics and autoupgrade to native IR. Clang patch to replace 512-bit vector and 64-bit element versions with native IR will follow. llvm-svn: 284955
2016-10-24 12:04:16 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
; CHECK-NEXT: vpmaxub %ymm1, %ymm0, %ymm2 {%k1} ## encoding: [0x62,0xf1,0x7d,0x29,0xde,0xd1]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpaddb %ymm3, %ymm2, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xed,0xfc,0xc3]
[AVX-512] Remove masked pmin/pmax intrinsics and autoupgrade to native IR. Clang patch to replace 512-bit vector and 64-bit element versions with native IR will follow. llvm-svn: 284955
2016-10-24 12:04:16 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <32 x i8> @llvm.x86.avx512.mask.pmaxu.b.256(<32 x i8> %x0, <32 x i8> %x1, <32 x i8> %x2, i32 %x3)
%res1 = call <32 x i8> @llvm.x86.avx512.mask.pmaxu.b.256(<32 x i8> %x0, <32 x i8> %x1, <32 x i8> %x2, i32 -1)
%res2 = add <32 x i8> %res, %res1
ret <32 x i8> %res2
}
declare <8 x i16> @llvm.x86.avx512.mask.pmaxu.w.128(<8 x i16>, <8 x i16>, <8 x i16>, i8)
define <8 x i16>@test_int_x86_avx512_mask_pmaxu_w_128(<8 x i16> %x0, <8 x i16> %x1, <8 x i16> %x2, i8 %x3) {
; CHECK-LABEL: test_int_x86_avx512_mask_pmaxu_w_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpmaxuw %xmm1, %xmm0, %xmm3 ## EVEX TO VEX Compression encoding: [0xc4,0xe2,0x79,0x3e,0xd9]
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
[AVX-512] Remove masked pmin/pmax intrinsics and autoupgrade to native IR. Clang patch to replace 512-bit vector and 64-bit element versions with native IR will follow. llvm-svn: 284955
2016-10-24 12:04:16 +08:00
; CHECK-NEXT: vpmaxuw %xmm1, %xmm0, %xmm2 {%k1} ## encoding: [0x62,0xf2,0x7d,0x09,0x3e,0xd1]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpaddw %xmm3, %xmm2, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xe9,0xfd,0xc3]
[AVX-512] Remove masked pmin/pmax intrinsics and autoupgrade to native IR. Clang patch to replace 512-bit vector and 64-bit element versions with native IR will follow. llvm-svn: 284955
2016-10-24 12:04:16 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <8 x i16> @llvm.x86.avx512.mask.pmaxu.w.128(<8 x i16> %x0, <8 x i16> %x1, <8 x i16> %x2, i8 %x3)
%res1 = call <8 x i16> @llvm.x86.avx512.mask.pmaxu.w.128(<8 x i16> %x0, <8 x i16> %x1, <8 x i16> %x2, i8 -1)
%res2 = add <8 x i16> %res, %res1
ret <8 x i16> %res2
}
declare <16 x i16> @llvm.x86.avx512.mask.pmaxu.w.256(<16 x i16>, <16 x i16>, <16 x i16>, i16)
define <16 x i16>@test_int_x86_avx512_mask_pmaxu_w_256(<16 x i16> %x0, <16 x i16> %x1, <16 x i16> %x2, i16 %mask) {
; CHECK-LABEL: test_int_x86_avx512_mask_pmaxu_w_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
[AVX-512] Remove masked pmin/pmax intrinsics and autoupgrade to native IR. Clang patch to replace 512-bit vector and 64-bit element versions with native IR will follow. llvm-svn: 284955
2016-10-24 12:04:16 +08:00
; CHECK-NEXT: vpmaxuw %ymm1, %ymm0, %ymm2 {%k1} ## encoding: [0x62,0xf2,0x7d,0x29,0x3e,0xd1]
; CHECK-NEXT: vpmaxuw %ymm1, %ymm0, %ymm0 {%k1} {z} ## encoding: [0x62,0xf2,0x7d,0xa9,0x3e,0xc1]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpaddw %ymm0, %ymm2, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xed,0xfd,0xc0]
[AVX-512] Remove masked pmin/pmax intrinsics and autoupgrade to native IR. Clang patch to replace 512-bit vector and 64-bit element versions with native IR will follow. llvm-svn: 284955
2016-10-24 12:04:16 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <16 x i16> @llvm.x86.avx512.mask.pmaxu.w.256(<16 x i16> %x0, <16 x i16> %x1, <16 x i16> %x2, i16 %mask)
%res1 = call <16 x i16> @llvm.x86.avx512.mask.pmaxu.w.256(<16 x i16> %x0, <16 x i16> %x1, <16 x i16> zeroinitializer, i16 %mask)
%res2 = add <16 x i16> %res, %res1
ret <16 x i16> %res2
}
declare <16 x i8> @llvm.x86.avx512.mask.pmins.b.128(<16 x i8>, <16 x i8>, <16 x i8>, i16)
define <16 x i8>@test_int_x86_avx512_mask_pmins_b_128(<16 x i8> %x0, <16 x i8> %x1, <16 x i8> %x2, i16 %mask) {
; CHECK-LABEL: test_int_x86_avx512_mask_pmins_b_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
[AVX-512] Remove masked pmin/pmax intrinsics and autoupgrade to native IR. Clang patch to replace 512-bit vector and 64-bit element versions with native IR will follow. llvm-svn: 284955
2016-10-24 12:04:16 +08:00
; CHECK-NEXT: vpminsb %xmm1, %xmm0, %xmm2 {%k1} ## encoding: [0x62,0xf2,0x7d,0x09,0x38,0xd1]
; CHECK-NEXT: vpminsb %xmm1, %xmm0, %xmm0 {%k1} {z} ## encoding: [0x62,0xf2,0x7d,0x89,0x38,0xc1]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpaddb %xmm0, %xmm2, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xe9,0xfc,0xc0]
[AVX-512] Remove masked pmin/pmax intrinsics and autoupgrade to native IR. Clang patch to replace 512-bit vector and 64-bit element versions with native IR will follow. llvm-svn: 284955
2016-10-24 12:04:16 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <16 x i8> @llvm.x86.avx512.mask.pmins.b.128(<16 x i8> %x0, <16 x i8> %x1, <16 x i8> %x2, i16 %mask)
%res1 = call <16 x i8> @llvm.x86.avx512.mask.pmins.b.128(<16 x i8> %x0, <16 x i8> %x1, <16 x i8> zeroinitializer, i16 %mask)
%res2 = add <16 x i8> %res, %res1
ret <16 x i8> %res2
}
declare <32 x i8> @llvm.x86.avx512.mask.pmins.b.256(<32 x i8>, <32 x i8>, <32 x i8>, i32)
define <32 x i8>@test_int_x86_avx512_mask_pmins_b_256(<32 x i8> %x0, <32 x i8> %x1, <32 x i8> %x2, i32 %x3) {
; CHECK-LABEL: test_int_x86_avx512_mask_pmins_b_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpminsb %ymm1, %ymm0, %ymm3 ## EVEX TO VEX Compression encoding: [0xc4,0xe2,0x7d,0x38,0xd9]
[AVX-512] Remove masked pmin/pmax intrinsics and autoupgrade to native IR. Clang patch to replace 512-bit vector and 64-bit element versions with native IR will follow. llvm-svn: 284955
2016-10-24 12:04:16 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
; CHECK-NEXT: vpminsb %ymm1, %ymm0, %ymm2 {%k1} ## encoding: [0x62,0xf2,0x7d,0x29,0x38,0xd1]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpaddb %ymm3, %ymm2, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xed,0xfc,0xc3]
[AVX-512] Remove masked pmin/pmax intrinsics and autoupgrade to native IR. Clang patch to replace 512-bit vector and 64-bit element versions with native IR will follow. llvm-svn: 284955
2016-10-24 12:04:16 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <32 x i8> @llvm.x86.avx512.mask.pmins.b.256(<32 x i8> %x0, <32 x i8> %x1, <32 x i8> %x2, i32 %x3)
%res1 = call <32 x i8> @llvm.x86.avx512.mask.pmins.b.256(<32 x i8> %x0, <32 x i8> %x1, <32 x i8> %x2, i32 -1)
%res2 = add <32 x i8> %res, %res1
ret <32 x i8> %res2
}
declare <8 x i16> @llvm.x86.avx512.mask.pmins.w.128(<8 x i16>, <8 x i16>, <8 x i16>, i8)
define <8 x i16>@test_int_x86_avx512_mask_pmins_w_128(<8 x i16> %x0, <8 x i16> %x1, <8 x i16> %x2, i8 %x3) {
; CHECK-LABEL: test_int_x86_avx512_mask_pmins_w_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpminsw %xmm1, %xmm0, %xmm3 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xea,0xd9]
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
[AVX-512] Remove masked pmin/pmax intrinsics and autoupgrade to native IR. Clang patch to replace 512-bit vector and 64-bit element versions with native IR will follow. llvm-svn: 284955
2016-10-24 12:04:16 +08:00
; CHECK-NEXT: vpminsw %xmm1, %xmm0, %xmm2 {%k1} ## encoding: [0x62,0xf1,0x7d,0x09,0xea,0xd1]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpaddw %xmm3, %xmm2, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xe9,0xfd,0xc3]
[AVX-512] Remove masked pmin/pmax intrinsics and autoupgrade to native IR. Clang patch to replace 512-bit vector and 64-bit element versions with native IR will follow. llvm-svn: 284955
2016-10-24 12:04:16 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <8 x i16> @llvm.x86.avx512.mask.pmins.w.128(<8 x i16> %x0, <8 x i16> %x1, <8 x i16> %x2, i8 %x3)
%res1 = call <8 x i16> @llvm.x86.avx512.mask.pmins.w.128(<8 x i16> %x0, <8 x i16> %x1, <8 x i16> %x2, i8 -1)
%res2 = add <8 x i16> %res, %res1
ret <8 x i16> %res2
}
declare <16 x i16> @llvm.x86.avx512.mask.pmins.w.256(<16 x i16>, <16 x i16>, <16 x i16>, i16)
define <16 x i16>@test_int_x86_avx512_mask_pmins_w_256(<16 x i16> %x0, <16 x i16> %x1, <16 x i16> %x2, i16 %mask) {
; CHECK-LABEL: test_int_x86_avx512_mask_pmins_w_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
[AVX-512] Remove masked pmin/pmax intrinsics and autoupgrade to native IR. Clang patch to replace 512-bit vector and 64-bit element versions with native IR will follow. llvm-svn: 284955
2016-10-24 12:04:16 +08:00
; CHECK-NEXT: vpminsw %ymm1, %ymm0, %ymm2 {%k1} ## encoding: [0x62,0xf1,0x7d,0x29,0xea,0xd1]
; CHECK-NEXT: vpminsw %ymm1, %ymm0, %ymm0 {%k1} {z} ## encoding: [0x62,0xf1,0x7d,0xa9,0xea,0xc1]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpaddw %ymm0, %ymm2, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xed,0xfd,0xc0]
[AVX-512] Remove masked pmin/pmax intrinsics and autoupgrade to native IR. Clang patch to replace 512-bit vector and 64-bit element versions with native IR will follow. llvm-svn: 284955
2016-10-24 12:04:16 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <16 x i16> @llvm.x86.avx512.mask.pmins.w.256(<16 x i16> %x0, <16 x i16> %x1, <16 x i16> %x2, i16 %mask)
%res1 = call <16 x i16> @llvm.x86.avx512.mask.pmins.w.256(<16 x i16> %x0, <16 x i16> %x1, <16 x i16> zeroinitializer, i16 %mask)
%res2 = add <16 x i16> %res, %res1
ret <16 x i16> %res2
}
declare <16 x i8> @llvm.x86.avx512.mask.pminu.b.128(<16 x i8>, <16 x i8>, <16 x i8>, i16)
define <16 x i8>@test_int_x86_avx512_mask_pminu_b_128(<16 x i8> %x0, <16 x i8> %x1, <16 x i8> %x2, i16 %mask) {
; CHECK-LABEL: test_int_x86_avx512_mask_pminu_b_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
[AVX-512] Remove masked pmin/pmax intrinsics and autoupgrade to native IR. Clang patch to replace 512-bit vector and 64-bit element versions with native IR will follow. llvm-svn: 284955
2016-10-24 12:04:16 +08:00
; CHECK-NEXT: vpminub %xmm1, %xmm0, %xmm2 {%k1} ## encoding: [0x62,0xf1,0x7d,0x09,0xda,0xd1]
; CHECK-NEXT: vpminub %xmm1, %xmm0, %xmm0 {%k1} {z} ## encoding: [0x62,0xf1,0x7d,0x89,0xda,0xc1]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpaddb %xmm0, %xmm2, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xe9,0xfc,0xc0]
[AVX-512] Remove masked pmin/pmax intrinsics and autoupgrade to native IR. Clang patch to replace 512-bit vector and 64-bit element versions with native IR will follow. llvm-svn: 284955
2016-10-24 12:04:16 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <16 x i8> @llvm.x86.avx512.mask.pminu.b.128(<16 x i8> %x0, <16 x i8> %x1, <16 x i8> %x2, i16 %mask)
%res1 = call <16 x i8> @llvm.x86.avx512.mask.pminu.b.128(<16 x i8> %x0, <16 x i8> %x1, <16 x i8> zeroinitializer, i16 %mask)
%res2 = add <16 x i8> %res, %res1
ret <16 x i8> %res2
}
declare <32 x i8> @llvm.x86.avx512.mask.pminu.b.256(<32 x i8>, <32 x i8>, <32 x i8>, i32)
define <32 x i8>@test_int_x86_avx512_mask_pminu_b_256(<32 x i8> %x0, <32 x i8> %x1, <32 x i8> %x2, i32 %x3) {
; CHECK-LABEL: test_int_x86_avx512_mask_pminu_b_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpminub %ymm1, %ymm0, %ymm3 ## EVEX TO VEX Compression encoding: [0xc5,0xfd,0xda,0xd9]
[AVX-512] Remove masked pmin/pmax intrinsics and autoupgrade to native IR. Clang patch to replace 512-bit vector and 64-bit element versions with native IR will follow. llvm-svn: 284955
2016-10-24 12:04:16 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
; CHECK-NEXT: vpminub %ymm1, %ymm0, %ymm2 {%k1} ## encoding: [0x62,0xf1,0x7d,0x29,0xda,0xd1]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpaddb %ymm3, %ymm2, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xed,0xfc,0xc3]
[AVX-512] Remove masked pmin/pmax intrinsics and autoupgrade to native IR. Clang patch to replace 512-bit vector and 64-bit element versions with native IR will follow. llvm-svn: 284955
2016-10-24 12:04:16 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <32 x i8> @llvm.x86.avx512.mask.pminu.b.256(<32 x i8> %x0, <32 x i8> %x1, <32 x i8> %x2, i32 %x3)
%res1 = call <32 x i8> @llvm.x86.avx512.mask.pminu.b.256(<32 x i8> %x0, <32 x i8> %x1, <32 x i8> %x2, i32 -1)
%res2 = add <32 x i8> %res, %res1
ret <32 x i8> %res2
}
declare <8 x i16> @llvm.x86.avx512.mask.pminu.w.128(<8 x i16>, <8 x i16>, <8 x i16>, i8)
define <8 x i16>@test_int_x86_avx512_mask_pminu_w_128(<8 x i16> %x0, <8 x i16> %x1, <8 x i16> %x2, i8 %x3) {
; CHECK-LABEL: test_int_x86_avx512_mask_pminu_w_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpminuw %xmm1, %xmm0, %xmm3 ## EVEX TO VEX Compression encoding: [0xc4,0xe2,0x79,0x3a,0xd9]
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
[AVX-512] Remove masked pmin/pmax intrinsics and autoupgrade to native IR. Clang patch to replace 512-bit vector and 64-bit element versions with native IR will follow. llvm-svn: 284955
2016-10-24 12:04:16 +08:00
; CHECK-NEXT: vpminuw %xmm1, %xmm0, %xmm2 {%k1} ## encoding: [0x62,0xf2,0x7d,0x09,0x3a,0xd1]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpaddw %xmm3, %xmm2, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xe9,0xfd,0xc3]
[AVX-512] Remove masked pmin/pmax intrinsics and autoupgrade to native IR. Clang patch to replace 512-bit vector and 64-bit element versions with native IR will follow. llvm-svn: 284955
2016-10-24 12:04:16 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <8 x i16> @llvm.x86.avx512.mask.pminu.w.128(<8 x i16> %x0, <8 x i16> %x1, <8 x i16> %x2, i8 %x3)
%res1 = call <8 x i16> @llvm.x86.avx512.mask.pminu.w.128(<8 x i16> %x0, <8 x i16> %x1, <8 x i16> %x2, i8 -1)
%res2 = add <8 x i16> %res, %res1
ret <8 x i16> %res2
}
declare <16 x i16> @llvm.x86.avx512.mask.pminu.w.256(<16 x i16>, <16 x i16>, <16 x i16>, i16)
define <16 x i16>@test_int_x86_avx512_mask_pminu_w_256(<16 x i16> %x0, <16 x i16> %x1, <16 x i16> %x2, i16 %mask) {
; CHECK-LABEL: test_int_x86_avx512_mask_pminu_w_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
[AVX-512] Remove masked pmin/pmax intrinsics and autoupgrade to native IR. Clang patch to replace 512-bit vector and 64-bit element versions with native IR will follow. llvm-svn: 284955
2016-10-24 12:04:16 +08:00
; CHECK-NEXT: vpminuw %ymm1, %ymm0, %ymm2 {%k1} ## encoding: [0x62,0xf2,0x7d,0x29,0x3a,0xd1]
; CHECK-NEXT: vpminuw %ymm1, %ymm0, %ymm0 {%k1} {z} ## encoding: [0x62,0xf2,0x7d,0xa9,0x3a,0xc1]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpaddw %ymm0, %ymm2, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xed,0xfd,0xc0]
[AVX-512] Remove masked pmin/pmax intrinsics and autoupgrade to native IR. Clang patch to replace 512-bit vector and 64-bit element versions with native IR will follow. llvm-svn: 284955
2016-10-24 12:04:16 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <16 x i16> @llvm.x86.avx512.mask.pminu.w.256(<16 x i16> %x0, <16 x i16> %x1, <16 x i16> %x2, i16 %mask)
%res1 = call <16 x i16> @llvm.x86.avx512.mask.pminu.w.256(<16 x i16> %x0, <16 x i16> %x1, <16 x i16> zeroinitializer, i16 %mask)
%res2 = add <16 x i16> %res, %res1
ret <16 x i16> %res2
}
[AVX-512] Remove intrinsics for 128/256-bit masked shift by single element in xmm. Instead upgrade them to a select and the older SSE/AVX2 intrinsic. llvm-svn: 286070
2016-11-07 00:29:08 +08:00
declare <8 x i16> @llvm.x86.avx512.mask.psrl.w.128(<8 x i16>, <8 x i16>, <8 x i16>, i8)
define <8 x i16>@test_int_x86_avx512_mask_psrl_w_128(<8 x i16> %x0, <8 x i16> %x1, <8 x i16> %x2, i8 %x3) {
; CHECK-LABEL: test_int_x86_avx512_mask_psrl_w_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpsrlw %xmm1, %xmm0, %xmm3 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xd1,0xd9]
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
[AVX-512] Remove intrinsics for 128/256-bit masked shift by single element in xmm. Instead upgrade them to a select and the older SSE/AVX2 intrinsic. llvm-svn: 286070
2016-11-07 00:29:08 +08:00
; CHECK-NEXT: vpsrlw %xmm1, %xmm0, %xmm2 {%k1} ## encoding: [0x62,0xf1,0x7d,0x09,0xd1,0xd1]
; CHECK-NEXT: vpsrlw %xmm1, %xmm0, %xmm0 {%k1} {z} ## encoding: [0x62,0xf1,0x7d,0x89,0xd1,0xc1]
[X86][SKX][KNL] Updated regression tests to use -mattr instead of -mcpu flag.NFC. NFC. Updated 8 regression tests to use -mattr instead of -mcpu flag as follows: -mcpu=knl --> -mattr=+avx512f -mcpu=skx --> -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq The updates are as part of the preparation of a large commit to add all instruction scheduling for the SKX target. Reviewers: delena, zvi, RKSimon Differential Revision: https://reviews.llvm.org/D38222 Change-Id: I2381c9b5bb75ecacfca017243c22d054f6eddd14 llvm-svn: 314306
2017-09-27 22:44:15 +08:00
; CHECK-NEXT: vpaddw %xmm0, %xmm3, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xe1,0xfd,0xc0]
; CHECK-NEXT: vpaddw %xmm0, %xmm2, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xe9,0xfd,0xc0]
[AVX-512] Remove intrinsics for 128/256-bit masked shift by single element in xmm. Instead upgrade them to a select and the older SSE/AVX2 intrinsic. llvm-svn: 286070
2016-11-07 00:29:08 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <8 x i16> @llvm.x86.avx512.mask.psrl.w.128(<8 x i16> %x0, <8 x i16> %x1, <8 x i16> %x2, i8 %x3)
%res1 = call <8 x i16> @llvm.x86.avx512.mask.psrl.w.128(<8 x i16> %x0, <8 x i16> %x1, <8 x i16> %x2, i8 -1)
%res2 = call <8 x i16> @llvm.x86.avx512.mask.psrl.w.128(<8 x i16> %x0, <8 x i16> %x1, <8 x i16> zeroinitializer, i8 %x3)
%res3 = add <8 x i16> %res, %res1
%res4 = add <8 x i16> %res2, %res3
ret <8 x i16> %res4
}
declare <16 x i16> @llvm.x86.avx512.mask.psrl.w.256(<16 x i16>, <8 x i16>, <16 x i16>, i16)
define <16 x i16>@test_int_x86_avx512_mask_psrl_w_256(<16 x i16> %x0, <8 x i16> %x1, <16 x i16> %x2, i16 %x3) {
; CHECK-LABEL: test_int_x86_avx512_mask_psrl_w_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpsrlw %xmm1, %ymm0, %ymm3 ## EVEX TO VEX Compression encoding: [0xc5,0xfd,0xd1,0xd9]
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
[AVX-512] Remove intrinsics for 128/256-bit masked shift by single element in xmm. Instead upgrade them to a select and the older SSE/AVX2 intrinsic. llvm-svn: 286070
2016-11-07 00:29:08 +08:00
; CHECK-NEXT: vpsrlw %xmm1, %ymm0, %ymm2 {%k1} ## encoding: [0x62,0xf1,0x7d,0x29,0xd1,0xd1]
; CHECK-NEXT: vpsrlw %xmm1, %ymm0, %ymm0 {%k1} {z} ## encoding: [0x62,0xf1,0x7d,0xa9,0xd1,0xc1]
[X86][SKX][KNL] Updated regression tests to use -mattr instead of -mcpu flag.NFC. NFC. Updated 8 regression tests to use -mattr instead of -mcpu flag as follows: -mcpu=knl --> -mattr=+avx512f -mcpu=skx --> -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq The updates are as part of the preparation of a large commit to add all instruction scheduling for the SKX target. Reviewers: delena, zvi, RKSimon Differential Revision: https://reviews.llvm.org/D38222 Change-Id: I2381c9b5bb75ecacfca017243c22d054f6eddd14 llvm-svn: 314306
2017-09-27 22:44:15 +08:00
; CHECK-NEXT: vpaddw %ymm0, %ymm3, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xe5,0xfd,0xc0]
Add LiveRangeShrink pass to shrink live range within BB. Summary: LiveRangeShrink pass moves instruction right after the definition with the same BB if the instruction and its operands all have more than one use. This pass is inexpensive and guarantees optimal live-range within BB. Reviewers: davidxl, wmi, hfinkel, MatzeB, andreadb Reviewed By: MatzeB, andreadb Subscribers: hiraditya, jyknight, sanjoy, skatkov, gberry, jholewinski, qcolombet, javed.absar, krytarowski, atrick, spatel, RKSimon, andreadb, MatzeB, mehdi_amini, mgorny, efriedma, davide, dberlin, llvm-commits Differential Revision: https://reviews.llvm.org/D32563 llvm-svn: 304371
2017-06-01 07:25:25 +08:00
; CHECK-NEXT: vpaddw %ymm0, %ymm2, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xed,0xfd,0xc0]
[AVX-512] Remove intrinsics for 128/256-bit masked shift by single element in xmm. Instead upgrade them to a select and the older SSE/AVX2 intrinsic. llvm-svn: 286070
2016-11-07 00:29:08 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <16 x i16> @llvm.x86.avx512.mask.psrl.w.256(<16 x i16> %x0, <8 x i16> %x1, <16 x i16> %x2, i16 %x3)
%res1 = call <16 x i16> @llvm.x86.avx512.mask.psrl.w.256(<16 x i16> %x0, <8 x i16> %x1, <16 x i16> %x2, i16 -1)
%res2 = call <16 x i16> @llvm.x86.avx512.mask.psrl.w.256(<16 x i16> %x0, <8 x i16> %x1, <16 x i16> zeroinitializer, i16 %x3)
%res3 = add <16 x i16> %res, %res1
%res4 = add <16 x i16> %res3, %res2
ret <16 x i16> %res4
}
declare <8 x i16> @llvm.x86.avx512.mask.psra.w.128(<8 x i16>, <8 x i16>, <8 x i16>, i8)
define <8 x i16>@test_int_x86_avx512_mask_psra_w_128(<8 x i16> %x0, <8 x i16> %x1, <8 x i16> %x2, i8 %x3) {
; CHECK-LABEL: test_int_x86_avx512_mask_psra_w_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpsraw %xmm1, %xmm0, %xmm3 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xe1,0xd9]
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
[AVX-512] Remove intrinsics for 128/256-bit masked shift by single element in xmm. Instead upgrade them to a select and the older SSE/AVX2 intrinsic. llvm-svn: 286070
2016-11-07 00:29:08 +08:00
; CHECK-NEXT: vpsraw %xmm1, %xmm0, %xmm2 {%k1} ## encoding: [0x62,0xf1,0x7d,0x09,0xe1,0xd1]
; CHECK-NEXT: vpsraw %xmm1, %xmm0, %xmm0 {%k1} {z} ## encoding: [0x62,0xf1,0x7d,0x89,0xe1,0xc1]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpaddw %xmm3, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xfd,0xc3]
[X86][SKX][KNL] Updated regression tests to use -mattr instead of -mcpu flag.NFC. NFC. Updated 8 regression tests to use -mattr instead of -mcpu flag as follows: -mcpu=knl --> -mattr=+avx512f -mcpu=skx --> -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq The updates are as part of the preparation of a large commit to add all instruction scheduling for the SKX target. Reviewers: delena, zvi, RKSimon Differential Revision: https://reviews.llvm.org/D38222 Change-Id: I2381c9b5bb75ecacfca017243c22d054f6eddd14 llvm-svn: 314306
2017-09-27 22:44:15 +08:00
; CHECK-NEXT: vpaddw %xmm0, %xmm2, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xe9,0xfd,0xc0]
[AVX-512] Remove intrinsics for 128/256-bit masked shift by single element in xmm. Instead upgrade them to a select and the older SSE/AVX2 intrinsic. llvm-svn: 286070
2016-11-07 00:29:08 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <8 x i16> @llvm.x86.avx512.mask.psra.w.128(<8 x i16> %x0, <8 x i16> %x1, <8 x i16> %x2, i8 %x3)
%res1 = call <8 x i16> @llvm.x86.avx512.mask.psra.w.128(<8 x i16> %x0, <8 x i16> %x1, <8 x i16> zeroinitializer, i8 %x3)
%res2 = call <8 x i16> @llvm.x86.avx512.mask.psra.w.128(<8 x i16> %x0, <8 x i16> %x1, <8 x i16> %x2, i8 -1)
%res3 = add <8 x i16> %res, %res1
%res4 = add <8 x i16> %res3, %res2
ret <8 x i16> %res4
}
declare <16 x i16> @llvm.x86.avx512.mask.psra.w.256(<16 x i16>, <8 x i16>, <16 x i16>, i16)
define <16 x i16>@test_int_x86_avx512_mask_psra_w_256(<16 x i16> %x0, <8 x i16> %x1, <16 x i16> %x2, i16 %x3) {
; CHECK-LABEL: test_int_x86_avx512_mask_psra_w_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpsraw %xmm1, %ymm0, %ymm3 ## EVEX TO VEX Compression encoding: [0xc5,0xfd,0xe1,0xd9]
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
[AVX-512] Remove intrinsics for 128/256-bit masked shift by single element in xmm. Instead upgrade them to a select and the older SSE/AVX2 intrinsic. llvm-svn: 286070
2016-11-07 00:29:08 +08:00
; CHECK-NEXT: vpsraw %xmm1, %ymm0, %ymm2 {%k1} ## encoding: [0x62,0xf1,0x7d,0x29,0xe1,0xd1]
; CHECK-NEXT: vpsraw %xmm1, %ymm0, %ymm0 {%k1} {z} ## encoding: [0x62,0xf1,0x7d,0xa9,0xe1,0xc1]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpaddw %ymm3, %ymm0, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xfd,0xfd,0xc3]
[X86][SKX][KNL] Updated regression tests to use -mattr instead of -mcpu flag.NFC. NFC. Updated 8 regression tests to use -mattr instead of -mcpu flag as follows: -mcpu=knl --> -mattr=+avx512f -mcpu=skx --> -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq The updates are as part of the preparation of a large commit to add all instruction scheduling for the SKX target. Reviewers: delena, zvi, RKSimon Differential Revision: https://reviews.llvm.org/D38222 Change-Id: I2381c9b5bb75ecacfca017243c22d054f6eddd14 llvm-svn: 314306
2017-09-27 22:44:15 +08:00
; CHECK-NEXT: vpaddw %ymm0, %ymm2, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xed,0xfd,0xc0]
[AVX-512] Remove intrinsics for 128/256-bit masked shift by single element in xmm. Instead upgrade them to a select and the older SSE/AVX2 intrinsic. llvm-svn: 286070
2016-11-07 00:29:08 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <16 x i16> @llvm.x86.avx512.mask.psra.w.256(<16 x i16> %x0, <8 x i16> %x1, <16 x i16> %x2, i16 %x3)
%res1 = call <16 x i16> @llvm.x86.avx512.mask.psra.w.256(<16 x i16> %x0, <8 x i16> %x1, <16 x i16> zeroinitializer, i16 %x3)
%res2 = call <16 x i16> @llvm.x86.avx512.mask.psra.w.256(<16 x i16> %x0, <8 x i16> %x1, <16 x i16> %x2, i16 -1)
%res3 = add <16 x i16> %res, %res1
%res4 = add <16 x i16> %res3, %res2
ret <16 x i16> %res4
}
declare <8 x i16> @llvm.x86.avx512.mask.psll.w.128(<8 x i16>, <8 x i16>, <8 x i16>, i8)
define <8 x i16>@test_int_x86_avx512_mask_psll_w_128(<8 x i16> %x0, <8 x i16> %x1, <8 x i16> %x2, i8 %x3) {
; CHECK-LABEL: test_int_x86_avx512_mask_psll_w_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpsllw %xmm1, %xmm0, %xmm3 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xf1,0xd9]
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
[AVX-512] Remove intrinsics for 128/256-bit masked shift by single element in xmm. Instead upgrade them to a select and the older SSE/AVX2 intrinsic. llvm-svn: 286070
2016-11-07 00:29:08 +08:00
; CHECK-NEXT: vpsllw %xmm1, %xmm0, %xmm2 {%k1} ## encoding: [0x62,0xf1,0x7d,0x09,0xf1,0xd1]
; CHECK-NEXT: vpsllw %xmm1, %xmm0, %xmm0 {%k1} {z} ## encoding: [0x62,0xf1,0x7d,0x89,0xf1,0xc1]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpaddw %xmm3, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xfd,0xc3]
[X86][SKX][KNL] Updated regression tests to use -mattr instead of -mcpu flag.NFC. NFC. Updated 8 regression tests to use -mattr instead of -mcpu flag as follows: -mcpu=knl --> -mattr=+avx512f -mcpu=skx --> -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq The updates are as part of the preparation of a large commit to add all instruction scheduling for the SKX target. Reviewers: delena, zvi, RKSimon Differential Revision: https://reviews.llvm.org/D38222 Change-Id: I2381c9b5bb75ecacfca017243c22d054f6eddd14 llvm-svn: 314306
2017-09-27 22:44:15 +08:00
; CHECK-NEXT: vpaddw %xmm0, %xmm2, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xe9,0xfd,0xc0]
[AVX-512] Remove intrinsics for 128/256-bit masked shift by single element in xmm. Instead upgrade them to a select and the older SSE/AVX2 intrinsic. llvm-svn: 286070
2016-11-07 00:29:08 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <8 x i16> @llvm.x86.avx512.mask.psll.w.128(<8 x i16> %x0, <8 x i16> %x1, <8 x i16> %x2, i8 %x3)
%res1 = call <8 x i16> @llvm.x86.avx512.mask.psll.w.128(<8 x i16> %x0, <8 x i16> %x1, <8 x i16> zeroinitializer, i8 %x3)
%res2 = call <8 x i16> @llvm.x86.avx512.mask.psll.w.128(<8 x i16> %x0, <8 x i16> %x1, <8 x i16> %x2, i8 -1)
%res3 = add <8 x i16> %res, %res1
%res4 = add <8 x i16> %res3, %res2
ret <8 x i16> %res4
}
declare <16 x i16> @llvm.x86.avx512.mask.psll.w.256(<16 x i16>, <8 x i16>, <16 x i16>, i16)
define <16 x i16>@test_int_x86_avx512_mask_psll_w_256(<16 x i16> %x0, <8 x i16> %x1, <16 x i16> %x2, i16 %x3) {
; CHECK-LABEL: test_int_x86_avx512_mask_psll_w_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpsllw %xmm1, %ymm0, %ymm3 ## EVEX TO VEX Compression encoding: [0xc5,0xfd,0xf1,0xd9]
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
[AVX-512] Remove intrinsics for 128/256-bit masked shift by single element in xmm. Instead upgrade them to a select and the older SSE/AVX2 intrinsic. llvm-svn: 286070
2016-11-07 00:29:08 +08:00
; CHECK-NEXT: vpsllw %xmm1, %ymm0, %ymm2 {%k1} ## encoding: [0x62,0xf1,0x7d,0x29,0xf1,0xd1]
; CHECK-NEXT: vpsllw %xmm1, %ymm0, %ymm0 {%k1} {z} ## encoding: [0x62,0xf1,0x7d,0xa9,0xf1,0xc1]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpaddw %ymm3, %ymm0, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xfd,0xfd,0xc3]
[X86][SKX][KNL] Updated regression tests to use -mattr instead of -mcpu flag.NFC. NFC. Updated 8 regression tests to use -mattr instead of -mcpu flag as follows: -mcpu=knl --> -mattr=+avx512f -mcpu=skx --> -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq The updates are as part of the preparation of a large commit to add all instruction scheduling for the SKX target. Reviewers: delena, zvi, RKSimon Differential Revision: https://reviews.llvm.org/D38222 Change-Id: I2381c9b5bb75ecacfca017243c22d054f6eddd14 llvm-svn: 314306
2017-09-27 22:44:15 +08:00
; CHECK-NEXT: vpaddw %ymm0, %ymm2, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xed,0xfd,0xc0]
[AVX-512] Remove intrinsics for 128/256-bit masked shift by single element in xmm. Instead upgrade them to a select and the older SSE/AVX2 intrinsic. llvm-svn: 286070
2016-11-07 00:29:08 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <16 x i16> @llvm.x86.avx512.mask.psll.w.256(<16 x i16> %x0, <8 x i16> %x1, <16 x i16> %x2, i16 %x3)
%res1 = call <16 x i16> @llvm.x86.avx512.mask.psll.w.256(<16 x i16> %x0, <8 x i16> %x1, <16 x i16> zeroinitializer, i16 %x3)
%res2 = call <16 x i16> @llvm.x86.avx512.mask.psll.w.256(<16 x i16> %x0, <8 x i16> %x1, <16 x i16> %x2, i16 -1)
%res3 = add <16 x i16> %res, %res1
%res4 = add <16 x i16> %res3, %res2
ret <16 x i16> %res4
}
[AVX-512] Remove intrinsics for 128/256-bit masked shift by immediate. Instead upgrade them to a select and the older SSE/AVX2 intrinsic. llvm-svn: 286072
2016-11-07 00:29:14 +08:00
declare <8 x i16> @llvm.x86.avx512.mask.psrl.wi.128(<8 x i16>, i32, <8 x i16>, i8)
define <8 x i16>@test_int_x86_avx512_mask_psrl_wi_128(<8 x i16> %x0, i32 %x1, <8 x i16> %x2, i8 %x3) {
; CHECK-LABEL: test_int_x86_avx512_mask_psrl_wi_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpsrlw $3, %xmm0, %xmm2 ## EVEX TO VEX Compression encoding: [0xc5,0xe9,0x71,0xd0,0x03]
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %esi, %k1 ## encoding: [0xc5,0xfb,0x92,0xce]
[AVX-512] Remove intrinsics for 128/256-bit masked shift by immediate. Instead upgrade them to a select and the older SSE/AVX2 intrinsic. llvm-svn: 286072
2016-11-07 00:29:14 +08:00
; CHECK-NEXT: vpsrlw $3, %xmm0, %xmm1 {%k1} ## encoding: [0x62,0xf1,0x75,0x09,0x71,0xd0,0x03]
Add LiveRangeShrink pass to shrink live range within BB. Summary: LiveRangeShrink pass moves instruction right after the definition with the same BB if the instruction and its operands all have more than one use. This pass is inexpensive and guarantees optimal live-range within BB. Reviewers: davidxl, wmi, hfinkel, MatzeB, andreadb Reviewed By: MatzeB, andreadb Subscribers: hiraditya, jyknight, sanjoy, skatkov, gberry, jholewinski, qcolombet, javed.absar, krytarowski, atrick, spatel, RKSimon, andreadb, MatzeB, mehdi_amini, mgorny, efriedma, davide, dberlin, llvm-commits Differential Revision: https://reviews.llvm.org/D32563 llvm-svn: 304371
2017-06-01 07:25:25 +08:00
; CHECK-NEXT: vpsrlw $3, %xmm0, %xmm0 {%k1} {z} ## encoding: [0x62,0xf1,0x7d,0x89,0x71,0xd0,0x03]
[X86][SKX][KNL] Updated regression tests to use -mattr instead of -mcpu flag.NFC. NFC. Updated 8 regression tests to use -mattr instead of -mcpu flag as follows: -mcpu=knl --> -mattr=+avx512f -mcpu=skx --> -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq The updates are as part of the preparation of a large commit to add all instruction scheduling for the SKX target. Reviewers: delena, zvi, RKSimon Differential Revision: https://reviews.llvm.org/D38222 Change-Id: I2381c9b5bb75ecacfca017243c22d054f6eddd14 llvm-svn: 314306
2017-09-27 22:44:15 +08:00
; CHECK-NEXT: vpaddw %xmm0, %xmm2, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xe9,0xfd,0xc0]
; CHECK-NEXT: vpaddw %xmm0, %xmm1, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf1,0xfd,0xc0]
[AVX-512] Remove intrinsics for 128/256-bit masked shift by immediate. Instead upgrade them to a select and the older SSE/AVX2 intrinsic. llvm-svn: 286072
2016-11-07 00:29:14 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <8 x i16> @llvm.x86.avx512.mask.psrl.wi.128(<8 x i16> %x0, i32 3, <8 x i16> %x2, i8 %x3)
%res1 = call <8 x i16> @llvm.x86.avx512.mask.psrl.wi.128(<8 x i16> %x0, i32 3, <8 x i16> %x2, i8 -1)
%res2 = call <8 x i16> @llvm.x86.avx512.mask.psrl.wi.128(<8 x i16> %x0, i32 3, <8 x i16> zeroinitializer, i8 %x3)
%res3 = add <8 x i16> %res, %res1
%res4 = add <8 x i16> %res2, %res3
ret <8 x i16> %res4
}
declare <16 x i16> @llvm.x86.avx512.mask.psrl.wi.256(<16 x i16>, i32, <16 x i16>, i16)
define <16 x i16>@test_int_x86_avx512_mask_psrl_wi_256(<16 x i16> %x0, i32 %x1, <16 x i16> %x2, i16 %x3) {
; CHECK-LABEL: test_int_x86_avx512_mask_psrl_wi_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpsrlw $3, %ymm0, %ymm2 ## EVEX TO VEX Compression encoding: [0xc5,0xed,0x71,0xd0,0x03]
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %esi, %k1 ## encoding: [0xc5,0xfb,0x92,0xce]
[AVX-512] Remove intrinsics for 128/256-bit masked shift by immediate. Instead upgrade them to a select and the older SSE/AVX2 intrinsic. llvm-svn: 286072
2016-11-07 00:29:14 +08:00
; CHECK-NEXT: vpsrlw $3, %ymm0, %ymm1 {%k1} ## encoding: [0x62,0xf1,0x75,0x29,0x71,0xd0,0x03]
Add LiveRangeShrink pass to shrink live range within BB. Summary: LiveRangeShrink pass moves instruction right after the definition with the same BB if the instruction and its operands all have more than one use. This pass is inexpensive and guarantees optimal live-range within BB. Reviewers: davidxl, wmi, hfinkel, MatzeB, andreadb Reviewed By: MatzeB, andreadb Subscribers: hiraditya, jyknight, sanjoy, skatkov, gberry, jholewinski, qcolombet, javed.absar, krytarowski, atrick, spatel, RKSimon, andreadb, MatzeB, mehdi_amini, mgorny, efriedma, davide, dberlin, llvm-commits Differential Revision: https://reviews.llvm.org/D32563 llvm-svn: 304371
2017-06-01 07:25:25 +08:00
; CHECK-NEXT: vpsrlw $3, %ymm0, %ymm0 {%k1} {z} ## encoding: [0x62,0xf1,0x7d,0xa9,0x71,0xd0,0x03]
[X86][SKX][KNL] Updated regression tests to use -mattr instead of -mcpu flag.NFC. NFC. Updated 8 regression tests to use -mattr instead of -mcpu flag as follows: -mcpu=knl --> -mattr=+avx512f -mcpu=skx --> -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq The updates are as part of the preparation of a large commit to add all instruction scheduling for the SKX target. Reviewers: delena, zvi, RKSimon Differential Revision: https://reviews.llvm.org/D38222 Change-Id: I2381c9b5bb75ecacfca017243c22d054f6eddd14 llvm-svn: 314306
2017-09-27 22:44:15 +08:00
; CHECK-NEXT: vpaddw %ymm0, %ymm2, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xed,0xfd,0xc0]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpaddw %ymm0, %ymm1, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf5,0xfd,0xc0]
[AVX-512] Remove intrinsics for 128/256-bit masked shift by immediate. Instead upgrade them to a select and the older SSE/AVX2 intrinsic. llvm-svn: 286072
2016-11-07 00:29:14 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <16 x i16> @llvm.x86.avx512.mask.psrl.wi.256(<16 x i16> %x0, i32 3, <16 x i16> %x2, i16 %x3)
%res1 = call <16 x i16> @llvm.x86.avx512.mask.psrl.wi.256(<16 x i16> %x0, i32 3, <16 x i16> %x2, i16 -1)
%res2 = call <16 x i16> @llvm.x86.avx512.mask.psrl.wi.256(<16 x i16> %x0, i32 3, <16 x i16> zeroinitializer, i16 %x3)
%res3 = add <16 x i16> %res, %res1
%res4 = add <16 x i16> %res3, %res2
ret <16 x i16> %res4
}
declare <8 x i16> @llvm.x86.avx512.mask.psra.wi.128(<8 x i16>, i32, <8 x i16>, i8)
define <8 x i16>@test_int_x86_avx512_mask_psra_wi_128(<8 x i16> %x0, i32 %x1, <8 x i16> %x2, i8 %x3) {
; CHECK-LABEL: test_int_x86_avx512_mask_psra_wi_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpsraw $3, %xmm0, %xmm2 ## EVEX TO VEX Compression encoding: [0xc5,0xe9,0x71,0xe0,0x03]
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %esi, %k1 ## encoding: [0xc5,0xfb,0x92,0xce]
[AVX-512] Remove intrinsics for 128/256-bit masked shift by immediate. Instead upgrade them to a select and the older SSE/AVX2 intrinsic. llvm-svn: 286072
2016-11-07 00:29:14 +08:00
; CHECK-NEXT: vpsraw $3, %xmm0, %xmm1 {%k1} ## encoding: [0x62,0xf1,0x75,0x09,0x71,0xe0,0x03]
; CHECK-NEXT: vpsraw $3, %xmm0, %xmm0 {%k1} {z} ## encoding: [0x62,0xf1,0x7d,0x89,0x71,0xe0,0x03]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpaddw %xmm2, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xfd,0xc2]
[X86][SKX][KNL] Updated regression tests to use -mattr instead of -mcpu flag.NFC. NFC. Updated 8 regression tests to use -mattr instead of -mcpu flag as follows: -mcpu=knl --> -mattr=+avx512f -mcpu=skx --> -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq The updates are as part of the preparation of a large commit to add all instruction scheduling for the SKX target. Reviewers: delena, zvi, RKSimon Differential Revision: https://reviews.llvm.org/D38222 Change-Id: I2381c9b5bb75ecacfca017243c22d054f6eddd14 llvm-svn: 314306
2017-09-27 22:44:15 +08:00
; CHECK-NEXT: vpaddw %xmm0, %xmm1, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf1,0xfd,0xc0]
[AVX-512] Remove intrinsics for 128/256-bit masked shift by immediate. Instead upgrade them to a select and the older SSE/AVX2 intrinsic. llvm-svn: 286072
2016-11-07 00:29:14 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <8 x i16> @llvm.x86.avx512.mask.psra.wi.128(<8 x i16> %x0, i32 3, <8 x i16> %x2, i8 %x3)
%res1 = call <8 x i16> @llvm.x86.avx512.mask.psra.wi.128(<8 x i16> %x0, i32 3, <8 x i16> zeroinitializer, i8 %x3)
%res2 = call <8 x i16> @llvm.x86.avx512.mask.psra.wi.128(<8 x i16> %x0, i32 3, <8 x i16> %x2, i8 -1)
%res3 = add <8 x i16> %res, %res1
%res4 = add <8 x i16> %res3, %res2
ret <8 x i16> %res4
}
declare <16 x i16> @llvm.x86.avx512.mask.psra.wi.256(<16 x i16>, i32, <16 x i16>, i16)
define <16 x i16>@test_int_x86_avx512_mask_psra_wi_256(<16 x i16> %x0, i32 %x1, <16 x i16> %x2, i16 %x3) {
; CHECK-LABEL: test_int_x86_avx512_mask_psra_wi_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpsraw $3, %ymm0, %ymm2 ## EVEX TO VEX Compression encoding: [0xc5,0xed,0x71,0xe0,0x03]
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %esi, %k1 ## encoding: [0xc5,0xfb,0x92,0xce]
[AVX-512] Remove intrinsics for 128/256-bit masked shift by immediate. Instead upgrade them to a select and the older SSE/AVX2 intrinsic. llvm-svn: 286072
2016-11-07 00:29:14 +08:00
; CHECK-NEXT: vpsraw $3, %ymm0, %ymm1 {%k1} ## encoding: [0x62,0xf1,0x75,0x29,0x71,0xe0,0x03]
; CHECK-NEXT: vpsraw $3, %ymm0, %ymm0 {%k1} {z} ## encoding: [0x62,0xf1,0x7d,0xa9,0x71,0xe0,0x03]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpaddw %ymm2, %ymm0, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xfd,0xfd,0xc2]
[X86][SKX][KNL] Updated regression tests to use -mattr instead of -mcpu flag.NFC. NFC. Updated 8 regression tests to use -mattr instead of -mcpu flag as follows: -mcpu=knl --> -mattr=+avx512f -mcpu=skx --> -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq The updates are as part of the preparation of a large commit to add all instruction scheduling for the SKX target. Reviewers: delena, zvi, RKSimon Differential Revision: https://reviews.llvm.org/D38222 Change-Id: I2381c9b5bb75ecacfca017243c22d054f6eddd14 llvm-svn: 314306
2017-09-27 22:44:15 +08:00
; CHECK-NEXT: vpaddw %ymm0, %ymm1, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf5,0xfd,0xc0]
[AVX-512] Remove intrinsics for 128/256-bit masked shift by immediate. Instead upgrade them to a select and the older SSE/AVX2 intrinsic. llvm-svn: 286072
2016-11-07 00:29:14 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <16 x i16> @llvm.x86.avx512.mask.psra.wi.256(<16 x i16> %x0, i32 3, <16 x i16> %x2, i16 %x3)
%res1 = call <16 x i16> @llvm.x86.avx512.mask.psra.wi.256(<16 x i16> %x0, i32 3, <16 x i16> zeroinitializer, i16 %x3)
%res2 = call <16 x i16> @llvm.x86.avx512.mask.psra.wi.256(<16 x i16> %x0, i32 3, <16 x i16> %x2, i16 -1)
%res3 = add <16 x i16> %res, %res1
%res4 = add <16 x i16> %res3, %res2
ret <16 x i16> %res4
}
declare <8 x i16> @llvm.x86.avx512.mask.psll.wi.128(<8 x i16>, i32, <8 x i16>, i8)
define <8 x i16>@test_int_x86_avx512_mask_psll_wi_128(<8 x i16> %x0, i32 %x1, <8 x i16> %x2, i8 %x3) {
; CHECK-LABEL: test_int_x86_avx512_mask_psll_wi_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpsllw $3, %xmm0, %xmm2 ## EVEX TO VEX Compression encoding: [0xc5,0xe9,0x71,0xf0,0x03]
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %esi, %k1 ## encoding: [0xc5,0xfb,0x92,0xce]
[AVX-512] Remove intrinsics for 128/256-bit masked shift by immediate. Instead upgrade them to a select and the older SSE/AVX2 intrinsic. llvm-svn: 286072
2016-11-07 00:29:14 +08:00
; CHECK-NEXT: vpsllw $3, %xmm0, %xmm1 {%k1} ## encoding: [0x62,0xf1,0x75,0x09,0x71,0xf0,0x03]
; CHECK-NEXT: vpsllw $3, %xmm0, %xmm0 {%k1} {z} ## encoding: [0x62,0xf1,0x7d,0x89,0x71,0xf0,0x03]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpaddw %xmm2, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xfd,0xc2]
[X86][SKX][KNL] Updated regression tests to use -mattr instead of -mcpu flag.NFC. NFC. Updated 8 regression tests to use -mattr instead of -mcpu flag as follows: -mcpu=knl --> -mattr=+avx512f -mcpu=skx --> -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq The updates are as part of the preparation of a large commit to add all instruction scheduling for the SKX target. Reviewers: delena, zvi, RKSimon Differential Revision: https://reviews.llvm.org/D38222 Change-Id: I2381c9b5bb75ecacfca017243c22d054f6eddd14 llvm-svn: 314306
2017-09-27 22:44:15 +08:00
; CHECK-NEXT: vpaddw %xmm0, %xmm1, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf1,0xfd,0xc0]
[AVX-512] Remove intrinsics for 128/256-bit masked shift by immediate. Instead upgrade them to a select and the older SSE/AVX2 intrinsic. llvm-svn: 286072
2016-11-07 00:29:14 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <8 x i16> @llvm.x86.avx512.mask.psll.wi.128(<8 x i16> %x0, i32 3, <8 x i16> %x2, i8 %x3)
%res1 = call <8 x i16> @llvm.x86.avx512.mask.psll.wi.128(<8 x i16> %x0, i32 3, <8 x i16> zeroinitializer, i8 %x3)
%res2 = call <8 x i16> @llvm.x86.avx512.mask.psll.wi.128(<8 x i16> %x0, i32 3, <8 x i16> %x2, i8 -1)
%res3 = add <8 x i16> %res, %res1
%res4 = add <8 x i16> %res3, %res2
ret <8 x i16> %res4
}
declare <16 x i16> @llvm.x86.avx512.mask.psll.wi.256(<16 x i16>, i32, <16 x i16>, i16)
define <16 x i16>@test_int_x86_avx512_mask_psll_wi_256(<16 x i16> %x0, i32 %x1, <16 x i16> %x2, i16 %x3) {
; CHECK-LABEL: test_int_x86_avx512_mask_psll_wi_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpsllw $3, %ymm0, %ymm2 ## EVEX TO VEX Compression encoding: [0xc5,0xed,0x71,0xf0,0x03]
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %esi, %k1 ## encoding: [0xc5,0xfb,0x92,0xce]
[AVX-512] Remove intrinsics for 128/256-bit masked shift by immediate. Instead upgrade them to a select and the older SSE/AVX2 intrinsic. llvm-svn: 286072
2016-11-07 00:29:14 +08:00
; CHECK-NEXT: vpsllw $3, %ymm0, %ymm1 {%k1} ## encoding: [0x62,0xf1,0x75,0x29,0x71,0xf0,0x03]
; CHECK-NEXT: vpsllw $3, %ymm0, %ymm0 {%k1} {z} ## encoding: [0x62,0xf1,0x7d,0xa9,0x71,0xf0,0x03]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpaddw %ymm2, %ymm0, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xfd,0xfd,0xc2]
[X86][SKX][KNL] Updated regression tests to use -mattr instead of -mcpu flag.NFC. NFC. Updated 8 regression tests to use -mattr instead of -mcpu flag as follows: -mcpu=knl --> -mattr=+avx512f -mcpu=skx --> -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq The updates are as part of the preparation of a large commit to add all instruction scheduling for the SKX target. Reviewers: delena, zvi, RKSimon Differential Revision: https://reviews.llvm.org/D38222 Change-Id: I2381c9b5bb75ecacfca017243c22d054f6eddd14 llvm-svn: 314306
2017-09-27 22:44:15 +08:00
; CHECK-NEXT: vpaddw %ymm0, %ymm1, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf5,0xfd,0xc0]
[AVX-512] Remove intrinsics for 128/256-bit masked shift by immediate. Instead upgrade them to a select and the older SSE/AVX2 intrinsic. llvm-svn: 286072
2016-11-07 00:29:14 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <16 x i16> @llvm.x86.avx512.mask.psll.wi.256(<16 x i16> %x0, i32 3, <16 x i16> %x2, i16 %x3)
%res1 = call <16 x i16> @llvm.x86.avx512.mask.psll.wi.256(<16 x i16> %x0, i32 3, <16 x i16> zeroinitializer, i16 %x3)
%res2 = call <16 x i16> @llvm.x86.avx512.mask.psll.wi.256(<16 x i16> %x0, i32 3, <16 x i16> %x2, i16 -1)
%res3 = add <16 x i16> %res, %res1
%res4 = add <16 x i16> %res3, %res2
ret <16 x i16> %res4
}
[AVX-512] Remove 128/256 masked pshufb intrinsics. Autoupgrade them to legacy intrinsics and a select. llvm-svn: 286089
2016-11-07 08:13:39 +08:00
declare <16 x i8> @llvm.x86.avx512.mask.pshuf.b.128(<16 x i8>, <16 x i8>, <16 x i8>, i16)
define <16 x i8>@test_int_x86_avx512_mask_pshuf_b_128(<16 x i8> %x0, <16 x i8> %x1, <16 x i8> %x2, i16 %x3) {
; CHECK-LABEL: test_int_x86_avx512_mask_pshuf_b_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpshufb %xmm1, %xmm0, %xmm3 ## EVEX TO VEX Compression encoding: [0xc4,0xe2,0x79,0x00,0xd9]
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
[AVX-512] Remove 128/256 masked pshufb intrinsics. Autoupgrade them to legacy intrinsics and a select. llvm-svn: 286089
2016-11-07 08:13:39 +08:00
; CHECK-NEXT: vpshufb %xmm1, %xmm0, %xmm2 {%k1} ## encoding: [0x62,0xf2,0x7d,0x09,0x00,0xd1]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpaddb %xmm3, %xmm2, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xe9,0xfc,0xc3]
[AVX-512] Remove 128/256 masked pshufb intrinsics. Autoupgrade them to legacy intrinsics and a select. llvm-svn: 286089
2016-11-07 08:13:39 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <16 x i8> @llvm.x86.avx512.mask.pshuf.b.128(<16 x i8> %x0, <16 x i8> %x1, <16 x i8> %x2, i16 %x3)
%res1 = call <16 x i8> @llvm.x86.avx512.mask.pshuf.b.128(<16 x i8> %x0, <16 x i8> %x1, <16 x i8> %x2, i16 -1)
%res2 = add <16 x i8> %res, %res1
ret <16 x i8> %res2
}
declare <32 x i8> @llvm.x86.avx512.mask.pshuf.b.256(<32 x i8>, <32 x i8>, <32 x i8>, i32)
define <32 x i8>@test_int_x86_avx512_mask_pshuf_b_256(<32 x i8> %x0, <32 x i8> %x1, <32 x i8> %x2, i32 %x3) {
; CHECK-LABEL: test_int_x86_avx512_mask_pshuf_b_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpshufb %ymm1, %ymm0, %ymm3 ## EVEX TO VEX Compression encoding: [0xc4,0xe2,0x7d,0x00,0xd9]
[AVX-512] Remove 128/256 masked pshufb intrinsics. Autoupgrade them to legacy intrinsics and a select. llvm-svn: 286089
2016-11-07 08:13:39 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
; CHECK-NEXT: vpshufb %ymm1, %ymm0, %ymm2 {%k1} ## encoding: [0x62,0xf2,0x7d,0x29,0x00,0xd1]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpaddb %ymm3, %ymm2, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xed,0xfc,0xc3]
[AVX-512] Remove 128/256 masked pshufb intrinsics. Autoupgrade them to legacy intrinsics and a select. llvm-svn: 286089
2016-11-07 08:13:39 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <32 x i8> @llvm.x86.avx512.mask.pshuf.b.256(<32 x i8> %x0, <32 x i8> %x1, <32 x i8> %x2, i32 %x3)
%res1 = call <32 x i8> @llvm.x86.avx512.mask.pshuf.b.256(<32 x i8> %x0, <32 x i8> %x1, <32 x i8> %x2, i32 -1)
%res2 = add <32 x i8> %res, %res1
ret <32 x i8> %res2
}
[AVX-512] Remove masked pmovzx/pmovsx builtins and autoupgrade them to selects and native zext/sext. This mostly reuses earlier autoupgrade support for the sse and avx equivalents. Just needed to add the code to add the select. llvm-svn: 286092
2016-11-07 10:12:57 +08:00
declare <8 x i16> @llvm.x86.avx512.mask.pmovzxb.w.128(<16 x i8>, <8 x i16>, i8)
define <8 x i16>@test_int_x86_avx512_mask_pmovzxb_w_128(<16 x i8> %x0, <8 x i16> %x1, i8 %x2) {
; CHECK-LABEL: test_int_x86_avx512_mask_pmovzxb_w_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpmovzxbw %xmm0, %xmm2 ## EVEX TO VEX Compression encoding: [0xc4,0xe2,0x79,0x30,0xd0]
[AVX-512] Remove masked pmovzx/pmovsx builtins and autoupgrade them to selects and native zext/sext. This mostly reuses earlier autoupgrade support for the sse and avx equivalents. Just needed to add the code to add the select. llvm-svn: 286092
2016-11-07 10:12:57 +08:00
; CHECK-NEXT: ## xmm2 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
[AVX-512] Remove masked pmovzx/pmovsx builtins and autoupgrade them to selects and native zext/sext. This mostly reuses earlier autoupgrade support for the sse and avx equivalents. Just needed to add the code to add the select. llvm-svn: 286092
2016-11-07 10:12:57 +08:00
; CHECK-NEXT: vpmovzxbw %xmm0, %xmm1 {%k1} ## encoding: [0x62,0xf2,0x7d,0x09,0x30,0xc8]
; CHECK-NEXT: ## xmm1 {%k1} = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
; CHECK-NEXT: vpmovzxbw %xmm0, %xmm0 {%k1} {z} ## encoding: [0x62,0xf2,0x7d,0x89,0x30,0xc0]
; CHECK-NEXT: ## xmm0 {%k1} {z} = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpaddw %xmm2, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xfd,0xc2]
[X86][SKX][KNL] Updated regression tests to use -mattr instead of -mcpu flag.NFC. NFC. Updated 8 regression tests to use -mattr instead of -mcpu flag as follows: -mcpu=knl --> -mattr=+avx512f -mcpu=skx --> -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq The updates are as part of the preparation of a large commit to add all instruction scheduling for the SKX target. Reviewers: delena, zvi, RKSimon Differential Revision: https://reviews.llvm.org/D38222 Change-Id: I2381c9b5bb75ecacfca017243c22d054f6eddd14 llvm-svn: 314306
2017-09-27 22:44:15 +08:00
; CHECK-NEXT: vpaddw %xmm0, %xmm1, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf1,0xfd,0xc0]
[AVX-512] Remove masked pmovzx/pmovsx builtins and autoupgrade them to selects and native zext/sext. This mostly reuses earlier autoupgrade support for the sse and avx equivalents. Just needed to add the code to add the select. llvm-svn: 286092
2016-11-07 10:12:57 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <8 x i16> @llvm.x86.avx512.mask.pmovzxb.w.128(<16 x i8> %x0, <8 x i16> %x1, i8 %x2)
%res1 = call <8 x i16> @llvm.x86.avx512.mask.pmovzxb.w.128(<16 x i8> %x0, <8 x i16> zeroinitializer, i8 %x2)
%res2 = call <8 x i16> @llvm.x86.avx512.mask.pmovzxb.w.128(<16 x i8> %x0, <8 x i16> %x1, i8 -1)
%res3 = add <8 x i16> %res, %res1
%res4 = add <8 x i16> %res3, %res2
ret <8 x i16> %res4
}
declare <16 x i16> @llvm.x86.avx512.mask.pmovzxb.w.256(<16 x i8>, <16 x i16>, i16)
define <16 x i16>@test_int_x86_avx512_mask_pmovzxb_w_256(<16 x i8> %x0, <16 x i16> %x1, i16 %x2) {
; CHECK-LABEL: test_int_x86_avx512_mask_pmovzxb_w_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpmovzxbw %xmm0, %ymm2 ## EVEX TO VEX Compression encoding: [0xc4,0xe2,0x7d,0x30,0xd0]
[AVX-512] Remove masked pmovzx/pmovsx builtins and autoupgrade them to selects and native zext/sext. This mostly reuses earlier autoupgrade support for the sse and avx equivalents. Just needed to add the code to add the select. llvm-svn: 286092
2016-11-07 10:12:57 +08:00
; CHECK-NEXT: ## ymm2 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero,xmm0[8],zero,xmm0[9],zero,xmm0[10],zero,xmm0[11],zero,xmm0[12],zero,xmm0[13],zero,xmm0[14],zero,xmm0[15],zero
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
[AVX-512] Remove masked pmovzx/pmovsx builtins and autoupgrade them to selects and native zext/sext. This mostly reuses earlier autoupgrade support for the sse and avx equivalents. Just needed to add the code to add the select. llvm-svn: 286092
2016-11-07 10:12:57 +08:00
; CHECK-NEXT: vpmovzxbw %xmm0, %ymm1 {%k1} ## encoding: [0x62,0xf2,0x7d,0x29,0x30,0xc8]
; CHECK-NEXT: ## ymm1 {%k1} = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero,xmm0[8],zero,xmm0[9],zero,xmm0[10],zero,xmm0[11],zero,xmm0[12],zero,xmm0[13],zero,xmm0[14],zero,xmm0[15],zero
; CHECK-NEXT: vpmovzxbw %xmm0, %ymm0 {%k1} {z} ## encoding: [0x62,0xf2,0x7d,0xa9,0x30,0xc0]
; CHECK-NEXT: ## ymm0 {%k1} {z} = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero,xmm0[8],zero,xmm0[9],zero,xmm0[10],zero,xmm0[11],zero,xmm0[12],zero,xmm0[13],zero,xmm0[14],zero,xmm0[15],zero
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpaddw %ymm2, %ymm0, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xfd,0xfd,0xc2]
[X86][SKX][KNL] Updated regression tests to use -mattr instead of -mcpu flag.NFC. NFC. Updated 8 regression tests to use -mattr instead of -mcpu flag as follows: -mcpu=knl --> -mattr=+avx512f -mcpu=skx --> -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq The updates are as part of the preparation of a large commit to add all instruction scheduling for the SKX target. Reviewers: delena, zvi, RKSimon Differential Revision: https://reviews.llvm.org/D38222 Change-Id: I2381c9b5bb75ecacfca017243c22d054f6eddd14 llvm-svn: 314306
2017-09-27 22:44:15 +08:00
; CHECK-NEXT: vpaddw %ymm0, %ymm1, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf5,0xfd,0xc0]
[AVX-512] Remove masked pmovzx/pmovsx builtins and autoupgrade them to selects and native zext/sext. This mostly reuses earlier autoupgrade support for the sse and avx equivalents. Just needed to add the code to add the select. llvm-svn: 286092
2016-11-07 10:12:57 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <16 x i16> @llvm.x86.avx512.mask.pmovzxb.w.256(<16 x i8> %x0, <16 x i16> %x1, i16 %x2)
%res1 = call <16 x i16> @llvm.x86.avx512.mask.pmovzxb.w.256(<16 x i8> %x0, <16 x i16> zeroinitializer, i16 %x2)
%res2 = call <16 x i16> @llvm.x86.avx512.mask.pmovzxb.w.256(<16 x i8> %x0, <16 x i16> %x1, i16 -1)
%res3 = add <16 x i16> %res, %res1
%res4 = add <16 x i16> %res3, %res2
ret <16 x i16> %res4
}
declare <8 x i16> @llvm.x86.avx512.mask.pmovsxb.w.128(<16 x i8>, <8 x i16>, i8)
define <8 x i16>@test_int_x86_avx512_mask_pmovsxb_w_128(<16 x i8> %x0, <8 x i16> %x1, i8 %x2) {
; CHECK-LABEL: test_int_x86_avx512_mask_pmovsxb_w_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpmovsxbw %xmm0, %xmm2 ## EVEX TO VEX Compression encoding: [0xc4,0xe2,0x79,0x20,0xd0]
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
[AVX-512] Remove masked pmovzx/pmovsx builtins and autoupgrade them to selects and native zext/sext. This mostly reuses earlier autoupgrade support for the sse and avx equivalents. Just needed to add the code to add the select. llvm-svn: 286092
2016-11-07 10:12:57 +08:00
; CHECK-NEXT: vpmovsxbw %xmm0, %xmm1 {%k1} ## encoding: [0x62,0xf2,0x7d,0x09,0x20,0xc8]
; CHECK-NEXT: vpmovsxbw %xmm0, %xmm0 {%k1} {z} ## encoding: [0x62,0xf2,0x7d,0x89,0x20,0xc0]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpaddw %xmm2, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xfd,0xc2]
[X86][SKX][KNL] Updated regression tests to use -mattr instead of -mcpu flag.NFC. NFC. Updated 8 regression tests to use -mattr instead of -mcpu flag as follows: -mcpu=knl --> -mattr=+avx512f -mcpu=skx --> -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq The updates are as part of the preparation of a large commit to add all instruction scheduling for the SKX target. Reviewers: delena, zvi, RKSimon Differential Revision: https://reviews.llvm.org/D38222 Change-Id: I2381c9b5bb75ecacfca017243c22d054f6eddd14 llvm-svn: 314306
2017-09-27 22:44:15 +08:00
; CHECK-NEXT: vpaddw %xmm0, %xmm1, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf1,0xfd,0xc0]
[AVX-512] Remove masked pmovzx/pmovsx builtins and autoupgrade them to selects and native zext/sext. This mostly reuses earlier autoupgrade support for the sse and avx equivalents. Just needed to add the code to add the select. llvm-svn: 286092
2016-11-07 10:12:57 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <8 x i16> @llvm.x86.avx512.mask.pmovsxb.w.128(<16 x i8> %x0, <8 x i16> %x1, i8 %x2)
%res1 = call <8 x i16> @llvm.x86.avx512.mask.pmovsxb.w.128(<16 x i8> %x0, <8 x i16> zeroinitializer, i8 %x2)
%res2 = call <8 x i16> @llvm.x86.avx512.mask.pmovsxb.w.128(<16 x i8> %x0, <8 x i16> %x1, i8 -1)
%res3 = add <8 x i16> %res, %res1
%res4 = add <8 x i16> %res3, %res2
ret <8 x i16> %res4
}
declare <16 x i16> @llvm.x86.avx512.mask.pmovsxb.w.256(<16 x i8>, <16 x i16>, i16)
define <16 x i16>@test_int_x86_avx512_mask_pmovsxb_w_256(<16 x i8> %x0, <16 x i16> %x1, i16 %x2) {
; CHECK-LABEL: test_int_x86_avx512_mask_pmovsxb_w_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpmovsxbw %xmm0, %ymm2 ## EVEX TO VEX Compression encoding: [0xc4,0xe2,0x7d,0x20,0xd0]
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
[AVX-512] Remove masked pmovzx/pmovsx builtins and autoupgrade them to selects and native zext/sext. This mostly reuses earlier autoupgrade support for the sse and avx equivalents. Just needed to add the code to add the select. llvm-svn: 286092
2016-11-07 10:12:57 +08:00
; CHECK-NEXT: vpmovsxbw %xmm0, %ymm1 {%k1} ## encoding: [0x62,0xf2,0x7d,0x29,0x20,0xc8]
; CHECK-NEXT: vpmovsxbw %xmm0, %ymm0 {%k1} {z} ## encoding: [0x62,0xf2,0x7d,0xa9,0x20,0xc0]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpaddw %ymm2, %ymm0, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xfd,0xfd,0xc2]
[X86][SKX][KNL] Updated regression tests to use -mattr instead of -mcpu flag.NFC. NFC. Updated 8 regression tests to use -mattr instead of -mcpu flag as follows: -mcpu=knl --> -mattr=+avx512f -mcpu=skx --> -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq The updates are as part of the preparation of a large commit to add all instruction scheduling for the SKX target. Reviewers: delena, zvi, RKSimon Differential Revision: https://reviews.llvm.org/D38222 Change-Id: I2381c9b5bb75ecacfca017243c22d054f6eddd14 llvm-svn: 314306
2017-09-27 22:44:15 +08:00
; CHECK-NEXT: vpaddw %ymm0, %ymm1, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf5,0xfd,0xc0]
[AVX-512] Remove masked pmovzx/pmovsx builtins and autoupgrade them to selects and native zext/sext. This mostly reuses earlier autoupgrade support for the sse and avx equivalents. Just needed to add the code to add the select. llvm-svn: 286092
2016-11-07 10:12:57 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <16 x i16> @llvm.x86.avx512.mask.pmovsxb.w.256(<16 x i8> %x0, <16 x i16> %x1, i16 %x2)
%res1 = call <16 x i16> @llvm.x86.avx512.mask.pmovsxb.w.256(<16 x i8> %x0, <16 x i16> zeroinitializer, i16 %x2)
%res2 = call <16 x i16> @llvm.x86.avx512.mask.pmovsxb.w.256(<16 x i8> %x0, <16 x i16> %x1, i16 -1)
%res3 = add <16 x i16> %res, %res1
%res4 = add <16 x i16> %res3, %res2
ret <16 x i16> %res4
}
declare <2 x i64> @llvm.x86.avx512.mask.pmovsxd.q.128(<4 x i32>, <2 x i64>, i8)
define <2 x i64>@test_int_x86_avx512_mask_pmovsxd_q_128(<4 x i32> %x0, <2 x i64> %x1, i8 %x2) {
; CHECK-LABEL: test_int_x86_avx512_mask_pmovsxd_q_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpmovsxdq %xmm0, %xmm2 ## EVEX TO VEX Compression encoding: [0xc4,0xe2,0x79,0x25,0xd0]
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
[AVX-512] Remove masked pmovzx/pmovsx builtins and autoupgrade them to selects and native zext/sext. This mostly reuses earlier autoupgrade support for the sse and avx equivalents. Just needed to add the code to add the select. llvm-svn: 286092
2016-11-07 10:12:57 +08:00
; CHECK-NEXT: vpmovsxdq %xmm0, %xmm1 {%k1} ## encoding: [0x62,0xf2,0x7d,0x09,0x25,0xc8]
; CHECK-NEXT: vpmovsxdq %xmm0, %xmm0 {%k1} {z} ## encoding: [0x62,0xf2,0x7d,0x89,0x25,0xc0]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpaddq %xmm2, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xd4,0xc2]
[X86][SKX][KNL] Updated regression tests to use -mattr instead of -mcpu flag.NFC. NFC. Updated 8 regression tests to use -mattr instead of -mcpu flag as follows: -mcpu=knl --> -mattr=+avx512f -mcpu=skx --> -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq The updates are as part of the preparation of a large commit to add all instruction scheduling for the SKX target. Reviewers: delena, zvi, RKSimon Differential Revision: https://reviews.llvm.org/D38222 Change-Id: I2381c9b5bb75ecacfca017243c22d054f6eddd14 llvm-svn: 314306
2017-09-27 22:44:15 +08:00
; CHECK-NEXT: vpaddq %xmm0, %xmm1, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf1,0xd4,0xc0]
[AVX-512] Remove masked pmovzx/pmovsx builtins and autoupgrade them to selects and native zext/sext. This mostly reuses earlier autoupgrade support for the sse and avx equivalents. Just needed to add the code to add the select. llvm-svn: 286092
2016-11-07 10:12:57 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <2 x i64> @llvm.x86.avx512.mask.pmovsxd.q.128(<4 x i32> %x0, <2 x i64> %x1, i8 %x2)
%res1 = call <2 x i64> @llvm.x86.avx512.mask.pmovsxd.q.128(<4 x i32> %x0, <2 x i64> zeroinitializer, i8 %x2)
%res2 = call <2 x i64> @llvm.x86.avx512.mask.pmovsxd.q.128(<4 x i32> %x0, <2 x i64> %x1, i8 -1)
%res3 = add <2 x i64> %res, %res1
%res4 = add <2 x i64> %res3, %res2
ret <2 x i64> %res4
}
declare <4 x i64> @llvm.x86.avx512.mask.pmovsxd.q.256(<4 x i32>, <4 x i64>, i8)
define <4 x i64>@test_int_x86_avx512_mask_pmovsxd_q_256(<4 x i32> %x0, <4 x i64> %x1, i8 %x2) {
; CHECK-LABEL: test_int_x86_avx512_mask_pmovsxd_q_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpmovsxdq %xmm0, %ymm2 ## EVEX TO VEX Compression encoding: [0xc4,0xe2,0x7d,0x25,0xd0]
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
[AVX-512] Remove masked pmovzx/pmovsx builtins and autoupgrade them to selects and native zext/sext. This mostly reuses earlier autoupgrade support for the sse and avx equivalents. Just needed to add the code to add the select. llvm-svn: 286092
2016-11-07 10:12:57 +08:00
; CHECK-NEXT: vpmovsxdq %xmm0, %ymm1 {%k1} ## encoding: [0x62,0xf2,0x7d,0x29,0x25,0xc8]
; CHECK-NEXT: vpmovsxdq %xmm0, %ymm0 {%k1} {z} ## encoding: [0x62,0xf2,0x7d,0xa9,0x25,0xc0]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible. There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers. The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled. Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky Differential Revision: https://reviews.llvm.org/D27901 llvm-svn: 290663
2016-12-28 18:12:48 +08:00
; CHECK-NEXT: vpaddq %ymm2, %ymm0, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xfd,0xd4,0xc2]
[X86][SKX][KNL] Updated regression tests to use -mattr instead of -mcpu flag.NFC. NFC. Updated 8 regression tests to use -mattr instead of -mcpu flag as follows: -mcpu=knl --> -mattr=+avx512f -mcpu=skx --> -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq The updates are as part of the preparation of a large commit to add all instruction scheduling for the SKX target. Reviewers: delena, zvi, RKSimon Differential Revision: https://reviews.llvm.org/D38222 Change-Id: I2381c9b5bb75ecacfca017243c22d054f6eddd14 llvm-svn: 314306
2017-09-27 22:44:15 +08:00
; CHECK-NEXT: vpaddq %ymm0, %ymm1, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf5,0xd4,0xc0]
[AVX-512] Remove masked pmovzx/pmovsx builtins and autoupgrade them to selects and native zext/sext. This mostly reuses earlier autoupgrade support for the sse and avx equivalents. Just needed to add the code to add the select. llvm-svn: 286092
2016-11-07 10:12:57 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <4 x i64> @llvm.x86.avx512.mask.pmovsxd.q.256(<4 x i32> %x0, <4 x i64> %x1, i8 %x2)
%res1 = call <4 x i64> @llvm.x86.avx512.mask.pmovsxd.q.256(<4 x i32> %x0, <4 x i64> zeroinitializer, i8 %x2)
%res2 = call <4 x i64> @llvm.x86.avx512.mask.pmovsxd.q.256(<4 x i32> %x0, <4 x i64> %x1, i8 -1)
%res3 = add <4 x i64> %res, %res1
%res4 = add <4 x i64> %res3, %res2
ret <4 x i64> %res4
}
[AVX-512] Remove 128/256 masked pshufb intrinsics. Autoupgrade them to legacy intrinsics and a select. llvm-svn: 286089
2016-11-07 08:13:39 +08:00
[X86][LLVM] Converting __mm{|256|512}_movm_epi{8|16|32|64} LLVMIR call into generic intrinsics. This patch is a part one of two reviews, one for the clang and the other for LLVM. The patch deletes the back-end intrinsics and adds support for them in the auto upgrade. Differential Revision: https://reviews.llvm.org/D31393 llvm-svn: 299432
2017-04-04 21:32:14 +08:00
declare <16 x i8> @llvm.x86.avx512.cvtmask2b.128(i16)
define <16 x i8>@test_int_x86_avx512_cvtmask2b_128(i16 %x0) {
; CHECK-LABEL: test_int_x86_avx512_cvtmask2b_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[X86][LLVM] Converting __mm{|256|512}_movm_epi{8|16|32|64} LLVMIR call into generic intrinsics. This patch is a part one of two reviews, one for the clang and the other for LLVM. The patch deletes the back-end intrinsics and adds support for them in the auto upgrade. Differential Revision: https://reviews.llvm.org/D31393 llvm-svn: 299432
2017-04-04 21:32:14 +08:00
; CHECK-NEXT: kmovd %edi, %k0 ## encoding: [0xc5,0xfb,0x92,0xc7]
; CHECK-NEXT: vpmovm2b %k0, %xmm0 ## encoding: [0x62,0xf2,0x7e,0x08,0x28,0xc0]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <16 x i8> @llvm.x86.avx512.cvtmask2b.128(i16 %x0)
ret <16 x i8> %res
}
declare <32 x i8> @llvm.x86.avx512.cvtmask2b.256(i32)
define <32 x i8>@test_int_x86_avx512_cvtmask2b_256(i32 %x0) {
; CHECK-LABEL: test_int_x86_avx512_cvtmask2b_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[X86][LLVM] Converting __mm{|256|512}_movm_epi{8|16|32|64} LLVMIR call into generic intrinsics. This patch is a part one of two reviews, one for the clang and the other for LLVM. The patch deletes the back-end intrinsics and adds support for them in the auto upgrade. Differential Revision: https://reviews.llvm.org/D31393 llvm-svn: 299432
2017-04-04 21:32:14 +08:00
; CHECK-NEXT: kmovd %edi, %k0 ## encoding: [0xc5,0xfb,0x92,0xc7]
; CHECK-NEXT: vpmovm2b %k0, %ymm0 ## encoding: [0x62,0xf2,0x7e,0x28,0x28,0xc0]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <32 x i8> @llvm.x86.avx512.cvtmask2b.256(i32 %x0)
ret <32 x i8> %res
}
declare <8 x i16> @llvm.x86.avx512.cvtmask2w.128(i8)
define <8 x i16>@test_int_x86_avx512_cvtmask2w_128(i8 %x0) {
; CHECK-LABEL: test_int_x86_avx512_cvtmask2w_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[X86][LLVM] Converting __mm{|256|512}_movm_epi{8|16|32|64} LLVMIR call into generic intrinsics. This patch is a part one of two reviews, one for the clang and the other for LLVM. The patch deletes the back-end intrinsics and adds support for them in the auto upgrade. Differential Revision: https://reviews.llvm.org/D31393 llvm-svn: 299432
2017-04-04 21:32:14 +08:00
; CHECK-NEXT: kmovd %edi, %k0 ## encoding: [0xc5,0xfb,0x92,0xc7]
; CHECK-NEXT: vpmovm2w %k0, %xmm0 ## encoding: [0x62,0xf2,0xfe,0x08,0x28,0xc0]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <8 x i16> @llvm.x86.avx512.cvtmask2w.128(i8 %x0)
ret <8 x i16> %res
}
declare <16 x i16> @llvm.x86.avx512.cvtmask2w.256(i16)
define <16 x i16>@test_int_x86_avx512_cvtmask2w_256(i16 %x0) {
; CHECK-LABEL: test_int_x86_avx512_cvtmask2w_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[X86][LLVM] Converting __mm{|256|512}_movm_epi{8|16|32|64} LLVMIR call into generic intrinsics. This patch is a part one of two reviews, one for the clang and the other for LLVM. The patch deletes the back-end intrinsics and adds support for them in the auto upgrade. Differential Revision: https://reviews.llvm.org/D31393 llvm-svn: 299432
2017-04-04 21:32:14 +08:00
; CHECK-NEXT: kmovd %edi, %k0 ## encoding: [0xc5,0xfb,0x92,0xc7]
; CHECK-NEXT: vpmovm2w %k0, %ymm0 ## encoding: [0x62,0xf2,0xfe,0x28,0x28,0xc0]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <16 x i16> @llvm.x86.avx512.cvtmask2w.256(i16 %x0)
ret <16 x i16> %res
}
[AVX-512] Remove masked packss/packus intrinsics and autoupgrade to unmasked intrinsics with select instructions. For 512-bit add new unmasked intrinsics. The new 512-bit unmasked intrinsics will make it easy to handle these with the SSE/AVX intrinsics in InstCombine where we currently have a TODO. llvm-svn: 295290
2017-02-16 14:31:54 +08:00
define <8 x i16> @test_mask_packs_epi32_rr_128(<4 x i32> %a, <4 x i32> %b) {
; CHECK-LABEL: test_mask_packs_epi32_rr_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove masked packss/packus intrinsics and autoupgrade to unmasked intrinsics with select instructions. For 512-bit add new unmasked intrinsics. The new 512-bit unmasked intrinsics will make it easy to handle these with the SSE/AVX intrinsics in InstCombine where we currently have a TODO. llvm-svn: 295290
2017-02-16 14:31:54 +08:00
; CHECK-NEXT: vpackssdw %xmm1, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6b,0xc1]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <8 x i16> @llvm.x86.avx512.mask.packssdw.128(<4 x i32> %a, <4 x i32> %b, <8 x i16> zeroinitializer, i8 -1)
ret <8 x i16> %res
}
define <8 x i16> @test_mask_packs_epi32_rrk_128(<4 x i32> %a, <4 x i32> %b, <8 x i16> %passThru, i8 %mask) {
; CHECK-LABEL: test_mask_packs_epi32_rrk_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
[AVX-512] Remove masked packss/packus intrinsics and autoupgrade to unmasked intrinsics with select instructions. For 512-bit add new unmasked intrinsics. The new 512-bit unmasked intrinsics will make it easy to handle these with the SSE/AVX intrinsics in InstCombine where we currently have a TODO. llvm-svn: 295290
2017-02-16 14:31:54 +08:00
; CHECK-NEXT: vpackssdw %xmm1, %xmm0, %xmm2 {%k1} ## encoding: [0x62,0xf1,0x7d,0x09,0x6b,0xd1]
; CHECK-NEXT: vmovdqa %xmm2, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6f,0xc2]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <8 x i16> @llvm.x86.avx512.mask.packssdw.128(<4 x i32> %a, <4 x i32> %b, <8 x i16> %passThru, i8 %mask)
ret <8 x i16> %res
}
define <8 x i16> @test_mask_packs_epi32_rrkz_128(<4 x i32> %a, <4 x i32> %b, i8 %mask) {
; CHECK-LABEL: test_mask_packs_epi32_rrkz_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
[AVX-512] Remove masked packss/packus intrinsics and autoupgrade to unmasked intrinsics with select instructions. For 512-bit add new unmasked intrinsics. The new 512-bit unmasked intrinsics will make it easy to handle these with the SSE/AVX intrinsics in InstCombine where we currently have a TODO. llvm-svn: 295290
2017-02-16 14:31:54 +08:00
; CHECK-NEXT: vpackssdw %xmm1, %xmm0, %xmm0 {%k1} {z} ## encoding: [0x62,0xf1,0x7d,0x89,0x6b,0xc1]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <8 x i16> @llvm.x86.avx512.mask.packssdw.128(<4 x i32> %a, <4 x i32> %b, <8 x i16> zeroinitializer, i8 %mask)
ret <8 x i16> %res
}
define <8 x i16> @test_mask_packs_epi32_rm_128(<4 x i32> %a, <4 x i32>* %ptr_b) {
; CHECK-LABEL: test_mask_packs_epi32_rm_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove masked packss/packus intrinsics and autoupgrade to unmasked intrinsics with select instructions. For 512-bit add new unmasked intrinsics. The new 512-bit unmasked intrinsics will make it easy to handle these with the SSE/AVX intrinsics in InstCombine where we currently have a TODO. llvm-svn: 295290
2017-02-16 14:31:54 +08:00
; CHECK-NEXT: vpackssdw (%rdi), %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6b,0x07]
; CHECK-NEXT: retq ## encoding: [0xc3]
%b = load <4 x i32>, <4 x i32>* %ptr_b
%res = call <8 x i16> @llvm.x86.avx512.mask.packssdw.128(<4 x i32> %a, <4 x i32> %b, <8 x i16> zeroinitializer, i8 -1)
ret <8 x i16> %res
}
define <8 x i16> @test_mask_packs_epi32_rmk_128(<4 x i32> %a, <4 x i32>* %ptr_b, <8 x i16> %passThru, i8 %mask) {
; CHECK-LABEL: test_mask_packs_epi32_rmk_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %esi, %k1 ## encoding: [0xc5,0xfb,0x92,0xce]
[AVX-512] Remove masked packss/packus intrinsics and autoupgrade to unmasked intrinsics with select instructions. For 512-bit add new unmasked intrinsics. The new 512-bit unmasked intrinsics will make it easy to handle these with the SSE/AVX intrinsics in InstCombine where we currently have a TODO. llvm-svn: 295290
2017-02-16 14:31:54 +08:00
; CHECK-NEXT: vpackssdw (%rdi), %xmm0, %xmm1 {%k1} ## encoding: [0x62,0xf1,0x7d,0x09,0x6b,0x0f]
; CHECK-NEXT: vmovdqa %xmm1, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6f,0xc1]
; CHECK-NEXT: retq ## encoding: [0xc3]
%b = load <4 x i32>, <4 x i32>* %ptr_b
%res = call <8 x i16> @llvm.x86.avx512.mask.packssdw.128(<4 x i32> %a, <4 x i32> %b, <8 x i16> %passThru, i8 %mask)
ret <8 x i16> %res
}
define <8 x i16> @test_mask_packs_epi32_rmkz_128(<4 x i32> %a, <4 x i32>* %ptr_b, i8 %mask) {
; CHECK-LABEL: test_mask_packs_epi32_rmkz_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %esi, %k1 ## encoding: [0xc5,0xfb,0x92,0xce]
[AVX-512] Remove masked packss/packus intrinsics and autoupgrade to unmasked intrinsics with select instructions. For 512-bit add new unmasked intrinsics. The new 512-bit unmasked intrinsics will make it easy to handle these with the SSE/AVX intrinsics in InstCombine where we currently have a TODO. llvm-svn: 295290
2017-02-16 14:31:54 +08:00
; CHECK-NEXT: vpackssdw (%rdi), %xmm0, %xmm0 {%k1} {z} ## encoding: [0x62,0xf1,0x7d,0x89,0x6b,0x07]
; CHECK-NEXT: retq ## encoding: [0xc3]
%b = load <4 x i32>, <4 x i32>* %ptr_b
%res = call <8 x i16> @llvm.x86.avx512.mask.packssdw.128(<4 x i32> %a, <4 x i32> %b, <8 x i16> zeroinitializer, i8 %mask)
ret <8 x i16> %res
}
define <8 x i16> @test_mask_packs_epi32_rmb_128(<4 x i32> %a, i32* %ptr_b) {
; CHECK-LABEL: test_mask_packs_epi32_rmb_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove masked packss/packus intrinsics and autoupgrade to unmasked intrinsics with select instructions. For 512-bit add new unmasked intrinsics. The new 512-bit unmasked intrinsics will make it easy to handle these with the SSE/AVX intrinsics in InstCombine where we currently have a TODO. llvm-svn: 295290
2017-02-16 14:31:54 +08:00
; CHECK-NEXT: vpackssdw (%rdi){1to4}, %xmm0, %xmm0 ## encoding: [0x62,0xf1,0x7d,0x18,0x6b,0x07]
; CHECK-NEXT: retq ## encoding: [0xc3]
%q = load i32, i32* %ptr_b
%vecinit.i = insertelement <4 x i32> undef, i32 %q, i32 0
%b = shufflevector <4 x i32> %vecinit.i, <4 x i32> undef, <4 x i32> zeroinitializer
%res = call <8 x i16> @llvm.x86.avx512.mask.packssdw.128(<4 x i32> %a, <4 x i32> %b, <8 x i16> zeroinitializer, i8 -1)
ret <8 x i16> %res
}
define <8 x i16> @test_mask_packs_epi32_rmbk_128(<4 x i32> %a, i32* %ptr_b, <8 x i16> %passThru, i8 %mask) {
; CHECK-LABEL: test_mask_packs_epi32_rmbk_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %esi, %k1 ## encoding: [0xc5,0xfb,0x92,0xce]
[AVX-512] Remove masked packss/packus intrinsics and autoupgrade to unmasked intrinsics with select instructions. For 512-bit add new unmasked intrinsics. The new 512-bit unmasked intrinsics will make it easy to handle these with the SSE/AVX intrinsics in InstCombine where we currently have a TODO. llvm-svn: 295290
2017-02-16 14:31:54 +08:00
; CHECK-NEXT: vpackssdw (%rdi){1to4}, %xmm0, %xmm1 {%k1} ## encoding: [0x62,0xf1,0x7d,0x19,0x6b,0x0f]
; CHECK-NEXT: vmovdqa %xmm1, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6f,0xc1]
; CHECK-NEXT: retq ## encoding: [0xc3]
%q = load i32, i32* %ptr_b
%vecinit.i = insertelement <4 x i32> undef, i32 %q, i32 0
%b = shufflevector <4 x i32> %vecinit.i, <4 x i32> undef, <4 x i32> zeroinitializer
%res = call <8 x i16> @llvm.x86.avx512.mask.packssdw.128(<4 x i32> %a, <4 x i32> %b, <8 x i16> %passThru, i8 %mask)
ret <8 x i16> %res
}
define <8 x i16> @test_mask_packs_epi32_rmbkz_128(<4 x i32> %a, i32* %ptr_b, i8 %mask) {
; CHECK-LABEL: test_mask_packs_epi32_rmbkz_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %esi, %k1 ## encoding: [0xc5,0xfb,0x92,0xce]
[AVX-512] Remove masked packss/packus intrinsics and autoupgrade to unmasked intrinsics with select instructions. For 512-bit add new unmasked intrinsics. The new 512-bit unmasked intrinsics will make it easy to handle these with the SSE/AVX intrinsics in InstCombine where we currently have a TODO. llvm-svn: 295290
2017-02-16 14:31:54 +08:00
; CHECK-NEXT: vpackssdw (%rdi){1to4}, %xmm0, %xmm0 {%k1} {z} ## encoding: [0x62,0xf1,0x7d,0x99,0x6b,0x07]
; CHECK-NEXT: retq ## encoding: [0xc3]
%q = load i32, i32* %ptr_b
%vecinit.i = insertelement <4 x i32> undef, i32 %q, i32 0
%b = shufflevector <4 x i32> %vecinit.i, <4 x i32> undef, <4 x i32> zeroinitializer
%res = call <8 x i16> @llvm.x86.avx512.mask.packssdw.128(<4 x i32> %a, <4 x i32> %b, <8 x i16> zeroinitializer, i8 %mask)
ret <8 x i16> %res
}
declare <8 x i16> @llvm.x86.avx512.mask.packssdw.128(<4 x i32>, <4 x i32>, <8 x i16>, i8)
define <16 x i16> @test_mask_packs_epi32_rr_256(<8 x i32> %a, <8 x i32> %b) {
; CHECK-LABEL: test_mask_packs_epi32_rr_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove masked packss/packus intrinsics and autoupgrade to unmasked intrinsics with select instructions. For 512-bit add new unmasked intrinsics. The new 512-bit unmasked intrinsics will make it easy to handle these with the SSE/AVX intrinsics in InstCombine where we currently have a TODO. llvm-svn: 295290
2017-02-16 14:31:54 +08:00
; CHECK-NEXT: vpackssdw %ymm1, %ymm0, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xfd,0x6b,0xc1]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <16 x i16> @llvm.x86.avx512.mask.packssdw.256(<8 x i32> %a, <8 x i32> %b, <16 x i16> zeroinitializer, i16 -1)
ret <16 x i16> %res
}
define <16 x i16> @test_mask_packs_epi32_rrk_256(<8 x i32> %a, <8 x i32> %b, <16 x i16> %passThru, i16 %mask) {
; CHECK-LABEL: test_mask_packs_epi32_rrk_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
[AVX-512] Remove masked packss/packus intrinsics and autoupgrade to unmasked intrinsics with select instructions. For 512-bit add new unmasked intrinsics. The new 512-bit unmasked intrinsics will make it easy to handle these with the SSE/AVX intrinsics in InstCombine where we currently have a TODO. llvm-svn: 295290
2017-02-16 14:31:54 +08:00
; CHECK-NEXT: vpackssdw %ymm1, %ymm0, %ymm2 {%k1} ## encoding: [0x62,0xf1,0x7d,0x29,0x6b,0xd1]
; CHECK-NEXT: vmovdqa %ymm2, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xfd,0x6f,0xc2]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <16 x i16> @llvm.x86.avx512.mask.packssdw.256(<8 x i32> %a, <8 x i32> %b, <16 x i16> %passThru, i16 %mask)
ret <16 x i16> %res
}
define <16 x i16> @test_mask_packs_epi32_rrkz_256(<8 x i32> %a, <8 x i32> %b, i16 %mask) {
; CHECK-LABEL: test_mask_packs_epi32_rrkz_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
[AVX-512] Remove masked packss/packus intrinsics and autoupgrade to unmasked intrinsics with select instructions. For 512-bit add new unmasked intrinsics. The new 512-bit unmasked intrinsics will make it easy to handle these with the SSE/AVX intrinsics in InstCombine where we currently have a TODO. llvm-svn: 295290
2017-02-16 14:31:54 +08:00
; CHECK-NEXT: vpackssdw %ymm1, %ymm0, %ymm0 {%k1} {z} ## encoding: [0x62,0xf1,0x7d,0xa9,0x6b,0xc1]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <16 x i16> @llvm.x86.avx512.mask.packssdw.256(<8 x i32> %a, <8 x i32> %b, <16 x i16> zeroinitializer, i16 %mask)
ret <16 x i16> %res
}
define <16 x i16> @test_mask_packs_epi32_rm_256(<8 x i32> %a, <8 x i32>* %ptr_b) {
; CHECK-LABEL: test_mask_packs_epi32_rm_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove masked packss/packus intrinsics and autoupgrade to unmasked intrinsics with select instructions. For 512-bit add new unmasked intrinsics. The new 512-bit unmasked intrinsics will make it easy to handle these with the SSE/AVX intrinsics in InstCombine where we currently have a TODO. llvm-svn: 295290
2017-02-16 14:31:54 +08:00
; CHECK-NEXT: vpackssdw (%rdi), %ymm0, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xfd,0x6b,0x07]
; CHECK-NEXT: retq ## encoding: [0xc3]
%b = load <8 x i32>, <8 x i32>* %ptr_b
%res = call <16 x i16> @llvm.x86.avx512.mask.packssdw.256(<8 x i32> %a, <8 x i32> %b, <16 x i16> zeroinitializer, i16 -1)
ret <16 x i16> %res
}
define <16 x i16> @test_mask_packs_epi32_rmk_256(<8 x i32> %a, <8 x i32>* %ptr_b, <16 x i16> %passThru, i16 %mask) {
; CHECK-LABEL: test_mask_packs_epi32_rmk_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %esi, %k1 ## encoding: [0xc5,0xfb,0x92,0xce]
[AVX-512] Remove masked packss/packus intrinsics and autoupgrade to unmasked intrinsics with select instructions. For 512-bit add new unmasked intrinsics. The new 512-bit unmasked intrinsics will make it easy to handle these with the SSE/AVX intrinsics in InstCombine where we currently have a TODO. llvm-svn: 295290
2017-02-16 14:31:54 +08:00
; CHECK-NEXT: vpackssdw (%rdi), %ymm0, %ymm1 {%k1} ## encoding: [0x62,0xf1,0x7d,0x29,0x6b,0x0f]
; CHECK-NEXT: vmovdqa %ymm1, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xfd,0x6f,0xc1]
; CHECK-NEXT: retq ## encoding: [0xc3]
%b = load <8 x i32>, <8 x i32>* %ptr_b
%res = call <16 x i16> @llvm.x86.avx512.mask.packssdw.256(<8 x i32> %a, <8 x i32> %b, <16 x i16> %passThru, i16 %mask)
ret <16 x i16> %res
}
define <16 x i16> @test_mask_packs_epi32_rmkz_256(<8 x i32> %a, <8 x i32>* %ptr_b, i16 %mask) {
; CHECK-LABEL: test_mask_packs_epi32_rmkz_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %esi, %k1 ## encoding: [0xc5,0xfb,0x92,0xce]
[AVX-512] Remove masked packss/packus intrinsics and autoupgrade to unmasked intrinsics with select instructions. For 512-bit add new unmasked intrinsics. The new 512-bit unmasked intrinsics will make it easy to handle these with the SSE/AVX intrinsics in InstCombine where we currently have a TODO. llvm-svn: 295290
2017-02-16 14:31:54 +08:00
; CHECK-NEXT: vpackssdw (%rdi), %ymm0, %ymm0 {%k1} {z} ## encoding: [0x62,0xf1,0x7d,0xa9,0x6b,0x07]
; CHECK-NEXT: retq ## encoding: [0xc3]
%b = load <8 x i32>, <8 x i32>* %ptr_b
%res = call <16 x i16> @llvm.x86.avx512.mask.packssdw.256(<8 x i32> %a, <8 x i32> %b, <16 x i16> zeroinitializer, i16 %mask)
ret <16 x i16> %res
}
define <16 x i16> @test_mask_packs_epi32_rmb_256(<8 x i32> %a, i32* %ptr_b) {
; CHECK-LABEL: test_mask_packs_epi32_rmb_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove masked packss/packus intrinsics and autoupgrade to unmasked intrinsics with select instructions. For 512-bit add new unmasked intrinsics. The new 512-bit unmasked intrinsics will make it easy to handle these with the SSE/AVX intrinsics in InstCombine where we currently have a TODO. llvm-svn: 295290
2017-02-16 14:31:54 +08:00
; CHECK-NEXT: vpackssdw (%rdi){1to8}, %ymm0, %ymm0 ## encoding: [0x62,0xf1,0x7d,0x38,0x6b,0x07]
; CHECK-NEXT: retq ## encoding: [0xc3]
%q = load i32, i32* %ptr_b
%vecinit.i = insertelement <8 x i32> undef, i32 %q, i32 0
%b = shufflevector <8 x i32> %vecinit.i, <8 x i32> undef, <8 x i32> zeroinitializer
%res = call <16 x i16> @llvm.x86.avx512.mask.packssdw.256(<8 x i32> %a, <8 x i32> %b, <16 x i16> zeroinitializer, i16 -1)
ret <16 x i16> %res
}
define <16 x i16> @test_mask_packs_epi32_rmbk_256(<8 x i32> %a, i32* %ptr_b, <16 x i16> %passThru, i16 %mask) {
; CHECK-LABEL: test_mask_packs_epi32_rmbk_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %esi, %k1 ## encoding: [0xc5,0xfb,0x92,0xce]
[AVX-512] Remove masked packss/packus intrinsics and autoupgrade to unmasked intrinsics with select instructions. For 512-bit add new unmasked intrinsics. The new 512-bit unmasked intrinsics will make it easy to handle these with the SSE/AVX intrinsics in InstCombine where we currently have a TODO. llvm-svn: 295290
2017-02-16 14:31:54 +08:00
; CHECK-NEXT: vpackssdw (%rdi){1to8}, %ymm0, %ymm1 {%k1} ## encoding: [0x62,0xf1,0x7d,0x39,0x6b,0x0f]
; CHECK-NEXT: vmovdqa %ymm1, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xfd,0x6f,0xc1]
; CHECK-NEXT: retq ## encoding: [0xc3]
%q = load i32, i32* %ptr_b
%vecinit.i = insertelement <8 x i32> undef, i32 %q, i32 0
%b = shufflevector <8 x i32> %vecinit.i, <8 x i32> undef, <8 x i32> zeroinitializer
%res = call <16 x i16> @llvm.x86.avx512.mask.packssdw.256(<8 x i32> %a, <8 x i32> %b, <16 x i16> %passThru, i16 %mask)
ret <16 x i16> %res
}
define <16 x i16> @test_mask_packs_epi32_rmbkz_256(<8 x i32> %a, i32* %ptr_b, i16 %mask) {
; CHECK-LABEL: test_mask_packs_epi32_rmbkz_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %esi, %k1 ## encoding: [0xc5,0xfb,0x92,0xce]
[AVX-512] Remove masked packss/packus intrinsics and autoupgrade to unmasked intrinsics with select instructions. For 512-bit add new unmasked intrinsics. The new 512-bit unmasked intrinsics will make it easy to handle these with the SSE/AVX intrinsics in InstCombine where we currently have a TODO. llvm-svn: 295290
2017-02-16 14:31:54 +08:00
; CHECK-NEXT: vpackssdw (%rdi){1to8}, %ymm0, %ymm0 {%k1} {z} ## encoding: [0x62,0xf1,0x7d,0xb9,0x6b,0x07]
; CHECK-NEXT: retq ## encoding: [0xc3]
%q = load i32, i32* %ptr_b
%vecinit.i = insertelement <8 x i32> undef, i32 %q, i32 0
%b = shufflevector <8 x i32> %vecinit.i, <8 x i32> undef, <8 x i32> zeroinitializer
%res = call <16 x i16> @llvm.x86.avx512.mask.packssdw.256(<8 x i32> %a, <8 x i32> %b, <16 x i16> zeroinitializer, i16 %mask)
ret <16 x i16> %res
}
declare <16 x i16> @llvm.x86.avx512.mask.packssdw.256(<8 x i32>, <8 x i32>, <16 x i16>, i16)
define <16 x i8> @test_mask_packs_epi16_rr_128(<8 x i16> %a, <8 x i16> %b) {
; CHECK-LABEL: test_mask_packs_epi16_rr_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove masked packss/packus intrinsics and autoupgrade to unmasked intrinsics with select instructions. For 512-bit add new unmasked intrinsics. The new 512-bit unmasked intrinsics will make it easy to handle these with the SSE/AVX intrinsics in InstCombine where we currently have a TODO. llvm-svn: 295290
2017-02-16 14:31:54 +08:00
; CHECK-NEXT: vpacksswb %xmm1, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0x63,0xc1]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <16 x i8> @llvm.x86.avx512.mask.packsswb.128(<8 x i16> %a, <8 x i16> %b, <16 x i8> zeroinitializer, i16 -1)
ret <16 x i8> %res
}
define <16 x i8> @test_mask_packs_epi16_rrk_128(<8 x i16> %a, <8 x i16> %b, <16 x i8> %passThru, i16 %mask) {
; CHECK-LABEL: test_mask_packs_epi16_rrk_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
[AVX-512] Remove masked packss/packus intrinsics and autoupgrade to unmasked intrinsics with select instructions. For 512-bit add new unmasked intrinsics. The new 512-bit unmasked intrinsics will make it easy to handle these with the SSE/AVX intrinsics in InstCombine where we currently have a TODO. llvm-svn: 295290
2017-02-16 14:31:54 +08:00
; CHECK-NEXT: vpacksswb %xmm1, %xmm0, %xmm2 {%k1} ## encoding: [0x62,0xf1,0x7d,0x09,0x63,0xd1]
; CHECK-NEXT: vmovdqa %xmm2, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6f,0xc2]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <16 x i8> @llvm.x86.avx512.mask.packsswb.128(<8 x i16> %a, <8 x i16> %b, <16 x i8> %passThru, i16 %mask)
ret <16 x i8> %res
}
define <16 x i8> @test_mask_packs_epi16_rrkz_128(<8 x i16> %a, <8 x i16> %b, i16 %mask) {
; CHECK-LABEL: test_mask_packs_epi16_rrkz_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
[AVX-512] Remove masked packss/packus intrinsics and autoupgrade to unmasked intrinsics with select instructions. For 512-bit add new unmasked intrinsics. The new 512-bit unmasked intrinsics will make it easy to handle these with the SSE/AVX intrinsics in InstCombine where we currently have a TODO. llvm-svn: 295290
2017-02-16 14:31:54 +08:00
; CHECK-NEXT: vpacksswb %xmm1, %xmm0, %xmm0 {%k1} {z} ## encoding: [0x62,0xf1,0x7d,0x89,0x63,0xc1]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <16 x i8> @llvm.x86.avx512.mask.packsswb.128(<8 x i16> %a, <8 x i16> %b, <16 x i8> zeroinitializer, i16 %mask)
ret <16 x i8> %res
}
define <16 x i8> @test_mask_packs_epi16_rm_128(<8 x i16> %a, <8 x i16>* %ptr_b) {
; CHECK-LABEL: test_mask_packs_epi16_rm_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove masked packss/packus intrinsics and autoupgrade to unmasked intrinsics with select instructions. For 512-bit add new unmasked intrinsics. The new 512-bit unmasked intrinsics will make it easy to handle these with the SSE/AVX intrinsics in InstCombine where we currently have a TODO. llvm-svn: 295290
2017-02-16 14:31:54 +08:00
; CHECK-NEXT: vpacksswb (%rdi), %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0x63,0x07]
; CHECK-NEXT: retq ## encoding: [0xc3]
%b = load <8 x i16>, <8 x i16>* %ptr_b
%res = call <16 x i8> @llvm.x86.avx512.mask.packsswb.128(<8 x i16> %a, <8 x i16> %b, <16 x i8> zeroinitializer, i16 -1)
ret <16 x i8> %res
}
define <16 x i8> @test_mask_packs_epi16_rmk_128(<8 x i16> %a, <8 x i16>* %ptr_b, <16 x i8> %passThru, i16 %mask) {
; CHECK-LABEL: test_mask_packs_epi16_rmk_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %esi, %k1 ## encoding: [0xc5,0xfb,0x92,0xce]
[AVX-512] Remove masked packss/packus intrinsics and autoupgrade to unmasked intrinsics with select instructions. For 512-bit add new unmasked intrinsics. The new 512-bit unmasked intrinsics will make it easy to handle these with the SSE/AVX intrinsics in InstCombine where we currently have a TODO. llvm-svn: 295290
2017-02-16 14:31:54 +08:00
; CHECK-NEXT: vpacksswb (%rdi), %xmm0, %xmm1 {%k1} ## encoding: [0x62,0xf1,0x7d,0x09,0x63,0x0f]
; CHECK-NEXT: vmovdqa %xmm1, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6f,0xc1]
; CHECK-NEXT: retq ## encoding: [0xc3]
%b = load <8 x i16>, <8 x i16>* %ptr_b
%res = call <16 x i8> @llvm.x86.avx512.mask.packsswb.128(<8 x i16> %a, <8 x i16> %b, <16 x i8> %passThru, i16 %mask)
ret <16 x i8> %res
}
define <16 x i8> @test_mask_packs_epi16_rmkz_128(<8 x i16> %a, <8 x i16>* %ptr_b, i16 %mask) {
; CHECK-LABEL: test_mask_packs_epi16_rmkz_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %esi, %k1 ## encoding: [0xc5,0xfb,0x92,0xce]
[AVX-512] Remove masked packss/packus intrinsics and autoupgrade to unmasked intrinsics with select instructions. For 512-bit add new unmasked intrinsics. The new 512-bit unmasked intrinsics will make it easy to handle these with the SSE/AVX intrinsics in InstCombine where we currently have a TODO. llvm-svn: 295290
2017-02-16 14:31:54 +08:00
; CHECK-NEXT: vpacksswb (%rdi), %xmm0, %xmm0 {%k1} {z} ## encoding: [0x62,0xf1,0x7d,0x89,0x63,0x07]
; CHECK-NEXT: retq ## encoding: [0xc3]
%b = load <8 x i16>, <8 x i16>* %ptr_b
%res = call <16 x i8> @llvm.x86.avx512.mask.packsswb.128(<8 x i16> %a, <8 x i16> %b, <16 x i8> zeroinitializer, i16 %mask)
ret <16 x i8> %res
}
declare <16 x i8> @llvm.x86.avx512.mask.packsswb.128(<8 x i16>, <8 x i16>, <16 x i8>, i16)
define <32 x i8> @test_mask_packs_epi16_rr_256(<16 x i16> %a, <16 x i16> %b) {
; CHECK-LABEL: test_mask_packs_epi16_rr_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove masked packss/packus intrinsics and autoupgrade to unmasked intrinsics with select instructions. For 512-bit add new unmasked intrinsics. The new 512-bit unmasked intrinsics will make it easy to handle these with the SSE/AVX intrinsics in InstCombine where we currently have a TODO. llvm-svn: 295290
2017-02-16 14:31:54 +08:00
; CHECK-NEXT: vpacksswb %ymm1, %ymm0, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xfd,0x63,0xc1]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <32 x i8> @llvm.x86.avx512.mask.packsswb.256(<16 x i16> %a, <16 x i16> %b, <32 x i8> zeroinitializer, i32 -1)
ret <32 x i8> %res
}
define <32 x i8> @test_mask_packs_epi16_rrk_256(<16 x i16> %a, <16 x i16> %b, <32 x i8> %passThru, i32 %mask) {
; CHECK-LABEL: test_mask_packs_epi16_rrk_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove masked packss/packus intrinsics and autoupgrade to unmasked intrinsics with select instructions. For 512-bit add new unmasked intrinsics. The new 512-bit unmasked intrinsics will make it easy to handle these with the SSE/AVX intrinsics in InstCombine where we currently have a TODO. llvm-svn: 295290
2017-02-16 14:31:54 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
; CHECK-NEXT: vpacksswb %ymm1, %ymm0, %ymm2 {%k1} ## encoding: [0x62,0xf1,0x7d,0x29,0x63,0xd1]
; CHECK-NEXT: vmovdqa %ymm2, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xfd,0x6f,0xc2]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <32 x i8> @llvm.x86.avx512.mask.packsswb.256(<16 x i16> %a, <16 x i16> %b, <32 x i8> %passThru, i32 %mask)
ret <32 x i8> %res
}
define <32 x i8> @test_mask_packs_epi16_rrkz_256(<16 x i16> %a, <16 x i16> %b, i32 %mask) {
; CHECK-LABEL: test_mask_packs_epi16_rrkz_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove masked packss/packus intrinsics and autoupgrade to unmasked intrinsics with select instructions. For 512-bit add new unmasked intrinsics. The new 512-bit unmasked intrinsics will make it easy to handle these with the SSE/AVX intrinsics in InstCombine where we currently have a TODO. llvm-svn: 295290
2017-02-16 14:31:54 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
; CHECK-NEXT: vpacksswb %ymm1, %ymm0, %ymm0 {%k1} {z} ## encoding: [0x62,0xf1,0x7d,0xa9,0x63,0xc1]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <32 x i8> @llvm.x86.avx512.mask.packsswb.256(<16 x i16> %a, <16 x i16> %b, <32 x i8> zeroinitializer, i32 %mask)
ret <32 x i8> %res
}
define <32 x i8> @test_mask_packs_epi16_rm_256(<16 x i16> %a, <16 x i16>* %ptr_b) {
; CHECK-LABEL: test_mask_packs_epi16_rm_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove masked packss/packus intrinsics and autoupgrade to unmasked intrinsics with select instructions. For 512-bit add new unmasked intrinsics. The new 512-bit unmasked intrinsics will make it easy to handle these with the SSE/AVX intrinsics in InstCombine where we currently have a TODO. llvm-svn: 295290
2017-02-16 14:31:54 +08:00
; CHECK-NEXT: vpacksswb (%rdi), %ymm0, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xfd,0x63,0x07]
; CHECK-NEXT: retq ## encoding: [0xc3]
%b = load <16 x i16>, <16 x i16>* %ptr_b
%res = call <32 x i8> @llvm.x86.avx512.mask.packsswb.256(<16 x i16> %a, <16 x i16> %b, <32 x i8> zeroinitializer, i32 -1)
ret <32 x i8> %res
}
define <32 x i8> @test_mask_packs_epi16_rmk_256(<16 x i16> %a, <16 x i16>* %ptr_b, <32 x i8> %passThru, i32 %mask) {
; CHECK-LABEL: test_mask_packs_epi16_rmk_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove masked packss/packus intrinsics and autoupgrade to unmasked intrinsics with select instructions. For 512-bit add new unmasked intrinsics. The new 512-bit unmasked intrinsics will make it easy to handle these with the SSE/AVX intrinsics in InstCombine where we currently have a TODO. llvm-svn: 295290
2017-02-16 14:31:54 +08:00
; CHECK-NEXT: kmovd %esi, %k1 ## encoding: [0xc5,0xfb,0x92,0xce]
; CHECK-NEXT: vpacksswb (%rdi), %ymm0, %ymm1 {%k1} ## encoding: [0x62,0xf1,0x7d,0x29,0x63,0x0f]
; CHECK-NEXT: vmovdqa %ymm1, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xfd,0x6f,0xc1]
; CHECK-NEXT: retq ## encoding: [0xc3]
%b = load <16 x i16>, <16 x i16>* %ptr_b
%res = call <32 x i8> @llvm.x86.avx512.mask.packsswb.256(<16 x i16> %a, <16 x i16> %b, <32 x i8> %passThru, i32 %mask)
ret <32 x i8> %res
}
define <32 x i8> @test_mask_packs_epi16_rmkz_256(<16 x i16> %a, <16 x i16>* %ptr_b, i32 %mask) {
; CHECK-LABEL: test_mask_packs_epi16_rmkz_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove masked packss/packus intrinsics and autoupgrade to unmasked intrinsics with select instructions. For 512-bit add new unmasked intrinsics. The new 512-bit unmasked intrinsics will make it easy to handle these with the SSE/AVX intrinsics in InstCombine where we currently have a TODO. llvm-svn: 295290
2017-02-16 14:31:54 +08:00
; CHECK-NEXT: kmovd %esi, %k1 ## encoding: [0xc5,0xfb,0x92,0xce]
; CHECK-NEXT: vpacksswb (%rdi), %ymm0, %ymm0 {%k1} {z} ## encoding: [0x62,0xf1,0x7d,0xa9,0x63,0x07]
; CHECK-NEXT: retq ## encoding: [0xc3]
%b = load <16 x i16>, <16 x i16>* %ptr_b
%res = call <32 x i8> @llvm.x86.avx512.mask.packsswb.256(<16 x i16> %a, <16 x i16> %b, <32 x i8> zeroinitializer, i32 %mask)
ret <32 x i8> %res
}
declare <32 x i8> @llvm.x86.avx512.mask.packsswb.256(<16 x i16>, <16 x i16>, <32 x i8>, i32)
define <8 x i16> @test_mask_packus_epi32_rr_128(<4 x i32> %a, <4 x i32> %b) {
; CHECK-LABEL: test_mask_packus_epi32_rr_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove masked packss/packus intrinsics and autoupgrade to unmasked intrinsics with select instructions. For 512-bit add new unmasked intrinsics. The new 512-bit unmasked intrinsics will make it easy to handle these with the SSE/AVX intrinsics in InstCombine where we currently have a TODO. llvm-svn: 295290
2017-02-16 14:31:54 +08:00
; CHECK-NEXT: vpackusdw %xmm1, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc4,0xe2,0x79,0x2b,0xc1]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <8 x i16> @llvm.x86.avx512.mask.packusdw.128(<4 x i32> %a, <4 x i32> %b, <8 x i16> zeroinitializer, i8 -1)
ret <8 x i16> %res
}
define <8 x i16> @test_mask_packus_epi32_rrk_128(<4 x i32> %a, <4 x i32> %b, <8 x i16> %passThru, i8 %mask) {
; CHECK-LABEL: test_mask_packus_epi32_rrk_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
[AVX-512] Remove masked packss/packus intrinsics and autoupgrade to unmasked intrinsics with select instructions. For 512-bit add new unmasked intrinsics. The new 512-bit unmasked intrinsics will make it easy to handle these with the SSE/AVX intrinsics in InstCombine where we currently have a TODO. llvm-svn: 295290
2017-02-16 14:31:54 +08:00
; CHECK-NEXT: vpackusdw %xmm1, %xmm0, %xmm2 {%k1} ## encoding: [0x62,0xf2,0x7d,0x09,0x2b,0xd1]
; CHECK-NEXT: vmovdqa %xmm2, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6f,0xc2]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <8 x i16> @llvm.x86.avx512.mask.packusdw.128(<4 x i32> %a, <4 x i32> %b, <8 x i16> %passThru, i8 %mask)
ret <8 x i16> %res
}
define <8 x i16> @test_mask_packus_epi32_rrkz_128(<4 x i32> %a, <4 x i32> %b, i8 %mask) {
; CHECK-LABEL: test_mask_packus_epi32_rrkz_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
[AVX-512] Remove masked packss/packus intrinsics and autoupgrade to unmasked intrinsics with select instructions. For 512-bit add new unmasked intrinsics. The new 512-bit unmasked intrinsics will make it easy to handle these with the SSE/AVX intrinsics in InstCombine where we currently have a TODO. llvm-svn: 295290
2017-02-16 14:31:54 +08:00
; CHECK-NEXT: vpackusdw %xmm1, %xmm0, %xmm0 {%k1} {z} ## encoding: [0x62,0xf2,0x7d,0x89,0x2b,0xc1]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <8 x i16> @llvm.x86.avx512.mask.packusdw.128(<4 x i32> %a, <4 x i32> %b, <8 x i16> zeroinitializer, i8 %mask)
ret <8 x i16> %res
}
define <8 x i16> @test_mask_packus_epi32_rm_128(<4 x i32> %a, <4 x i32>* %ptr_b) {
; CHECK-LABEL: test_mask_packus_epi32_rm_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove masked packss/packus intrinsics and autoupgrade to unmasked intrinsics with select instructions. For 512-bit add new unmasked intrinsics. The new 512-bit unmasked intrinsics will make it easy to handle these with the SSE/AVX intrinsics in InstCombine where we currently have a TODO. llvm-svn: 295290
2017-02-16 14:31:54 +08:00
; CHECK-NEXT: vpackusdw (%rdi), %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc4,0xe2,0x79,0x2b,0x07]
; CHECK-NEXT: retq ## encoding: [0xc3]
%b = load <4 x i32>, <4 x i32>* %ptr_b
%res = call <8 x i16> @llvm.x86.avx512.mask.packusdw.128(<4 x i32> %a, <4 x i32> %b, <8 x i16> zeroinitializer, i8 -1)
ret <8 x i16> %res
}
define <8 x i16> @test_mask_packus_epi32_rmk_128(<4 x i32> %a, <4 x i32>* %ptr_b, <8 x i16> %passThru, i8 %mask) {
; CHECK-LABEL: test_mask_packus_epi32_rmk_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %esi, %k1 ## encoding: [0xc5,0xfb,0x92,0xce]
[AVX-512] Remove masked packss/packus intrinsics and autoupgrade to unmasked intrinsics with select instructions. For 512-bit add new unmasked intrinsics. The new 512-bit unmasked intrinsics will make it easy to handle these with the SSE/AVX intrinsics in InstCombine where we currently have a TODO. llvm-svn: 295290
2017-02-16 14:31:54 +08:00
; CHECK-NEXT: vpackusdw (%rdi), %xmm0, %xmm1 {%k1} ## encoding: [0x62,0xf2,0x7d,0x09,0x2b,0x0f]
; CHECK-NEXT: vmovdqa %xmm1, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6f,0xc1]
; CHECK-NEXT: retq ## encoding: [0xc3]
%b = load <4 x i32>, <4 x i32>* %ptr_b
%res = call <8 x i16> @llvm.x86.avx512.mask.packusdw.128(<4 x i32> %a, <4 x i32> %b, <8 x i16> %passThru, i8 %mask)
ret <8 x i16> %res
}
define <8 x i16> @test_mask_packus_epi32_rmkz_128(<4 x i32> %a, <4 x i32>* %ptr_b, i8 %mask) {
; CHECK-LABEL: test_mask_packus_epi32_rmkz_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %esi, %k1 ## encoding: [0xc5,0xfb,0x92,0xce]
[AVX-512] Remove masked packss/packus intrinsics and autoupgrade to unmasked intrinsics with select instructions. For 512-bit add new unmasked intrinsics. The new 512-bit unmasked intrinsics will make it easy to handle these with the SSE/AVX intrinsics in InstCombine where we currently have a TODO. llvm-svn: 295290
2017-02-16 14:31:54 +08:00
; CHECK-NEXT: vpackusdw (%rdi), %xmm0, %xmm0 {%k1} {z} ## encoding: [0x62,0xf2,0x7d,0x89,0x2b,0x07]
; CHECK-NEXT: retq ## encoding: [0xc3]
%b = load <4 x i32>, <4 x i32>* %ptr_b
%res = call <8 x i16> @llvm.x86.avx512.mask.packusdw.128(<4 x i32> %a, <4 x i32> %b, <8 x i16> zeroinitializer, i8 %mask)
ret <8 x i16> %res
}
define <8 x i16> @test_mask_packus_epi32_rmb_128(<4 x i32> %a, i32* %ptr_b) {
; CHECK-LABEL: test_mask_packus_epi32_rmb_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove masked packss/packus intrinsics and autoupgrade to unmasked intrinsics with select instructions. For 512-bit add new unmasked intrinsics. The new 512-bit unmasked intrinsics will make it easy to handle these with the SSE/AVX intrinsics in InstCombine where we currently have a TODO. llvm-svn: 295290
2017-02-16 14:31:54 +08:00
; CHECK-NEXT: vpackusdw (%rdi){1to4}, %xmm0, %xmm0 ## encoding: [0x62,0xf2,0x7d,0x18,0x2b,0x07]
; CHECK-NEXT: retq ## encoding: [0xc3]
%q = load i32, i32* %ptr_b
%vecinit.i = insertelement <4 x i32> undef, i32 %q, i32 0
%b = shufflevector <4 x i32> %vecinit.i, <4 x i32> undef, <4 x i32> zeroinitializer
%res = call <8 x i16> @llvm.x86.avx512.mask.packusdw.128(<4 x i32> %a, <4 x i32> %b, <8 x i16> zeroinitializer, i8 -1)
ret <8 x i16> %res
}
define <8 x i16> @test_mask_packus_epi32_rmbk_128(<4 x i32> %a, i32* %ptr_b, <8 x i16> %passThru, i8 %mask) {
; CHECK-LABEL: test_mask_packus_epi32_rmbk_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %esi, %k1 ## encoding: [0xc5,0xfb,0x92,0xce]
[AVX-512] Remove masked packss/packus intrinsics and autoupgrade to unmasked intrinsics with select instructions. For 512-bit add new unmasked intrinsics. The new 512-bit unmasked intrinsics will make it easy to handle these with the SSE/AVX intrinsics in InstCombine where we currently have a TODO. llvm-svn: 295290
2017-02-16 14:31:54 +08:00
; CHECK-NEXT: vpackusdw (%rdi){1to4}, %xmm0, %xmm1 {%k1} ## encoding: [0x62,0xf2,0x7d,0x19,0x2b,0x0f]
; CHECK-NEXT: vmovdqa %xmm1, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6f,0xc1]
; CHECK-NEXT: retq ## encoding: [0xc3]
%q = load i32, i32* %ptr_b
%vecinit.i = insertelement <4 x i32> undef, i32 %q, i32 0
%b = shufflevector <4 x i32> %vecinit.i, <4 x i32> undef, <4 x i32> zeroinitializer
%res = call <8 x i16> @llvm.x86.avx512.mask.packusdw.128(<4 x i32> %a, <4 x i32> %b, <8 x i16> %passThru, i8 %mask)
ret <8 x i16> %res
}
define <8 x i16> @test_mask_packus_epi32_rmbkz_128(<4 x i32> %a, i32* %ptr_b, i8 %mask) {
; CHECK-LABEL: test_mask_packus_epi32_rmbkz_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %esi, %k1 ## encoding: [0xc5,0xfb,0x92,0xce]
[AVX-512] Remove masked packss/packus intrinsics and autoupgrade to unmasked intrinsics with select instructions. For 512-bit add new unmasked intrinsics. The new 512-bit unmasked intrinsics will make it easy to handle these with the SSE/AVX intrinsics in InstCombine where we currently have a TODO. llvm-svn: 295290
2017-02-16 14:31:54 +08:00
; CHECK-NEXT: vpackusdw (%rdi){1to4}, %xmm0, %xmm0 {%k1} {z} ## encoding: [0x62,0xf2,0x7d,0x99,0x2b,0x07]
; CHECK-NEXT: retq ## encoding: [0xc3]
%q = load i32, i32* %ptr_b
%vecinit.i = insertelement <4 x i32> undef, i32 %q, i32 0
%b = shufflevector <4 x i32> %vecinit.i, <4 x i32> undef, <4 x i32> zeroinitializer
%res = call <8 x i16> @llvm.x86.avx512.mask.packusdw.128(<4 x i32> %a, <4 x i32> %b, <8 x i16> zeroinitializer, i8 %mask)
ret <8 x i16> %res
}
declare <8 x i16> @llvm.x86.avx512.mask.packusdw.128(<4 x i32>, <4 x i32>, <8 x i16>, i8)
define <16 x i16> @test_mask_packus_epi32_rr_256(<8 x i32> %a, <8 x i32> %b) {
; CHECK-LABEL: test_mask_packus_epi32_rr_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove masked packss/packus intrinsics and autoupgrade to unmasked intrinsics with select instructions. For 512-bit add new unmasked intrinsics. The new 512-bit unmasked intrinsics will make it easy to handle these with the SSE/AVX intrinsics in InstCombine where we currently have a TODO. llvm-svn: 295290
2017-02-16 14:31:54 +08:00
; CHECK-NEXT: vpackusdw %ymm1, %ymm0, %ymm0 ## EVEX TO VEX Compression encoding: [0xc4,0xe2,0x7d,0x2b,0xc1]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <16 x i16> @llvm.x86.avx512.mask.packusdw.256(<8 x i32> %a, <8 x i32> %b, <16 x i16> zeroinitializer, i16 -1)
ret <16 x i16> %res
}
define <16 x i16> @test_mask_packus_epi32_rrk_256(<8 x i32> %a, <8 x i32> %b, <16 x i16> %passThru, i16 %mask) {
; CHECK-LABEL: test_mask_packus_epi32_rrk_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
[AVX-512] Remove masked packss/packus intrinsics and autoupgrade to unmasked intrinsics with select instructions. For 512-bit add new unmasked intrinsics. The new 512-bit unmasked intrinsics will make it easy to handle these with the SSE/AVX intrinsics in InstCombine where we currently have a TODO. llvm-svn: 295290
2017-02-16 14:31:54 +08:00
; CHECK-NEXT: vpackusdw %ymm1, %ymm0, %ymm2 {%k1} ## encoding: [0x62,0xf2,0x7d,0x29,0x2b,0xd1]
; CHECK-NEXT: vmovdqa %ymm2, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xfd,0x6f,0xc2]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <16 x i16> @llvm.x86.avx512.mask.packusdw.256(<8 x i32> %a, <8 x i32> %b, <16 x i16> %passThru, i16 %mask)
ret <16 x i16> %res
}
define <16 x i16> @test_mask_packus_epi32_rrkz_256(<8 x i32> %a, <8 x i32> %b, i16 %mask) {
; CHECK-LABEL: test_mask_packus_epi32_rrkz_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
[AVX-512] Remove masked packss/packus intrinsics and autoupgrade to unmasked intrinsics with select instructions. For 512-bit add new unmasked intrinsics. The new 512-bit unmasked intrinsics will make it easy to handle these with the SSE/AVX intrinsics in InstCombine where we currently have a TODO. llvm-svn: 295290
2017-02-16 14:31:54 +08:00
; CHECK-NEXT: vpackusdw %ymm1, %ymm0, %ymm0 {%k1} {z} ## encoding: [0x62,0xf2,0x7d,0xa9,0x2b,0xc1]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <16 x i16> @llvm.x86.avx512.mask.packusdw.256(<8 x i32> %a, <8 x i32> %b, <16 x i16> zeroinitializer, i16 %mask)
ret <16 x i16> %res
}
define <16 x i16> @test_mask_packus_epi32_rm_256(<8 x i32> %a, <8 x i32>* %ptr_b) {
; CHECK-LABEL: test_mask_packus_epi32_rm_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove masked packss/packus intrinsics and autoupgrade to unmasked intrinsics with select instructions. For 512-bit add new unmasked intrinsics. The new 512-bit unmasked intrinsics will make it easy to handle these with the SSE/AVX intrinsics in InstCombine where we currently have a TODO. llvm-svn: 295290
2017-02-16 14:31:54 +08:00
; CHECK-NEXT: vpackusdw (%rdi), %ymm0, %ymm0 ## EVEX TO VEX Compression encoding: [0xc4,0xe2,0x7d,0x2b,0x07]
; CHECK-NEXT: retq ## encoding: [0xc3]
%b = load <8 x i32>, <8 x i32>* %ptr_b
%res = call <16 x i16> @llvm.x86.avx512.mask.packusdw.256(<8 x i32> %a, <8 x i32> %b, <16 x i16> zeroinitializer, i16 -1)
ret <16 x i16> %res
}
define <16 x i16> @test_mask_packus_epi32_rmk_256(<8 x i32> %a, <8 x i32>* %ptr_b, <16 x i16> %passThru, i16 %mask) {
; CHECK-LABEL: test_mask_packus_epi32_rmk_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %esi, %k1 ## encoding: [0xc5,0xfb,0x92,0xce]
[AVX-512] Remove masked packss/packus intrinsics and autoupgrade to unmasked intrinsics with select instructions. For 512-bit add new unmasked intrinsics. The new 512-bit unmasked intrinsics will make it easy to handle these with the SSE/AVX intrinsics in InstCombine where we currently have a TODO. llvm-svn: 295290
2017-02-16 14:31:54 +08:00
; CHECK-NEXT: vpackusdw (%rdi), %ymm0, %ymm1 {%k1} ## encoding: [0x62,0xf2,0x7d,0x29,0x2b,0x0f]
; CHECK-NEXT: vmovdqa %ymm1, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xfd,0x6f,0xc1]
; CHECK-NEXT: retq ## encoding: [0xc3]
%b = load <8 x i32>, <8 x i32>* %ptr_b
%res = call <16 x i16> @llvm.x86.avx512.mask.packusdw.256(<8 x i32> %a, <8 x i32> %b, <16 x i16> %passThru, i16 %mask)
ret <16 x i16> %res
}
define <16 x i16> @test_mask_packus_epi32_rmkz_256(<8 x i32> %a, <8 x i32>* %ptr_b, i16 %mask) {
; CHECK-LABEL: test_mask_packus_epi32_rmkz_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %esi, %k1 ## encoding: [0xc5,0xfb,0x92,0xce]
[AVX-512] Remove masked packss/packus intrinsics and autoupgrade to unmasked intrinsics with select instructions. For 512-bit add new unmasked intrinsics. The new 512-bit unmasked intrinsics will make it easy to handle these with the SSE/AVX intrinsics in InstCombine where we currently have a TODO. llvm-svn: 295290
2017-02-16 14:31:54 +08:00
; CHECK-NEXT: vpackusdw (%rdi), %ymm0, %ymm0 {%k1} {z} ## encoding: [0x62,0xf2,0x7d,0xa9,0x2b,0x07]
; CHECK-NEXT: retq ## encoding: [0xc3]
%b = load <8 x i32>, <8 x i32>* %ptr_b
%res = call <16 x i16> @llvm.x86.avx512.mask.packusdw.256(<8 x i32> %a, <8 x i32> %b, <16 x i16> zeroinitializer, i16 %mask)
ret <16 x i16> %res
}
define <16 x i16> @test_mask_packus_epi32_rmb_256(<8 x i32> %a, i32* %ptr_b) {
; CHECK-LABEL: test_mask_packus_epi32_rmb_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove masked packss/packus intrinsics and autoupgrade to unmasked intrinsics with select instructions. For 512-bit add new unmasked intrinsics. The new 512-bit unmasked intrinsics will make it easy to handle these with the SSE/AVX intrinsics in InstCombine where we currently have a TODO. llvm-svn: 295290
2017-02-16 14:31:54 +08:00
; CHECK-NEXT: vpackusdw (%rdi){1to8}, %ymm0, %ymm0 ## encoding: [0x62,0xf2,0x7d,0x38,0x2b,0x07]
; CHECK-NEXT: retq ## encoding: [0xc3]
%q = load i32, i32* %ptr_b
%vecinit.i = insertelement <8 x i32> undef, i32 %q, i32 0
%b = shufflevector <8 x i32> %vecinit.i, <8 x i32> undef, <8 x i32> zeroinitializer
%res = call <16 x i16> @llvm.x86.avx512.mask.packusdw.256(<8 x i32> %a, <8 x i32> %b, <16 x i16> zeroinitializer, i16 -1)
ret <16 x i16> %res
}
define <16 x i16> @test_mask_packus_epi32_rmbk_256(<8 x i32> %a, i32* %ptr_b, <16 x i16> %passThru, i16 %mask) {
; CHECK-LABEL: test_mask_packus_epi32_rmbk_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %esi, %k1 ## encoding: [0xc5,0xfb,0x92,0xce]
[AVX-512] Remove masked packss/packus intrinsics and autoupgrade to unmasked intrinsics with select instructions. For 512-bit add new unmasked intrinsics. The new 512-bit unmasked intrinsics will make it easy to handle these with the SSE/AVX intrinsics in InstCombine where we currently have a TODO. llvm-svn: 295290
2017-02-16 14:31:54 +08:00
; CHECK-NEXT: vpackusdw (%rdi){1to8}, %ymm0, %ymm1 {%k1} ## encoding: [0x62,0xf2,0x7d,0x39,0x2b,0x0f]
; CHECK-NEXT: vmovdqa %ymm1, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xfd,0x6f,0xc1]
; CHECK-NEXT: retq ## encoding: [0xc3]
%q = load i32, i32* %ptr_b
%vecinit.i = insertelement <8 x i32> undef, i32 %q, i32 0
%b = shufflevector <8 x i32> %vecinit.i, <8 x i32> undef, <8 x i32> zeroinitializer
%res = call <16 x i16> @llvm.x86.avx512.mask.packusdw.256(<8 x i32> %a, <8 x i32> %b, <16 x i16> %passThru, i16 %mask)
ret <16 x i16> %res
}
define <16 x i16> @test_mask_packus_epi32_rmbkz_256(<8 x i32> %a, i32* %ptr_b, i16 %mask) {
; CHECK-LABEL: test_mask_packus_epi32_rmbkz_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %esi, %k1 ## encoding: [0xc5,0xfb,0x92,0xce]
[AVX-512] Remove masked packss/packus intrinsics and autoupgrade to unmasked intrinsics with select instructions. For 512-bit add new unmasked intrinsics. The new 512-bit unmasked intrinsics will make it easy to handle these with the SSE/AVX intrinsics in InstCombine where we currently have a TODO. llvm-svn: 295290
2017-02-16 14:31:54 +08:00
; CHECK-NEXT: vpackusdw (%rdi){1to8}, %ymm0, %ymm0 {%k1} {z} ## encoding: [0x62,0xf2,0x7d,0xb9,0x2b,0x07]
; CHECK-NEXT: retq ## encoding: [0xc3]
%q = load i32, i32* %ptr_b
%vecinit.i = insertelement <8 x i32> undef, i32 %q, i32 0
%b = shufflevector <8 x i32> %vecinit.i, <8 x i32> undef, <8 x i32> zeroinitializer
%res = call <16 x i16> @llvm.x86.avx512.mask.packusdw.256(<8 x i32> %a, <8 x i32> %b, <16 x i16> zeroinitializer, i16 %mask)
ret <16 x i16> %res
}
declare <16 x i16> @llvm.x86.avx512.mask.packusdw.256(<8 x i32>, <8 x i32>, <16 x i16>, i16)
define <16 x i8> @test_mask_packus_epi16_rr_128(<8 x i16> %a, <8 x i16> %b) {
; CHECK-LABEL: test_mask_packus_epi16_rr_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove masked packss/packus intrinsics and autoupgrade to unmasked intrinsics with select instructions. For 512-bit add new unmasked intrinsics. The new 512-bit unmasked intrinsics will make it easy to handle these with the SSE/AVX intrinsics in InstCombine where we currently have a TODO. llvm-svn: 295290
2017-02-16 14:31:54 +08:00
; CHECK-NEXT: vpackuswb %xmm1, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0x67,0xc1]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <16 x i8> @llvm.x86.avx512.mask.packuswb.128(<8 x i16> %a, <8 x i16> %b, <16 x i8> zeroinitializer, i16 -1)
ret <16 x i8> %res
}
define <16 x i8> @test_mask_packus_epi16_rrk_128(<8 x i16> %a, <8 x i16> %b, <16 x i8> %passThru, i16 %mask) {
; CHECK-LABEL: test_mask_packus_epi16_rrk_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
[AVX-512] Remove masked packss/packus intrinsics and autoupgrade to unmasked intrinsics with select instructions. For 512-bit add new unmasked intrinsics. The new 512-bit unmasked intrinsics will make it easy to handle these with the SSE/AVX intrinsics in InstCombine where we currently have a TODO. llvm-svn: 295290
2017-02-16 14:31:54 +08:00
; CHECK-NEXT: vpackuswb %xmm1, %xmm0, %xmm2 {%k1} ## encoding: [0x62,0xf1,0x7d,0x09,0x67,0xd1]
; CHECK-NEXT: vmovdqa %xmm2, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6f,0xc2]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <16 x i8> @llvm.x86.avx512.mask.packuswb.128(<8 x i16> %a, <8 x i16> %b, <16 x i8> %passThru, i16 %mask)
ret <16 x i8> %res
}
define <16 x i8> @test_mask_packus_epi16_rrkz_128(<8 x i16> %a, <8 x i16> %b, i16 %mask) {
; CHECK-LABEL: test_mask_packus_epi16_rrkz_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
[AVX-512] Remove masked packss/packus intrinsics and autoupgrade to unmasked intrinsics with select instructions. For 512-bit add new unmasked intrinsics. The new 512-bit unmasked intrinsics will make it easy to handle these with the SSE/AVX intrinsics in InstCombine where we currently have a TODO. llvm-svn: 295290
2017-02-16 14:31:54 +08:00
; CHECK-NEXT: vpackuswb %xmm1, %xmm0, %xmm0 {%k1} {z} ## encoding: [0x62,0xf1,0x7d,0x89,0x67,0xc1]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <16 x i8> @llvm.x86.avx512.mask.packuswb.128(<8 x i16> %a, <8 x i16> %b, <16 x i8> zeroinitializer, i16 %mask)
ret <16 x i8> %res
}
define <16 x i8> @test_mask_packus_epi16_rm_128(<8 x i16> %a, <8 x i16>* %ptr_b) {
; CHECK-LABEL: test_mask_packus_epi16_rm_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove masked packss/packus intrinsics and autoupgrade to unmasked intrinsics with select instructions. For 512-bit add new unmasked intrinsics. The new 512-bit unmasked intrinsics will make it easy to handle these with the SSE/AVX intrinsics in InstCombine where we currently have a TODO. llvm-svn: 295290
2017-02-16 14:31:54 +08:00
; CHECK-NEXT: vpackuswb (%rdi), %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0x67,0x07]
; CHECK-NEXT: retq ## encoding: [0xc3]
%b = load <8 x i16>, <8 x i16>* %ptr_b
%res = call <16 x i8> @llvm.x86.avx512.mask.packuswb.128(<8 x i16> %a, <8 x i16> %b, <16 x i8> zeroinitializer, i16 -1)
ret <16 x i8> %res
}
define <16 x i8> @test_mask_packus_epi16_rmk_128(<8 x i16> %a, <8 x i16>* %ptr_b, <16 x i8> %passThru, i16 %mask) {
; CHECK-LABEL: test_mask_packus_epi16_rmk_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %esi, %k1 ## encoding: [0xc5,0xfb,0x92,0xce]
[AVX-512] Remove masked packss/packus intrinsics and autoupgrade to unmasked intrinsics with select instructions. For 512-bit add new unmasked intrinsics. The new 512-bit unmasked intrinsics will make it easy to handle these with the SSE/AVX intrinsics in InstCombine where we currently have a TODO. llvm-svn: 295290
2017-02-16 14:31:54 +08:00
; CHECK-NEXT: vpackuswb (%rdi), %xmm0, %xmm1 {%k1} ## encoding: [0x62,0xf1,0x7d,0x09,0x67,0x0f]
; CHECK-NEXT: vmovdqa %xmm1, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6f,0xc1]
; CHECK-NEXT: retq ## encoding: [0xc3]
%b = load <8 x i16>, <8 x i16>* %ptr_b
%res = call <16 x i8> @llvm.x86.avx512.mask.packuswb.128(<8 x i16> %a, <8 x i16> %b, <16 x i8> %passThru, i16 %mask)
ret <16 x i8> %res
}
define <16 x i8> @test_mask_packus_epi16_rmkz_128(<8 x i16> %a, <8 x i16>* %ptr_b, i16 %mask) {
; CHECK-LABEL: test_mask_packus_epi16_rmkz_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovd %esi, %k1 ## encoding: [0xc5,0xfb,0x92,0xce]
[AVX-512] Remove masked packss/packus intrinsics and autoupgrade to unmasked intrinsics with select instructions. For 512-bit add new unmasked intrinsics. The new 512-bit unmasked intrinsics will make it easy to handle these with the SSE/AVX intrinsics in InstCombine where we currently have a TODO. llvm-svn: 295290
2017-02-16 14:31:54 +08:00
; CHECK-NEXT: vpackuswb (%rdi), %xmm0, %xmm0 {%k1} {z} ## encoding: [0x62,0xf1,0x7d,0x89,0x67,0x07]
; CHECK-NEXT: retq ## encoding: [0xc3]
%b = load <8 x i16>, <8 x i16>* %ptr_b
%res = call <16 x i8> @llvm.x86.avx512.mask.packuswb.128(<8 x i16> %a, <8 x i16> %b, <16 x i8> zeroinitializer, i16 %mask)
ret <16 x i8> %res
}
declare <16 x i8> @llvm.x86.avx512.mask.packuswb.128(<8 x i16>, <8 x i16>, <16 x i8>, i16)
define <32 x i8> @test_mask_packus_epi16_rr_256(<16 x i16> %a, <16 x i16> %b) {
; CHECK-LABEL: test_mask_packus_epi16_rr_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove masked packss/packus intrinsics and autoupgrade to unmasked intrinsics with select instructions. For 512-bit add new unmasked intrinsics. The new 512-bit unmasked intrinsics will make it easy to handle these with the SSE/AVX intrinsics in InstCombine where we currently have a TODO. llvm-svn: 295290
2017-02-16 14:31:54 +08:00
; CHECK-NEXT: vpackuswb %ymm1, %ymm0, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xfd,0x67,0xc1]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <32 x i8> @llvm.x86.avx512.mask.packuswb.256(<16 x i16> %a, <16 x i16> %b, <32 x i8> zeroinitializer, i32 -1)
ret <32 x i8> %res
}
define <32 x i8> @test_mask_packus_epi16_rrk_256(<16 x i16> %a, <16 x i16> %b, <32 x i8> %passThru, i32 %mask) {
; CHECK-LABEL: test_mask_packus_epi16_rrk_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove masked packss/packus intrinsics and autoupgrade to unmasked intrinsics with select instructions. For 512-bit add new unmasked intrinsics. The new 512-bit unmasked intrinsics will make it easy to handle these with the SSE/AVX intrinsics in InstCombine where we currently have a TODO. llvm-svn: 295290
2017-02-16 14:31:54 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
; CHECK-NEXT: vpackuswb %ymm1, %ymm0, %ymm2 {%k1} ## encoding: [0x62,0xf1,0x7d,0x29,0x67,0xd1]
; CHECK-NEXT: vmovdqa %ymm2, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xfd,0x6f,0xc2]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <32 x i8> @llvm.x86.avx512.mask.packuswb.256(<16 x i16> %a, <16 x i16> %b, <32 x i8> %passThru, i32 %mask)
ret <32 x i8> %res
}
define <32 x i8> @test_mask_packus_epi16_rrkz_256(<16 x i16> %a, <16 x i16> %b, i32 %mask) {
; CHECK-LABEL: test_mask_packus_epi16_rrkz_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove masked packss/packus intrinsics and autoupgrade to unmasked intrinsics with select instructions. For 512-bit add new unmasked intrinsics. The new 512-bit unmasked intrinsics will make it easy to handle these with the SSE/AVX intrinsics in InstCombine where we currently have a TODO. llvm-svn: 295290
2017-02-16 14:31:54 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
; CHECK-NEXT: vpackuswb %ymm1, %ymm0, %ymm0 {%k1} {z} ## encoding: [0x62,0xf1,0x7d,0xa9,0x67,0xc1]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <32 x i8> @llvm.x86.avx512.mask.packuswb.256(<16 x i16> %a, <16 x i16> %b, <32 x i8> zeroinitializer, i32 %mask)
ret <32 x i8> %res
}
define <32 x i8> @test_mask_packus_epi16_rm_256(<16 x i16> %a, <16 x i16>* %ptr_b) {
; CHECK-LABEL: test_mask_packus_epi16_rm_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove masked packss/packus intrinsics and autoupgrade to unmasked intrinsics with select instructions. For 512-bit add new unmasked intrinsics. The new 512-bit unmasked intrinsics will make it easy to handle these with the SSE/AVX intrinsics in InstCombine where we currently have a TODO. llvm-svn: 295290
2017-02-16 14:31:54 +08:00
; CHECK-NEXT: vpackuswb (%rdi), %ymm0, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xfd,0x67,0x07]
; CHECK-NEXT: retq ## encoding: [0xc3]
%b = load <16 x i16>, <16 x i16>* %ptr_b
%res = call <32 x i8> @llvm.x86.avx512.mask.packuswb.256(<16 x i16> %a, <16 x i16> %b, <32 x i8> zeroinitializer, i32 -1)
ret <32 x i8> %res
}
define <32 x i8> @test_mask_packus_epi16_rmk_256(<16 x i16> %a, <16 x i16>* %ptr_b, <32 x i8> %passThru, i32 %mask) {
; CHECK-LABEL: test_mask_packus_epi16_rmk_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove masked packss/packus intrinsics and autoupgrade to unmasked intrinsics with select instructions. For 512-bit add new unmasked intrinsics. The new 512-bit unmasked intrinsics will make it easy to handle these with the SSE/AVX intrinsics in InstCombine where we currently have a TODO. llvm-svn: 295290
2017-02-16 14:31:54 +08:00
; CHECK-NEXT: kmovd %esi, %k1 ## encoding: [0xc5,0xfb,0x92,0xce]
; CHECK-NEXT: vpackuswb (%rdi), %ymm0, %ymm1 {%k1} ## encoding: [0x62,0xf1,0x7d,0x29,0x67,0x0f]
; CHECK-NEXT: vmovdqa %ymm1, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xfd,0x6f,0xc1]
; CHECK-NEXT: retq ## encoding: [0xc3]
%b = load <16 x i16>, <16 x i16>* %ptr_b
%res = call <32 x i8> @llvm.x86.avx512.mask.packuswb.256(<16 x i16> %a, <16 x i16> %b, <32 x i8> %passThru, i32 %mask)
ret <32 x i8> %res
}
define <32 x i8> @test_mask_packus_epi16_rmkz_256(<16 x i16> %a, <16 x i16>* %ptr_b, i32 %mask) {
; CHECK-LABEL: test_mask_packus_epi16_rmkz_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove masked packss/packus intrinsics and autoupgrade to unmasked intrinsics with select instructions. For 512-bit add new unmasked intrinsics. The new 512-bit unmasked intrinsics will make it easy to handle these with the SSE/AVX intrinsics in InstCombine where we currently have a TODO. llvm-svn: 295290
2017-02-16 14:31:54 +08:00
; CHECK-NEXT: kmovd %esi, %k1 ## encoding: [0xc5,0xfb,0x92,0xce]
; CHECK-NEXT: vpackuswb (%rdi), %ymm0, %ymm0 {%k1} {z} ## encoding: [0x62,0xf1,0x7d,0xa9,0x67,0x07]
; CHECK-NEXT: retq ## encoding: [0xc3]
%b = load <16 x i16>, <16 x i16>* %ptr_b
%res = call <32 x i8> @llvm.x86.avx512.mask.packuswb.256(<16 x i16> %a, <16 x i16> %b, <32 x i8> zeroinitializer, i32 %mask)
ret <32 x i8> %res
}
declare <32 x i8> @llvm.x86.avx512.mask.packuswb.256(<16 x i16>, <16 x i16>, <32 x i8>, i32)
[AVX-512] Remove and autoupgrade the masked integer compare intrinsics Summary: These intrinsics aren't used by clang and haven't been for a while. There's some really terrible codegen in the 32-bit target for avx512bw due to i64 not being legal. But as I said these intrinsics aren't used by clang even before this patch so this codegen reflects our clang behavior today. Reviewers: spatel, RKSimon, zvi, igorb Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D34389 llvm-svn: 306047
2017-06-23 04:11:01 +08:00
define <8 x i32> @test_cmp_b_256(<32 x i8> %a0, <32 x i8> %a1) {
; CHECK-LABEL: test_cmp_b_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
Updated and extended the information about each instruction in HSW and SNB to include the following data: •static latency •number of uOps from which the instructions consists •all ports used by the instruction Reviewers:  RKSimon zvi aymanmus m_zuckerman Differential Revision: https://reviews.llvm.org/D33897 llvm-svn: 306414
2017-06-27 23:05:13 +08:00
; CHECK-NEXT: vpcmpeqb %ymm1, %ymm0, %k0 ## encoding: [0x62,0xf1,0x7d,0x28,0x74,0xc1]
Reverting commit 306414 on behalf of @gadi.haber llvm-svn: 306532
2017-06-28 19:23:31 +08:00
; CHECK-NEXT: kmovd %k0, %r8d ## encoding: [0xc5,0x7b,0x93,0xc0]
Updated and extended the information about each instruction in HSW and SNB to include the following data: •static latency •number of uOps from which the instructions consists •all ports used by the instruction Reviewers:  RKSimon zvi aymanmus m_zuckerman Differential Revision: https://reviews.llvm.org/D33897 llvm-svn: 306414
2017-06-27 23:05:13 +08:00
; CHECK-NEXT: vpcmpgtb %ymm0, %ymm1, %k0 ## encoding: [0x62,0xf1,0x75,0x28,0x64,0xc0]
; CHECK-NEXT: kmovd %k0, %ecx ## encoding: [0xc5,0xfb,0x93,0xc8]
; CHECK-NEXT: vpcmpleb %ymm1, %ymm0, %k0 ## encoding: [0x62,0xf3,0x7d,0x28,0x3f,0xc1,0x02]
; CHECK-NEXT: kmovd %k0, %edx ## encoding: [0xc5,0xfb,0x93,0xd0]
Reverting commit 306414 on behalf of @gadi.haber llvm-svn: 306532
2017-06-28 19:23:31 +08:00
; CHECK-NEXT: vpcmpneqb %ymm1, %ymm0, %k0 ## encoding: [0x62,0xf3,0x7d,0x28,0x3f,0xc1,0x04]
; CHECK-NEXT: kmovd %k0, %esi ## encoding: [0xc5,0xfb,0x93,0xf0]
; CHECK-NEXT: vpcmpleb %ymm0, %ymm1, %k0 ## encoding: [0x62,0xf3,0x75,0x28,0x3f,0xc0,0x02]
; CHECK-NEXT: kmovd %k0, %edi ## encoding: [0xc5,0xfb,0x93,0xf8]
; CHECK-NEXT: vpcmpgtb %ymm1, %ymm0, %k0 ## encoding: [0x62,0xf1,0x7d,0x28,0x64,0xc1]
; CHECK-NEXT: kmovd %k0, %eax ## encoding: [0xc5,0xfb,0x93,0xc0]
; CHECK-NEXT: vmovd %esi, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6e,0xc6]
; CHECK-NEXT: vpinsrd $1, %edi, %xmm0, %xmm0 ## encoding: [0xc4,0xe3,0x79,0x22,0xc7,0x01]
; CHECK-NEXT: vpinsrd $2, %eax, %xmm0, %xmm0 ## encoding: [0xc4,0xe3,0x79,0x22,0xc0,0x02]
; CHECK-NEXT: kxnord %k0, %k0, %k0 ## encoding: [0xc4,0xe1,0xfd,0x46,0xc0]
; CHECK-NEXT: kmovd %k0, %eax ## encoding: [0xc5,0xfb,0x93,0xc0]
; CHECK-NEXT: vpinsrd $3, %eax, %xmm0, %xmm0 ## encoding: [0xc4,0xe3,0x79,0x22,0xc0,0x03]
; CHECK-NEXT: vmovd %ecx, %xmm1 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6e,0xc9]
; CHECK-NEXT: vmovd %r8d, %xmm2 ## EVEX TO VEX Compression encoding: [0xc4,0xc1,0x79,0x6e,0xd0]
; CHECK-NEXT: vpunpckldq %xmm1, %xmm2, %xmm1 ## EVEX TO VEX Compression encoding: [0xc5,0xe9,0x62,0xc9]
; CHECK-NEXT: ## xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]
; CHECK-NEXT: vmovd %edx, %xmm2 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6e,0xd2]
; CHECK-NEXT: vpunpcklqdq %xmm2, %xmm1, %xmm1 ## EVEX TO VEX Compression encoding: [0xc5,0xf1,0x6c,0xca]
; CHECK-NEXT: ## xmm1 = xmm1[0],xmm2[0]
; CHECK-NEXT: vinserti128 $1, %xmm0, %ymm1, %ymm0 ## EVEX TO VEX Compression encoding: [0xc4,0xe3,0x75,0x38,0xc0,0x01]
[AVX-512] Remove and autoupgrade the masked integer compare intrinsics Summary: These intrinsics aren't used by clang and haven't been for a while. There's some really terrible codegen in the 32-bit target for avx512bw due to i64 not being legal. But as I said these intrinsics aren't used by clang even before this patch so this codegen reflects our clang behavior today. Reviewers: spatel, RKSimon, zvi, igorb Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D34389 llvm-svn: 306047
2017-06-23 04:11:01 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res0 = call i32 @llvm.x86.avx512.mask.cmp.b.256(<32 x i8> %a0, <32 x i8> %a1, i32 0, i32 -1)
%vec0 = insertelement <8 x i32> undef, i32 %res0, i32 0
%res1 = call i32 @llvm.x86.avx512.mask.cmp.b.256(<32 x i8> %a0, <32 x i8> %a1, i32 1, i32 -1)
%vec1 = insertelement <8 x i32> %vec0, i32 %res1, i32 1
%res2 = call i32 @llvm.x86.avx512.mask.cmp.b.256(<32 x i8> %a0, <32 x i8> %a1, i32 2, i32 -1)
%vec2 = insertelement <8 x i32> %vec1, i32 %res2, i32 2
%res3 = call i32 @llvm.x86.avx512.mask.cmp.b.256(<32 x i8> %a0, <32 x i8> %a1, i32 3, i32 -1)
%vec3 = insertelement <8 x i32> %vec2, i32 %res3, i32 3
%res4 = call i32 @llvm.x86.avx512.mask.cmp.b.256(<32 x i8> %a0, <32 x i8> %a1, i32 4, i32 -1)
%vec4 = insertelement <8 x i32> %vec3, i32 %res4, i32 4
%res5 = call i32 @llvm.x86.avx512.mask.cmp.b.256(<32 x i8> %a0, <32 x i8> %a1, i32 5, i32 -1)
%vec5 = insertelement <8 x i32> %vec4, i32 %res5, i32 5
%res6 = call i32 @llvm.x86.avx512.mask.cmp.b.256(<32 x i8> %a0, <32 x i8> %a1, i32 6, i32 -1)
%vec6 = insertelement <8 x i32> %vec5, i32 %res6, i32 6
%res7 = call i32 @llvm.x86.avx512.mask.cmp.b.256(<32 x i8> %a0, <32 x i8> %a1, i32 7, i32 -1)
%vec7 = insertelement <8 x i32> %vec6, i32 %res7, i32 7
ret <8 x i32> %vec7
}
define <8 x i32> @test_mask_cmp_b_256(<32 x i8> %a0, <32 x i8> %a1, i32 %mask) {
; CHECK-LABEL: test_mask_cmp_b_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove and autoupgrade the masked integer compare intrinsics Summary: These intrinsics aren't used by clang and haven't been for a while. There's some really terrible codegen in the 32-bit target for avx512bw due to i64 not being legal. But as I said these intrinsics aren't used by clang even before this patch so this codegen reflects our clang behavior today. Reviewers: spatel, RKSimon, zvi, igorb Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D34389 llvm-svn: 306047
2017-06-23 04:11:01 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
; CHECK-NEXT: vpcmpeqb %ymm1, %ymm0, %k0 {%k1} ## encoding: [0x62,0xf1,0x7d,0x29,0x74,0xc1]
; CHECK-NEXT: kmovd %k0, %r8d ## encoding: [0xc5,0x7b,0x93,0xc0]
; CHECK-NEXT: vpcmpgtb %ymm0, %ymm1, %k0 {%k1} ## encoding: [0x62,0xf1,0x75,0x29,0x64,0xc0]
; CHECK-NEXT: kmovd %k0, %r9d ## encoding: [0xc5,0x7b,0x93,0xc8]
; CHECK-NEXT: vpcmpleb %ymm1, %ymm0, %k0 {%k1} ## encoding: [0x62,0xf3,0x7d,0x29,0x3f,0xc1,0x02]
[X86][SKX][KNL] Updated regression tests to use -mattr instead of -mcpu flag.NFC. NFC. Updated 8 regression tests to use -mattr instead of -mcpu flag as follows: -mcpu=knl --> -mattr=+avx512f -mcpu=skx --> -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq The updates are as part of the preparation of a large commit to add all instruction scheduling for the SKX target. Reviewers: delena, zvi, RKSimon Differential Revision: https://reviews.llvm.org/D38222 Change-Id: I2381c9b5bb75ecacfca017243c22d054f6eddd14 llvm-svn: 314306
2017-09-27 22:44:15 +08:00
; CHECK-NEXT: kmovd %k0, %r10d ## encoding: [0xc5,0x7b,0x93,0xd0]
; CHECK-NEXT: kxord %k0, %k0, %k0 ## encoding: [0xc4,0xe1,0xfd,0x47,0xc0]
[X86][Haswell] Updating HSW instruction scheduling information This patch completely replaces the instruction scheduling information for the Haswell architecture target by modifying the file X86SchedHaswell.td located under the X86 Target. We used the scheduling information retrieved from the Haswell architects in order to replace and modify the existing scheduling. The patch continues the scheduling replacement effort started with the SNB target in r307529 and r310792. Information includes latency, number of micro-Ops and used ports by each HSW instruction. Please expect some performance fluctuations due to code alignment effects. Reviewers: RKSimon, zvi, aymanmus, craig.topper, m_zuckerman, igorb, dim, chandlerc, aaboud Differential Revision: https://reviews.llvm.org/D36663 llvm-svn: 311879
2017-08-28 18:04:16 +08:00
; CHECK-NEXT: kmovd %k0, %esi ## encoding: [0xc5,0xfb,0x93,0xf0]
[X86][SKX][KNL] Updated regression tests to use -mattr instead of -mcpu flag.NFC. NFC. Updated 8 regression tests to use -mattr instead of -mcpu flag as follows: -mcpu=knl --> -mattr=+avx512f -mcpu=skx --> -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq The updates are as part of the preparation of a large commit to add all instruction scheduling for the SKX target. Reviewers: delena, zvi, RKSimon Differential Revision: https://reviews.llvm.org/D38222 Change-Id: I2381c9b5bb75ecacfca017243c22d054f6eddd14 llvm-svn: 314306
2017-09-27 22:44:15 +08:00
; CHECK-NEXT: vpcmpneqb %ymm1, %ymm0, %k0 {%k1} ## encoding: [0x62,0xf3,0x7d,0x29,0x3f,0xc1,0x04]
[X86][Haswell] Updating HSW instruction scheduling information This patch completely replaces the instruction scheduling information for the Haswell architecture target by modifying the file X86SchedHaswell.td located under the X86 Target. We used the scheduling information retrieved from the Haswell architects in order to replace and modify the existing scheduling. The patch continues the scheduling replacement effort started with the SNB target in r307529 and r310792. Information includes latency, number of micro-Ops and used ports by each HSW instruction. Please expect some performance fluctuations due to code alignment effects. Reviewers: RKSimon, zvi, aymanmus, craig.topper, m_zuckerman, igorb, dim, chandlerc, aaboud Differential Revision: https://reviews.llvm.org/D36663 llvm-svn: 311879
2017-08-28 18:04:16 +08:00
; CHECK-NEXT: kmovd %k0, %eax ## encoding: [0xc5,0xfb,0x93,0xc0]
[X86][SKX][KNL] Updated regression tests to use -mattr instead of -mcpu flag.NFC. NFC. Updated 8 regression tests to use -mattr instead of -mcpu flag as follows: -mcpu=knl --> -mattr=+avx512f -mcpu=skx --> -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq The updates are as part of the preparation of a large commit to add all instruction scheduling for the SKX target. Reviewers: delena, zvi, RKSimon Differential Revision: https://reviews.llvm.org/D38222 Change-Id: I2381c9b5bb75ecacfca017243c22d054f6eddd14 llvm-svn: 314306
2017-09-27 22:44:15 +08:00
; CHECK-NEXT: vpcmpleb %ymm0, %ymm1, %k0 {%k1} ## encoding: [0x62,0xf3,0x75,0x29,0x3f,0xc0,0x02]
[X86][Haswell] Updating HSW instruction scheduling information This patch completely replaces the instruction scheduling information for the Haswell architecture target by modifying the file X86SchedHaswell.td located under the X86 Target. We used the scheduling information retrieved from the Haswell architects in order to replace and modify the existing scheduling. The patch continues the scheduling replacement effort started with the SNB target in r307529 and r310792. Information includes latency, number of micro-Ops and used ports by each HSW instruction. Please expect some performance fluctuations due to code alignment effects. Reviewers: RKSimon, zvi, aymanmus, craig.topper, m_zuckerman, igorb, dim, chandlerc, aaboud Differential Revision: https://reviews.llvm.org/D36663 llvm-svn: 311879
2017-08-28 18:04:16 +08:00
; CHECK-NEXT: kmovd %k0, %ecx ## encoding: [0xc5,0xfb,0x93,0xc8]
[X86][SKX][KNL] Updated regression tests to use -mattr instead of -mcpu flag.NFC. NFC. Updated 8 regression tests to use -mattr instead of -mcpu flag as follows: -mcpu=knl --> -mattr=+avx512f -mcpu=skx --> -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq The updates are as part of the preparation of a large commit to add all instruction scheduling for the SKX target. Reviewers: delena, zvi, RKSimon Differential Revision: https://reviews.llvm.org/D38222 Change-Id: I2381c9b5bb75ecacfca017243c22d054f6eddd14 llvm-svn: 314306
2017-09-27 22:44:15 +08:00
; CHECK-NEXT: vpcmpgtb %ymm1, %ymm0, %k0 {%k1} ## encoding: [0x62,0xf1,0x7d,0x29,0x64,0xc1]
; CHECK-NEXT: kmovd %k0, %edx ## encoding: [0xc5,0xfb,0x93,0xd0]
; CHECK-NEXT: vmovd %eax, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6e,0xc0]
; CHECK-NEXT: vpinsrd $1, %ecx, %xmm0, %xmm0 ## encoding: [0xc4,0xe3,0x79,0x22,0xc1,0x01]
; CHECK-NEXT: vpinsrd $2, %edx, %xmm0, %xmm0 ## encoding: [0xc4,0xe3,0x79,0x22,0xc2,0x02]
[AVX-512] Remove and autoupgrade the masked integer compare intrinsics Summary: These intrinsics aren't used by clang and haven't been for a while. There's some really terrible codegen in the 32-bit target for avx512bw due to i64 not being legal. But as I said these intrinsics aren't used by clang even before this patch so this codegen reflects our clang behavior today. Reviewers: spatel, RKSimon, zvi, igorb Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D34389 llvm-svn: 306047
2017-06-23 04:11:01 +08:00
; CHECK-NEXT: vpinsrd $3, %edi, %xmm0, %xmm0 ## encoding: [0xc4,0xe3,0x79,0x22,0xc7,0x03]
; CHECK-NEXT: vmovd %r8d, %xmm1 ## EVEX TO VEX Compression encoding: [0xc4,0xc1,0x79,0x6e,0xc8]
; CHECK-NEXT: vpinsrd $1, %r9d, %xmm1, %xmm1 ## encoding: [0xc4,0xc3,0x71,0x22,0xc9,0x01]
[X86][SKX][KNL] Updated regression tests to use -mattr instead of -mcpu flag.NFC. NFC. Updated 8 regression tests to use -mattr instead of -mcpu flag as follows: -mcpu=knl --> -mattr=+avx512f -mcpu=skx --> -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq The updates are as part of the preparation of a large commit to add all instruction scheduling for the SKX target. Reviewers: delena, zvi, RKSimon Differential Revision: https://reviews.llvm.org/D38222 Change-Id: I2381c9b5bb75ecacfca017243c22d054f6eddd14 llvm-svn: 314306
2017-09-27 22:44:15 +08:00
; CHECK-NEXT: vpinsrd $2, %r10d, %xmm1, %xmm1 ## encoding: [0xc4,0xc3,0x71,0x22,0xca,0x02]
; CHECK-NEXT: vpinsrd $3, %esi, %xmm1, %xmm1 ## encoding: [0xc4,0xe3,0x71,0x22,0xce,0x03]
[AVX-512] Remove and autoupgrade the masked integer compare intrinsics Summary: These intrinsics aren't used by clang and haven't been for a while. There's some really terrible codegen in the 32-bit target for avx512bw due to i64 not being legal. But as I said these intrinsics aren't used by clang even before this patch so this codegen reflects our clang behavior today. Reviewers: spatel, RKSimon, zvi, igorb Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D34389 llvm-svn: 306047
2017-06-23 04:11:01 +08:00
; CHECK-NEXT: vinserti128 $1, %xmm0, %ymm1, %ymm0 ## EVEX TO VEX Compression encoding: [0xc4,0xe3,0x75,0x38,0xc0,0x01]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res0 = call i32 @llvm.x86.avx512.mask.cmp.b.256(<32 x i8> %a0, <32 x i8> %a1, i32 0, i32 %mask)
%vec0 = insertelement <8 x i32> undef, i32 %res0, i32 0
%res1 = call i32 @llvm.x86.avx512.mask.cmp.b.256(<32 x i8> %a0, <32 x i8> %a1, i32 1, i32 %mask)
%vec1 = insertelement <8 x i32> %vec0, i32 %res1, i32 1
%res2 = call i32 @llvm.x86.avx512.mask.cmp.b.256(<32 x i8> %a0, <32 x i8> %a1, i32 2, i32 %mask)
%vec2 = insertelement <8 x i32> %vec1, i32 %res2, i32 2
%res3 = call i32 @llvm.x86.avx512.mask.cmp.b.256(<32 x i8> %a0, <32 x i8> %a1, i32 3, i32 %mask)
%vec3 = insertelement <8 x i32> %vec2, i32 %res3, i32 3
%res4 = call i32 @llvm.x86.avx512.mask.cmp.b.256(<32 x i8> %a0, <32 x i8> %a1, i32 4, i32 %mask)
%vec4 = insertelement <8 x i32> %vec3, i32 %res4, i32 4
%res5 = call i32 @llvm.x86.avx512.mask.cmp.b.256(<32 x i8> %a0, <32 x i8> %a1, i32 5, i32 %mask)
%vec5 = insertelement <8 x i32> %vec4, i32 %res5, i32 5
%res6 = call i32 @llvm.x86.avx512.mask.cmp.b.256(<32 x i8> %a0, <32 x i8> %a1, i32 6, i32 %mask)
%vec6 = insertelement <8 x i32> %vec5, i32 %res6, i32 6
%res7 = call i32 @llvm.x86.avx512.mask.cmp.b.256(<32 x i8> %a0, <32 x i8> %a1, i32 7, i32 %mask)
%vec7 = insertelement <8 x i32> %vec6, i32 %res7, i32 7
ret <8 x i32> %vec7
}
declare i32 @llvm.x86.avx512.mask.cmp.b.256(<32 x i8>, <32 x i8>, i32, i32) nounwind readnone
define <8 x i32> @test_ucmp_b_256(<32 x i8> %a0, <32 x i8> %a1) {
; CHECK-LABEL: test_ucmp_b_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
Updated and extended the information about each instruction in HSW and SNB to include the following data: •static latency •number of uOps from which the instructions consists •all ports used by the instruction Reviewers:  RKSimon zvi aymanmus m_zuckerman Differential Revision: https://reviews.llvm.org/D33897 llvm-svn: 306414
2017-06-27 23:05:13 +08:00
; CHECK-NEXT: vpcmpeqb %ymm1, %ymm0, %k0 ## encoding: [0x62,0xf1,0x7d,0x28,0x74,0xc1]
Reverting commit 306414 on behalf of @gadi.haber llvm-svn: 306532
2017-06-28 19:23:31 +08:00
; CHECK-NEXT: kmovd %k0, %r8d ## encoding: [0xc5,0x7b,0x93,0xc0]
Updated and extended the information about each instruction in HSW and SNB to include the following data: •static latency •number of uOps from which the instructions consists •all ports used by the instruction Reviewers:  RKSimon zvi aymanmus m_zuckerman Differential Revision: https://reviews.llvm.org/D33897 llvm-svn: 306414
2017-06-27 23:05:13 +08:00
; CHECK-NEXT: vpcmpltub %ymm1, %ymm0, %k0 ## encoding: [0x62,0xf3,0x7d,0x28,0x3e,0xc1,0x01]
; CHECK-NEXT: kmovd %k0, %ecx ## encoding: [0xc5,0xfb,0x93,0xc8]
; CHECK-NEXT: vpcmpleub %ymm1, %ymm0, %k0 ## encoding: [0x62,0xf3,0x7d,0x28,0x3e,0xc1,0x02]
; CHECK-NEXT: kmovd %k0, %edx ## encoding: [0xc5,0xfb,0x93,0xd0]
Reverting commit 306414 on behalf of @gadi.haber llvm-svn: 306532
2017-06-28 19:23:31 +08:00
; CHECK-NEXT: vpcmpneqb %ymm1, %ymm0, %k0 ## encoding: [0x62,0xf3,0x7d,0x28,0x3f,0xc1,0x04]
; CHECK-NEXT: kmovd %k0, %esi ## encoding: [0xc5,0xfb,0x93,0xf0]
; CHECK-NEXT: vpcmpnltub %ymm1, %ymm0, %k0 ## encoding: [0x62,0xf3,0x7d,0x28,0x3e,0xc1,0x05]
; CHECK-NEXT: kmovd %k0, %edi ## encoding: [0xc5,0xfb,0x93,0xf8]
; CHECK-NEXT: vpcmpnleub %ymm1, %ymm0, %k0 ## encoding: [0x62,0xf3,0x7d,0x28,0x3e,0xc1,0x06]
; CHECK-NEXT: kmovd %k0, %eax ## encoding: [0xc5,0xfb,0x93,0xc0]
; CHECK-NEXT: vmovd %esi, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6e,0xc6]
; CHECK-NEXT: vpinsrd $1, %edi, %xmm0, %xmm0 ## encoding: [0xc4,0xe3,0x79,0x22,0xc7,0x01]
; CHECK-NEXT: vpinsrd $2, %eax, %xmm0, %xmm0 ## encoding: [0xc4,0xe3,0x79,0x22,0xc0,0x02]
; CHECK-NEXT: kxnord %k0, %k0, %k0 ## encoding: [0xc4,0xe1,0xfd,0x46,0xc0]
; CHECK-NEXT: kmovd %k0, %eax ## encoding: [0xc5,0xfb,0x93,0xc0]
; CHECK-NEXT: vpinsrd $3, %eax, %xmm0, %xmm0 ## encoding: [0xc4,0xe3,0x79,0x22,0xc0,0x03]
; CHECK-NEXT: vmovd %ecx, %xmm1 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6e,0xc9]
; CHECK-NEXT: vmovd %r8d, %xmm2 ## EVEX TO VEX Compression encoding: [0xc4,0xc1,0x79,0x6e,0xd0]
; CHECK-NEXT: vpunpckldq %xmm1, %xmm2, %xmm1 ## EVEX TO VEX Compression encoding: [0xc5,0xe9,0x62,0xc9]
; CHECK-NEXT: ## xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]
; CHECK-NEXT: vmovd %edx, %xmm2 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6e,0xd2]
; CHECK-NEXT: vpunpcklqdq %xmm2, %xmm1, %xmm1 ## EVEX TO VEX Compression encoding: [0xc5,0xf1,0x6c,0xca]
; CHECK-NEXT: ## xmm1 = xmm1[0],xmm2[0]
; CHECK-NEXT: vinserti128 $1, %xmm0, %ymm1, %ymm0 ## EVEX TO VEX Compression encoding: [0xc4,0xe3,0x75,0x38,0xc0,0x01]
[AVX-512] Remove and autoupgrade the masked integer compare intrinsics Summary: These intrinsics aren't used by clang and haven't been for a while. There's some really terrible codegen in the 32-bit target for avx512bw due to i64 not being legal. But as I said these intrinsics aren't used by clang even before this patch so this codegen reflects our clang behavior today. Reviewers: spatel, RKSimon, zvi, igorb Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D34389 llvm-svn: 306047
2017-06-23 04:11:01 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res0 = call i32 @llvm.x86.avx512.mask.ucmp.b.256(<32 x i8> %a0, <32 x i8> %a1, i32 0, i32 -1)
%vec0 = insertelement <8 x i32> undef, i32 %res0, i32 0
%res1 = call i32 @llvm.x86.avx512.mask.ucmp.b.256(<32 x i8> %a0, <32 x i8> %a1, i32 1, i32 -1)
%vec1 = insertelement <8 x i32> %vec0, i32 %res1, i32 1
%res2 = call i32 @llvm.x86.avx512.mask.ucmp.b.256(<32 x i8> %a0, <32 x i8> %a1, i32 2, i32 -1)
%vec2 = insertelement <8 x i32> %vec1, i32 %res2, i32 2
%res3 = call i32 @llvm.x86.avx512.mask.ucmp.b.256(<32 x i8> %a0, <32 x i8> %a1, i32 3, i32 -1)
%vec3 = insertelement <8 x i32> %vec2, i32 %res3, i32 3
%res4 = call i32 @llvm.x86.avx512.mask.ucmp.b.256(<32 x i8> %a0, <32 x i8> %a1, i32 4, i32 -1)
%vec4 = insertelement <8 x i32> %vec3, i32 %res4, i32 4
%res5 = call i32 @llvm.x86.avx512.mask.ucmp.b.256(<32 x i8> %a0, <32 x i8> %a1, i32 5, i32 -1)
%vec5 = insertelement <8 x i32> %vec4, i32 %res5, i32 5
%res6 = call i32 @llvm.x86.avx512.mask.ucmp.b.256(<32 x i8> %a0, <32 x i8> %a1, i32 6, i32 -1)
%vec6 = insertelement <8 x i32> %vec5, i32 %res6, i32 6
%res7 = call i32 @llvm.x86.avx512.mask.ucmp.b.256(<32 x i8> %a0, <32 x i8> %a1, i32 7, i32 -1)
%vec7 = insertelement <8 x i32> %vec6, i32 %res7, i32 7
ret <8 x i32> %vec7
}
define <8 x i32> @test_mask_ucmp_b_256(<32 x i8> %a0, <32 x i8> %a1, i32 %mask) {
; CHECK-LABEL: test_mask_ucmp_b_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove and autoupgrade the masked integer compare intrinsics Summary: These intrinsics aren't used by clang and haven't been for a while. There's some really terrible codegen in the 32-bit target for avx512bw due to i64 not being legal. But as I said these intrinsics aren't used by clang even before this patch so this codegen reflects our clang behavior today. Reviewers: spatel, RKSimon, zvi, igorb Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D34389 llvm-svn: 306047
2017-06-23 04:11:01 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
; CHECK-NEXT: vpcmpeqb %ymm1, %ymm0, %k0 {%k1} ## encoding: [0x62,0xf1,0x7d,0x29,0x74,0xc1]
; CHECK-NEXT: kmovd %k0, %r8d ## encoding: [0xc5,0x7b,0x93,0xc0]
; CHECK-NEXT: vpcmpltub %ymm1, %ymm0, %k0 {%k1} ## encoding: [0x62,0xf3,0x7d,0x29,0x3e,0xc1,0x01]
; CHECK-NEXT: kmovd %k0, %r9d ## encoding: [0xc5,0x7b,0x93,0xc8]
; CHECK-NEXT: vpcmpleub %ymm1, %ymm0, %k0 {%k1} ## encoding: [0x62,0xf3,0x7d,0x29,0x3e,0xc1,0x02]
[X86][SKX][KNL] Updated regression tests to use -mattr instead of -mcpu flag.NFC. NFC. Updated 8 regression tests to use -mattr instead of -mcpu flag as follows: -mcpu=knl --> -mattr=+avx512f -mcpu=skx --> -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq The updates are as part of the preparation of a large commit to add all instruction scheduling for the SKX target. Reviewers: delena, zvi, RKSimon Differential Revision: https://reviews.llvm.org/D38222 Change-Id: I2381c9b5bb75ecacfca017243c22d054f6eddd14 llvm-svn: 314306
2017-09-27 22:44:15 +08:00
; CHECK-NEXT: kmovd %k0, %r10d ## encoding: [0xc5,0x7b,0x93,0xd0]
; CHECK-NEXT: kxord %k0, %k0, %k0 ## encoding: [0xc4,0xe1,0xfd,0x47,0xc0]
[X86][Haswell] Updating HSW instruction scheduling information This patch completely replaces the instruction scheduling information for the Haswell architecture target by modifying the file X86SchedHaswell.td located under the X86 Target. We used the scheduling information retrieved from the Haswell architects in order to replace and modify the existing scheduling. The patch continues the scheduling replacement effort started with the SNB target in r307529 and r310792. Information includes latency, number of micro-Ops and used ports by each HSW instruction. Please expect some performance fluctuations due to code alignment effects. Reviewers: RKSimon, zvi, aymanmus, craig.topper, m_zuckerman, igorb, dim, chandlerc, aaboud Differential Revision: https://reviews.llvm.org/D36663 llvm-svn: 311879
2017-08-28 18:04:16 +08:00
; CHECK-NEXT: kmovd %k0, %esi ## encoding: [0xc5,0xfb,0x93,0xf0]
[X86][SKX][KNL] Updated regression tests to use -mattr instead of -mcpu flag.NFC. NFC. Updated 8 regression tests to use -mattr instead of -mcpu flag as follows: -mcpu=knl --> -mattr=+avx512f -mcpu=skx --> -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq The updates are as part of the preparation of a large commit to add all instruction scheduling for the SKX target. Reviewers: delena, zvi, RKSimon Differential Revision: https://reviews.llvm.org/D38222 Change-Id: I2381c9b5bb75ecacfca017243c22d054f6eddd14 llvm-svn: 314306
2017-09-27 22:44:15 +08:00
; CHECK-NEXT: vpcmpneqb %ymm1, %ymm0, %k0 {%k1} ## encoding: [0x62,0xf3,0x7d,0x29,0x3f,0xc1,0x04]
[X86][Haswell] Updating HSW instruction scheduling information This patch completely replaces the instruction scheduling information for the Haswell architecture target by modifying the file X86SchedHaswell.td located under the X86 Target. We used the scheduling information retrieved from the Haswell architects in order to replace and modify the existing scheduling. The patch continues the scheduling replacement effort started with the SNB target in r307529 and r310792. Information includes latency, number of micro-Ops and used ports by each HSW instruction. Please expect some performance fluctuations due to code alignment effects. Reviewers: RKSimon, zvi, aymanmus, craig.topper, m_zuckerman, igorb, dim, chandlerc, aaboud Differential Revision: https://reviews.llvm.org/D36663 llvm-svn: 311879
2017-08-28 18:04:16 +08:00
; CHECK-NEXT: kmovd %k0, %eax ## encoding: [0xc5,0xfb,0x93,0xc0]
[X86][SKX][KNL] Updated regression tests to use -mattr instead of -mcpu flag.NFC. NFC. Updated 8 regression tests to use -mattr instead of -mcpu flag as follows: -mcpu=knl --> -mattr=+avx512f -mcpu=skx --> -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq The updates are as part of the preparation of a large commit to add all instruction scheduling for the SKX target. Reviewers: delena, zvi, RKSimon Differential Revision: https://reviews.llvm.org/D38222 Change-Id: I2381c9b5bb75ecacfca017243c22d054f6eddd14 llvm-svn: 314306
2017-09-27 22:44:15 +08:00
; CHECK-NEXT: vpcmpnltub %ymm1, %ymm0, %k0 {%k1} ## encoding: [0x62,0xf3,0x7d,0x29,0x3e,0xc1,0x05]
[X86][Haswell] Updating HSW instruction scheduling information This patch completely replaces the instruction scheduling information for the Haswell architecture target by modifying the file X86SchedHaswell.td located under the X86 Target. We used the scheduling information retrieved from the Haswell architects in order to replace and modify the existing scheduling. The patch continues the scheduling replacement effort started with the SNB target in r307529 and r310792. Information includes latency, number of micro-Ops and used ports by each HSW instruction. Please expect some performance fluctuations due to code alignment effects. Reviewers: RKSimon, zvi, aymanmus, craig.topper, m_zuckerman, igorb, dim, chandlerc, aaboud Differential Revision: https://reviews.llvm.org/D36663 llvm-svn: 311879
2017-08-28 18:04:16 +08:00
; CHECK-NEXT: kmovd %k0, %ecx ## encoding: [0xc5,0xfb,0x93,0xc8]
[X86][SKX][KNL] Updated regression tests to use -mattr instead of -mcpu flag.NFC. NFC. Updated 8 regression tests to use -mattr instead of -mcpu flag as follows: -mcpu=knl --> -mattr=+avx512f -mcpu=skx --> -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq The updates are as part of the preparation of a large commit to add all instruction scheduling for the SKX target. Reviewers: delena, zvi, RKSimon Differential Revision: https://reviews.llvm.org/D38222 Change-Id: I2381c9b5bb75ecacfca017243c22d054f6eddd14 llvm-svn: 314306
2017-09-27 22:44:15 +08:00
; CHECK-NEXT: vpcmpnleub %ymm1, %ymm0, %k0 {%k1} ## encoding: [0x62,0xf3,0x7d,0x29,0x3e,0xc1,0x06]
; CHECK-NEXT: kmovd %k0, %edx ## encoding: [0xc5,0xfb,0x93,0xd0]
; CHECK-NEXT: vmovd %eax, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6e,0xc0]
; CHECK-NEXT: vpinsrd $1, %ecx, %xmm0, %xmm0 ## encoding: [0xc4,0xe3,0x79,0x22,0xc1,0x01]
; CHECK-NEXT: vpinsrd $2, %edx, %xmm0, %xmm0 ## encoding: [0xc4,0xe3,0x79,0x22,0xc2,0x02]
[AVX-512] Remove and autoupgrade the masked integer compare intrinsics Summary: These intrinsics aren't used by clang and haven't been for a while. There's some really terrible codegen in the 32-bit target for avx512bw due to i64 not being legal. But as I said these intrinsics aren't used by clang even before this patch so this codegen reflects our clang behavior today. Reviewers: spatel, RKSimon, zvi, igorb Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D34389 llvm-svn: 306047
2017-06-23 04:11:01 +08:00
; CHECK-NEXT: vpinsrd $3, %edi, %xmm0, %xmm0 ## encoding: [0xc4,0xe3,0x79,0x22,0xc7,0x03]
; CHECK-NEXT: vmovd %r8d, %xmm1 ## EVEX TO VEX Compression encoding: [0xc4,0xc1,0x79,0x6e,0xc8]
; CHECK-NEXT: vpinsrd $1, %r9d, %xmm1, %xmm1 ## encoding: [0xc4,0xc3,0x71,0x22,0xc9,0x01]
[X86][SKX][KNL] Updated regression tests to use -mattr instead of -mcpu flag.NFC. NFC. Updated 8 regression tests to use -mattr instead of -mcpu flag as follows: -mcpu=knl --> -mattr=+avx512f -mcpu=skx --> -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq The updates are as part of the preparation of a large commit to add all instruction scheduling for the SKX target. Reviewers: delena, zvi, RKSimon Differential Revision: https://reviews.llvm.org/D38222 Change-Id: I2381c9b5bb75ecacfca017243c22d054f6eddd14 llvm-svn: 314306
2017-09-27 22:44:15 +08:00
; CHECK-NEXT: vpinsrd $2, %r10d, %xmm1, %xmm1 ## encoding: [0xc4,0xc3,0x71,0x22,0xca,0x02]
; CHECK-NEXT: vpinsrd $3, %esi, %xmm1, %xmm1 ## encoding: [0xc4,0xe3,0x71,0x22,0xce,0x03]
[AVX-512] Remove and autoupgrade the masked integer compare intrinsics Summary: These intrinsics aren't used by clang and haven't been for a while. There's some really terrible codegen in the 32-bit target for avx512bw due to i64 not being legal. But as I said these intrinsics aren't used by clang even before this patch so this codegen reflects our clang behavior today. Reviewers: spatel, RKSimon, zvi, igorb Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D34389 llvm-svn: 306047
2017-06-23 04:11:01 +08:00
; CHECK-NEXT: vinserti128 $1, %xmm0, %ymm1, %ymm0 ## EVEX TO VEX Compression encoding: [0xc4,0xe3,0x75,0x38,0xc0,0x01]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res0 = call i32 @llvm.x86.avx512.mask.ucmp.b.256(<32 x i8> %a0, <32 x i8> %a1, i32 0, i32 %mask)
%vec0 = insertelement <8 x i32> undef, i32 %res0, i32 0
%res1 = call i32 @llvm.x86.avx512.mask.ucmp.b.256(<32 x i8> %a0, <32 x i8> %a1, i32 1, i32 %mask)
%vec1 = insertelement <8 x i32> %vec0, i32 %res1, i32 1
%res2 = call i32 @llvm.x86.avx512.mask.ucmp.b.256(<32 x i8> %a0, <32 x i8> %a1, i32 2, i32 %mask)
%vec2 = insertelement <8 x i32> %vec1, i32 %res2, i32 2
%res3 = call i32 @llvm.x86.avx512.mask.ucmp.b.256(<32 x i8> %a0, <32 x i8> %a1, i32 3, i32 %mask)
%vec3 = insertelement <8 x i32> %vec2, i32 %res3, i32 3
%res4 = call i32 @llvm.x86.avx512.mask.ucmp.b.256(<32 x i8> %a0, <32 x i8> %a1, i32 4, i32 %mask)
%vec4 = insertelement <8 x i32> %vec3, i32 %res4, i32 4
%res5 = call i32 @llvm.x86.avx512.mask.ucmp.b.256(<32 x i8> %a0, <32 x i8> %a1, i32 5, i32 %mask)
%vec5 = insertelement <8 x i32> %vec4, i32 %res5, i32 5
%res6 = call i32 @llvm.x86.avx512.mask.ucmp.b.256(<32 x i8> %a0, <32 x i8> %a1, i32 6, i32 %mask)
%vec6 = insertelement <8 x i32> %vec5, i32 %res6, i32 6
%res7 = call i32 @llvm.x86.avx512.mask.ucmp.b.256(<32 x i8> %a0, <32 x i8> %a1, i32 7, i32 %mask)
%vec7 = insertelement <8 x i32> %vec6, i32 %res7, i32 7
ret <8 x i32> %vec7
}
declare i32 @llvm.x86.avx512.mask.ucmp.b.256(<32 x i8>, <32 x i8>, i32, i32) nounwind readnone
define <8 x i16> @test_cmp_w_256(<16 x i16> %a0, <16 x i16> %a1) {
; CHECK-LABEL: test_cmp_w_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove and autoupgrade the masked integer compare intrinsics Summary: These intrinsics aren't used by clang and haven't been for a while. There's some really terrible codegen in the 32-bit target for avx512bw due to i64 not being legal. But as I said these intrinsics aren't used by clang even before this patch so this codegen reflects our clang behavior today. Reviewers: spatel, RKSimon, zvi, igorb Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D34389 llvm-svn: 306047
2017-06-23 04:11:01 +08:00
; CHECK-NEXT: vpcmpeqw %ymm1, %ymm0, %k0 ## encoding: [0x62,0xf1,0x7d,0x28,0x75,0xc1]
; CHECK-NEXT: vpcmpgtw %ymm0, %ymm1, %k1 ## encoding: [0x62,0xf1,0x75,0x28,0x65,0xc8]
; CHECK-NEXT: vpcmplew %ymm1, %ymm0, %k2 ## encoding: [0x62,0xf3,0xfd,0x28,0x3f,0xd1,0x02]
; CHECK-NEXT: vpcmpneqw %ymm1, %ymm0, %k3 ## encoding: [0x62,0xf3,0xfd,0x28,0x3f,0xd9,0x04]
; CHECK-NEXT: vpcmplew %ymm0, %ymm1, %k4 ## encoding: [0x62,0xf3,0xf5,0x28,0x3f,0xe0,0x02]
; CHECK-NEXT: vpcmpgtw %ymm1, %ymm0, %k5 ## encoding: [0x62,0xf1,0x7d,0x28,0x65,0xe9]
; CHECK-NEXT: kmovd %k0, %eax ## encoding: [0xc5,0xfb,0x93,0xc0]
; CHECK-NEXT: vpxor %xmm0, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xef,0xc0]
; CHECK-NEXT: vpinsrw $0, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x00]
; CHECK-NEXT: kmovd %k1, %eax ## encoding: [0xc5,0xfb,0x93,0xc1]
; CHECK-NEXT: vpinsrw $1, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x01]
; CHECK-NEXT: kmovd %k2, %eax ## encoding: [0xc5,0xfb,0x93,0xc2]
; CHECK-NEXT: vpinsrw $2, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x02]
; CHECK-NEXT: kmovd %k3, %eax ## encoding: [0xc5,0xfb,0x93,0xc3]
; CHECK-NEXT: vpinsrw $4, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x04]
; CHECK-NEXT: kmovd %k4, %eax ## encoding: [0xc5,0xfb,0x93,0xc4]
; CHECK-NEXT: vpinsrw $5, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x05]
; CHECK-NEXT: kmovd %k5, %eax ## encoding: [0xc5,0xfb,0x93,0xc5]
; CHECK-NEXT: vpinsrw $6, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x06]
; CHECK-NEXT: kxnorw %k0, %k0, %k0 ## encoding: [0xc5,0xfc,0x46,0xc0]
; CHECK-NEXT: kmovd %k0, %eax ## encoding: [0xc5,0xfb,0x93,0xc0]
; CHECK-NEXT: vpinsrw $7, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x07]
[X86][SKX][KNL] Updated regression tests to use -mattr instead of -mcpu flag.NFC. NFC. Updated 8 regression tests to use -mattr instead of -mcpu flag as follows: -mcpu=knl --> -mattr=+avx512f -mcpu=skx --> -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq The updates are as part of the preparation of a large commit to add all instruction scheduling for the SKX target. Reviewers: delena, zvi, RKSimon Differential Revision: https://reviews.llvm.org/D38222 Change-Id: I2381c9b5bb75ecacfca017243c22d054f6eddd14 llvm-svn: 314306
2017-09-27 22:44:15 +08:00
; CHECK-NEXT: vzeroupper ## encoding: [0xc5,0xf8,0x77]
[AVX-512] Remove and autoupgrade the masked integer compare intrinsics Summary: These intrinsics aren't used by clang and haven't been for a while. There's some really terrible codegen in the 32-bit target for avx512bw due to i64 not being legal. But as I said these intrinsics aren't used by clang even before this patch so this codegen reflects our clang behavior today. Reviewers: spatel, RKSimon, zvi, igorb Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D34389 llvm-svn: 306047
2017-06-23 04:11:01 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res0 = call i16 @llvm.x86.avx512.mask.cmp.w.256(<16 x i16> %a0, <16 x i16> %a1, i32 0, i16 -1)
%vec0 = insertelement <8 x i16> undef, i16 %res0, i32 0
%res1 = call i16 @llvm.x86.avx512.mask.cmp.w.256(<16 x i16> %a0, <16 x i16> %a1, i32 1, i16 -1)
%vec1 = insertelement <8 x i16> %vec0, i16 %res1, i32 1
%res2 = call i16 @llvm.x86.avx512.mask.cmp.w.256(<16 x i16> %a0, <16 x i16> %a1, i32 2, i16 -1)
%vec2 = insertelement <8 x i16> %vec1, i16 %res2, i32 2
%res3 = call i16 @llvm.x86.avx512.mask.cmp.w.256(<16 x i16> %a0, <16 x i16> %a1, i32 3, i16 -1)
%vec3 = insertelement <8 x i16> %vec2, i16 %res3, i32 3
%res4 = call i16 @llvm.x86.avx512.mask.cmp.w.256(<16 x i16> %a0, <16 x i16> %a1, i32 4, i16 -1)
%vec4 = insertelement <8 x i16> %vec3, i16 %res4, i32 4
%res5 = call i16 @llvm.x86.avx512.mask.cmp.w.256(<16 x i16> %a0, <16 x i16> %a1, i32 5, i16 -1)
%vec5 = insertelement <8 x i16> %vec4, i16 %res5, i32 5
%res6 = call i16 @llvm.x86.avx512.mask.cmp.w.256(<16 x i16> %a0, <16 x i16> %a1, i32 6, i16 -1)
%vec6 = insertelement <8 x i16> %vec5, i16 %res6, i32 6
%res7 = call i16 @llvm.x86.avx512.mask.cmp.w.256(<16 x i16> %a0, <16 x i16> %a1, i32 7, i16 -1)
%vec7 = insertelement <8 x i16> %vec6, i16 %res7, i32 7
ret <8 x i16> %vec7
}
define <8 x i16> @test_mask_cmp_w_256(<16 x i16> %a0, <16 x i16> %a1, i16 %mask) {
; CHECK-LABEL: test_mask_cmp_w_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove and autoupgrade the masked integer compare intrinsics Summary: These intrinsics aren't used by clang and haven't been for a while. There's some really terrible codegen in the 32-bit target for avx512bw due to i64 not being legal. But as I said these intrinsics aren't used by clang even before this patch so this codegen reflects our clang behavior today. Reviewers: spatel, RKSimon, zvi, igorb Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D34389 llvm-svn: 306047
2017-06-23 04:11:01 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
; CHECK-NEXT: vpcmpeqw %ymm1, %ymm0, %k0 {%k1} ## encoding: [0x62,0xf1,0x7d,0x29,0x75,0xc1]
; CHECK-NEXT: vpcmpgtw %ymm0, %ymm1, %k2 {%k1} ## encoding: [0x62,0xf1,0x75,0x29,0x65,0xd0]
; CHECK-NEXT: vpcmplew %ymm1, %ymm0, %k3 {%k1} ## encoding: [0x62,0xf3,0xfd,0x29,0x3f,0xd9,0x02]
; CHECK-NEXT: kxorw %k0, %k0, %k4 ## encoding: [0xc5,0xfc,0x47,0xe0]
; CHECK-NEXT: vpcmpneqw %ymm1, %ymm0, %k5 {%k1} ## encoding: [0x62,0xf3,0xfd,0x29,0x3f,0xe9,0x04]
; CHECK-NEXT: vpcmplew %ymm0, %ymm1, %k6 {%k1} ## encoding: [0x62,0xf3,0xf5,0x29,0x3f,0xf0,0x02]
; CHECK-NEXT: vpcmpgtw %ymm1, %ymm0, %k1 {%k1} ## encoding: [0x62,0xf1,0x7d,0x29,0x65,0xc9]
; CHECK-NEXT: kmovd %k2, %eax ## encoding: [0xc5,0xfb,0x93,0xc2]
; CHECK-NEXT: kmovd %k0, %ecx ## encoding: [0xc5,0xfb,0x93,0xc8]
; CHECK-NEXT: vmovd %ecx, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6e,0xc1]
; CHECK-NEXT: vpinsrw $1, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x01]
; CHECK-NEXT: kmovd %k3, %eax ## encoding: [0xc5,0xfb,0x93,0xc3]
; CHECK-NEXT: vpinsrw $2, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x02]
; CHECK-NEXT: kmovd %k4, %eax ## encoding: [0xc5,0xfb,0x93,0xc4]
; CHECK-NEXT: vpinsrw $3, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x03]
; CHECK-NEXT: kmovd %k5, %eax ## encoding: [0xc5,0xfb,0x93,0xc5]
; CHECK-NEXT: vpinsrw $4, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x04]
; CHECK-NEXT: kmovd %k6, %eax ## encoding: [0xc5,0xfb,0x93,0xc6]
; CHECK-NEXT: vpinsrw $5, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x05]
; CHECK-NEXT: kmovd %k1, %eax ## encoding: [0xc5,0xfb,0x93,0xc1]
; CHECK-NEXT: vpinsrw $6, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x06]
; CHECK-NEXT: vpinsrw $7, %edi, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc7,0x07]
[X86][SKX][KNL] Updated regression tests to use -mattr instead of -mcpu flag.NFC. NFC. Updated 8 regression tests to use -mattr instead of -mcpu flag as follows: -mcpu=knl --> -mattr=+avx512f -mcpu=skx --> -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq The updates are as part of the preparation of a large commit to add all instruction scheduling for the SKX target. Reviewers: delena, zvi, RKSimon Differential Revision: https://reviews.llvm.org/D38222 Change-Id: I2381c9b5bb75ecacfca017243c22d054f6eddd14 llvm-svn: 314306
2017-09-27 22:44:15 +08:00
; CHECK-NEXT: vzeroupper ## encoding: [0xc5,0xf8,0x77]
[AVX-512] Remove and autoupgrade the masked integer compare intrinsics Summary: These intrinsics aren't used by clang and haven't been for a while. There's some really terrible codegen in the 32-bit target for avx512bw due to i64 not being legal. But as I said these intrinsics aren't used by clang even before this patch so this codegen reflects our clang behavior today. Reviewers: spatel, RKSimon, zvi, igorb Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D34389 llvm-svn: 306047
2017-06-23 04:11:01 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res0 = call i16 @llvm.x86.avx512.mask.cmp.w.256(<16 x i16> %a0, <16 x i16> %a1, i32 0, i16 %mask)
%vec0 = insertelement <8 x i16> undef, i16 %res0, i32 0
%res1 = call i16 @llvm.x86.avx512.mask.cmp.w.256(<16 x i16> %a0, <16 x i16> %a1, i32 1, i16 %mask)
%vec1 = insertelement <8 x i16> %vec0, i16 %res1, i32 1
%res2 = call i16 @llvm.x86.avx512.mask.cmp.w.256(<16 x i16> %a0, <16 x i16> %a1, i32 2, i16 %mask)
%vec2 = insertelement <8 x i16> %vec1, i16 %res2, i32 2
%res3 = call i16 @llvm.x86.avx512.mask.cmp.w.256(<16 x i16> %a0, <16 x i16> %a1, i32 3, i16 %mask)
%vec3 = insertelement <8 x i16> %vec2, i16 %res3, i32 3
%res4 = call i16 @llvm.x86.avx512.mask.cmp.w.256(<16 x i16> %a0, <16 x i16> %a1, i32 4, i16 %mask)
%vec4 = insertelement <8 x i16> %vec3, i16 %res4, i32 4
%res5 = call i16 @llvm.x86.avx512.mask.cmp.w.256(<16 x i16> %a0, <16 x i16> %a1, i32 5, i16 %mask)
%vec5 = insertelement <8 x i16> %vec4, i16 %res5, i32 5
%res6 = call i16 @llvm.x86.avx512.mask.cmp.w.256(<16 x i16> %a0, <16 x i16> %a1, i32 6, i16 %mask)
%vec6 = insertelement <8 x i16> %vec5, i16 %res6, i32 6
%res7 = call i16 @llvm.x86.avx512.mask.cmp.w.256(<16 x i16> %a0, <16 x i16> %a1, i32 7, i16 %mask)
%vec7 = insertelement <8 x i16> %vec6, i16 %res7, i32 7
ret <8 x i16> %vec7
}
declare i16 @llvm.x86.avx512.mask.cmp.w.256(<16 x i16>, <16 x i16>, i32, i16) nounwind readnone
define <8 x i16> @test_ucmp_w_256(<16 x i16> %a0, <16 x i16> %a1) {
; CHECK-LABEL: test_ucmp_w_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove and autoupgrade the masked integer compare intrinsics Summary: These intrinsics aren't used by clang and haven't been for a while. There's some really terrible codegen in the 32-bit target for avx512bw due to i64 not being legal. But as I said these intrinsics aren't used by clang even before this patch so this codegen reflects our clang behavior today. Reviewers: spatel, RKSimon, zvi, igorb Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D34389 llvm-svn: 306047
2017-06-23 04:11:01 +08:00
; CHECK-NEXT: vpcmpeqw %ymm1, %ymm0, %k0 ## encoding: [0x62,0xf1,0x7d,0x28,0x75,0xc1]
; CHECK-NEXT: vpcmpltuw %ymm1, %ymm0, %k1 ## encoding: [0x62,0xf3,0xfd,0x28,0x3e,0xc9,0x01]
; CHECK-NEXT: vpcmpleuw %ymm1, %ymm0, %k2 ## encoding: [0x62,0xf3,0xfd,0x28,0x3e,0xd1,0x02]
; CHECK-NEXT: vpcmpneqw %ymm1, %ymm0, %k3 ## encoding: [0x62,0xf3,0xfd,0x28,0x3f,0xd9,0x04]
; CHECK-NEXT: vpcmpnltuw %ymm1, %ymm0, %k4 ## encoding: [0x62,0xf3,0xfd,0x28,0x3e,0xe1,0x05]
; CHECK-NEXT: vpcmpnleuw %ymm1, %ymm0, %k5 ## encoding: [0x62,0xf3,0xfd,0x28,0x3e,0xe9,0x06]
; CHECK-NEXT: kmovd %k0, %eax ## encoding: [0xc5,0xfb,0x93,0xc0]
; CHECK-NEXT: vpxor %xmm0, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xef,0xc0]
; CHECK-NEXT: vpinsrw $0, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x00]
; CHECK-NEXT: kmovd %k1, %eax ## encoding: [0xc5,0xfb,0x93,0xc1]
; CHECK-NEXT: vpinsrw $1, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x01]
; CHECK-NEXT: kmovd %k2, %eax ## encoding: [0xc5,0xfb,0x93,0xc2]
; CHECK-NEXT: vpinsrw $2, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x02]
; CHECK-NEXT: kmovd %k3, %eax ## encoding: [0xc5,0xfb,0x93,0xc3]
; CHECK-NEXT: vpinsrw $4, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x04]
; CHECK-NEXT: kmovd %k4, %eax ## encoding: [0xc5,0xfb,0x93,0xc4]
; CHECK-NEXT: vpinsrw $5, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x05]
; CHECK-NEXT: kmovd %k5, %eax ## encoding: [0xc5,0xfb,0x93,0xc5]
; CHECK-NEXT: vpinsrw $6, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x06]
; CHECK-NEXT: kxnorw %k0, %k0, %k0 ## encoding: [0xc5,0xfc,0x46,0xc0]
; CHECK-NEXT: kmovd %k0, %eax ## encoding: [0xc5,0xfb,0x93,0xc0]
; CHECK-NEXT: vpinsrw $7, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x07]
[X86][SKX][KNL] Updated regression tests to use -mattr instead of -mcpu flag.NFC. NFC. Updated 8 regression tests to use -mattr instead of -mcpu flag as follows: -mcpu=knl --> -mattr=+avx512f -mcpu=skx --> -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq The updates are as part of the preparation of a large commit to add all instruction scheduling for the SKX target. Reviewers: delena, zvi, RKSimon Differential Revision: https://reviews.llvm.org/D38222 Change-Id: I2381c9b5bb75ecacfca017243c22d054f6eddd14 llvm-svn: 314306
2017-09-27 22:44:15 +08:00
; CHECK-NEXT: vzeroupper ## encoding: [0xc5,0xf8,0x77]
[AVX-512] Remove and autoupgrade the masked integer compare intrinsics Summary: These intrinsics aren't used by clang and haven't been for a while. There's some really terrible codegen in the 32-bit target for avx512bw due to i64 not being legal. But as I said these intrinsics aren't used by clang even before this patch so this codegen reflects our clang behavior today. Reviewers: spatel, RKSimon, zvi, igorb Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D34389 llvm-svn: 306047
2017-06-23 04:11:01 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res0 = call i16 @llvm.x86.avx512.mask.ucmp.w.256(<16 x i16> %a0, <16 x i16> %a1, i32 0, i16 -1)
%vec0 = insertelement <8 x i16> undef, i16 %res0, i32 0
%res1 = call i16 @llvm.x86.avx512.mask.ucmp.w.256(<16 x i16> %a0, <16 x i16> %a1, i32 1, i16 -1)
%vec1 = insertelement <8 x i16> %vec0, i16 %res1, i32 1
%res2 = call i16 @llvm.x86.avx512.mask.ucmp.w.256(<16 x i16> %a0, <16 x i16> %a1, i32 2, i16 -1)
%vec2 = insertelement <8 x i16> %vec1, i16 %res2, i32 2
%res3 = call i16 @llvm.x86.avx512.mask.ucmp.w.256(<16 x i16> %a0, <16 x i16> %a1, i32 3, i16 -1)
%vec3 = insertelement <8 x i16> %vec2, i16 %res3, i32 3
%res4 = call i16 @llvm.x86.avx512.mask.ucmp.w.256(<16 x i16> %a0, <16 x i16> %a1, i32 4, i16 -1)
%vec4 = insertelement <8 x i16> %vec3, i16 %res4, i32 4
%res5 = call i16 @llvm.x86.avx512.mask.ucmp.w.256(<16 x i16> %a0, <16 x i16> %a1, i32 5, i16 -1)
%vec5 = insertelement <8 x i16> %vec4, i16 %res5, i32 5
%res6 = call i16 @llvm.x86.avx512.mask.ucmp.w.256(<16 x i16> %a0, <16 x i16> %a1, i32 6, i16 -1)
%vec6 = insertelement <8 x i16> %vec5, i16 %res6, i32 6
%res7 = call i16 @llvm.x86.avx512.mask.ucmp.w.256(<16 x i16> %a0, <16 x i16> %a1, i32 7, i16 -1)
%vec7 = insertelement <8 x i16> %vec6, i16 %res7, i32 7
ret <8 x i16> %vec7
}
define <8 x i16> @test_mask_ucmp_w_256(<16 x i16> %a0, <16 x i16> %a1, i16 %mask) {
; CHECK-LABEL: test_mask_ucmp_w_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove and autoupgrade the masked integer compare intrinsics Summary: These intrinsics aren't used by clang and haven't been for a while. There's some really terrible codegen in the 32-bit target for avx512bw due to i64 not being legal. But as I said these intrinsics aren't used by clang even before this patch so this codegen reflects our clang behavior today. Reviewers: spatel, RKSimon, zvi, igorb Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D34389 llvm-svn: 306047
2017-06-23 04:11:01 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
; CHECK-NEXT: vpcmpeqw %ymm1, %ymm0, %k0 {%k1} ## encoding: [0x62,0xf1,0x7d,0x29,0x75,0xc1]
; CHECK-NEXT: vpcmpltuw %ymm1, %ymm0, %k2 {%k1} ## encoding: [0x62,0xf3,0xfd,0x29,0x3e,0xd1,0x01]
; CHECK-NEXT: vpcmpleuw %ymm1, %ymm0, %k3 {%k1} ## encoding: [0x62,0xf3,0xfd,0x29,0x3e,0xd9,0x02]
; CHECK-NEXT: kxorw %k0, %k0, %k4 ## encoding: [0xc5,0xfc,0x47,0xe0]
; CHECK-NEXT: vpcmpneqw %ymm1, %ymm0, %k5 {%k1} ## encoding: [0x62,0xf3,0xfd,0x29,0x3f,0xe9,0x04]
; CHECK-NEXT: vpcmpnltuw %ymm1, %ymm0, %k6 {%k1} ## encoding: [0x62,0xf3,0xfd,0x29,0x3e,0xf1,0x05]
; CHECK-NEXT: vpcmpnleuw %ymm1, %ymm0, %k1 {%k1} ## encoding: [0x62,0xf3,0xfd,0x29,0x3e,0xc9,0x06]
; CHECK-NEXT: kmovd %k2, %eax ## encoding: [0xc5,0xfb,0x93,0xc2]
; CHECK-NEXT: kmovd %k0, %ecx ## encoding: [0xc5,0xfb,0x93,0xc8]
; CHECK-NEXT: vmovd %ecx, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6e,0xc1]
; CHECK-NEXT: vpinsrw $1, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x01]
; CHECK-NEXT: kmovd %k3, %eax ## encoding: [0xc5,0xfb,0x93,0xc3]
; CHECK-NEXT: vpinsrw $2, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x02]
; CHECK-NEXT: kmovd %k4, %eax ## encoding: [0xc5,0xfb,0x93,0xc4]
; CHECK-NEXT: vpinsrw $3, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x03]
; CHECK-NEXT: kmovd %k5, %eax ## encoding: [0xc5,0xfb,0x93,0xc5]
; CHECK-NEXT: vpinsrw $4, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x04]
; CHECK-NEXT: kmovd %k6, %eax ## encoding: [0xc5,0xfb,0x93,0xc6]
; CHECK-NEXT: vpinsrw $5, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x05]
; CHECK-NEXT: kmovd %k1, %eax ## encoding: [0xc5,0xfb,0x93,0xc1]
; CHECK-NEXT: vpinsrw $6, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x06]
; CHECK-NEXT: vpinsrw $7, %edi, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc7,0x07]
[X86][SKX][KNL] Updated regression tests to use -mattr instead of -mcpu flag.NFC. NFC. Updated 8 regression tests to use -mattr instead of -mcpu flag as follows: -mcpu=knl --> -mattr=+avx512f -mcpu=skx --> -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq The updates are as part of the preparation of a large commit to add all instruction scheduling for the SKX target. Reviewers: delena, zvi, RKSimon Differential Revision: https://reviews.llvm.org/D38222 Change-Id: I2381c9b5bb75ecacfca017243c22d054f6eddd14 llvm-svn: 314306
2017-09-27 22:44:15 +08:00
; CHECK-NEXT: vzeroupper ## encoding: [0xc5,0xf8,0x77]
[AVX-512] Remove and autoupgrade the masked integer compare intrinsics Summary: These intrinsics aren't used by clang and haven't been for a while. There's some really terrible codegen in the 32-bit target for avx512bw due to i64 not being legal. But as I said these intrinsics aren't used by clang even before this patch so this codegen reflects our clang behavior today. Reviewers: spatel, RKSimon, zvi, igorb Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D34389 llvm-svn: 306047
2017-06-23 04:11:01 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res0 = call i16 @llvm.x86.avx512.mask.ucmp.w.256(<16 x i16> %a0, <16 x i16> %a1, i32 0, i16 %mask)
%vec0 = insertelement <8 x i16> undef, i16 %res0, i32 0
%res1 = call i16 @llvm.x86.avx512.mask.ucmp.w.256(<16 x i16> %a0, <16 x i16> %a1, i32 1, i16 %mask)
%vec1 = insertelement <8 x i16> %vec0, i16 %res1, i32 1
%res2 = call i16 @llvm.x86.avx512.mask.ucmp.w.256(<16 x i16> %a0, <16 x i16> %a1, i32 2, i16 %mask)
%vec2 = insertelement <8 x i16> %vec1, i16 %res2, i32 2
%res3 = call i16 @llvm.x86.avx512.mask.ucmp.w.256(<16 x i16> %a0, <16 x i16> %a1, i32 3, i16 %mask)
%vec3 = insertelement <8 x i16> %vec2, i16 %res3, i32 3
%res4 = call i16 @llvm.x86.avx512.mask.ucmp.w.256(<16 x i16> %a0, <16 x i16> %a1, i32 4, i16 %mask)
%vec4 = insertelement <8 x i16> %vec3, i16 %res4, i32 4
%res5 = call i16 @llvm.x86.avx512.mask.ucmp.w.256(<16 x i16> %a0, <16 x i16> %a1, i32 5, i16 %mask)
%vec5 = insertelement <8 x i16> %vec4, i16 %res5, i32 5
%res6 = call i16 @llvm.x86.avx512.mask.ucmp.w.256(<16 x i16> %a0, <16 x i16> %a1, i32 6, i16 %mask)
%vec6 = insertelement <8 x i16> %vec5, i16 %res6, i32 6
%res7 = call i16 @llvm.x86.avx512.mask.ucmp.w.256(<16 x i16> %a0, <16 x i16> %a1, i32 7, i16 %mask)
%vec7 = insertelement <8 x i16> %vec6, i16 %res7, i32 7
ret <8 x i16> %vec7
}
declare i16 @llvm.x86.avx512.mask.ucmp.w.256(<16 x i16>, <16 x i16>, i32, i16) nounwind readnone
define <8 x i16> @test_cmp_b_128(<16 x i8> %a0, <16 x i8> %a1) {
; CHECK-LABEL: test_cmp_b_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove and autoupgrade the masked integer compare intrinsics Summary: These intrinsics aren't used by clang and haven't been for a while. There's some really terrible codegen in the 32-bit target for avx512bw due to i64 not being legal. But as I said these intrinsics aren't used by clang even before this patch so this codegen reflects our clang behavior today. Reviewers: spatel, RKSimon, zvi, igorb Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D34389 llvm-svn: 306047
2017-06-23 04:11:01 +08:00
; CHECK-NEXT: vpcmpeqb %xmm1, %xmm0, %k0 ## encoding: [0x62,0xf1,0x7d,0x08,0x74,0xc1]
; CHECK-NEXT: vpcmpgtb %xmm0, %xmm1, %k1 ## encoding: [0x62,0xf1,0x75,0x08,0x64,0xc8]
; CHECK-NEXT: vpcmpleb %xmm1, %xmm0, %k2 ## encoding: [0x62,0xf3,0x7d,0x08,0x3f,0xd1,0x02]
; CHECK-NEXT: vpcmpneqb %xmm1, %xmm0, %k3 ## encoding: [0x62,0xf3,0x7d,0x08,0x3f,0xd9,0x04]
; CHECK-NEXT: vpcmpleb %xmm0, %xmm1, %k4 ## encoding: [0x62,0xf3,0x75,0x08,0x3f,0xe0,0x02]
; CHECK-NEXT: vpcmpgtb %xmm1, %xmm0, %k5 ## encoding: [0x62,0xf1,0x7d,0x08,0x64,0xe9]
; CHECK-NEXT: kmovd %k0, %eax ## encoding: [0xc5,0xfb,0x93,0xc0]
; CHECK-NEXT: vpxor %xmm0, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xef,0xc0]
; CHECK-NEXT: vpinsrw $0, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x00]
; CHECK-NEXT: kmovd %k1, %eax ## encoding: [0xc5,0xfb,0x93,0xc1]
; CHECK-NEXT: vpinsrw $1, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x01]
; CHECK-NEXT: kmovd %k2, %eax ## encoding: [0xc5,0xfb,0x93,0xc2]
; CHECK-NEXT: vpinsrw $2, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x02]
; CHECK-NEXT: kmovd %k3, %eax ## encoding: [0xc5,0xfb,0x93,0xc3]
; CHECK-NEXT: vpinsrw $4, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x04]
; CHECK-NEXT: kmovd %k4, %eax ## encoding: [0xc5,0xfb,0x93,0xc4]
; CHECK-NEXT: vpinsrw $5, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x05]
; CHECK-NEXT: kmovd %k5, %eax ## encoding: [0xc5,0xfb,0x93,0xc5]
; CHECK-NEXT: vpinsrw $6, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x06]
; CHECK-NEXT: kxnorw %k0, %k0, %k0 ## encoding: [0xc5,0xfc,0x46,0xc0]
; CHECK-NEXT: kmovd %k0, %eax ## encoding: [0xc5,0xfb,0x93,0xc0]
; CHECK-NEXT: vpinsrw $7, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x07]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res0 = call i16 @llvm.x86.avx512.mask.cmp.b.128(<16 x i8> %a0, <16 x i8> %a1, i32 0, i16 -1)
%vec0 = insertelement <8 x i16> undef, i16 %res0, i32 0
%res1 = call i16 @llvm.x86.avx512.mask.cmp.b.128(<16 x i8> %a0, <16 x i8> %a1, i32 1, i16 -1)
%vec1 = insertelement <8 x i16> %vec0, i16 %res1, i32 1
%res2 = call i16 @llvm.x86.avx512.mask.cmp.b.128(<16 x i8> %a0, <16 x i8> %a1, i32 2, i16 -1)
%vec2 = insertelement <8 x i16> %vec1, i16 %res2, i32 2
%res3 = call i16 @llvm.x86.avx512.mask.cmp.b.128(<16 x i8> %a0, <16 x i8> %a1, i32 3, i16 -1)
%vec3 = insertelement <8 x i16> %vec2, i16 %res3, i32 3
%res4 = call i16 @llvm.x86.avx512.mask.cmp.b.128(<16 x i8> %a0, <16 x i8> %a1, i32 4, i16 -1)
%vec4 = insertelement <8 x i16> %vec3, i16 %res4, i32 4
%res5 = call i16 @llvm.x86.avx512.mask.cmp.b.128(<16 x i8> %a0, <16 x i8> %a1, i32 5, i16 -1)
%vec5 = insertelement <8 x i16> %vec4, i16 %res5, i32 5
%res6 = call i16 @llvm.x86.avx512.mask.cmp.b.128(<16 x i8> %a0, <16 x i8> %a1, i32 6, i16 -1)
%vec6 = insertelement <8 x i16> %vec5, i16 %res6, i32 6
%res7 = call i16 @llvm.x86.avx512.mask.cmp.b.128(<16 x i8> %a0, <16 x i8> %a1, i32 7, i16 -1)
%vec7 = insertelement <8 x i16> %vec6, i16 %res7, i32 7
ret <8 x i16> %vec7
}
define <8 x i16> @test_mask_cmp_b_128(<16 x i8> %a0, <16 x i8> %a1, i16 %mask) {
; CHECK-LABEL: test_mask_cmp_b_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove and autoupgrade the masked integer compare intrinsics Summary: These intrinsics aren't used by clang and haven't been for a while. There's some really terrible codegen in the 32-bit target for avx512bw due to i64 not being legal. But as I said these intrinsics aren't used by clang even before this patch so this codegen reflects our clang behavior today. Reviewers: spatel, RKSimon, zvi, igorb Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D34389 llvm-svn: 306047
2017-06-23 04:11:01 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
; CHECK-NEXT: vpcmpeqb %xmm1, %xmm0, %k0 {%k1} ## encoding: [0x62,0xf1,0x7d,0x09,0x74,0xc1]
; CHECK-NEXT: vpcmpgtb %xmm0, %xmm1, %k2 {%k1} ## encoding: [0x62,0xf1,0x75,0x09,0x64,0xd0]
; CHECK-NEXT: vpcmpleb %xmm1, %xmm0, %k3 {%k1} ## encoding: [0x62,0xf3,0x7d,0x09,0x3f,0xd9,0x02]
; CHECK-NEXT: kxorw %k0, %k0, %k4 ## encoding: [0xc5,0xfc,0x47,0xe0]
; CHECK-NEXT: vpcmpneqb %xmm1, %xmm0, %k5 {%k1} ## encoding: [0x62,0xf3,0x7d,0x09,0x3f,0xe9,0x04]
; CHECK-NEXT: vpcmpleb %xmm0, %xmm1, %k6 {%k1} ## encoding: [0x62,0xf3,0x75,0x09,0x3f,0xf0,0x02]
; CHECK-NEXT: vpcmpgtb %xmm1, %xmm0, %k1 {%k1} ## encoding: [0x62,0xf1,0x7d,0x09,0x64,0xc9]
; CHECK-NEXT: kmovd %k2, %eax ## encoding: [0xc5,0xfb,0x93,0xc2]
; CHECK-NEXT: kmovd %k0, %ecx ## encoding: [0xc5,0xfb,0x93,0xc8]
; CHECK-NEXT: vmovd %ecx, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6e,0xc1]
; CHECK-NEXT: vpinsrw $1, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x01]
; CHECK-NEXT: kmovd %k3, %eax ## encoding: [0xc5,0xfb,0x93,0xc3]
; CHECK-NEXT: vpinsrw $2, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x02]
; CHECK-NEXT: kmovd %k4, %eax ## encoding: [0xc5,0xfb,0x93,0xc4]
; CHECK-NEXT: vpinsrw $3, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x03]
; CHECK-NEXT: kmovd %k5, %eax ## encoding: [0xc5,0xfb,0x93,0xc5]
; CHECK-NEXT: vpinsrw $4, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x04]
; CHECK-NEXT: kmovd %k6, %eax ## encoding: [0xc5,0xfb,0x93,0xc6]
; CHECK-NEXT: vpinsrw $5, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x05]
; CHECK-NEXT: kmovd %k1, %eax ## encoding: [0xc5,0xfb,0x93,0xc1]
; CHECK-NEXT: vpinsrw $6, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x06]
; CHECK-NEXT: vpinsrw $7, %edi, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc7,0x07]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res0 = call i16 @llvm.x86.avx512.mask.cmp.b.128(<16 x i8> %a0, <16 x i8> %a1, i32 0, i16 %mask)
%vec0 = insertelement <8 x i16> undef, i16 %res0, i32 0
%res1 = call i16 @llvm.x86.avx512.mask.cmp.b.128(<16 x i8> %a0, <16 x i8> %a1, i32 1, i16 %mask)
%vec1 = insertelement <8 x i16> %vec0, i16 %res1, i32 1
%res2 = call i16 @llvm.x86.avx512.mask.cmp.b.128(<16 x i8> %a0, <16 x i8> %a1, i32 2, i16 %mask)
%vec2 = insertelement <8 x i16> %vec1, i16 %res2, i32 2
%res3 = call i16 @llvm.x86.avx512.mask.cmp.b.128(<16 x i8> %a0, <16 x i8> %a1, i32 3, i16 %mask)
%vec3 = insertelement <8 x i16> %vec2, i16 %res3, i32 3
%res4 = call i16 @llvm.x86.avx512.mask.cmp.b.128(<16 x i8> %a0, <16 x i8> %a1, i32 4, i16 %mask)
%vec4 = insertelement <8 x i16> %vec3, i16 %res4, i32 4
%res5 = call i16 @llvm.x86.avx512.mask.cmp.b.128(<16 x i8> %a0, <16 x i8> %a1, i32 5, i16 %mask)
%vec5 = insertelement <8 x i16> %vec4, i16 %res5, i32 5
%res6 = call i16 @llvm.x86.avx512.mask.cmp.b.128(<16 x i8> %a0, <16 x i8> %a1, i32 6, i16 %mask)
%vec6 = insertelement <8 x i16> %vec5, i16 %res6, i32 6
%res7 = call i16 @llvm.x86.avx512.mask.cmp.b.128(<16 x i8> %a0, <16 x i8> %a1, i32 7, i16 %mask)
%vec7 = insertelement <8 x i16> %vec6, i16 %res7, i32 7
ret <8 x i16> %vec7
}
declare i16 @llvm.x86.avx512.mask.cmp.b.128(<16 x i8>, <16 x i8>, i32, i16) nounwind readnone
define <8 x i16> @test_ucmp_b_128(<16 x i8> %a0, <16 x i8> %a1) {
; CHECK-LABEL: test_ucmp_b_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove and autoupgrade the masked integer compare intrinsics Summary: These intrinsics aren't used by clang and haven't been for a while. There's some really terrible codegen in the 32-bit target for avx512bw due to i64 not being legal. But as I said these intrinsics aren't used by clang even before this patch so this codegen reflects our clang behavior today. Reviewers: spatel, RKSimon, zvi, igorb Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D34389 llvm-svn: 306047
2017-06-23 04:11:01 +08:00
; CHECK-NEXT: vpcmpeqb %xmm1, %xmm0, %k0 ## encoding: [0x62,0xf1,0x7d,0x08,0x74,0xc1]
; CHECK-NEXT: vpcmpltub %xmm1, %xmm0, %k1 ## encoding: [0x62,0xf3,0x7d,0x08,0x3e,0xc9,0x01]
; CHECK-NEXT: vpcmpleub %xmm1, %xmm0, %k2 ## encoding: [0x62,0xf3,0x7d,0x08,0x3e,0xd1,0x02]
; CHECK-NEXT: vpcmpneqb %xmm1, %xmm0, %k3 ## encoding: [0x62,0xf3,0x7d,0x08,0x3f,0xd9,0x04]
; CHECK-NEXT: vpcmpnltub %xmm1, %xmm0, %k4 ## encoding: [0x62,0xf3,0x7d,0x08,0x3e,0xe1,0x05]
; CHECK-NEXT: vpcmpnleub %xmm1, %xmm0, %k5 ## encoding: [0x62,0xf3,0x7d,0x08,0x3e,0xe9,0x06]
; CHECK-NEXT: kmovd %k0, %eax ## encoding: [0xc5,0xfb,0x93,0xc0]
; CHECK-NEXT: vpxor %xmm0, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xef,0xc0]
; CHECK-NEXT: vpinsrw $0, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x00]
; CHECK-NEXT: kmovd %k1, %eax ## encoding: [0xc5,0xfb,0x93,0xc1]
; CHECK-NEXT: vpinsrw $1, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x01]
; CHECK-NEXT: kmovd %k2, %eax ## encoding: [0xc5,0xfb,0x93,0xc2]
; CHECK-NEXT: vpinsrw $2, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x02]
; CHECK-NEXT: kmovd %k3, %eax ## encoding: [0xc5,0xfb,0x93,0xc3]
; CHECK-NEXT: vpinsrw $4, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x04]
; CHECK-NEXT: kmovd %k4, %eax ## encoding: [0xc5,0xfb,0x93,0xc4]
; CHECK-NEXT: vpinsrw $5, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x05]
; CHECK-NEXT: kmovd %k5, %eax ## encoding: [0xc5,0xfb,0x93,0xc5]
; CHECK-NEXT: vpinsrw $6, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x06]
; CHECK-NEXT: kxnorw %k0, %k0, %k0 ## encoding: [0xc5,0xfc,0x46,0xc0]
; CHECK-NEXT: kmovd %k0, %eax ## encoding: [0xc5,0xfb,0x93,0xc0]
; CHECK-NEXT: vpinsrw $7, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x07]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res0 = call i16 @llvm.x86.avx512.mask.ucmp.b.128(<16 x i8> %a0, <16 x i8> %a1, i32 0, i16 -1)
%vec0 = insertelement <8 x i16> undef, i16 %res0, i32 0
%res1 = call i16 @llvm.x86.avx512.mask.ucmp.b.128(<16 x i8> %a0, <16 x i8> %a1, i32 1, i16 -1)
%vec1 = insertelement <8 x i16> %vec0, i16 %res1, i32 1
%res2 = call i16 @llvm.x86.avx512.mask.ucmp.b.128(<16 x i8> %a0, <16 x i8> %a1, i32 2, i16 -1)
%vec2 = insertelement <8 x i16> %vec1, i16 %res2, i32 2
%res3 = call i16 @llvm.x86.avx512.mask.ucmp.b.128(<16 x i8> %a0, <16 x i8> %a1, i32 3, i16 -1)
%vec3 = insertelement <8 x i16> %vec2, i16 %res3, i32 3
%res4 = call i16 @llvm.x86.avx512.mask.ucmp.b.128(<16 x i8> %a0, <16 x i8> %a1, i32 4, i16 -1)
%vec4 = insertelement <8 x i16> %vec3, i16 %res4, i32 4
%res5 = call i16 @llvm.x86.avx512.mask.ucmp.b.128(<16 x i8> %a0, <16 x i8> %a1, i32 5, i16 -1)
%vec5 = insertelement <8 x i16> %vec4, i16 %res5, i32 5
%res6 = call i16 @llvm.x86.avx512.mask.ucmp.b.128(<16 x i8> %a0, <16 x i8> %a1, i32 6, i16 -1)
%vec6 = insertelement <8 x i16> %vec5, i16 %res6, i32 6
%res7 = call i16 @llvm.x86.avx512.mask.ucmp.b.128(<16 x i8> %a0, <16 x i8> %a1, i32 7, i16 -1)
%vec7 = insertelement <8 x i16> %vec6, i16 %res7, i32 7
ret <8 x i16> %vec7
}
define <8 x i16> @test_mask_ucmp_b_128(<16 x i8> %a0, <16 x i8> %a1, i16 %mask) {
; CHECK-LABEL: test_mask_ucmp_b_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove and autoupgrade the masked integer compare intrinsics Summary: These intrinsics aren't used by clang and haven't been for a while. There's some really terrible codegen in the 32-bit target for avx512bw due to i64 not being legal. But as I said these intrinsics aren't used by clang even before this patch so this codegen reflects our clang behavior today. Reviewers: spatel, RKSimon, zvi, igorb Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D34389 llvm-svn: 306047
2017-06-23 04:11:01 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
; CHECK-NEXT: vpcmpeqb %xmm1, %xmm0, %k0 {%k1} ## encoding: [0x62,0xf1,0x7d,0x09,0x74,0xc1]
; CHECK-NEXT: vpcmpltub %xmm1, %xmm0, %k2 {%k1} ## encoding: [0x62,0xf3,0x7d,0x09,0x3e,0xd1,0x01]
; CHECK-NEXT: vpcmpleub %xmm1, %xmm0, %k3 {%k1} ## encoding: [0x62,0xf3,0x7d,0x09,0x3e,0xd9,0x02]
; CHECK-NEXT: kxorw %k0, %k0, %k4 ## encoding: [0xc5,0xfc,0x47,0xe0]
; CHECK-NEXT: vpcmpneqb %xmm1, %xmm0, %k5 {%k1} ## encoding: [0x62,0xf3,0x7d,0x09,0x3f,0xe9,0x04]
; CHECK-NEXT: vpcmpnltub %xmm1, %xmm0, %k6 {%k1} ## encoding: [0x62,0xf3,0x7d,0x09,0x3e,0xf1,0x05]
; CHECK-NEXT: vpcmpnleub %xmm1, %xmm0, %k1 {%k1} ## encoding: [0x62,0xf3,0x7d,0x09,0x3e,0xc9,0x06]
; CHECK-NEXT: kmovd %k2, %eax ## encoding: [0xc5,0xfb,0x93,0xc2]
; CHECK-NEXT: kmovd %k0, %ecx ## encoding: [0xc5,0xfb,0x93,0xc8]
; CHECK-NEXT: vmovd %ecx, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6e,0xc1]
; CHECK-NEXT: vpinsrw $1, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x01]
; CHECK-NEXT: kmovd %k3, %eax ## encoding: [0xc5,0xfb,0x93,0xc3]
; CHECK-NEXT: vpinsrw $2, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x02]
; CHECK-NEXT: kmovd %k4, %eax ## encoding: [0xc5,0xfb,0x93,0xc4]
; CHECK-NEXT: vpinsrw $3, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x03]
; CHECK-NEXT: kmovd %k5, %eax ## encoding: [0xc5,0xfb,0x93,0xc5]
; CHECK-NEXT: vpinsrw $4, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x04]
; CHECK-NEXT: kmovd %k6, %eax ## encoding: [0xc5,0xfb,0x93,0xc6]
; CHECK-NEXT: vpinsrw $5, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x05]
; CHECK-NEXT: kmovd %k1, %eax ## encoding: [0xc5,0xfb,0x93,0xc1]
; CHECK-NEXT: vpinsrw $6, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x06]
; CHECK-NEXT: vpinsrw $7, %edi, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc7,0x07]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res0 = call i16 @llvm.x86.avx512.mask.ucmp.b.128(<16 x i8> %a0, <16 x i8> %a1, i32 0, i16 %mask)
%vec0 = insertelement <8 x i16> undef, i16 %res0, i32 0
%res1 = call i16 @llvm.x86.avx512.mask.ucmp.b.128(<16 x i8> %a0, <16 x i8> %a1, i32 1, i16 %mask)
%vec1 = insertelement <8 x i16> %vec0, i16 %res1, i32 1
%res2 = call i16 @llvm.x86.avx512.mask.ucmp.b.128(<16 x i8> %a0, <16 x i8> %a1, i32 2, i16 %mask)
%vec2 = insertelement <8 x i16> %vec1, i16 %res2, i32 2
%res3 = call i16 @llvm.x86.avx512.mask.ucmp.b.128(<16 x i8> %a0, <16 x i8> %a1, i32 3, i16 %mask)
%vec3 = insertelement <8 x i16> %vec2, i16 %res3, i32 3
%res4 = call i16 @llvm.x86.avx512.mask.ucmp.b.128(<16 x i8> %a0, <16 x i8> %a1, i32 4, i16 %mask)
%vec4 = insertelement <8 x i16> %vec3, i16 %res4, i32 4
%res5 = call i16 @llvm.x86.avx512.mask.ucmp.b.128(<16 x i8> %a0, <16 x i8> %a1, i32 5, i16 %mask)
%vec5 = insertelement <8 x i16> %vec4, i16 %res5, i32 5
%res6 = call i16 @llvm.x86.avx512.mask.ucmp.b.128(<16 x i8> %a0, <16 x i8> %a1, i32 6, i16 %mask)
%vec6 = insertelement <8 x i16> %vec5, i16 %res6, i32 6
%res7 = call i16 @llvm.x86.avx512.mask.ucmp.b.128(<16 x i8> %a0, <16 x i8> %a1, i32 7, i16 %mask)
%vec7 = insertelement <8 x i16> %vec6, i16 %res7, i32 7
ret <8 x i16> %vec7
}
declare i16 @llvm.x86.avx512.mask.ucmp.b.128(<16 x i8>, <16 x i8>, i32, i16) nounwind readnone
define <8 x i8> @test_cmp_w_128(<8 x i16> %a0, <8 x i16> %a1) {
; CHECK-LABEL: test_cmp_w_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove and autoupgrade the masked integer compare intrinsics Summary: These intrinsics aren't used by clang and haven't been for a while. There's some really terrible codegen in the 32-bit target for avx512bw due to i64 not being legal. But as I said these intrinsics aren't used by clang even before this patch so this codegen reflects our clang behavior today. Reviewers: spatel, RKSimon, zvi, igorb Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D34389 llvm-svn: 306047
2017-06-23 04:11:01 +08:00
; CHECK-NEXT: vpcmpeqw %xmm1, %xmm0, %k0 ## encoding: [0x62,0xf1,0x7d,0x08,0x75,0xc1]
; CHECK-NEXT: vpcmpgtw %xmm0, %xmm1, %k1 ## encoding: [0x62,0xf1,0x75,0x08,0x65,0xc8]
; CHECK-NEXT: vpcmplew %xmm1, %xmm0, %k2 ## encoding: [0x62,0xf3,0xfd,0x08,0x3f,0xd1,0x02]
; CHECK-NEXT: vpcmpneqw %xmm1, %xmm0, %k3 ## encoding: [0x62,0xf3,0xfd,0x08,0x3f,0xd9,0x04]
; CHECK-NEXT: vpcmplew %xmm0, %xmm1, %k4 ## encoding: [0x62,0xf3,0xf5,0x08,0x3f,0xe0,0x02]
; CHECK-NEXT: vpcmpgtw %xmm1, %xmm0, %k5 ## encoding: [0x62,0xf1,0x7d,0x08,0x65,0xe9]
; CHECK-NEXT: kmovd %k0, %eax ## encoding: [0xc5,0xfb,0x93,0xc0]
; CHECK-NEXT: vpxor %xmm0, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xef,0xc0]
; CHECK-NEXT: vpinsrw $0, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x00]
; CHECK-NEXT: kmovd %k1, %eax ## encoding: [0xc5,0xfb,0x93,0xc1]
; CHECK-NEXT: vpinsrw $1, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x01]
; CHECK-NEXT: kmovd %k2, %eax ## encoding: [0xc5,0xfb,0x93,0xc2]
; CHECK-NEXT: vpinsrw $2, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x02]
; CHECK-NEXT: kmovd %k3, %eax ## encoding: [0xc5,0xfb,0x93,0xc3]
; CHECK-NEXT: vpinsrw $4, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x04]
; CHECK-NEXT: kmovd %k4, %eax ## encoding: [0xc5,0xfb,0x93,0xc4]
; CHECK-NEXT: vpinsrw $5, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x05]
; CHECK-NEXT: kmovd %k5, %eax ## encoding: [0xc5,0xfb,0x93,0xc5]
; CHECK-NEXT: vpinsrw $6, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x06]
; CHECK-NEXT: kxnorw %k0, %k0, %k0 ## encoding: [0xc5,0xfc,0x46,0xc0]
; CHECK-NEXT: kmovd %k0, %eax ## encoding: [0xc5,0xfb,0x93,0xc0]
; CHECK-NEXT: vpinsrw $7, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x07]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res0 = call i8 @llvm.x86.avx512.mask.cmp.w.128(<8 x i16> %a0, <8 x i16> %a1, i32 0, i8 -1)
%vec0 = insertelement <8 x i8> undef, i8 %res0, i32 0
%res1 = call i8 @llvm.x86.avx512.mask.cmp.w.128(<8 x i16> %a0, <8 x i16> %a1, i32 1, i8 -1)
%vec1 = insertelement <8 x i8> %vec0, i8 %res1, i32 1
%res2 = call i8 @llvm.x86.avx512.mask.cmp.w.128(<8 x i16> %a0, <8 x i16> %a1, i32 2, i8 -1)
%vec2 = insertelement <8 x i8> %vec1, i8 %res2, i32 2
%res3 = call i8 @llvm.x86.avx512.mask.cmp.w.128(<8 x i16> %a0, <8 x i16> %a1, i32 3, i8 -1)
%vec3 = insertelement <8 x i8> %vec2, i8 %res3, i32 3
%res4 = call i8 @llvm.x86.avx512.mask.cmp.w.128(<8 x i16> %a0, <8 x i16> %a1, i32 4, i8 -1)
%vec4 = insertelement <8 x i8> %vec3, i8 %res4, i32 4
%res5 = call i8 @llvm.x86.avx512.mask.cmp.w.128(<8 x i16> %a0, <8 x i16> %a1, i32 5, i8 -1)
%vec5 = insertelement <8 x i8> %vec4, i8 %res5, i32 5
%res6 = call i8 @llvm.x86.avx512.mask.cmp.w.128(<8 x i16> %a0, <8 x i16> %a1, i32 6, i8 -1)
%vec6 = insertelement <8 x i8> %vec5, i8 %res6, i32 6
%res7 = call i8 @llvm.x86.avx512.mask.cmp.w.128(<8 x i16> %a0, <8 x i16> %a1, i32 7, i8 -1)
%vec7 = insertelement <8 x i8> %vec6, i8 %res7, i32 7
ret <8 x i8> %vec7
}
define <8 x i8> @test_mask_cmp_w_128(<8 x i16> %a0, <8 x i16> %a1, i8 %mask) {
; CHECK-LABEL: test_mask_cmp_w_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove and autoupgrade the masked integer compare intrinsics Summary: These intrinsics aren't used by clang and haven't been for a while. There's some really terrible codegen in the 32-bit target for avx512bw due to i64 not being legal. But as I said these intrinsics aren't used by clang even before this patch so this codegen reflects our clang behavior today. Reviewers: spatel, RKSimon, zvi, igorb Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D34389 llvm-svn: 306047
2017-06-23 04:11:01 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
; CHECK-NEXT: vpcmpeqw %xmm1, %xmm0, %k0 {%k1} ## encoding: [0x62,0xf1,0x7d,0x09,0x75,0xc1]
; CHECK-NEXT: vpcmpgtw %xmm0, %xmm1, %k2 {%k1} ## encoding: [0x62,0xf1,0x75,0x09,0x65,0xd0]
; CHECK-NEXT: vpcmplew %xmm1, %xmm0, %k3 {%k1} ## encoding: [0x62,0xf3,0xfd,0x09,0x3f,0xd9,0x02]
; CHECK-NEXT: kxorw %k0, %k0, %k4 ## encoding: [0xc5,0xfc,0x47,0xe0]
; CHECK-NEXT: vpcmpneqw %xmm1, %xmm0, %k5 {%k1} ## encoding: [0x62,0xf3,0xfd,0x09,0x3f,0xe9,0x04]
; CHECK-NEXT: vpcmplew %xmm0, %xmm1, %k6 {%k1} ## encoding: [0x62,0xf3,0xf5,0x09,0x3f,0xf0,0x02]
; CHECK-NEXT: vpcmpgtw %xmm1, %xmm0, %k1 {%k1} ## encoding: [0x62,0xf1,0x7d,0x09,0x65,0xc9]
; CHECK-NEXT: kmovd %k2, %eax ## encoding: [0xc5,0xfb,0x93,0xc2]
; CHECK-NEXT: kmovd %k0, %ecx ## encoding: [0xc5,0xfb,0x93,0xc8]
; CHECK-NEXT: vmovd %ecx, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6e,0xc1]
; CHECK-NEXT: vpinsrb $2, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc4,0xe3,0x79,0x20,0xc0,0x02]
; CHECK-NEXT: kmovd %k3, %eax ## encoding: [0xc5,0xfb,0x93,0xc3]
; CHECK-NEXT: vpinsrb $4, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc4,0xe3,0x79,0x20,0xc0,0x04]
; CHECK-NEXT: kmovd %k4, %eax ## encoding: [0xc5,0xfb,0x93,0xc4]
; CHECK-NEXT: vpinsrb $6, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc4,0xe3,0x79,0x20,0xc0,0x06]
; CHECK-NEXT: kmovd %k5, %eax ## encoding: [0xc5,0xfb,0x93,0xc5]
; CHECK-NEXT: vpinsrb $8, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc4,0xe3,0x79,0x20,0xc0,0x08]
; CHECK-NEXT: kmovd %k6, %eax ## encoding: [0xc5,0xfb,0x93,0xc6]
; CHECK-NEXT: vpinsrb $10, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc4,0xe3,0x79,0x20,0xc0,0x0a]
; CHECK-NEXT: kmovd %k1, %eax ## encoding: [0xc5,0xfb,0x93,0xc1]
; CHECK-NEXT: vpinsrb $12, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc4,0xe3,0x79,0x20,0xc0,0x0c]
; CHECK-NEXT: vpinsrb $14, %edi, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc4,0xe3,0x79,0x20,0xc7,0x0e]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res0 = call i8 @llvm.x86.avx512.mask.cmp.w.128(<8 x i16> %a0, <8 x i16> %a1, i32 0, i8 %mask)
%vec0 = insertelement <8 x i8> undef, i8 %res0, i32 0
%res1 = call i8 @llvm.x86.avx512.mask.cmp.w.128(<8 x i16> %a0, <8 x i16> %a1, i32 1, i8 %mask)
%vec1 = insertelement <8 x i8> %vec0, i8 %res1, i32 1
%res2 = call i8 @llvm.x86.avx512.mask.cmp.w.128(<8 x i16> %a0, <8 x i16> %a1, i32 2, i8 %mask)
%vec2 = insertelement <8 x i8> %vec1, i8 %res2, i32 2
%res3 = call i8 @llvm.x86.avx512.mask.cmp.w.128(<8 x i16> %a0, <8 x i16> %a1, i32 3, i8 %mask)
%vec3 = insertelement <8 x i8> %vec2, i8 %res3, i32 3
%res4 = call i8 @llvm.x86.avx512.mask.cmp.w.128(<8 x i16> %a0, <8 x i16> %a1, i32 4, i8 %mask)
%vec4 = insertelement <8 x i8> %vec3, i8 %res4, i32 4
%res5 = call i8 @llvm.x86.avx512.mask.cmp.w.128(<8 x i16> %a0, <8 x i16> %a1, i32 5, i8 %mask)
%vec5 = insertelement <8 x i8> %vec4, i8 %res5, i32 5
%res6 = call i8 @llvm.x86.avx512.mask.cmp.w.128(<8 x i16> %a0, <8 x i16> %a1, i32 6, i8 %mask)
%vec6 = insertelement <8 x i8> %vec5, i8 %res6, i32 6
%res7 = call i8 @llvm.x86.avx512.mask.cmp.w.128(<8 x i16> %a0, <8 x i16> %a1, i32 7, i8 %mask)
%vec7 = insertelement <8 x i8> %vec6, i8 %res7, i32 7
ret <8 x i8> %vec7
}
declare i8 @llvm.x86.avx512.mask.cmp.w.128(<8 x i16>, <8 x i16>, i32, i8) nounwind readnone
define <8 x i8> @test_ucmp_w_128(<8 x i16> %a0, <8 x i16> %a1) {
; CHECK-LABEL: test_ucmp_w_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove and autoupgrade the masked integer compare intrinsics Summary: These intrinsics aren't used by clang and haven't been for a while. There's some really terrible codegen in the 32-bit target for avx512bw due to i64 not being legal. But as I said these intrinsics aren't used by clang even before this patch so this codegen reflects our clang behavior today. Reviewers: spatel, RKSimon, zvi, igorb Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D34389 llvm-svn: 306047
2017-06-23 04:11:01 +08:00
; CHECK-NEXT: vpcmpeqw %xmm1, %xmm0, %k0 ## encoding: [0x62,0xf1,0x7d,0x08,0x75,0xc1]
; CHECK-NEXT: vpcmpltuw %xmm1, %xmm0, %k1 ## encoding: [0x62,0xf3,0xfd,0x08,0x3e,0xc9,0x01]
; CHECK-NEXT: vpcmpleuw %xmm1, %xmm0, %k2 ## encoding: [0x62,0xf3,0xfd,0x08,0x3e,0xd1,0x02]
; CHECK-NEXT: vpcmpneqw %xmm1, %xmm0, %k3 ## encoding: [0x62,0xf3,0xfd,0x08,0x3f,0xd9,0x04]
; CHECK-NEXT: vpcmpnltuw %xmm1, %xmm0, %k4 ## encoding: [0x62,0xf3,0xfd,0x08,0x3e,0xe1,0x05]
; CHECK-NEXT: vpcmpnleuw %xmm1, %xmm0, %k5 ## encoding: [0x62,0xf3,0xfd,0x08,0x3e,0xe9,0x06]
; CHECK-NEXT: kmovd %k0, %eax ## encoding: [0xc5,0xfb,0x93,0xc0]
; CHECK-NEXT: vpxor %xmm0, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xef,0xc0]
; CHECK-NEXT: vpinsrw $0, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x00]
; CHECK-NEXT: kmovd %k1, %eax ## encoding: [0xc5,0xfb,0x93,0xc1]
; CHECK-NEXT: vpinsrw $1, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x01]
; CHECK-NEXT: kmovd %k2, %eax ## encoding: [0xc5,0xfb,0x93,0xc2]
; CHECK-NEXT: vpinsrw $2, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x02]
; CHECK-NEXT: kmovd %k3, %eax ## encoding: [0xc5,0xfb,0x93,0xc3]
; CHECK-NEXT: vpinsrw $4, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x04]
; CHECK-NEXT: kmovd %k4, %eax ## encoding: [0xc5,0xfb,0x93,0xc4]
; CHECK-NEXT: vpinsrw $5, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x05]
; CHECK-NEXT: kmovd %k5, %eax ## encoding: [0xc5,0xfb,0x93,0xc5]
; CHECK-NEXT: vpinsrw $6, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x06]
; CHECK-NEXT: kxnorw %k0, %k0, %k0 ## encoding: [0xc5,0xfc,0x46,0xc0]
; CHECK-NEXT: kmovd %k0, %eax ## encoding: [0xc5,0xfb,0x93,0xc0]
; CHECK-NEXT: vpinsrw $7, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xc4,0xc0,0x07]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res0 = call i8 @llvm.x86.avx512.mask.ucmp.w.128(<8 x i16> %a0, <8 x i16> %a1, i32 0, i8 -1)
%vec0 = insertelement <8 x i8> undef, i8 %res0, i32 0
%res1 = call i8 @llvm.x86.avx512.mask.ucmp.w.128(<8 x i16> %a0, <8 x i16> %a1, i32 1, i8 -1)
%vec1 = insertelement <8 x i8> %vec0, i8 %res1, i32 1
%res2 = call i8 @llvm.x86.avx512.mask.ucmp.w.128(<8 x i16> %a0, <8 x i16> %a1, i32 2, i8 -1)
%vec2 = insertelement <8 x i8> %vec1, i8 %res2, i32 2
%res3 = call i8 @llvm.x86.avx512.mask.ucmp.w.128(<8 x i16> %a0, <8 x i16> %a1, i32 3, i8 -1)
%vec3 = insertelement <8 x i8> %vec2, i8 %res3, i32 3
%res4 = call i8 @llvm.x86.avx512.mask.ucmp.w.128(<8 x i16> %a0, <8 x i16> %a1, i32 4, i8 -1)
%vec4 = insertelement <8 x i8> %vec3, i8 %res4, i32 4
%res5 = call i8 @llvm.x86.avx512.mask.ucmp.w.128(<8 x i16> %a0, <8 x i16> %a1, i32 5, i8 -1)
%vec5 = insertelement <8 x i8> %vec4, i8 %res5, i32 5
%res6 = call i8 @llvm.x86.avx512.mask.ucmp.w.128(<8 x i16> %a0, <8 x i16> %a1, i32 6, i8 -1)
%vec6 = insertelement <8 x i8> %vec5, i8 %res6, i32 6
%res7 = call i8 @llvm.x86.avx512.mask.ucmp.w.128(<8 x i16> %a0, <8 x i16> %a1, i32 7, i8 -1)
%vec7 = insertelement <8 x i8> %vec6, i8 %res7, i32 7
ret <8 x i8> %vec7
}
define <8 x i8> @test_mask_ucmp_w_128(<8 x i16> %a0, <8 x i16> %a1, i8 %mask) {
; CHECK-LABEL: test_mask_ucmp_w_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[AVX-512] Remove and autoupgrade the masked integer compare intrinsics Summary: These intrinsics aren't used by clang and haven't been for a while. There's some really terrible codegen in the 32-bit target for avx512bw due to i64 not being legal. But as I said these intrinsics aren't used by clang even before this patch so this codegen reflects our clang behavior today. Reviewers: spatel, RKSimon, zvi, igorb Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D34389 llvm-svn: 306047
2017-06-23 04:11:01 +08:00
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
; CHECK-NEXT: vpcmpeqw %xmm1, %xmm0, %k0 {%k1} ## encoding: [0x62,0xf1,0x7d,0x09,0x75,0xc1]
; CHECK-NEXT: vpcmpltuw %xmm1, %xmm0, %k2 {%k1} ## encoding: [0x62,0xf3,0xfd,0x09,0x3e,0xd1,0x01]
; CHECK-NEXT: vpcmpleuw %xmm1, %xmm0, %k3 {%k1} ## encoding: [0x62,0xf3,0xfd,0x09,0x3e,0xd9,0x02]
; CHECK-NEXT: kxorw %k0, %k0, %k4 ## encoding: [0xc5,0xfc,0x47,0xe0]
; CHECK-NEXT: vpcmpneqw %xmm1, %xmm0, %k5 {%k1} ## encoding: [0x62,0xf3,0xfd,0x09,0x3f,0xe9,0x04]
; CHECK-NEXT: vpcmpnltuw %xmm1, %xmm0, %k6 {%k1} ## encoding: [0x62,0xf3,0xfd,0x09,0x3e,0xf1,0x05]
; CHECK-NEXT: vpcmpnleuw %xmm1, %xmm0, %k1 {%k1} ## encoding: [0x62,0xf3,0xfd,0x09,0x3e,0xc9,0x06]
; CHECK-NEXT: kmovd %k2, %eax ## encoding: [0xc5,0xfb,0x93,0xc2]
; CHECK-NEXT: kmovd %k0, %ecx ## encoding: [0xc5,0xfb,0x93,0xc8]
; CHECK-NEXT: vmovd %ecx, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6e,0xc1]
; CHECK-NEXT: vpinsrb $2, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc4,0xe3,0x79,0x20,0xc0,0x02]
; CHECK-NEXT: kmovd %k3, %eax ## encoding: [0xc5,0xfb,0x93,0xc3]
; CHECK-NEXT: vpinsrb $4, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc4,0xe3,0x79,0x20,0xc0,0x04]
; CHECK-NEXT: kmovd %k4, %eax ## encoding: [0xc5,0xfb,0x93,0xc4]
; CHECK-NEXT: vpinsrb $6, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc4,0xe3,0x79,0x20,0xc0,0x06]
; CHECK-NEXT: kmovd %k5, %eax ## encoding: [0xc5,0xfb,0x93,0xc5]
; CHECK-NEXT: vpinsrb $8, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc4,0xe3,0x79,0x20,0xc0,0x08]
; CHECK-NEXT: kmovd %k6, %eax ## encoding: [0xc5,0xfb,0x93,0xc6]
; CHECK-NEXT: vpinsrb $10, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc4,0xe3,0x79,0x20,0xc0,0x0a]
; CHECK-NEXT: kmovd %k1, %eax ## encoding: [0xc5,0xfb,0x93,0xc1]
; CHECK-NEXT: vpinsrb $12, %eax, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc4,0xe3,0x79,0x20,0xc0,0x0c]
; CHECK-NEXT: vpinsrb $14, %edi, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc4,0xe3,0x79,0x20,0xc7,0x0e]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res0 = call i8 @llvm.x86.avx512.mask.ucmp.w.128(<8 x i16> %a0, <8 x i16> %a1, i32 0, i8 %mask)
%vec0 = insertelement <8 x i8> undef, i8 %res0, i32 0
%res1 = call i8 @llvm.x86.avx512.mask.ucmp.w.128(<8 x i16> %a0, <8 x i16> %a1, i32 1, i8 %mask)
%vec1 = insertelement <8 x i8> %vec0, i8 %res1, i32 1
%res2 = call i8 @llvm.x86.avx512.mask.ucmp.w.128(<8 x i16> %a0, <8 x i16> %a1, i32 2, i8 %mask)
%vec2 = insertelement <8 x i8> %vec1, i8 %res2, i32 2
%res3 = call i8 @llvm.x86.avx512.mask.ucmp.w.128(<8 x i16> %a0, <8 x i16> %a1, i32 3, i8 %mask)
%vec3 = insertelement <8 x i8> %vec2, i8 %res3, i32 3
%res4 = call i8 @llvm.x86.avx512.mask.ucmp.w.128(<8 x i16> %a0, <8 x i16> %a1, i32 4, i8 %mask)
%vec4 = insertelement <8 x i8> %vec3, i8 %res4, i32 4
%res5 = call i8 @llvm.x86.avx512.mask.ucmp.w.128(<8 x i16> %a0, <8 x i16> %a1, i32 5, i8 %mask)
%vec5 = insertelement <8 x i8> %vec4, i8 %res5, i32 5
%res6 = call i8 @llvm.x86.avx512.mask.ucmp.w.128(<8 x i16> %a0, <8 x i16> %a1, i32 6, i8 %mask)
%vec6 = insertelement <8 x i8> %vec5, i8 %res6, i32 6
%res7 = call i8 @llvm.x86.avx512.mask.ucmp.w.128(<8 x i16> %a0, <8 x i16> %a1, i32 7, i8 %mask)
%vec7 = insertelement <8 x i8> %vec6, i8 %res7, i32 7
ret <8 x i8> %vec7
}
declare i8 @llvm.x86.avx512.mask.ucmp.w.128(<8 x i16>, <8 x i16>, i32, i8) nounwind readnone
[X86] Lower _mm[256|512]_[mask[z]]_avg_epu[8|16] intrinsics to native llvm IR Differential Revision: https://reviews.llvm.org/D37560 llvm-svn: 313013
2017-09-12 15:50:35 +08:00
define <16 x i8>@mm_mask_avg_epu8(<16 x i8> %x0, <16 x i8> %x1, <16 x i8> %x2, i16 %x3) {
; CHECK-LABEL: mm_mask_avg_epu8:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[X86] Lower _mm[256|512]_[mask[z]]_avg_epu[8|16] intrinsics to native llvm IR Differential Revision: https://reviews.llvm.org/D37560 llvm-svn: 313013
2017-09-12 15:50:35 +08:00
; CHECK-NEXT: vpavgb %xmm1, %xmm0, %xmm3 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xe0,0xd9]
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
; CHECK-NEXT: vpavgb %xmm1, %xmm0, %xmm2 {%k1} ## encoding: [0x62,0xf1,0x7d,0x09,0xe0,0xd1]
; CHECK-NEXT: vpaddb %xmm3, %xmm2, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xe9,0xfc,0xc3]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <16 x i8> @llvm.x86.avx512.mask.pavg.b.128(<16 x i8> %x0, <16 x i8> %x1, <16 x i8> %x2, i16 %x3)
%res1 = call <16 x i8> @llvm.x86.avx512.mask.pavg.b.128(<16 x i8> %x0, <16 x i8> %x1, <16 x i8> %x2, i16 -1)
[X86] [PATCH] [intrinsics] Lowering X86 ABS intrinsics to IR. (llvm) This patch, together with a matching clang patch (https://reviews.llvm.org/D37694), implements the lowering of X86 ABS intrinsics to IR. differential revision: https://reviews.llvm.org/D37693. llvm-svn: 313134
2017-09-13 17:02:36 +08:00
%res2 = add <16 x i8> %res, %res1
ret <16 x i8> %res2
}
declare <16 x i8> @llvm.x86.avx512.mask.pabs.b.128(<16 x i8>, <16 x i8>, i16)
define <16 x i8>@test_int_x86_avx512_mask_pabs_b_128(<16 x i8> %x0, <16 x i8> %x1, i16 %x2) {
; CHECK-LABEL: test_int_x86_avx512_mask_pabs_b_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[X86] [PATCH] [intrinsics] Lowering X86 ABS intrinsics to IR. (llvm) This patch, together with a matching clang patch (https://reviews.llvm.org/D37694), implements the lowering of X86 ABS intrinsics to IR. differential revision: https://reviews.llvm.org/D37693. llvm-svn: 313134
2017-09-13 17:02:36 +08:00
; CHECK-NEXT: vpabsb %xmm0, %xmm2 ## EVEX TO VEX Compression encoding: [0xc4,0xe2,0x79,0x1c,0xd0]
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
; CHECK-NEXT: vpabsb %xmm0, %xmm1 {%k1} ## encoding: [0x62,0xf2,0x7d,0x09,0x1c,0xc8]
; CHECK-NEXT: vpaddb %xmm2, %xmm1, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf1,0xfc,0xc2]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <16 x i8> @llvm.x86.avx512.mask.pabs.b.128(<16 x i8> %x0, <16 x i8> %x1, i16 %x2)
%res1 = call <16 x i8> @llvm.x86.avx512.mask.pabs.b.128(<16 x i8> %x0, <16 x i8> %x1, i16 -1)
[X86] Lower _mm[256|512]_[mask[z]]_avg_epu[8|16] intrinsics to native llvm IR Differential Revision: https://reviews.llvm.org/D37560 llvm-svn: 313013
2017-09-12 15:50:35 +08:00
%res2 = add <16 x i8> %res, %res1
ret <16 x i8> %res2
}
declare <16 x i8> @llvm.x86.avx512.mask.pavg.b.128(<16 x i8>, <16 x i8>, <16 x i8>, i16)
define <32 x i8>@mm256_mask_avg_epu8(<32 x i8> %x0, <32 x i8> %x1, <32 x i8> %x2, i32 %x3) {
; CHECK-LABEL: mm256_mask_avg_epu8:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[X86] Lower _mm[256|512]_[mask[z]]_avg_epu[8|16] intrinsics to native llvm IR Differential Revision: https://reviews.llvm.org/D37560 llvm-svn: 313013
2017-09-12 15:50:35 +08:00
; CHECK-NEXT: vpavgb %ymm1, %ymm0, %ymm3 ## EVEX TO VEX Compression encoding: [0xc5,0xfd,0xe0,0xd9]
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
; CHECK-NEXT: vpavgb %ymm1, %ymm0, %ymm2 {%k1} ## encoding: [0x62,0xf1,0x7d,0x29,0xe0,0xd1]
; CHECK-NEXT: vpaddb %ymm3, %ymm2, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xed,0xfc,0xc3]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <32 x i8> @llvm.x86.avx512.mask.pavg.b.256(<32 x i8> %x0, <32 x i8> %x1, <32 x i8> %x2, i32 %x3)
%res1 = call <32 x i8> @llvm.x86.avx512.mask.pavg.b.256(<32 x i8> %x0, <32 x i8> %x1, <32 x i8> %x2, i32 -1)
%res2 = add <32 x i8> %res, %res1
ret <32 x i8> %res2
}
[X86] [PATCH] [intrinsics] Lowering X86 ABS intrinsics to IR. (llvm) This patch, together with a matching clang patch (https://reviews.llvm.org/D37694), implements the lowering of X86 ABS intrinsics to IR. differential revision: https://reviews.llvm.org/D37693. llvm-svn: 313134
2017-09-13 17:02:36 +08:00
declare <32 x i8> @llvm.x86.avx512.mask.pabs.b.256(<32 x i8>, <32 x i8>, i32)
define <32 x i8>@test_int_x86_avx512_mask_pabs_b_256(<32 x i8> %x0, <32 x i8> %x1, i32 %x2) {
; CHECK-LABEL: test_int_x86_avx512_mask_pabs_b_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[X86] [PATCH] [intrinsics] Lowering X86 ABS intrinsics to IR. (llvm) This patch, together with a matching clang patch (https://reviews.llvm.org/D37694), implements the lowering of X86 ABS intrinsics to IR. differential revision: https://reviews.llvm.org/D37693. llvm-svn: 313134
2017-09-13 17:02:36 +08:00
; CHECK-NEXT: vpabsb %ymm0, %ymm2 ## EVEX TO VEX Compression encoding: [0xc4,0xe2,0x7d,0x1c,0xd0]
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
; CHECK-NEXT: vpabsb %ymm0, %ymm1 {%k1} ## encoding: [0x62,0xf2,0x7d,0x29,0x1c,0xc8]
; CHECK-NEXT: vpaddb %ymm2, %ymm1, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf5,0xfc,0xc2]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <32 x i8> @llvm.x86.avx512.mask.pabs.b.256(<32 x i8> %x0, <32 x i8> %x1, i32 %x2)
%res1 = call <32 x i8> @llvm.x86.avx512.mask.pabs.b.256(<32 x i8> %x0, <32 x i8> %x1, i32 -1)
%res2 = add <32 x i8> %res, %res1
ret <32 x i8> %res2
}
[X86] Lower _mm[256|512]_[mask[z]]_avg_epu[8|16] intrinsics to native llvm IR Differential Revision: https://reviews.llvm.org/D37560 llvm-svn: 313013
2017-09-12 15:50:35 +08:00
declare <32 x i8> @llvm.x86.avx512.mask.pavg.b.256(<32 x i8>, <32 x i8>, <32 x i8>, i32)
define <8 x i16>@mm_mask_avg_epu16(<8 x i16> %x0, <8 x i16> %x1, <8 x i16> %x2, i8 %x3) {
; CHECK-LABEL: mm_mask_avg_epu16:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[X86] Lower _mm[256|512]_[mask[z]]_avg_epu[8|16] intrinsics to native llvm IR Differential Revision: https://reviews.llvm.org/D37560 llvm-svn: 313013
2017-09-12 15:50:35 +08:00
; CHECK-NEXT: vpavgw %xmm1, %xmm0, %xmm3 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xe3,0xd9]
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
; CHECK-NEXT: vpavgw %xmm1, %xmm0, %xmm2 {%k1} ## encoding: [0x62,0xf1,0x7d,0x09,0xe3,0xd1]
; CHECK-NEXT: vpaddw %xmm3, %xmm2, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xe9,0xfd,0xc3]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <8 x i16> @llvm.x86.avx512.mask.pavg.w.128(<8 x i16> %x0, <8 x i16> %x1, <8 x i16> %x2, i8 %x3)
%res1 = call <8 x i16> @llvm.x86.avx512.mask.pavg.w.128(<8 x i16> %x0, <8 x i16> %x1, <8 x i16> %x2, i8 -1)
%res2 = add <8 x i16> %res, %res1
ret <8 x i16> %res2
}
[X86] [PATCH] [intrinsics] Lowering X86 ABS intrinsics to IR. (llvm) This patch, together with a matching clang patch (https://reviews.llvm.org/D37694), implements the lowering of X86 ABS intrinsics to IR. differential revision: https://reviews.llvm.org/D37693. llvm-svn: 313134
2017-09-13 17:02:36 +08:00
declare <8 x i16> @llvm.x86.avx512.mask.pabs.w.128(<8 x i16>, <8 x i16>, i8)
define <8 x i16>@test_int_x86_avx512_mask_pabs_w_128(<8 x i16> %x0, <8 x i16> %x1, i8 %x2) {
; CHECK-LABEL: test_int_x86_avx512_mask_pabs_w_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[X86] [PATCH] [intrinsics] Lowering X86 ABS intrinsics to IR. (llvm) This patch, together with a matching clang patch (https://reviews.llvm.org/D37694), implements the lowering of X86 ABS intrinsics to IR. differential revision: https://reviews.llvm.org/D37693. llvm-svn: 313134
2017-09-13 17:02:36 +08:00
; CHECK-NEXT: vpabsw %xmm0, %xmm2 ## EVEX TO VEX Compression encoding: [0xc4,0xe2,0x79,0x1d,0xd0]
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
; CHECK-NEXT: vpabsw %xmm0, %xmm1 {%k1} ## encoding: [0x62,0xf2,0x7d,0x09,0x1d,0xc8]
; CHECK-NEXT: vpaddw %xmm2, %xmm1, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf1,0xfd,0xc2]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <8 x i16> @llvm.x86.avx512.mask.pabs.w.128(<8 x i16> %x0, <8 x i16> %x1, i8 %x2)
%res1 = call <8 x i16> @llvm.x86.avx512.mask.pabs.w.128(<8 x i16> %x0, <8 x i16> %x1, i8 -1)
%res2 = add <8 x i16> %res, %res1
ret <8 x i16> %res2
}
[X86] Lower _mm[256|512]_[mask[z]]_avg_epu[8|16] intrinsics to native llvm IR Differential Revision: https://reviews.llvm.org/D37560 llvm-svn: 313013
2017-09-12 15:50:35 +08:00
declare <8 x i16> @llvm.x86.avx512.mask.pavg.w.128(<8 x i16>, <8 x i16>, <8 x i16>, i8)
define <16 x i16>@mm256_mask_avg_epu16(<16 x i16> %x0, <16 x i16> %x1, <16 x i16> %x2, i16 %x3) {
; CHECK-LABEL: mm256_mask_avg_epu16:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[X86] Lower _mm[256|512]_[mask[z]]_avg_epu[8|16] intrinsics to native llvm IR Differential Revision: https://reviews.llvm.org/D37560 llvm-svn: 313013
2017-09-12 15:50:35 +08:00
; CHECK-NEXT: vpavgw %ymm1, %ymm0, %ymm3 ## EVEX TO VEX Compression encoding: [0xc5,0xfd,0xe3,0xd9]
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
; CHECK-NEXT: vpavgw %ymm1, %ymm0, %ymm2 {%k1} ## encoding: [0x62,0xf1,0x7d,0x29,0xe3,0xd1]
; CHECK-NEXT: vpaddw %ymm3, %ymm2, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xed,0xfd,0xc3]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <16 x i16> @llvm.x86.avx512.mask.pavg.w.256(<16 x i16> %x0, <16 x i16> %x1, <16 x i16> %x2, i16 %x3)
%res1 = call <16 x i16> @llvm.x86.avx512.mask.pavg.w.256(<16 x i16> %x0, <16 x i16> %x1, <16 x i16> %x2, i16 -1)
%res2 = add <16 x i16> %res, %res1
ret <16 x i16> %res2
}
[X86] [PATCH] [intrinsics] Lowering X86 ABS intrinsics to IR. (llvm) This patch, together with a matching clang patch (https://reviews.llvm.org/D37694), implements the lowering of X86 ABS intrinsics to IR. differential revision: https://reviews.llvm.org/D37693. llvm-svn: 313134
2017-09-13 17:02:36 +08:00
declare <16 x i16> @llvm.x86.avx512.mask.pabs.w.256(<16 x i16>, <16 x i16>, i16)
define <16 x i16>@test_int_x86_avx512_mask_pabs_w_256(<16 x i16> %x0, <16 x i16> %x1, i16 %x2) {
; CHECK-LABEL: test_int_x86_avx512_mask_pabs_w_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[X86] [PATCH] [intrinsics] Lowering X86 ABS intrinsics to IR. (llvm) This patch, together with a matching clang patch (https://reviews.llvm.org/D37694), implements the lowering of X86 ABS intrinsics to IR. differential revision: https://reviews.llvm.org/D37693. llvm-svn: 313134
2017-09-13 17:02:36 +08:00
; CHECK-NEXT: vpabsw %ymm0, %ymm2 ## EVEX TO VEX Compression encoding: [0xc4,0xe2,0x7d,0x1d,0xd0]
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
; CHECK-NEXT: vpabsw %ymm0, %ymm1 {%k1} ## encoding: [0x62,0xf2,0x7d,0x29,0x1d,0xc8]
; CHECK-NEXT: vpaddw %ymm2, %ymm1, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf5,0xfd,0xc2]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call <16 x i16> @llvm.x86.avx512.mask.pabs.w.256(<16 x i16> %x0, <16 x i16> %x1, i16 %x2)
%res1 = call <16 x i16> @llvm.x86.avx512.mask.pabs.w.256(<16 x i16> %x0, <16 x i16> %x1, i16 -1)
%res2 = add <16 x i16> %res, %res1
ret <16 x i16> %res2
}
[X86] Lower _mm[256|512]_[mask[z]]_avg_epu[8|16] intrinsics to native llvm IR Differential Revision: https://reviews.llvm.org/D37560 llvm-svn: 313013
2017-09-12 15:50:35 +08:00
declare <16 x i16> @llvm.x86.avx512.mask.pavg.w.256(<16 x i16>, <16 x i16>, <16 x i16>, i16)
[X86] [PATCH] [intrinsics] Lowering X86 ABS intrinsics to IR. (llvm) This patch, together with a matching clang patch (https://reviews.llvm.org/D37694), implements the lowering of X86 ABS intrinsics to IR. differential revision: https://reviews.llvm.org/D37693. llvm-svn: 313134
2017-09-13 17:02:36 +08:00
[X86] test/testn intrinsics lowering to IR. llvm part. Remove builtins from llvm and add AutoUpgrade support. Also add fast-isel tests for the TEST and TESTN instructions. Differential Revision: https://reviews.llvm.org/D38736 llvm-svn: 318036
2017-11-13 20:51:18 +08:00
declare i16 @llvm.x86.avx512.ptestm.b.128(<16 x i8>, <16 x i8>, i16)
define i16@test_int_x86_avx512_ptestm_b_128(<16 x i8> %x0, <16 x i8> %x1, i16 %x2) {
; CHECK-LABEL: test_int_x86_avx512_ptestm_b_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[X86] test/testn intrinsics lowering to IR. llvm part. Remove builtins from llvm and add AutoUpgrade support. Also add fast-isel tests for the TEST and TESTN instructions. Differential Revision: https://reviews.llvm.org/D38736 llvm-svn: 318036
2017-11-13 20:51:18 +08:00
; CHECK-NEXT: vptestmb %xmm1, %xmm0, %k0 ## encoding: [0x62,0xf2,0x7d,0x08,0x26,0xc1]
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
; CHECK-NEXT: vptestmb %xmm1, %xmm0, %k1 {%k1} ## encoding: [0x62,0xf2,0x7d,0x09,0x26,0xc9]
; CHECK-NEXT: kmovd %k1, %ecx ## encoding: [0xc5,0xfb,0x93,0xc9]
; CHECK-NEXT: kmovd %k0, %eax ## encoding: [0xc5,0xfb,0x93,0xc0]
; CHECK-NEXT: addl %ecx, %eax ## encoding: [0x01,0xc8]
[CodeGen] Use MachineOperand::print in the MIRPrinter for MO_Register. Work towards the unification of MIR and debug output by refactoring the interfaces. For MachineOperand::print, keep a simple version that can be easily called from `dump()`, and a more complex one which will be called from both the MIRPrinter and MachineInstr::print. Add extra checks inside MachineOperand for detached operands (operands with getParent() == nullptr). https://reviews.llvm.org/D40836 * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/kill: ([^ ]+) ([^ ]+)<def> ([^ ]+)/kill: \1 def \2 \3/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/kill: ([^ ]+) ([^ ]+) ([^ ]+)<def>/kill: \1 \2 def \3/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/kill: def ([^ ]+) ([^ ]+) ([^ ]+)<def>/kill: def \1 \2 def \3/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/<def>//g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<kill>/killed \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-use,kill>/implicit killed \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<dead>/dead \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<def[ ]*,[ ]*dead>/dead \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-def[ ]*,[ ]*dead>/implicit-def dead \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-def>/implicit-def \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-use>/implicit \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<internal>/internal \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<undef>/undef \1/g' llvm-svn: 320022
2017-12-07 18:40:31 +08:00
; CHECK-NEXT: ## kill: def %ax killed %ax killed %eax
[X86] test/testn intrinsics lowering to IR. llvm part. Remove builtins from llvm and add AutoUpgrade support. Also add fast-isel tests for the TEST and TESTN instructions. Differential Revision: https://reviews.llvm.org/D38736 llvm-svn: 318036
2017-11-13 20:51:18 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call i16 @llvm.x86.avx512.ptestm.b.128(<16 x i8> %x0, <16 x i8> %x1, i16 %x2)
%res1 = call i16 @llvm.x86.avx512.ptestm.b.128(<16 x i8> %x0, <16 x i8> %x1, i16-1)
%res2 = add i16 %res, %res1
ret i16 %res2
}
declare i32 @llvm.x86.avx512.ptestm.b.256(<32 x i8>, <32 x i8>, i32)
define i32@test_int_x86_avx512_ptestm_b_256(<32 x i8> %x0, <32 x i8> %x1, i32 %x2) {
; CHECK-LABEL: test_int_x86_avx512_ptestm_b_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[X86] test/testn intrinsics lowering to IR. llvm part. Remove builtins from llvm and add AutoUpgrade support. Also add fast-isel tests for the TEST and TESTN instructions. Differential Revision: https://reviews.llvm.org/D38736 llvm-svn: 318036
2017-11-13 20:51:18 +08:00
; CHECK-NEXT: vptestmb %ymm1, %ymm0, %k0 ## encoding: [0x62,0xf2,0x7d,0x28,0x26,0xc1]
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
; CHECK-NEXT: vptestmb %ymm1, %ymm0, %k1 {%k1} ## encoding: [0x62,0xf2,0x7d,0x29,0x26,0xc9]
; CHECK-NEXT: kmovd %k1, %ecx ## encoding: [0xc5,0xfb,0x93,0xc9]
; CHECK-NEXT: kmovd %k0, %eax ## encoding: [0xc5,0xfb,0x93,0xc0]
; CHECK-NEXT: addl %ecx, %eax ## encoding: [0x01,0xc8]
; CHECK-NEXT: vzeroupper ## encoding: [0xc5,0xf8,0x77]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call i32 @llvm.x86.avx512.ptestm.b.256(<32 x i8> %x0, <32 x i8> %x1, i32 %x2)
%res1 = call i32 @llvm.x86.avx512.ptestm.b.256(<32 x i8> %x0, <32 x i8> %x1, i32-1)
%res2 = add i32 %res, %res1
ret i32 %res2
}
declare i8 @llvm.x86.avx512.ptestm.w.128(<8 x i16>, <8 x i16>, i8)
define i8@test_int_x86_avx512_ptestm_w_128(<8 x i16> %x0, <8 x i16> %x1, i8 %x2) {
; CHECK-LABEL: test_int_x86_avx512_ptestm_w_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[X86] test/testn intrinsics lowering to IR. llvm part. Remove builtins from llvm and add AutoUpgrade support. Also add fast-isel tests for the TEST and TESTN instructions. Differential Revision: https://reviews.llvm.org/D38736 llvm-svn: 318036
2017-11-13 20:51:18 +08:00
; CHECK-NEXT: vptestmw %xmm1, %xmm0, %k0 ## encoding: [0x62,0xf2,0xfd,0x08,0x26,0xc1]
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
; CHECK-NEXT: vptestmw %xmm1, %xmm0, %k1 {%k1} ## encoding: [0x62,0xf2,0xfd,0x09,0x26,0xc9]
; CHECK-NEXT: kmovd %k1, %ecx ## encoding: [0xc5,0xfb,0x93,0xc9]
; CHECK-NEXT: kmovd %k0, %eax ## encoding: [0xc5,0xfb,0x93,0xc0]
; CHECK-NEXT: addb %cl, %al ## encoding: [0x00,0xc8]
[CodeGen] Use MachineOperand::print in the MIRPrinter for MO_Register. Work towards the unification of MIR and debug output by refactoring the interfaces. For MachineOperand::print, keep a simple version that can be easily called from `dump()`, and a more complex one which will be called from both the MIRPrinter and MachineInstr::print. Add extra checks inside MachineOperand for detached operands (operands with getParent() == nullptr). https://reviews.llvm.org/D40836 * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/kill: ([^ ]+) ([^ ]+)<def> ([^ ]+)/kill: \1 def \2 \3/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/kill: ([^ ]+) ([^ ]+) ([^ ]+)<def>/kill: \1 \2 def \3/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/kill: def ([^ ]+) ([^ ]+) ([^ ]+)<def>/kill: def \1 \2 def \3/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/<def>//g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<kill>/killed \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-use,kill>/implicit killed \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<dead>/dead \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<def[ ]*,[ ]*dead>/dead \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-def[ ]*,[ ]*dead>/implicit-def dead \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-def>/implicit-def \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-use>/implicit \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<internal>/internal \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<undef>/undef \1/g' llvm-svn: 320022
2017-12-07 18:40:31 +08:00
; CHECK-NEXT: ## kill: def %al killed %al killed %eax
[X86] test/testn intrinsics lowering to IR. llvm part. Remove builtins from llvm and add AutoUpgrade support. Also add fast-isel tests for the TEST and TESTN instructions. Differential Revision: https://reviews.llvm.org/D38736 llvm-svn: 318036
2017-11-13 20:51:18 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call i8 @llvm.x86.avx512.ptestm.w.128(<8 x i16> %x0, <8 x i16> %x1, i8 %x2)
%res1 = call i8 @llvm.x86.avx512.ptestm.w.128(<8 x i16> %x0, <8 x i16> %x1, i8-1)
%res2 = add i8 %res, %res1
ret i8 %res2
}
declare i16 @llvm.x86.avx512.ptestm.w.256(<16 x i16>, <16 x i16>, i16)
define i16@test_int_x86_avx512_ptestm_w_256(<16 x i16> %x0, <16 x i16> %x1, i16 %x2) {
; CHECK-LABEL: test_int_x86_avx512_ptestm_w_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[X86] test/testn intrinsics lowering to IR. llvm part. Remove builtins from llvm and add AutoUpgrade support. Also add fast-isel tests for the TEST and TESTN instructions. Differential Revision: https://reviews.llvm.org/D38736 llvm-svn: 318036
2017-11-13 20:51:18 +08:00
; CHECK-NEXT: vptestmw %ymm1, %ymm0, %k0 ## encoding: [0x62,0xf2,0xfd,0x28,0x26,0xc1]
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
; CHECK-NEXT: vptestmw %ymm1, %ymm0, %k1 {%k1} ## encoding: [0x62,0xf2,0xfd,0x29,0x26,0xc9]
; CHECK-NEXT: kmovd %k1, %ecx ## encoding: [0xc5,0xfb,0x93,0xc9]
; CHECK-NEXT: kmovd %k0, %eax ## encoding: [0xc5,0xfb,0x93,0xc0]
; CHECK-NEXT: addl %ecx, %eax ## encoding: [0x01,0xc8]
[CodeGen] Use MachineOperand::print in the MIRPrinter for MO_Register. Work towards the unification of MIR and debug output by refactoring the interfaces. For MachineOperand::print, keep a simple version that can be easily called from `dump()`, and a more complex one which will be called from both the MIRPrinter and MachineInstr::print. Add extra checks inside MachineOperand for detached operands (operands with getParent() == nullptr). https://reviews.llvm.org/D40836 * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/kill: ([^ ]+) ([^ ]+)<def> ([^ ]+)/kill: \1 def \2 \3/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/kill: ([^ ]+) ([^ ]+) ([^ ]+)<def>/kill: \1 \2 def \3/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/kill: def ([^ ]+) ([^ ]+) ([^ ]+)<def>/kill: def \1 \2 def \3/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/<def>//g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<kill>/killed \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-use,kill>/implicit killed \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<dead>/dead \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<def[ ]*,[ ]*dead>/dead \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-def[ ]*,[ ]*dead>/implicit-def dead \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-def>/implicit-def \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-use>/implicit \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<internal>/internal \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<undef>/undef \1/g' llvm-svn: 320022
2017-12-07 18:40:31 +08:00
; CHECK-NEXT: ## kill: def %ax killed %ax killed %eax
[X86] test/testn intrinsics lowering to IR. llvm part. Remove builtins from llvm and add AutoUpgrade support. Also add fast-isel tests for the TEST and TESTN instructions. Differential Revision: https://reviews.llvm.org/D38736 llvm-svn: 318036
2017-11-13 20:51:18 +08:00
; CHECK-NEXT: vzeroupper ## encoding: [0xc5,0xf8,0x77]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call i16 @llvm.x86.avx512.ptestm.w.256(<16 x i16> %x0, <16 x i16> %x1, i16 %x2)
%res1 = call i16 @llvm.x86.avx512.ptestm.w.256(<16 x i16> %x0, <16 x i16> %x1, i16-1)
%res2 = add i16 %res, %res1
ret i16 %res2
}
declare i16 @llvm.x86.avx512.ptestnm.b.128(<16 x i8>, <16 x i8>, i16)
define i16@test_int_x86_avx512_ptestnm_b_128(<16 x i8> %x0, <16 x i8> %x1, i16 %x2) {
; CHECK-LABEL: test_int_x86_avx512_ptestnm_b_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[X86] test/testn intrinsics lowering to IR. llvm part. Remove builtins from llvm and add AutoUpgrade support. Also add fast-isel tests for the TEST and TESTN instructions. Differential Revision: https://reviews.llvm.org/D38736 llvm-svn: 318036
2017-11-13 20:51:18 +08:00
; CHECK-NEXT: vptestnmb %xmm1, %xmm0, %k0 ## encoding: [0x62,0xf2,0x7e,0x08,0x26,0xc1]
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
; CHECK-NEXT: vptestnmb %xmm1, %xmm0, %k1 {%k1} ## encoding: [0x62,0xf2,0x7e,0x09,0x26,0xc9]
; CHECK-NEXT: kmovd %k1, %ecx ## encoding: [0xc5,0xfb,0x93,0xc9]
; CHECK-NEXT: kmovd %k0, %eax ## encoding: [0xc5,0xfb,0x93,0xc0]
; CHECK-NEXT: addl %ecx, %eax ## encoding: [0x01,0xc8]
[CodeGen] Use MachineOperand::print in the MIRPrinter for MO_Register. Work towards the unification of MIR and debug output by refactoring the interfaces. For MachineOperand::print, keep a simple version that can be easily called from `dump()`, and a more complex one which will be called from both the MIRPrinter and MachineInstr::print. Add extra checks inside MachineOperand for detached operands (operands with getParent() == nullptr). https://reviews.llvm.org/D40836 * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/kill: ([^ ]+) ([^ ]+)<def> ([^ ]+)/kill: \1 def \2 \3/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/kill: ([^ ]+) ([^ ]+) ([^ ]+)<def>/kill: \1 \2 def \3/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/kill: def ([^ ]+) ([^ ]+) ([^ ]+)<def>/kill: def \1 \2 def \3/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/<def>//g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<kill>/killed \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-use,kill>/implicit killed \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<dead>/dead \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<def[ ]*,[ ]*dead>/dead \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-def[ ]*,[ ]*dead>/implicit-def dead \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-def>/implicit-def \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-use>/implicit \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<internal>/internal \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<undef>/undef \1/g' llvm-svn: 320022
2017-12-07 18:40:31 +08:00
; CHECK-NEXT: ## kill: def %ax killed %ax killed %eax
[X86] test/testn intrinsics lowering to IR. llvm part. Remove builtins from llvm and add AutoUpgrade support. Also add fast-isel tests for the TEST and TESTN instructions. Differential Revision: https://reviews.llvm.org/D38736 llvm-svn: 318036
2017-11-13 20:51:18 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call i16 @llvm.x86.avx512.ptestnm.b.128(<16 x i8> %x0, <16 x i8> %x1, i16 %x2)
%res1 = call i16 @llvm.x86.avx512.ptestnm.b.128(<16 x i8> %x0, <16 x i8> %x1, i16-1)
%res2 = add i16 %res, %res1
ret i16 %res2
}
declare i32 @llvm.x86.avx512.ptestnm.b.256(<32 x i8>, <32 x i8>, i32)
define i32@test_int_x86_avx512_ptestnm_b_256(<32 x i8> %x0, <32 x i8> %x1, i32 %x2) {
; CHECK-LABEL: test_int_x86_avx512_ptestnm_b_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[X86] test/testn intrinsics lowering to IR. llvm part. Remove builtins from llvm and add AutoUpgrade support. Also add fast-isel tests for the TEST and TESTN instructions. Differential Revision: https://reviews.llvm.org/D38736 llvm-svn: 318036
2017-11-13 20:51:18 +08:00
; CHECK-NEXT: vptestnmb %ymm1, %ymm0, %k0 ## encoding: [0x62,0xf2,0x7e,0x28,0x26,0xc1]
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
; CHECK-NEXT: vptestnmb %ymm1, %ymm0, %k1 {%k1} ## encoding: [0x62,0xf2,0x7e,0x29,0x26,0xc9]
; CHECK-NEXT: kmovd %k1, %ecx ## encoding: [0xc5,0xfb,0x93,0xc9]
; CHECK-NEXT: kmovd %k0, %eax ## encoding: [0xc5,0xfb,0x93,0xc0]
; CHECK-NEXT: addl %ecx, %eax ## encoding: [0x01,0xc8]
; CHECK-NEXT: vzeroupper ## encoding: [0xc5,0xf8,0x77]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call i32 @llvm.x86.avx512.ptestnm.b.256(<32 x i8> %x0, <32 x i8> %x1, i32 %x2)
%res1 = call i32 @llvm.x86.avx512.ptestnm.b.256(<32 x i8> %x0, <32 x i8> %x1, i32-1)
%res2 = add i32 %res, %res1
ret i32 %res2
}
declare i8 @llvm.x86.avx512.ptestnm.w.128(<8 x i16>, <8 x i16>, i8 %x2)
define i8@test_int_x86_avx512_ptestnm_w_128(<8 x i16> %x0, <8 x i16> %x1, i8 %x2) {
; CHECK-LABEL: test_int_x86_avx512_ptestnm_w_128:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[X86] test/testn intrinsics lowering to IR. llvm part. Remove builtins from llvm and add AutoUpgrade support. Also add fast-isel tests for the TEST and TESTN instructions. Differential Revision: https://reviews.llvm.org/D38736 llvm-svn: 318036
2017-11-13 20:51:18 +08:00
; CHECK-NEXT: vptestnmw %xmm1, %xmm0, %k0 ## encoding: [0x62,0xf2,0xfe,0x08,0x26,0xc1]
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
; CHECK-NEXT: vptestnmw %xmm1, %xmm0, %k1 {%k1} ## encoding: [0x62,0xf2,0xfe,0x09,0x26,0xc9]
; CHECK-NEXT: kmovd %k1, %ecx ## encoding: [0xc5,0xfb,0x93,0xc9]
; CHECK-NEXT: kmovd %k0, %eax ## encoding: [0xc5,0xfb,0x93,0xc0]
; CHECK-NEXT: addb %cl, %al ## encoding: [0x00,0xc8]
[CodeGen] Use MachineOperand::print in the MIRPrinter for MO_Register. Work towards the unification of MIR and debug output by refactoring the interfaces. For MachineOperand::print, keep a simple version that can be easily called from `dump()`, and a more complex one which will be called from both the MIRPrinter and MachineInstr::print. Add extra checks inside MachineOperand for detached operands (operands with getParent() == nullptr). https://reviews.llvm.org/D40836 * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/kill: ([^ ]+) ([^ ]+)<def> ([^ ]+)/kill: \1 def \2 \3/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/kill: ([^ ]+) ([^ ]+) ([^ ]+)<def>/kill: \1 \2 def \3/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/kill: def ([^ ]+) ([^ ]+) ([^ ]+)<def>/kill: def \1 \2 def \3/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/<def>//g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<kill>/killed \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-use,kill>/implicit killed \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<dead>/dead \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<def[ ]*,[ ]*dead>/dead \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-def[ ]*,[ ]*dead>/implicit-def dead \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-def>/implicit-def \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-use>/implicit \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<internal>/internal \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<undef>/undef \1/g' llvm-svn: 320022
2017-12-07 18:40:31 +08:00
; CHECK-NEXT: ## kill: def %al killed %al killed %eax
[X86] test/testn intrinsics lowering to IR. llvm part. Remove builtins from llvm and add AutoUpgrade support. Also add fast-isel tests for the TEST and TESTN instructions. Differential Revision: https://reviews.llvm.org/D38736 llvm-svn: 318036
2017-11-13 20:51:18 +08:00
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call i8 @llvm.x86.avx512.ptestnm.w.128(<8 x i16> %x0, <8 x i16> %x1, i8 %x2)
%res1 = call i8 @llvm.x86.avx512.ptestnm.w.128(<8 x i16> %x0, <8 x i16> %x1, i8-1)
%res2 = add i8 %res, %res1
ret i8 %res2
}
declare i16 @llvm.x86.avx512.ptestnm.w.256(<16 x i16>, <16 x i16>, i16 %x2)
define i16@test_int_x86_avx512_ptestnm_w_256(<16 x i16> %x0, <16 x i16> %x1, i16 %x2) {
; CHECK-LABEL: test_int_x86_avx512_ptestnm_w_256:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; CHECK: ## %bb.0:
[X86] test/testn intrinsics lowering to IR. llvm part. Remove builtins from llvm and add AutoUpgrade support. Also add fast-isel tests for the TEST and TESTN instructions. Differential Revision: https://reviews.llvm.org/D38736 llvm-svn: 318036
2017-11-13 20:51:18 +08:00
; CHECK-NEXT: vptestnmw %ymm1, %ymm0, %k0 ## encoding: [0x62,0xf2,0xfe,0x28,0x26,0xc1]
; CHECK-NEXT: kmovd %edi, %k1 ## encoding: [0xc5,0xfb,0x92,0xcf]
; CHECK-NEXT: vptestnmw %ymm1, %ymm0, %k1 {%k1} ## encoding: [0x62,0xf2,0xfe,0x29,0x26,0xc9]
; CHECK-NEXT: kmovd %k1, %ecx ## encoding: [0xc5,0xfb,0x93,0xc9]
; CHECK-NEXT: kmovd %k0, %eax ## encoding: [0xc5,0xfb,0x93,0xc0]
; CHECK-NEXT: addl %ecx, %eax ## encoding: [0x01,0xc8]
[CodeGen] Use MachineOperand::print in the MIRPrinter for MO_Register. Work towards the unification of MIR and debug output by refactoring the interfaces. For MachineOperand::print, keep a simple version that can be easily called from `dump()`, and a more complex one which will be called from both the MIRPrinter and MachineInstr::print. Add extra checks inside MachineOperand for detached operands (operands with getParent() == nullptr). https://reviews.llvm.org/D40836 * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/kill: ([^ ]+) ([^ ]+)<def> ([^ ]+)/kill: \1 def \2 \3/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/kill: ([^ ]+) ([^ ]+) ([^ ]+)<def>/kill: \1 \2 def \3/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/kill: def ([^ ]+) ([^ ]+) ([^ ]+)<def>/kill: def \1 \2 def \3/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/<def>//g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<kill>/killed \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-use,kill>/implicit killed \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<dead>/dead \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<def[ ]*,[ ]*dead>/dead \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-def[ ]*,[ ]*dead>/implicit-def dead \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-def>/implicit-def \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<imp-use>/implicit \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<internal>/internal \1/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" -o -name "*.s" \) -type f -print0 | xargs -0 sed -i '' -E 's/([^ ]+)<undef>/undef \1/g' llvm-svn: 320022
2017-12-07 18:40:31 +08:00
; CHECK-NEXT: ## kill: def %ax killed %ax killed %eax
[X86] test/testn intrinsics lowering to IR. llvm part. Remove builtins from llvm and add AutoUpgrade support. Also add fast-isel tests for the TEST and TESTN instructions. Differential Revision: https://reviews.llvm.org/D38736 llvm-svn: 318036
2017-11-13 20:51:18 +08:00
; CHECK-NEXT: vzeroupper ## encoding: [0xc5,0xf8,0x77]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call i16 @llvm.x86.avx512.ptestnm.w.256(<16 x i16> %x0, <16 x i16> %x1, i16 %x2)
%res1 = call i16 @llvm.x86.avx512.ptestnm.w.256(<16 x i16> %x0, <16 x i16> %x1, i16-1)
%res2 = add i16 %res, %res1
ret i16 %res2
}