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/avx512-insert-extract.ll

2256 lines
86 KiB
LLVM
Raw Normal View History

VirtRegMap: Replace some identity copies with KILL instructions. An identity COPY like this: %AL = COPY %AL, %EAX<imp-def> has no semantic effect, but encodes liveness information: Further users of %EAX only depend on this instruction even though it does not define the full register. Replace the COPY with a KILL instruction in those cases to maintain this liveness information. (This reverts a small part of r238588 but this time adds a comment explaining why a KILL instruction is useful). llvm-svn: 274952
2016-07-09 08:19:07 +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 | FileCheck --check-prefix=CHECK --check-prefix=KNL %s
; RUN: llc < %s -mtriple=x86_64-apple-darwin -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq | FileCheck --check-prefix=CHECK --check-prefix=SKX --check-prefix=SKX_ONLY %s
; RUN: llc < %s -mtriple=x86_64-apple-darwin -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq,+avx512vbmi | FileCheck --check-prefix=CHECK --check-prefix=SKX --check-prefix=SKX_VBMI %s
Added INSERT and EXTRACT intructions from AVX-512 ISA. All insertf*/extractf* functions replaced with insert/extract since we have insertf and inserti forms. Added lowering for INSERT_VECTOR_ELT / EXTRACT_VECTOR_ELT for 512-bit vectors. Added lowering for EXTRACT/INSERT subvector for 512-bit vectors. Added a test. llvm-svn: 187491
2013-07-31 19:35:14 +08:00
define <16 x float> @test1(<16 x float> %x, float* %br, float %y) nounwind {
[AVX512] Add a common prefix to avx512-insert-extract.ll so we can reduce the number of check lines on some test cases. This was pointed out during the review for D313804. llvm-svn: 309629
2017-08-01 05:20:06 +08:00
; CHECK-LABEL: test1:
[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] Add a common prefix to avx512-insert-extract.ll so we can reduce the number of check lines on some test cases. This was pointed out during the review for D313804. llvm-svn: 309629
2017-08-01 05:20:06 +08:00
; CHECK-NEXT: vinsertps {{.*#+}} xmm2 = xmm0[0],mem[0],xmm0[2,3]
; CHECK-NEXT: vinsertf32x4 $0, %xmm2, %zmm0, %zmm2
; CHECK-NEXT: vextractf32x4 $3, %zmm0, %xmm0
; CHECK-NEXT: vinsertps {{.*#+}} xmm0 = xmm0[0,1],xmm1[0],xmm0[3]
; CHECK-NEXT: vinsertf32x4 $3, %xmm0, %zmm2, %zmm0
; CHECK-NEXT: retq
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 llvm-svn: 230794
2015-02-28 05:17:42 +08:00
%rrr = load float, float* %br
Added INSERT and EXTRACT intructions from AVX-512 ISA. All insertf*/extractf* functions replaced with insert/extract since we have insertf and inserti forms. Added lowering for INSERT_VECTOR_ELT / EXTRACT_VECTOR_ELT for 512-bit vectors. Added lowering for EXTRACT/INSERT subvector for 512-bit vectors. Added a test. llvm-svn: 187491
2013-07-31 19:35:14 +08:00
%rrr2 = insertelement <16 x float> %x, float %rrr, i32 1
%rrr3 = insertelement <16 x float> %rrr2, float %y, i32 14
ret <16 x float> %rrr3
}
define <8 x double> @test2(<8 x double> %x, double* %br, double %y) nounwind {
[AVX512] Don't switch unmasked subvector insert/extract instructions when AVX512DQI is enabled. There's no reason to switch instructions with and without DQI. It just creates extra isel patterns and test divergences. There is however value in enabling the masked version of the instructions with DQI. This required introducing some new multiclasses to enabling this splitting. Differential Revision: https://reviews.llvm.org/D36661 llvm-svn: 311091
2017-08-17 23:40:25 +08:00
; CHECK-LABEL: test2:
[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] Don't switch unmasked subvector insert/extract instructions when AVX512DQI is enabled. There's no reason to switch instructions with and without DQI. It just creates extra isel patterns and test divergences. There is however value in enabling the masked version of the instructions with DQI. This required introducing some new multiclasses to enabling this splitting. Differential Revision: https://reviews.llvm.org/D36661 llvm-svn: 311091
2017-08-17 23:40:25 +08:00
; CHECK-NEXT: vmovhpd {{.*#+}} xmm2 = xmm0[0],mem[0]
; CHECK-NEXT: vinsertf32x4 $0, %xmm2, %zmm0, %zmm2
; CHECK-NEXT: vextractf32x4 $3, %zmm0, %xmm0
[X86] Prefer blendi over movss/sd when avx512 is enabled unless optimizing for size. AVX512 doesn't have an immediate controlled blend instruction. But blend throughput is still better than movss/sd on SKX. This commit changes AVX512 to use the AVX blend instructions instead of MOVSS/MOVSD. This constrains the register allocation since it won't be able to use XMM16-31, but hopefully the increased throughput and reduced port 5 pressure makes up for that. llvm-svn: 337083
2018-07-14 10:05:08 +08:00
; CHECK-NEXT: vblendpd {{.*#+}} xmm0 = xmm1[0],xmm0[1]
[AVX512] Don't switch unmasked subvector insert/extract instructions when AVX512DQI is enabled. There's no reason to switch instructions with and without DQI. It just creates extra isel patterns and test divergences. There is however value in enabling the masked version of the instructions with DQI. This required introducing some new multiclasses to enabling this splitting. Differential Revision: https://reviews.llvm.org/D36661 llvm-svn: 311091
2017-08-17 23:40:25 +08:00
; CHECK-NEXT: vinsertf32x4 $3, %xmm0, %zmm2, %zmm0
; CHECK-NEXT: retq
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 llvm-svn: 230794
2015-02-28 05:17:42 +08:00
%rrr = load double, double* %br
Added INSERT and EXTRACT intructions from AVX-512 ISA. All insertf*/extractf* functions replaced with insert/extract since we have insertf and inserti forms. Added lowering for INSERT_VECTOR_ELT / EXTRACT_VECTOR_ELT for 512-bit vectors. Added lowering for EXTRACT/INSERT subvector for 512-bit vectors. Added a test. llvm-svn: 187491
2013-07-31 19:35:14 +08:00
%rrr2 = insertelement <8 x double> %x, double %rrr, i32 1
%rrr3 = insertelement <8 x double> %rrr2, double %y, i32 6
ret <8 x double> %rrr3
}
define <16 x float> @test3(<16 x float> %x) nounwind {
[AVX512] Add a common prefix to avx512-insert-extract.ll so we can reduce the number of check lines on some test cases. This was pointed out during the review for D313804. llvm-svn: 309629
2017-08-01 05:20:06 +08:00
; CHECK-LABEL: test3:
[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] Use 256-bit extract instructions for extracting bits [255:128] from a 512-bit register This enables the use of a smaller encoding by using a VEX instruction when possible. Differential Revision: https://reviews.llvm.org/D37092 llvm-svn: 312100
2017-08-30 15:26:12 +08:00
; CHECK-NEXT: vextractf128 $1, %ymm0, %xmm1
[AVX512] Add a common prefix to avx512-insert-extract.ll so we can reduce the number of check lines on some test cases. This was pointed out during the review for D313804. llvm-svn: 309629
2017-08-01 05:20:06 +08:00
; CHECK-NEXT: vinsertps {{.*#+}} xmm1 = xmm0[0],xmm1[0],xmm0[2,3]
; CHECK-NEXT: vinsertf32x4 $0, %xmm1, %zmm0, %zmm0
; CHECK-NEXT: retq
Added INSERT and EXTRACT intructions from AVX-512 ISA. All insertf*/extractf* functions replaced with insert/extract since we have insertf and inserti forms. Added lowering for INSERT_VECTOR_ELT / EXTRACT_VECTOR_ELT for 512-bit vectors. Added lowering for EXTRACT/INSERT subvector for 512-bit vectors. Added a test. llvm-svn: 187491
2013-07-31 19:35:14 +08:00
%eee = extractelement <16 x float> %x, i32 4
%rrr2 = insertelement <16 x float> %x, float %eee, i32 1
ret <16 x float> %rrr2
}
define <8 x i64> @test4(<8 x i64> %x) nounwind {
[AVX512] Don't switch unmasked subvector insert/extract instructions when AVX512DQI is enabled. There's no reason to switch instructions with and without DQI. It just creates extra isel patterns and test divergences. There is however value in enabling the masked version of the instructions with DQI. This required introducing some new multiclasses to enabling this splitting. Differential Revision: https://reviews.llvm.org/D36661 llvm-svn: 311091
2017-08-17 23:40:25 +08:00
; CHECK-LABEL: test4:
[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] Don't switch unmasked subvector insert/extract instructions when AVX512DQI is enabled. There's no reason to switch instructions with and without DQI. It just creates extra isel patterns and test divergences. There is however value in enabling the masked version of the instructions with DQI. This required introducing some new multiclasses to enabling this splitting. Differential Revision: https://reviews.llvm.org/D36661 llvm-svn: 311091
2017-08-17 23:40:25 +08:00
; CHECK-NEXT: vextracti32x4 $2, %zmm0, %xmm1
; CHECK-NEXT: vmovq %xmm1, %rax
; CHECK-NEXT: vpinsrq $1, %rax, %xmm0, %xmm1
; CHECK-NEXT: vinserti32x4 $0, %xmm1, %zmm0, %zmm0
; CHECK-NEXT: retq
Added INSERT and EXTRACT intructions from AVX-512 ISA. All insertf*/extractf* functions replaced with insert/extract since we have insertf and inserti forms. Added lowering for INSERT_VECTOR_ELT / EXTRACT_VECTOR_ELT for 512-bit vectors. Added lowering for EXTRACT/INSERT subvector for 512-bit vectors. Added a test. llvm-svn: 187491
2013-07-31 19:35:14 +08:00
%eee = extractelement <8 x i64> %x, i32 4
%rrr2 = insertelement <8 x i64> %x, i64 %eee, i32 1
ret <8 x i64> %rrr2
AVX-512 set: added VEXTRACTPS instruction llvm-svn: 187705
2013-08-04 18:46:07 +08:00
}
define i32 @test5(<4 x float> %x) nounwind {
[AVX512] Add a common prefix to avx512-insert-extract.ll so we can reduce the number of check lines on some test cases. This was pointed out during the review for D313804. llvm-svn: 309629
2017-08-01 05:20:06 +08:00
; CHECK-LABEL: test5:
[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] Add a common prefix to avx512-insert-extract.ll so we can reduce the number of check lines on some test cases. This was pointed out during the review for D313804. llvm-svn: 309629
2017-08-01 05:20:06 +08:00
; CHECK-NEXT: vextractps $3, %xmm0, %eax
; CHECK-NEXT: retq
AVX-512 set: added VEXTRACTPS instruction llvm-svn: 187705
2013-08-04 18:46:07 +08:00
%ef = extractelement <4 x float> %x, i32 3
%ei = bitcast float %ef to i32
ret i32 %ei
}
define void @test6(<4 x float> %x, float* %out) nounwind {
[AVX512] Add a common prefix to avx512-insert-extract.ll so we can reduce the number of check lines on some test cases. This was pointed out during the review for D313804. llvm-svn: 309629
2017-08-01 05:20:06 +08:00
; CHECK-LABEL: test6:
[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] Add a common prefix to avx512-insert-extract.ll so we can reduce the number of check lines on some test cases. This was pointed out during the review for D313804. llvm-svn: 309629
2017-08-01 05:20:06 +08:00
; CHECK-NEXT: vextractps $3, %xmm0, (%rdi)
; CHECK-NEXT: retq
AVX-512 set: added VEXTRACTPS instruction llvm-svn: 187705
2013-08-04 18:46:07 +08:00
%ef = extractelement <4 x float> %x, i32 3
store float %ef, float* %out, align 4
ret void
}
AVX-512: implemented extractelement with variable index. Added parsing of mask register and "zeroing" semantic, like {%k1} {z}. llvm-svn: 190595
2013-09-12 16:55:00 +08:00
define float @test7(<16 x float> %x, i32 %ind) nounwind {
[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-LABEL: test7:
[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][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: pushq %rbp
; CHECK-NEXT: movq %rsp, %rbp
; CHECK-NEXT: andq $-64, %rsp
; CHECK-NEXT: subq $128, %rsp
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; CHECK-NEXT: ## kill: def $edi killed $edi def $rdi
[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: vmovaps %zmm0, (%rsp)
; CHECK-NEXT: andl $15, %edi
; CHECK-NEXT: vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero
; CHECK-NEXT: movq %rbp, %rsp
; CHECK-NEXT: popq %rbp
; CHECK-NEXT: vzeroupper
; CHECK-NEXT: retq
AVX-512: implemented extractelement with variable index. Added parsing of mask register and "zeroing" semantic, like {%k1} {z}. llvm-svn: 190595
2013-09-12 16:55:00 +08:00
%e = extractelement <16 x float> %x, i32 %ind
ret float %e
}
define double @test8(<8 x double> %x, i32 %ind) nounwind {
[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-LABEL: test8:
[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][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: pushq %rbp
; CHECK-NEXT: movq %rsp, %rbp
; CHECK-NEXT: andq $-64, %rsp
; CHECK-NEXT: subq $128, %rsp
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; CHECK-NEXT: ## kill: def $edi killed $edi def $rdi
[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: vmovaps %zmm0, (%rsp)
; CHECK-NEXT: andl $7, %edi
; CHECK-NEXT: vmovsd {{.*#+}} xmm0 = mem[0],zero
; CHECK-NEXT: movq %rbp, %rsp
; CHECK-NEXT: popq %rbp
; CHECK-NEXT: vzeroupper
; CHECK-NEXT: retq
AVX-512: implemented extractelement with variable index. Added parsing of mask register and "zeroing" semantic, like {%k1} {z}. llvm-svn: 190595
2013-09-12 16:55:00 +08:00
%e = extractelement <8 x double> %x, i32 %ind
ret double %e
}
define float @test9(<8 x float> %x, i32 %ind) nounwind {
[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-LABEL: test9:
[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][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: pushq %rbp
; CHECK-NEXT: movq %rsp, %rbp
; CHECK-NEXT: andq $-32, %rsp
; CHECK-NEXT: subq $64, %rsp
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; CHECK-NEXT: ## kill: def $edi killed $edi def $rdi
[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: vmovaps %ymm0, (%rsp)
; CHECK-NEXT: andl $7, %edi
; CHECK-NEXT: vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero
; CHECK-NEXT: movq %rbp, %rsp
; CHECK-NEXT: popq %rbp
; CHECK-NEXT: vzeroupper
; CHECK-NEXT: retq
AVX-512: implemented extractelement with variable index. Added parsing of mask register and "zeroing" semantic, like {%k1} {z}. llvm-svn: 190595
2013-09-12 16:55:00 +08:00
%e = extractelement <8 x float> %x, i32 %ind
ret float %e
}
define i32 @test10(<16 x i32> %x, i32 %ind) nounwind {
[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-LABEL: test10:
[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][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: pushq %rbp
; CHECK-NEXT: movq %rsp, %rbp
; CHECK-NEXT: andq $-64, %rsp
; CHECK-NEXT: subq $128, %rsp
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; CHECK-NEXT: ## kill: def $edi killed $edi def $rdi
[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: vmovaps %zmm0, (%rsp)
; CHECK-NEXT: andl $15, %edi
; CHECK-NEXT: movl (%rsp,%rdi,4), %eax
; CHECK-NEXT: movq %rbp, %rsp
; CHECK-NEXT: popq %rbp
; CHECK-NEXT: vzeroupper
; CHECK-NEXT: retq
AVX-512: implemented extractelement with variable index. Added parsing of mask register and "zeroing" semantic, like {%k1} {z}. llvm-svn: 190595
2013-09-12 16:55:00 +08:00
%e = extractelement <16 x i32> %x, i32 %ind
ret i32 %e
}
AVX-512: Handled extractelement from mask vector; Added VMOSHDUP/VMOVSLDUP shuffle instructions. llvm-svn: 194691
2013-11-14 19:29:27 +08:00
define <16 x i32> @test11(<16 x i32>%a, <16 x i32>%b) {
VirtRegMap: Replace some identity copies with KILL instructions. An identity COPY like this: %AL = COPY %AL, %EAX<imp-def> has no semantic effect, but encodes liveness information: Further users of %EAX only depend on this instruction even though it does not define the full register. Replace the COPY with a KILL instruction in those cases to maintain this liveness information. (This reverts a small part of r238588 but this time adds a comment explaining why a KILL instruction is useful). llvm-svn: 274952
2016-07-09 08:19:07 +08:00
; KNL-LABEL: test11:
[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
; KNL: ## %bb.0:
VirtRegMap: Replace some identity copies with KILL instructions. An identity COPY like this: %AL = COPY %AL, %EAX<imp-def> has no semantic effect, but encodes liveness information: Further users of %EAX only depend on this instruction even though it does not define the full register. Replace the COPY with a KILL instruction in those cases to maintain this liveness information. (This reverts a small part of r238588 but this time adds a comment explaining why a KILL instruction is useful). llvm-svn: 274952
2016-07-09 08:19:07 +08:00
; KNL-NEXT: vpcmpltud %zmm1, %zmm0, %k0
[X86] Don't zero the upper bits of the k-register before extracting a single bit from a vXi1. This doesn't match the semantics of the extract_vector_elt operation. Nothing downstream knows the bits were zeroed so they still get masked or sign extended after the extrat anyway. llvm-svn: 320723
2017-12-15 02:35:25 +08:00
; KNL-NEXT: kshiftrw $4, %k0, %k0
revert r279960. https://llvm.org/bugs/show_bug.cgi?id=30249 llvm-svn: 280625
2016-09-04 22:03:52 +08:00
; KNL-NEXT: kmovw %k0, %eax
[X86][AVX512] Make i1 illegal in the CodeGen This patch defines the i1 type as illegal in the X86 backend for AVX512. For DAG operations on <N x i1> types (build vector, extract vector element, ...) i8 is used, and should be truncated/extended. This should produce better scalar code for i1 types since GPRs will be used instead of mask registers. Differential Revision: https://reviews.llvm.org/D32273 llvm-svn: 303421
2017-05-19 20:35:15 +08:00
; KNL-NEXT: testb $1, %al
VirtRegMap: Replace some identity copies with KILL instructions. An identity COPY like this: %AL = COPY %AL, %EAX<imp-def> has no semantic effect, but encodes liveness information: Further users of %EAX only depend on this instruction even though it does not define the full register. Replace the COPY with a KILL instruction in those cases to maintain this liveness information. (This reverts a small part of r238588 but this time adds a comment explaining why a KILL instruction is useful). llvm-svn: 274952
2016-07-09 08:19:07 +08:00
; KNL-NEXT: je LBB10_2
[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
; KNL-NEXT: ## %bb.1: ## %A
[AVX512] Add initial support for the Execution Domain fixing pass to change some EVEX instructions. llvm-svn: 276393
2016-07-22 13:00:52 +08:00
; KNL-NEXT: vmovdqa64 %zmm1, %zmm0
VirtRegMap: Replace some identity copies with KILL instructions. An identity COPY like this: %AL = COPY %AL, %EAX<imp-def> has no semantic effect, but encodes liveness information: Further users of %EAX only depend on this instruction even though it does not define the full register. Replace the COPY with a KILL instruction in those cases to maintain this liveness information. (This reverts a small part of r238588 but this time adds a comment explaining why a KILL instruction is useful). llvm-svn: 274952
2016-07-09 08:19:07 +08:00
; KNL-NEXT: retq
; KNL-NEXT: LBB10_2: ## %B
; KNL-NEXT: vpaddd %zmm0, %zmm1, %zmm0
; KNL-NEXT: retq
;
; SKX-LABEL: test11:
[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
; SKX: ## %bb.0:
VirtRegMap: Replace some identity copies with KILL instructions. An identity COPY like this: %AL = COPY %AL, %EAX<imp-def> has no semantic effect, but encodes liveness information: Further users of %EAX only depend on this instruction even though it does not define the full register. Replace the COPY with a KILL instruction in those cases to maintain this liveness information. (This reverts a small part of r238588 but this time adds a comment explaining why a KILL instruction is useful). llvm-svn: 274952
2016-07-09 08:19:07 +08:00
; SKX-NEXT: vpcmpltud %zmm1, %zmm0, %k0
[X86] Don't zero the upper bits of the k-register before extracting a single bit from a vXi1. This doesn't match the semantics of the extract_vector_elt operation. Nothing downstream knows the bits were zeroed so they still get masked or sign extended after the extrat anyway. llvm-svn: 320723
2017-12-15 02:35:25 +08:00
; SKX-NEXT: kshiftrw $4, %k0, %k0
[AVX-512] Remove explicit KMOVWrk/KMOVWKr instructions from patterns where we can just use COPY_TO_REGCLASS instead. This will result in a KMOVW or KMOVD being emitted during register allocation. And in at least some cases this might allow the register coalescer to remove the copy all together. llvm-svn: 298984
2017-03-29 14:55:28 +08:00
; SKX-NEXT: kmovd %k0, %eax
[X86][AVX512] Make i1 illegal in the CodeGen This patch defines the i1 type as illegal in the X86 backend for AVX512. For DAG operations on <N x i1> types (build vector, extract vector element, ...) i8 is used, and should be truncated/extended. This should produce better scalar code for i1 types since GPRs will be used instead of mask registers. Differential Revision: https://reviews.llvm.org/D32273 llvm-svn: 303421
2017-05-19 20:35:15 +08:00
; SKX-NEXT: testb $1, %al
VirtRegMap: Replace some identity copies with KILL instructions. An identity COPY like this: %AL = COPY %AL, %EAX<imp-def> has no semantic effect, but encodes liveness information: Further users of %EAX only depend on this instruction even though it does not define the full register. Replace the COPY with a KILL instruction in those cases to maintain this liveness information. (This reverts a small part of r238588 but this time adds a comment explaining why a KILL instruction is useful). llvm-svn: 274952
2016-07-09 08:19:07 +08:00
; SKX-NEXT: je LBB10_2
[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
; SKX-NEXT: ## %bb.1: ## %A
[AVX512] Add initial support for the Execution Domain fixing pass to change some EVEX instructions. llvm-svn: 276393
2016-07-22 13:00:52 +08:00
; SKX-NEXT: vmovdqa64 %zmm1, %zmm0
VirtRegMap: Replace some identity copies with KILL instructions. An identity COPY like this: %AL = COPY %AL, %EAX<imp-def> has no semantic effect, but encodes liveness information: Further users of %EAX only depend on this instruction even though it does not define the full register. Replace the COPY with a KILL instruction in those cases to maintain this liveness information. (This reverts a small part of r238588 but this time adds a comment explaining why a KILL instruction is useful). llvm-svn: 274952
2016-07-09 08:19:07 +08:00
; SKX-NEXT: retq
; SKX-NEXT: LBB10_2: ## %B
; SKX-NEXT: vpaddd %zmm0, %zmm1, %zmm0
; SKX-NEXT: retq
AVX-512: Handled extractelement from mask vector; Added VMOSHDUP/VMOVSLDUP shuffle instructions. llvm-svn: 194691
2013-11-14 19:29:27 +08:00
%cmp_res = icmp ult <16 x i32> %a, %b
%ia = extractelement <16 x i1> %cmp_res, i32 4
br i1 %ia, label %A, label %B
A:
ret <16 x i32>%b
B:
%c = add <16 x i32>%b, %a
AVX-512: Added legal type MVT::i1 and VK1 register for it. Added scalar compare VCMPSS, VCMPSD. Implemented LowerSELECT for scalar FP operations. I replaced FSETCCss, FSETCCsd with one node type FSETCCs. Node extract_vector_elt(v16i1/v8i1, idx) returns an element of type i1. llvm-svn: 197384
2013-12-16 21:52:35 +08:00
ret <16 x i32>%c
AVX-512: Handled extractelement from mask vector; Added VMOSHDUP/VMOVSLDUP shuffle instructions. llvm-svn: 194691
2013-11-14 19:29:27 +08:00
}
AVX-512: Fixed extract_vector_elt for v16i1 and v8i1 vectors. llvm-svn: 201066
2014-02-10 15:02:39 +08:00
define i64 @test12(<16 x i64>%a, <16 x i64>%b, i64 %a1, i64 %b1) {
VirtRegMap: Replace some identity copies with KILL instructions. An identity COPY like this: %AL = COPY %AL, %EAX<imp-def> has no semantic effect, but encodes liveness information: Further users of %EAX only depend on this instruction even though it does not define the full register. Replace the COPY with a KILL instruction in those cases to maintain this liveness information. (This reverts a small part of r238588 but this time adds a comment explaining why a KILL instruction is useful). llvm-svn: 274952
2016-07-09 08:19:07 +08:00
; KNL-LABEL: test12:
[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
; KNL: ## %bb.0:
VirtRegMap: Replace some identity copies with KILL instructions. An identity COPY like this: %AL = COPY %AL, %EAX<imp-def> has no semantic effect, but encodes liveness information: Further users of %EAX only depend on this instruction even though it does not define the full register. Replace the COPY with a KILL instruction in those cases to maintain this liveness information. (This reverts a small part of r238588 but this time adds a comment explaining why a KILL instruction is useful). llvm-svn: 274952
2016-07-09 08:19:07 +08:00
; KNL-NEXT: movq %rdi, %rax
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; KNL-NEXT: vpcmpgtq %zmm0, %zmm2, %k0
; KNL-NEXT: kmovw %k0, %ecx
; KNL-NEXT: testb $1, %cl
; KNL-NEXT: cmoveq %rsi, %rax
[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
; KNL-NEXT: vzeroupper
VirtRegMap: Replace some identity copies with KILL instructions. An identity COPY like this: %AL = COPY %AL, %EAX<imp-def> has no semantic effect, but encodes liveness information: Further users of %EAX only depend on this instruction even though it does not define the full register. Replace the COPY with a KILL instruction in those cases to maintain this liveness information. (This reverts a small part of r238588 but this time adds a comment explaining why a KILL instruction is useful). llvm-svn: 274952
2016-07-09 08:19:07 +08:00
; KNL-NEXT: retq
;
; SKX-LABEL: test12:
[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
; SKX: ## %bb.0:
VirtRegMap: Replace some identity copies with KILL instructions. An identity COPY like this: %AL = COPY %AL, %EAX<imp-def> has no semantic effect, but encodes liveness information: Further users of %EAX only depend on this instruction even though it does not define the full register. Replace the COPY with a KILL instruction in those cases to maintain this liveness information. (This reverts a small part of r238588 but this time adds a comment explaining why a KILL instruction is useful). llvm-svn: 274952
2016-07-09 08:19:07 +08:00
; SKX-NEXT: movq %rdi, %rax
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; SKX-NEXT: vpcmpgtq %zmm0, %zmm2, %k0
; SKX-NEXT: kmovd %k0, %ecx
; SKX-NEXT: testb $1, %cl
; SKX-NEXT: cmoveq %rsi, %rax
[X86] Generate VZEROUPPER for Skylake-avx512. VZEROUPPER should not be issued on Knights Landing (KNL), but on Skylake-avx512 it should be. Differential Revision: https://reviews.llvm.org/D29874 llvm-svn: 296859
2017-03-03 17:03:24 +08:00
; SKX-NEXT: vzeroupper
VirtRegMap: Replace some identity copies with KILL instructions. An identity COPY like this: %AL = COPY %AL, %EAX<imp-def> has no semantic effect, but encodes liveness information: Further users of %EAX only depend on this instruction even though it does not define the full register. Replace the COPY with a KILL instruction in those cases to maintain this liveness information. (This reverts a small part of r238588 but this time adds a comment explaining why a KILL instruction is useful). llvm-svn: 274952
2016-07-09 08:19:07 +08:00
; SKX-NEXT: retq
AVX-512: Fixed extract_vector_elt for v16i1 and v8i1 vectors. llvm-svn: 201066
2014-02-10 15:02:39 +08:00
%cmpvector_func.i = icmp slt <16 x i64> %a, %b
%extract24vector_func.i = extractelement <16 x i1> %cmpvector_func.i, i32 0
%res = select i1 %extract24vector_func.i, i64 %a1, i64 %b1
ret i64 %res
}
AVX-512: simpyfied BUILD_VECTOR for masks; fixed cmp/test sequence llvm-svn: 201487
2014-02-16 19:34:23 +08:00
define i16 @test13(i32 %a, i32 %b) {
VirtRegMap: Replace some identity copies with KILL instructions. An identity COPY like this: %AL = COPY %AL, %EAX<imp-def> has no semantic effect, but encodes liveness information: Further users of %EAX only depend on this instruction even though it does not define the full register. Replace the COPY with a KILL instruction in those cases to maintain this liveness information. (This reverts a small part of r238588 but this time adds a comment explaining why a KILL instruction is useful). llvm-svn: 274952
2016-07-09 08:19:07 +08:00
; KNL-LABEL: test13:
[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
; KNL: ## %bb.0:
VirtRegMap: Replace some identity copies with KILL instructions. An identity COPY like this: %AL = COPY %AL, %EAX<imp-def> has no semantic effect, but encodes liveness information: Further users of %EAX only depend on this instruction even though it does not define the full register. Replace the COPY with a KILL instruction in those cases to maintain this liveness information. (This reverts a small part of r238588 but this time adds a comment explaining why a KILL instruction is useful). llvm-svn: 274952
2016-07-09 08:19:07 +08:00
; KNL-NEXT: cmpl %esi, %edi
[x86] don't blindly transform SETB into SBB I noticed unnecessary 'sbb' instructions in D30472 and while looking at 'ptest' codegen recently. This happens because we were transforming any 'setb' - even when we only wanted a single-bit result. This patch moves those transforms under visitAdd/visitSub, so we we're only creating sbb/adc when it is a win. I don't know why we need a SETCC_CARRY node type, but I'm not proposing to change that existing behavior in this patch. Also, I'm skeptical that sbb/adc are a win for all micro-arches, so I added comments to the test files where this transform still fires. The test changes here are all cases where we no longer produce sbb/adc. Avoiding partial register stalls (generating an xor to clear a register) is not handled in some cases, but that's a separate issue. Differential Revision: https://reviews.llvm.org/D30611 llvm-svn: 297586
2017-03-13 02:28:48 +08:00
; KNL-NEXT: setb %al
[X86][AVX512] Make i1 illegal in the CodeGen This patch defines the i1 type as illegal in the X86 backend for AVX512. For DAG operations on <N x i1> types (build vector, extract vector element, ...) i8 is used, and should be truncated/extended. This should produce better scalar code for i1 types since GPRs will be used instead of mask registers. Differential Revision: https://reviews.llvm.org/D32273 llvm-svn: 303421
2017-05-19 20:35:15 +08:00
; KNL-NEXT: movw $-4, %cx
; KNL-NEXT: kmovw %ecx, %k0
; KNL-NEXT: kshiftrw $1, %k0, %k0
; KNL-NEXT: kshiftlw $1, %k0, %k0
Revert r274613 because it breaks the test suite with AVX512 This reverts most of r274613 (AKA r274626) and its follow-ups (r276347, r277289), due to miscompiles in the test suite. The FastISel change was left in, because it apparently fixes an unrelated issue. (Recommit of r279782 which was broken due to a bad merge.) This fixes 4 out of the 5 test failures in PR29112. llvm-svn: 279788
2016-08-26 06:48:11 +08:00
; KNL-NEXT: andl $1, %eax
VirtRegMap: Replace some identity copies with KILL instructions. An identity COPY like this: %AL = COPY %AL, %EAX<imp-def> has no semantic effect, but encodes liveness information: Further users of %EAX only depend on this instruction even though it does not define the full register. Replace the COPY with a KILL instruction in those cases to maintain this liveness information. (This reverts a small part of r238588 but this time adds a comment explaining why a KILL instruction is useful). llvm-svn: 274952
2016-07-09 08:19:07 +08:00
; KNL-NEXT: kmovw %eax, %k1
[X86][AVX512] Make i1 illegal in the CodeGen This patch defines the i1 type as illegal in the X86 backend for AVX512. For DAG operations on <N x i1> types (build vector, extract vector element, ...) i8 is used, and should be truncated/extended. This should produce better scalar code for i1 types since GPRs will be used instead of mask registers. Differential Revision: https://reviews.llvm.org/D32273 llvm-svn: 303421
2017-05-19 20:35:15 +08:00
; KNL-NEXT: korw %k1, %k0, %k0
VirtRegMap: Replace some identity copies with KILL instructions. An identity COPY like this: %AL = COPY %AL, %EAX<imp-def> has no semantic effect, but encodes liveness information: Further users of %EAX only depend on this instruction even though it does not define the full register. Replace the COPY with a KILL instruction in those cases to maintain this liveness information. (This reverts a small part of r238588 but this time adds a comment explaining why a KILL instruction is useful). llvm-svn: 274952
2016-07-09 08:19:07 +08:00
; KNL-NEXT: kmovw %k0, %eax
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; KNL-NEXT: ## kill: def $ax killed $ax killed $eax
VirtRegMap: Replace some identity copies with KILL instructions. An identity COPY like this: %AL = COPY %AL, %EAX<imp-def> has no semantic effect, but encodes liveness information: Further users of %EAX only depend on this instruction even though it does not define the full register. Replace the COPY with a KILL instruction in those cases to maintain this liveness information. (This reverts a small part of r238588 but this time adds a comment explaining why a KILL instruction is useful). llvm-svn: 274952
2016-07-09 08:19:07 +08:00
; KNL-NEXT: retq
;
; SKX-LABEL: test13:
[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
; SKX: ## %bb.0:
VirtRegMap: Replace some identity copies with KILL instructions. An identity COPY like this: %AL = COPY %AL, %EAX<imp-def> has no semantic effect, but encodes liveness information: Further users of %EAX only depend on this instruction even though it does not define the full register. Replace the COPY with a KILL instruction in those cases to maintain this liveness information. (This reverts a small part of r238588 but this time adds a comment explaining why a KILL instruction is useful). llvm-svn: 274952
2016-07-09 08:19:07 +08:00
; SKX-NEXT: cmpl %esi, %edi
[x86] don't blindly transform SETB into SBB I noticed unnecessary 'sbb' instructions in D30472 and while looking at 'ptest' codegen recently. This happens because we were transforming any 'setb' - even when we only wanted a single-bit result. This patch moves those transforms under visitAdd/visitSub, so we we're only creating sbb/adc when it is a win. I don't know why we need a SETCC_CARRY node type, but I'm not proposing to change that existing behavior in this patch. Also, I'm skeptical that sbb/adc are a win for all micro-arches, so I added comments to the test files where this transform still fires. The test changes here are all cases where we no longer produce sbb/adc. Avoiding partial register stalls (generating an xor to clear a register) is not handled in some cases, but that's a separate issue. Differential Revision: https://reviews.llvm.org/D30611 llvm-svn: 297586
2017-03-13 02:28:48 +08:00
; SKX-NEXT: setb %al
[X86][AVX512] Make i1 illegal in the CodeGen This patch defines the i1 type as illegal in the X86 backend for AVX512. For DAG operations on <N x i1> types (build vector, extract vector element, ...) i8 is used, and should be truncated/extended. This should produce better scalar code for i1 types since GPRs will be used instead of mask registers. Differential Revision: https://reviews.llvm.org/D32273 llvm-svn: 303421
2017-05-19 20:35:15 +08:00
; SKX-NEXT: movw $-4, %cx
; SKX-NEXT: kmovd %ecx, %k0
; SKX-NEXT: kshiftrw $1, %k0, %k0
; SKX-NEXT: kshiftlw $1, %k0, %k0
Revert r274613 because it breaks the test suite with AVX512 This reverts most of r274613 (AKA r274626) and its follow-ups (r276347, r277289), due to miscompiles in the test suite. The FastISel change was left in, because it apparently fixes an unrelated issue. (Recommit of r279782 which was broken due to a bad merge.) This fixes 4 out of the 5 test failures in PR29112. llvm-svn: 279788
2016-08-26 06:48:11 +08:00
; SKX-NEXT: andl $1, %eax
[X86][AVX512] Make i1 illegal in the CodeGen This patch defines the i1 type as illegal in the X86 backend for AVX512. For DAG operations on <N x i1> types (build vector, extract vector element, ...) i8 is used, and should be truncated/extended. This should produce better scalar code for i1 types since GPRs will be used instead of mask registers. Differential Revision: https://reviews.llvm.org/D32273 llvm-svn: 303421
2017-05-19 20:35:15 +08:00
; SKX-NEXT: kmovw %eax, %k1
; SKX-NEXT: korw %k1, %k0, %k0
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; SKX-NEXT: kmovd %k0, %eax
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; SKX-NEXT: ## kill: def $ax killed $ax killed $eax
VirtRegMap: Replace some identity copies with KILL instructions. An identity COPY like this: %AL = COPY %AL, %EAX<imp-def> has no semantic effect, but encodes liveness information: Further users of %EAX only depend on this instruction even though it does not define the full register. Replace the COPY with a KILL instruction in those cases to maintain this liveness information. (This reverts a small part of r238588 but this time adds a comment explaining why a KILL instruction is useful). llvm-svn: 274952
2016-07-09 08:19:07 +08:00
; SKX-NEXT: retq
AVX-512: simpyfied BUILD_VECTOR for masks; fixed cmp/test sequence llvm-svn: 201487
2014-02-16 19:34:23 +08:00
%cmp_res = icmp ult i32 %a, %b
%maskv = insertelement <16 x i1> <i1 true, i1 false, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true>, i1 %cmp_res, i32 0
%res = bitcast <16 x i1> %maskv to i16
ret i16 %res
}
AVX-512: Fixed extract_vector_elt for v8i1 vector llvm-svn: 202624
2014-03-02 17:19:44 +08:00
define i64 @test14(<8 x i64>%a, <8 x i64>%b, i64 %a1, i64 %b1) {
VirtRegMap: Replace some identity copies with KILL instructions. An identity COPY like this: %AL = COPY %AL, %EAX<imp-def> has no semantic effect, but encodes liveness information: Further users of %EAX only depend on this instruction even though it does not define the full register. Replace the COPY with a KILL instruction in those cases to maintain this liveness information. (This reverts a small part of r238588 but this time adds a comment explaining why a KILL instruction is useful). llvm-svn: 274952
2016-07-09 08:19:07 +08:00
; KNL-LABEL: test14:
[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
; KNL: ## %bb.0:
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; KNL-NEXT: movq %rdi, %rax
VirtRegMap: Replace some identity copies with KILL instructions. An identity COPY like this: %AL = COPY %AL, %EAX<imp-def> has no semantic effect, but encodes liveness information: Further users of %EAX only depend on this instruction even though it does not define the full register. Replace the COPY with a KILL instruction in those cases to maintain this liveness information. (This reverts a small part of r238588 but this time adds a comment explaining why a KILL instruction is useful). llvm-svn: 274952
2016-07-09 08:19:07 +08:00
; KNL-NEXT: vpcmpgtq %zmm0, %zmm1, %k0
[X86] Don't zero the upper bits of the k-register before extracting a single bit from a vXi1. This doesn't match the semantics of the extract_vector_elt operation. Nothing downstream knows the bits were zeroed so they still get masked or sign extended after the extrat anyway. llvm-svn: 320723
2017-12-15 02:35:25 +08:00
; KNL-NEXT: kshiftrw $4, %k0, %k0
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; KNL-NEXT: kmovw %k0, %ecx
; KNL-NEXT: testb $1, %cl
; KNL-NEXT: cmoveq %rsi, %rax
[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
; KNL-NEXT: vzeroupper
VirtRegMap: Replace some identity copies with KILL instructions. An identity COPY like this: %AL = COPY %AL, %EAX<imp-def> has no semantic effect, but encodes liveness information: Further users of %EAX only depend on this instruction even though it does not define the full register. Replace the COPY with a KILL instruction in those cases to maintain this liveness information. (This reverts a small part of r238588 but this time adds a comment explaining why a KILL instruction is useful). llvm-svn: 274952
2016-07-09 08:19:07 +08:00
; KNL-NEXT: retq
;
; SKX-LABEL: test14:
[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
; SKX: ## %bb.0:
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; SKX-NEXT: movq %rdi, %rax
VirtRegMap: Replace some identity copies with KILL instructions. An identity COPY like this: %AL = COPY %AL, %EAX<imp-def> has no semantic effect, but encodes liveness information: Further users of %EAX only depend on this instruction even though it does not define the full register. Replace the COPY with a KILL instruction in those cases to maintain this liveness information. (This reverts a small part of r238588 but this time adds a comment explaining why a KILL instruction is useful). llvm-svn: 274952
2016-07-09 08:19:07 +08:00
; SKX-NEXT: vpcmpgtq %zmm0, %zmm1, %k0
[X86] Lower extract_element from k-registers by bitcasting from v16i1 to i16 and extending/truncating. This is equivalent to what isel was doing anyway but by canonicalizing earlier we can remove some patterns. llvm-svn: 326375
2018-03-01 06:23:55 +08:00
; SKX-NEXT: kshiftrw $4, %k0, %k0
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; SKX-NEXT: kmovd %k0, %ecx
; SKX-NEXT: testb $1, %cl
; SKX-NEXT: cmoveq %rsi, %rax
[X86] Generate VZEROUPPER for Skylake-avx512. VZEROUPPER should not be issued on Knights Landing (KNL), but on Skylake-avx512 it should be. Differential Revision: https://reviews.llvm.org/D29874 llvm-svn: 296859
2017-03-03 17:03:24 +08:00
; SKX-NEXT: vzeroupper
VirtRegMap: Replace some identity copies with KILL instructions. An identity COPY like this: %AL = COPY %AL, %EAX<imp-def> has no semantic effect, but encodes liveness information: Further users of %EAX only depend on this instruction even though it does not define the full register. Replace the COPY with a KILL instruction in those cases to maintain this liveness information. (This reverts a small part of r238588 but this time adds a comment explaining why a KILL instruction is useful). llvm-svn: 274952
2016-07-09 08:19:07 +08:00
; SKX-NEXT: retq
AVX-512: Fixed extract_vector_elt for v8i1 vector llvm-svn: 202624
2014-03-02 17:19:44 +08:00
%cmpvector_func.i = icmp slt <8 x i64> %a, %b
%extract24vector_func.i = extractelement <8 x i1> %cmpvector_func.i, i32 4
%res = select i1 %extract24vector_func.i, i64 %a1, i64 %b1
ret i64 %res
}
AVX-512: insert element to mask vector; store i1 data Implemented INSERT_VECTOR_ELT operation for v16i1 and v8i1 vectors; Implemented "store" for i1 type llvm-svn: 205850
2014-04-09 20:37:50 +08:00
define i16 @test15(i1 *%addr) {
[AVX512] Add a common prefix to avx512-insert-extract.ll so we can reduce the number of check lines on some test cases. This was pointed out during the review for D313804. llvm-svn: 309629
2017-08-01 05:20:06 +08:00
; CHECK-LABEL: test15:
[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] Add a common prefix to avx512-insert-extract.ll so we can reduce the number of check lines on some test cases. This was pointed out during the review for D313804. llvm-svn: 309629
2017-08-01 05:20:06 +08:00
; CHECK-NEXT: xorl %ecx, %ecx
[X86] Change some compare patterns to use loadi8/loadi16/loadi32/loadi64 helper fragments. This enables CMP8mi to fold zextloadi8i1 which in all tests allows us to avoid creating a TEST8rr that peephole can't fold. llvm-svn: 324863
2018-02-12 10:48:42 +08:00
; CHECK-NEXT: cmpb $0, (%rdi)
[X86] Promote 16-bit cmovs to 32-bits This allows us to avoid an opsize prefix. And forcing some move immediates to i32 avoids a length changing prefix on those instructions. This mostly replaces the existing combine we had for zext/sext+cmov of constants. I left in a case for sign extending a 32 bit cmov of constants to 64 bits. Differential Revision: https://reviews.llvm.org/D43327 llvm-svn: 325601
2018-02-21 01:41:00 +08:00
; CHECK-NEXT: movl $65535, %eax ## imm = 0xFFFF
; CHECK-NEXT: cmovel %ecx, %eax
; CHECK-NEXT: ## kill: def $ax killed $ax killed $eax
[AVX512] Add a common prefix to avx512-insert-extract.ll so we can reduce the number of check lines on some test cases. This was pointed out during the review for D313804. llvm-svn: 309629
2017-08-01 05:20:06 +08:00
; CHECK-NEXT: retq
AVX-512: fixed algorithm of building vectors of i1 elements fixed extract-insert i1 element, load i1, zextload i1 should be with "and $1, %reg" to prevent loading garbage. added a bunch of new tests. llvm-svn: 237793
2015-05-20 22:32:03 +08:00
%x = load i1 , i1 * %addr, align 1
AVX-512: insert element to mask vector; store i1 data Implemented INSERT_VECTOR_ELT operation for v16i1 and v8i1 vectors; Implemented "store" for i1 type llvm-svn: 205850
2014-04-09 20:37:50 +08:00
%x1 = insertelement <16 x i1> undef, i1 %x, i32 10
%x2 = bitcast <16 x i1>%x1 to i16
ret i16 %x2
}
define i16 @test16(i1 *%addr, i16 %a) {
VirtRegMap: Replace some identity copies with KILL instructions. An identity COPY like this: %AL = COPY %AL, %EAX<imp-def> has no semantic effect, but encodes liveness information: Further users of %EAX only depend on this instruction even though it does not define the full register. Replace the COPY with a KILL instruction in those cases to maintain this liveness information. (This reverts a small part of r238588 but this time adds a comment explaining why a KILL instruction is useful). llvm-svn: 274952
2016-07-09 08:19:07 +08:00
; KNL-LABEL: test16:
[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
; KNL: ## %bb.0:
[X86][AVX512] Make i1 illegal in the CodeGen This patch defines the i1 type as illegal in the X86 backend for AVX512. For DAG operations on <N x i1> types (build vector, extract vector element, ...) i8 is used, and should be truncated/extended. This should produce better scalar code for i1 types since GPRs will be used instead of mask registers. Differential Revision: https://reviews.llvm.org/D32273 llvm-svn: 303421
2017-05-19 20:35:15 +08:00
; KNL-NEXT: movb (%rdi), %al
[X86] Handle alls version of vXi1 insert_vector_elt with a constant index without falling back to shuffles. We previously only supported inserting to the LSB or MSB where it was easy to zero to perform an OR to insert. This change effectively extracts the old value and the new value, xors them together and then xors that single bit with the correct location in the original vector. This will cancel out the old value in the first xor leaving the new value in the position. The way I've implemented this uses 3 shifts and two xors and uses an additional register. We can avoid the additional register at the cost of another shift. llvm-svn: 320120
2017-12-08 08:16:09 +08:00
; KNL-NEXT: kmovw %esi, %k0
; KNL-NEXT: kmovw %eax, %k1
; KNL-NEXT: kshiftrw $10, %k0, %k2
; KNL-NEXT: kxorw %k1, %k2, %k1
; KNL-NEXT: kshiftlw $15, %k1, %k1
; KNL-NEXT: kshiftrw $5, %k1, %k1
[X86] Rewrite vXi1 element insertion by using a vXi1 scalar_to_vector and inserting into a vXi1 vector. The existing code was already doing something very similar to subvector insertion so this allows us to remove the nearly duplicate code. This patch is a little larger than it should be due to differences between the DQI handling between the two today. llvm-svn: 323212
2018-01-23 23:56:36 +08:00
; KNL-NEXT: kxorw %k1, %k0, %k0
VirtRegMap: Replace some identity copies with KILL instructions. An identity COPY like this: %AL = COPY %AL, %EAX<imp-def> has no semantic effect, but encodes liveness information: Further users of %EAX only depend on this instruction even though it does not define the full register. Replace the COPY with a KILL instruction in those cases to maintain this liveness information. (This reverts a small part of r238588 but this time adds a comment explaining why a KILL instruction is useful). llvm-svn: 274952
2016-07-09 08:19:07 +08:00
; KNL-NEXT: kmovw %k0, %eax
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; KNL-NEXT: ## kill: def $ax killed $ax killed $eax
VirtRegMap: Replace some identity copies with KILL instructions. An identity COPY like this: %AL = COPY %AL, %EAX<imp-def> has no semantic effect, but encodes liveness information: Further users of %EAX only depend on this instruction even though it does not define the full register. Replace the COPY with a KILL instruction in those cases to maintain this liveness information. (This reverts a small part of r238588 but this time adds a comment explaining why a KILL instruction is useful). llvm-svn: 274952
2016-07-09 08:19:07 +08:00
; KNL-NEXT: retq
;
; SKX-LABEL: test16:
[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
; SKX: ## %bb.0:
[X86] Add a pass to convert instruction chains between domains. The pass scans the function to find instruction chains that define registers in the same domain (closures). It then calculates the cost of converting the closure to another domain. If found profitable, the instructions are converted to instructions in the other domain and the register classes are changed accordingly. This commit adds the pass infrastructure and a simple conversion from the GPR domain to the Mask domain. Differential Revision: https://reviews.llvm.org/D37251 Change-Id: Ic2cf1d76598110401168326d411128ae2580a604 llvm-svn: 316288
2017-10-22 19:43:08 +08:00
; SKX-NEXT: kmovb (%rdi), %k0
; SKX-NEXT: kmovd %esi, %k1
[X86] Handle alls version of vXi1 insert_vector_elt with a constant index without falling back to shuffles. We previously only supported inserting to the LSB or MSB where it was easy to zero to perform an OR to insert. This change effectively extracts the old value and the new value, xors them together and then xors that single bit with the correct location in the original vector. This will cancel out the old value in the first xor leaving the new value in the position. The way I've implemented this uses 3 shifts and two xors and uses an additional register. We can avoid the additional register at the cost of another shift. llvm-svn: 320120
2017-12-08 08:16:09 +08:00
; SKX-NEXT: kshiftrw $10, %k1, %k2
; SKX-NEXT: kxorw %k0, %k2, %k0
; SKX-NEXT: kshiftlw $15, %k0, %k0
; SKX-NEXT: kshiftrw $5, %k0, %k0
[X86] Rewrite vXi1 element insertion by using a vXi1 scalar_to_vector and inserting into a vXi1 vector. The existing code was already doing something very similar to subvector insertion so this allows us to remove the nearly duplicate code. This patch is a little larger than it should be due to differences between the DQI handling between the two today. llvm-svn: 323212
2018-01-23 23:56:36 +08:00
; SKX-NEXT: kxorw %k0, %k1, %k0
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; SKX-NEXT: kmovd %k0, %eax
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; SKX-NEXT: ## kill: def $ax killed $ax killed $eax
VirtRegMap: Replace some identity copies with KILL instructions. An identity COPY like this: %AL = COPY %AL, %EAX<imp-def> has no semantic effect, but encodes liveness information: Further users of %EAX only depend on this instruction even though it does not define the full register. Replace the COPY with a KILL instruction in those cases to maintain this liveness information. (This reverts a small part of r238588 but this time adds a comment explaining why a KILL instruction is useful). llvm-svn: 274952
2016-07-09 08:19:07 +08:00
; SKX-NEXT: retq
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 llvm-svn: 230794
2015-02-28 05:17:42 +08:00
%x = load i1 , i1 * %addr, align 128
AVX-512: insert element to mask vector; store i1 data Implemented INSERT_VECTOR_ELT operation for v16i1 and v8i1 vectors; Implemented "store" for i1 type llvm-svn: 205850
2014-04-09 20:37:50 +08:00
%a1 = bitcast i16 %a to <16 x i1>
%x1 = insertelement <16 x i1> %a1, i1 %x, i32 10
%x2 = bitcast <16 x i1>%x1 to i16
ret i16 %x2
}
define i8 @test17(i1 *%addr, i8 %a) {
VirtRegMap: Replace some identity copies with KILL instructions. An identity COPY like this: %AL = COPY %AL, %EAX<imp-def> has no semantic effect, but encodes liveness information: Further users of %EAX only depend on this instruction even though it does not define the full register. Replace the COPY with a KILL instruction in those cases to maintain this liveness information. (This reverts a small part of r238588 but this time adds a comment explaining why a KILL instruction is useful). llvm-svn: 274952
2016-07-09 08:19:07 +08:00
; KNL-LABEL: test17:
[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
; KNL: ## %bb.0:
[X86][AVX512] Make i1 illegal in the CodeGen This patch defines the i1 type as illegal in the X86 backend for AVX512. For DAG operations on <N x i1> types (build vector, extract vector element, ...) i8 is used, and should be truncated/extended. This should produce better scalar code for i1 types since GPRs will be used instead of mask registers. Differential Revision: https://reviews.llvm.org/D32273 llvm-svn: 303421
2017-05-19 20:35:15 +08:00
; KNL-NEXT: movb (%rdi), %al
[X86] Handle alls version of vXi1 insert_vector_elt with a constant index without falling back to shuffles. We previously only supported inserting to the LSB or MSB where it was easy to zero to perform an OR to insert. This change effectively extracts the old value and the new value, xors them together and then xors that single bit with the correct location in the original vector. This will cancel out the old value in the first xor leaving the new value in the position. The way I've implemented this uses 3 shifts and two xors and uses an additional register. We can avoid the additional register at the cost of another shift. llvm-svn: 320120
2017-12-08 08:16:09 +08:00
; KNL-NEXT: kmovw %esi, %k0
; KNL-NEXT: kmovw %eax, %k1
; KNL-NEXT: kshiftrw $4, %k0, %k2
; KNL-NEXT: kxorw %k1, %k2, %k1
; KNL-NEXT: kshiftlw $15, %k1, %k1
; KNL-NEXT: kshiftrw $11, %k1, %k1
[X86] Rewrite vXi1 element insertion by using a vXi1 scalar_to_vector and inserting into a vXi1 vector. The existing code was already doing something very similar to subvector insertion so this allows us to remove the nearly duplicate code. This patch is a little larger than it should be due to differences between the DQI handling between the two today. llvm-svn: 323212
2018-01-23 23:56:36 +08:00
; KNL-NEXT: kxorw %k1, %k0, %k0
VirtRegMap: Replace some identity copies with KILL instructions. An identity COPY like this: %AL = COPY %AL, %EAX<imp-def> has no semantic effect, but encodes liveness information: Further users of %EAX only depend on this instruction even though it does not define the full register. Replace the COPY with a KILL instruction in those cases to maintain this liveness information. (This reverts a small part of r238588 but this time adds a comment explaining why a KILL instruction is useful). llvm-svn: 274952
2016-07-09 08:19:07 +08:00
; KNL-NEXT: kmovw %k0, %eax
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; KNL-NEXT: ## kill: def $al killed $al killed $eax
VirtRegMap: Replace some identity copies with KILL instructions. An identity COPY like this: %AL = COPY %AL, %EAX<imp-def> has no semantic effect, but encodes liveness information: Further users of %EAX only depend on this instruction even though it does not define the full register. Replace the COPY with a KILL instruction in those cases to maintain this liveness information. (This reverts a small part of r238588 but this time adds a comment explaining why a KILL instruction is useful). llvm-svn: 274952
2016-07-09 08:19:07 +08:00
; KNL-NEXT: retq
;
; SKX-LABEL: test17:
[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
; SKX: ## %bb.0:
[X86] Add a pass to convert instruction chains between domains. The pass scans the function to find instruction chains that define registers in the same domain (closures). It then calculates the cost of converting the closure to another domain. If found profitable, the instructions are converted to instructions in the other domain and the register classes are changed accordingly. This commit adds the pass infrastructure and a simple conversion from the GPR domain to the Mask domain. Differential Revision: https://reviews.llvm.org/D37251 Change-Id: Ic2cf1d76598110401168326d411128ae2580a604 llvm-svn: 316288
2017-10-22 19:43:08 +08:00
; SKX-NEXT: kmovb (%rdi), %k0
; SKX-NEXT: kmovd %esi, %k1
[X86] Handle alls version of vXi1 insert_vector_elt with a constant index without falling back to shuffles. We previously only supported inserting to the LSB or MSB where it was easy to zero to perform an OR to insert. This change effectively extracts the old value and the new value, xors them together and then xors that single bit with the correct location in the original vector. This will cancel out the old value in the first xor leaving the new value in the position. The way I've implemented this uses 3 shifts and two xors and uses an additional register. We can avoid the additional register at the cost of another shift. llvm-svn: 320120
2017-12-08 08:16:09 +08:00
; SKX-NEXT: kshiftrb $4, %k1, %k2
; SKX-NEXT: kxorb %k0, %k2, %k0
; SKX-NEXT: kshiftlb $7, %k0, %k0
; SKX-NEXT: kshiftrb $3, %k0, %k0
[X86] Rewrite vXi1 element insertion by using a vXi1 scalar_to_vector and inserting into a vXi1 vector. The existing code was already doing something very similar to subvector insertion so this allows us to remove the nearly duplicate code. This patch is a little larger than it should be due to differences between the DQI handling between the two today. llvm-svn: 323212
2018-01-23 23:56:36 +08:00
; SKX-NEXT: kxorb %k0, %k1, %k0
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; SKX-NEXT: kmovd %k0, %eax
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; SKX-NEXT: ## kill: def $al killed $al killed $eax
VirtRegMap: Replace some identity copies with KILL instructions. An identity COPY like this: %AL = COPY %AL, %EAX<imp-def> has no semantic effect, but encodes liveness information: Further users of %EAX only depend on this instruction even though it does not define the full register. Replace the COPY with a KILL instruction in those cases to maintain this liveness information. (This reverts a small part of r238588 but this time adds a comment explaining why a KILL instruction is useful). llvm-svn: 274952
2016-07-09 08:19:07 +08:00
; SKX-NEXT: retq
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 llvm-svn: 230794
2015-02-28 05:17:42 +08:00
%x = load i1 , i1 * %addr, align 128
AVX-512: insert element to mask vector; store i1 data Implemented INSERT_VECTOR_ELT operation for v16i1 and v8i1 vectors; Implemented "store" for i1 type llvm-svn: 205850
2014-04-09 20:37:50 +08:00
%a1 = bitcast i8 %a to <8 x i1>
[AVX512] Fix test to produce a defined value We're inserting into a 8 wide vector, so the index should be < 8. llvm-svn: 219872
2014-10-16 07:42:11 +08:00
%x1 = insertelement <8 x i1> %a1, i1 %x, i32 4
AVX-512: insert element to mask vector; store i1 data Implemented INSERT_VECTOR_ELT operation for v16i1 and v8i1 vectors; Implemented "store" for i1 type llvm-svn: 205850
2014-04-09 20:37:50 +08:00
%x2 = bitcast <8 x i1>%x1 to i8
ret i8 %x2
}
AVX512: vpextrb/w/d/q and vpinsrb/w/d/q implementation. This instructions doesn't have intrincis. Added tests for lowering and encoding. Differential Revision: http://reviews.llvm.org/D12317 llvm-svn: 249688
2015-10-08 20:55:01 +08:00
define i64 @extract_v8i64(<8 x i64> %x, i64* %dst) {
[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-LABEL: extract_v8i64:
[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][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: vpextrq $1, %xmm0, %rax
; CHECK-NEXT: vextracti128 $1, %ymm0, %xmm0
; CHECK-NEXT: vpextrq $1, %xmm0, (%rdi)
; CHECK-NEXT: vzeroupper
; CHECK-NEXT: retq
AVX512: vpextrb/w/d/q and vpinsrb/w/d/q implementation. This instructions doesn't have intrincis. Added tests for lowering and encoding. Differential Revision: http://reviews.llvm.org/D12317 llvm-svn: 249688
2015-10-08 20:55:01 +08:00
%r1 = extractelement <8 x i64> %x, i32 1
%r2 = extractelement <8 x i64> %x, i32 3
store i64 %r2, i64* %dst, align 1
ret i64 %r1
}
define i64 @extract_v4i64(<4 x i64> %x, i64* %dst) {
[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-LABEL: extract_v4i64:
[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][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: vpextrq $1, %xmm0, %rax
; CHECK-NEXT: vextracti128 $1, %ymm0, %xmm0
; CHECK-NEXT: vpextrq $1, %xmm0, (%rdi)
; CHECK-NEXT: vzeroupper
; CHECK-NEXT: retq
AVX512: vpextrb/w/d/q and vpinsrb/w/d/q implementation. This instructions doesn't have intrincis. Added tests for lowering and encoding. Differential Revision: http://reviews.llvm.org/D12317 llvm-svn: 249688
2015-10-08 20:55:01 +08:00
%r1 = extractelement <4 x i64> %x, i32 1
%r2 = extractelement <4 x i64> %x, i32 3
store i64 %r2, i64* %dst, align 1
ret i64 %r1
}
define i64 @extract_v2i64(<2 x i64> %x, i64* %dst) {
[AVX512] Add a common prefix to avx512-insert-extract.ll so we can reduce the number of check lines on some test cases. This was pointed out during the review for D313804. llvm-svn: 309629
2017-08-01 05:20:06 +08:00
; CHECK-LABEL: extract_v2i64:
[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] Add a common prefix to avx512-insert-extract.ll so we can reduce the number of check lines on some test cases. This was pointed out during the review for D313804. llvm-svn: 309629
2017-08-01 05:20:06 +08:00
; CHECK-NEXT: vmovq %xmm0, %rax
; CHECK-NEXT: vpextrq $1, %xmm0, (%rdi)
; CHECK-NEXT: retq
AVX512: vpextrb/w/d/q and vpinsrb/w/d/q implementation. This instructions doesn't have intrincis. Added tests for lowering and encoding. Differential Revision: http://reviews.llvm.org/D12317 llvm-svn: 249688
2015-10-08 20:55:01 +08:00
%r1 = extractelement <2 x i64> %x, i32 0
%r2 = extractelement <2 x i64> %x, i32 1
store i64 %r2, i64* %dst, align 1
ret i64 %r1
}
define i32 @extract_v16i32(<16 x i32> %x, i32* %dst) {
[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-LABEL: extract_v16i32:
[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][SSE] Add extractps/pextrd equivalence to domain tables Differential Revision: https://reviews.llvm.org/D39135 llvm-svn: 316274
2017-10-22 04:19:48 +08:00
; CHECK-NEXT: vextractps $1, %xmm0, %eax
; CHECK-NEXT: vextractf128 $1, %ymm0, %xmm0
; CHECK-NEXT: vextractps $1, %xmm0, (%rdi)
[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
; CHECK-NEXT: retq
AVX512: vpextrb/w/d/q and vpinsrb/w/d/q implementation. This instructions doesn't have intrincis. Added tests for lowering and encoding. Differential Revision: http://reviews.llvm.org/D12317 llvm-svn: 249688
2015-10-08 20:55:01 +08:00
%r1 = extractelement <16 x i32> %x, i32 1
%r2 = extractelement <16 x i32> %x, i32 5
store i32 %r2, i32* %dst, align 1
ret i32 %r1
}
define i32 @extract_v8i32(<8 x i32> %x, i32* %dst) {
[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-LABEL: extract_v8i32:
[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][SSE] Add extractps/pextrd equivalence to domain tables Differential Revision: https://reviews.llvm.org/D39135 llvm-svn: 316274
2017-10-22 04:19:48 +08:00
; CHECK-NEXT: vextractps $1, %xmm0, %eax
; CHECK-NEXT: vextractf128 $1, %ymm0, %xmm0
; CHECK-NEXT: vextractps $1, %xmm0, (%rdi)
[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
; CHECK-NEXT: retq
AVX512: vpextrb/w/d/q and vpinsrb/w/d/q implementation. This instructions doesn't have intrincis. Added tests for lowering and encoding. Differential Revision: http://reviews.llvm.org/D12317 llvm-svn: 249688
2015-10-08 20:55:01 +08:00
%r1 = extractelement <8 x i32> %x, i32 1
%r2 = extractelement <8 x i32> %x, i32 5
store i32 %r2, i32* %dst, align 1
ret i32 %r1
}
define i32 @extract_v4i32(<4 x i32> %x, i32* %dst) {
[AVX512] Add a common prefix to avx512-insert-extract.ll so we can reduce the number of check lines on some test cases. This was pointed out during the review for D313804. llvm-svn: 309629
2017-08-01 05:20:06 +08:00
; CHECK-LABEL: extract_v4i32:
[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][SSE] Add extractps/pextrd equivalence to domain tables Differential Revision: https://reviews.llvm.org/D39135 llvm-svn: 316274
2017-10-22 04:19:48 +08:00
; CHECK-NEXT: vextractps $1, %xmm0, %eax
; CHECK-NEXT: vextractps $3, %xmm0, (%rdi)
[AVX512] Add a common prefix to avx512-insert-extract.ll so we can reduce the number of check lines on some test cases. This was pointed out during the review for D313804. llvm-svn: 309629
2017-08-01 05:20:06 +08:00
; CHECK-NEXT: retq
AVX512: vpextrb/w/d/q and vpinsrb/w/d/q implementation. This instructions doesn't have intrincis. Added tests for lowering and encoding. Differential Revision: http://reviews.llvm.org/D12317 llvm-svn: 249688
2015-10-08 20:55:01 +08:00
%r1 = extractelement <4 x i32> %x, i32 1
%r2 = extractelement <4 x i32> %x, i32 3
store i32 %r2, i32* %dst, align 1
ret i32 %r1
}
define i16 @extract_v32i16(<32 x i16> %x, i16* %dst) {
[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-LABEL: extract_v32i16:
[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][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: vpextrw $1, %xmm0, %eax
; CHECK-NEXT: vextracti128 $1, %ymm0, %xmm0
; CHECK-NEXT: vpextrw $1, %xmm0, (%rdi)
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +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
; CHECK-NEXT: retq
AVX512: vpextrb/w/d/q and vpinsrb/w/d/q implementation. This instructions doesn't have intrincis. Added tests for lowering and encoding. Differential Revision: http://reviews.llvm.org/D12317 llvm-svn: 249688
2015-10-08 20:55:01 +08:00
%r1 = extractelement <32 x i16> %x, i32 1
%r2 = extractelement <32 x i16> %x, i32 9
store i16 %r2, i16* %dst, align 1
ret i16 %r1
}
define i16 @extract_v16i16(<16 x i16> %x, i16* %dst) {
[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-LABEL: extract_v16i16:
[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][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: vpextrw $1, %xmm0, %eax
; CHECK-NEXT: vextracti128 $1, %ymm0, %xmm0
; CHECK-NEXT: vpextrw $1, %xmm0, (%rdi)
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +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
; CHECK-NEXT: retq
AVX512: vpextrb/w/d/q and vpinsrb/w/d/q implementation. This instructions doesn't have intrincis. Added tests for lowering and encoding. Differential Revision: http://reviews.llvm.org/D12317 llvm-svn: 249688
2015-10-08 20:55:01 +08:00
%r1 = extractelement <16 x i16> %x, i32 1
%r2 = extractelement <16 x i16> %x, i32 9
store i16 %r2, i16* %dst, align 1
ret i16 %r1
}
define i16 @extract_v8i16(<8 x i16> %x, i16* %dst) {
[AVX512] Add a common prefix to avx512-insert-extract.ll so we can reduce the number of check lines on some test cases. This was pointed out during the review for D313804. llvm-svn: 309629
2017-08-01 05:20:06 +08:00
; CHECK-LABEL: extract_v8i16:
[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] Add a common prefix to avx512-insert-extract.ll so we can reduce the number of check lines on some test cases. This was pointed out during the review for D313804. llvm-svn: 309629
2017-08-01 05:20:06 +08:00
; CHECK-NEXT: vpextrw $1, %xmm0, %eax
; CHECK-NEXT: vpextrw $3, %xmm0, (%rdi)
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; CHECK-NEXT: ## kill: def $ax killed $ax killed $eax
[AVX512] Add a common prefix to avx512-insert-extract.ll so we can reduce the number of check lines on some test cases. This was pointed out during the review for D313804. llvm-svn: 309629
2017-08-01 05:20:06 +08:00
; CHECK-NEXT: retq
AVX512: vpextrb/w/d/q and vpinsrb/w/d/q implementation. This instructions doesn't have intrincis. Added tests for lowering and encoding. Differential Revision: http://reviews.llvm.org/D12317 llvm-svn: 249688
2015-10-08 20:55:01 +08:00
%r1 = extractelement <8 x i16> %x, i32 1
%r2 = extractelement <8 x i16> %x, i32 3
store i16 %r2, i16* %dst, align 1
ret i16 %r1
}
define i8 @extract_v64i8(<64 x i8> %x, i8* %dst) {
[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-LABEL: extract_v64i8:
[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][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: vpextrb $1, %xmm0, %eax
; CHECK-NEXT: vextracti128 $1, %ymm0, %xmm0
; CHECK-NEXT: vpextrb $1, %xmm0, (%rdi)
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; CHECK-NEXT: ## kill: def $al killed $al 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
; CHECK-NEXT: retq
AVX512: vpextrb/w/d/q and vpinsrb/w/d/q implementation. This instructions doesn't have intrincis. Added tests for lowering and encoding. Differential Revision: http://reviews.llvm.org/D12317 llvm-svn: 249688
2015-10-08 20:55:01 +08:00
%r1 = extractelement <64 x i8> %x, i32 1
%r2 = extractelement <64 x i8> %x, i32 17
store i8 %r2, i8* %dst, align 1
ret i8 %r1
}
define i8 @extract_v32i8(<32 x i8> %x, i8* %dst) {
[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-LABEL: extract_v32i8:
[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][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: vpextrb $1, %xmm0, %eax
; CHECK-NEXT: vextracti128 $1, %ymm0, %xmm0
; CHECK-NEXT: vpextrb $1, %xmm0, (%rdi)
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; CHECK-NEXT: ## kill: def $al killed $al 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
; CHECK-NEXT: retq
AVX512: vpextrb/w/d/q and vpinsrb/w/d/q implementation. This instructions doesn't have intrincis. Added tests for lowering and encoding. Differential Revision: http://reviews.llvm.org/D12317 llvm-svn: 249688
2015-10-08 20:55:01 +08:00
%r1 = extractelement <32 x i8> %x, i32 1
%r2 = extractelement <32 x i8> %x, i32 17
store i8 %r2, i8* %dst, align 1
ret i8 %r1
}
define i8 @extract_v16i8(<16 x i8> %x, i8* %dst) {
[AVX512] Add a common prefix to avx512-insert-extract.ll so we can reduce the number of check lines on some test cases. This was pointed out during the review for D313804. llvm-svn: 309629
2017-08-01 05:20:06 +08:00
; CHECK-LABEL: extract_v16i8:
[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] Add a common prefix to avx512-insert-extract.ll so we can reduce the number of check lines on some test cases. This was pointed out during the review for D313804. llvm-svn: 309629
2017-08-01 05:20:06 +08:00
; CHECK-NEXT: vpextrb $1, %xmm0, %eax
; CHECK-NEXT: vpextrb $3, %xmm0, (%rdi)
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; CHECK-NEXT: ## kill: def $al killed $al killed $eax
[AVX512] Add a common prefix to avx512-insert-extract.ll so we can reduce the number of check lines on some test cases. This was pointed out during the review for D313804. llvm-svn: 309629
2017-08-01 05:20:06 +08:00
; CHECK-NEXT: retq
AVX512: vpextrb/w/d/q and vpinsrb/w/d/q implementation. This instructions doesn't have intrincis. Added tests for lowering and encoding. Differential Revision: http://reviews.llvm.org/D12317 llvm-svn: 249688
2015-10-08 20:55:01 +08:00
%r1 = extractelement <16 x i8> %x, i32 1
%r2 = extractelement <16 x i8> %x, i32 3
store i8 %r2, i8* %dst, align 1
ret i8 %r1
}
define <8 x i64> @insert_v8i64(<8 x i64> %x, i64 %y , i64* %ptr) {
[AVX512] Don't switch unmasked subvector insert/extract instructions when AVX512DQI is enabled. There's no reason to switch instructions with and without DQI. It just creates extra isel patterns and test divergences. There is however value in enabling the masked version of the instructions with DQI. This required introducing some new multiclasses to enabling this splitting. Differential Revision: https://reviews.llvm.org/D36661 llvm-svn: 311091
2017-08-17 23:40:25 +08:00
; CHECK-LABEL: insert_v8i64:
[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] Don't switch unmasked subvector insert/extract instructions when AVX512DQI is enabled. There's no reason to switch instructions with and without DQI. It just creates extra isel patterns and test divergences. There is however value in enabling the masked version of the instructions with DQI. This required introducing some new multiclasses to enabling this splitting. Differential Revision: https://reviews.llvm.org/D36661 llvm-svn: 311091
2017-08-17 23:40:25 +08:00
; CHECK-NEXT: vpinsrq $1, (%rsi), %xmm0, %xmm1
; CHECK-NEXT: vinserti32x4 $0, %xmm1, %zmm0, %zmm1
[AVX512] Use 256-bit extract instructions for extracting bits [255:128] from a 512-bit register This enables the use of a smaller encoding by using a VEX instruction when possible. Differential Revision: https://reviews.llvm.org/D37092 llvm-svn: 312100
2017-08-30 15:26:12 +08:00
; CHECK-NEXT: vextracti128 $1, %ymm0, %xmm0
[AVX512] Don't switch unmasked subvector insert/extract instructions when AVX512DQI is enabled. There's no reason to switch instructions with and without DQI. It just creates extra isel patterns and test divergences. There is however value in enabling the masked version of the instructions with DQI. This required introducing some new multiclasses to enabling this splitting. Differential Revision: https://reviews.llvm.org/D36661 llvm-svn: 311091
2017-08-17 23:40:25 +08:00
; CHECK-NEXT: vpinsrq $1, %rdi, %xmm0, %xmm0
; CHECK-NEXT: vinserti32x4 $1, %xmm0, %zmm1, %zmm0
; CHECK-NEXT: retq
AVX512: vpextrb/w/d/q and vpinsrb/w/d/q implementation. This instructions doesn't have intrincis. Added tests for lowering and encoding. Differential Revision: http://reviews.llvm.org/D12317 llvm-svn: 249688
2015-10-08 20:55:01 +08:00
%val = load i64, i64* %ptr
%r1 = insertelement <8 x i64> %x, i64 %val, i32 1
%r2 = insertelement <8 x i64> %r1, i64 %y, i32 3
ret <8 x i64> %r2
}
define <4 x i64> @insert_v4i64(<4 x i64> %x, i64 %y , i64* %ptr) {
[AVX512] Add a common prefix to avx512-insert-extract.ll so we can reduce the number of check lines on some test cases. This was pointed out during the review for D313804. llvm-svn: 309629
2017-08-01 05:20:06 +08:00
; CHECK-LABEL: insert_v4i64:
[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] Add a common prefix to avx512-insert-extract.ll so we can reduce the number of check lines on some test cases. This was pointed out during the review for D313804. llvm-svn: 309629
2017-08-01 05:20:06 +08:00
; CHECK-NEXT: vpinsrq $1, (%rsi), %xmm0, %xmm1
; CHECK-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; CHECK-NEXT: vextracti128 $1, %ymm0, %xmm0
; CHECK-NEXT: vpinsrq $1, %rdi, %xmm0, %xmm0
; CHECK-NEXT: vinserti128 $1, %xmm0, %ymm1, %ymm0
; CHECK-NEXT: retq
AVX512: vpextrb/w/d/q and vpinsrb/w/d/q implementation. This instructions doesn't have intrincis. Added tests for lowering and encoding. Differential Revision: http://reviews.llvm.org/D12317 llvm-svn: 249688
2015-10-08 20:55:01 +08:00
%val = load i64, i64* %ptr
%r1 = insertelement <4 x i64> %x, i64 %val, i32 1
%r2 = insertelement <4 x i64> %r1, i64 %y, i32 3
ret <4 x i64> %r2
}
define <2 x i64> @insert_v2i64(<2 x i64> %x, i64 %y , i64* %ptr) {
[AVX512] Add a common prefix to avx512-insert-extract.ll so we can reduce the number of check lines on some test cases. This was pointed out during the review for D313804. llvm-svn: 309629
2017-08-01 05:20:06 +08:00
; CHECK-LABEL: insert_v2i64:
[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] Add a common prefix to avx512-insert-extract.ll so we can reduce the number of check lines on some test cases. This was pointed out during the review for D313804. llvm-svn: 309629
2017-08-01 05:20:06 +08:00
; CHECK-NEXT: vpinsrq $0, %rdi, %xmm0, %xmm0
; CHECK-NEXT: vpinsrq $1, (%rsi), %xmm0, %xmm0
; CHECK-NEXT: retq
AVX512: vpextrb/w/d/q and vpinsrb/w/d/q implementation. This instructions doesn't have intrincis. Added tests for lowering and encoding. Differential Revision: http://reviews.llvm.org/D12317 llvm-svn: 249688
2015-10-08 20:55:01 +08:00
%val = load i64, i64* %ptr
%r1 = insertelement <2 x i64> %x, i64 %val, i32 1
DAG: Fold out out of bounds insert_vector_elt getNode already prevents formation of out of bounds constant extract_vector_elts. Do the same for insert_vector_elt. llvm-svn: 288603
2016-12-04 07:03:26 +08:00
%r2 = insertelement <2 x i64> %r1, i64 %y, i32 0
AVX512: vpextrb/w/d/q and vpinsrb/w/d/q implementation. This instructions doesn't have intrincis. Added tests for lowering and encoding. Differential Revision: http://reviews.llvm.org/D12317 llvm-svn: 249688
2015-10-08 20:55:01 +08:00
ret <2 x i64> %r2
}
define <16 x i32> @insert_v16i32(<16 x i32> %x, i32 %y, i32* %ptr) {
[AVX512] Add a common prefix to avx512-insert-extract.ll so we can reduce the number of check lines on some test cases. This was pointed out during the review for D313804. llvm-svn: 309629
2017-08-01 05:20:06 +08:00
; CHECK-LABEL: insert_v16i32:
[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] Add a common prefix to avx512-insert-extract.ll so we can reduce the number of check lines on some test cases. This was pointed out during the review for D313804. llvm-svn: 309629
2017-08-01 05:20:06 +08:00
; CHECK-NEXT: vpinsrd $1, (%rsi), %xmm0, %xmm1
; CHECK-NEXT: vinserti32x4 $0, %xmm1, %zmm0, %zmm1
[AVX512] Use 256-bit extract instructions for extracting bits [255:128] from a 512-bit register This enables the use of a smaller encoding by using a VEX instruction when possible. Differential Revision: https://reviews.llvm.org/D37092 llvm-svn: 312100
2017-08-30 15:26:12 +08:00
; CHECK-NEXT: vextracti128 $1, %ymm0, %xmm0
[AVX512] Add a common prefix to avx512-insert-extract.ll so we can reduce the number of check lines on some test cases. This was pointed out during the review for D313804. llvm-svn: 309629
2017-08-01 05:20:06 +08:00
; CHECK-NEXT: vpinsrd $1, %edi, %xmm0, %xmm0
; CHECK-NEXT: vinserti32x4 $1, %xmm0, %zmm1, %zmm0
; CHECK-NEXT: retq
AVX512: vpextrb/w/d/q and vpinsrb/w/d/q implementation. This instructions doesn't have intrincis. Added tests for lowering and encoding. Differential Revision: http://reviews.llvm.org/D12317 llvm-svn: 249688
2015-10-08 20:55:01 +08:00
%val = load i32, i32* %ptr
%r1 = insertelement <16 x i32> %x, i32 %val, i32 1
%r2 = insertelement <16 x i32> %r1, i32 %y, i32 5
ret <16 x i32> %r2
}
define <8 x i32> @insert_v8i32(<8 x i32> %x, i32 %y, i32* %ptr) {
[AVX512] Add a common prefix to avx512-insert-extract.ll so we can reduce the number of check lines on some test cases. This was pointed out during the review for D313804. llvm-svn: 309629
2017-08-01 05:20:06 +08:00
; CHECK-LABEL: insert_v8i32:
[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] Add a common prefix to avx512-insert-extract.ll so we can reduce the number of check lines on some test cases. This was pointed out during the review for D313804. llvm-svn: 309629
2017-08-01 05:20:06 +08:00
; CHECK-NEXT: vpinsrd $1, (%rsi), %xmm0, %xmm1
; CHECK-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; CHECK-NEXT: vextracti128 $1, %ymm0, %xmm0
; CHECK-NEXT: vpinsrd $1, %edi, %xmm0, %xmm0
; CHECK-NEXT: vinserti128 $1, %xmm0, %ymm1, %ymm0
; CHECK-NEXT: retq
AVX512: vpextrb/w/d/q and vpinsrb/w/d/q implementation. This instructions doesn't have intrincis. Added tests for lowering and encoding. Differential Revision: http://reviews.llvm.org/D12317 llvm-svn: 249688
2015-10-08 20:55:01 +08:00
%val = load i32, i32* %ptr
%r1 = insertelement <8 x i32> %x, i32 %val, i32 1
%r2 = insertelement <8 x i32> %r1, i32 %y, i32 5
ret <8 x i32> %r2
}
define <4 x i32> @insert_v4i32(<4 x i32> %x, i32 %y, i32* %ptr) {
[AVX512] Add a common prefix to avx512-insert-extract.ll so we can reduce the number of check lines on some test cases. This was pointed out during the review for D313804. llvm-svn: 309629
2017-08-01 05:20:06 +08:00
; CHECK-LABEL: insert_v4i32:
[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] Add a common prefix to avx512-insert-extract.ll so we can reduce the number of check lines on some test cases. This was pointed out during the review for D313804. llvm-svn: 309629
2017-08-01 05:20:06 +08:00
; CHECK-NEXT: vpinsrd $1, (%rsi), %xmm0, %xmm0
; CHECK-NEXT: vpinsrd $3, %edi, %xmm0, %xmm0
; CHECK-NEXT: retq
AVX512: vpextrb/w/d/q and vpinsrb/w/d/q implementation. This instructions doesn't have intrincis. Added tests for lowering and encoding. Differential Revision: http://reviews.llvm.org/D12317 llvm-svn: 249688
2015-10-08 20:55:01 +08:00
%val = load i32, i32* %ptr
%r1 = insertelement <4 x i32> %x, i32 %val, i32 1
%r2 = insertelement <4 x i32> %r1, i32 %y, i32 3
ret <4 x i32> %r2
}
define <32 x i16> @insert_v32i16(<32 x i16> %x, i16 %y, i16* %ptr) {
; KNL-LABEL: insert_v32i16:
[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
; KNL: ## %bb.0:
AVX512: vpextrb/w/d/q and vpinsrb/w/d/q implementation. This instructions doesn't have intrincis. Added tests for lowering and encoding. Differential Revision: http://reviews.llvm.org/D12317 llvm-svn: 249688
2015-10-08 20:55:01 +08:00
; KNL-NEXT: vpinsrw $1, (%rsi), %xmm0, %xmm2
[X86] Update test case I missed in r294876. llvm-svn: 294878
2017-02-12 07:23:11 +08:00
; KNL-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm0[4,5,6,7]
; KNL-NEXT: vextracti128 $1, %ymm0, %xmm0
; KNL-NEXT: vpinsrw $1, %edi, %xmm0, %xmm0
; KNL-NEXT: vinserti128 $1, %xmm0, %ymm2, %ymm0
AVX512: vpextrb/w/d/q and vpinsrb/w/d/q implementation. This instructions doesn't have intrincis. Added tests for lowering and encoding. Differential Revision: http://reviews.llvm.org/D12317 llvm-svn: 249688
2015-10-08 20:55:01 +08:00
; KNL-NEXT: retq
;
; SKX-LABEL: insert_v32i16:
[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
; SKX: ## %bb.0:
AVX512: vpextrb/w/d/q and vpinsrb/w/d/q implementation. This instructions doesn't have intrincis. Added tests for lowering and encoding. Differential Revision: http://reviews.llvm.org/D12317 llvm-svn: 249688
2015-10-08 20:55:01 +08:00
; SKX-NEXT: vpinsrw $1, (%rsi), %xmm0, %xmm1
[DAGCombiner] Remove the half vector width check for the combine of EXTRACT_SUBVECTOR from an INSERT_SUBVECTOR. This gives more parallelism opportunities for AVX-512 when dealing with 128-bit extracts from 512-bit vectors. llvm-svn: 294930
2017-02-13 07:49:49 +08:00
; SKX-NEXT: vinserti32x4 $0, %xmm1, %zmm0, %zmm1
[AVX512] Use 256-bit extract instructions for extracting bits [255:128] from a 512-bit register This enables the use of a smaller encoding by using a VEX instruction when possible. Differential Revision: https://reviews.llvm.org/D37092 llvm-svn: 312100
2017-08-30 15:26:12 +08:00
; SKX-NEXT: vextracti128 $1, %ymm0, %xmm0
[DAGCombiner] Remove the half vector width check for the combine of EXTRACT_SUBVECTOR from an INSERT_SUBVECTOR. This gives more parallelism opportunities for AVX-512 when dealing with 128-bit extracts from 512-bit vectors. llvm-svn: 294930
2017-02-13 07:49:49 +08:00
; SKX-NEXT: vpinsrw $1, %edi, %xmm0, %xmm0
; SKX-NEXT: vinserti32x4 $1, %xmm0, %zmm1, %zmm0
AVX512: vpextrb/w/d/q and vpinsrb/w/d/q implementation. This instructions doesn't have intrincis. Added tests for lowering and encoding. Differential Revision: http://reviews.llvm.org/D12317 llvm-svn: 249688
2015-10-08 20:55:01 +08:00
; SKX-NEXT: retq
%val = load i16, i16* %ptr
%r1 = insertelement <32 x i16> %x, i16 %val, i32 1
%r2 = insertelement <32 x i16> %r1, i16 %y, i32 9
ret <32 x i16> %r2
}
define <16 x i16> @insert_v16i16(<16 x i16> %x, i16 %y, i16* %ptr) {
[AVX512] Add a common prefix to avx512-insert-extract.ll so we can reduce the number of check lines on some test cases. This was pointed out during the review for D313804. llvm-svn: 309629
2017-08-01 05:20:06 +08:00
; CHECK-LABEL: insert_v16i16:
[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] Add a common prefix to avx512-insert-extract.ll so we can reduce the number of check lines on some test cases. This was pointed out during the review for D313804. llvm-svn: 309629
2017-08-01 05:20:06 +08:00
; CHECK-NEXT: vpinsrw $1, (%rsi), %xmm0, %xmm1
; CHECK-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; CHECK-NEXT: vextracti128 $1, %ymm0, %xmm0
; CHECK-NEXT: vpinsrw $1, %edi, %xmm0, %xmm0
; CHECK-NEXT: vinserti128 $1, %xmm0, %ymm1, %ymm0
; CHECK-NEXT: retq
AVX512: vpextrb/w/d/q and vpinsrb/w/d/q implementation. This instructions doesn't have intrincis. Added tests for lowering and encoding. Differential Revision: http://reviews.llvm.org/D12317 llvm-svn: 249688
2015-10-08 20:55:01 +08:00
%val = load i16, i16* %ptr
%r1 = insertelement <16 x i16> %x, i16 %val, i32 1
%r2 = insertelement <16 x i16> %r1, i16 %y, i32 9
ret <16 x i16> %r2
}
define <8 x i16> @insert_v8i16(<8 x i16> %x, i16 %y, i16* %ptr) {
[AVX512] Add a common prefix to avx512-insert-extract.ll so we can reduce the number of check lines on some test cases. This was pointed out during the review for D313804. llvm-svn: 309629
2017-08-01 05:20:06 +08:00
; CHECK-LABEL: insert_v8i16:
[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] Add a common prefix to avx512-insert-extract.ll so we can reduce the number of check lines on some test cases. This was pointed out during the review for D313804. llvm-svn: 309629
2017-08-01 05:20:06 +08:00
; CHECK-NEXT: vpinsrw $1, (%rsi), %xmm0, %xmm0
; CHECK-NEXT: vpinsrw $5, %edi, %xmm0, %xmm0
; CHECK-NEXT: retq
AVX512: vpextrb/w/d/q and vpinsrb/w/d/q implementation. This instructions doesn't have intrincis. Added tests for lowering and encoding. Differential Revision: http://reviews.llvm.org/D12317 llvm-svn: 249688
2015-10-08 20:55:01 +08:00
%val = load i16, i16* %ptr
%r1 = insertelement <8 x i16> %x, i16 %val, i32 1
%r2 = insertelement <8 x i16> %r1, i16 %y, i32 5
ret <8 x i16> %r2
}
define <64 x i8> @insert_v64i8(<64 x i8> %x, i8 %y, i8* %ptr) {
; KNL-LABEL: insert_v64i8:
[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
; KNL: ## %bb.0:
AVX512: vpextrb/w/d/q and vpinsrb/w/d/q implementation. This instructions doesn't have intrincis. Added tests for lowering and encoding. Differential Revision: http://reviews.llvm.org/D12317 llvm-svn: 249688
2015-10-08 20:55:01 +08:00
; KNL-NEXT: vpinsrb $1, (%rsi), %xmm0, %xmm2
; KNL-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]
; KNL-NEXT: vextracti128 $1, %ymm1, %xmm2
; KNL-NEXT: vpinsrb $2, %edi, %xmm2, %xmm2
; KNL-NEXT: vinserti128 $1, %xmm2, %ymm1, %ymm1
; KNL-NEXT: retq
;
; SKX-LABEL: insert_v64i8:
[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
; SKX: ## %bb.0:
AVX512: vpextrb/w/d/q and vpinsrb/w/d/q implementation. This instructions doesn't have intrincis. Added tests for lowering and encoding. Differential Revision: http://reviews.llvm.org/D12317 llvm-svn: 249688
2015-10-08 20:55:01 +08:00
; SKX-NEXT: vpinsrb $1, (%rsi), %xmm0, %xmm1
[DAGCombiner] Remove the half vector width check for the combine of EXTRACT_SUBVECTOR from an INSERT_SUBVECTOR. This gives more parallelism opportunities for AVX-512 when dealing with 128-bit extracts from 512-bit vectors. llvm-svn: 294930
2017-02-13 07:49:49 +08:00
; SKX-NEXT: vinserti32x4 $0, %xmm1, %zmm0, %zmm1
; SKX-NEXT: vextracti32x4 $3, %zmm0, %xmm0
; SKX-NEXT: vpinsrb $2, %edi, %xmm0, %xmm0
; SKX-NEXT: vinserti32x4 $3, %xmm0, %zmm1, %zmm0
AVX512: vpextrb/w/d/q and vpinsrb/w/d/q implementation. This instructions doesn't have intrincis. Added tests for lowering and encoding. Differential Revision: http://reviews.llvm.org/D12317 llvm-svn: 249688
2015-10-08 20:55:01 +08:00
; SKX-NEXT: retq
%val = load i8, i8* %ptr
%r1 = insertelement <64 x i8> %x, i8 %val, i32 1
%r2 = insertelement <64 x i8> %r1, i8 %y, i32 50
ret <64 x i8> %r2
}
define <32 x i8> @insert_v32i8(<32 x i8> %x, i8 %y, i8* %ptr) {
[AVX512] Add a common prefix to avx512-insert-extract.ll so we can reduce the number of check lines on some test cases. This was pointed out during the review for D313804. llvm-svn: 309629
2017-08-01 05:20:06 +08:00
; CHECK-LABEL: insert_v32i8:
[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] Add a common prefix to avx512-insert-extract.ll so we can reduce the number of check lines on some test cases. This was pointed out during the review for D313804. llvm-svn: 309629
2017-08-01 05:20:06 +08:00
; CHECK-NEXT: vpinsrb $1, (%rsi), %xmm0, %xmm1
; CHECK-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; CHECK-NEXT: vextracti128 $1, %ymm0, %xmm0
; CHECK-NEXT: vpinsrb $1, %edi, %xmm0, %xmm0
; CHECK-NEXT: vinserti128 $1, %xmm0, %ymm1, %ymm0
; CHECK-NEXT: retq
AVX512: vpextrb/w/d/q and vpinsrb/w/d/q implementation. This instructions doesn't have intrincis. Added tests for lowering and encoding. Differential Revision: http://reviews.llvm.org/D12317 llvm-svn: 249688
2015-10-08 20:55:01 +08:00
%val = load i8, i8* %ptr
%r1 = insertelement <32 x i8> %x, i8 %val, i32 1
%r2 = insertelement <32 x i8> %r1, i8 %y, i32 17
ret <32 x i8> %r2
}
define <16 x i8> @insert_v16i8(<16 x i8> %x, i8 %y, i8* %ptr) {
[AVX512] Add a common prefix to avx512-insert-extract.ll so we can reduce the number of check lines on some test cases. This was pointed out during the review for D313804. llvm-svn: 309629
2017-08-01 05:20:06 +08:00
; CHECK-LABEL: insert_v16i8:
[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] Add a common prefix to avx512-insert-extract.ll so we can reduce the number of check lines on some test cases. This was pointed out during the review for D313804. llvm-svn: 309629
2017-08-01 05:20:06 +08:00
; CHECK-NEXT: vpinsrb $3, (%rsi), %xmm0, %xmm0
; CHECK-NEXT: vpinsrb $10, %edi, %xmm0, %xmm0
; CHECK-NEXT: retq
AVX512: vpextrb/w/d/q and vpinsrb/w/d/q implementation. This instructions doesn't have intrincis. Added tests for lowering and encoding. Differential Revision: http://reviews.llvm.org/D12317 llvm-svn: 249688
2015-10-08 20:55:01 +08:00
%val = load i8, i8* %ptr
%r1 = insertelement <16 x i8> %x, i8 %val, i32 3
%r2 = insertelement <16 x i8> %r1, i8 %y, i32 10
ret <16 x i8> %r2
}
AVX512: Implement instructions encoding, lowering and intrinsics vinserti64x4, vinserti64x2, vinserti32x8, vinserti32x4, vinsertf64x4, vinsertf64x2, vinsertf32x8, vinsertf32x4 Added tests for encoding, lowering and intrinsics. Differential Revision: http://reviews.llvm.org/D11893 llvm-svn: 248111
2015-09-20 14:52:42 +08:00
define <8 x i64> @test_insert_128_v8i64(<8 x i64> %x, i64 %y) {
[AVX512] Don't switch unmasked subvector insert/extract instructions when AVX512DQI is enabled. There's no reason to switch instructions with and without DQI. It just creates extra isel patterns and test divergences. There is however value in enabling the masked version of the instructions with DQI. This required introducing some new multiclasses to enabling this splitting. Differential Revision: https://reviews.llvm.org/D36661 llvm-svn: 311091
2017-08-17 23:40:25 +08:00
; CHECK-LABEL: test_insert_128_v8i64:
[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] Don't switch unmasked subvector insert/extract instructions when AVX512DQI is enabled. There's no reason to switch instructions with and without DQI. It just creates extra isel patterns and test divergences. There is however value in enabling the masked version of the instructions with DQI. This required introducing some new multiclasses to enabling this splitting. Differential Revision: https://reviews.llvm.org/D36661 llvm-svn: 311091
2017-08-17 23:40:25 +08:00
; CHECK-NEXT: vpinsrq $1, %rdi, %xmm0, %xmm1
; CHECK-NEXT: vinserti32x4 $0, %xmm1, %zmm0, %zmm0
; CHECK-NEXT: retq
AVX512: Implement instructions encoding, lowering and intrinsics vinserti64x4, vinserti64x2, vinserti32x8, vinserti32x4, vinsertf64x4, vinsertf64x2, vinsertf32x8, vinsertf32x4 Added tests for encoding, lowering and intrinsics. Differential Revision: http://reviews.llvm.org/D11893 llvm-svn: 248111
2015-09-20 14:52:42 +08:00
%r = insertelement <8 x i64> %x, i64 %y, i32 1
ret <8 x i64> %r
}
define <16 x i32> @test_insert_128_v16i32(<16 x i32> %x, i32 %y) {
[AVX512] Add a common prefix to avx512-insert-extract.ll so we can reduce the number of check lines on some test cases. This was pointed out during the review for D313804. llvm-svn: 309629
2017-08-01 05:20:06 +08:00
; CHECK-LABEL: test_insert_128_v16i32:
[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] Add a common prefix to avx512-insert-extract.ll so we can reduce the number of check lines on some test cases. This was pointed out during the review for D313804. llvm-svn: 309629
2017-08-01 05:20:06 +08:00
; CHECK-NEXT: vpinsrd $1, %edi, %xmm0, %xmm1
; CHECK-NEXT: vinserti32x4 $0, %xmm1, %zmm0, %zmm0
; CHECK-NEXT: retq
AVX512: Implement instructions encoding, lowering and intrinsics vinserti64x4, vinserti64x2, vinserti32x8, vinserti32x4, vinsertf64x4, vinsertf64x2, vinsertf32x8, vinsertf32x4 Added tests for encoding, lowering and intrinsics. Differential Revision: http://reviews.llvm.org/D11893 llvm-svn: 248111
2015-09-20 14:52:42 +08:00
%r = insertelement <16 x i32> %x, i32 %y, i32 1
ret <16 x i32> %r
}
define <8 x double> @test_insert_128_v8f64(<8 x double> %x, double %y) {
[AVX512] Don't switch unmasked subvector insert/extract instructions when AVX512DQI is enabled. There's no reason to switch instructions with and without DQI. It just creates extra isel patterns and test divergences. There is however value in enabling the masked version of the instructions with DQI. This required introducing some new multiclasses to enabling this splitting. Differential Revision: https://reviews.llvm.org/D36661 llvm-svn: 311091
2017-08-17 23:40:25 +08:00
; CHECK-LABEL: test_insert_128_v8f64:
[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] Teach the execution domain fixing tables to use movlhps inplace of unpcklpd for the packed single domain. MOVLHPS has a smaller encoding than UNPCKLPD in the legacy encodings. With VEX and EVEX encodings it doesn't matter. llvm-svn: 313509
2017-09-18 12:40:58 +08:00
; CHECK-NEXT: vmovlhps {{.*#+}} xmm1 = xmm0[0],xmm1[0]
[AVX512] Don't switch unmasked subvector insert/extract instructions when AVX512DQI is enabled. There's no reason to switch instructions with and without DQI. It just creates extra isel patterns and test divergences. There is however value in enabling the masked version of the instructions with DQI. This required introducing some new multiclasses to enabling this splitting. Differential Revision: https://reviews.llvm.org/D36661 llvm-svn: 311091
2017-08-17 23:40:25 +08:00
; CHECK-NEXT: vinsertf32x4 $0, %xmm1, %zmm0, %zmm0
; CHECK-NEXT: retq
AVX512: Implement instructions encoding, lowering and intrinsics vinserti64x4, vinserti64x2, vinserti32x8, vinserti32x4, vinsertf64x4, vinsertf64x2, vinsertf32x8, vinsertf32x4 Added tests for encoding, lowering and intrinsics. Differential Revision: http://reviews.llvm.org/D11893 llvm-svn: 248111
2015-09-20 14:52:42 +08:00
%r = insertelement <8 x double> %x, double %y, i32 1
ret <8 x double> %r
}
define <16 x float> @test_insert_128_v16f32(<16 x float> %x, float %y) {
[AVX512] Add a common prefix to avx512-insert-extract.ll so we can reduce the number of check lines on some test cases. This was pointed out during the review for D313804. llvm-svn: 309629
2017-08-01 05:20:06 +08:00
; CHECK-LABEL: test_insert_128_v16f32:
[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] Add a common prefix to avx512-insert-extract.ll so we can reduce the number of check lines on some test cases. This was pointed out during the review for D313804. llvm-svn: 309629
2017-08-01 05:20:06 +08:00
; CHECK-NEXT: vinsertps {{.*#+}} xmm1 = xmm0[0],xmm1[0],xmm0[2,3]
; CHECK-NEXT: vinsertf32x4 $0, %xmm1, %zmm0, %zmm0
; CHECK-NEXT: retq
AVX512: Implement instructions encoding, lowering and intrinsics vinserti64x4, vinserti64x2, vinserti32x8, vinserti32x4, vinsertf64x4, vinsertf64x2, vinsertf32x8, vinsertf32x4 Added tests for encoding, lowering and intrinsics. Differential Revision: http://reviews.llvm.org/D11893 llvm-svn: 248111
2015-09-20 14:52:42 +08:00
%r = insertelement <16 x float> %x, float %y, i32 1
ret <16 x float> %r
}
define <16 x i16> @test_insert_128_v16i16(<16 x i16> %x, i16 %y) {
[AVX512] Add a common prefix to avx512-insert-extract.ll so we can reduce the number of check lines on some test cases. This was pointed out during the review for D313804. llvm-svn: 309629
2017-08-01 05:20:06 +08:00
; CHECK-LABEL: test_insert_128_v16i16:
[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] Add a common prefix to avx512-insert-extract.ll so we can reduce the number of check lines on some test cases. This was pointed out during the review for D313804. llvm-svn: 309629
2017-08-01 05:20:06 +08:00
; CHECK-NEXT: vextracti128 $1, %ymm0, %xmm1
; CHECK-NEXT: vpinsrw $2, %edi, %xmm1, %xmm1
; CHECK-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm0
; CHECK-NEXT: retq
AVX512: Implement instructions encoding, lowering and intrinsics vinserti64x4, vinserti64x2, vinserti32x8, vinserti32x4, vinsertf64x4, vinsertf64x2, vinsertf32x8, vinsertf32x4 Added tests for encoding, lowering and intrinsics. Differential Revision: http://reviews.llvm.org/D11893 llvm-svn: 248111
2015-09-20 14:52:42 +08:00
%r = insertelement <16 x i16> %x, i16 %y, i32 10
ret <16 x i16> %r
}
define <32 x i8> @test_insert_128_v32i8(<32 x i8> %x, i8 %y) {
[AVX512] Add a common prefix to avx512-insert-extract.ll so we can reduce the number of check lines on some test cases. This was pointed out during the review for D313804. llvm-svn: 309629
2017-08-01 05:20:06 +08:00
; CHECK-LABEL: test_insert_128_v32i8:
[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] Add a common prefix to avx512-insert-extract.ll so we can reduce the number of check lines on some test cases. This was pointed out during the review for D313804. llvm-svn: 309629
2017-08-01 05:20:06 +08:00
; CHECK-NEXT: vextracti128 $1, %ymm0, %xmm1
; CHECK-NEXT: vpinsrb $4, %edi, %xmm1, %xmm1
; CHECK-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm0
; CHECK-NEXT: retq
AVX512: Implement instructions encoding, lowering and intrinsics vinserti64x4, vinserti64x2, vinserti32x8, vinserti32x4, vinsertf64x4, vinsertf64x2, vinsertf32x8, vinsertf32x4 Added tests for encoding, lowering and intrinsics. Differential Revision: http://reviews.llvm.org/D11893 llvm-svn: 248111
2015-09-20 14:52:42 +08:00
%r = insertelement <32 x i8> %x, i8 %y, i32 20
ret <32 x i8> %r
}
[AVX512] Fix extractelement i1 lowering. The previous implementation (not custom) doesn't enforce zeroing off upper bits. The assumption is that i1 PRODUCER (truncate and extractelement) must zero all upper bits, so i1 CONSUMER instructions ( test, zext, save, etc) can be done without additional zeroing. Make extractelement i1 lowering custom for all vector i1. Differential Revision: http://reviews.llvm.org/D23246 llvm-svn: 278328
2016-08-11 20:13:46 +08:00
[AVX512] Fix insertelement i1 lowering. 1. Use shuffle to insert element i1 into vector. The previous implementation was incorrect ( dest_bit OR src_bit , it doesn't clear the bit if src_bit=0 ) 2. Improve shuffle i1 vector, use CVT2MASK if supported instead TRUNCATE. Differential Revision: http://reviews.llvm.org/D23347 llvm-svn: 278623
2016-08-14 13:25:07 +08:00
define i32 @test_insertelement_v32i1(i32 %a, i32 %b, <32 x i32> %x , <32 x i32> %y) {
Revert r274613 because it breaks the test suite with AVX512 This reverts most of r274613 (AKA r274626) and its follow-ups (r276347, r277289), due to miscompiles in the test suite. The FastISel change was left in, because it apparently fixes an unrelated issue. (Recommit of r279782 which was broken due to a bad merge.) This fixes 4 out of the 5 test failures in PR29112. llvm-svn: 279788
2016-08-26 06:48:11 +08:00
; KNL-LABEL: test_insertelement_v32i1:
[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
; KNL: ## %bb.0:
Reverting commit 306414 on behalf of @gadi.haber llvm-svn: 306532
2017-06-28 19:23:31 +08:00
; KNL-NEXT: cmpl %esi, %edi
[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
; KNL-NEXT: setb %al
[X86] Legalize v32i1 without BWI via splitting to v16i1 rather than the default of promoting to v32i8. Summary: For the most part its better to keep v32i1 as a mask type of a narrower width than trying to promote it to a ymm register. I had to add some overrides to the methods that get the types for the calling convention so that we still use v32i8 for argument/return purposes. There are still some regressions in here. I definitely saw some around shuffles. I think we probably should move vXi1 shuffle from lowering to a DAG combine where I think the extend and truncate we have to emit would be better combined. I think we also need a DAG combine to remove trunc from (extract_vector_elt (trunc)) Overall this removes something like 13000 CHECK lines from lit tests. Reviewers: zvi, RKSimon, delena, spatel Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D42031 llvm-svn: 323201
2018-01-23 22:25:39 +08:00
; KNL-NEXT: vpcmpltud %zmm3, %zmm1, %k0
; KNL-NEXT: kmovw %k0, %ecx
; KNL-NEXT: shll $16, %ecx
; KNL-NEXT: vpcmpltud %zmm2, %zmm0, %k0
; KNL-NEXT: kshiftrw $4, %k0, %k1
; KNL-NEXT: kmovw %eax, %k2
; KNL-NEXT: kxorw %k2, %k1, %k1
; KNL-NEXT: kshiftlw $15, %k1, %k1
; KNL-NEXT: kshiftrw $11, %k1, %k1
[X86] Rewrite vXi1 element insertion by using a vXi1 scalar_to_vector and inserting into a vXi1 vector. The existing code was already doing something very similar to subvector insertion so this allows us to remove the nearly duplicate code. This patch is a little larger than it should be due to differences between the DQI handling between the two today. llvm-svn: 323212
2018-01-23 23:56:36 +08:00
; KNL-NEXT: kxorw %k1, %k0, %k0
[X86] Legalize v32i1 without BWI via splitting to v16i1 rather than the default of promoting to v32i8. Summary: For the most part its better to keep v32i1 as a mask type of a narrower width than trying to promote it to a ymm register. I had to add some overrides to the methods that get the types for the calling convention so that we still use v32i8 for argument/return purposes. There are still some regressions in here. I definitely saw some around shuffles. I think we probably should move vXi1 shuffle from lowering to a DAG combine where I think the extend and truncate we have to emit would be better combined. I think we also need a DAG combine to remove trunc from (extract_vector_elt (trunc)) Overall this removes something like 13000 CHECK lines from lit tests. Reviewers: zvi, RKSimon, delena, spatel Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D42031 llvm-svn: 323201
2018-01-23 22:25:39 +08:00
; KNL-NEXT: kmovw %k0, %eax
; KNL-NEXT: orl %ecx, %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
; KNL-NEXT: vzeroupper
Revert r274613 because it breaks the test suite with AVX512 This reverts most of r274613 (AKA r274626) and its follow-ups (r276347, r277289), due to miscompiles in the test suite. The FastISel change was left in, because it apparently fixes an unrelated issue. (Recommit of r279782 which was broken due to a bad merge.) This fixes 4 out of the 5 test failures in PR29112. llvm-svn: 279788
2016-08-26 06:48:11 +08:00
; KNL-NEXT: retq
;
[AVX512] Fix insertelement i1 lowering. 1. Use shuffle to insert element i1 into vector. The previous implementation was incorrect ( dest_bit OR src_bit , it doesn't clear the bit if src_bit=0 ) 2. Improve shuffle i1 vector, use CVT2MASK if supported instead TRUNCATE. Differential Revision: http://reviews.llvm.org/D23347 llvm-svn: 278623
2016-08-14 13:25:07 +08:00
; SKX-LABEL: test_insertelement_v32i1:
[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
; SKX: ## %bb.0:
[AVX512] Fix insertelement i1 lowering. 1. Use shuffle to insert element i1 into vector. The previous implementation was incorrect ( dest_bit OR src_bit , it doesn't clear the bit if src_bit=0 ) 2. Improve shuffle i1 vector, use CVT2MASK if supported instead TRUNCATE. Differential Revision: http://reviews.llvm.org/D23347 llvm-svn: 278623
2016-08-14 13:25:07 +08:00
; SKX-NEXT: cmpl %esi, %edi
[x86] don't blindly transform SETB into SBB I noticed unnecessary 'sbb' instructions in D30472 and while looking at 'ptest' codegen recently. This happens because we were transforming any 'setb' - even when we only wanted a single-bit result. This patch moves those transforms under visitAdd/visitSub, so we we're only creating sbb/adc when it is a win. I don't know why we need a SETCC_CARRY node type, but I'm not proposing to change that existing behavior in this patch. Also, I'm skeptical that sbb/adc are a win for all micro-arches, so I added comments to the test files where this transform still fires. The test changes here are all cases where we no longer produce sbb/adc. Avoiding partial register stalls (generating an xor to clear a register) is not handled in some cases, but that's a separate issue. Differential Revision: https://reviews.llvm.org/D30611 llvm-svn: 297586
2017-03-13 02:28:48 +08:00
; SKX-NEXT: setb %al
[X86][AVX512] Make i1 illegal in the CodeGen This patch defines the i1 type as illegal in the X86 backend for AVX512. For DAG operations on <N x i1> types (build vector, extract vector element, ...) i8 is used, and should be truncated/extended. This should produce better scalar code for i1 types since GPRs will be used instead of mask registers. Differential Revision: https://reviews.llvm.org/D32273 llvm-svn: 303421
2017-05-19 20:35:15 +08:00
; SKX-NEXT: vpcmpltud %zmm2, %zmm0, %k0
; SKX-NEXT: vpcmpltud %zmm3, %zmm1, %k1
; SKX-NEXT: kunpckwd %k0, %k1, %k0
[X86] Handle alls version of vXi1 insert_vector_elt with a constant index without falling back to shuffles. We previously only supported inserting to the LSB or MSB where it was easy to zero to perform an OR to insert. This change effectively extracts the old value and the new value, xors them together and then xors that single bit with the correct location in the original vector. This will cancel out the old value in the first xor leaving the new value in the position. The way I've implemented this uses 3 shifts and two xors and uses an additional register. We can avoid the additional register at the cost of another shift. llvm-svn: 320120
2017-12-08 08:16:09 +08:00
; SKX-NEXT: kshiftrd $4, %k0, %k1
; SKX-NEXT: kmovd %eax, %k2
; SKX-NEXT: kxord %k2, %k1, %k1
; SKX-NEXT: kshiftld $31, %k1, %k1
; SKX-NEXT: kshiftrd $27, %k1, %k1
[X86] Rewrite vXi1 element insertion by using a vXi1 scalar_to_vector and inserting into a vXi1 vector. The existing code was already doing something very similar to subvector insertion so this allows us to remove the nearly duplicate code. This patch is a little larger than it should be due to differences between the DQI handling between the two today. llvm-svn: 323212
2018-01-23 23:56:36 +08:00
; SKX-NEXT: kxord %k1, %k0, %k0
[AVX512] Fix insertelement i1 lowering. 1. Use shuffle to insert element i1 into vector. The previous implementation was incorrect ( dest_bit OR src_bit , it doesn't clear the bit if src_bit=0 ) 2. Improve shuffle i1 vector, use CVT2MASK if supported instead TRUNCATE. Differential Revision: http://reviews.llvm.org/D23347 llvm-svn: 278623
2016-08-14 13:25:07 +08:00
; SKX-NEXT: kmovd %k0, %eax
[X86] Generate VZEROUPPER for Skylake-avx512. VZEROUPPER should not be issued on Knights Landing (KNL), but on Skylake-avx512 it should be. Differential Revision: https://reviews.llvm.org/D29874 llvm-svn: 296859
2017-03-03 17:03:24 +08:00
; SKX-NEXT: vzeroupper
[AVX512] Fix insertelement i1 lowering. 1. Use shuffle to insert element i1 into vector. The previous implementation was incorrect ( dest_bit OR src_bit , it doesn't clear the bit if src_bit=0 ) 2. Improve shuffle i1 vector, use CVT2MASK if supported instead TRUNCATE. Differential Revision: http://reviews.llvm.org/D23347 llvm-svn: 278623
2016-08-14 13:25:07 +08:00
; SKX-NEXT: retq
%cmp_res_i1 = icmp ult i32 %a, %b
%cmp_cmp_vec = icmp ult <32 x i32> %x, %y
%maskv = insertelement <32 x i1> %cmp_cmp_vec, i1 %cmp_res_i1, i32 4
%res = bitcast <32 x i1> %maskv to i32
ret i32 %res
}
define i8 @test_iinsertelement_v4i1(i32 %a, i32 %b, <4 x i32> %x , <4 x i32> %y) {
Revert r274613 because it breaks the test suite with AVX512 This reverts most of r274613 (AKA r274626) and its follow-ups (r276347, r277289), due to miscompiles in the test suite. The FastISel change was left in, because it apparently fixes an unrelated issue. (Recommit of r279782 which was broken due to a bad merge.) This fixes 4 out of the 5 test failures in PR29112. llvm-svn: 279788
2016-08-26 06:48:11 +08:00
; KNL-LABEL: test_iinsertelement_v4i1:
[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
; KNL: ## %bb.0:
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; KNL-NEXT: ## kill: def $xmm1 killed $xmm1 def $zmm1
; KNL-NEXT: ## kill: def $xmm0 killed $xmm0 def $zmm0
Revert r274613 because it breaks the test suite with AVX512 This reverts most of r274613 (AKA r274626) and its follow-ups (r276347, r277289), due to miscompiles in the test suite. The FastISel change was left in, because it apparently fixes an unrelated issue. (Recommit of r279782 which was broken due to a bad merge.) This fixes 4 out of the 5 test failures in PR29112. llvm-svn: 279788
2016-08-26 06:48:11 +08:00
; KNL-NEXT: cmpl %esi, %edi
[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
; KNL-NEXT: setb %al
[X86] Make v2i1 and v4i1 legal types without VLX Summary: There are few oddities that occur due to v1i1, v8i1, v16i1 being legal without v2i1 and v4i1 being legal when we don't have VLX. Particularly during legalization of v2i32/v4i32/v2i64/v4i64 masked gather/scatter/load/store. We end up promoting the mask argument to these during type legalization and then have to widen the promoted type to v8iX/v16iX and truncate it to get the element size back down to v8i1/v16i1 to use a 512-bit operation. Since need to fill the upper bits of the mask we have to fill with 0s at the promoted type. It would be better if we could just have the v2i1/v4i1 types as legal so they don't undergo any promotion. Then we can just widen with 0s directly in a k register. There are no real v4i1/v2i1 instructions anyway. Everything is done on a larger register anyway. This also fixes an issue that we couldn't implement a masked vextractf32x4 from zmm to xmm properly. We now have to support widening more compares to 512-bit to get a mask result out so new tablegen patterns got added. I had to hack the legalizer for widening the operand of a setcc a bit so it didn't try create a setcc returning v4i32, extract from it, then try to promote it using a sign extend to v2i1. Now we create the setcc with v4i1 if the original setcc's result type is v2i1. Then extract that and don't sign extend it at all. There's definitely room for improvement with some follow up patches. Reviewers: RKSimon, zvi, guyblank Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D41560 llvm-svn: 321967
2018-01-08 02:20:37 +08:00
; KNL-NEXT: vpcmpltud %zmm1, %zmm0, %k0
[X86] Handle alls version of vXi1 insert_vector_elt with a constant index without falling back to shuffles. We previously only supported inserting to the LSB or MSB where it was easy to zero to perform an OR to insert. This change effectively extracts the old value and the new value, xors them together and then xors that single bit with the correct location in the original vector. This will cancel out the old value in the first xor leaving the new value in the position. The way I've implemented this uses 3 shifts and two xors and uses an additional register. We can avoid the additional register at the cost of another shift. llvm-svn: 320120
2017-12-08 08:16:09 +08:00
; KNL-NEXT: kshiftrw $2, %k0, %k1
; KNL-NEXT: kmovw %eax, %k2
; KNL-NEXT: kxorw %k2, %k1, %k1
; KNL-NEXT: kshiftlw $15, %k1, %k1
; KNL-NEXT: kshiftrw $13, %k1, %k1
[X86] Rewrite vXi1 element insertion by using a vXi1 scalar_to_vector and inserting into a vXi1 vector. The existing code was already doing something very similar to subvector insertion so this allows us to remove the nearly duplicate code. This patch is a little larger than it should be due to differences between the DQI handling between the two today. llvm-svn: 323212
2018-01-23 23:56:36 +08:00
; KNL-NEXT: kxorw %k1, %k0, %k0
Revert r274613 because it breaks the test suite with AVX512 This reverts most of r274613 (AKA r274626) and its follow-ups (r276347, r277289), due to miscompiles in the test suite. The FastISel change was left in, because it apparently fixes an unrelated issue. (Recommit of r279782 which was broken due to a bad merge.) This fixes 4 out of the 5 test failures in PR29112. llvm-svn: 279788
2016-08-26 06:48:11 +08:00
; KNL-NEXT: kmovw %k0, %eax
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; KNL-NEXT: ## kill: def $al killed $al killed $eax
[X86] Make v2i1 and v4i1 legal types without VLX Summary: There are few oddities that occur due to v1i1, v8i1, v16i1 being legal without v2i1 and v4i1 being legal when we don't have VLX. Particularly during legalization of v2i32/v4i32/v2i64/v4i64 masked gather/scatter/load/store. We end up promoting the mask argument to these during type legalization and then have to widen the promoted type to v8iX/v16iX and truncate it to get the element size back down to v8i1/v16i1 to use a 512-bit operation. Since need to fill the upper bits of the mask we have to fill with 0s at the promoted type. It would be better if we could just have the v2i1/v4i1 types as legal so they don't undergo any promotion. Then we can just widen with 0s directly in a k register. There are no real v4i1/v2i1 instructions anyway. Everything is done on a larger register anyway. This also fixes an issue that we couldn't implement a masked vextractf32x4 from zmm to xmm properly. We now have to support widening more compares to 512-bit to get a mask result out so new tablegen patterns got added. I had to hack the legalizer for widening the operand of a setcc a bit so it didn't try create a setcc returning v4i32, extract from it, then try to promote it using a sign extend to v2i1. Now we create the setcc with v4i1 if the original setcc's result type is v2i1. Then extract that and don't sign extend it at all. There's definitely room for improvement with some follow up patches. Reviewers: RKSimon, zvi, guyblank Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D41560 llvm-svn: 321967
2018-01-08 02:20:37 +08:00
; KNL-NEXT: vzeroupper
Revert r274613 because it breaks the test suite with AVX512 This reverts most of r274613 (AKA r274626) and its follow-ups (r276347, r277289), due to miscompiles in the test suite. The FastISel change was left in, because it apparently fixes an unrelated issue. (Recommit of r279782 which was broken due to a bad merge.) This fixes 4 out of the 5 test failures in PR29112. llvm-svn: 279788
2016-08-26 06:48:11 +08:00
; KNL-NEXT: retq
;
[AVX512] Fix insertelement i1 lowering. 1. Use shuffle to insert element i1 into vector. The previous implementation was incorrect ( dest_bit OR src_bit , it doesn't clear the bit if src_bit=0 ) 2. Improve shuffle i1 vector, use CVT2MASK if supported instead TRUNCATE. Differential Revision: http://reviews.llvm.org/D23347 llvm-svn: 278623
2016-08-14 13:25:07 +08:00
; SKX-LABEL: test_iinsertelement_v4i1:
[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
; SKX: ## %bb.0:
[AVX512] Fix insertelement i1 lowering. 1. Use shuffle to insert element i1 into vector. The previous implementation was incorrect ( dest_bit OR src_bit , it doesn't clear the bit if src_bit=0 ) 2. Improve shuffle i1 vector, use CVT2MASK if supported instead TRUNCATE. Differential Revision: http://reviews.llvm.org/D23347 llvm-svn: 278623
2016-08-14 13:25:07 +08:00
; SKX-NEXT: cmpl %esi, %edi
[x86] don't blindly transform SETB into SBB I noticed unnecessary 'sbb' instructions in D30472 and while looking at 'ptest' codegen recently. This happens because we were transforming any 'setb' - even when we only wanted a single-bit result. This patch moves those transforms under visitAdd/visitSub, so we we're only creating sbb/adc when it is a win. I don't know why we need a SETCC_CARRY node type, but I'm not proposing to change that existing behavior in this patch. Also, I'm skeptical that sbb/adc are a win for all micro-arches, so I added comments to the test files where this transform still fires. The test changes here are all cases where we no longer produce sbb/adc. Avoiding partial register stalls (generating an xor to clear a register) is not handled in some cases, but that's a separate issue. Differential Revision: https://reviews.llvm.org/D30611 llvm-svn: 297586
2017-03-13 02:28:48 +08:00
; SKX-NEXT: setb %al
[X86][AVX512] Make i1 illegal in the CodeGen This patch defines the i1 type as illegal in the X86 backend for AVX512. For DAG operations on <N x i1> types (build vector, extract vector element, ...) i8 is used, and should be truncated/extended. This should produce better scalar code for i1 types since GPRs will be used instead of mask registers. Differential Revision: https://reviews.llvm.org/D32273 llvm-svn: 303421
2017-05-19 20:35:15 +08:00
; SKX-NEXT: vpcmpltud %xmm1, %xmm0, %k0
[X86] Rewrite vXi1 element insertion by using a vXi1 scalar_to_vector and inserting into a vXi1 vector. The existing code was already doing something very similar to subvector insertion so this allows us to remove the nearly duplicate code. This patch is a little larger than it should be due to differences between the DQI handling between the two today. llvm-svn: 323212
2018-01-23 23:56:36 +08:00
; SKX-NEXT: kshiftrb $2, %k0, %k1
[X86] Handle alls version of vXi1 insert_vector_elt with a constant index without falling back to shuffles. We previously only supported inserting to the LSB or MSB where it was easy to zero to perform an OR to insert. This change effectively extracts the old value and the new value, xors them together and then xors that single bit with the correct location in the original vector. This will cancel out the old value in the first xor leaving the new value in the position. The way I've implemented this uses 3 shifts and two xors and uses an additional register. We can avoid the additional register at the cost of another shift. llvm-svn: 320120
2017-12-08 08:16:09 +08:00
; SKX-NEXT: kmovd %eax, %k2
[X86] Rewrite vXi1 element insertion by using a vXi1 scalar_to_vector and inserting into a vXi1 vector. The existing code was already doing something very similar to subvector insertion so this allows us to remove the nearly duplicate code. This patch is a little larger than it should be due to differences between the DQI handling between the two today. llvm-svn: 323212
2018-01-23 23:56:36 +08:00
; SKX-NEXT: kxorb %k2, %k1, %k1
; SKX-NEXT: kshiftlb $7, %k1, %k1
; SKX-NEXT: kshiftrb $5, %k1, %k1
[DAGCombiner] narrow vector binops when extraction is cheap Narrowing vector binops came up in the demanded bits discussion in D52912. I don't think we're going to be able to do this transform in IR as a canonicalization because of the risk of creating unsupported widths for vector ops, but we already have a DAG TLI hook to allow what I was hoping for: isExtractSubvectorCheap(). This is currently enabled for x86, ARM, and AArch64 (although only x86 has existing regression test diffs). This is artificially limited to not look through bitcasts because there are so many test diffs already, but that's marked with a TODO and is a small follow-up. Differential Revision: https://reviews.llvm.org/D53784 llvm-svn: 345602
2018-10-30 22:14:34 +08:00
; SKX-NEXT: kxorw %k1, %k0, %k0
[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
; SKX-NEXT: kmovd %k0, %eax
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; SKX-NEXT: ## kill: def $al killed $al killed $eax
[AVX512] Fix insertelement i1 lowering. 1. Use shuffle to insert element i1 into vector. The previous implementation was incorrect ( dest_bit OR src_bit , it doesn't clear the bit if src_bit=0 ) 2. Improve shuffle i1 vector, use CVT2MASK if supported instead TRUNCATE. Differential Revision: http://reviews.llvm.org/D23347 llvm-svn: 278623
2016-08-14 13:25:07 +08:00
; SKX-NEXT: retq
%cmp_res_i1 = icmp ult i32 %a, %b
%cmp_cmp_vec = icmp ult <4 x i32> %x, %y
%maskv = insertelement <4 x i1> %cmp_cmp_vec, i1 %cmp_res_i1, i32 2
%res0 = shufflevector <4 x i1> %maskv, <4 x i1> undef , <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 4, i32 4, i32 4>
%res = bitcast <8 x i1> %res0 to i8
ret i8 %res
}
define i8 @test_iinsertelement_v2i1(i32 %a, i32 %b, <2 x i64> %x , <2 x i64> %y) {
Revert r274613 because it breaks the test suite with AVX512 This reverts most of r274613 (AKA r274626) and its follow-ups (r276347, r277289), due to miscompiles in the test suite. The FastISel change was left in, because it apparently fixes an unrelated issue. (Recommit of r279782 which was broken due to a bad merge.) This fixes 4 out of the 5 test failures in PR29112. llvm-svn: 279788
2016-08-26 06:48:11 +08:00
; KNL-LABEL: test_iinsertelement_v2i1:
[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
; KNL: ## %bb.0:
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; KNL-NEXT: ## kill: def $xmm1 killed $xmm1 def $zmm1
; KNL-NEXT: ## kill: def $xmm0 killed $xmm0 def $zmm0
Revert r274613 because it breaks the test suite with AVX512 This reverts most of r274613 (AKA r274626) and its follow-ups (r276347, r277289), due to miscompiles in the test suite. The FastISel change was left in, because it apparently fixes an unrelated issue. (Recommit of r279782 which was broken due to a bad merge.) This fixes 4 out of the 5 test failures in PR29112. llvm-svn: 279788
2016-08-26 06:48:11 +08:00
; KNL-NEXT: cmpl %esi, %edi
[x86] don't blindly transform SETB into SBB I noticed unnecessary 'sbb' instructions in D30472 and while looking at 'ptest' codegen recently. This happens because we were transforming any 'setb' - even when we only wanted a single-bit result. This patch moves those transforms under visitAdd/visitSub, so we we're only creating sbb/adc when it is a win. I don't know why we need a SETCC_CARRY node type, but I'm not proposing to change that existing behavior in this patch. Also, I'm skeptical that sbb/adc are a win for all micro-arches, so I added comments to the test files where this transform still fires. The test changes here are all cases where we no longer produce sbb/adc. Avoiding partial register stalls (generating an xor to clear a register) is not handled in some cases, but that's a separate issue. Differential Revision: https://reviews.llvm.org/D30611 llvm-svn: 297586
2017-03-13 02:28:48 +08:00
; KNL-NEXT: setb %al
[X86] Make v2i1 and v4i1 legal types without VLX Summary: There are few oddities that occur due to v1i1, v8i1, v16i1 being legal without v2i1 and v4i1 being legal when we don't have VLX. Particularly during legalization of v2i32/v4i32/v2i64/v4i64 masked gather/scatter/load/store. We end up promoting the mask argument to these during type legalization and then have to widen the promoted type to v8iX/v16iX and truncate it to get the element size back down to v8i1/v16i1 to use a 512-bit operation. Since need to fill the upper bits of the mask we have to fill with 0s at the promoted type. It would be better if we could just have the v2i1/v4i1 types as legal so they don't undergo any promotion. Then we can just widen with 0s directly in a k register. There are no real v4i1/v2i1 instructions anyway. Everything is done on a larger register anyway. This also fixes an issue that we couldn't implement a masked vextractf32x4 from zmm to xmm properly. We now have to support widening more compares to 512-bit to get a mask result out so new tablegen patterns got added. I had to hack the legalizer for widening the operand of a setcc a bit so it didn't try create a setcc returning v4i32, extract from it, then try to promote it using a sign extend to v2i1. Now we create the setcc with v4i1 if the original setcc's result type is v2i1. Then extract that and don't sign extend it at all. There's definitely room for improvement with some follow up patches. Reviewers: RKSimon, zvi, guyblank Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D41560 llvm-svn: 321967
2018-01-08 02:20:37 +08:00
; KNL-NEXT: vpcmpltuq %zmm1, %zmm0, %k0
[X86] Rewrite vXi1 element insertion by using a vXi1 scalar_to_vector and inserting into a vXi1 vector. The existing code was already doing something very similar to subvector insertion so this allows us to remove the nearly duplicate code. This patch is a little larger than it should be due to differences between the DQI handling between the two today. llvm-svn: 323212
2018-01-23 23:56:36 +08:00
; KNL-NEXT: kshiftlw $15, %k0, %k0
; KNL-NEXT: kshiftrw $15, %k0, %k0
; KNL-NEXT: kmovw %eax, %k1
; KNL-NEXT: kshiftlw $1, %k1, %k1
; KNL-NEXT: korw %k1, %k0, %k0
Revert r274613 because it breaks the test suite with AVX512 This reverts most of r274613 (AKA r274626) and its follow-ups (r276347, r277289), due to miscompiles in the test suite. The FastISel change was left in, because it apparently fixes an unrelated issue. (Recommit of r279782 which was broken due to a bad merge.) This fixes 4 out of the 5 test failures in PR29112. llvm-svn: 279788
2016-08-26 06:48:11 +08:00
; KNL-NEXT: kmovw %k0, %eax
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; KNL-NEXT: ## kill: def $al killed $al killed $eax
[X86] Make v2i1 and v4i1 legal types without VLX Summary: There are few oddities that occur due to v1i1, v8i1, v16i1 being legal without v2i1 and v4i1 being legal when we don't have VLX. Particularly during legalization of v2i32/v4i32/v2i64/v4i64 masked gather/scatter/load/store. We end up promoting the mask argument to these during type legalization and then have to widen the promoted type to v8iX/v16iX and truncate it to get the element size back down to v8i1/v16i1 to use a 512-bit operation. Since need to fill the upper bits of the mask we have to fill with 0s at the promoted type. It would be better if we could just have the v2i1/v4i1 types as legal so they don't undergo any promotion. Then we can just widen with 0s directly in a k register. There are no real v4i1/v2i1 instructions anyway. Everything is done on a larger register anyway. This also fixes an issue that we couldn't implement a masked vextractf32x4 from zmm to xmm properly. We now have to support widening more compares to 512-bit to get a mask result out so new tablegen patterns got added. I had to hack the legalizer for widening the operand of a setcc a bit so it didn't try create a setcc returning v4i32, extract from it, then try to promote it using a sign extend to v2i1. Now we create the setcc with v4i1 if the original setcc's result type is v2i1. Then extract that and don't sign extend it at all. There's definitely room for improvement with some follow up patches. Reviewers: RKSimon, zvi, guyblank Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D41560 llvm-svn: 321967
2018-01-08 02:20:37 +08:00
; KNL-NEXT: vzeroupper
Revert r274613 because it breaks the test suite with AVX512 This reverts most of r274613 (AKA r274626) and its follow-ups (r276347, r277289), due to miscompiles in the test suite. The FastISel change was left in, because it apparently fixes an unrelated issue. (Recommit of r279782 which was broken due to a bad merge.) This fixes 4 out of the 5 test failures in PR29112. llvm-svn: 279788
2016-08-26 06:48:11 +08:00
; KNL-NEXT: retq
;
[AVX512] Fix insertelement i1 lowering. 1. Use shuffle to insert element i1 into vector. The previous implementation was incorrect ( dest_bit OR src_bit , it doesn't clear the bit if src_bit=0 ) 2. Improve shuffle i1 vector, use CVT2MASK if supported instead TRUNCATE. Differential Revision: http://reviews.llvm.org/D23347 llvm-svn: 278623
2016-08-14 13:25:07 +08:00
; SKX-LABEL: test_iinsertelement_v2i1:
[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
; SKX: ## %bb.0:
[AVX512] Fix insertelement i1 lowering. 1. Use shuffle to insert element i1 into vector. The previous implementation was incorrect ( dest_bit OR src_bit , it doesn't clear the bit if src_bit=0 ) 2. Improve shuffle i1 vector, use CVT2MASK if supported instead TRUNCATE. Differential Revision: http://reviews.llvm.org/D23347 llvm-svn: 278623
2016-08-14 13:25:07 +08:00
; SKX-NEXT: cmpl %esi, %edi
[x86] don't blindly transform SETB into SBB I noticed unnecessary 'sbb' instructions in D30472 and while looking at 'ptest' codegen recently. This happens because we were transforming any 'setb' - even when we only wanted a single-bit result. This patch moves those transforms under visitAdd/visitSub, so we we're only creating sbb/adc when it is a win. I don't know why we need a SETCC_CARRY node type, but I'm not proposing to change that existing behavior in this patch. Also, I'm skeptical that sbb/adc are a win for all micro-arches, so I added comments to the test files where this transform still fires. The test changes here are all cases where we no longer produce sbb/adc. Avoiding partial register stalls (generating an xor to clear a register) is not handled in some cases, but that's a separate issue. Differential Revision: https://reviews.llvm.org/D30611 llvm-svn: 297586
2017-03-13 02:28:48 +08:00
; SKX-NEXT: setb %al
[X86][AVX512] Make i1 illegal in the CodeGen This patch defines the i1 type as illegal in the X86 backend for AVX512. For DAG operations on <N x i1> types (build vector, extract vector element, ...) i8 is used, and should be truncated/extended. This should produce better scalar code for i1 types since GPRs will be used instead of mask registers. Differential Revision: https://reviews.llvm.org/D32273 llvm-svn: 303421
2017-05-19 20:35:15 +08:00
; SKX-NEXT: vpcmpltuq %xmm1, %xmm0, %k0
[X86] Rewrite vXi1 element insertion by using a vXi1 scalar_to_vector and inserting into a vXi1 vector. The existing code was already doing something very similar to subvector insertion so this allows us to remove the nearly duplicate code. This patch is a little larger than it should be due to differences between the DQI handling between the two today. llvm-svn: 323212
2018-01-23 23:56:36 +08:00
; SKX-NEXT: kshiftlb $7, %k0, %k0
; SKX-NEXT: kshiftrb $7, %k0, %k0
; SKX-NEXT: kmovd %eax, %k1
; SKX-NEXT: kshiftlb $1, %k1, %k1
[DAGCombiner] narrow vector binops when extraction is cheap Narrowing vector binops came up in the demanded bits discussion in D52912. I don't think we're going to be able to do this transform in IR as a canonicalization because of the risk of creating unsupported widths for vector ops, but we already have a DAG TLI hook to allow what I was hoping for: isExtractSubvectorCheap(). This is currently enabled for x86, ARM, and AArch64 (although only x86 has existing regression test diffs). This is artificially limited to not look through bitcasts because there are so many test diffs already, but that's marked with a TODO and is a small follow-up. Differential Revision: https://reviews.llvm.org/D53784 llvm-svn: 345602
2018-10-30 22:14:34 +08:00
; SKX-NEXT: korw %k1, %k0, %k0
[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
; SKX-NEXT: kmovd %k0, %eax
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; SKX-NEXT: ## kill: def $al killed $al killed $eax
[AVX512] Fix insertelement i1 lowering. 1. Use shuffle to insert element i1 into vector. The previous implementation was incorrect ( dest_bit OR src_bit , it doesn't clear the bit if src_bit=0 ) 2. Improve shuffle i1 vector, use CVT2MASK if supported instead TRUNCATE. Differential Revision: http://reviews.llvm.org/D23347 llvm-svn: 278623
2016-08-14 13:25:07 +08:00
; SKX-NEXT: retq
%cmp_res_i1 = icmp ult i32 %a, %b
%cmp_cmp_vec = icmp ult <2 x i64> %x, %y
%maskv = insertelement <2 x i1> %cmp_cmp_vec, i1 %cmp_res_i1, i32 1
%res0 = shufflevector <2 x i1> %maskv, <2 x i1> undef , <8 x i32> <i32 0, i32 1, i32 2, i32 2, i32 2, i32 2, i32 2, i32 2>
%res = bitcast <8 x i1> %res0 to i8
ret i8 %res
}
[AVX512] Fix extractelement i1 lowering. The previous implementation (not custom) doesn't enforce zeroing off upper bits. The assumption is that i1 PRODUCER (truncate and extractelement) must zero all upper bits, so i1 CONSUMER instructions ( test, zext, save, etc) can be done without additional zeroing. Make extractelement i1 lowering custom for all vector i1. Differential Revision: http://reviews.llvm.org/D23246 llvm-svn: 278328
2016-08-11 20:13:46 +08:00
define zeroext i8 @test_extractelement_v2i1(<2 x i64> %a, <2 x i64> %b) {
; KNL-LABEL: test_extractelement_v2i1:
[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
; KNL: ## %bb.0:
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; KNL-NEXT: ## kill: def $xmm1 killed $xmm1 def $zmm1
; KNL-NEXT: ## kill: def $xmm0 killed $xmm0 def $zmm0
[X86] Make v2i1 and v4i1 legal types without VLX Summary: There are few oddities that occur due to v1i1, v8i1, v16i1 being legal without v2i1 and v4i1 being legal when we don't have VLX. Particularly during legalization of v2i32/v4i32/v2i64/v4i64 masked gather/scatter/load/store. We end up promoting the mask argument to these during type legalization and then have to widen the promoted type to v8iX/v16iX and truncate it to get the element size back down to v8i1/v16i1 to use a 512-bit operation. Since need to fill the upper bits of the mask we have to fill with 0s at the promoted type. It would be better if we could just have the v2i1/v4i1 types as legal so they don't undergo any promotion. Then we can just widen with 0s directly in a k register. There are no real v4i1/v2i1 instructions anyway. Everything is done on a larger register anyway. This also fixes an issue that we couldn't implement a masked vextractf32x4 from zmm to xmm properly. We now have to support widening more compares to 512-bit to get a mask result out so new tablegen patterns got added. I had to hack the legalizer for widening the operand of a setcc a bit so it didn't try create a setcc returning v4i32, extract from it, then try to promote it using a sign extend to v2i1. Now we create the setcc with v4i1 if the original setcc's result type is v2i1. Then extract that and don't sign extend it at all. There's definitely room for improvement with some follow up patches. Reviewers: RKSimon, zvi, guyblank Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D41560 llvm-svn: 321967
2018-01-08 02:20:37 +08:00
; KNL-NEXT: vpcmpnleuq %zmm1, %zmm0, %k0
; KNL-NEXT: kmovw %k0, %eax
[X86][AVX512] Make i1 illegal in the CodeGen This patch defines the i1 type as illegal in the X86 backend for AVX512. For DAG operations on <N x i1> types (build vector, extract vector element, ...) i8 is used, and should be truncated/extended. This should produce better scalar code for i1 types since GPRs will be used instead of mask registers. Differential Revision: https://reviews.llvm.org/D32273 llvm-svn: 303421
2017-05-19 20:35:15 +08:00
; KNL-NEXT: andb $1, %al
; KNL-NEXT: movb $4, %cl
; KNL-NEXT: subb %al, %cl
; KNL-NEXT: movzbl %cl, %eax
[X86] Make v2i1 and v4i1 legal types without VLX Summary: There are few oddities that occur due to v1i1, v8i1, v16i1 being legal without v2i1 and v4i1 being legal when we don't have VLX. Particularly during legalization of v2i32/v4i32/v2i64/v4i64 masked gather/scatter/load/store. We end up promoting the mask argument to these during type legalization and then have to widen the promoted type to v8iX/v16iX and truncate it to get the element size back down to v8i1/v16i1 to use a 512-bit operation. Since need to fill the upper bits of the mask we have to fill with 0s at the promoted type. It would be better if we could just have the v2i1/v4i1 types as legal so they don't undergo any promotion. Then we can just widen with 0s directly in a k register. There are no real v4i1/v2i1 instructions anyway. Everything is done on a larger register anyway. This also fixes an issue that we couldn't implement a masked vextractf32x4 from zmm to xmm properly. We now have to support widening more compares to 512-bit to get a mask result out so new tablegen patterns got added. I had to hack the legalizer for widening the operand of a setcc a bit so it didn't try create a setcc returning v4i32, extract from it, then try to promote it using a sign extend to v2i1. Now we create the setcc with v4i1 if the original setcc's result type is v2i1. Then extract that and don't sign extend it at all. There's definitely room for improvement with some follow up patches. Reviewers: RKSimon, zvi, guyblank Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D41560 llvm-svn: 321967
2018-01-08 02:20:37 +08:00
; KNL-NEXT: vzeroupper
[AVX512] Fix extractelement i1 lowering. The previous implementation (not custom) doesn't enforce zeroing off upper bits. The assumption is that i1 PRODUCER (truncate and extractelement) must zero all upper bits, so i1 CONSUMER instructions ( test, zext, save, etc) can be done without additional zeroing. Make extractelement i1 lowering custom for all vector i1. Differential Revision: http://reviews.llvm.org/D23246 llvm-svn: 278328
2016-08-11 20:13:46 +08:00
; KNL-NEXT: retq
;
; SKX-LABEL: test_extractelement_v2i1:
[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
; SKX: ## %bb.0:
[AVX512] Fix extractelement i1 lowering. The previous implementation (not custom) doesn't enforce zeroing off upper bits. The assumption is that i1 PRODUCER (truncate and extractelement) must zero all upper bits, so i1 CONSUMER instructions ( test, zext, save, etc) can be done without additional zeroing. Make extractelement i1 lowering custom for all vector i1. Differential Revision: http://reviews.llvm.org/D23246 llvm-svn: 278328
2016-08-11 20:13:46 +08:00
; SKX-NEXT: vpcmpnleuq %xmm1, %xmm0, %k0
[AVX-512] Remove explicit KMOVWrk/KMOVWKr instructions from patterns where we can just use COPY_TO_REGCLASS instead. This will result in a KMOVW or KMOVD being emitted during register allocation. And in at least some cases this might allow the register coalescer to remove the copy all together. llvm-svn: 298984
2017-03-29 14:55:28 +08:00
; SKX-NEXT: kmovd %k0, %eax
[X86][AVX512] Make i1 illegal in the CodeGen This patch defines the i1 type as illegal in the X86 backend for AVX512. For DAG operations on <N x i1> types (build vector, extract vector element, ...) i8 is used, and should be truncated/extended. This should produce better scalar code for i1 types since GPRs will be used instead of mask registers. Differential Revision: https://reviews.llvm.org/D32273 llvm-svn: 303421
2017-05-19 20:35:15 +08:00
; SKX-NEXT: andb $1, %al
; SKX-NEXT: movb $4, %cl
; SKX-NEXT: subb %al, %cl
; SKX-NEXT: movzbl %cl, %eax
[AVX512] Fix extractelement i1 lowering. The previous implementation (not custom) doesn't enforce zeroing off upper bits. The assumption is that i1 PRODUCER (truncate and extractelement) must zero all upper bits, so i1 CONSUMER instructions ( test, zext, save, etc) can be done without additional zeroing. Make extractelement i1 lowering custom for all vector i1. Differential Revision: http://reviews.llvm.org/D23246 llvm-svn: 278328
2016-08-11 20:13:46 +08:00
; SKX-NEXT: retq
%t1 = icmp ugt <2 x i64> %a, %b
%t2 = extractelement <2 x i1> %t1, i32 0
%res = select i1 %t2, i8 3, i8 4
ret i8 %res
}
[x86] add alternate IR tests for select of constants; NFC llvm-svn: 296496
2017-03-01 02:02:38 +08:00
define zeroext i8 @extractelement_v2i1_alt(<2 x i64> %a, <2 x i64> %b) {
; KNL-LABEL: extractelement_v2i1_alt:
[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
; KNL: ## %bb.0:
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; KNL-NEXT: ## kill: def $xmm1 killed $xmm1 def $zmm1
; KNL-NEXT: ## kill: def $xmm0 killed $xmm0 def $zmm0
[X86] Make v2i1 and v4i1 legal types without VLX Summary: There are few oddities that occur due to v1i1, v8i1, v16i1 being legal without v2i1 and v4i1 being legal when we don't have VLX. Particularly during legalization of v2i32/v4i32/v2i64/v4i64 masked gather/scatter/load/store. We end up promoting the mask argument to these during type legalization and then have to widen the promoted type to v8iX/v16iX and truncate it to get the element size back down to v8i1/v16i1 to use a 512-bit operation. Since need to fill the upper bits of the mask we have to fill with 0s at the promoted type. It would be better if we could just have the v2i1/v4i1 types as legal so they don't undergo any promotion. Then we can just widen with 0s directly in a k register. There are no real v4i1/v2i1 instructions anyway. Everything is done on a larger register anyway. This also fixes an issue that we couldn't implement a masked vextractf32x4 from zmm to xmm properly. We now have to support widening more compares to 512-bit to get a mask result out so new tablegen patterns got added. I had to hack the legalizer for widening the operand of a setcc a bit so it didn't try create a setcc returning v4i32, extract from it, then try to promote it using a sign extend to v2i1. Now we create the setcc with v4i1 if the original setcc's result type is v2i1. Then extract that and don't sign extend it at all. There's definitely room for improvement with some follow up patches. Reviewers: RKSimon, zvi, guyblank Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D41560 llvm-svn: 321967
2018-01-08 02:20:37 +08:00
; KNL-NEXT: vpcmpnleuq %zmm1, %zmm0, %k0
; KNL-NEXT: kmovw %k0, %eax
[X86][AVX512] Make i1 illegal in the CodeGen This patch defines the i1 type as illegal in the X86 backend for AVX512. For DAG operations on <N x i1> types (build vector, extract vector element, ...) i8 is used, and should be truncated/extended. This should produce better scalar code for i1 types since GPRs will be used instead of mask registers. Differential Revision: https://reviews.llvm.org/D32273 llvm-svn: 303421
2017-05-19 20:35:15 +08:00
; KNL-NEXT: andb $1, %al
; KNL-NEXT: movb $4, %cl
; KNL-NEXT: subb %al, %cl
; KNL-NEXT: movzbl %cl, %eax
[X86] Make v2i1 and v4i1 legal types without VLX Summary: There are few oddities that occur due to v1i1, v8i1, v16i1 being legal without v2i1 and v4i1 being legal when we don't have VLX. Particularly during legalization of v2i32/v4i32/v2i64/v4i64 masked gather/scatter/load/store. We end up promoting the mask argument to these during type legalization and then have to widen the promoted type to v8iX/v16iX and truncate it to get the element size back down to v8i1/v16i1 to use a 512-bit operation. Since need to fill the upper bits of the mask we have to fill with 0s at the promoted type. It would be better if we could just have the v2i1/v4i1 types as legal so they don't undergo any promotion. Then we can just widen with 0s directly in a k register. There are no real v4i1/v2i1 instructions anyway. Everything is done on a larger register anyway. This also fixes an issue that we couldn't implement a masked vextractf32x4 from zmm to xmm properly. We now have to support widening more compares to 512-bit to get a mask result out so new tablegen patterns got added. I had to hack the legalizer for widening the operand of a setcc a bit so it didn't try create a setcc returning v4i32, extract from it, then try to promote it using a sign extend to v2i1. Now we create the setcc with v4i1 if the original setcc's result type is v2i1. Then extract that and don't sign extend it at all. There's definitely room for improvement with some follow up patches. Reviewers: RKSimon, zvi, guyblank Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D41560 llvm-svn: 321967
2018-01-08 02:20:37 +08:00
; KNL-NEXT: vzeroupper
[x86] add alternate IR tests for select of constants; NFC llvm-svn: 296496
2017-03-01 02:02:38 +08:00
; KNL-NEXT: retq
;
; SKX-LABEL: extractelement_v2i1_alt:
[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
; SKX: ## %bb.0:
[x86] add alternate IR tests for select of constants; NFC llvm-svn: 296496
2017-03-01 02:02:38 +08:00
; SKX-NEXT: vpcmpnleuq %xmm1, %xmm0, %k0
[AVX-512] Remove explicit KMOVWrk/KMOVWKr instructions from patterns where we can just use COPY_TO_REGCLASS instead. This will result in a KMOVW or KMOVD being emitted during register allocation. And in at least some cases this might allow the register coalescer to remove the copy all together. llvm-svn: 298984
2017-03-29 14:55:28 +08:00
; SKX-NEXT: kmovd %k0, %eax
[X86][AVX512] Make i1 illegal in the CodeGen This patch defines the i1 type as illegal in the X86 backend for AVX512. For DAG operations on <N x i1> types (build vector, extract vector element, ...) i8 is used, and should be truncated/extended. This should produce better scalar code for i1 types since GPRs will be used instead of mask registers. Differential Revision: https://reviews.llvm.org/D32273 llvm-svn: 303421
2017-05-19 20:35:15 +08:00
; SKX-NEXT: andb $1, %al
; SKX-NEXT: movb $4, %cl
; SKX-NEXT: subb %al, %cl
; SKX-NEXT: movzbl %cl, %eax
[x86] add alternate IR tests for select of constants; NFC llvm-svn: 296496
2017-03-01 02:02:38 +08:00
; SKX-NEXT: retq
%t1 = icmp ugt <2 x i64> %a, %b
%t2 = extractelement <2 x i1> %t1, i32 0
%sext = sext i1 %t2 to i8
%res = add i8 %sext, 4
ret i8 %res
}
[AVX512] Fix extractelement i1 lowering. The previous implementation (not custom) doesn't enforce zeroing off upper bits. The assumption is that i1 PRODUCER (truncate and extractelement) must zero all upper bits, so i1 CONSUMER instructions ( test, zext, save, etc) can be done without additional zeroing. Make extractelement i1 lowering custom for all vector i1. Differential Revision: http://reviews.llvm.org/D23246 llvm-svn: 278328
2016-08-11 20:13:46 +08:00
define zeroext i8 @test_extractelement_v4i1(<4 x i32> %a, <4 x i32> %b) {
; KNL-LABEL: test_extractelement_v4i1:
[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
; KNL: ## %bb.0:
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; KNL-NEXT: ## kill: def $xmm1 killed $xmm1 def $zmm1
; KNL-NEXT: ## kill: def $xmm0 killed $xmm0 def $zmm0
[X86] Make v2i1 and v4i1 legal types without VLX Summary: There are few oddities that occur due to v1i1, v8i1, v16i1 being legal without v2i1 and v4i1 being legal when we don't have VLX. Particularly during legalization of v2i32/v4i32/v2i64/v4i64 masked gather/scatter/load/store. We end up promoting the mask argument to these during type legalization and then have to widen the promoted type to v8iX/v16iX and truncate it to get the element size back down to v8i1/v16i1 to use a 512-bit operation. Since need to fill the upper bits of the mask we have to fill with 0s at the promoted type. It would be better if we could just have the v2i1/v4i1 types as legal so they don't undergo any promotion. Then we can just widen with 0s directly in a k register. There are no real v4i1/v2i1 instructions anyway. Everything is done on a larger register anyway. This also fixes an issue that we couldn't implement a masked vextractf32x4 from zmm to xmm properly. We now have to support widening more compares to 512-bit to get a mask result out so new tablegen patterns got added. I had to hack the legalizer for widening the operand of a setcc a bit so it didn't try create a setcc returning v4i32, extract from it, then try to promote it using a sign extend to v2i1. Now we create the setcc with v4i1 if the original setcc's result type is v2i1. Then extract that and don't sign extend it at all. There's definitely room for improvement with some follow up patches. Reviewers: RKSimon, zvi, guyblank Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D41560 llvm-svn: 321967
2018-01-08 02:20:37 +08:00
; KNL-NEXT: vpcmpnleud %zmm1, %zmm0, %k0
; KNL-NEXT: kshiftrw $3, %k0, %k0
; KNL-NEXT: kmovw %k0, %eax
[AVX512] Fix extractelement i1 lowering. The previous implementation (not custom) doesn't enforce zeroing off upper bits. The assumption is that i1 PRODUCER (truncate and extractelement) must zero all upper bits, so i1 CONSUMER instructions ( test, zext, save, etc) can be done without additional zeroing. Make extractelement i1 lowering custom for all vector i1. Differential Revision: http://reviews.llvm.org/D23246 llvm-svn: 278328
2016-08-11 20:13:46 +08:00
; KNL-NEXT: andl $1, %eax
[X86] Make v2i1 and v4i1 legal types without VLX Summary: There are few oddities that occur due to v1i1, v8i1, v16i1 being legal without v2i1 and v4i1 being legal when we don't have VLX. Particularly during legalization of v2i32/v4i32/v2i64/v4i64 masked gather/scatter/load/store. We end up promoting the mask argument to these during type legalization and then have to widen the promoted type to v8iX/v16iX and truncate it to get the element size back down to v8i1/v16i1 to use a 512-bit operation. Since need to fill the upper bits of the mask we have to fill with 0s at the promoted type. It would be better if we could just have the v2i1/v4i1 types as legal so they don't undergo any promotion. Then we can just widen with 0s directly in a k register. There are no real v4i1/v2i1 instructions anyway. Everything is done on a larger register anyway. This also fixes an issue that we couldn't implement a masked vextractf32x4 from zmm to xmm properly. We now have to support widening more compares to 512-bit to get a mask result out so new tablegen patterns got added. I had to hack the legalizer for widening the operand of a setcc a bit so it didn't try create a setcc returning v4i32, extract from it, then try to promote it using a sign extend to v2i1. Now we create the setcc with v4i1 if the original setcc's result type is v2i1. Then extract that and don't sign extend it at all. There's definitely room for improvement with some follow up patches. Reviewers: RKSimon, zvi, guyblank Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D41560 llvm-svn: 321967
2018-01-08 02:20:37 +08:00
; KNL-NEXT: vzeroupper
[AVX512] Fix extractelement i1 lowering. The previous implementation (not custom) doesn't enforce zeroing off upper bits. The assumption is that i1 PRODUCER (truncate and extractelement) must zero all upper bits, so i1 CONSUMER instructions ( test, zext, save, etc) can be done without additional zeroing. Make extractelement i1 lowering custom for all vector i1. Differential Revision: http://reviews.llvm.org/D23246 llvm-svn: 278328
2016-08-11 20:13:46 +08:00
; KNL-NEXT: retq
;
; SKX-LABEL: test_extractelement_v4i1:
[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
; SKX: ## %bb.0:
[AVX512] Fix extractelement i1 lowering. The previous implementation (not custom) doesn't enforce zeroing off upper bits. The assumption is that i1 PRODUCER (truncate and extractelement) must zero all upper bits, so i1 CONSUMER instructions ( test, zext, save, etc) can be done without additional zeroing. Make extractelement i1 lowering custom for all vector i1. Differential Revision: http://reviews.llvm.org/D23246 llvm-svn: 278328
2016-08-11 20:13:46 +08:00
; SKX-NEXT: vpcmpnleud %xmm1, %xmm0, %k0
[X86] Don't zero the upper bits of the k-register before extracting a single bit from a vXi1. This doesn't match the semantics of the extract_vector_elt operation. Nothing downstream knows the bits were zeroed so they still get masked or sign extended after the extrat anyway. llvm-svn: 320723
2017-12-15 02:35:25 +08:00
; SKX-NEXT: kshiftrw $3, %k0, %k0
[AVX-512] Remove explicit KMOVWrk/KMOVWKr instructions from patterns where we can just use COPY_TO_REGCLASS instead. This will result in a KMOVW or KMOVD being emitted during register allocation. And in at least some cases this might allow the register coalescer to remove the copy all together. llvm-svn: 298984
2017-03-29 14:55:28 +08:00
; SKX-NEXT: kmovd %k0, %eax
Revert r274613 because it breaks the test suite with AVX512 This reverts most of r274613 (AKA r274626) and its follow-ups (r276347, r277289), due to miscompiles in the test suite. The FastISel change was left in, because it apparently fixes an unrelated issue. (Recommit of r279782 which was broken due to a bad merge.) This fixes 4 out of the 5 test failures in PR29112. llvm-svn: 279788
2016-08-26 06:48:11 +08:00
; SKX-NEXT: andl $1, %eax
[AVX512] Fix extractelement i1 lowering. The previous implementation (not custom) doesn't enforce zeroing off upper bits. The assumption is that i1 PRODUCER (truncate and extractelement) must zero all upper bits, so i1 CONSUMER instructions ( test, zext, save, etc) can be done without additional zeroing. Make extractelement i1 lowering custom for all vector i1. Differential Revision: http://reviews.llvm.org/D23246 llvm-svn: 278328
2016-08-11 20:13:46 +08:00
; SKX-NEXT: retq
%t1 = icmp ugt <4 x i32> %a, %b
%t2 = extractelement <4 x i1> %t1, i32 3
%res = zext i1 %t2 to i8
ret i8 %res
}
define zeroext i8 @test_extractelement_v32i1(<32 x i8> %a, <32 x i8> %b) {
; KNL-LABEL: test_extractelement_v32i1:
[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
; KNL: ## %bb.0:
[X86] Use min/max for vector ult/ugt compares if avoids a sign flip. Summary: Currently we only use min/max to help with ule/uge compares because it removes an invert of the result that would otherwise be needed. But we can also use it for ult/ugt compares if it will prevent the need for a sign bit flip needed to use pcmpgt at the cost of requiring an invert after the compare. I also refactored the code so that the max/min code is self contained and does its own return instead of setting up a flag to manipulate the rest of the function's behavior. Most of the test cases look ok with this. I did notice that we added instructions when one of the operands being sign flipped is a constant vector that we were able to constant fold the flip into. I also noticed that sometimes the SSE min/max clobbers a register that is needed after the compare. This resulted in an extra move being inserted before the min/max to preserve the register. We could try to detect this and switch from min to max and change the compare operands to use the operand that gets reused in the compare. Reviewers: spatel, RKSimon Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D42935 llvm-svn: 324842
2018-02-12 01:11:40 +08:00
; KNL-NEXT: vpminub %ymm1, %ymm0, %ymm1
; KNL-NEXT: vpcmpeqb %ymm1, %ymm0, %ymm0
; KNL-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
[X86] Legalize v32i1 without BWI via splitting to v16i1 rather than the default of promoting to v32i8. Summary: For the most part its better to keep v32i1 as a mask type of a narrower width than trying to promote it to a ymm register. I had to add some overrides to the methods that get the types for the calling convention so that we still use v32i8 for argument/return purposes. There are still some regressions in here. I definitely saw some around shuffles. I think we probably should move vXi1 shuffle from lowering to a DAG combine where I think the extend and truncate we have to emit would be better combined. I think we also need a DAG combine to remove trunc from (extract_vector_elt (trunc)) Overall this removes something like 13000 CHECK lines from lit tests. Reviewers: zvi, RKSimon, delena, spatel Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D42031 llvm-svn: 323201
2018-01-23 22:25:39 +08:00
; KNL-NEXT: vpmovsxbd %xmm0, %zmm0
; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
; KNL-NEXT: kshiftrw $2, %k0, %k0
; KNL-NEXT: kmovw %k0, %eax
[AVX512] Fix extractelement i1 lowering. The previous implementation (not custom) doesn't enforce zeroing off upper bits. The assumption is that i1 PRODUCER (truncate and extractelement) must zero all upper bits, so i1 CONSUMER instructions ( test, zext, save, etc) can be done without additional zeroing. Make extractelement i1 lowering custom for all vector i1. Differential Revision: http://reviews.llvm.org/D23246 llvm-svn: 278328
2016-08-11 20:13:46 +08:00
; KNL-NEXT: andl $1, %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
; KNL-NEXT: vzeroupper
[AVX512] Fix extractelement i1 lowering. The previous implementation (not custom) doesn't enforce zeroing off upper bits. The assumption is that i1 PRODUCER (truncate and extractelement) must zero all upper bits, so i1 CONSUMER instructions ( test, zext, save, etc) can be done without additional zeroing. Make extractelement i1 lowering custom for all vector i1. Differential Revision: http://reviews.llvm.org/D23246 llvm-svn: 278328
2016-08-11 20:13:46 +08:00
; KNL-NEXT: retq
;
; SKX-LABEL: test_extractelement_v32i1:
[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
; SKX: ## %bb.0:
[AVX512] Fix extractelement i1 lowering. The previous implementation (not custom) doesn't enforce zeroing off upper bits. The assumption is that i1 PRODUCER (truncate and extractelement) must zero all upper bits, so i1 CONSUMER instructions ( test, zext, save, etc) can be done without additional zeroing. Make extractelement i1 lowering custom for all vector i1. Differential Revision: http://reviews.llvm.org/D23246 llvm-svn: 278328
2016-08-11 20:13:46 +08:00
; SKX-NEXT: vpcmpnleub %ymm1, %ymm0, %k0
[X86] Don't zero the upper bits of the k-register before extracting a single bit from a vXi1. This doesn't match the semantics of the extract_vector_elt operation. Nothing downstream knows the bits were zeroed so they still get masked or sign extended after the extrat anyway. llvm-svn: 320723
2017-12-15 02:35:25 +08:00
; SKX-NEXT: kshiftrd $2, %k0, %k0
[AVX-512] Remove explicit KMOVWrk/KMOVWKr instructions from patterns where we can just use COPY_TO_REGCLASS instead. This will result in a KMOVW or KMOVD being emitted during register allocation. And in at least some cases this might allow the register coalescer to remove the copy all together. llvm-svn: 298984
2017-03-29 14:55:28 +08:00
; SKX-NEXT: kmovd %k0, %eax
Revert r274613 because it breaks the test suite with AVX512 This reverts most of r274613 (AKA r274626) and its follow-ups (r276347, r277289), due to miscompiles in the test suite. The FastISel change was left in, because it apparently fixes an unrelated issue. (Recommit of r279782 which was broken due to a bad merge.) This fixes 4 out of the 5 test failures in PR29112. llvm-svn: 279788
2016-08-26 06:48:11 +08:00
; SKX-NEXT: andl $1, %eax
[X86] Generate VZEROUPPER for Skylake-avx512. VZEROUPPER should not be issued on Knights Landing (KNL), but on Skylake-avx512 it should be. Differential Revision: https://reviews.llvm.org/D29874 llvm-svn: 296859
2017-03-03 17:03:24 +08:00
; SKX-NEXT: vzeroupper
[AVX512] Fix extractelement i1 lowering. The previous implementation (not custom) doesn't enforce zeroing off upper bits. The assumption is that i1 PRODUCER (truncate and extractelement) must zero all upper bits, so i1 CONSUMER instructions ( test, zext, save, etc) can be done without additional zeroing. Make extractelement i1 lowering custom for all vector i1. Differential Revision: http://reviews.llvm.org/D23246 llvm-svn: 278328
2016-08-11 20:13:46 +08:00
; SKX-NEXT: retq
%t1 = icmp ugt <32 x i8> %a, %b
%t2 = extractelement <32 x i1> %t1, i32 2
%res = zext i1 %t2 to i8
ret i8 %res
}
define zeroext i8 @test_extractelement_v64i1(<64 x i8> %a, <64 x i8> %b) {
; KNL-LABEL: test_extractelement_v64i1:
[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
; KNL: ## %bb.0:
[X86] Use min/max for vector ult/ugt compares if avoids a sign flip. Summary: Currently we only use min/max to help with ule/uge compares because it removes an invert of the result that would otherwise be needed. But we can also use it for ult/ugt compares if it will prevent the need for a sign bit flip needed to use pcmpgt at the cost of requiring an invert after the compare. I also refactored the code so that the max/min code is self contained and does its own return instead of setting up a flag to manipulate the rest of the function's behavior. Most of the test cases look ok with this. I did notice that we added instructions when one of the operands being sign flipped is a constant vector that we were able to constant fold the flip into. I also noticed that sometimes the SSE min/max clobbers a register that is needed after the compare. This resulted in an extra move being inserted before the min/max to preserve the register. We could try to detect this and switch from min to max and change the compare operands to use the operand that gets reused in the compare. Reviewers: spatel, RKSimon Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D42935 llvm-svn: 324842
2018-02-12 01:11:40 +08:00
; KNL-NEXT: vpminub %ymm3, %ymm1, %ymm0
; KNL-NEXT: vpcmpeqb %ymm0, %ymm1, %ymm0
[AVX512] Fix extractelement i1 lowering. The previous implementation (not custom) doesn't enforce zeroing off upper bits. The assumption is that i1 PRODUCER (truncate and extractelement) must zero all upper bits, so i1 CONSUMER instructions ( test, zext, save, etc) can be done without additional zeroing. Make extractelement i1 lowering custom for all vector i1. Differential Revision: http://reviews.llvm.org/D23246 llvm-svn: 278328
2016-08-11 20:13:46 +08:00
; KNL-NEXT: vextracti128 $1, %ymm0, %xmm0
[DAGCombiner] narrow vector binops when extraction is cheap Narrowing vector binops came up in the demanded bits discussion in D52912. I don't think we're going to be able to do this transform in IR as a canonicalization because of the risk of creating unsupported widths for vector ops, but we already have a DAG TLI hook to allow what I was hoping for: isExtractSubvectorCheap(). This is currently enabled for x86, ARM, and AArch64 (although only x86 has existing regression test diffs). This is artificially limited to not look through bitcasts because there are so many test diffs already, but that's marked with a TODO and is a small follow-up. Differential Revision: https://reviews.llvm.org/D53784 llvm-svn: 345602
2018-10-30 22:14:34 +08:00
; KNL-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
[X86] Legalize v32i1 without BWI via splitting to v16i1 rather than the default of promoting to v32i8. Summary: For the most part its better to keep v32i1 as a mask type of a narrower width than trying to promote it to a ymm register. I had to add some overrides to the methods that get the types for the calling convention so that we still use v32i8 for argument/return purposes. There are still some regressions in here. I definitely saw some around shuffles. I think we probably should move vXi1 shuffle from lowering to a DAG combine where I think the extend and truncate we have to emit would be better combined. I think we also need a DAG combine to remove trunc from (extract_vector_elt (trunc)) Overall this removes something like 13000 CHECK lines from lit tests. Reviewers: zvi, RKSimon, delena, spatel Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D42031 llvm-svn: 323201
2018-01-23 22:25:39 +08:00
; KNL-NEXT: vpmovsxbd %xmm0, %zmm0
; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
; KNL-NEXT: kshiftrw $15, %k0, %k0
; KNL-NEXT: kmovw %k0, %eax
[X86][AVX512] Make i1 illegal in the CodeGen This patch defines the i1 type as illegal in the X86 backend for AVX512. For DAG operations on <N x i1> types (build vector, extract vector element, ...) i8 is used, and should be truncated/extended. This should produce better scalar code for i1 types since GPRs will be used instead of mask registers. Differential Revision: https://reviews.llvm.org/D32273 llvm-svn: 303421
2017-05-19 20:35:15 +08:00
; KNL-NEXT: andb $1, %al
; KNL-NEXT: movb $4, %cl
; KNL-NEXT: subb %al, %cl
; KNL-NEXT: movzbl %cl, %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
; KNL-NEXT: vzeroupper
[AVX512] Fix extractelement i1 lowering. The previous implementation (not custom) doesn't enforce zeroing off upper bits. The assumption is that i1 PRODUCER (truncate and extractelement) must zero all upper bits, so i1 CONSUMER instructions ( test, zext, save, etc) can be done without additional zeroing. Make extractelement i1 lowering custom for all vector i1. Differential Revision: http://reviews.llvm.org/D23246 llvm-svn: 278328
2016-08-11 20:13:46 +08:00
; KNL-NEXT: retq
;
; SKX-LABEL: test_extractelement_v64i1:
[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
; SKX: ## %bb.0:
[AVX512] Fix extractelement i1 lowering. The previous implementation (not custom) doesn't enforce zeroing off upper bits. The assumption is that i1 PRODUCER (truncate and extractelement) must zero all upper bits, so i1 CONSUMER instructions ( test, zext, save, etc) can be done without additional zeroing. Make extractelement i1 lowering custom for all vector i1. Differential Revision: http://reviews.llvm.org/D23246 llvm-svn: 278328
2016-08-11 20:13:46 +08:00
; SKX-NEXT: vpcmpnleub %zmm1, %zmm0, %k0
; SKX-NEXT: kshiftrq $63, %k0, %k0
[AVX-512] Remove explicit KMOVWrk/KMOVWKr instructions from patterns where we can just use COPY_TO_REGCLASS instead. This will result in a KMOVW or KMOVD being emitted during register allocation. And in at least some cases this might allow the register coalescer to remove the copy all together. llvm-svn: 298984
2017-03-29 14:55:28 +08:00
; SKX-NEXT: kmovd %k0, %eax
[X86][AVX512] Make i1 illegal in the CodeGen This patch defines the i1 type as illegal in the X86 backend for AVX512. For DAG operations on <N x i1> types (build vector, extract vector element, ...) i8 is used, and should be truncated/extended. This should produce better scalar code for i1 types since GPRs will be used instead of mask registers. Differential Revision: https://reviews.llvm.org/D32273 llvm-svn: 303421
2017-05-19 20:35:15 +08:00
; SKX-NEXT: andb $1, %al
; SKX-NEXT: movb $4, %cl
; SKX-NEXT: subb %al, %cl
; SKX-NEXT: movzbl %cl, %eax
[X86] Generate VZEROUPPER for Skylake-avx512. VZEROUPPER should not be issued on Knights Landing (KNL), but on Skylake-avx512 it should be. Differential Revision: https://reviews.llvm.org/D29874 llvm-svn: 296859
2017-03-03 17:03:24 +08:00
; SKX-NEXT: vzeroupper
[AVX512] Fix extractelement i1 lowering. The previous implementation (not custom) doesn't enforce zeroing off upper bits. The assumption is that i1 PRODUCER (truncate and extractelement) must zero all upper bits, so i1 CONSUMER instructions ( test, zext, save, etc) can be done without additional zeroing. Make extractelement i1 lowering custom for all vector i1. Differential Revision: http://reviews.llvm.org/D23246 llvm-svn: 278328
2016-08-11 20:13:46 +08:00
; SKX-NEXT: retq
%t1 = icmp ugt <64 x i8> %a, %b
%t2 = extractelement <64 x i1> %t1, i32 63
%res = select i1 %t2, i8 3, i8 4
ret i8 %res
}
Add new tests for EXTRACT_VECTOR_ELT (vector of packed i8/16/i32/i64/ps/pd data) llvm-svn: 294565
2017-02-09 15:39:19 +08:00
[x86] add alternate IR tests for select of constants; NFC llvm-svn: 296496
2017-03-01 02:02:38 +08:00
define zeroext i8 @extractelement_v64i1_alt(<64 x i8> %a, <64 x i8> %b) {
; KNL-LABEL: extractelement_v64i1_alt:
[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
; KNL: ## %bb.0:
[X86] Use min/max for vector ult/ugt compares if avoids a sign flip. Summary: Currently we only use min/max to help with ule/uge compares because it removes an invert of the result that would otherwise be needed. But we can also use it for ult/ugt compares if it will prevent the need for a sign bit flip needed to use pcmpgt at the cost of requiring an invert after the compare. I also refactored the code so that the max/min code is self contained and does its own return instead of setting up a flag to manipulate the rest of the function's behavior. Most of the test cases look ok with this. I did notice that we added instructions when one of the operands being sign flipped is a constant vector that we were able to constant fold the flip into. I also noticed that sometimes the SSE min/max clobbers a register that is needed after the compare. This resulted in an extra move being inserted before the min/max to preserve the register. We could try to detect this and switch from min to max and change the compare operands to use the operand that gets reused in the compare. Reviewers: spatel, RKSimon Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D42935 llvm-svn: 324842
2018-02-12 01:11:40 +08:00
; KNL-NEXT: vpminub %ymm3, %ymm1, %ymm0
; KNL-NEXT: vpcmpeqb %ymm0, %ymm1, %ymm0
[x86] add alternate IR tests for select of constants; NFC llvm-svn: 296496
2017-03-01 02:02:38 +08:00
; KNL-NEXT: vextracti128 $1, %ymm0, %xmm0
[DAGCombiner] narrow vector binops when extraction is cheap Narrowing vector binops came up in the demanded bits discussion in D52912. I don't think we're going to be able to do this transform in IR as a canonicalization because of the risk of creating unsupported widths for vector ops, but we already have a DAG TLI hook to allow what I was hoping for: isExtractSubvectorCheap(). This is currently enabled for x86, ARM, and AArch64 (although only x86 has existing regression test diffs). This is artificially limited to not look through bitcasts because there are so many test diffs already, but that's marked with a TODO and is a small follow-up. Differential Revision: https://reviews.llvm.org/D53784 llvm-svn: 345602
2018-10-30 22:14:34 +08:00
; KNL-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
[X86] Legalize v32i1 without BWI via splitting to v16i1 rather than the default of promoting to v32i8. Summary: For the most part its better to keep v32i1 as a mask type of a narrower width than trying to promote it to a ymm register. I had to add some overrides to the methods that get the types for the calling convention so that we still use v32i8 for argument/return purposes. There are still some regressions in here. I definitely saw some around shuffles. I think we probably should move vXi1 shuffle from lowering to a DAG combine where I think the extend and truncate we have to emit would be better combined. I think we also need a DAG combine to remove trunc from (extract_vector_elt (trunc)) Overall this removes something like 13000 CHECK lines from lit tests. Reviewers: zvi, RKSimon, delena, spatel Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D42031 llvm-svn: 323201
2018-01-23 22:25:39 +08:00
; KNL-NEXT: vpmovsxbd %xmm0, %zmm0
; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
; KNL-NEXT: kshiftrw $15, %k0, %k0
; KNL-NEXT: kmovw %k0, %eax
[X86][AVX512] Make i1 illegal in the CodeGen This patch defines the i1 type as illegal in the X86 backend for AVX512. For DAG operations on <N x i1> types (build vector, extract vector element, ...) i8 is used, and should be truncated/extended. This should produce better scalar code for i1 types since GPRs will be used instead of mask registers. Differential Revision: https://reviews.llvm.org/D32273 llvm-svn: 303421
2017-05-19 20:35:15 +08:00
; KNL-NEXT: andb $1, %al
; KNL-NEXT: movb $4, %cl
; KNL-NEXT: subb %al, %cl
; KNL-NEXT: movzbl %cl, %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
; KNL-NEXT: vzeroupper
[x86] add alternate IR tests for select of constants; NFC llvm-svn: 296496
2017-03-01 02:02:38 +08:00
; KNL-NEXT: retq
;
; SKX-LABEL: extractelement_v64i1_alt:
[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
; SKX: ## %bb.0:
[x86] add alternate IR tests for select of constants; NFC llvm-svn: 296496
2017-03-01 02:02:38 +08:00
; SKX-NEXT: vpcmpnleub %zmm1, %zmm0, %k0
; SKX-NEXT: kshiftrq $63, %k0, %k0
[AVX-512] Remove explicit KMOVWrk/KMOVWKr instructions from patterns where we can just use COPY_TO_REGCLASS instead. This will result in a KMOVW or KMOVD being emitted during register allocation. And in at least some cases this might allow the register coalescer to remove the copy all together. llvm-svn: 298984
2017-03-29 14:55:28 +08:00
; SKX-NEXT: kmovd %k0, %eax
[X86][AVX512] Make i1 illegal in the CodeGen This patch defines the i1 type as illegal in the X86 backend for AVX512. For DAG operations on <N x i1> types (build vector, extract vector element, ...) i8 is used, and should be truncated/extended. This should produce better scalar code for i1 types since GPRs will be used instead of mask registers. Differential Revision: https://reviews.llvm.org/D32273 llvm-svn: 303421
2017-05-19 20:35:15 +08:00
; SKX-NEXT: andb $1, %al
; SKX-NEXT: movb $4, %cl
; SKX-NEXT: subb %al, %cl
; SKX-NEXT: movzbl %cl, %eax
[X86] Generate VZEROUPPER for Skylake-avx512. VZEROUPPER should not be issued on Knights Landing (KNL), but on Skylake-avx512 it should be. Differential Revision: https://reviews.llvm.org/D29874 llvm-svn: 296859
2017-03-03 17:03:24 +08:00
; SKX-NEXT: vzeroupper
[x86] add alternate IR tests for select of constants; NFC llvm-svn: 296496
2017-03-01 02:02:38 +08:00
; SKX-NEXT: retq
%t1 = icmp ugt <64 x i8> %a, %b
%t2 = extractelement <64 x i1> %t1, i32 63
%sext = sext i1 %t2 to i8
%res = add i8 %sext, 4
ret i8 %res
}
Add new tests for EXTRACT_VECTOR_ELT (vector of packed i8/16/i32/i64/ps/pd data) llvm-svn: 294565
2017-02-09 15:39:19 +08:00
define i64 @test_extractelement_variable_v2i64(<2 x i64> %t1, i32 %index) {
[AVX512] Add a common prefix to avx512-insert-extract.ll so we can reduce the number of check lines on some test cases. This was pointed out during the review for D313804. llvm-svn: 309629
2017-08-01 05:20:06 +08:00
; CHECK-LABEL: test_extractelement_variable_v2i64:
[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:
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; CHECK-NEXT: ## kill: def $edi killed $edi def $rdi
[AVX512] Add a common prefix to avx512-insert-extract.ll so we can reduce the number of check lines on some test cases. This was pointed out during the review for D313804. llvm-svn: 309629
2017-08-01 05:20:06 +08:00
; CHECK-NEXT: vmovaps %xmm0, -{{[0-9]+}}(%rsp)
; CHECK-NEXT: andl $1, %edi
; CHECK-NEXT: movq -24(%rsp,%rdi,8), %rax
; CHECK-NEXT: retq
Add new tests for EXTRACT_VECTOR_ELT (vector of packed i8/16/i32/i64/ps/pd data) llvm-svn: 294565
2017-02-09 15:39:19 +08:00
%t2 = extractelement <2 x i64> %t1, i32 %index
ret i64 %t2
}
define i64 @test_extractelement_variable_v4i64(<4 x i64> %t1, i32 %index) {
[X86][AVX512] Regenerate element insertion/extraction tests llvm-svn: 315322
2017-10-10 23:58:54 +08:00
; CHECK-LABEL: test_extractelement_variable_v4i64:
[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] Regenerate element insertion/extraction tests llvm-svn: 315322
2017-10-10 23:58:54 +08:00
; CHECK-NEXT: pushq %rbp
; CHECK-NEXT: .cfi_def_cfa_offset 16
; CHECK-NEXT: .cfi_offset %rbp, -16
; CHECK-NEXT: movq %rsp, %rbp
; CHECK-NEXT: .cfi_def_cfa_register %rbp
; CHECK-NEXT: andq $-32, %rsp
; CHECK-NEXT: subq $64, %rsp
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; CHECK-NEXT: ## kill: def $edi killed $edi def $rdi
[X86][AVX512] Regenerate element insertion/extraction tests llvm-svn: 315322
2017-10-10 23:58:54 +08:00
; CHECK-NEXT: vmovaps %ymm0, (%rsp)
; CHECK-NEXT: andl $3, %edi
; CHECK-NEXT: movq (%rsp,%rdi,8), %rax
; CHECK-NEXT: movq %rbp, %rsp
; CHECK-NEXT: popq %rbp
; CHECK-NEXT: vzeroupper
; CHECK-NEXT: retq
Add new tests for EXTRACT_VECTOR_ELT (vector of packed i8/16/i32/i64/ps/pd data) llvm-svn: 294565
2017-02-09 15:39:19 +08:00
%t2 = extractelement <4 x i64> %t1, i32 %index
ret i64 %t2
}
define i64 @test_extractelement_variable_v8i64(<8 x i64> %t1, i32 %index) {
[X86][AVX512] Regenerate element insertion/extraction tests llvm-svn: 315322
2017-10-10 23:58:54 +08:00
; CHECK-LABEL: test_extractelement_variable_v8i64:
[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] Regenerate element insertion/extraction tests llvm-svn: 315322
2017-10-10 23:58:54 +08:00
; CHECK-NEXT: pushq %rbp
; CHECK-NEXT: .cfi_def_cfa_offset 16
; CHECK-NEXT: .cfi_offset %rbp, -16
; CHECK-NEXT: movq %rsp, %rbp
; CHECK-NEXT: .cfi_def_cfa_register %rbp
; CHECK-NEXT: andq $-64, %rsp
; CHECK-NEXT: subq $128, %rsp
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; CHECK-NEXT: ## kill: def $edi killed $edi def $rdi
[X86][AVX512] Regenerate element insertion/extraction tests llvm-svn: 315322
2017-10-10 23:58:54 +08:00
; CHECK-NEXT: vmovaps %zmm0, (%rsp)
; CHECK-NEXT: andl $7, %edi
; CHECK-NEXT: movq (%rsp,%rdi,8), %rax
; CHECK-NEXT: movq %rbp, %rsp
; CHECK-NEXT: popq %rbp
; CHECK-NEXT: vzeroupper
; CHECK-NEXT: retq
Add new tests for EXTRACT_VECTOR_ELT (vector of packed i8/16/i32/i64/ps/pd data) llvm-svn: 294565
2017-02-09 15:39:19 +08:00
%t2 = extractelement <8 x i64> %t1, i32 %index
ret i64 %t2
}
define double @test_extractelement_variable_v2f64(<2 x double> %t1, i32 %index) {
[AVX512] Add a common prefix to avx512-insert-extract.ll so we can reduce the number of check lines on some test cases. This was pointed out during the review for D313804. llvm-svn: 309629
2017-08-01 05:20:06 +08:00
; CHECK-LABEL: test_extractelement_variable_v2f64:
[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:
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; CHECK-NEXT: ## kill: def $edi killed $edi def $rdi
[AVX512] Add a common prefix to avx512-insert-extract.ll so we can reduce the number of check lines on some test cases. This was pointed out during the review for D313804. llvm-svn: 309629
2017-08-01 05:20:06 +08:00
; CHECK-NEXT: vmovaps %xmm0, -{{[0-9]+}}(%rsp)
; CHECK-NEXT: andl $1, %edi
; CHECK-NEXT: vmovsd {{.*#+}} xmm0 = mem[0],zero
; CHECK-NEXT: retq
Add new tests for EXTRACT_VECTOR_ELT (vector of packed i8/16/i32/i64/ps/pd data) llvm-svn: 294565
2017-02-09 15:39:19 +08:00
%t2 = extractelement <2 x double> %t1, i32 %index
ret double %t2
}
define double @test_extractelement_variable_v4f64(<4 x double> %t1, i32 %index) {
[X86][AVX512] Regenerate element insertion/extraction tests llvm-svn: 315322
2017-10-10 23:58:54 +08:00
; CHECK-LABEL: test_extractelement_variable_v4f64:
[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] Regenerate element insertion/extraction tests llvm-svn: 315322
2017-10-10 23:58:54 +08:00
; CHECK-NEXT: pushq %rbp
; CHECK-NEXT: .cfi_def_cfa_offset 16
; CHECK-NEXT: .cfi_offset %rbp, -16
; CHECK-NEXT: movq %rsp, %rbp
; CHECK-NEXT: .cfi_def_cfa_register %rbp
; CHECK-NEXT: andq $-32, %rsp
; CHECK-NEXT: subq $64, %rsp
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; CHECK-NEXT: ## kill: def $edi killed $edi def $rdi
[X86][AVX512] Regenerate element insertion/extraction tests llvm-svn: 315322
2017-10-10 23:58:54 +08:00
; CHECK-NEXT: vmovaps %ymm0, (%rsp)
; CHECK-NEXT: andl $3, %edi
; CHECK-NEXT: vmovsd {{.*#+}} xmm0 = mem[0],zero
; CHECK-NEXT: movq %rbp, %rsp
; CHECK-NEXT: popq %rbp
; CHECK-NEXT: vzeroupper
; CHECK-NEXT: retq
Add new tests for EXTRACT_VECTOR_ELT (vector of packed i8/16/i32/i64/ps/pd data) llvm-svn: 294565
2017-02-09 15:39:19 +08:00
%t2 = extractelement <4 x double> %t1, i32 %index
ret double %t2
}
define double @test_extractelement_variable_v8f64(<8 x double> %t1, i32 %index) {
[X86][AVX512] Regenerate element insertion/extraction tests llvm-svn: 315322
2017-10-10 23:58:54 +08:00
; CHECK-LABEL: test_extractelement_variable_v8f64:
[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] Regenerate element insertion/extraction tests llvm-svn: 315322
2017-10-10 23:58:54 +08:00
; CHECK-NEXT: pushq %rbp
; CHECK-NEXT: .cfi_def_cfa_offset 16
; CHECK-NEXT: .cfi_offset %rbp, -16
; CHECK-NEXT: movq %rsp, %rbp
; CHECK-NEXT: .cfi_def_cfa_register %rbp
; CHECK-NEXT: andq $-64, %rsp
; CHECK-NEXT: subq $128, %rsp
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; CHECK-NEXT: ## kill: def $edi killed $edi def $rdi
[X86][AVX512] Regenerate element insertion/extraction tests llvm-svn: 315322
2017-10-10 23:58:54 +08:00
; CHECK-NEXT: vmovaps %zmm0, (%rsp)
; CHECK-NEXT: andl $7, %edi
; CHECK-NEXT: vmovsd {{.*#+}} xmm0 = mem[0],zero
; CHECK-NEXT: movq %rbp, %rsp
; CHECK-NEXT: popq %rbp
; CHECK-NEXT: vzeroupper
; CHECK-NEXT: retq
Add new tests for EXTRACT_VECTOR_ELT (vector of packed i8/16/i32/i64/ps/pd data) llvm-svn: 294565
2017-02-09 15:39:19 +08:00
%t2 = extractelement <8 x double> %t1, i32 %index
ret double %t2
}
define i32 @test_extractelement_variable_v4i32(<4 x i32> %t1, i32 %index) {
[AVX512] Add a common prefix to avx512-insert-extract.ll so we can reduce the number of check lines on some test cases. This was pointed out during the review for D313804. llvm-svn: 309629
2017-08-01 05:20:06 +08:00
; CHECK-LABEL: test_extractelement_variable_v4i32:
[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:
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; CHECK-NEXT: ## kill: def $edi killed $edi def $rdi
[AVX512] Add a common prefix to avx512-insert-extract.ll so we can reduce the number of check lines on some test cases. This was pointed out during the review for D313804. llvm-svn: 309629
2017-08-01 05:20:06 +08:00
; CHECK-NEXT: vmovaps %xmm0, -{{[0-9]+}}(%rsp)
; CHECK-NEXT: andl $3, %edi
; CHECK-NEXT: movl -24(%rsp,%rdi,4), %eax
; CHECK-NEXT: retq
Add new tests for EXTRACT_VECTOR_ELT (vector of packed i8/16/i32/i64/ps/pd data) llvm-svn: 294565
2017-02-09 15:39:19 +08:00
%t2 = extractelement <4 x i32> %t1, i32 %index
ret i32 %t2
}
define i32 @test_extractelement_variable_v8i32(<8 x i32> %t1, i32 %index) {
[X86][AVX512] Regenerate element insertion/extraction tests llvm-svn: 315322
2017-10-10 23:58:54 +08:00
; CHECK-LABEL: test_extractelement_variable_v8i32:
[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] Regenerate element insertion/extraction tests llvm-svn: 315322
2017-10-10 23:58:54 +08:00
; CHECK-NEXT: pushq %rbp
; CHECK-NEXT: .cfi_def_cfa_offset 16
; CHECK-NEXT: .cfi_offset %rbp, -16
; CHECK-NEXT: movq %rsp, %rbp
; CHECK-NEXT: .cfi_def_cfa_register %rbp
; CHECK-NEXT: andq $-32, %rsp
; CHECK-NEXT: subq $64, %rsp
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; CHECK-NEXT: ## kill: def $edi killed $edi def $rdi
[X86][AVX512] Regenerate element insertion/extraction tests llvm-svn: 315322
2017-10-10 23:58:54 +08:00
; CHECK-NEXT: vmovaps %ymm0, (%rsp)
; CHECK-NEXT: andl $7, %edi
; CHECK-NEXT: movl (%rsp,%rdi,4), %eax
; CHECK-NEXT: movq %rbp, %rsp
; CHECK-NEXT: popq %rbp
; CHECK-NEXT: vzeroupper
; CHECK-NEXT: retq
Add new tests for EXTRACT_VECTOR_ELT (vector of packed i8/16/i32/i64/ps/pd data) llvm-svn: 294565
2017-02-09 15:39:19 +08:00
%t2 = extractelement <8 x i32> %t1, i32 %index
ret i32 %t2
}
define i32 @test_extractelement_variable_v16i32(<16 x i32> %t1, i32 %index) {
[X86][AVX512] Regenerate element insertion/extraction tests llvm-svn: 315322
2017-10-10 23:58:54 +08:00
; CHECK-LABEL: test_extractelement_variable_v16i32:
[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] Regenerate element insertion/extraction tests llvm-svn: 315322
2017-10-10 23:58:54 +08:00
; CHECK-NEXT: pushq %rbp
; CHECK-NEXT: .cfi_def_cfa_offset 16
; CHECK-NEXT: .cfi_offset %rbp, -16
; CHECK-NEXT: movq %rsp, %rbp
; CHECK-NEXT: .cfi_def_cfa_register %rbp
; CHECK-NEXT: andq $-64, %rsp
; CHECK-NEXT: subq $128, %rsp
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; CHECK-NEXT: ## kill: def $edi killed $edi def $rdi
[X86][AVX512] Regenerate element insertion/extraction tests llvm-svn: 315322
2017-10-10 23:58:54 +08:00
; CHECK-NEXT: vmovaps %zmm0, (%rsp)
; CHECK-NEXT: andl $15, %edi
; CHECK-NEXT: movl (%rsp,%rdi,4), %eax
; CHECK-NEXT: movq %rbp, %rsp
; CHECK-NEXT: popq %rbp
; CHECK-NEXT: vzeroupper
; CHECK-NEXT: retq
Add new tests for EXTRACT_VECTOR_ELT (vector of packed i8/16/i32/i64/ps/pd data) llvm-svn: 294565
2017-02-09 15:39:19 +08:00
%t2 = extractelement <16 x i32> %t1, i32 %index
ret i32 %t2
}
define float @test_extractelement_variable_v4f32(<4 x float> %t1, i32 %index) {
[AVX512] Add a common prefix to avx512-insert-extract.ll so we can reduce the number of check lines on some test cases. This was pointed out during the review for D313804. llvm-svn: 309629
2017-08-01 05:20:06 +08:00
; CHECK-LABEL: test_extractelement_variable_v4f32:
[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:
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; CHECK-NEXT: ## kill: def $edi killed $edi def $rdi
[AVX512] Add a common prefix to avx512-insert-extract.ll so we can reduce the number of check lines on some test cases. This was pointed out during the review for D313804. llvm-svn: 309629
2017-08-01 05:20:06 +08:00
; CHECK-NEXT: vmovaps %xmm0, -{{[0-9]+}}(%rsp)
; CHECK-NEXT: andl $3, %edi
; CHECK-NEXT: vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero
; CHECK-NEXT: retq
Add new tests for EXTRACT_VECTOR_ELT (vector of packed i8/16/i32/i64/ps/pd data) llvm-svn: 294565
2017-02-09 15:39:19 +08:00
%t2 = extractelement <4 x float> %t1, i32 %index
ret float %t2
}
define float @test_extractelement_variable_v8f32(<8 x float> %t1, i32 %index) {
[X86][AVX512] Regenerate element insertion/extraction tests llvm-svn: 315322
2017-10-10 23:58:54 +08:00
; CHECK-LABEL: test_extractelement_variable_v8f32:
[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] Regenerate element insertion/extraction tests llvm-svn: 315322
2017-10-10 23:58:54 +08:00
; CHECK-NEXT: pushq %rbp
; CHECK-NEXT: .cfi_def_cfa_offset 16
; CHECK-NEXT: .cfi_offset %rbp, -16
; CHECK-NEXT: movq %rsp, %rbp
; CHECK-NEXT: .cfi_def_cfa_register %rbp
; CHECK-NEXT: andq $-32, %rsp
; CHECK-NEXT: subq $64, %rsp
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; CHECK-NEXT: ## kill: def $edi killed $edi def $rdi
[X86][AVX512] Regenerate element insertion/extraction tests llvm-svn: 315322
2017-10-10 23:58:54 +08:00
; CHECK-NEXT: vmovaps %ymm0, (%rsp)
; CHECK-NEXT: andl $7, %edi
; CHECK-NEXT: vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero
; CHECK-NEXT: movq %rbp, %rsp
; CHECK-NEXT: popq %rbp
; CHECK-NEXT: vzeroupper
; CHECK-NEXT: retq
Add new tests for EXTRACT_VECTOR_ELT (vector of packed i8/16/i32/i64/ps/pd data) llvm-svn: 294565
2017-02-09 15:39:19 +08:00
%t2 = extractelement <8 x float> %t1, i32 %index
ret float %t2
}
define float @test_extractelement_variable_v16f32(<16 x float> %t1, i32 %index) {
[X86][AVX512] Regenerate element insertion/extraction tests llvm-svn: 315322
2017-10-10 23:58:54 +08:00
; CHECK-LABEL: test_extractelement_variable_v16f32:
[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] Regenerate element insertion/extraction tests llvm-svn: 315322
2017-10-10 23:58:54 +08:00
; CHECK-NEXT: pushq %rbp
; CHECK-NEXT: .cfi_def_cfa_offset 16
; CHECK-NEXT: .cfi_offset %rbp, -16
; CHECK-NEXT: movq %rsp, %rbp
; CHECK-NEXT: .cfi_def_cfa_register %rbp
; CHECK-NEXT: andq $-64, %rsp
; CHECK-NEXT: subq $128, %rsp
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; CHECK-NEXT: ## kill: def $edi killed $edi def $rdi
[X86][AVX512] Regenerate element insertion/extraction tests llvm-svn: 315322
2017-10-10 23:58:54 +08:00
; CHECK-NEXT: vmovaps %zmm0, (%rsp)
; CHECK-NEXT: andl $15, %edi
; CHECK-NEXT: vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero
; CHECK-NEXT: movq %rbp, %rsp
; CHECK-NEXT: popq %rbp
; CHECK-NEXT: vzeroupper
; CHECK-NEXT: retq
Add new tests for EXTRACT_VECTOR_ELT (vector of packed i8/16/i32/i64/ps/pd data) llvm-svn: 294565
2017-02-09 15:39:19 +08:00
%t2 = extractelement <16 x float> %t1, i32 %index
ret float %t2
}
define i16 @test_extractelement_variable_v8i16(<8 x i16> %t1, i32 %index) {
[AVX-512] Don't use unmasked VMOVDQU8/16 for 8-bit or 16-bit element stores even when BWI instructions are supported. Always use VMOVDQA32/VMOVDQU32. We were already using the 32 bit element opcode if BWI isn't enabled, but there's no reason to change opcode if we have BWI. We will still use the 8/16 opcodes for masked stores though. This allows us to use the aligned opcode when we can which makes our test output more consistent between different modes. It also reduces the number of isel patterns we need. This is a slight inconsistency with loads which default to 64 bit element opcodes. I'll probably rectify that in a future patch. Differential Revision: https://reviews.llvm.org/D35978 llvm-svn: 309693
2017-08-01 23:31:24 +08:00
; CHECK-LABEL: test_extractelement_variable_v8i16:
[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:
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; CHECK-NEXT: ## kill: def $edi killed $edi def $rdi
[AVX-512] Don't use unmasked VMOVDQU8/16 for 8-bit or 16-bit element stores even when BWI instructions are supported. Always use VMOVDQA32/VMOVDQU32. We were already using the 32 bit element opcode if BWI isn't enabled, but there's no reason to change opcode if we have BWI. We will still use the 8/16 opcodes for masked stores though. This allows us to use the aligned opcode when we can which makes our test output more consistent between different modes. It also reduces the number of isel patterns we need. This is a slight inconsistency with loads which default to 64 bit element opcodes. I'll probably rectify that in a future patch. Differential Revision: https://reviews.llvm.org/D35978 llvm-svn: 309693
2017-08-01 23:31:24 +08:00
; CHECK-NEXT: vmovaps %xmm0, -{{[0-9]+}}(%rsp)
; CHECK-NEXT: andl $7, %edi
; CHECK-NEXT: movzwl -24(%rsp,%rdi,2), %eax
; CHECK-NEXT: retq
Add new tests for EXTRACT_VECTOR_ELT (vector of packed i8/16/i32/i64/ps/pd data) llvm-svn: 294565
2017-02-09 15:39:19 +08:00
%t2 = extractelement <8 x i16> %t1, i32 %index
ret i16 %t2
}
define i16 @test_extractelement_variable_v16i16(<16 x i16> %t1, i32 %index) {
[X86][AVX512] Regenerate element insertion/extraction tests llvm-svn: 315322
2017-10-10 23:58:54 +08:00
; CHECK-LABEL: test_extractelement_variable_v16i16:
[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] Regenerate element insertion/extraction tests llvm-svn: 315322
2017-10-10 23:58:54 +08:00
; CHECK-NEXT: pushq %rbp
; CHECK-NEXT: .cfi_def_cfa_offset 16
; CHECK-NEXT: .cfi_offset %rbp, -16
; CHECK-NEXT: movq %rsp, %rbp
; CHECK-NEXT: .cfi_def_cfa_register %rbp
; CHECK-NEXT: andq $-32, %rsp
; CHECK-NEXT: subq $64, %rsp
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; CHECK-NEXT: ## kill: def $edi killed $edi def $rdi
[X86][AVX512] Regenerate element insertion/extraction tests llvm-svn: 315322
2017-10-10 23:58:54 +08:00
; CHECK-NEXT: vmovaps %ymm0, (%rsp)
; CHECK-NEXT: andl $15, %edi
; CHECK-NEXT: movzwl (%rsp,%rdi,2), %eax
; CHECK-NEXT: movq %rbp, %rsp
; CHECK-NEXT: popq %rbp
; CHECK-NEXT: vzeroupper
; CHECK-NEXT: retq
Add new tests for EXTRACT_VECTOR_ELT (vector of packed i8/16/i32/i64/ps/pd data) llvm-svn: 294565
2017-02-09 15:39:19 +08:00
%t2 = extractelement <16 x i16> %t1, i32 %index
ret i16 %t2
}
[X86] Fix EXTRACT_VECTOR_ELT with variable index from v32i16 and v64i8 vector. Its more profitable to go through memory (1 cycles throughput) than using VMOVD + VPERMV/PSHUFB sequence ( 2/3 cycles throughput) to implement EXTRACT_VECTOR_ELT with variable index. IACA tool was used to get performace estimation (https://software.intel.com/en-us/articles/intel-architecture-code-analyzer) For example for var_shuffle_v16i8_v16i8_xxxxxxxxxxxxxxxx_i8 test from vector-shuffle-variable-128.ll I get 26 cycles vs 79 cycles. Removing the VINSERT node, we don't need it any more. Differential Revision: https://reviews.llvm.org/D29690 llvm-svn: 295660
2017-02-20 22:16:29 +08:00
define i16 @test_extractelement_variable_v32i16(<32 x i16> %t1, i32 %index) {
; KNL-LABEL: test_extractelement_variable_v32i16:
[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
; KNL: ## %bb.0:
[X86] Fix EXTRACT_VECTOR_ELT with variable index from v32i16 and v64i8 vector. Its more profitable to go through memory (1 cycles throughput) than using VMOVD + VPERMV/PSHUFB sequence ( 2/3 cycles throughput) to implement EXTRACT_VECTOR_ELT with variable index. IACA tool was used to get performace estimation (https://software.intel.com/en-us/articles/intel-architecture-code-analyzer) For example for var_shuffle_v16i8_v16i8_xxxxxxxxxxxxxxxx_i8 test from vector-shuffle-variable-128.ll I get 26 cycles vs 79 cycles. Removing the VINSERT node, we don't need it any more. Differential Revision: https://reviews.llvm.org/D29690 llvm-svn: 295660
2017-02-20 22:16:29 +08:00
; KNL-NEXT: pushq %rbp
; KNL-NEXT: .cfi_def_cfa_offset 16
; KNL-NEXT: .cfi_offset %rbp, -16
; KNL-NEXT: movq %rsp, %rbp
; KNL-NEXT: .cfi_def_cfa_register %rbp
; KNL-NEXT: andq $-64, %rsp
; KNL-NEXT: subq $128, %rsp
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; KNL-NEXT: ## kill: def $edi killed $edi def $rdi
[X86] Fix EXTRACT_VECTOR_ELT with variable index from v32i16 and v64i8 vector. Its more profitable to go through memory (1 cycles throughput) than using VMOVD + VPERMV/PSHUFB sequence ( 2/3 cycles throughput) to implement EXTRACT_VECTOR_ELT with variable index. IACA tool was used to get performace estimation (https://software.intel.com/en-us/articles/intel-architecture-code-analyzer) For example for var_shuffle_v16i8_v16i8_xxxxxxxxxxxxxxxx_i8 test from vector-shuffle-variable-128.ll I get 26 cycles vs 79 cycles. Removing the VINSERT node, we don't need it any more. Differential Revision: https://reviews.llvm.org/D29690 llvm-svn: 295660
2017-02-20 22:16:29 +08:00
; KNL-NEXT: vmovaps %ymm1, {{[0-9]+}}(%rsp)
; KNL-NEXT: vmovaps %ymm0, (%rsp)
; KNL-NEXT: andl $31, %edi
; KNL-NEXT: movzwl (%rsp,%rdi,2), %eax
; KNL-NEXT: movq %rbp, %rsp
; KNL-NEXT: popq %rbp
[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
; KNL-NEXT: vzeroupper
[X86] Fix EXTRACT_VECTOR_ELT with variable index from v32i16 and v64i8 vector. Its more profitable to go through memory (1 cycles throughput) than using VMOVD + VPERMV/PSHUFB sequence ( 2/3 cycles throughput) to implement EXTRACT_VECTOR_ELT with variable index. IACA tool was used to get performace estimation (https://software.intel.com/en-us/articles/intel-architecture-code-analyzer) For example for var_shuffle_v16i8_v16i8_xxxxxxxxxxxxxxxx_i8 test from vector-shuffle-variable-128.ll I get 26 cycles vs 79 cycles. Removing the VINSERT node, we don't need it any more. Differential Revision: https://reviews.llvm.org/D29690 llvm-svn: 295660
2017-02-20 22:16:29 +08:00
; KNL-NEXT: retq
;
; SKX-LABEL: test_extractelement_variable_v32i16:
[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
; SKX: ## %bb.0:
[X86] Fix EXTRACT_VECTOR_ELT with variable index from v32i16 and v64i8 vector. Its more profitable to go through memory (1 cycles throughput) than using VMOVD + VPERMV/PSHUFB sequence ( 2/3 cycles throughput) to implement EXTRACT_VECTOR_ELT with variable index. IACA tool was used to get performace estimation (https://software.intel.com/en-us/articles/intel-architecture-code-analyzer) For example for var_shuffle_v16i8_v16i8_xxxxxxxxxxxxxxxx_i8 test from vector-shuffle-variable-128.ll I get 26 cycles vs 79 cycles. Removing the VINSERT node, we don't need it any more. Differential Revision: https://reviews.llvm.org/D29690 llvm-svn: 295660
2017-02-20 22:16:29 +08:00
; SKX-NEXT: pushq %rbp
; SKX-NEXT: .cfi_def_cfa_offset 16
; SKX-NEXT: .cfi_offset %rbp, -16
; SKX-NEXT: movq %rsp, %rbp
; SKX-NEXT: .cfi_def_cfa_register %rbp
; SKX-NEXT: andq $-64, %rsp
; SKX-NEXT: subq $128, %rsp
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; SKX-NEXT: ## kill: def $edi killed $edi def $rdi
[AVX-512] Don't use unmasked VMOVDQU8/16 for 8-bit or 16-bit element stores even when BWI instructions are supported. Always use VMOVDQA32/VMOVDQU32. We were already using the 32 bit element opcode if BWI isn't enabled, but there's no reason to change opcode if we have BWI. We will still use the 8/16 opcodes for masked stores though. This allows us to use the aligned opcode when we can which makes our test output more consistent between different modes. It also reduces the number of isel patterns we need. This is a slight inconsistency with loads which default to 64 bit element opcodes. I'll probably rectify that in a future patch. Differential Revision: https://reviews.llvm.org/D35978 llvm-svn: 309693
2017-08-01 23:31:24 +08:00
; SKX-NEXT: vmovaps %zmm0, (%rsp)
[X86] Fix EXTRACT_VECTOR_ELT with variable index from v32i16 and v64i8 vector. Its more profitable to go through memory (1 cycles throughput) than using VMOVD + VPERMV/PSHUFB sequence ( 2/3 cycles throughput) to implement EXTRACT_VECTOR_ELT with variable index. IACA tool was used to get performace estimation (https://software.intel.com/en-us/articles/intel-architecture-code-analyzer) For example for var_shuffle_v16i8_v16i8_xxxxxxxxxxxxxxxx_i8 test from vector-shuffle-variable-128.ll I get 26 cycles vs 79 cycles. Removing the VINSERT node, we don't need it any more. Differential Revision: https://reviews.llvm.org/D29690 llvm-svn: 295660
2017-02-20 22:16:29 +08:00
; SKX-NEXT: andl $31, %edi
; SKX-NEXT: movzwl (%rsp,%rdi,2), %eax
; SKX-NEXT: movq %rbp, %rsp
; SKX-NEXT: popq %rbp
[X86] Generate VZEROUPPER for Skylake-avx512. VZEROUPPER should not be issued on Knights Landing (KNL), but on Skylake-avx512 it should be. Differential Revision: https://reviews.llvm.org/D29874 llvm-svn: 296859
2017-03-03 17:03:24 +08:00
; SKX-NEXT: vzeroupper
[X86] Fix EXTRACT_VECTOR_ELT with variable index from v32i16 and v64i8 vector. Its more profitable to go through memory (1 cycles throughput) than using VMOVD + VPERMV/PSHUFB sequence ( 2/3 cycles throughput) to implement EXTRACT_VECTOR_ELT with variable index. IACA tool was used to get performace estimation (https://software.intel.com/en-us/articles/intel-architecture-code-analyzer) For example for var_shuffle_v16i8_v16i8_xxxxxxxxxxxxxxxx_i8 test from vector-shuffle-variable-128.ll I get 26 cycles vs 79 cycles. Removing the VINSERT node, we don't need it any more. Differential Revision: https://reviews.llvm.org/D29690 llvm-svn: 295660
2017-02-20 22:16:29 +08:00
; SKX-NEXT: retq
%t2 = extractelement <32 x i16> %t1, i32 %index
ret i16 %t2
}
Add new tests for EXTRACT_VECTOR_ELT (vector of packed i8/16/i32/i64/ps/pd data) llvm-svn: 294565
2017-02-09 15:39:19 +08:00
define i8 @test_extractelement_variable_v16i8(<16 x i8> %t1, i32 %index) {
[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-LABEL: test_extractelement_variable_v16i8:
[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:
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; CHECK-NEXT: ## kill: def $edi killed $edi def $rdi
[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: vmovaps %xmm0, -{{[0-9]+}}(%rsp)
; CHECK-NEXT: andl $15, %edi
[SelectionDAG] Reverse the order of operands in the ISD::ADD created by TargetLowering::getVectorElementPointer so that the FrameIndex is on the left. This seems to improve X86's ability to match this into an address computation. Otherwise the other operand gets assigned to the base register and the stack pointer + frame index ends up in the index register. But index registers can't encode ESP/RSP so we end up having to move it into another register to meet the constraint. I could try to improve the address matcher in X86, but swapping the producer seemed easier. Several other places already have the operands in this order so this is at least consistent. llvm-svn: 321370
2017-12-23 01:18:13 +08:00
; CHECK-NEXT: movb -24(%rsp,%rdi), %al
[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: retq
Add new tests for EXTRACT_VECTOR_ELT (vector of packed i8/16/i32/i64/ps/pd data) llvm-svn: 294565
2017-02-09 15:39:19 +08:00
%t2 = extractelement <16 x i8> %t1, i32 %index
ret i8 %t2
}
define i8 @test_extractelement_variable_v32i8(<32 x i8> %t1, i32 %index) {
[X86][AVX512] Regenerate element insertion/extraction tests llvm-svn: 315322
2017-10-10 23:58:54 +08:00
; CHECK-LABEL: test_extractelement_variable_v32i8:
[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] Regenerate element insertion/extraction tests llvm-svn: 315322
2017-10-10 23:58:54 +08:00
; CHECK-NEXT: pushq %rbp
; CHECK-NEXT: .cfi_def_cfa_offset 16
; CHECK-NEXT: .cfi_offset %rbp, -16
; CHECK-NEXT: movq %rsp, %rbp
; CHECK-NEXT: .cfi_def_cfa_register %rbp
; CHECK-NEXT: andq $-32, %rsp
; CHECK-NEXT: subq $64, %rsp
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; CHECK-NEXT: ## kill: def $edi killed $edi def $rdi
[X86][AVX512] Regenerate element insertion/extraction tests llvm-svn: 315322
2017-10-10 23:58:54 +08:00
; CHECK-NEXT: vmovaps %ymm0, (%rsp)
; CHECK-NEXT: andl $31, %edi
[SelectionDAG] Reverse the order of operands in the ISD::ADD created by TargetLowering::getVectorElementPointer so that the FrameIndex is on the left. This seems to improve X86's ability to match this into an address computation. Otherwise the other operand gets assigned to the base register and the stack pointer + frame index ends up in the index register. But index registers can't encode ESP/RSP so we end up having to move it into another register to meet the constraint. I could try to improve the address matcher in X86, but swapping the producer seemed easier. Several other places already have the operands in this order so this is at least consistent. llvm-svn: 321370
2017-12-23 01:18:13 +08:00
; CHECK-NEXT: movb (%rsp,%rdi), %al
[X86][AVX512] Regenerate element insertion/extraction tests llvm-svn: 315322
2017-10-10 23:58:54 +08:00
; CHECK-NEXT: movq %rbp, %rsp
; CHECK-NEXT: popq %rbp
; CHECK-NEXT: vzeroupper
; CHECK-NEXT: retq
Add new tests for EXTRACT_VECTOR_ELT (vector of packed i8/16/i32/i64/ps/pd data) llvm-svn: 294565
2017-02-09 15:39:19 +08:00
%t2 = extractelement <32 x i8> %t1, i32 %index
ret i8 %t2
}
[X86] Fix EXTRACT_VECTOR_ELT with variable index from v32i16 and v64i8 vector. Its more profitable to go through memory (1 cycles throughput) than using VMOVD + VPERMV/PSHUFB sequence ( 2/3 cycles throughput) to implement EXTRACT_VECTOR_ELT with variable index. IACA tool was used to get performace estimation (https://software.intel.com/en-us/articles/intel-architecture-code-analyzer) For example for var_shuffle_v16i8_v16i8_xxxxxxxxxxxxxxxx_i8 test from vector-shuffle-variable-128.ll I get 26 cycles vs 79 cycles. Removing the VINSERT node, we don't need it any more. Differential Revision: https://reviews.llvm.org/D29690 llvm-svn: 295660
2017-02-20 22:16:29 +08:00
define i8 @test_extractelement_variable_v64i8(<64 x i8> %t1, i32 %index) {
; KNL-LABEL: test_extractelement_variable_v64i8:
[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
; KNL: ## %bb.0:
[X86] Fix EXTRACT_VECTOR_ELT with variable index from v32i16 and v64i8 vector. Its more profitable to go through memory (1 cycles throughput) than using VMOVD + VPERMV/PSHUFB sequence ( 2/3 cycles throughput) to implement EXTRACT_VECTOR_ELT with variable index. IACA tool was used to get performace estimation (https://software.intel.com/en-us/articles/intel-architecture-code-analyzer) For example for var_shuffle_v16i8_v16i8_xxxxxxxxxxxxxxxx_i8 test from vector-shuffle-variable-128.ll I get 26 cycles vs 79 cycles. Removing the VINSERT node, we don't need it any more. Differential Revision: https://reviews.llvm.org/D29690 llvm-svn: 295660
2017-02-20 22:16:29 +08:00
; KNL-NEXT: pushq %rbp
; KNL-NEXT: .cfi_def_cfa_offset 16
; KNL-NEXT: .cfi_offset %rbp, -16
; KNL-NEXT: movq %rsp, %rbp
; KNL-NEXT: .cfi_def_cfa_register %rbp
; KNL-NEXT: andq $-64, %rsp
; KNL-NEXT: subq $128, %rsp
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; KNL-NEXT: ## kill: def $edi killed $edi def $rdi
[X86] Fix EXTRACT_VECTOR_ELT with variable index from v32i16 and v64i8 vector. Its more profitable to go through memory (1 cycles throughput) than using VMOVD + VPERMV/PSHUFB sequence ( 2/3 cycles throughput) to implement EXTRACT_VECTOR_ELT with variable index. IACA tool was used to get performace estimation (https://software.intel.com/en-us/articles/intel-architecture-code-analyzer) For example for var_shuffle_v16i8_v16i8_xxxxxxxxxxxxxxxx_i8 test from vector-shuffle-variable-128.ll I get 26 cycles vs 79 cycles. Removing the VINSERT node, we don't need it any more. Differential Revision: https://reviews.llvm.org/D29690 llvm-svn: 295660
2017-02-20 22:16:29 +08:00
; KNL-NEXT: vmovaps %ymm1, {{[0-9]+}}(%rsp)
; KNL-NEXT: vmovaps %ymm0, (%rsp)
; KNL-NEXT: andl $63, %edi
[SelectionDAG] Reverse the order of operands in the ISD::ADD created by TargetLowering::getVectorElementPointer so that the FrameIndex is on the left. This seems to improve X86's ability to match this into an address computation. Otherwise the other operand gets assigned to the base register and the stack pointer + frame index ends up in the index register. But index registers can't encode ESP/RSP so we end up having to move it into another register to meet the constraint. I could try to improve the address matcher in X86, but swapping the producer seemed easier. Several other places already have the operands in this order so this is at least consistent. llvm-svn: 321370
2017-12-23 01:18:13 +08:00
; KNL-NEXT: movb (%rsp,%rdi), %al
[X86] Fix EXTRACT_VECTOR_ELT with variable index from v32i16 and v64i8 vector. Its more profitable to go through memory (1 cycles throughput) than using VMOVD + VPERMV/PSHUFB sequence ( 2/3 cycles throughput) to implement EXTRACT_VECTOR_ELT with variable index. IACA tool was used to get performace estimation (https://software.intel.com/en-us/articles/intel-architecture-code-analyzer) For example for var_shuffle_v16i8_v16i8_xxxxxxxxxxxxxxxx_i8 test from vector-shuffle-variable-128.ll I get 26 cycles vs 79 cycles. Removing the VINSERT node, we don't need it any more. Differential Revision: https://reviews.llvm.org/D29690 llvm-svn: 295660
2017-02-20 22:16:29 +08:00
; KNL-NEXT: movq %rbp, %rsp
; KNL-NEXT: popq %rbp
[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
; KNL-NEXT: vzeroupper
[X86] Fix EXTRACT_VECTOR_ELT with variable index from v32i16 and v64i8 vector. Its more profitable to go through memory (1 cycles throughput) than using VMOVD + VPERMV/PSHUFB sequence ( 2/3 cycles throughput) to implement EXTRACT_VECTOR_ELT with variable index. IACA tool was used to get performace estimation (https://software.intel.com/en-us/articles/intel-architecture-code-analyzer) For example for var_shuffle_v16i8_v16i8_xxxxxxxxxxxxxxxx_i8 test from vector-shuffle-variable-128.ll I get 26 cycles vs 79 cycles. Removing the VINSERT node, we don't need it any more. Differential Revision: https://reviews.llvm.org/D29690 llvm-svn: 295660
2017-02-20 22:16:29 +08:00
; KNL-NEXT: retq
;
; SKX-LABEL: test_extractelement_variable_v64i8:
[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
; SKX: ## %bb.0:
[X86] Fix EXTRACT_VECTOR_ELT with variable index from v32i16 and v64i8 vector. Its more profitable to go through memory (1 cycles throughput) than using VMOVD + VPERMV/PSHUFB sequence ( 2/3 cycles throughput) to implement EXTRACT_VECTOR_ELT with variable index. IACA tool was used to get performace estimation (https://software.intel.com/en-us/articles/intel-architecture-code-analyzer) For example for var_shuffle_v16i8_v16i8_xxxxxxxxxxxxxxxx_i8 test from vector-shuffle-variable-128.ll I get 26 cycles vs 79 cycles. Removing the VINSERT node, we don't need it any more. Differential Revision: https://reviews.llvm.org/D29690 llvm-svn: 295660
2017-02-20 22:16:29 +08:00
; SKX-NEXT: pushq %rbp
; SKX-NEXT: .cfi_def_cfa_offset 16
; SKX-NEXT: .cfi_offset %rbp, -16
; SKX-NEXT: movq %rsp, %rbp
; SKX-NEXT: .cfi_def_cfa_register %rbp
; SKX-NEXT: andq $-64, %rsp
; SKX-NEXT: subq $128, %rsp
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; SKX-NEXT: ## kill: def $edi killed $edi def $rdi
[AVX-512] Don't use unmasked VMOVDQU8/16 for 8-bit or 16-bit element stores even when BWI instructions are supported. Always use VMOVDQA32/VMOVDQU32. We were already using the 32 bit element opcode if BWI isn't enabled, but there's no reason to change opcode if we have BWI. We will still use the 8/16 opcodes for masked stores though. This allows us to use the aligned opcode when we can which makes our test output more consistent between different modes. It also reduces the number of isel patterns we need. This is a slight inconsistency with loads which default to 64 bit element opcodes. I'll probably rectify that in a future patch. Differential Revision: https://reviews.llvm.org/D35978 llvm-svn: 309693
2017-08-01 23:31:24 +08:00
; SKX-NEXT: vmovaps %zmm0, (%rsp)
[X86] Fix EXTRACT_VECTOR_ELT with variable index from v32i16 and v64i8 vector. Its more profitable to go through memory (1 cycles throughput) than using VMOVD + VPERMV/PSHUFB sequence ( 2/3 cycles throughput) to implement EXTRACT_VECTOR_ELT with variable index. IACA tool was used to get performace estimation (https://software.intel.com/en-us/articles/intel-architecture-code-analyzer) For example for var_shuffle_v16i8_v16i8_xxxxxxxxxxxxxxxx_i8 test from vector-shuffle-variable-128.ll I get 26 cycles vs 79 cycles. Removing the VINSERT node, we don't need it any more. Differential Revision: https://reviews.llvm.org/D29690 llvm-svn: 295660
2017-02-20 22:16:29 +08:00
; SKX-NEXT: andl $63, %edi
[SelectionDAG] Reverse the order of operands in the ISD::ADD created by TargetLowering::getVectorElementPointer so that the FrameIndex is on the left. This seems to improve X86's ability to match this into an address computation. Otherwise the other operand gets assigned to the base register and the stack pointer + frame index ends up in the index register. But index registers can't encode ESP/RSP so we end up having to move it into another register to meet the constraint. I could try to improve the address matcher in X86, but swapping the producer seemed easier. Several other places already have the operands in this order so this is at least consistent. llvm-svn: 321370
2017-12-23 01:18:13 +08:00
; SKX-NEXT: movb (%rsp,%rdi), %al
[X86] Fix EXTRACT_VECTOR_ELT with variable index from v32i16 and v64i8 vector. Its more profitable to go through memory (1 cycles throughput) than using VMOVD + VPERMV/PSHUFB sequence ( 2/3 cycles throughput) to implement EXTRACT_VECTOR_ELT with variable index. IACA tool was used to get performace estimation (https://software.intel.com/en-us/articles/intel-architecture-code-analyzer) For example for var_shuffle_v16i8_v16i8_xxxxxxxxxxxxxxxx_i8 test from vector-shuffle-variable-128.ll I get 26 cycles vs 79 cycles. Removing the VINSERT node, we don't need it any more. Differential Revision: https://reviews.llvm.org/D29690 llvm-svn: 295660
2017-02-20 22:16:29 +08:00
; SKX-NEXT: movq %rbp, %rsp
; SKX-NEXT: popq %rbp
[X86] Generate VZEROUPPER for Skylake-avx512. VZEROUPPER should not be issued on Knights Landing (KNL), but on Skylake-avx512 it should be. Differential Revision: https://reviews.llvm.org/D29874 llvm-svn: 296859
2017-03-03 17:03:24 +08:00
; SKX-NEXT: vzeroupper
[X86] Fix EXTRACT_VECTOR_ELT with variable index from v32i16 and v64i8 vector. Its more profitable to go through memory (1 cycles throughput) than using VMOVD + VPERMV/PSHUFB sequence ( 2/3 cycles throughput) to implement EXTRACT_VECTOR_ELT with variable index. IACA tool was used to get performace estimation (https://software.intel.com/en-us/articles/intel-architecture-code-analyzer) For example for var_shuffle_v16i8_v16i8_xxxxxxxxxxxxxxxx_i8 test from vector-shuffle-variable-128.ll I get 26 cycles vs 79 cycles. Removing the VINSERT node, we don't need it any more. Differential Revision: https://reviews.llvm.org/D29690 llvm-svn: 295660
2017-02-20 22:16:29 +08:00
; SKX-NEXT: retq
%t2 = extractelement <64 x i8> %t1, i32 %index
ret i8 %t2
}
define i8 @test_extractelement_variable_v64i8_indexi8(<64 x i8> %t1, i8 %index) {
; KNL-LABEL: test_extractelement_variable_v64i8_indexi8:
[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
; KNL: ## %bb.0:
[X86] Fix EXTRACT_VECTOR_ELT with variable index from v32i16 and v64i8 vector. Its more profitable to go through memory (1 cycles throughput) than using VMOVD + VPERMV/PSHUFB sequence ( 2/3 cycles throughput) to implement EXTRACT_VECTOR_ELT with variable index. IACA tool was used to get performace estimation (https://software.intel.com/en-us/articles/intel-architecture-code-analyzer) For example for var_shuffle_v16i8_v16i8_xxxxxxxxxxxxxxxx_i8 test from vector-shuffle-variable-128.ll I get 26 cycles vs 79 cycles. Removing the VINSERT node, we don't need it any more. Differential Revision: https://reviews.llvm.org/D29690 llvm-svn: 295660
2017-02-20 22:16:29 +08:00
; KNL-NEXT: pushq %rbp
; KNL-NEXT: .cfi_def_cfa_offset 16
; KNL-NEXT: .cfi_offset %rbp, -16
; KNL-NEXT: movq %rsp, %rbp
; KNL-NEXT: .cfi_def_cfa_register %rbp
; KNL-NEXT: andq $-64, %rsp
; KNL-NEXT: subq $128, %rsp
; KNL-NEXT: addb %dil, %dil
; KNL-NEXT: vmovaps %ymm1, {{[0-9]+}}(%rsp)
; KNL-NEXT: vmovaps %ymm0, (%rsp)
; KNL-NEXT: movzbl %dil, %eax
; KNL-NEXT: andl $63, %eax
[SelectionDAG] Reverse the order of operands in the ISD::ADD created by TargetLowering::getVectorElementPointer so that the FrameIndex is on the left. This seems to improve X86's ability to match this into an address computation. Otherwise the other operand gets assigned to the base register and the stack pointer + frame index ends up in the index register. But index registers can't encode ESP/RSP so we end up having to move it into another register to meet the constraint. I could try to improve the address matcher in X86, but swapping the producer seemed easier. Several other places already have the operands in this order so this is at least consistent. llvm-svn: 321370
2017-12-23 01:18:13 +08:00
; KNL-NEXT: movb (%rsp,%rax), %al
[X86] Fix EXTRACT_VECTOR_ELT with variable index from v32i16 and v64i8 vector. Its more profitable to go through memory (1 cycles throughput) than using VMOVD + VPERMV/PSHUFB sequence ( 2/3 cycles throughput) to implement EXTRACT_VECTOR_ELT with variable index. IACA tool was used to get performace estimation (https://software.intel.com/en-us/articles/intel-architecture-code-analyzer) For example for var_shuffle_v16i8_v16i8_xxxxxxxxxxxxxxxx_i8 test from vector-shuffle-variable-128.ll I get 26 cycles vs 79 cycles. Removing the VINSERT node, we don't need it any more. Differential Revision: https://reviews.llvm.org/D29690 llvm-svn: 295660
2017-02-20 22:16:29 +08:00
; KNL-NEXT: movq %rbp, %rsp
; KNL-NEXT: popq %rbp
[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
; KNL-NEXT: vzeroupper
[X86] Fix EXTRACT_VECTOR_ELT with variable index from v32i16 and v64i8 vector. Its more profitable to go through memory (1 cycles throughput) than using VMOVD + VPERMV/PSHUFB sequence ( 2/3 cycles throughput) to implement EXTRACT_VECTOR_ELT with variable index. IACA tool was used to get performace estimation (https://software.intel.com/en-us/articles/intel-architecture-code-analyzer) For example for var_shuffle_v16i8_v16i8_xxxxxxxxxxxxxxxx_i8 test from vector-shuffle-variable-128.ll I get 26 cycles vs 79 cycles. Removing the VINSERT node, we don't need it any more. Differential Revision: https://reviews.llvm.org/D29690 llvm-svn: 295660
2017-02-20 22:16:29 +08:00
; KNL-NEXT: retq
;
; SKX-LABEL: test_extractelement_variable_v64i8_indexi8:
[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
; SKX: ## %bb.0:
[X86] Fix EXTRACT_VECTOR_ELT with variable index from v32i16 and v64i8 vector. Its more profitable to go through memory (1 cycles throughput) than using VMOVD + VPERMV/PSHUFB sequence ( 2/3 cycles throughput) to implement EXTRACT_VECTOR_ELT with variable index. IACA tool was used to get performace estimation (https://software.intel.com/en-us/articles/intel-architecture-code-analyzer) For example for var_shuffle_v16i8_v16i8_xxxxxxxxxxxxxxxx_i8 test from vector-shuffle-variable-128.ll I get 26 cycles vs 79 cycles. Removing the VINSERT node, we don't need it any more. Differential Revision: https://reviews.llvm.org/D29690 llvm-svn: 295660
2017-02-20 22:16:29 +08:00
; SKX-NEXT: pushq %rbp
; SKX-NEXT: .cfi_def_cfa_offset 16
; SKX-NEXT: .cfi_offset %rbp, -16
; SKX-NEXT: movq %rsp, %rbp
; SKX-NEXT: .cfi_def_cfa_register %rbp
; SKX-NEXT: andq $-64, %rsp
; SKX-NEXT: subq $128, %rsp
; SKX-NEXT: addb %dil, %dil
[AVX-512] Don't use unmasked VMOVDQU8/16 for 8-bit or 16-bit element stores even when BWI instructions are supported. Always use VMOVDQA32/VMOVDQU32. We were already using the 32 bit element opcode if BWI isn't enabled, but there's no reason to change opcode if we have BWI. We will still use the 8/16 opcodes for masked stores though. This allows us to use the aligned opcode when we can which makes our test output more consistent between different modes. It also reduces the number of isel patterns we need. This is a slight inconsistency with loads which default to 64 bit element opcodes. I'll probably rectify that in a future patch. Differential Revision: https://reviews.llvm.org/D35978 llvm-svn: 309693
2017-08-01 23:31:24 +08:00
; SKX-NEXT: vmovaps %zmm0, (%rsp)
[X86] Fix EXTRACT_VECTOR_ELT with variable index from v32i16 and v64i8 vector. Its more profitable to go through memory (1 cycles throughput) than using VMOVD + VPERMV/PSHUFB sequence ( 2/3 cycles throughput) to implement EXTRACT_VECTOR_ELT with variable index. IACA tool was used to get performace estimation (https://software.intel.com/en-us/articles/intel-architecture-code-analyzer) For example for var_shuffle_v16i8_v16i8_xxxxxxxxxxxxxxxx_i8 test from vector-shuffle-variable-128.ll I get 26 cycles vs 79 cycles. Removing the VINSERT node, we don't need it any more. Differential Revision: https://reviews.llvm.org/D29690 llvm-svn: 295660
2017-02-20 22:16:29 +08:00
; SKX-NEXT: movzbl %dil, %eax
; SKX-NEXT: andl $63, %eax
[SelectionDAG] Reverse the order of operands in the ISD::ADD created by TargetLowering::getVectorElementPointer so that the FrameIndex is on the left. This seems to improve X86's ability to match this into an address computation. Otherwise the other operand gets assigned to the base register and the stack pointer + frame index ends up in the index register. But index registers can't encode ESP/RSP so we end up having to move it into another register to meet the constraint. I could try to improve the address matcher in X86, but swapping the producer seemed easier. Several other places already have the operands in this order so this is at least consistent. llvm-svn: 321370
2017-12-23 01:18:13 +08:00
; SKX-NEXT: movb (%rsp,%rax), %al
[X86] Fix EXTRACT_VECTOR_ELT with variable index from v32i16 and v64i8 vector. Its more profitable to go through memory (1 cycles throughput) than using VMOVD + VPERMV/PSHUFB sequence ( 2/3 cycles throughput) to implement EXTRACT_VECTOR_ELT with variable index. IACA tool was used to get performace estimation (https://software.intel.com/en-us/articles/intel-architecture-code-analyzer) For example for var_shuffle_v16i8_v16i8_xxxxxxxxxxxxxxxx_i8 test from vector-shuffle-variable-128.ll I get 26 cycles vs 79 cycles. Removing the VINSERT node, we don't need it any more. Differential Revision: https://reviews.llvm.org/D29690 llvm-svn: 295660
2017-02-20 22:16:29 +08:00
; SKX-NEXT: movq %rbp, %rsp
; SKX-NEXT: popq %rbp
[X86] Generate VZEROUPPER for Skylake-avx512. VZEROUPPER should not be issued on Knights Landing (KNL), but on Skylake-avx512 it should be. Differential Revision: https://reviews.llvm.org/D29874 llvm-svn: 296859
2017-03-03 17:03:24 +08:00
; SKX-NEXT: vzeroupper
[X86] Fix EXTRACT_VECTOR_ELT with variable index from v32i16 and v64i8 vector. Its more profitable to go through memory (1 cycles throughput) than using VMOVD + VPERMV/PSHUFB sequence ( 2/3 cycles throughput) to implement EXTRACT_VECTOR_ELT with variable index. IACA tool was used to get performace estimation (https://software.intel.com/en-us/articles/intel-architecture-code-analyzer) For example for var_shuffle_v16i8_v16i8_xxxxxxxxxxxxxxxx_i8 test from vector-shuffle-variable-128.ll I get 26 cycles vs 79 cycles. Removing the VINSERT node, we don't need it any more. Differential Revision: https://reviews.llvm.org/D29690 llvm-svn: 295660
2017-02-20 22:16:29 +08:00
; SKX-NEXT: retq
%i = add i8 %index, %index
%t2 = extractelement <64 x i8> %t1, i8 %i
ret i8 %t2
}
[AVX512] Fix EXTRACT_VECTOR_ELT for v2i1/v4i1/v32i1/v64i1 with variable index. Differential Revision: https://reviews.llvm.org/D30189 llvm-svn: 295718
2017-02-21 22:01:25 +08:00
define zeroext i8 @test_extractelement_varible_v2i1(<2 x i64> %a, <2 x i64> %b, i32 %index) {
; KNL-LABEL: test_extractelement_varible_v2i1:
[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
; KNL: ## %bb.0:
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; KNL-NEXT: ## kill: def $edi killed $edi def $rdi
; KNL-NEXT: ## kill: def $xmm1 killed $xmm1 def $zmm1
; KNL-NEXT: ## kill: def $xmm0 killed $xmm0 def $zmm0
[X86] Make v2i1 and v4i1 legal types without VLX Summary: There are few oddities that occur due to v1i1, v8i1, v16i1 being legal without v2i1 and v4i1 being legal when we don't have VLX. Particularly during legalization of v2i32/v4i32/v2i64/v4i64 masked gather/scatter/load/store. We end up promoting the mask argument to these during type legalization and then have to widen the promoted type to v8iX/v16iX and truncate it to get the element size back down to v8i1/v16i1 to use a 512-bit operation. Since need to fill the upper bits of the mask we have to fill with 0s at the promoted type. It would be better if we could just have the v2i1/v4i1 types as legal so they don't undergo any promotion. Then we can just widen with 0s directly in a k register. There are no real v4i1/v2i1 instructions anyway. Everything is done on a larger register anyway. This also fixes an issue that we couldn't implement a masked vextractf32x4 from zmm to xmm properly. We now have to support widening more compares to 512-bit to get a mask result out so new tablegen patterns got added. I had to hack the legalizer for widening the operand of a setcc a bit so it didn't try create a setcc returning v4i32, extract from it, then try to promote it using a sign extend to v2i1. Now we create the setcc with v4i1 if the original setcc's result type is v2i1. Then extract that and don't sign extend it at all. There's definitely room for improvement with some follow up patches. Reviewers: RKSimon, zvi, guyblank Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D41560 llvm-svn: 321967
2018-01-08 02:20:37 +08:00
; KNL-NEXT: vpcmpnleuq %zmm1, %zmm0, %k1
; KNL-NEXT: vpternlogq $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; KNL-NEXT: vextracti32x4 $0, %zmm0, -{{[0-9]+}}(%rsp)
[AVX512] Fix EXTRACT_VECTOR_ELT for v2i1/v4i1/v32i1/v64i1 with variable index. Differential Revision: https://reviews.llvm.org/D30189 llvm-svn: 295718
2017-02-21 22:01:25 +08:00
; KNL-NEXT: andl $1, %edi
[X86] Make v2i1 and v4i1 legal types without VLX Summary: There are few oddities that occur due to v1i1, v8i1, v16i1 being legal without v2i1 and v4i1 being legal when we don't have VLX. Particularly during legalization of v2i32/v4i32/v2i64/v4i64 masked gather/scatter/load/store. We end up promoting the mask argument to these during type legalization and then have to widen the promoted type to v8iX/v16iX and truncate it to get the element size back down to v8i1/v16i1 to use a 512-bit operation. Since need to fill the upper bits of the mask we have to fill with 0s at the promoted type. It would be better if we could just have the v2i1/v4i1 types as legal so they don't undergo any promotion. Then we can just widen with 0s directly in a k register. There are no real v4i1/v2i1 instructions anyway. Everything is done on a larger register anyway. This also fixes an issue that we couldn't implement a masked vextractf32x4 from zmm to xmm properly. We now have to support widening more compares to 512-bit to get a mask result out so new tablegen patterns got added. I had to hack the legalizer for widening the operand of a setcc a bit so it didn't try create a setcc returning v4i32, extract from it, then try to promote it using a sign extend to v2i1. Now we create the setcc with v4i1 if the original setcc's result type is v2i1. Then extract that and don't sign extend it at all. There's definitely room for improvement with some follow up patches. Reviewers: RKSimon, zvi, guyblank Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D41560 llvm-svn: 321967
2018-01-08 02:20:37 +08:00
; KNL-NEXT: movzbl -24(%rsp,%rdi,8), %eax
[AVX512] Fix EXTRACT_VECTOR_ELT for v2i1/v4i1/v32i1/v64i1 with variable index. Differential Revision: https://reviews.llvm.org/D30189 llvm-svn: 295718
2017-02-21 22:01:25 +08:00
; KNL-NEXT: andl $1, %eax
[X86] Make v2i1 and v4i1 legal types without VLX Summary: There are few oddities that occur due to v1i1, v8i1, v16i1 being legal without v2i1 and v4i1 being legal when we don't have VLX. Particularly during legalization of v2i32/v4i32/v2i64/v4i64 masked gather/scatter/load/store. We end up promoting the mask argument to these during type legalization and then have to widen the promoted type to v8iX/v16iX and truncate it to get the element size back down to v8i1/v16i1 to use a 512-bit operation. Since need to fill the upper bits of the mask we have to fill with 0s at the promoted type. It would be better if we could just have the v2i1/v4i1 types as legal so they don't undergo any promotion. Then we can just widen with 0s directly in a k register. There are no real v4i1/v2i1 instructions anyway. Everything is done on a larger register anyway. This also fixes an issue that we couldn't implement a masked vextractf32x4 from zmm to xmm properly. We now have to support widening more compares to 512-bit to get a mask result out so new tablegen patterns got added. I had to hack the legalizer for widening the operand of a setcc a bit so it didn't try create a setcc returning v4i32, extract from it, then try to promote it using a sign extend to v2i1. Now we create the setcc with v4i1 if the original setcc's result type is v2i1. Then extract that and don't sign extend it at all. There's definitely room for improvement with some follow up patches. Reviewers: RKSimon, zvi, guyblank Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D41560 llvm-svn: 321967
2018-01-08 02:20:37 +08:00
; KNL-NEXT: vzeroupper
[AVX512] Fix EXTRACT_VECTOR_ELT for v2i1/v4i1/v32i1/v64i1 with variable index. Differential Revision: https://reviews.llvm.org/D30189 llvm-svn: 295718
2017-02-21 22:01:25 +08:00
; KNL-NEXT: retq
;
; SKX-LABEL: test_extractelement_varible_v2i1:
[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
; SKX: ## %bb.0:
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; SKX-NEXT: ## kill: def $edi killed $edi def $rdi
[AVX512] Fix EXTRACT_VECTOR_ELT for v2i1/v4i1/v32i1/v64i1 with variable index. Differential Revision: https://reviews.llvm.org/D30189 llvm-svn: 295718
2017-02-21 22:01:25 +08:00
; SKX-NEXT: vpcmpnleuq %xmm1, %xmm0, %k0
; SKX-NEXT: vpmovm2q %k0, %xmm0
; SKX-NEXT: vmovdqa %xmm0, -{{[0-9]+}}(%rsp)
; SKX-NEXT: andl $1, %edi
[X86][AVX512] Make i1 illegal in the CodeGen This patch defines the i1 type as illegal in the X86 backend for AVX512. For DAG operations on <N x i1> types (build vector, extract vector element, ...) i8 is used, and should be truncated/extended. This should produce better scalar code for i1 types since GPRs will be used instead of mask registers. Differential Revision: https://reviews.llvm.org/D32273 llvm-svn: 303421
2017-05-19 20:35:15 +08:00
; SKX-NEXT: movzbl -24(%rsp,%rdi,8), %eax
[AVX512] Fix EXTRACT_VECTOR_ELT for v2i1/v4i1/v32i1/v64i1 with variable index. Differential Revision: https://reviews.llvm.org/D30189 llvm-svn: 295718
2017-02-21 22:01:25 +08:00
; SKX-NEXT: andl $1, %eax
; SKX-NEXT: retq
%t1 = icmp ugt <2 x i64> %a, %b
%t2 = extractelement <2 x i1> %t1, i32 %index
%res = zext i1 %t2 to i8
ret i8 %res
}
define zeroext i8 @test_extractelement_varible_v4i1(<4 x i32> %a, <4 x i32> %b, i32 %index) {
; KNL-LABEL: test_extractelement_varible_v4i1:
[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
; KNL: ## %bb.0:
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; KNL-NEXT: ## kill: def $edi killed $edi def $rdi
; KNL-NEXT: ## kill: def $xmm1 killed $xmm1 def $zmm1
; KNL-NEXT: ## kill: def $xmm0 killed $xmm0 def $zmm0
[X86] Make v2i1 and v4i1 legal types without VLX Summary: There are few oddities that occur due to v1i1, v8i1, v16i1 being legal without v2i1 and v4i1 being legal when we don't have VLX. Particularly during legalization of v2i32/v4i32/v2i64/v4i64 masked gather/scatter/load/store. We end up promoting the mask argument to these during type legalization and then have to widen the promoted type to v8iX/v16iX and truncate it to get the element size back down to v8i1/v16i1 to use a 512-bit operation. Since need to fill the upper bits of the mask we have to fill with 0s at the promoted type. It would be better if we could just have the v2i1/v4i1 types as legal so they don't undergo any promotion. Then we can just widen with 0s directly in a k register. There are no real v4i1/v2i1 instructions anyway. Everything is done on a larger register anyway. This also fixes an issue that we couldn't implement a masked vextractf32x4 from zmm to xmm properly. We now have to support widening more compares to 512-bit to get a mask result out so new tablegen patterns got added. I had to hack the legalizer for widening the operand of a setcc a bit so it didn't try create a setcc returning v4i32, extract from it, then try to promote it using a sign extend to v2i1. Now we create the setcc with v4i1 if the original setcc's result type is v2i1. Then extract that and don't sign extend it at all. There's definitely room for improvement with some follow up patches. Reviewers: RKSimon, zvi, guyblank Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D41560 llvm-svn: 321967
2018-01-08 02:20:37 +08:00
; KNL-NEXT: vpcmpnleud %zmm1, %zmm0, %k1
; KNL-NEXT: vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; KNL-NEXT: vextracti32x4 $0, %zmm0, -{{[0-9]+}}(%rsp)
[AVX512] Fix EXTRACT_VECTOR_ELT for v2i1/v4i1/v32i1/v64i1 with variable index. Differential Revision: https://reviews.llvm.org/D30189 llvm-svn: 295718
2017-02-21 22:01:25 +08:00
; KNL-NEXT: andl $3, %edi
[X86] Make v2i1 and v4i1 legal types without VLX Summary: There are few oddities that occur due to v1i1, v8i1, v16i1 being legal without v2i1 and v4i1 being legal when we don't have VLX. Particularly during legalization of v2i32/v4i32/v2i64/v4i64 masked gather/scatter/load/store. We end up promoting the mask argument to these during type legalization and then have to widen the promoted type to v8iX/v16iX and truncate it to get the element size back down to v8i1/v16i1 to use a 512-bit operation. Since need to fill the upper bits of the mask we have to fill with 0s at the promoted type. It would be better if we could just have the v2i1/v4i1 types as legal so they don't undergo any promotion. Then we can just widen with 0s directly in a k register. There are no real v4i1/v2i1 instructions anyway. Everything is done on a larger register anyway. This also fixes an issue that we couldn't implement a masked vextractf32x4 from zmm to xmm properly. We now have to support widening more compares to 512-bit to get a mask result out so new tablegen patterns got added. I had to hack the legalizer for widening the operand of a setcc a bit so it didn't try create a setcc returning v4i32, extract from it, then try to promote it using a sign extend to v2i1. Now we create the setcc with v4i1 if the original setcc's result type is v2i1. Then extract that and don't sign extend it at all. There's definitely room for improvement with some follow up patches. Reviewers: RKSimon, zvi, guyblank Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D41560 llvm-svn: 321967
2018-01-08 02:20:37 +08:00
; KNL-NEXT: movzbl -24(%rsp,%rdi,4), %eax
[AVX512] Fix EXTRACT_VECTOR_ELT for v2i1/v4i1/v32i1/v64i1 with variable index. Differential Revision: https://reviews.llvm.org/D30189 llvm-svn: 295718
2017-02-21 22:01:25 +08:00
; KNL-NEXT: andl $1, %eax
[X86] Make v2i1 and v4i1 legal types without VLX Summary: There are few oddities that occur due to v1i1, v8i1, v16i1 being legal without v2i1 and v4i1 being legal when we don't have VLX. Particularly during legalization of v2i32/v4i32/v2i64/v4i64 masked gather/scatter/load/store. We end up promoting the mask argument to these during type legalization and then have to widen the promoted type to v8iX/v16iX and truncate it to get the element size back down to v8i1/v16i1 to use a 512-bit operation. Since need to fill the upper bits of the mask we have to fill with 0s at the promoted type. It would be better if we could just have the v2i1/v4i1 types as legal so they don't undergo any promotion. Then we can just widen with 0s directly in a k register. There are no real v4i1/v2i1 instructions anyway. Everything is done on a larger register anyway. This also fixes an issue that we couldn't implement a masked vextractf32x4 from zmm to xmm properly. We now have to support widening more compares to 512-bit to get a mask result out so new tablegen patterns got added. I had to hack the legalizer for widening the operand of a setcc a bit so it didn't try create a setcc returning v4i32, extract from it, then try to promote it using a sign extend to v2i1. Now we create the setcc with v4i1 if the original setcc's result type is v2i1. Then extract that and don't sign extend it at all. There's definitely room for improvement with some follow up patches. Reviewers: RKSimon, zvi, guyblank Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D41560 llvm-svn: 321967
2018-01-08 02:20:37 +08:00
; KNL-NEXT: vzeroupper
[AVX512] Fix EXTRACT_VECTOR_ELT for v2i1/v4i1/v32i1/v64i1 with variable index. Differential Revision: https://reviews.llvm.org/D30189 llvm-svn: 295718
2017-02-21 22:01:25 +08:00
; KNL-NEXT: retq
;
; SKX-LABEL: test_extractelement_varible_v4i1:
[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
; SKX: ## %bb.0:
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; SKX-NEXT: ## kill: def $edi killed $edi def $rdi
[AVX512] Fix EXTRACT_VECTOR_ELT for v2i1/v4i1/v32i1/v64i1 with variable index. Differential Revision: https://reviews.llvm.org/D30189 llvm-svn: 295718
2017-02-21 22:01:25 +08:00
; SKX-NEXT: vpcmpnleud %xmm1, %xmm0, %k0
; SKX-NEXT: vpmovm2d %k0, %xmm0
; SKX-NEXT: vmovdqa %xmm0, -{{[0-9]+}}(%rsp)
; SKX-NEXT: andl $3, %edi
[X86][AVX512] Make i1 illegal in the CodeGen This patch defines the i1 type as illegal in the X86 backend for AVX512. For DAG operations on <N x i1> types (build vector, extract vector element, ...) i8 is used, and should be truncated/extended. This should produce better scalar code for i1 types since GPRs will be used instead of mask registers. Differential Revision: https://reviews.llvm.org/D32273 llvm-svn: 303421
2017-05-19 20:35:15 +08:00
; SKX-NEXT: movzbl -24(%rsp,%rdi,4), %eax
[AVX512] Fix EXTRACT_VECTOR_ELT for v2i1/v4i1/v32i1/v64i1 with variable index. Differential Revision: https://reviews.llvm.org/D30189 llvm-svn: 295718
2017-02-21 22:01:25 +08:00
; SKX-NEXT: andl $1, %eax
; SKX-NEXT: retq
%t1 = icmp ugt <4 x i32> %a, %b
%t2 = extractelement <4 x i1> %t1, i32 %index
%res = zext i1 %t2 to i8
ret i8 %res
}
define zeroext i8 @test_extractelement_varible_v8i1(<8 x i32> %a, <8 x i32> %b, i32 %index) {
; KNL-LABEL: test_extractelement_varible_v8i1:
[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
; KNL: ## %bb.0:
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; KNL-NEXT: ## kill: def $edi killed $edi def $rdi
; KNL-NEXT: ## kill: def $ymm1 killed $ymm1 def $zmm1
; KNL-NEXT: ## kill: def $ymm0 killed $ymm0 def $zmm0
[AVX512] Fix EXTRACT_VECTOR_ELT for v2i1/v4i1/v32i1/v64i1 with variable index. Differential Revision: https://reviews.llvm.org/D30189 llvm-svn: 295718
2017-02-21 22:01:25 +08:00
; KNL-NEXT: vpcmpnleud %zmm1, %zmm0, %k1
[X86] When lowering insert_vector_elt/extract_vector_elt of vXi1 with a non-constant index just use either a 128-bit type or the vXi8 type with the correct number of elements. Despite what the comment said there isn't better codegen for 512-bit vectors. The 128/256/512 bit implementation jus stores to memory and loads an element. There's no advantage to doing that with a larger size. In fact in many cases it causes a stack realignment and generates worse code. llvm-svn: 321369
2017-12-23 01:18:11 +08:00
; KNL-NEXT: vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; KNL-NEXT: vpmovdw %zmm0, %ymm0
; KNL-NEXT: vmovdqa %xmm0, -{{[0-9]+}}(%rsp)
[AVX512] Fix EXTRACT_VECTOR_ELT for v2i1/v4i1/v32i1/v64i1 with variable index. Differential Revision: https://reviews.llvm.org/D30189 llvm-svn: 295718
2017-02-21 22:01:25 +08:00
; KNL-NEXT: andl $7, %edi
[X86] When lowering insert_vector_elt/extract_vector_elt of vXi1 with a non-constant index just use either a 128-bit type or the vXi8 type with the correct number of elements. Despite what the comment said there isn't better codegen for 512-bit vectors. The 128/256/512 bit implementation jus stores to memory and loads an element. There's no advantage to doing that with a larger size. In fact in many cases it causes a stack realignment and generates worse code. llvm-svn: 321369
2017-12-23 01:18:11 +08:00
; KNL-NEXT: movzbl -24(%rsp,%rdi,2), %eax
[AVX512] Fix EXTRACT_VECTOR_ELT for v2i1/v4i1/v32i1/v64i1 with variable index. Differential Revision: https://reviews.llvm.org/D30189 llvm-svn: 295718
2017-02-21 22:01:25 +08:00
; KNL-NEXT: andl $1, %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
; KNL-NEXT: vzeroupper
[AVX512] Fix EXTRACT_VECTOR_ELT for v2i1/v4i1/v32i1/v64i1 with variable index. Differential Revision: https://reviews.llvm.org/D30189 llvm-svn: 295718
2017-02-21 22:01:25 +08:00
; KNL-NEXT: retq
;
; SKX-LABEL: test_extractelement_varible_v8i1:
[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
; SKX: ## %bb.0:
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; SKX-NEXT: ## kill: def $edi killed $edi def $rdi
[AVX512] Fix EXTRACT_VECTOR_ELT for v2i1/v4i1/v32i1/v64i1 with variable index. Differential Revision: https://reviews.llvm.org/D30189 llvm-svn: 295718
2017-02-21 22:01:25 +08:00
; SKX-NEXT: vpcmpnleud %ymm1, %ymm0, %k0
[X86] When lowering insert_vector_elt/extract_vector_elt of vXi1 with a non-constant index just use either a 128-bit type or the vXi8 type with the correct number of elements. Despite what the comment said there isn't better codegen for 512-bit vectors. The 128/256/512 bit implementation jus stores to memory and loads an element. There's no advantage to doing that with a larger size. In fact in many cases it causes a stack realignment and generates worse code. llvm-svn: 321369
2017-12-23 01:18:11 +08:00
; SKX-NEXT: vpmovm2w %k0, %xmm0
; SKX-NEXT: vmovdqa %xmm0, -{{[0-9]+}}(%rsp)
[AVX512] Fix EXTRACT_VECTOR_ELT for v2i1/v4i1/v32i1/v64i1 with variable index. Differential Revision: https://reviews.llvm.org/D30189 llvm-svn: 295718
2017-02-21 22:01:25 +08:00
; SKX-NEXT: andl $7, %edi
[X86] When lowering insert_vector_elt/extract_vector_elt of vXi1 with a non-constant index just use either a 128-bit type or the vXi8 type with the correct number of elements. Despite what the comment said there isn't better codegen for 512-bit vectors. The 128/256/512 bit implementation jus stores to memory and loads an element. There's no advantage to doing that with a larger size. In fact in many cases it causes a stack realignment and generates worse code. llvm-svn: 321369
2017-12-23 01:18:11 +08:00
; SKX-NEXT: movzbl -24(%rsp,%rdi,2), %eax
[AVX512] Fix EXTRACT_VECTOR_ELT for v2i1/v4i1/v32i1/v64i1 with variable index. Differential Revision: https://reviews.llvm.org/D30189 llvm-svn: 295718
2017-02-21 22:01:25 +08:00
; SKX-NEXT: andl $1, %eax
[X86] Generate VZEROUPPER for Skylake-avx512. VZEROUPPER should not be issued on Knights Landing (KNL), but on Skylake-avx512 it should be. Differential Revision: https://reviews.llvm.org/D29874 llvm-svn: 296859
2017-03-03 17:03:24 +08:00
; SKX-NEXT: vzeroupper
[AVX512] Fix EXTRACT_VECTOR_ELT for v2i1/v4i1/v32i1/v64i1 with variable index. Differential Revision: https://reviews.llvm.org/D30189 llvm-svn: 295718
2017-02-21 22:01:25 +08:00
; SKX-NEXT: retq
%t1 = icmp ugt <8 x i32> %a, %b
%t2 = extractelement <8 x i1> %t1, i32 %index
%res = zext i1 %t2 to i8
ret i8 %res
}
define zeroext i8 @test_extractelement_varible_v16i1(<16 x i32> %a, <16 x i32> %b, i32 %index) {
; KNL-LABEL: test_extractelement_varible_v16i1:
[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
; KNL: ## %bb.0:
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; KNL-NEXT: ## kill: def $edi killed $edi def $rdi
[AVX512] Fix EXTRACT_VECTOR_ELT for v2i1/v4i1/v32i1/v64i1 with variable index. Differential Revision: https://reviews.llvm.org/D30189 llvm-svn: 295718
2017-02-21 22:01:25 +08:00
; KNL-NEXT: vpcmpnleud %zmm1, %zmm0, %k1
; KNL-NEXT: vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
[X86] Use ISD::TRUNCATE instead of X86ISD::VTRUNC when input and output types have the same number of elements. llvm-svn: 322455
2018-01-14 16:11:36 +08:00
; KNL-NEXT: vpmovdb %zmm0, -{{[0-9]+}}(%rsp)
[AVX512] Fix EXTRACT_VECTOR_ELT for v2i1/v4i1/v32i1/v64i1 with variable index. Differential Revision: https://reviews.llvm.org/D30189 llvm-svn: 295718
2017-02-21 22:01:25 +08:00
; KNL-NEXT: andl $15, %edi
[SelectionDAG] Reverse the order of operands in the ISD::ADD created by TargetLowering::getVectorElementPointer so that the FrameIndex is on the left. This seems to improve X86's ability to match this into an address computation. Otherwise the other operand gets assigned to the base register and the stack pointer + frame index ends up in the index register. But index registers can't encode ESP/RSP so we end up having to move it into another register to meet the constraint. I could try to improve the address matcher in X86, but swapping the producer seemed easier. Several other places already have the operands in this order so this is at least consistent. llvm-svn: 321370
2017-12-23 01:18:13 +08:00
; KNL-NEXT: movzbl -24(%rsp,%rdi), %eax
[AVX512] Fix EXTRACT_VECTOR_ELT for v2i1/v4i1/v32i1/v64i1 with variable index. Differential Revision: https://reviews.llvm.org/D30189 llvm-svn: 295718
2017-02-21 22:01:25 +08:00
; KNL-NEXT: andl $1, %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
; KNL-NEXT: vzeroupper
[AVX512] Fix EXTRACT_VECTOR_ELT for v2i1/v4i1/v32i1/v64i1 with variable index. Differential Revision: https://reviews.llvm.org/D30189 llvm-svn: 295718
2017-02-21 22:01:25 +08:00
; KNL-NEXT: retq
;
; SKX-LABEL: test_extractelement_varible_v16i1:
[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
; SKX: ## %bb.0:
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; SKX-NEXT: ## kill: def $edi killed $edi def $rdi
[AVX512] Fix EXTRACT_VECTOR_ELT for v2i1/v4i1/v32i1/v64i1 with variable index. Differential Revision: https://reviews.llvm.org/D30189 llvm-svn: 295718
2017-02-21 22:01:25 +08:00
; SKX-NEXT: vpcmpnleud %zmm1, %zmm0, %k0
[X86] When lowering insert_vector_elt/extract_vector_elt of vXi1 with a non-constant index just use either a 128-bit type or the vXi8 type with the correct number of elements. Despite what the comment said there isn't better codegen for 512-bit vectors. The 128/256/512 bit implementation jus stores to memory and loads an element. There's no advantage to doing that with a larger size. In fact in many cases it causes a stack realignment and generates worse code. llvm-svn: 321369
2017-12-23 01:18:11 +08:00
; SKX-NEXT: vpmovm2b %k0, %xmm0
; SKX-NEXT: vmovdqa %xmm0, -{{[0-9]+}}(%rsp)
[AVX512] Fix EXTRACT_VECTOR_ELT for v2i1/v4i1/v32i1/v64i1 with variable index. Differential Revision: https://reviews.llvm.org/D30189 llvm-svn: 295718
2017-02-21 22:01:25 +08:00
; SKX-NEXT: andl $15, %edi
[SelectionDAG] Reverse the order of operands in the ISD::ADD created by TargetLowering::getVectorElementPointer so that the FrameIndex is on the left. This seems to improve X86's ability to match this into an address computation. Otherwise the other operand gets assigned to the base register and the stack pointer + frame index ends up in the index register. But index registers can't encode ESP/RSP so we end up having to move it into another register to meet the constraint. I could try to improve the address matcher in X86, but swapping the producer seemed easier. Several other places already have the operands in this order so this is at least consistent. llvm-svn: 321370
2017-12-23 01:18:13 +08:00
; SKX-NEXT: movzbl -24(%rsp,%rdi), %eax
[AVX512] Fix EXTRACT_VECTOR_ELT for v2i1/v4i1/v32i1/v64i1 with variable index. Differential Revision: https://reviews.llvm.org/D30189 llvm-svn: 295718
2017-02-21 22:01:25 +08:00
; SKX-NEXT: andl $1, %eax
[X86] Generate VZEROUPPER for Skylake-avx512. VZEROUPPER should not be issued on Knights Landing (KNL), but on Skylake-avx512 it should be. Differential Revision: https://reviews.llvm.org/D29874 llvm-svn: 296859
2017-03-03 17:03:24 +08:00
; SKX-NEXT: vzeroupper
[AVX512] Fix EXTRACT_VECTOR_ELT for v2i1/v4i1/v32i1/v64i1 with variable index. Differential Revision: https://reviews.llvm.org/D30189 llvm-svn: 295718
2017-02-21 22:01:25 +08:00
; SKX-NEXT: retq
%t1 = icmp ugt <16 x i32> %a, %b
%t2 = extractelement <16 x i1> %t1, i32 %index
%res = zext i1 %t2 to i8
ret i8 %res
}
define zeroext i8 @test_extractelement_varible_v32i1(<32 x i8> %a, <32 x i8> %b, i32 %index) {
; KNL-LABEL: test_extractelement_varible_v32i1:
[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
; KNL: ## %bb.0:
[AVX512] Fix EXTRACT_VECTOR_ELT for v2i1/v4i1/v32i1/v64i1 with variable index. Differential Revision: https://reviews.llvm.org/D30189 llvm-svn: 295718
2017-02-21 22:01:25 +08:00
; KNL-NEXT: pushq %rbp
; KNL-NEXT: .cfi_def_cfa_offset 16
; KNL-NEXT: .cfi_offset %rbp, -16
; KNL-NEXT: movq %rsp, %rbp
; KNL-NEXT: .cfi_def_cfa_register %rbp
; KNL-NEXT: andq $-32, %rsp
; KNL-NEXT: subq $64, %rsp
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; KNL-NEXT: ## kill: def $edi killed $edi def $rdi
[X86] Use min/max for vector ult/ugt compares if avoids a sign flip. Summary: Currently we only use min/max to help with ule/uge compares because it removes an invert of the result that would otherwise be needed. But we can also use it for ult/ugt compares if it will prevent the need for a sign bit flip needed to use pcmpgt at the cost of requiring an invert after the compare. I also refactored the code so that the max/min code is self contained and does its own return instead of setting up a flag to manipulate the rest of the function's behavior. Most of the test cases look ok with this. I did notice that we added instructions when one of the operands being sign flipped is a constant vector that we were able to constant fold the flip into. I also noticed that sometimes the SSE min/max clobbers a register that is needed after the compare. This resulted in an extra move being inserted before the min/max to preserve the register. We could try to detect this and switch from min to max and change the compare operands to use the operand that gets reused in the compare. Reviewers: spatel, RKSimon Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D42935 llvm-svn: 324842
2018-02-12 01:11:40 +08:00
; KNL-NEXT: vpminub %ymm1, %ymm0, %ymm1
; KNL-NEXT: vpcmpeqb %ymm1, %ymm0, %ymm0
; KNL-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
[AVX512] Fix EXTRACT_VECTOR_ELT for v2i1/v4i1/v32i1/v64i1 with variable index. Differential Revision: https://reviews.llvm.org/D30189 llvm-svn: 295718
2017-02-21 22:01:25 +08:00
; KNL-NEXT: vmovdqa %ymm0, (%rsp)
; KNL-NEXT: andl $31, %edi
[SelectionDAG] Reverse the order of operands in the ISD::ADD created by TargetLowering::getVectorElementPointer so that the FrameIndex is on the left. This seems to improve X86's ability to match this into an address computation. Otherwise the other operand gets assigned to the base register and the stack pointer + frame index ends up in the index register. But index registers can't encode ESP/RSP so we end up having to move it into another register to meet the constraint. I could try to improve the address matcher in X86, but swapping the producer seemed easier. Several other places already have the operands in this order so this is at least consistent. llvm-svn: 321370
2017-12-23 01:18:13 +08:00
; KNL-NEXT: movzbl (%rsp,%rdi), %eax
[X86][AVX512] Make i1 illegal in the CodeGen This patch defines the i1 type as illegal in the X86 backend for AVX512. For DAG operations on <N x i1> types (build vector, extract vector element, ...) i8 is used, and should be truncated/extended. This should produce better scalar code for i1 types since GPRs will be used instead of mask registers. Differential Revision: https://reviews.llvm.org/D32273 llvm-svn: 303421
2017-05-19 20:35:15 +08:00
; KNL-NEXT: andl $1, %eax
[AVX512] Fix EXTRACT_VECTOR_ELT for v2i1/v4i1/v32i1/v64i1 with variable index. Differential Revision: https://reviews.llvm.org/D30189 llvm-svn: 295718
2017-02-21 22:01:25 +08:00
; KNL-NEXT: movq %rbp, %rsp
; KNL-NEXT: popq %rbp
[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
; KNL-NEXT: vzeroupper
[AVX512] Fix EXTRACT_VECTOR_ELT for v2i1/v4i1/v32i1/v64i1 with variable index. Differential Revision: https://reviews.llvm.org/D30189 llvm-svn: 295718
2017-02-21 22:01:25 +08:00
; KNL-NEXT: retq
;
; SKX-LABEL: test_extractelement_varible_v32i1:
[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
; SKX: ## %bb.0:
[AVX512] Fix EXTRACT_VECTOR_ELT for v2i1/v4i1/v32i1/v64i1 with variable index. Differential Revision: https://reviews.llvm.org/D30189 llvm-svn: 295718
2017-02-21 22:01:25 +08:00
; SKX-NEXT: pushq %rbp
; SKX-NEXT: .cfi_def_cfa_offset 16
; SKX-NEXT: .cfi_offset %rbp, -16
; SKX-NEXT: movq %rsp, %rbp
; SKX-NEXT: .cfi_def_cfa_register %rbp
[X86] When lowering insert_vector_elt/extract_vector_elt of vXi1 with a non-constant index just use either a 128-bit type or the vXi8 type with the correct number of elements. Despite what the comment said there isn't better codegen for 512-bit vectors. The 128/256/512 bit implementation jus stores to memory and loads an element. There's no advantage to doing that with a larger size. In fact in many cases it causes a stack realignment and generates worse code. llvm-svn: 321369
2017-12-23 01:18:11 +08:00
; SKX-NEXT: andq $-32, %rsp
; SKX-NEXT: subq $64, %rsp
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; SKX-NEXT: ## kill: def $edi killed $edi def $rdi
[AVX512] Fix EXTRACT_VECTOR_ELT for v2i1/v4i1/v32i1/v64i1 with variable index. Differential Revision: https://reviews.llvm.org/D30189 llvm-svn: 295718
2017-02-21 22:01:25 +08:00
; SKX-NEXT: vpcmpnleub %ymm1, %ymm0, %k0
[X86] When lowering insert_vector_elt/extract_vector_elt of vXi1 with a non-constant index just use either a 128-bit type or the vXi8 type with the correct number of elements. Despite what the comment said there isn't better codegen for 512-bit vectors. The 128/256/512 bit implementation jus stores to memory and loads an element. There's no advantage to doing that with a larger size. In fact in many cases it causes a stack realignment and generates worse code. llvm-svn: 321369
2017-12-23 01:18:11 +08:00
; SKX-NEXT: vpmovm2b %k0, %ymm0
; SKX-NEXT: vmovdqa %ymm0, (%rsp)
[AVX512] Fix EXTRACT_VECTOR_ELT for v2i1/v4i1/v32i1/v64i1 with variable index. Differential Revision: https://reviews.llvm.org/D30189 llvm-svn: 295718
2017-02-21 22:01:25 +08:00
; SKX-NEXT: andl $31, %edi
[SelectionDAG] Reverse the order of operands in the ISD::ADD created by TargetLowering::getVectorElementPointer so that the FrameIndex is on the left. This seems to improve X86's ability to match this into an address computation. Otherwise the other operand gets assigned to the base register and the stack pointer + frame index ends up in the index register. But index registers can't encode ESP/RSP so we end up having to move it into another register to meet the constraint. I could try to improve the address matcher in X86, but swapping the producer seemed easier. Several other places already have the operands in this order so this is at least consistent. llvm-svn: 321370
2017-12-23 01:18:13 +08:00
; SKX-NEXT: movzbl (%rsp,%rdi), %eax
[AVX512] Fix EXTRACT_VECTOR_ELT for v2i1/v4i1/v32i1/v64i1 with variable index. Differential Revision: https://reviews.llvm.org/D30189 llvm-svn: 295718
2017-02-21 22:01:25 +08:00
; SKX-NEXT: andl $1, %eax
; SKX-NEXT: movq %rbp, %rsp
; SKX-NEXT: popq %rbp
[X86] Generate VZEROUPPER for Skylake-avx512. VZEROUPPER should not be issued on Knights Landing (KNL), but on Skylake-avx512 it should be. Differential Revision: https://reviews.llvm.org/D29874 llvm-svn: 296859
2017-03-03 17:03:24 +08:00
; SKX-NEXT: vzeroupper
[AVX512] Fix EXTRACT_VECTOR_ELT for v2i1/v4i1/v32i1/v64i1 with variable index. Differential Revision: https://reviews.llvm.org/D30189 llvm-svn: 295718
2017-02-21 22:01:25 +08:00
; SKX-NEXT: retq
%t1 = icmp ugt <32 x i8> %a, %b
%t2 = extractelement <32 x i1> %t1, i32 %index
%res = zext i1 %t2 to i8
ret i8 %res
}
[X86] Use correct subvector index when combining two insert subvectors featuring zero vectors. Previously we were using one of the subvector indices twice. The included test case causes an assert without this change. Thanks to Simon Pilgrim for catching this. llvm-svn: 314429
2017-09-29 00:53:16 +08:00
define <8 x i64> @insert_double_zero(<2 x i64> %a) nounwind {
; CHECK-LABEL: insert_double_zero:
[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] Use correct subvector index when combining two insert subvectors featuring zero vectors. Previously we were using one of the subvector indices twice. The included test case causes an assert without this change. Thanks to Simon Pilgrim for catching this. llvm-svn: 314429
2017-09-29 00:53:16 +08:00
; CHECK-NEXT: vxorps %xmm1, %xmm1, %xmm1
; CHECK-NEXT: vinsertf32x4 $2, %xmm0, %zmm1, %zmm0
; CHECK-NEXT: retq
%b = shufflevector <2 x i64> %a, <2 x i64> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
%d = shufflevector <4 x i64> %b, <4 x i64> undef, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
%e = shufflevector <8 x i64> %d, <8 x i64> zeroinitializer, <8 x i32> <i32 8, i32 9, i32 10, i32 11, i32 0, i32 1, i32 2, i32 3>
ret <8 x i64> %e
}
[SelectionDAG][X86] Fix insert_vector_elt lowering for v32i1/v64i1 with non-constant index Summary: Currently we don't handle v32i1/v64i1 insert_vector_elt correctly as we fail to look at the number of elements closely and assume it can only be v16i1 or v8i1. We also can't type legalize v64i1 insert_vector_elt correctly on KNL due to the type not being byte addressable as required by the legalizing through memory accesses path requires. For the first issue, the patch now tries to pick a 512-bit register with the correct number of elements and promotes to that. For the second issue, we now extend the vector to a byte addressable type, do the stores to memory, load the two halves, and then truncate the halves back to the original type. Technically since we changed the type, we may not need two loads, but actually checking that is more work and for the v64i1 case we do need them. Reviewers: RKSimon, delena, spatel, zvi Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D40942 llvm-svn: 320849
2017-12-16 03:35:22 +08:00
define i32 @test_insertelement_variable_v32i1(<32 x i8> %a, i8 %b, i32 %index) {
; KNL-LABEL: test_insertelement_variable_v32i1:
; KNL: ## %bb.0:
; KNL-NEXT: pushq %rbp
; KNL-NEXT: .cfi_def_cfa_offset 16
; KNL-NEXT: .cfi_offset %rbp, -16
; KNL-NEXT: movq %rsp, %rbp
; KNL-NEXT: .cfi_def_cfa_register %rbp
; KNL-NEXT: andq $-32, %rsp
[X86] Legalize v32i1 without BWI via splitting to v16i1 rather than the default of promoting to v32i8. Summary: For the most part its better to keep v32i1 as a mask type of a narrower width than trying to promote it to a ymm register. I had to add some overrides to the methods that get the types for the calling convention so that we still use v32i8 for argument/return purposes. There are still some regressions in here. I definitely saw some around shuffles. I think we probably should move vXi1 shuffle from lowering to a DAG combine where I think the extend and truncate we have to emit would be better combined. I think we also need a DAG combine to remove trunc from (extract_vector_elt (trunc)) Overall this removes something like 13000 CHECK lines from lit tests. Reviewers: zvi, RKSimon, delena, spatel Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D42031 llvm-svn: 323201
2018-01-23 22:25:39 +08:00
; KNL-NEXT: subq $64, %rsp
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; KNL-NEXT: ## kill: def $esi killed $esi def $rsi
[X86] Use min/max for vector ult/ugt compares if avoids a sign flip. Summary: Currently we only use min/max to help with ule/uge compares because it removes an invert of the result that would otherwise be needed. But we can also use it for ult/ugt compares if it will prevent the need for a sign bit flip needed to use pcmpgt at the cost of requiring an invert after the compare. I also refactored the code so that the max/min code is self contained and does its own return instead of setting up a flag to manipulate the rest of the function's behavior. Most of the test cases look ok with this. I did notice that we added instructions when one of the operands being sign flipped is a constant vector that we were able to constant fold the flip into. I also noticed that sometimes the SSE min/max clobbers a register that is needed after the compare. This resulted in an extra move being inserted before the min/max to preserve the register. We could try to detect this and switch from min to max and change the compare operands to use the operand that gets reused in the compare. Reviewers: spatel, RKSimon Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D42935 llvm-svn: 324842
2018-02-12 01:11:40 +08:00
; KNL-NEXT: vpxor %xmm1, %xmm1, %xmm1
; KNL-NEXT: vpcmpeqb %ymm1, %ymm0, %ymm0
; KNL-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
[SelectionDAG][X86] Fix insert_vector_elt lowering for v32i1/v64i1 with non-constant index Summary: Currently we don't handle v32i1/v64i1 insert_vector_elt correctly as we fail to look at the number of elements closely and assume it can only be v16i1 or v8i1. We also can't type legalize v64i1 insert_vector_elt correctly on KNL due to the type not being byte addressable as required by the legalizing through memory accesses path requires. For the first issue, the patch now tries to pick a 512-bit register with the correct number of elements and promotes to that. For the second issue, we now extend the vector to a byte addressable type, do the stores to memory, load the two halves, and then truncate the halves back to the original type. Technically since we changed the type, we may not need two loads, but actually checking that is more work and for the v64i1 case we do need them. Reviewers: RKSimon, delena, spatel, zvi Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D40942 llvm-svn: 320849
2017-12-16 03:35:22 +08:00
; KNL-NEXT: andl $31, %esi
; KNL-NEXT: testb %dil, %dil
[X86] Legalize v32i1 without BWI via splitting to v16i1 rather than the default of promoting to v32i8. Summary: For the most part its better to keep v32i1 as a mask type of a narrower width than trying to promote it to a ymm register. I had to add some overrides to the methods that get the types for the calling convention so that we still use v32i8 for argument/return purposes. There are still some regressions in here. I definitely saw some around shuffles. I think we probably should move vXi1 shuffle from lowering to a DAG combine where I think the extend and truncate we have to emit would be better combined. I think we also need a DAG combine to remove trunc from (extract_vector_elt (trunc)) Overall this removes something like 13000 CHECK lines from lit tests. Reviewers: zvi, RKSimon, delena, spatel Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D42031 llvm-svn: 323201
2018-01-23 22:25:39 +08:00
; KNL-NEXT: vmovdqa %ymm0, (%rsp)
; KNL-NEXT: setne (%rsp,%rsi)
[X86] Turn X86ISD::VSEXT into X86ISD::VZEXT if the upper bits aren't demanded. This makes X86ISD::VSEXT more similar to ISD::SIGN_EXTEND and ISD::ZERO_EXTEND. I'm hoping to replace X86ISD::VSEXT/VZEXT with target independent nodes. Making the target specific nodes similar to the target independent nodes helps minimize test diffs in that patch. llvm-svn: 346539
2018-11-10 03:05:51 +08:00
; KNL-NEXT: vpmovzxbd {{.*#+}} zmm0 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero,mem[4],zero,zero,zero,mem[5],zero,zero,zero,mem[6],zero,zero,zero,mem[7],zero,zero,zero,mem[8],zero,zero,zero,mem[9],zero,zero,zero,mem[10],zero,zero,zero,mem[11],zero,zero,zero,mem[12],zero,zero,zero,mem[13],zero,zero,zero,mem[14],zero,zero,zero,mem[15],zero,zero,zero
[SelectionDAG][X86] Fix insert_vector_elt lowering for v32i1/v64i1 with non-constant index Summary: Currently we don't handle v32i1/v64i1 insert_vector_elt correctly as we fail to look at the number of elements closely and assume it can only be v16i1 or v8i1. We also can't type legalize v64i1 insert_vector_elt correctly on KNL due to the type not being byte addressable as required by the legalizing through memory accesses path requires. For the first issue, the patch now tries to pick a 512-bit register with the correct number of elements and promotes to that. For the second issue, we now extend the vector to a byte addressable type, do the stores to memory, load the two halves, and then truncate the halves back to the original type. Technically since we changed the type, we may not need two loads, but actually checking that is more work and for the v64i1 case we do need them. Reviewers: RKSimon, delena, spatel, zvi Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D40942 llvm-svn: 320849
2017-12-16 03:35:22 +08:00
; KNL-NEXT: vpslld $31, %zmm0, %zmm0
; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
[X86] Legalize v32i1 without BWI via splitting to v16i1 rather than the default of promoting to v32i8. Summary: For the most part its better to keep v32i1 as a mask type of a narrower width than trying to promote it to a ymm register. I had to add some overrides to the methods that get the types for the calling convention so that we still use v32i8 for argument/return purposes. There are still some regressions in here. I definitely saw some around shuffles. I think we probably should move vXi1 shuffle from lowering to a DAG combine where I think the extend and truncate we have to emit would be better combined. I think we also need a DAG combine to remove trunc from (extract_vector_elt (trunc)) Overall this removes something like 13000 CHECK lines from lit tests. Reviewers: zvi, RKSimon, delena, spatel Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D42031 llvm-svn: 323201
2018-01-23 22:25:39 +08:00
; KNL-NEXT: kmovw %k0, %ecx
[X86] Turn X86ISD::VSEXT into X86ISD::VZEXT if the upper bits aren't demanded. This makes X86ISD::VSEXT more similar to ISD::SIGN_EXTEND and ISD::ZERO_EXTEND. I'm hoping to replace X86ISD::VSEXT/VZEXT with target independent nodes. Making the target specific nodes similar to the target independent nodes helps minimize test diffs in that patch. llvm-svn: 346539
2018-11-10 03:05:51 +08:00
; KNL-NEXT: vpmovzxbd {{.*#+}} zmm0 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero,mem[4],zero,zero,zero,mem[5],zero,zero,zero,mem[6],zero,zero,zero,mem[7],zero,zero,zero,mem[8],zero,zero,zero,mem[9],zero,zero,zero,mem[10],zero,zero,zero,mem[11],zero,zero,zero,mem[12],zero,zero,zero,mem[13],zero,zero,zero,mem[14],zero,zero,zero,mem[15],zero,zero,zero
[X86] Legalize v32i1 without BWI via splitting to v16i1 rather than the default of promoting to v32i8. Summary: For the most part its better to keep v32i1 as a mask type of a narrower width than trying to promote it to a ymm register. I had to add some overrides to the methods that get the types for the calling convention so that we still use v32i8 for argument/return purposes. There are still some regressions in here. I definitely saw some around shuffles. I think we probably should move vXi1 shuffle from lowering to a DAG combine where I think the extend and truncate we have to emit would be better combined. I think we also need a DAG combine to remove trunc from (extract_vector_elt (trunc)) Overall this removes something like 13000 CHECK lines from lit tests. Reviewers: zvi, RKSimon, delena, spatel Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D42031 llvm-svn: 323201
2018-01-23 22:25:39 +08:00
; KNL-NEXT: vpslld $31, %zmm0, %zmm0
; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
; KNL-NEXT: kmovw %k0, %eax
; KNL-NEXT: shll $16, %eax
; KNL-NEXT: orl %ecx, %eax
[SelectionDAG][X86] Fix insert_vector_elt lowering for v32i1/v64i1 with non-constant index Summary: Currently we don't handle v32i1/v64i1 insert_vector_elt correctly as we fail to look at the number of elements closely and assume it can only be v16i1 or v8i1. We also can't type legalize v64i1 insert_vector_elt correctly on KNL due to the type not being byte addressable as required by the legalizing through memory accesses path requires. For the first issue, the patch now tries to pick a 512-bit register with the correct number of elements and promotes to that. For the second issue, we now extend the vector to a byte addressable type, do the stores to memory, load the two halves, and then truncate the halves back to the original type. Technically since we changed the type, we may not need two loads, but actually checking that is more work and for the v64i1 case we do need them. Reviewers: RKSimon, delena, spatel, zvi Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D40942 llvm-svn: 320849
2017-12-16 03:35:22 +08:00
; KNL-NEXT: movq %rbp, %rsp
; KNL-NEXT: popq %rbp
; KNL-NEXT: vzeroupper
; KNL-NEXT: retq
;
; SKX-LABEL: test_insertelement_variable_v32i1:
; SKX: ## %bb.0:
; SKX-NEXT: pushq %rbp
; SKX-NEXT: .cfi_def_cfa_offset 16
; SKX-NEXT: .cfi_offset %rbp, -16
; SKX-NEXT: movq %rsp, %rbp
; SKX-NEXT: .cfi_def_cfa_register %rbp
[X86] When lowering insert_vector_elt/extract_vector_elt of vXi1 with a non-constant index just use either a 128-bit type or the vXi8 type with the correct number of elements. Despite what the comment said there isn't better codegen for 512-bit vectors. The 128/256/512 bit implementation jus stores to memory and loads an element. There's no advantage to doing that with a larger size. In fact in many cases it causes a stack realignment and generates worse code. llvm-svn: 321369
2017-12-23 01:18:11 +08:00
; SKX-NEXT: andq $-32, %rsp
; SKX-NEXT: subq $64, %rsp
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; SKX-NEXT: ## kill: def $esi killed $esi def $rsi
[SelectionDAG] Support some SimplifySetCC cases for comparing against vector splats of constants. This supports things like (setcc ugt X, 0) -> (setcc ne X, 0) I've restricted to only make changes to vectors before legalize ops because I doubt all targets have accurate condition code legality information for vectors given how little we did before. Differential Revision: https://reviews.llvm.org/D42948 llvm-svn: 326495
2018-03-02 06:15:39 +08:00
; SKX-NEXT: vptestmb %ymm0, %ymm0, %k0
[SelectionDAG][X86] Fix insert_vector_elt lowering for v32i1/v64i1 with non-constant index Summary: Currently we don't handle v32i1/v64i1 insert_vector_elt correctly as we fail to look at the number of elements closely and assume it can only be v16i1 or v8i1. We also can't type legalize v64i1 insert_vector_elt correctly on KNL due to the type not being byte addressable as required by the legalizing through memory accesses path requires. For the first issue, the patch now tries to pick a 512-bit register with the correct number of elements and promotes to that. For the second issue, we now extend the vector to a byte addressable type, do the stores to memory, load the two halves, and then truncate the halves back to the original type. Technically since we changed the type, we may not need two loads, but actually checking that is more work and for the v64i1 case we do need them. Reviewers: RKSimon, delena, spatel, zvi Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D40942 llvm-svn: 320849
2017-12-16 03:35:22 +08:00
; SKX-NEXT: andl $31, %esi
[X86] When lowering insert_vector_elt/extract_vector_elt of vXi1 with a non-constant index just use either a 128-bit type or the vXi8 type with the correct number of elements. Despite what the comment said there isn't better codegen for 512-bit vectors. The 128/256/512 bit implementation jus stores to memory and loads an element. There's no advantage to doing that with a larger size. In fact in many cases it causes a stack realignment and generates worse code. llvm-svn: 321369
2017-12-23 01:18:11 +08:00
; SKX-NEXT: testb %dil, %dil
; SKX-NEXT: vpmovm2b %k0, %ymm0
; SKX-NEXT: vmovdqa %ymm0, (%rsp)
[SelectionDAG] Reverse the order of operands in the ISD::ADD created by TargetLowering::getVectorElementPointer so that the FrameIndex is on the left. This seems to improve X86's ability to match this into an address computation. Otherwise the other operand gets assigned to the base register and the stack pointer + frame index ends up in the index register. But index registers can't encode ESP/RSP so we end up having to move it into another register to meet the constraint. I could try to improve the address matcher in X86, but swapping the producer seemed easier. Several other places already have the operands in this order so this is at least consistent. llvm-svn: 321370
2017-12-23 01:18:13 +08:00
; SKX-NEXT: setne (%rsp,%rsi)
[X86] When lowering insert_vector_elt/extract_vector_elt of vXi1 with a non-constant index just use either a 128-bit type or the vXi8 type with the correct number of elements. Despite what the comment said there isn't better codegen for 512-bit vectors. The 128/256/512 bit implementation jus stores to memory and loads an element. There's no advantage to doing that with a larger size. In fact in many cases it causes a stack realignment and generates worse code. llvm-svn: 321369
2017-12-23 01:18:11 +08:00
; SKX-NEXT: vpsllw $7, (%rsp), %ymm0
; SKX-NEXT: vpmovb2m %ymm0, %k0
[SelectionDAG][X86] Fix insert_vector_elt lowering for v32i1/v64i1 with non-constant index Summary: Currently we don't handle v32i1/v64i1 insert_vector_elt correctly as we fail to look at the number of elements closely and assume it can only be v16i1 or v8i1. We also can't type legalize v64i1 insert_vector_elt correctly on KNL due to the type not being byte addressable as required by the legalizing through memory accesses path requires. For the first issue, the patch now tries to pick a 512-bit register with the correct number of elements and promotes to that. For the second issue, we now extend the vector to a byte addressable type, do the stores to memory, load the two halves, and then truncate the halves back to the original type. Technically since we changed the type, we may not need two loads, but actually checking that is more work and for the v64i1 case we do need them. Reviewers: RKSimon, delena, spatel, zvi Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D40942 llvm-svn: 320849
2017-12-16 03:35:22 +08:00
; SKX-NEXT: kmovd %k0, %eax
; SKX-NEXT: movq %rbp, %rsp
; SKX-NEXT: popq %rbp
; SKX-NEXT: vzeroupper
; SKX-NEXT: retq
%t1 = icmp ugt <32 x i8> %a, zeroinitializer
%t2 = icmp ugt i8 %b, 0
%t3 = insertelement <32 x i1> %t1, i1 %t2, i32 %index
%t4 = bitcast <32 x i1> %t3 to i32
ret i32 %t4
}
define i64 @test_insertelement_variable_v64i1(<64 x i8> %a, i8 %b, i32 %index) {
; KNL-LABEL: test_insertelement_variable_v64i1:
; KNL: ## %bb.0:
; KNL-NEXT: pushq %rbp
; KNL-NEXT: .cfi_def_cfa_offset 16
; KNL-NEXT: .cfi_offset %rbp, -16
; KNL-NEXT: movq %rsp, %rbp
; KNL-NEXT: .cfi_def_cfa_register %rbp
; KNL-NEXT: andq $-64, %rsp
[X86] Legalize v32i1 without BWI via splitting to v16i1 rather than the default of promoting to v32i8. Summary: For the most part its better to keep v32i1 as a mask type of a narrower width than trying to promote it to a ymm register. I had to add some overrides to the methods that get the types for the calling convention so that we still use v32i8 for argument/return purposes. There are still some regressions in here. I definitely saw some around shuffles. I think we probably should move vXi1 shuffle from lowering to a DAG combine where I think the extend and truncate we have to emit would be better combined. I think we also need a DAG combine to remove trunc from (extract_vector_elt (trunc)) Overall this removes something like 13000 CHECK lines from lit tests. Reviewers: zvi, RKSimon, delena, spatel Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D42031 llvm-svn: 323201
2018-01-23 22:25:39 +08:00
; KNL-NEXT: subq $128, %rsp
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; KNL-NEXT: ## kill: def $esi killed $esi def $rsi
[X86] Use min/max for vector ult/ugt compares if avoids a sign flip. Summary: Currently we only use min/max to help with ule/uge compares because it removes an invert of the result that would otherwise be needed. But we can also use it for ult/ugt compares if it will prevent the need for a sign bit flip needed to use pcmpgt at the cost of requiring an invert after the compare. I also refactored the code so that the max/min code is self contained and does its own return instead of setting up a flag to manipulate the rest of the function's behavior. Most of the test cases look ok with this. I did notice that we added instructions when one of the operands being sign flipped is a constant vector that we were able to constant fold the flip into. I also noticed that sometimes the SSE min/max clobbers a register that is needed after the compare. This resulted in an extra move being inserted before the min/max to preserve the register. We could try to detect this and switch from min to max and change the compare operands to use the operand that gets reused in the compare. Reviewers: spatel, RKSimon Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D42935 llvm-svn: 324842
2018-02-12 01:11:40 +08:00
; KNL-NEXT: vpxor %xmm2, %xmm2, %xmm2
[SelectionDAG] Support some SimplifySetCC cases for comparing against vector splats of constants. This supports things like (setcc ugt X, 0) -> (setcc ne X, 0) I've restricted to only make changes to vectors before legalize ops because I doubt all targets have accurate condition code legality information for vectors given how little we did before. Differential Revision: https://reviews.llvm.org/D42948 llvm-svn: 326495
2018-03-02 06:15:39 +08:00
; KNL-NEXT: vpcmpeqb %ymm2, %ymm0, %ymm0
[X86] Use min/max for vector ult/ugt compares if avoids a sign flip. Summary: Currently we only use min/max to help with ule/uge compares because it removes an invert of the result that would otherwise be needed. But we can also use it for ult/ugt compares if it will prevent the need for a sign bit flip needed to use pcmpgt at the cost of requiring an invert after the compare. I also refactored the code so that the max/min code is self contained and does its own return instead of setting up a flag to manipulate the rest of the function's behavior. Most of the test cases look ok with this. I did notice that we added instructions when one of the operands being sign flipped is a constant vector that we were able to constant fold the flip into. I also noticed that sometimes the SSE min/max clobbers a register that is needed after the compare. This resulted in an extra move being inserted before the min/max to preserve the register. We could try to detect this and switch from min to max and change the compare operands to use the operand that gets reused in the compare. Reviewers: spatel, RKSimon Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D42935 llvm-svn: 324842
2018-02-12 01:11:40 +08:00
; KNL-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
; KNL-NEXT: vpcmpeqb %ymm2, %ymm1, %ymm1
; KNL-NEXT: vpternlogq $15, %zmm1, %zmm1, %zmm1
[SelectionDAG][X86] Fix insert_vector_elt lowering for v32i1/v64i1 with non-constant index Summary: Currently we don't handle v32i1/v64i1 insert_vector_elt correctly as we fail to look at the number of elements closely and assume it can only be v16i1 or v8i1. We also can't type legalize v64i1 insert_vector_elt correctly on KNL due to the type not being byte addressable as required by the legalizing through memory accesses path requires. For the first issue, the patch now tries to pick a 512-bit register with the correct number of elements and promotes to that. For the second issue, we now extend the vector to a byte addressable type, do the stores to memory, load the two halves, and then truncate the halves back to the original type. Technically since we changed the type, we may not need two loads, but actually checking that is more work and for the v64i1 case we do need them. Reviewers: RKSimon, delena, spatel, zvi Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D40942 llvm-svn: 320849
2017-12-16 03:35:22 +08:00
; KNL-NEXT: andl $63, %esi
; KNL-NEXT: testb %dil, %dil
; KNL-NEXT: vmovdqa %ymm1, {{[0-9]+}}(%rsp)
[X86] Legalize v32i1 without BWI via splitting to v16i1 rather than the default of promoting to v32i8. Summary: For the most part its better to keep v32i1 as a mask type of a narrower width than trying to promote it to a ymm register. I had to add some overrides to the methods that get the types for the calling convention so that we still use v32i8 for argument/return purposes. There are still some regressions in here. I definitely saw some around shuffles. I think we probably should move vXi1 shuffle from lowering to a DAG combine where I think the extend and truncate we have to emit would be better combined. I think we also need a DAG combine to remove trunc from (extract_vector_elt (trunc)) Overall this removes something like 13000 CHECK lines from lit tests. Reviewers: zvi, RKSimon, delena, spatel Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D42031 llvm-svn: 323201
2018-01-23 22:25:39 +08:00
; KNL-NEXT: vmovdqa %ymm0, (%rsp)
; KNL-NEXT: setne (%rsp,%rsi)
[x86] allow vector load narrowing with multi-use values This is a long-awaited follow-up suggested in D33578. Since then, we've picked up even more opportunities for vector narrowing from changes like D53784, so there are a lot of test diffs. Apart from 2-3 strange cases, these are all wins. I've structured this to be no-functional-change-intended for any target except for x86 because I couldn't tell if AArch64, ARM, and AMDGPU would improve or not. All of those targets have existing regression tests (4, 4, 10 files respectively) that would be affected. Also, Hexagon overrides the shouldReduceLoadWidth() hook, but doesn't show any regression test diffs. The trade-off is deciding if an extra vector load is better than a single wide load + extract_subvector. For x86, this is almost always better (on paper at least) because we often can fold loads into subsequent ops and not increase the official instruction count. There's also some unknown -- but potentially large -- benefit from using narrower vector ops if wide ops are implemented with multiple uops and/or frequency throttling is avoided. Differential Revision: https://reviews.llvm.org/D54073 llvm-svn: 346595
2018-11-11 04:05:31 +08:00
; KNL-NEXT: vpmovzxbd {{.*#+}} zmm0 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero,mem[4],zero,zero,zero,mem[5],zero,zero,zero,mem[6],zero,zero,zero,mem[7],zero,zero,zero,mem[8],zero,zero,zero,mem[9],zero,zero,zero,mem[10],zero,zero,zero,mem[11],zero,zero,zero,mem[12],zero,zero,zero,mem[13],zero,zero,zero,mem[14],zero,zero,zero,mem[15],zero,zero,zero
; KNL-NEXT: vpslld $31, %zmm0, %zmm0
; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
[X86] Legalize v32i1 without BWI via splitting to v16i1 rather than the default of promoting to v32i8. Summary: For the most part its better to keep v32i1 as a mask type of a narrower width than trying to promote it to a ymm register. I had to add some overrides to the methods that get the types for the calling convention so that we still use v32i8 for argument/return purposes. There are still some regressions in here. I definitely saw some around shuffles. I think we probably should move vXi1 shuffle from lowering to a DAG combine where I think the extend and truncate we have to emit would be better combined. I think we also need a DAG combine to remove trunc from (extract_vector_elt (trunc)) Overall this removes something like 13000 CHECK lines from lit tests. Reviewers: zvi, RKSimon, delena, spatel Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D42031 llvm-svn: 323201
2018-01-23 22:25:39 +08:00
; KNL-NEXT: kmovw %k0, %eax
[x86] allow vector load narrowing with multi-use values This is a long-awaited follow-up suggested in D33578. Since then, we've picked up even more opportunities for vector narrowing from changes like D53784, so there are a lot of test diffs. Apart from 2-3 strange cases, these are all wins. I've structured this to be no-functional-change-intended for any target except for x86 because I couldn't tell if AArch64, ARM, and AMDGPU would improve or not. All of those targets have existing regression tests (4, 4, 10 files respectively) that would be affected. Also, Hexagon overrides the shouldReduceLoadWidth() hook, but doesn't show any regression test diffs. The trade-off is deciding if an extra vector load is better than a single wide load + extract_subvector. For x86, this is almost always better (on paper at least) because we often can fold loads into subsequent ops and not increase the official instruction count. There's also some unknown -- but potentially large -- benefit from using narrower vector ops if wide ops are implemented with multiple uops and/or frequency throttling is avoided. Differential Revision: https://reviews.llvm.org/D54073 llvm-svn: 346595
2018-11-11 04:05:31 +08:00
; KNL-NEXT: vpmovzxbd {{.*#+}} zmm0 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero,mem[4],zero,zero,zero,mem[5],zero,zero,zero,mem[6],zero,zero,zero,mem[7],zero,zero,zero,mem[8],zero,zero,zero,mem[9],zero,zero,zero,mem[10],zero,zero,zero,mem[11],zero,zero,zero,mem[12],zero,zero,zero,mem[13],zero,zero,zero,mem[14],zero,zero,zero,mem[15],zero,zero,zero
[SelectionDAG][X86] Fix insert_vector_elt lowering for v32i1/v64i1 with non-constant index Summary: Currently we don't handle v32i1/v64i1 insert_vector_elt correctly as we fail to look at the number of elements closely and assume it can only be v16i1 or v8i1. We also can't type legalize v64i1 insert_vector_elt correctly on KNL due to the type not being byte addressable as required by the legalizing through memory accesses path requires. For the first issue, the patch now tries to pick a 512-bit register with the correct number of elements and promotes to that. For the second issue, we now extend the vector to a byte addressable type, do the stores to memory, load the two halves, and then truncate the halves back to the original type. Technically since we changed the type, we may not need two loads, but actually checking that is more work and for the v64i1 case we do need them. Reviewers: RKSimon, delena, spatel, zvi Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D40942 llvm-svn: 320849
2017-12-16 03:35:22 +08:00
; KNL-NEXT: vpslld $31, %zmm0, %zmm0
; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
[X86] Legalize v32i1 without BWI via splitting to v16i1 rather than the default of promoting to v32i8. Summary: For the most part its better to keep v32i1 as a mask type of a narrower width than trying to promote it to a ymm register. I had to add some overrides to the methods that get the types for the calling convention so that we still use v32i8 for argument/return purposes. There are still some regressions in here. I definitely saw some around shuffles. I think we probably should move vXi1 shuffle from lowering to a DAG combine where I think the extend and truncate we have to emit would be better combined. I think we also need a DAG combine to remove trunc from (extract_vector_elt (trunc)) Overall this removes something like 13000 CHECK lines from lit tests. Reviewers: zvi, RKSimon, delena, spatel Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D42031 llvm-svn: 323201
2018-01-23 22:25:39 +08:00
; KNL-NEXT: kmovw %k0, %ecx
; KNL-NEXT: shll $16, %ecx
; KNL-NEXT: orl %eax, %ecx
[x86] allow vector load narrowing with multi-use values This is a long-awaited follow-up suggested in D33578. Since then, we've picked up even more opportunities for vector narrowing from changes like D53784, so there are a lot of test diffs. Apart from 2-3 strange cases, these are all wins. I've structured this to be no-functional-change-intended for any target except for x86 because I couldn't tell if AArch64, ARM, and AMDGPU would improve or not. All of those targets have existing regression tests (4, 4, 10 files respectively) that would be affected. Also, Hexagon overrides the shouldReduceLoadWidth() hook, but doesn't show any regression test diffs. The trade-off is deciding if an extra vector load is better than a single wide load + extract_subvector. For x86, this is almost always better (on paper at least) because we often can fold loads into subsequent ops and not increase the official instruction count. There's also some unknown -- but potentially large -- benefit from using narrower vector ops if wide ops are implemented with multiple uops and/or frequency throttling is avoided. Differential Revision: https://reviews.llvm.org/D54073 llvm-svn: 346595
2018-11-11 04:05:31 +08:00
; KNL-NEXT: vpmovzxbd {{.*#+}} zmm0 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero,mem[4],zero,zero,zero,mem[5],zero,zero,zero,mem[6],zero,zero,zero,mem[7],zero,zero,zero,mem[8],zero,zero,zero,mem[9],zero,zero,zero,mem[10],zero,zero,zero,mem[11],zero,zero,zero,mem[12],zero,zero,zero,mem[13],zero,zero,zero,mem[14],zero,zero,zero,mem[15],zero,zero,zero
[SelectionDAG][X86] Fix insert_vector_elt lowering for v32i1/v64i1 with non-constant index Summary: Currently we don't handle v32i1/v64i1 insert_vector_elt correctly as we fail to look at the number of elements closely and assume it can only be v16i1 or v8i1. We also can't type legalize v64i1 insert_vector_elt correctly on KNL due to the type not being byte addressable as required by the legalizing through memory accesses path requires. For the first issue, the patch now tries to pick a 512-bit register with the correct number of elements and promotes to that. For the second issue, we now extend the vector to a byte addressable type, do the stores to memory, load the two halves, and then truncate the halves back to the original type. Technically since we changed the type, we may not need two loads, but actually checking that is more work and for the v64i1 case we do need them. Reviewers: RKSimon, delena, spatel, zvi Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D40942 llvm-svn: 320849
2017-12-16 03:35:22 +08:00
; KNL-NEXT: vpslld $31, %zmm0, %zmm0
; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
[X86] Legalize v32i1 without BWI via splitting to v16i1 rather than the default of promoting to v32i8. Summary: For the most part its better to keep v32i1 as a mask type of a narrower width than trying to promote it to a ymm register. I had to add some overrides to the methods that get the types for the calling convention so that we still use v32i8 for argument/return purposes. There are still some regressions in here. I definitely saw some around shuffles. I think we probably should move vXi1 shuffle from lowering to a DAG combine where I think the extend and truncate we have to emit would be better combined. I think we also need a DAG combine to remove trunc from (extract_vector_elt (trunc)) Overall this removes something like 13000 CHECK lines from lit tests. Reviewers: zvi, RKSimon, delena, spatel Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D42031 llvm-svn: 323201
2018-01-23 22:25:39 +08:00
; KNL-NEXT: kmovw %k0, %edx
[x86] allow vector load narrowing with multi-use values This is a long-awaited follow-up suggested in D33578. Since then, we've picked up even more opportunities for vector narrowing from changes like D53784, so there are a lot of test diffs. Apart from 2-3 strange cases, these are all wins. I've structured this to be no-functional-change-intended for any target except for x86 because I couldn't tell if AArch64, ARM, and AMDGPU would improve or not. All of those targets have existing regression tests (4, 4, 10 files respectively) that would be affected. Also, Hexagon overrides the shouldReduceLoadWidth() hook, but doesn't show any regression test diffs. The trade-off is deciding if an extra vector load is better than a single wide load + extract_subvector. For x86, this is almost always better (on paper at least) because we often can fold loads into subsequent ops and not increase the official instruction count. There's also some unknown -- but potentially large -- benefit from using narrower vector ops if wide ops are implemented with multiple uops and/or frequency throttling is avoided. Differential Revision: https://reviews.llvm.org/D54073 llvm-svn: 346595
2018-11-11 04:05:31 +08:00
; KNL-NEXT: vpmovzxbd {{.*#+}} zmm0 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero,mem[4],zero,zero,zero,mem[5],zero,zero,zero,mem[6],zero,zero,zero,mem[7],zero,zero,zero,mem[8],zero,zero,zero,mem[9],zero,zero,zero,mem[10],zero,zero,zero,mem[11],zero,zero,zero,mem[12],zero,zero,zero,mem[13],zero,zero,zero,mem[14],zero,zero,zero,mem[15],zero,zero,zero
[SelectionDAG][X86] Fix insert_vector_elt lowering for v32i1/v64i1 with non-constant index Summary: Currently we don't handle v32i1/v64i1 insert_vector_elt correctly as we fail to look at the number of elements closely and assume it can only be v16i1 or v8i1. We also can't type legalize v64i1 insert_vector_elt correctly on KNL due to the type not being byte addressable as required by the legalizing through memory accesses path requires. For the first issue, the patch now tries to pick a 512-bit register with the correct number of elements and promotes to that. For the second issue, we now extend the vector to a byte addressable type, do the stores to memory, load the two halves, and then truncate the halves back to the original type. Technically since we changed the type, we may not need two loads, but actually checking that is more work and for the v64i1 case we do need them. Reviewers: RKSimon, delena, spatel, zvi Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D40942 llvm-svn: 320849
2017-12-16 03:35:22 +08:00
; KNL-NEXT: vpslld $31, %zmm0, %zmm0
; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
[X86] Legalize v32i1 without BWI via splitting to v16i1 rather than the default of promoting to v32i8. Summary: For the most part its better to keep v32i1 as a mask type of a narrower width than trying to promote it to a ymm register. I had to add some overrides to the methods that get the types for the calling convention so that we still use v32i8 for argument/return purposes. There are still some regressions in here. I definitely saw some around shuffles. I think we probably should move vXi1 shuffle from lowering to a DAG combine where I think the extend and truncate we have to emit would be better combined. I think we also need a DAG combine to remove trunc from (extract_vector_elt (trunc)) Overall this removes something like 13000 CHECK lines from lit tests. Reviewers: zvi, RKSimon, delena, spatel Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D42031 llvm-svn: 323201
2018-01-23 22:25:39 +08:00
; KNL-NEXT: kmovw %k0, %eax
; KNL-NEXT: shll $16, %eax
; KNL-NEXT: orl %edx, %eax
[SelectionDAG][X86] Fix insert_vector_elt lowering for v32i1/v64i1 with non-constant index Summary: Currently we don't handle v32i1/v64i1 insert_vector_elt correctly as we fail to look at the number of elements closely and assume it can only be v16i1 or v8i1. We also can't type legalize v64i1 insert_vector_elt correctly on KNL due to the type not being byte addressable as required by the legalizing through memory accesses path requires. For the first issue, the patch now tries to pick a 512-bit register with the correct number of elements and promotes to that. For the second issue, we now extend the vector to a byte addressable type, do the stores to memory, load the two halves, and then truncate the halves back to the original type. Technically since we changed the type, we may not need two loads, but actually checking that is more work and for the v64i1 case we do need them. Reviewers: RKSimon, delena, spatel, zvi Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D40942 llvm-svn: 320849
2017-12-16 03:35:22 +08:00
; KNL-NEXT: shlq $32, %rax
; KNL-NEXT: orq %rcx, %rax
; KNL-NEXT: movq %rbp, %rsp
; KNL-NEXT: popq %rbp
; KNL-NEXT: vzeroupper
; KNL-NEXT: retq
;
; SKX-LABEL: test_insertelement_variable_v64i1:
; SKX: ## %bb.0:
; SKX-NEXT: pushq %rbp
; SKX-NEXT: .cfi_def_cfa_offset 16
; SKX-NEXT: .cfi_offset %rbp, -16
; SKX-NEXT: movq %rsp, %rbp
; SKX-NEXT: .cfi_def_cfa_register %rbp
; SKX-NEXT: andq $-64, %rsp
; SKX-NEXT: subq $128, %rsp
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; SKX-NEXT: ## kill: def $esi killed $esi def $rsi
[SelectionDAG] Support some SimplifySetCC cases for comparing against vector splats of constants. This supports things like (setcc ugt X, 0) -> (setcc ne X, 0) I've restricted to only make changes to vectors before legalize ops because I doubt all targets have accurate condition code legality information for vectors given how little we did before. Differential Revision: https://reviews.llvm.org/D42948 llvm-svn: 326495
2018-03-02 06:15:39 +08:00
; SKX-NEXT: vptestmb %zmm0, %zmm0, %k0
[SelectionDAG][X86] Fix insert_vector_elt lowering for v32i1/v64i1 with non-constant index Summary: Currently we don't handle v32i1/v64i1 insert_vector_elt correctly as we fail to look at the number of elements closely and assume it can only be v16i1 or v8i1. We also can't type legalize v64i1 insert_vector_elt correctly on KNL due to the type not being byte addressable as required by the legalizing through memory accesses path requires. For the first issue, the patch now tries to pick a 512-bit register with the correct number of elements and promotes to that. For the second issue, we now extend the vector to a byte addressable type, do the stores to memory, load the two halves, and then truncate the halves back to the original type. Technically since we changed the type, we may not need two loads, but actually checking that is more work and for the v64i1 case we do need them. Reviewers: RKSimon, delena, spatel, zvi Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D40942 llvm-svn: 320849
2017-12-16 03:35:22 +08:00
; SKX-NEXT: andl $63, %esi
; SKX-NEXT: testb %dil, %dil
[X86] Use SIGN_EXTEND rather than ZERO_EXTEND for lowering extract_vector_elt from vXi1 with a non-const index. We have a better range of instructions we can use if we can fill with the value i1 value rather than zeroing. llvm-svn: 321315
2017-12-22 06:08:23 +08:00
; SKX-NEXT: vpmovm2b %k0, %zmm0
[X86] Use vmovdqu64/vmovdqa64 for unmasked integer vector stores for consistency with loads. Previously we used 64 for vXi64 stores and 32 for everything else. This change uses 64 for everything just like do for loads. llvm-svn: 322820
2018-01-18 15:44:09 +08:00
; SKX-NEXT: vmovdqa64 %zmm0, (%rsp)
[SelectionDAG] Reverse the order of operands in the ISD::ADD created by TargetLowering::getVectorElementPointer so that the FrameIndex is on the left. This seems to improve X86's ability to match this into an address computation. Otherwise the other operand gets assigned to the base register and the stack pointer + frame index ends up in the index register. But index registers can't encode ESP/RSP so we end up having to move it into another register to meet the constraint. I could try to improve the address matcher in X86, but swapping the producer seemed easier. Several other places already have the operands in this order so this is at least consistent. llvm-svn: 321370
2017-12-23 01:18:13 +08:00
; SKX-NEXT: setne (%rsp,%rsi)
[SelectionDAG][X86] Fix insert_vector_elt lowering for v32i1/v64i1 with non-constant index Summary: Currently we don't handle v32i1/v64i1 insert_vector_elt correctly as we fail to look at the number of elements closely and assume it can only be v16i1 or v8i1. We also can't type legalize v64i1 insert_vector_elt correctly on KNL due to the type not being byte addressable as required by the legalizing through memory accesses path requires. For the first issue, the patch now tries to pick a 512-bit register with the correct number of elements and promotes to that. For the second issue, we now extend the vector to a byte addressable type, do the stores to memory, load the two halves, and then truncate the halves back to the original type. Technically since we changed the type, we may not need two loads, but actually checking that is more work and for the v64i1 case we do need them. Reviewers: RKSimon, delena, spatel, zvi Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D40942 llvm-svn: 320849
2017-12-16 03:35:22 +08:00
; SKX-NEXT: vpsllw $7, (%rsp), %zmm0
; SKX-NEXT: vpmovb2m %zmm0, %k0
; SKX-NEXT: kmovq %k0, %rax
; SKX-NEXT: movq %rbp, %rsp
; SKX-NEXT: popq %rbp
; SKX-NEXT: vzeroupper
; SKX-NEXT: retq
%t1 = icmp ugt <64 x i8> %a, zeroinitializer
%t2 = icmp ugt i8 %b, 0
%t3 = insertelement <64 x i1> %t1, i1 %t2, i32 %index
%t4 = bitcast <64 x i1> %t3 to i64
ret i64 %t4
}
define i96 @test_insertelement_variable_v96i1(<96 x i8> %a, i8 %b, i32 %index) {
; KNL-LABEL: test_insertelement_variable_v96i1:
; KNL: ## %bb.0:
; KNL-NEXT: pushq %rbp
; KNL-NEXT: .cfi_def_cfa_offset 16
; KNL-NEXT: .cfi_offset %rbp, -16
; KNL-NEXT: movq %rsp, %rbp
; KNL-NEXT: .cfi_def_cfa_register %rbp
; KNL-NEXT: andq $-128, %rsp
[X86] Legalize v32i1 without BWI via splitting to v16i1 rather than the default of promoting to v32i8. Summary: For the most part its better to keep v32i1 as a mask type of a narrower width than trying to promote it to a ymm register. I had to add some overrides to the methods that get the types for the calling convention so that we still use v32i8 for argument/return purposes. There are still some regressions in here. I definitely saw some around shuffles. I think we probably should move vXi1 shuffle from lowering to a DAG combine where I think the extend and truncate we have to emit would be better combined. I think we also need a DAG combine to remove trunc from (extract_vector_elt (trunc)) Overall this removes something like 13000 CHECK lines from lit tests. Reviewers: zvi, RKSimon, delena, spatel Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D42031 llvm-svn: 323201
2018-01-23 22:25:39 +08:00
; KNL-NEXT: subq $256, %rsp ## imm = 0x100
[SelectionDAG][X86] Fix insert_vector_elt lowering for v32i1/v64i1 with non-constant index Summary: Currently we don't handle v32i1/v64i1 insert_vector_elt correctly as we fail to look at the number of elements closely and assume it can only be v16i1 or v8i1. We also can't type legalize v64i1 insert_vector_elt correctly on KNL due to the type not being byte addressable as required by the legalizing through memory accesses path requires. For the first issue, the patch now tries to pick a 512-bit register with the correct number of elements and promotes to that. For the second issue, we now extend the vector to a byte addressable type, do the stores to memory, load the two halves, and then truncate the halves back to the original type. Technically since we changed the type, we may not need two loads, but actually checking that is more work and for the v64i1 case we do need them. Reviewers: RKSimon, delena, spatel, zvi Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D40942 llvm-svn: 320849
2017-12-16 03:35:22 +08:00
; KNL-NEXT: vmovd {{.*#+}} xmm0 = mem[0],zero,zero,zero
; KNL-NEXT: vpinsrb $1, 488(%rbp), %xmm0, %xmm0
; KNL-NEXT: vpinsrb $2, 496(%rbp), %xmm0, %xmm0
; KNL-NEXT: vpinsrb $3, 504(%rbp), %xmm0, %xmm0
; KNL-NEXT: vpinsrb $4, 512(%rbp), %xmm0, %xmm0
; KNL-NEXT: vpinsrb $5, 520(%rbp), %xmm0, %xmm0
; KNL-NEXT: vpinsrb $6, 528(%rbp), %xmm0, %xmm0
; KNL-NEXT: vpinsrb $7, 536(%rbp), %xmm0, %xmm0
; KNL-NEXT: vpinsrb $8, 544(%rbp), %xmm0, %xmm0
; KNL-NEXT: vpinsrb $9, 552(%rbp), %xmm0, %xmm0
; KNL-NEXT: vpinsrb $10, 560(%rbp), %xmm0, %xmm0
; KNL-NEXT: vpinsrb $11, 568(%rbp), %xmm0, %xmm0
; KNL-NEXT: vpinsrb $12, 576(%rbp), %xmm0, %xmm0
; KNL-NEXT: vpinsrb $13, 584(%rbp), %xmm0, %xmm0
; KNL-NEXT: vpinsrb $14, 592(%rbp), %xmm0, %xmm0
; KNL-NEXT: vpinsrb $15, 600(%rbp), %xmm0, %xmm0
; KNL-NEXT: vmovd {{.*#+}} xmm1 = mem[0],zero,zero,zero
; KNL-NEXT: vpinsrb $1, 616(%rbp), %xmm1, %xmm1
; KNL-NEXT: vpinsrb $2, 624(%rbp), %xmm1, %xmm1
; KNL-NEXT: vpinsrb $3, 632(%rbp), %xmm1, %xmm1
; KNL-NEXT: vpinsrb $4, 640(%rbp), %xmm1, %xmm1
; KNL-NEXT: vpinsrb $5, 648(%rbp), %xmm1, %xmm1
; KNL-NEXT: vpinsrb $6, 656(%rbp), %xmm1, %xmm1
; KNL-NEXT: vpinsrb $7, 664(%rbp), %xmm1, %xmm1
; KNL-NEXT: vpinsrb $8, 672(%rbp), %xmm1, %xmm1
; KNL-NEXT: vpinsrb $9, 680(%rbp), %xmm1, %xmm1
; KNL-NEXT: vpinsrb $10, 688(%rbp), %xmm1, %xmm1
; KNL-NEXT: vpinsrb $11, 696(%rbp), %xmm1, %xmm1
; KNL-NEXT: vpinsrb $12, 704(%rbp), %xmm1, %xmm1
; KNL-NEXT: vpinsrb $13, 712(%rbp), %xmm1, %xmm1
; KNL-NEXT: vpinsrb $14, 720(%rbp), %xmm1, %xmm1
; KNL-NEXT: vpinsrb $15, 728(%rbp), %xmm1, %xmm1
; KNL-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm0
; KNL-NEXT: vmovd {{.*#+}} xmm1 = mem[0],zero,zero,zero
; KNL-NEXT: vpinsrb $1, 232(%rbp), %xmm1, %xmm1
; KNL-NEXT: vpinsrb $2, 240(%rbp), %xmm1, %xmm1
; KNL-NEXT: vpinsrb $3, 248(%rbp), %xmm1, %xmm1
; KNL-NEXT: vpinsrb $4, 256(%rbp), %xmm1, %xmm1
; KNL-NEXT: vpinsrb $5, 264(%rbp), %xmm1, %xmm1
; KNL-NEXT: vpinsrb $6, 272(%rbp), %xmm1, %xmm1
; KNL-NEXT: vpinsrb $7, 280(%rbp), %xmm1, %xmm1
; KNL-NEXT: vpinsrb $8, 288(%rbp), %xmm1, %xmm1
; KNL-NEXT: vpinsrb $9, 296(%rbp), %xmm1, %xmm1
; KNL-NEXT: vpinsrb $10, 304(%rbp), %xmm1, %xmm1
; KNL-NEXT: vpinsrb $11, 312(%rbp), %xmm1, %xmm1
; KNL-NEXT: vpinsrb $12, 320(%rbp), %xmm1, %xmm1
; KNL-NEXT: vpinsrb $13, 328(%rbp), %xmm1, %xmm1
; KNL-NEXT: vpinsrb $14, 336(%rbp), %xmm1, %xmm1
; KNL-NEXT: vpinsrb $15, 344(%rbp), %xmm1, %xmm1
; KNL-NEXT: vmovd {{.*#+}} xmm2 = mem[0],zero,zero,zero
; KNL-NEXT: vpinsrb $1, 360(%rbp), %xmm2, %xmm2
; KNL-NEXT: vpinsrb $2, 368(%rbp), %xmm2, %xmm2
; KNL-NEXT: vpinsrb $3, 376(%rbp), %xmm2, %xmm2
; KNL-NEXT: vpinsrb $4, 384(%rbp), %xmm2, %xmm2
; KNL-NEXT: vpinsrb $5, 392(%rbp), %xmm2, %xmm2
; KNL-NEXT: vpinsrb $6, 400(%rbp), %xmm2, %xmm2
; KNL-NEXT: vpinsrb $7, 408(%rbp), %xmm2, %xmm2
; KNL-NEXT: vpinsrb $8, 416(%rbp), %xmm2, %xmm2
; KNL-NEXT: vpinsrb $9, 424(%rbp), %xmm2, %xmm2
; KNL-NEXT: vpinsrb $10, 432(%rbp), %xmm2, %xmm2
; KNL-NEXT: vpinsrb $11, 440(%rbp), %xmm2, %xmm2
; KNL-NEXT: vpinsrb $12, 448(%rbp), %xmm2, %xmm2
; KNL-NEXT: vpinsrb $13, 456(%rbp), %xmm2, %xmm2
; KNL-NEXT: vpinsrb $14, 464(%rbp), %xmm2, %xmm2
; KNL-NEXT: vpinsrb $15, 472(%rbp), %xmm2, %xmm2
; KNL-NEXT: vinserti128 $1, %xmm2, %ymm1, %ymm1
; KNL-NEXT: vmovd %edi, %xmm2
; KNL-NEXT: vpinsrb $1, %esi, %xmm2, %xmm2
; KNL-NEXT: vpinsrb $2, %edx, %xmm2, %xmm2
; KNL-NEXT: vpinsrb $3, %ecx, %xmm2, %xmm2
; KNL-NEXT: vpinsrb $4, %r8d, %xmm2, %xmm2
; KNL-NEXT: vpinsrb $5, %r9d, %xmm2, %xmm2
; KNL-NEXT: vpinsrb $6, 16(%rbp), %xmm2, %xmm2
; KNL-NEXT: vpinsrb $7, 24(%rbp), %xmm2, %xmm2
; KNL-NEXT: vpinsrb $8, 32(%rbp), %xmm2, %xmm2
; KNL-NEXT: vpinsrb $9, 40(%rbp), %xmm2, %xmm2
; KNL-NEXT: vpinsrb $10, 48(%rbp), %xmm2, %xmm2
; KNL-NEXT: vpinsrb $11, 56(%rbp), %xmm2, %xmm2
; KNL-NEXT: vpinsrb $12, 64(%rbp), %xmm2, %xmm2
; KNL-NEXT: vpinsrb $13, 72(%rbp), %xmm2, %xmm2
; KNL-NEXT: vpinsrb $14, 80(%rbp), %xmm2, %xmm2
; KNL-NEXT: vpinsrb $15, 88(%rbp), %xmm2, %xmm2
; KNL-NEXT: vmovd {{.*#+}} xmm3 = mem[0],zero,zero,zero
; KNL-NEXT: vpinsrb $1, 104(%rbp), %xmm3, %xmm3
; KNL-NEXT: vpinsrb $2, 112(%rbp), %xmm3, %xmm3
; KNL-NEXT: vpinsrb $3, 120(%rbp), %xmm3, %xmm3
; KNL-NEXT: vpinsrb $4, 128(%rbp), %xmm3, %xmm3
; KNL-NEXT: vpinsrb $5, 136(%rbp), %xmm3, %xmm3
; KNL-NEXT: vpinsrb $6, 144(%rbp), %xmm3, %xmm3
; KNL-NEXT: vpinsrb $7, 152(%rbp), %xmm3, %xmm3
; KNL-NEXT: vpinsrb $8, 160(%rbp), %xmm3, %xmm3
; KNL-NEXT: vpinsrb $9, 168(%rbp), %xmm3, %xmm3
; KNL-NEXT: vpinsrb $10, 176(%rbp), %xmm3, %xmm3
; KNL-NEXT: vpinsrb $11, 184(%rbp), %xmm3, %xmm3
; KNL-NEXT: vpinsrb $12, 192(%rbp), %xmm3, %xmm3
; KNL-NEXT: vpinsrb $13, 200(%rbp), %xmm3, %xmm3
; KNL-NEXT: vpinsrb $14, 208(%rbp), %xmm3, %xmm3
; KNL-NEXT: vpinsrb $15, 216(%rbp), %xmm3, %xmm3
; KNL-NEXT: vinserti128 $1, %xmm3, %ymm2, %ymm2
[DAGCombiner] Set the right SDLoc on a newly-created sextload (6/N) This teaches tryToFoldExtOfLoad to set the right location on a newly-created extload. With that in place, the logic for performing a certain ([s|z]ext (load ...)) combine becomes identical for sexts and zexts, and we can get rid of one copy of the logic. The test case churn is due to dependencies on IROrders inherited from the wrong SDLoc. Part of: llvm.org/PR37262 Differential Revision: https://reviews.llvm.org/D46158 llvm-svn: 332118
2018-05-12 02:40:08 +08:00
; KNL-NEXT: movl 744(%rbp), %eax
; KNL-NEXT: andl $127, %eax
[SelectionDAG][X86] Fix insert_vector_elt lowering for v32i1/v64i1 with non-constant index Summary: Currently we don't handle v32i1/v64i1 insert_vector_elt correctly as we fail to look at the number of elements closely and assume it can only be v16i1 or v8i1. We also can't type legalize v64i1 insert_vector_elt correctly on KNL due to the type not being byte addressable as required by the legalizing through memory accesses path requires. For the first issue, the patch now tries to pick a 512-bit register with the correct number of elements and promotes to that. For the second issue, we now extend the vector to a byte addressable type, do the stores to memory, load the two halves, and then truncate the halves back to the original type. Technically since we changed the type, we may not need two loads, but actually checking that is more work and for the v64i1 case we do need them. Reviewers: RKSimon, delena, spatel, zvi Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D40942 llvm-svn: 320849
2017-12-16 03:35:22 +08:00
; KNL-NEXT: vpxor %xmm3, %xmm3, %xmm3
[SelectionDAG] Support some SimplifySetCC cases for comparing against vector splats of constants. This supports things like (setcc ugt X, 0) -> (setcc ne X, 0) I've restricted to only make changes to vectors before legalize ops because I doubt all targets have accurate condition code legality information for vectors given how little we did before. Differential Revision: https://reviews.llvm.org/D42948 llvm-svn: 326495
2018-03-02 06:15:39 +08:00
; KNL-NEXT: vpcmpeqb %ymm3, %ymm2, %ymm2
[X86] Use min/max for vector ult/ugt compares if avoids a sign flip. Summary: Currently we only use min/max to help with ule/uge compares because it removes an invert of the result that would otherwise be needed. But we can also use it for ult/ugt compares if it will prevent the need for a sign bit flip needed to use pcmpgt at the cost of requiring an invert after the compare. I also refactored the code so that the max/min code is self contained and does its own return instead of setting up a flag to manipulate the rest of the function's behavior. Most of the test cases look ok with this. I did notice that we added instructions when one of the operands being sign flipped is a constant vector that we were able to constant fold the flip into. I also noticed that sometimes the SSE min/max clobbers a register that is needed after the compare. This resulted in an extra move being inserted before the min/max to preserve the register. We could try to detect this and switch from min to max and change the compare operands to use the operand that gets reused in the compare. Reviewers: spatel, RKSimon Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D42935 llvm-svn: 324842
2018-02-12 01:11:40 +08:00
; KNL-NEXT: vpternlogq $15, %zmm2, %zmm2, %zmm2
[SelectionDAG] Support some SimplifySetCC cases for comparing against vector splats of constants. This supports things like (setcc ugt X, 0) -> (setcc ne X, 0) I've restricted to only make changes to vectors before legalize ops because I doubt all targets have accurate condition code legality information for vectors given how little we did before. Differential Revision: https://reviews.llvm.org/D42948 llvm-svn: 326495
2018-03-02 06:15:39 +08:00
; KNL-NEXT: vpcmpeqb %ymm3, %ymm1, %ymm1
[X86] Use min/max for vector ult/ugt compares if avoids a sign flip. Summary: Currently we only use min/max to help with ule/uge compares because it removes an invert of the result that would otherwise be needed. But we can also use it for ult/ugt compares if it will prevent the need for a sign bit flip needed to use pcmpgt at the cost of requiring an invert after the compare. I also refactored the code so that the max/min code is self contained and does its own return instead of setting up a flag to manipulate the rest of the function's behavior. Most of the test cases look ok with this. I did notice that we added instructions when one of the operands being sign flipped is a constant vector that we were able to constant fold the flip into. I also noticed that sometimes the SSE min/max clobbers a register that is needed after the compare. This resulted in an extra move being inserted before the min/max to preserve the register. We could try to detect this and switch from min to max and change the compare operands to use the operand that gets reused in the compare. Reviewers: spatel, RKSimon Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D42935 llvm-svn: 324842
2018-02-12 01:11:40 +08:00
; KNL-NEXT: vpternlogq $15, %zmm1, %zmm1, %zmm1
[SelectionDAG] Support some SimplifySetCC cases for comparing against vector splats of constants. This supports things like (setcc ugt X, 0) -> (setcc ne X, 0) I've restricted to only make changes to vectors before legalize ops because I doubt all targets have accurate condition code legality information for vectors given how little we did before. Differential Revision: https://reviews.llvm.org/D42948 llvm-svn: 326495
2018-03-02 06:15:39 +08:00
; KNL-NEXT: vpcmpeqb %ymm3, %ymm0, %ymm0
[X86] Use min/max for vector ult/ugt compares if avoids a sign flip. Summary: Currently we only use min/max to help with ule/uge compares because it removes an invert of the result that would otherwise be needed. But we can also use it for ult/ugt compares if it will prevent the need for a sign bit flip needed to use pcmpgt at the cost of requiring an invert after the compare. I also refactored the code so that the max/min code is self contained and does its own return instead of setting up a flag to manipulate the rest of the function's behavior. Most of the test cases look ok with this. I did notice that we added instructions when one of the operands being sign flipped is a constant vector that we were able to constant fold the flip into. I also noticed that sometimes the SSE min/max clobbers a register that is needed after the compare. This resulted in an extra move being inserted before the min/max to preserve the register. We could try to detect this and switch from min to max and change the compare operands to use the operand that gets reused in the compare. Reviewers: spatel, RKSimon Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D42935 llvm-svn: 324842
2018-02-12 01:11:40 +08:00
; KNL-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
[SelectionDAG][X86] Fix insert_vector_elt lowering for v32i1/v64i1 with non-constant index Summary: Currently we don't handle v32i1/v64i1 insert_vector_elt correctly as we fail to look at the number of elements closely and assume it can only be v16i1 or v8i1. We also can't type legalize v64i1 insert_vector_elt correctly on KNL due to the type not being byte addressable as required by the legalizing through memory accesses path requires. For the first issue, the patch now tries to pick a 512-bit register with the correct number of elements and promotes to that. For the second issue, we now extend the vector to a byte addressable type, do the stores to memory, load the two halves, and then truncate the halves back to the original type. Technically since we changed the type, we may not need two loads, but actually checking that is more work and for the v64i1 case we do need them. Reviewers: RKSimon, delena, spatel, zvi Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D40942 llvm-svn: 320849
2017-12-16 03:35:22 +08:00
; KNL-NEXT: cmpb $0, 736(%rbp)
; KNL-NEXT: vmovdqa %ymm3, {{[0-9]+}}(%rsp)
; KNL-NEXT: vmovdqa %ymm0, {{[0-9]+}}(%rsp)
; KNL-NEXT: vmovdqa %ymm1, {{[0-9]+}}(%rsp)
[X86] Legalize v32i1 without BWI via splitting to v16i1 rather than the default of promoting to v32i8. Summary: For the most part its better to keep v32i1 as a mask type of a narrower width than trying to promote it to a ymm register. I had to add some overrides to the methods that get the types for the calling convention so that we still use v32i8 for argument/return purposes. There are still some regressions in here. I definitely saw some around shuffles. I think we probably should move vXi1 shuffle from lowering to a DAG combine where I think the extend and truncate we have to emit would be better combined. I think we also need a DAG combine to remove trunc from (extract_vector_elt (trunc)) Overall this removes something like 13000 CHECK lines from lit tests. Reviewers: zvi, RKSimon, delena, spatel Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D42031 llvm-svn: 323201
2018-01-23 22:25:39 +08:00
; KNL-NEXT: vmovdqa %ymm2, (%rsp)
; KNL-NEXT: setne (%rsp,%rax)
[x86] allow vector load narrowing with multi-use values This is a long-awaited follow-up suggested in D33578. Since then, we've picked up even more opportunities for vector narrowing from changes like D53784, so there are a lot of test diffs. Apart from 2-3 strange cases, these are all wins. I've structured this to be no-functional-change-intended for any target except for x86 because I couldn't tell if AArch64, ARM, and AMDGPU would improve or not. All of those targets have existing regression tests (4, 4, 10 files respectively) that would be affected. Also, Hexagon overrides the shouldReduceLoadWidth() hook, but doesn't show any regression test diffs. The trade-off is deciding if an extra vector load is better than a single wide load + extract_subvector. For x86, this is almost always better (on paper at least) because we often can fold loads into subsequent ops and not increase the official instruction count. There's also some unknown -- but potentially large -- benefit from using narrower vector ops if wide ops are implemented with multiple uops and/or frequency throttling is avoided. Differential Revision: https://reviews.llvm.org/D54073 llvm-svn: 346595
2018-11-11 04:05:31 +08:00
; KNL-NEXT: vpmovzxbd {{.*#+}} zmm0 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero,mem[4],zero,zero,zero,mem[5],zero,zero,zero,mem[6],zero,zero,zero,mem[7],zero,zero,zero,mem[8],zero,zero,zero,mem[9],zero,zero,zero,mem[10],zero,zero,zero,mem[11],zero,zero,zero,mem[12],zero,zero,zero,mem[13],zero,zero,zero,mem[14],zero,zero,zero,mem[15],zero,zero,zero
; KNL-NEXT: vpslld $31, %zmm0, %zmm0
; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
[X86] Legalize v32i1 without BWI via splitting to v16i1 rather than the default of promoting to v32i8. Summary: For the most part its better to keep v32i1 as a mask type of a narrower width than trying to promote it to a ymm register. I had to add some overrides to the methods that get the types for the calling convention so that we still use v32i8 for argument/return purposes. There are still some regressions in here. I definitely saw some around shuffles. I think we probably should move vXi1 shuffle from lowering to a DAG combine where I think the extend and truncate we have to emit would be better combined. I think we also need a DAG combine to remove trunc from (extract_vector_elt (trunc)) Overall this removes something like 13000 CHECK lines from lit tests. Reviewers: zvi, RKSimon, delena, spatel Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D42031 llvm-svn: 323201
2018-01-23 22:25:39 +08:00
; KNL-NEXT: kmovw %k0, %eax
[x86] allow vector load narrowing with multi-use values This is a long-awaited follow-up suggested in D33578. Since then, we've picked up even more opportunities for vector narrowing from changes like D53784, so there are a lot of test diffs. Apart from 2-3 strange cases, these are all wins. I've structured this to be no-functional-change-intended for any target except for x86 because I couldn't tell if AArch64, ARM, and AMDGPU would improve or not. All of those targets have existing regression tests (4, 4, 10 files respectively) that would be affected. Also, Hexagon overrides the shouldReduceLoadWidth() hook, but doesn't show any regression test diffs. The trade-off is deciding if an extra vector load is better than a single wide load + extract_subvector. For x86, this is almost always better (on paper at least) because we often can fold loads into subsequent ops and not increase the official instruction count. There's also some unknown -- but potentially large -- benefit from using narrower vector ops if wide ops are implemented with multiple uops and/or frequency throttling is avoided. Differential Revision: https://reviews.llvm.org/D54073 llvm-svn: 346595
2018-11-11 04:05:31 +08:00
; KNL-NEXT: vpmovzxbd {{.*#+}} zmm0 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero,mem[4],zero,zero,zero,mem[5],zero,zero,zero,mem[6],zero,zero,zero,mem[7],zero,zero,zero,mem[8],zero,zero,zero,mem[9],zero,zero,zero,mem[10],zero,zero,zero,mem[11],zero,zero,zero,mem[12],zero,zero,zero,mem[13],zero,zero,zero,mem[14],zero,zero,zero,mem[15],zero,zero,zero
; KNL-NEXT: vpslld $31, %zmm0, %zmm0
; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
[X86] Legalize v32i1 without BWI via splitting to v16i1 rather than the default of promoting to v32i8. Summary: For the most part its better to keep v32i1 as a mask type of a narrower width than trying to promote it to a ymm register. I had to add some overrides to the methods that get the types for the calling convention so that we still use v32i8 for argument/return purposes. There are still some regressions in here. I definitely saw some around shuffles. I think we probably should move vXi1 shuffle from lowering to a DAG combine where I think the extend and truncate we have to emit would be better combined. I think we also need a DAG combine to remove trunc from (extract_vector_elt (trunc)) Overall this removes something like 13000 CHECK lines from lit tests. Reviewers: zvi, RKSimon, delena, spatel Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D42031 llvm-svn: 323201
2018-01-23 22:25:39 +08:00
; KNL-NEXT: kmovw %k0, %ecx
; KNL-NEXT: shll $16, %ecx
; KNL-NEXT: orl %eax, %ecx
[x86] allow vector load narrowing with multi-use values This is a long-awaited follow-up suggested in D33578. Since then, we've picked up even more opportunities for vector narrowing from changes like D53784, so there are a lot of test diffs. Apart from 2-3 strange cases, these are all wins. I've structured this to be no-functional-change-intended for any target except for x86 because I couldn't tell if AArch64, ARM, and AMDGPU would improve or not. All of those targets have existing regression tests (4, 4, 10 files respectively) that would be affected. Also, Hexagon overrides the shouldReduceLoadWidth() hook, but doesn't show any regression test diffs. The trade-off is deciding if an extra vector load is better than a single wide load + extract_subvector. For x86, this is almost always better (on paper at least) because we often can fold loads into subsequent ops and not increase the official instruction count. There's also some unknown -- but potentially large -- benefit from using narrower vector ops if wide ops are implemented with multiple uops and/or frequency throttling is avoided. Differential Revision: https://reviews.llvm.org/D54073 llvm-svn: 346595
2018-11-11 04:05:31 +08:00
; KNL-NEXT: vpmovzxbd {{.*#+}} zmm0 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero,mem[4],zero,zero,zero,mem[5],zero,zero,zero,mem[6],zero,zero,zero,mem[7],zero,zero,zero,mem[8],zero,zero,zero,mem[9],zero,zero,zero,mem[10],zero,zero,zero,mem[11],zero,zero,zero,mem[12],zero,zero,zero,mem[13],zero,zero,zero,mem[14],zero,zero,zero,mem[15],zero,zero,zero
; KNL-NEXT: vpslld $31, %zmm0, %zmm0
; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
[X86] Legalize v32i1 without BWI via splitting to v16i1 rather than the default of promoting to v32i8. Summary: For the most part its better to keep v32i1 as a mask type of a narrower width than trying to promote it to a ymm register. I had to add some overrides to the methods that get the types for the calling convention so that we still use v32i8 for argument/return purposes. There are still some regressions in here. I definitely saw some around shuffles. I think we probably should move vXi1 shuffle from lowering to a DAG combine where I think the extend and truncate we have to emit would be better combined. I think we also need a DAG combine to remove trunc from (extract_vector_elt (trunc)) Overall this removes something like 13000 CHECK lines from lit tests. Reviewers: zvi, RKSimon, delena, spatel Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D42031 llvm-svn: 323201
2018-01-23 22:25:39 +08:00
; KNL-NEXT: kmovw %k0, %edx
[x86] allow vector load narrowing with multi-use values This is a long-awaited follow-up suggested in D33578. Since then, we've picked up even more opportunities for vector narrowing from changes like D53784, so there are a lot of test diffs. Apart from 2-3 strange cases, these are all wins. I've structured this to be no-functional-change-intended for any target except for x86 because I couldn't tell if AArch64, ARM, and AMDGPU would improve or not. All of those targets have existing regression tests (4, 4, 10 files respectively) that would be affected. Also, Hexagon overrides the shouldReduceLoadWidth() hook, but doesn't show any regression test diffs. The trade-off is deciding if an extra vector load is better than a single wide load + extract_subvector. For x86, this is almost always better (on paper at least) because we often can fold loads into subsequent ops and not increase the official instruction count. There's also some unknown -- but potentially large -- benefit from using narrower vector ops if wide ops are implemented with multiple uops and/or frequency throttling is avoided. Differential Revision: https://reviews.llvm.org/D54073 llvm-svn: 346595
2018-11-11 04:05:31 +08:00
; KNL-NEXT: vpmovzxbd {{.*#+}} zmm0 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero,mem[4],zero,zero,zero,mem[5],zero,zero,zero,mem[6],zero,zero,zero,mem[7],zero,zero,zero,mem[8],zero,zero,zero,mem[9],zero,zero,zero,mem[10],zero,zero,zero,mem[11],zero,zero,zero,mem[12],zero,zero,zero,mem[13],zero,zero,zero,mem[14],zero,zero,zero,mem[15],zero,zero,zero
; KNL-NEXT: vpslld $31, %zmm0, %zmm0
; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
[X86] Legalize v32i1 without BWI via splitting to v16i1 rather than the default of promoting to v32i8. Summary: For the most part its better to keep v32i1 as a mask type of a narrower width than trying to promote it to a ymm register. I had to add some overrides to the methods that get the types for the calling convention so that we still use v32i8 for argument/return purposes. There are still some regressions in here. I definitely saw some around shuffles. I think we probably should move vXi1 shuffle from lowering to a DAG combine where I think the extend and truncate we have to emit would be better combined. I think we also need a DAG combine to remove trunc from (extract_vector_elt (trunc)) Overall this removes something like 13000 CHECK lines from lit tests. Reviewers: zvi, RKSimon, delena, spatel Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D42031 llvm-svn: 323201
2018-01-23 22:25:39 +08:00
; KNL-NEXT: kmovw %k0, %eax
; KNL-NEXT: shll $16, %eax
; KNL-NEXT: orl %edx, %eax
; KNL-NEXT: shlq $32, %rax
; KNL-NEXT: orq %rcx, %rax
[x86] allow vector load narrowing with multi-use values This is a long-awaited follow-up suggested in D33578. Since then, we've picked up even more opportunities for vector narrowing from changes like D53784, so there are a lot of test diffs. Apart from 2-3 strange cases, these are all wins. I've structured this to be no-functional-change-intended for any target except for x86 because I couldn't tell if AArch64, ARM, and AMDGPU would improve or not. All of those targets have existing regression tests (4, 4, 10 files respectively) that would be affected. Also, Hexagon overrides the shouldReduceLoadWidth() hook, but doesn't show any regression test diffs. The trade-off is deciding if an extra vector load is better than a single wide load + extract_subvector. For x86, this is almost always better (on paper at least) because we often can fold loads into subsequent ops and not increase the official instruction count. There's also some unknown -- but potentially large -- benefit from using narrower vector ops if wide ops are implemented with multiple uops and/or frequency throttling is avoided. Differential Revision: https://reviews.llvm.org/D54073 llvm-svn: 346595
2018-11-11 04:05:31 +08:00
; KNL-NEXT: vpmovzxbd {{.*#+}} zmm0 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero,mem[4],zero,zero,zero,mem[5],zero,zero,zero,mem[6],zero,zero,zero,mem[7],zero,zero,zero,mem[8],zero,zero,zero,mem[9],zero,zero,zero,mem[10],zero,zero,zero,mem[11],zero,zero,zero,mem[12],zero,zero,zero,mem[13],zero,zero,zero,mem[14],zero,zero,zero,mem[15],zero,zero,zero
; KNL-NEXT: vpslld $31, %zmm0, %zmm0
; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
[X86] Legalize v32i1 without BWI via splitting to v16i1 rather than the default of promoting to v32i8. Summary: For the most part its better to keep v32i1 as a mask type of a narrower width than trying to promote it to a ymm register. I had to add some overrides to the methods that get the types for the calling convention so that we still use v32i8 for argument/return purposes. There are still some regressions in here. I definitely saw some around shuffles. I think we probably should move vXi1 shuffle from lowering to a DAG combine where I think the extend and truncate we have to emit would be better combined. I think we also need a DAG combine to remove trunc from (extract_vector_elt (trunc)) Overall this removes something like 13000 CHECK lines from lit tests. Reviewers: zvi, RKSimon, delena, spatel Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D42031 llvm-svn: 323201
2018-01-23 22:25:39 +08:00
; KNL-NEXT: kmovw %k0, %ecx
[x86] allow vector load narrowing with multi-use values This is a long-awaited follow-up suggested in D33578. Since then, we've picked up even more opportunities for vector narrowing from changes like D53784, so there are a lot of test diffs. Apart from 2-3 strange cases, these are all wins. I've structured this to be no-functional-change-intended for any target except for x86 because I couldn't tell if AArch64, ARM, and AMDGPU would improve or not. All of those targets have existing regression tests (4, 4, 10 files respectively) that would be affected. Also, Hexagon overrides the shouldReduceLoadWidth() hook, but doesn't show any regression test diffs. The trade-off is deciding if an extra vector load is better than a single wide load + extract_subvector. For x86, this is almost always better (on paper at least) because we often can fold loads into subsequent ops and not increase the official instruction count. There's also some unknown -- but potentially large -- benefit from using narrower vector ops if wide ops are implemented with multiple uops and/or frequency throttling is avoided. Differential Revision: https://reviews.llvm.org/D54073 llvm-svn: 346595
2018-11-11 04:05:31 +08:00
; KNL-NEXT: vpmovzxbd {{.*#+}} zmm0 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero,mem[4],zero,zero,zero,mem[5],zero,zero,zero,mem[6],zero,zero,zero,mem[7],zero,zero,zero,mem[8],zero,zero,zero,mem[9],zero,zero,zero,mem[10],zero,zero,zero,mem[11],zero,zero,zero,mem[12],zero,zero,zero,mem[13],zero,zero,zero,mem[14],zero,zero,zero,mem[15],zero,zero,zero
; KNL-NEXT: vpslld $31, %zmm0, %zmm0
; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
[X86] Legalize v32i1 without BWI via splitting to v16i1 rather than the default of promoting to v32i8. Summary: For the most part its better to keep v32i1 as a mask type of a narrower width than trying to promote it to a ymm register. I had to add some overrides to the methods that get the types for the calling convention so that we still use v32i8 for argument/return purposes. There are still some regressions in here. I definitely saw some around shuffles. I think we probably should move vXi1 shuffle from lowering to a DAG combine where I think the extend and truncate we have to emit would be better combined. I think we also need a DAG combine to remove trunc from (extract_vector_elt (trunc)) Overall this removes something like 13000 CHECK lines from lit tests. Reviewers: zvi, RKSimon, delena, spatel Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D42031 llvm-svn: 323201
2018-01-23 22:25:39 +08:00
; KNL-NEXT: kmovw %k0, %esi
; KNL-NEXT: shll $16, %esi
; KNL-NEXT: orl %ecx, %esi
[x86] allow vector load narrowing with multi-use values This is a long-awaited follow-up suggested in D33578. Since then, we've picked up even more opportunities for vector narrowing from changes like D53784, so there are a lot of test diffs. Apart from 2-3 strange cases, these are all wins. I've structured this to be no-functional-change-intended for any target except for x86 because I couldn't tell if AArch64, ARM, and AMDGPU would improve or not. All of those targets have existing regression tests (4, 4, 10 files respectively) that would be affected. Also, Hexagon overrides the shouldReduceLoadWidth() hook, but doesn't show any regression test diffs. The trade-off is deciding if an extra vector load is better than a single wide load + extract_subvector. For x86, this is almost always better (on paper at least) because we often can fold loads into subsequent ops and not increase the official instruction count. There's also some unknown -- but potentially large -- benefit from using narrower vector ops if wide ops are implemented with multiple uops and/or frequency throttling is avoided. Differential Revision: https://reviews.llvm.org/D54073 llvm-svn: 346595
2018-11-11 04:05:31 +08:00
; KNL-NEXT: vpmovzxbd {{.*#+}} zmm0 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero,mem[4],zero,zero,zero,mem[5],zero,zero,zero,mem[6],zero,zero,zero,mem[7],zero,zero,zero,mem[8],zero,zero,zero,mem[9],zero,zero,zero,mem[10],zero,zero,zero,mem[11],zero,zero,zero,mem[12],zero,zero,zero,mem[13],zero,zero,zero,mem[14],zero,zero,zero,mem[15],zero,zero,zero
; KNL-NEXT: vpslld $31, %zmm0, %zmm0
; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
[X86] Legalize v32i1 without BWI via splitting to v16i1 rather than the default of promoting to v32i8. Summary: For the most part its better to keep v32i1 as a mask type of a narrower width than trying to promote it to a ymm register. I had to add some overrides to the methods that get the types for the calling convention so that we still use v32i8 for argument/return purposes. There are still some regressions in here. I definitely saw some around shuffles. I think we probably should move vXi1 shuffle from lowering to a DAG combine where I think the extend and truncate we have to emit would be better combined. I think we also need a DAG combine to remove trunc from (extract_vector_elt (trunc)) Overall this removes something like 13000 CHECK lines from lit tests. Reviewers: zvi, RKSimon, delena, spatel Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D42031 llvm-svn: 323201
2018-01-23 22:25:39 +08:00
; KNL-NEXT: kmovw %k0, %ecx
[x86] allow vector load narrowing with multi-use values This is a long-awaited follow-up suggested in D33578. Since then, we've picked up even more opportunities for vector narrowing from changes like D53784, so there are a lot of test diffs. Apart from 2-3 strange cases, these are all wins. I've structured this to be no-functional-change-intended for any target except for x86 because I couldn't tell if AArch64, ARM, and AMDGPU would improve or not. All of those targets have existing regression tests (4, 4, 10 files respectively) that would be affected. Also, Hexagon overrides the shouldReduceLoadWidth() hook, but doesn't show any regression test diffs. The trade-off is deciding if an extra vector load is better than a single wide load + extract_subvector. For x86, this is almost always better (on paper at least) because we often can fold loads into subsequent ops and not increase the official instruction count. There's also some unknown -- but potentially large -- benefit from using narrower vector ops if wide ops are implemented with multiple uops and/or frequency throttling is avoided. Differential Revision: https://reviews.llvm.org/D54073 llvm-svn: 346595
2018-11-11 04:05:31 +08:00
; KNL-NEXT: vpmovzxbd {{.*#+}} zmm0 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero,mem[4],zero,zero,zero,mem[5],zero,zero,zero,mem[6],zero,zero,zero,mem[7],zero,zero,zero,mem[8],zero,zero,zero,mem[9],zero,zero,zero,mem[10],zero,zero,zero,mem[11],zero,zero,zero,mem[12],zero,zero,zero,mem[13],zero,zero,zero,mem[14],zero,zero,zero,mem[15],zero,zero,zero
[SelectionDAG][X86] Fix insert_vector_elt lowering for v32i1/v64i1 with non-constant index Summary: Currently we don't handle v32i1/v64i1 insert_vector_elt correctly as we fail to look at the number of elements closely and assume it can only be v16i1 or v8i1. We also can't type legalize v64i1 insert_vector_elt correctly on KNL due to the type not being byte addressable as required by the legalizing through memory accesses path requires. For the first issue, the patch now tries to pick a 512-bit register with the correct number of elements and promotes to that. For the second issue, we now extend the vector to a byte addressable type, do the stores to memory, load the two halves, and then truncate the halves back to the original type. Technically since we changed the type, we may not need two loads, but actually checking that is more work and for the v64i1 case we do need them. Reviewers: RKSimon, delena, spatel, zvi Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D40942 llvm-svn: 320849
2017-12-16 03:35:22 +08:00
; KNL-NEXT: vpslld $31, %zmm0, %zmm0
; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
[X86] Legalize v32i1 without BWI via splitting to v16i1 rather than the default of promoting to v32i8. Summary: For the most part its better to keep v32i1 as a mask type of a narrower width than trying to promote it to a ymm register. I had to add some overrides to the methods that get the types for the calling convention so that we still use v32i8 for argument/return purposes. There are still some regressions in here. I definitely saw some around shuffles. I think we probably should move vXi1 shuffle from lowering to a DAG combine where I think the extend and truncate we have to emit would be better combined. I think we also need a DAG combine to remove trunc from (extract_vector_elt (trunc)) Overall this removes something like 13000 CHECK lines from lit tests. Reviewers: zvi, RKSimon, delena, spatel Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D42031 llvm-svn: 323201
2018-01-23 22:25:39 +08:00
; KNL-NEXT: kmovw %k0, %edx
; KNL-NEXT: shll $16, %edx
; KNL-NEXT: orl %ecx, %edx
[SelectionDAG][X86] Fix insert_vector_elt lowering for v32i1/v64i1 with non-constant index Summary: Currently we don't handle v32i1/v64i1 insert_vector_elt correctly as we fail to look at the number of elements closely and assume it can only be v16i1 or v8i1. We also can't type legalize v64i1 insert_vector_elt correctly on KNL due to the type not being byte addressable as required by the legalizing through memory accesses path requires. For the first issue, the patch now tries to pick a 512-bit register with the correct number of elements and promotes to that. For the second issue, we now extend the vector to a byte addressable type, do the stores to memory, load the two halves, and then truncate the halves back to the original type. Technically since we changed the type, we may not need two loads, but actually checking that is more work and for the v64i1 case we do need them. Reviewers: RKSimon, delena, spatel, zvi Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D40942 llvm-svn: 320849
2017-12-16 03:35:22 +08:00
; KNL-NEXT: shlq $32, %rdx
[X86] Legalize v32i1 without BWI via splitting to v16i1 rather than the default of promoting to v32i8. Summary: For the most part its better to keep v32i1 as a mask type of a narrower width than trying to promote it to a ymm register. I had to add some overrides to the methods that get the types for the calling convention so that we still use v32i8 for argument/return purposes. There are still some regressions in here. I definitely saw some around shuffles. I think we probably should move vXi1 shuffle from lowering to a DAG combine where I think the extend and truncate we have to emit would be better combined. I think we also need a DAG combine to remove trunc from (extract_vector_elt (trunc)) Overall this removes something like 13000 CHECK lines from lit tests. Reviewers: zvi, RKSimon, delena, spatel Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D42031 llvm-svn: 323201
2018-01-23 22:25:39 +08:00
; KNL-NEXT: orq %rsi, %rdx
[SelectionDAG][X86] Fix insert_vector_elt lowering for v32i1/v64i1 with non-constant index Summary: Currently we don't handle v32i1/v64i1 insert_vector_elt correctly as we fail to look at the number of elements closely and assume it can only be v16i1 or v8i1. We also can't type legalize v64i1 insert_vector_elt correctly on KNL due to the type not being byte addressable as required by the legalizing through memory accesses path requires. For the first issue, the patch now tries to pick a 512-bit register with the correct number of elements and promotes to that. For the second issue, we now extend the vector to a byte addressable type, do the stores to memory, load the two halves, and then truncate the halves back to the original type. Technically since we changed the type, we may not need two loads, but actually checking that is more work and for the v64i1 case we do need them. Reviewers: RKSimon, delena, spatel, zvi Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D40942 llvm-svn: 320849
2017-12-16 03:35:22 +08:00
; KNL-NEXT: movq %rbp, %rsp
; KNL-NEXT: popq %rbp
; KNL-NEXT: vzeroupper
; KNL-NEXT: retq
;
; SKX-LABEL: test_insertelement_variable_v96i1:
; SKX: ## %bb.0:
; SKX-NEXT: pushq %rbp
; SKX-NEXT: .cfi_def_cfa_offset 16
; SKX-NEXT: .cfi_offset %rbp, -16
; SKX-NEXT: movq %rsp, %rbp
; SKX-NEXT: .cfi_def_cfa_register %rbp
; SKX-NEXT: andq $-128, %rsp
; SKX-NEXT: subq $256, %rsp ## imm = 0x100
; SKX-NEXT: vmovd {{.*#+}} xmm0 = mem[0],zero,zero,zero
; SKX-NEXT: vpinsrb $1, 232(%rbp), %xmm0, %xmm0
; SKX-NEXT: vpinsrb $2, 240(%rbp), %xmm0, %xmm0
; SKX-NEXT: vpinsrb $3, 248(%rbp), %xmm0, %xmm0
; SKX-NEXT: vpinsrb $4, 256(%rbp), %xmm0, %xmm0
; SKX-NEXT: vpinsrb $5, 264(%rbp), %xmm0, %xmm0
; SKX-NEXT: vpinsrb $6, 272(%rbp), %xmm0, %xmm0
; SKX-NEXT: vpinsrb $7, 280(%rbp), %xmm0, %xmm0
; SKX-NEXT: vpinsrb $8, 288(%rbp), %xmm0, %xmm0
; SKX-NEXT: vpinsrb $9, 296(%rbp), %xmm0, %xmm0
; SKX-NEXT: vpinsrb $10, 304(%rbp), %xmm0, %xmm0
; SKX-NEXT: vpinsrb $11, 312(%rbp), %xmm0, %xmm0
; SKX-NEXT: vpinsrb $12, 320(%rbp), %xmm0, %xmm0
; SKX-NEXT: vpinsrb $13, 328(%rbp), %xmm0, %xmm0
; SKX-NEXT: vpinsrb $14, 336(%rbp), %xmm0, %xmm0
; SKX-NEXT: vpinsrb $15, 344(%rbp), %xmm0, %xmm0
; SKX-NEXT: vmovd {{.*#+}} xmm1 = mem[0],zero,zero,zero
; SKX-NEXT: vpinsrb $1, 360(%rbp), %xmm1, %xmm1
; SKX-NEXT: vpinsrb $2, 368(%rbp), %xmm1, %xmm1
; SKX-NEXT: vpinsrb $3, 376(%rbp), %xmm1, %xmm1
; SKX-NEXT: vpinsrb $4, 384(%rbp), %xmm1, %xmm1
; SKX-NEXT: vpinsrb $5, 392(%rbp), %xmm1, %xmm1
; SKX-NEXT: vpinsrb $6, 400(%rbp), %xmm1, %xmm1
; SKX-NEXT: vpinsrb $7, 408(%rbp), %xmm1, %xmm1
; SKX-NEXT: vpinsrb $8, 416(%rbp), %xmm1, %xmm1
; SKX-NEXT: vpinsrb $9, 424(%rbp), %xmm1, %xmm1
; SKX-NEXT: vpinsrb $10, 432(%rbp), %xmm1, %xmm1
; SKX-NEXT: vpinsrb $11, 440(%rbp), %xmm1, %xmm1
; SKX-NEXT: vpinsrb $12, 448(%rbp), %xmm1, %xmm1
; SKX-NEXT: vpinsrb $13, 456(%rbp), %xmm1, %xmm1
; SKX-NEXT: vpinsrb $14, 464(%rbp), %xmm1, %xmm1
; SKX-NEXT: vpinsrb $15, 472(%rbp), %xmm1, %xmm1
; SKX-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm0
; SKX-NEXT: vmovd %edi, %xmm1
; SKX-NEXT: vpinsrb $1, %esi, %xmm1, %xmm1
; SKX-NEXT: vpinsrb $2, %edx, %xmm1, %xmm1
; SKX-NEXT: vpinsrb $3, %ecx, %xmm1, %xmm1
; SKX-NEXT: vpinsrb $4, %r8d, %xmm1, %xmm1
; SKX-NEXT: vpinsrb $5, %r9d, %xmm1, %xmm1
; SKX-NEXT: vpinsrb $6, 16(%rbp), %xmm1, %xmm1
; SKX-NEXT: vpinsrb $7, 24(%rbp), %xmm1, %xmm1
; SKX-NEXT: vpinsrb $8, 32(%rbp), %xmm1, %xmm1
; SKX-NEXT: vpinsrb $9, 40(%rbp), %xmm1, %xmm1
; SKX-NEXT: vpinsrb $10, 48(%rbp), %xmm1, %xmm1
; SKX-NEXT: vpinsrb $11, 56(%rbp), %xmm1, %xmm1
; SKX-NEXT: vpinsrb $12, 64(%rbp), %xmm1, %xmm1
; SKX-NEXT: vpinsrb $13, 72(%rbp), %xmm1, %xmm1
; SKX-NEXT: vpinsrb $14, 80(%rbp), %xmm1, %xmm1
; SKX-NEXT: vpinsrb $15, 88(%rbp), %xmm1, %xmm1
; SKX-NEXT: vmovd {{.*#+}} xmm2 = mem[0],zero,zero,zero
; SKX-NEXT: vpinsrb $1, 104(%rbp), %xmm2, %xmm2
; SKX-NEXT: vpinsrb $2, 112(%rbp), %xmm2, %xmm2
; SKX-NEXT: vpinsrb $3, 120(%rbp), %xmm2, %xmm2
; SKX-NEXT: vpinsrb $4, 128(%rbp), %xmm2, %xmm2
; SKX-NEXT: vpinsrb $5, 136(%rbp), %xmm2, %xmm2
; SKX-NEXT: vpinsrb $6, 144(%rbp), %xmm2, %xmm2
; SKX-NEXT: vpinsrb $7, 152(%rbp), %xmm2, %xmm2
; SKX-NEXT: vpinsrb $8, 160(%rbp), %xmm2, %xmm2
; SKX-NEXT: vpinsrb $9, 168(%rbp), %xmm2, %xmm2
; SKX-NEXT: vpinsrb $10, 176(%rbp), %xmm2, %xmm2
; SKX-NEXT: vpinsrb $11, 184(%rbp), %xmm2, %xmm2
; SKX-NEXT: vpinsrb $12, 192(%rbp), %xmm2, %xmm2
; SKX-NEXT: vpinsrb $13, 200(%rbp), %xmm2, %xmm2
; SKX-NEXT: vpinsrb $14, 208(%rbp), %xmm2, %xmm2
; SKX-NEXT: vpinsrb $15, 216(%rbp), %xmm2, %xmm2
; SKX-NEXT: vinserti128 $1, %xmm2, %ymm1, %ymm1
; SKX-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
; SKX-NEXT: vmovd {{.*#+}} xmm1 = mem[0],zero,zero,zero
; SKX-NEXT: vpinsrb $1, 488(%rbp), %xmm1, %xmm1
; SKX-NEXT: vpinsrb $2, 496(%rbp), %xmm1, %xmm1
; SKX-NEXT: vpinsrb $3, 504(%rbp), %xmm1, %xmm1
; SKX-NEXT: vpinsrb $4, 512(%rbp), %xmm1, %xmm1
; SKX-NEXT: vpinsrb $5, 520(%rbp), %xmm1, %xmm1
; SKX-NEXT: vpinsrb $6, 528(%rbp), %xmm1, %xmm1
; SKX-NEXT: vpinsrb $7, 536(%rbp), %xmm1, %xmm1
; SKX-NEXT: vpinsrb $8, 544(%rbp), %xmm1, %xmm1
; SKX-NEXT: vpinsrb $9, 552(%rbp), %xmm1, %xmm1
; SKX-NEXT: vpinsrb $10, 560(%rbp), %xmm1, %xmm1
; SKX-NEXT: vpinsrb $11, 568(%rbp), %xmm1, %xmm1
; SKX-NEXT: vpinsrb $12, 576(%rbp), %xmm1, %xmm1
; SKX-NEXT: vpinsrb $13, 584(%rbp), %xmm1, %xmm1
; SKX-NEXT: vpinsrb $14, 592(%rbp), %xmm1, %xmm1
; SKX-NEXT: vpinsrb $15, 600(%rbp), %xmm1, %xmm1
; SKX-NEXT: vmovd {{.*#+}} xmm2 = mem[0],zero,zero,zero
; SKX-NEXT: vpinsrb $1, 616(%rbp), %xmm2, %xmm2
; SKX-NEXT: vpinsrb $2, 624(%rbp), %xmm2, %xmm2
; SKX-NEXT: vpinsrb $3, 632(%rbp), %xmm2, %xmm2
; SKX-NEXT: vpinsrb $4, 640(%rbp), %xmm2, %xmm2
; SKX-NEXT: vpinsrb $5, 648(%rbp), %xmm2, %xmm2
; SKX-NEXT: vpinsrb $6, 656(%rbp), %xmm2, %xmm2
; SKX-NEXT: vpinsrb $7, 664(%rbp), %xmm2, %xmm2
; SKX-NEXT: vpinsrb $8, 672(%rbp), %xmm2, %xmm2
; SKX-NEXT: vpinsrb $9, 680(%rbp), %xmm2, %xmm2
; SKX-NEXT: vpinsrb $10, 688(%rbp), %xmm2, %xmm2
; SKX-NEXT: vpinsrb $11, 696(%rbp), %xmm2, %xmm2
; SKX-NEXT: vpinsrb $12, 704(%rbp), %xmm2, %xmm2
; SKX-NEXT: vpinsrb $13, 712(%rbp), %xmm2, %xmm2
; SKX-NEXT: vpinsrb $14, 720(%rbp), %xmm2, %xmm2
; SKX-NEXT: vpinsrb $15, 728(%rbp), %xmm2, %xmm2
; SKX-NEXT: vinserti128 $1, %xmm2, %ymm1, %ymm1
; SKX-NEXT: movl 744(%rbp), %eax
; SKX-NEXT: andl $127, %eax
[DAGCombiner] Set the right SDLoc on a newly-created sextload (6/N) This teaches tryToFoldExtOfLoad to set the right location on a newly-created extload. With that in place, the logic for performing a certain ([s|z]ext (load ...)) combine becomes identical for sexts and zexts, and we can get rid of one copy of the logic. The test case churn is due to dependencies on IROrders inherited from the wrong SDLoc. Part of: llvm.org/PR37262 Differential Revision: https://reviews.llvm.org/D46158 llvm-svn: 332118
2018-05-12 02:40:08 +08:00
; SKX-NEXT: vptestmb %zmm0, %zmm0, %k0
; SKX-NEXT: vptestmb %zmm1, %zmm1, %k1
[SelectionDAG][X86] Fix insert_vector_elt lowering for v32i1/v64i1 with non-constant index Summary: Currently we don't handle v32i1/v64i1 insert_vector_elt correctly as we fail to look at the number of elements closely and assume it can only be v16i1 or v8i1. We also can't type legalize v64i1 insert_vector_elt correctly on KNL due to the type not being byte addressable as required by the legalizing through memory accesses path requires. For the first issue, the patch now tries to pick a 512-bit register with the correct number of elements and promotes to that. For the second issue, we now extend the vector to a byte addressable type, do the stores to memory, load the two halves, and then truncate the halves back to the original type. Technically since we changed the type, we may not need two loads, but actually checking that is more work and for the v64i1 case we do need them. Reviewers: RKSimon, delena, spatel, zvi Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D40942 llvm-svn: 320849
2017-12-16 03:35:22 +08:00
; SKX-NEXT: cmpb $0, 736(%rbp)
[X86] Use SIGN_EXTEND to implement ANY_EXTEND from vXi1. llvm-svn: 321334
2017-12-22 10:30:26 +08:00
; SKX-NEXT: vpmovm2b %k1, %zmm0
[X86] Use vmovdqu64/vmovdqa64 for unmasked integer vector stores for consistency with loads. Previously we used 64 for vXi64 stores and 32 for everything else. This change uses 64 for everything just like do for loads. llvm-svn: 322820
2018-01-18 15:44:09 +08:00
; SKX-NEXT: vmovdqa64 %zmm0, {{[0-9]+}}(%rsp)
[X86] Use SIGN_EXTEND to implement ANY_EXTEND from vXi1. llvm-svn: 321334
2017-12-22 10:30:26 +08:00
; SKX-NEXT: vpmovm2b %k0, %zmm0
[X86] Use vmovdqu64/vmovdqa64 for unmasked integer vector stores for consistency with loads. Previously we used 64 for vXi64 stores and 32 for everything else. This change uses 64 for everything just like do for loads. llvm-svn: 322820
2018-01-18 15:44:09 +08:00
; SKX-NEXT: vmovdqa64 %zmm0, (%rsp)
[SelectionDAG] Reverse the order of operands in the ISD::ADD created by TargetLowering::getVectorElementPointer so that the FrameIndex is on the left. This seems to improve X86's ability to match this into an address computation. Otherwise the other operand gets assigned to the base register and the stack pointer + frame index ends up in the index register. But index registers can't encode ESP/RSP so we end up having to move it into another register to meet the constraint. I could try to improve the address matcher in X86, but swapping the producer seemed easier. Several other places already have the operands in this order so this is at least consistent. llvm-svn: 321370
2017-12-23 01:18:13 +08:00
; SKX-NEXT: setne (%rsp,%rax)
[SelectionDAG][X86] Fix insert_vector_elt lowering for v32i1/v64i1 with non-constant index Summary: Currently we don't handle v32i1/v64i1 insert_vector_elt correctly as we fail to look at the number of elements closely and assume it can only be v16i1 or v8i1. We also can't type legalize v64i1 insert_vector_elt correctly on KNL due to the type not being byte addressable as required by the legalizing through memory accesses path requires. For the first issue, the patch now tries to pick a 512-bit register with the correct number of elements and promotes to that. For the second issue, we now extend the vector to a byte addressable type, do the stores to memory, load the two halves, and then truncate the halves back to the original type. Technically since we changed the type, we may not need two loads, but actually checking that is more work and for the v64i1 case we do need them. Reviewers: RKSimon, delena, spatel, zvi Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D40942 llvm-svn: 320849
2017-12-16 03:35:22 +08:00
; SKX-NEXT: vpsllw $7, {{[0-9]+}}(%rsp), %zmm0
; SKX-NEXT: vpmovb2m %zmm0, %k0
; SKX-NEXT: vpsllw $7, (%rsp), %zmm0
; SKX-NEXT: vpmovb2m %zmm0, %k1
; SKX-NEXT: kmovq %k1, %rax
; SKX-NEXT: kmovq %k0, %rdx
; SKX-NEXT: movq %rbp, %rsp
; SKX-NEXT: popq %rbp
; SKX-NEXT: vzeroupper
; SKX-NEXT: retq
%t1 = icmp ugt <96 x i8> %a, zeroinitializer
%t2 = icmp ugt i8 %b, 0
%t3 = insertelement <96 x i1> %t1, i1 %t2, i32 %index
%t4 = bitcast <96 x i1> %t3 to i96
ret i96 %t4
}
define i128 @test_insertelement_variable_v128i1(<128 x i8> %a, i8 %b, i32 %index) {
; KNL-LABEL: test_insertelement_variable_v128i1:
; KNL: ## %bb.0:
; KNL-NEXT: pushq %rbp
; KNL-NEXT: .cfi_def_cfa_offset 16
; KNL-NEXT: .cfi_offset %rbp, -16
; KNL-NEXT: movq %rsp, %rbp
; KNL-NEXT: .cfi_def_cfa_register %rbp
; KNL-NEXT: andq $-128, %rsp
[X86] Legalize v32i1 without BWI via splitting to v16i1 rather than the default of promoting to v32i8. Summary: For the most part its better to keep v32i1 as a mask type of a narrower width than trying to promote it to a ymm register. I had to add some overrides to the methods that get the types for the calling convention so that we still use v32i8 for argument/return purposes. There are still some regressions in here. I definitely saw some around shuffles. I think we probably should move vXi1 shuffle from lowering to a DAG combine where I think the extend and truncate we have to emit would be better combined. I think we also need a DAG combine to remove trunc from (extract_vector_elt (trunc)) Overall this removes something like 13000 CHECK lines from lit tests. Reviewers: zvi, RKSimon, delena, spatel Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D42031 llvm-svn: 323201
2018-01-23 22:25:39 +08:00
; KNL-NEXT: subq $256, %rsp ## imm = 0x100
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; KNL-NEXT: ## kill: def $esi killed $esi def $rsi
[X86] Use min/max for vector ult/ugt compares if avoids a sign flip. Summary: Currently we only use min/max to help with ule/uge compares because it removes an invert of the result that would otherwise be needed. But we can also use it for ult/ugt compares if it will prevent the need for a sign bit flip needed to use pcmpgt at the cost of requiring an invert after the compare. I also refactored the code so that the max/min code is self contained and does its own return instead of setting up a flag to manipulate the rest of the function's behavior. Most of the test cases look ok with this. I did notice that we added instructions when one of the operands being sign flipped is a constant vector that we were able to constant fold the flip into. I also noticed that sometimes the SSE min/max clobbers a register that is needed after the compare. This resulted in an extra move being inserted before the min/max to preserve the register. We could try to detect this and switch from min to max and change the compare operands to use the operand that gets reused in the compare. Reviewers: spatel, RKSimon Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D42935 llvm-svn: 324842
2018-02-12 01:11:40 +08:00
; KNL-NEXT: vpxor %xmm4, %xmm4, %xmm4
[SelectionDAG] Support some SimplifySetCC cases for comparing against vector splats of constants. This supports things like (setcc ugt X, 0) -> (setcc ne X, 0) I've restricted to only make changes to vectors before legalize ops because I doubt all targets have accurate condition code legality information for vectors given how little we did before. Differential Revision: https://reviews.llvm.org/D42948 llvm-svn: 326495
2018-03-02 06:15:39 +08:00
; KNL-NEXT: vpcmpeqb %ymm4, %ymm0, %ymm0
[X86] Use min/max for vector ult/ugt compares if avoids a sign flip. Summary: Currently we only use min/max to help with ule/uge compares because it removes an invert of the result that would otherwise be needed. But we can also use it for ult/ugt compares if it will prevent the need for a sign bit flip needed to use pcmpgt at the cost of requiring an invert after the compare. I also refactored the code so that the max/min code is self contained and does its own return instead of setting up a flag to manipulate the rest of the function's behavior. Most of the test cases look ok with this. I did notice that we added instructions when one of the operands being sign flipped is a constant vector that we were able to constant fold the flip into. I also noticed that sometimes the SSE min/max clobbers a register that is needed after the compare. This resulted in an extra move being inserted before the min/max to preserve the register. We could try to detect this and switch from min to max and change the compare operands to use the operand that gets reused in the compare. Reviewers: spatel, RKSimon Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D42935 llvm-svn: 324842
2018-02-12 01:11:40 +08:00
; KNL-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
[SelectionDAG] Support some SimplifySetCC cases for comparing against vector splats of constants. This supports things like (setcc ugt X, 0) -> (setcc ne X, 0) I've restricted to only make changes to vectors before legalize ops because I doubt all targets have accurate condition code legality information for vectors given how little we did before. Differential Revision: https://reviews.llvm.org/D42948 llvm-svn: 326495
2018-03-02 06:15:39 +08:00
; KNL-NEXT: vpcmpeqb %ymm4, %ymm1, %ymm1
[X86] Use min/max for vector ult/ugt compares if avoids a sign flip. Summary: Currently we only use min/max to help with ule/uge compares because it removes an invert of the result that would otherwise be needed. But we can also use it for ult/ugt compares if it will prevent the need for a sign bit flip needed to use pcmpgt at the cost of requiring an invert after the compare. I also refactored the code so that the max/min code is self contained and does its own return instead of setting up a flag to manipulate the rest of the function's behavior. Most of the test cases look ok with this. I did notice that we added instructions when one of the operands being sign flipped is a constant vector that we were able to constant fold the flip into. I also noticed that sometimes the SSE min/max clobbers a register that is needed after the compare. This resulted in an extra move being inserted before the min/max to preserve the register. We could try to detect this and switch from min to max and change the compare operands to use the operand that gets reused in the compare. Reviewers: spatel, RKSimon Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D42935 llvm-svn: 324842
2018-02-12 01:11:40 +08:00
; KNL-NEXT: vpternlogq $15, %zmm1, %zmm1, %zmm1
[SelectionDAG] Support some SimplifySetCC cases for comparing against vector splats of constants. This supports things like (setcc ugt X, 0) -> (setcc ne X, 0) I've restricted to only make changes to vectors before legalize ops because I doubt all targets have accurate condition code legality information for vectors given how little we did before. Differential Revision: https://reviews.llvm.org/D42948 llvm-svn: 326495
2018-03-02 06:15:39 +08:00
; KNL-NEXT: vpcmpeqb %ymm4, %ymm2, %ymm2
[X86] Use min/max for vector ult/ugt compares if avoids a sign flip. Summary: Currently we only use min/max to help with ule/uge compares because it removes an invert of the result that would otherwise be needed. But we can also use it for ult/ugt compares if it will prevent the need for a sign bit flip needed to use pcmpgt at the cost of requiring an invert after the compare. I also refactored the code so that the max/min code is self contained and does its own return instead of setting up a flag to manipulate the rest of the function's behavior. Most of the test cases look ok with this. I did notice that we added instructions when one of the operands being sign flipped is a constant vector that we were able to constant fold the flip into. I also noticed that sometimes the SSE min/max clobbers a register that is needed after the compare. This resulted in an extra move being inserted before the min/max to preserve the register. We could try to detect this and switch from min to max and change the compare operands to use the operand that gets reused in the compare. Reviewers: spatel, RKSimon Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D42935 llvm-svn: 324842
2018-02-12 01:11:40 +08:00
; KNL-NEXT: vpternlogq $15, %zmm2, %zmm2, %zmm2
; KNL-NEXT: vpcmpeqb %ymm4, %ymm3, %ymm3
; KNL-NEXT: vpternlogq $15, %zmm3, %zmm3, %zmm3
[SelectionDAG][X86] Fix insert_vector_elt lowering for v32i1/v64i1 with non-constant index Summary: Currently we don't handle v32i1/v64i1 insert_vector_elt correctly as we fail to look at the number of elements closely and assume it can only be v16i1 or v8i1. We also can't type legalize v64i1 insert_vector_elt correctly on KNL due to the type not being byte addressable as required by the legalizing through memory accesses path requires. For the first issue, the patch now tries to pick a 512-bit register with the correct number of elements and promotes to that. For the second issue, we now extend the vector to a byte addressable type, do the stores to memory, load the two halves, and then truncate the halves back to the original type. Technically since we changed the type, we may not need two loads, but actually checking that is more work and for the v64i1 case we do need them. Reviewers: RKSimon, delena, spatel, zvi Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D40942 llvm-svn: 320849
2017-12-16 03:35:22 +08:00
; KNL-NEXT: andl $127, %esi
; KNL-NEXT: testb %dil, %dil
; KNL-NEXT: vmovdqa %ymm3, {{[0-9]+}}(%rsp)
; KNL-NEXT: vmovdqa %ymm2, {{[0-9]+}}(%rsp)
; KNL-NEXT: vmovdqa %ymm1, {{[0-9]+}}(%rsp)
[X86] Legalize v32i1 without BWI via splitting to v16i1 rather than the default of promoting to v32i8. Summary: For the most part its better to keep v32i1 as a mask type of a narrower width than trying to promote it to a ymm register. I had to add some overrides to the methods that get the types for the calling convention so that we still use v32i8 for argument/return purposes. There are still some regressions in here. I definitely saw some around shuffles. I think we probably should move vXi1 shuffle from lowering to a DAG combine where I think the extend and truncate we have to emit would be better combined. I think we also need a DAG combine to remove trunc from (extract_vector_elt (trunc)) Overall this removes something like 13000 CHECK lines from lit tests. Reviewers: zvi, RKSimon, delena, spatel Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D42031 llvm-svn: 323201
2018-01-23 22:25:39 +08:00
; KNL-NEXT: vmovdqa %ymm0, (%rsp)
; KNL-NEXT: setne (%rsp,%rsi)
[x86] allow vector load narrowing with multi-use values This is a long-awaited follow-up suggested in D33578. Since then, we've picked up even more opportunities for vector narrowing from changes like D53784, so there are a lot of test diffs. Apart from 2-3 strange cases, these are all wins. I've structured this to be no-functional-change-intended for any target except for x86 because I couldn't tell if AArch64, ARM, and AMDGPU would improve or not. All of those targets have existing regression tests (4, 4, 10 files respectively) that would be affected. Also, Hexagon overrides the shouldReduceLoadWidth() hook, but doesn't show any regression test diffs. The trade-off is deciding if an extra vector load is better than a single wide load + extract_subvector. For x86, this is almost always better (on paper at least) because we often can fold loads into subsequent ops and not increase the official instruction count. There's also some unknown -- but potentially large -- benefit from using narrower vector ops if wide ops are implemented with multiple uops and/or frequency throttling is avoided. Differential Revision: https://reviews.llvm.org/D54073 llvm-svn: 346595
2018-11-11 04:05:31 +08:00
; KNL-NEXT: vpmovzxbd {{.*#+}} zmm0 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero,mem[4],zero,zero,zero,mem[5],zero,zero,zero,mem[6],zero,zero,zero,mem[7],zero,zero,zero,mem[8],zero,zero,zero,mem[9],zero,zero,zero,mem[10],zero,zero,zero,mem[11],zero,zero,zero,mem[12],zero,zero,zero,mem[13],zero,zero,zero,mem[14],zero,zero,zero,mem[15],zero,zero,zero
; KNL-NEXT: vpslld $31, %zmm0, %zmm0
; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
[X86] Legalize v32i1 without BWI via splitting to v16i1 rather than the default of promoting to v32i8. Summary: For the most part its better to keep v32i1 as a mask type of a narrower width than trying to promote it to a ymm register. I had to add some overrides to the methods that get the types for the calling convention so that we still use v32i8 for argument/return purposes. There are still some regressions in here. I definitely saw some around shuffles. I think we probably should move vXi1 shuffle from lowering to a DAG combine where I think the extend and truncate we have to emit would be better combined. I think we also need a DAG combine to remove trunc from (extract_vector_elt (trunc)) Overall this removes something like 13000 CHECK lines from lit tests. Reviewers: zvi, RKSimon, delena, spatel Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D42031 llvm-svn: 323201
2018-01-23 22:25:39 +08:00
; KNL-NEXT: kmovw %k0, %eax
[x86] allow vector load narrowing with multi-use values This is a long-awaited follow-up suggested in D33578. Since then, we've picked up even more opportunities for vector narrowing from changes like D53784, so there are a lot of test diffs. Apart from 2-3 strange cases, these are all wins. I've structured this to be no-functional-change-intended for any target except for x86 because I couldn't tell if AArch64, ARM, and AMDGPU would improve or not. All of those targets have existing regression tests (4, 4, 10 files respectively) that would be affected. Also, Hexagon overrides the shouldReduceLoadWidth() hook, but doesn't show any regression test diffs. The trade-off is deciding if an extra vector load is better than a single wide load + extract_subvector. For x86, this is almost always better (on paper at least) because we often can fold loads into subsequent ops and not increase the official instruction count. There's also some unknown -- but potentially large -- benefit from using narrower vector ops if wide ops are implemented with multiple uops and/or frequency throttling is avoided. Differential Revision: https://reviews.llvm.org/D54073 llvm-svn: 346595
2018-11-11 04:05:31 +08:00
; KNL-NEXT: vpmovzxbd {{.*#+}} zmm0 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero,mem[4],zero,zero,zero,mem[5],zero,zero,zero,mem[6],zero,zero,zero,mem[7],zero,zero,zero,mem[8],zero,zero,zero,mem[9],zero,zero,zero,mem[10],zero,zero,zero,mem[11],zero,zero,zero,mem[12],zero,zero,zero,mem[13],zero,zero,zero,mem[14],zero,zero,zero,mem[15],zero,zero,zero
; KNL-NEXT: vpslld $31, %zmm0, %zmm0
; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
[X86] Legalize v32i1 without BWI via splitting to v16i1 rather than the default of promoting to v32i8. Summary: For the most part its better to keep v32i1 as a mask type of a narrower width than trying to promote it to a ymm register. I had to add some overrides to the methods that get the types for the calling convention so that we still use v32i8 for argument/return purposes. There are still some regressions in here. I definitely saw some around shuffles. I think we probably should move vXi1 shuffle from lowering to a DAG combine where I think the extend and truncate we have to emit would be better combined. I think we also need a DAG combine to remove trunc from (extract_vector_elt (trunc)) Overall this removes something like 13000 CHECK lines from lit tests. Reviewers: zvi, RKSimon, delena, spatel Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D42031 llvm-svn: 323201
2018-01-23 22:25:39 +08:00
; KNL-NEXT: kmovw %k0, %ecx
; KNL-NEXT: shll $16, %ecx
; KNL-NEXT: orl %eax, %ecx
[x86] allow vector load narrowing with multi-use values This is a long-awaited follow-up suggested in D33578. Since then, we've picked up even more opportunities for vector narrowing from changes like D53784, so there are a lot of test diffs. Apart from 2-3 strange cases, these are all wins. I've structured this to be no-functional-change-intended for any target except for x86 because I couldn't tell if AArch64, ARM, and AMDGPU would improve or not. All of those targets have existing regression tests (4, 4, 10 files respectively) that would be affected. Also, Hexagon overrides the shouldReduceLoadWidth() hook, but doesn't show any regression test diffs. The trade-off is deciding if an extra vector load is better than a single wide load + extract_subvector. For x86, this is almost always better (on paper at least) because we often can fold loads into subsequent ops and not increase the official instruction count. There's also some unknown -- but potentially large -- benefit from using narrower vector ops if wide ops are implemented with multiple uops and/or frequency throttling is avoided. Differential Revision: https://reviews.llvm.org/D54073 llvm-svn: 346595
2018-11-11 04:05:31 +08:00
; KNL-NEXT: vpmovzxbd {{.*#+}} zmm0 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero,mem[4],zero,zero,zero,mem[5],zero,zero,zero,mem[6],zero,zero,zero,mem[7],zero,zero,zero,mem[8],zero,zero,zero,mem[9],zero,zero,zero,mem[10],zero,zero,zero,mem[11],zero,zero,zero,mem[12],zero,zero,zero,mem[13],zero,zero,zero,mem[14],zero,zero,zero,mem[15],zero,zero,zero
; KNL-NEXT: vpslld $31, %zmm0, %zmm0
; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
[X86] Legalize v32i1 without BWI via splitting to v16i1 rather than the default of promoting to v32i8. Summary: For the most part its better to keep v32i1 as a mask type of a narrower width than trying to promote it to a ymm register. I had to add some overrides to the methods that get the types for the calling convention so that we still use v32i8 for argument/return purposes. There are still some regressions in here. I definitely saw some around shuffles. I think we probably should move vXi1 shuffle from lowering to a DAG combine where I think the extend and truncate we have to emit would be better combined. I think we also need a DAG combine to remove trunc from (extract_vector_elt (trunc)) Overall this removes something like 13000 CHECK lines from lit tests. Reviewers: zvi, RKSimon, delena, spatel Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D42031 llvm-svn: 323201
2018-01-23 22:25:39 +08:00
; KNL-NEXT: kmovw %k0, %edx
[x86] allow vector load narrowing with multi-use values This is a long-awaited follow-up suggested in D33578. Since then, we've picked up even more opportunities for vector narrowing from changes like D53784, so there are a lot of test diffs. Apart from 2-3 strange cases, these are all wins. I've structured this to be no-functional-change-intended for any target except for x86 because I couldn't tell if AArch64, ARM, and AMDGPU would improve or not. All of those targets have existing regression tests (4, 4, 10 files respectively) that would be affected. Also, Hexagon overrides the shouldReduceLoadWidth() hook, but doesn't show any regression test diffs. The trade-off is deciding if an extra vector load is better than a single wide load + extract_subvector. For x86, this is almost always better (on paper at least) because we often can fold loads into subsequent ops and not increase the official instruction count. There's also some unknown -- but potentially large -- benefit from using narrower vector ops if wide ops are implemented with multiple uops and/or frequency throttling is avoided. Differential Revision: https://reviews.llvm.org/D54073 llvm-svn: 346595
2018-11-11 04:05:31 +08:00
; KNL-NEXT: vpmovzxbd {{.*#+}} zmm0 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero,mem[4],zero,zero,zero,mem[5],zero,zero,zero,mem[6],zero,zero,zero,mem[7],zero,zero,zero,mem[8],zero,zero,zero,mem[9],zero,zero,zero,mem[10],zero,zero,zero,mem[11],zero,zero,zero,mem[12],zero,zero,zero,mem[13],zero,zero,zero,mem[14],zero,zero,zero,mem[15],zero,zero,zero
; KNL-NEXT: vpslld $31, %zmm0, %zmm0
; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
[X86] Legalize v32i1 without BWI via splitting to v16i1 rather than the default of promoting to v32i8. Summary: For the most part its better to keep v32i1 as a mask type of a narrower width than trying to promote it to a ymm register. I had to add some overrides to the methods that get the types for the calling convention so that we still use v32i8 for argument/return purposes. There are still some regressions in here. I definitely saw some around shuffles. I think we probably should move vXi1 shuffle from lowering to a DAG combine where I think the extend and truncate we have to emit would be better combined. I think we also need a DAG combine to remove trunc from (extract_vector_elt (trunc)) Overall this removes something like 13000 CHECK lines from lit tests. Reviewers: zvi, RKSimon, delena, spatel Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D42031 llvm-svn: 323201
2018-01-23 22:25:39 +08:00
; KNL-NEXT: kmovw %k0, %eax
; KNL-NEXT: shll $16, %eax
; KNL-NEXT: orl %edx, %eax
; KNL-NEXT: shlq $32, %rax
; KNL-NEXT: orq %rcx, %rax
[x86] allow vector load narrowing with multi-use values This is a long-awaited follow-up suggested in D33578. Since then, we've picked up even more opportunities for vector narrowing from changes like D53784, so there are a lot of test diffs. Apart from 2-3 strange cases, these are all wins. I've structured this to be no-functional-change-intended for any target except for x86 because I couldn't tell if AArch64, ARM, and AMDGPU would improve or not. All of those targets have existing regression tests (4, 4, 10 files respectively) that would be affected. Also, Hexagon overrides the shouldReduceLoadWidth() hook, but doesn't show any regression test diffs. The trade-off is deciding if an extra vector load is better than a single wide load + extract_subvector. For x86, this is almost always better (on paper at least) because we often can fold loads into subsequent ops and not increase the official instruction count. There's also some unknown -- but potentially large -- benefit from using narrower vector ops if wide ops are implemented with multiple uops and/or frequency throttling is avoided. Differential Revision: https://reviews.llvm.org/D54073 llvm-svn: 346595
2018-11-11 04:05:31 +08:00
; KNL-NEXT: vpmovzxbd {{.*#+}} zmm0 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero,mem[4],zero,zero,zero,mem[5],zero,zero,zero,mem[6],zero,zero,zero,mem[7],zero,zero,zero,mem[8],zero,zero,zero,mem[9],zero,zero,zero,mem[10],zero,zero,zero,mem[11],zero,zero,zero,mem[12],zero,zero,zero,mem[13],zero,zero,zero,mem[14],zero,zero,zero,mem[15],zero,zero,zero
; KNL-NEXT: vpslld $31, %zmm0, %zmm0
; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
[X86] Legalize v32i1 without BWI via splitting to v16i1 rather than the default of promoting to v32i8. Summary: For the most part its better to keep v32i1 as a mask type of a narrower width than trying to promote it to a ymm register. I had to add some overrides to the methods that get the types for the calling convention so that we still use v32i8 for argument/return purposes. There are still some regressions in here. I definitely saw some around shuffles. I think we probably should move vXi1 shuffle from lowering to a DAG combine where I think the extend and truncate we have to emit would be better combined. I think we also need a DAG combine to remove trunc from (extract_vector_elt (trunc)) Overall this removes something like 13000 CHECK lines from lit tests. Reviewers: zvi, RKSimon, delena, spatel Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D42031 llvm-svn: 323201
2018-01-23 22:25:39 +08:00
; KNL-NEXT: kmovw %k0, %ecx
[x86] allow vector load narrowing with multi-use values This is a long-awaited follow-up suggested in D33578. Since then, we've picked up even more opportunities for vector narrowing from changes like D53784, so there are a lot of test diffs. Apart from 2-3 strange cases, these are all wins. I've structured this to be no-functional-change-intended for any target except for x86 because I couldn't tell if AArch64, ARM, and AMDGPU would improve or not. All of those targets have existing regression tests (4, 4, 10 files respectively) that would be affected. Also, Hexagon overrides the shouldReduceLoadWidth() hook, but doesn't show any regression test diffs. The trade-off is deciding if an extra vector load is better than a single wide load + extract_subvector. For x86, this is almost always better (on paper at least) because we often can fold loads into subsequent ops and not increase the official instruction count. There's also some unknown -- but potentially large -- benefit from using narrower vector ops if wide ops are implemented with multiple uops and/or frequency throttling is avoided. Differential Revision: https://reviews.llvm.org/D54073 llvm-svn: 346595
2018-11-11 04:05:31 +08:00
; KNL-NEXT: vpmovzxbd {{.*#+}} zmm0 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero,mem[4],zero,zero,zero,mem[5],zero,zero,zero,mem[6],zero,zero,zero,mem[7],zero,zero,zero,mem[8],zero,zero,zero,mem[9],zero,zero,zero,mem[10],zero,zero,zero,mem[11],zero,zero,zero,mem[12],zero,zero,zero,mem[13],zero,zero,zero,mem[14],zero,zero,zero,mem[15],zero,zero,zero
; KNL-NEXT: vpslld $31, %zmm0, %zmm0
; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
[X86] Legalize v32i1 without BWI via splitting to v16i1 rather than the default of promoting to v32i8. Summary: For the most part its better to keep v32i1 as a mask type of a narrower width than trying to promote it to a ymm register. I had to add some overrides to the methods that get the types for the calling convention so that we still use v32i8 for argument/return purposes. There are still some regressions in here. I definitely saw some around shuffles. I think we probably should move vXi1 shuffle from lowering to a DAG combine where I think the extend and truncate we have to emit would be better combined. I think we also need a DAG combine to remove trunc from (extract_vector_elt (trunc)) Overall this removes something like 13000 CHECK lines from lit tests. Reviewers: zvi, RKSimon, delena, spatel Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D42031 llvm-svn: 323201
2018-01-23 22:25:39 +08:00
; KNL-NEXT: kmovw %k0, %esi
; KNL-NEXT: shll $16, %esi
; KNL-NEXT: orl %ecx, %esi
[x86] allow vector load narrowing with multi-use values This is a long-awaited follow-up suggested in D33578. Since then, we've picked up even more opportunities for vector narrowing from changes like D53784, so there are a lot of test diffs. Apart from 2-3 strange cases, these are all wins. I've structured this to be no-functional-change-intended for any target except for x86 because I couldn't tell if AArch64, ARM, and AMDGPU would improve or not. All of those targets have existing regression tests (4, 4, 10 files respectively) that would be affected. Also, Hexagon overrides the shouldReduceLoadWidth() hook, but doesn't show any regression test diffs. The trade-off is deciding if an extra vector load is better than a single wide load + extract_subvector. For x86, this is almost always better (on paper at least) because we often can fold loads into subsequent ops and not increase the official instruction count. There's also some unknown -- but potentially large -- benefit from using narrower vector ops if wide ops are implemented with multiple uops and/or frequency throttling is avoided. Differential Revision: https://reviews.llvm.org/D54073 llvm-svn: 346595
2018-11-11 04:05:31 +08:00
; KNL-NEXT: vpmovzxbd {{.*#+}} zmm0 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero,mem[4],zero,zero,zero,mem[5],zero,zero,zero,mem[6],zero,zero,zero,mem[7],zero,zero,zero,mem[8],zero,zero,zero,mem[9],zero,zero,zero,mem[10],zero,zero,zero,mem[11],zero,zero,zero,mem[12],zero,zero,zero,mem[13],zero,zero,zero,mem[14],zero,zero,zero,mem[15],zero,zero,zero
; KNL-NEXT: vpslld $31, %zmm0, %zmm0
; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
[X86] Legalize v32i1 without BWI via splitting to v16i1 rather than the default of promoting to v32i8. Summary: For the most part its better to keep v32i1 as a mask type of a narrower width than trying to promote it to a ymm register. I had to add some overrides to the methods that get the types for the calling convention so that we still use v32i8 for argument/return purposes. There are still some regressions in here. I definitely saw some around shuffles. I think we probably should move vXi1 shuffle from lowering to a DAG combine where I think the extend and truncate we have to emit would be better combined. I think we also need a DAG combine to remove trunc from (extract_vector_elt (trunc)) Overall this removes something like 13000 CHECK lines from lit tests. Reviewers: zvi, RKSimon, delena, spatel Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D42031 llvm-svn: 323201
2018-01-23 22:25:39 +08:00
; KNL-NEXT: kmovw %k0, %ecx
[x86] allow vector load narrowing with multi-use values This is a long-awaited follow-up suggested in D33578. Since then, we've picked up even more opportunities for vector narrowing from changes like D53784, so there are a lot of test diffs. Apart from 2-3 strange cases, these are all wins. I've structured this to be no-functional-change-intended for any target except for x86 because I couldn't tell if AArch64, ARM, and AMDGPU would improve or not. All of those targets have existing regression tests (4, 4, 10 files respectively) that would be affected. Also, Hexagon overrides the shouldReduceLoadWidth() hook, but doesn't show any regression test diffs. The trade-off is deciding if an extra vector load is better than a single wide load + extract_subvector. For x86, this is almost always better (on paper at least) because we often can fold loads into subsequent ops and not increase the official instruction count. There's also some unknown -- but potentially large -- benefit from using narrower vector ops if wide ops are implemented with multiple uops and/or frequency throttling is avoided. Differential Revision: https://reviews.llvm.org/D54073 llvm-svn: 346595
2018-11-11 04:05:31 +08:00
; KNL-NEXT: vpmovzxbd {{.*#+}} zmm0 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero,mem[4],zero,zero,zero,mem[5],zero,zero,zero,mem[6],zero,zero,zero,mem[7],zero,zero,zero,mem[8],zero,zero,zero,mem[9],zero,zero,zero,mem[10],zero,zero,zero,mem[11],zero,zero,zero,mem[12],zero,zero,zero,mem[13],zero,zero,zero,mem[14],zero,zero,zero,mem[15],zero,zero,zero
[SelectionDAG][X86] Fix insert_vector_elt lowering for v32i1/v64i1 with non-constant index Summary: Currently we don't handle v32i1/v64i1 insert_vector_elt correctly as we fail to look at the number of elements closely and assume it can only be v16i1 or v8i1. We also can't type legalize v64i1 insert_vector_elt correctly on KNL due to the type not being byte addressable as required by the legalizing through memory accesses path requires. For the first issue, the patch now tries to pick a 512-bit register with the correct number of elements and promotes to that. For the second issue, we now extend the vector to a byte addressable type, do the stores to memory, load the two halves, and then truncate the halves back to the original type. Technically since we changed the type, we may not need two loads, but actually checking that is more work and for the v64i1 case we do need them. Reviewers: RKSimon, delena, spatel, zvi Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D40942 llvm-svn: 320849
2017-12-16 03:35:22 +08:00
; KNL-NEXT: vpslld $31, %zmm0, %zmm0
; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
[X86] Legalize v32i1 without BWI via splitting to v16i1 rather than the default of promoting to v32i8. Summary: For the most part its better to keep v32i1 as a mask type of a narrower width than trying to promote it to a ymm register. I had to add some overrides to the methods that get the types for the calling convention so that we still use v32i8 for argument/return purposes. There are still some regressions in here. I definitely saw some around shuffles. I think we probably should move vXi1 shuffle from lowering to a DAG combine where I think the extend and truncate we have to emit would be better combined. I think we also need a DAG combine to remove trunc from (extract_vector_elt (trunc)) Overall this removes something like 13000 CHECK lines from lit tests. Reviewers: zvi, RKSimon, delena, spatel Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D42031 llvm-svn: 323201
2018-01-23 22:25:39 +08:00
; KNL-NEXT: kmovw %k0, %edx
; KNL-NEXT: shll $16, %edx
; KNL-NEXT: orl %ecx, %edx
[SelectionDAG][X86] Fix insert_vector_elt lowering for v32i1/v64i1 with non-constant index Summary: Currently we don't handle v32i1/v64i1 insert_vector_elt correctly as we fail to look at the number of elements closely and assume it can only be v16i1 or v8i1. We also can't type legalize v64i1 insert_vector_elt correctly on KNL due to the type not being byte addressable as required by the legalizing through memory accesses path requires. For the first issue, the patch now tries to pick a 512-bit register with the correct number of elements and promotes to that. For the second issue, we now extend the vector to a byte addressable type, do the stores to memory, load the two halves, and then truncate the halves back to the original type. Technically since we changed the type, we may not need two loads, but actually checking that is more work and for the v64i1 case we do need them. Reviewers: RKSimon, delena, spatel, zvi Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D40942 llvm-svn: 320849
2017-12-16 03:35:22 +08:00
; KNL-NEXT: shlq $32, %rdx
[X86] Legalize v32i1 without BWI via splitting to v16i1 rather than the default of promoting to v32i8. Summary: For the most part its better to keep v32i1 as a mask type of a narrower width than trying to promote it to a ymm register. I had to add some overrides to the methods that get the types for the calling convention so that we still use v32i8 for argument/return purposes. There are still some regressions in here. I definitely saw some around shuffles. I think we probably should move vXi1 shuffle from lowering to a DAG combine where I think the extend and truncate we have to emit would be better combined. I think we also need a DAG combine to remove trunc from (extract_vector_elt (trunc)) Overall this removes something like 13000 CHECK lines from lit tests. Reviewers: zvi, RKSimon, delena, spatel Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D42031 llvm-svn: 323201
2018-01-23 22:25:39 +08:00
; KNL-NEXT: orq %rsi, %rdx
[SelectionDAG][X86] Fix insert_vector_elt lowering for v32i1/v64i1 with non-constant index Summary: Currently we don't handle v32i1/v64i1 insert_vector_elt correctly as we fail to look at the number of elements closely and assume it can only be v16i1 or v8i1. We also can't type legalize v64i1 insert_vector_elt correctly on KNL due to the type not being byte addressable as required by the legalizing through memory accesses path requires. For the first issue, the patch now tries to pick a 512-bit register with the correct number of elements and promotes to that. For the second issue, we now extend the vector to a byte addressable type, do the stores to memory, load the two halves, and then truncate the halves back to the original type. Technically since we changed the type, we may not need two loads, but actually checking that is more work and for the v64i1 case we do need them. Reviewers: RKSimon, delena, spatel, zvi Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D40942 llvm-svn: 320849
2017-12-16 03:35:22 +08:00
; KNL-NEXT: movq %rbp, %rsp
; KNL-NEXT: popq %rbp
; KNL-NEXT: vzeroupper
; KNL-NEXT: retq
;
; SKX-LABEL: test_insertelement_variable_v128i1:
; SKX: ## %bb.0:
; SKX-NEXT: pushq %rbp
; SKX-NEXT: .cfi_def_cfa_offset 16
; SKX-NEXT: .cfi_offset %rbp, -16
; SKX-NEXT: movq %rsp, %rbp
; SKX-NEXT: .cfi_def_cfa_register %rbp
; SKX-NEXT: andq $-128, %rsp
; SKX-NEXT: subq $256, %rsp ## imm = 0x100
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; SKX-NEXT: ## kill: def $esi killed $esi def $rsi
[SelectionDAG] Support some SimplifySetCC cases for comparing against vector splats of constants. This supports things like (setcc ugt X, 0) -> (setcc ne X, 0) I've restricted to only make changes to vectors before legalize ops because I doubt all targets have accurate condition code legality information for vectors given how little we did before. Differential Revision: https://reviews.llvm.org/D42948 llvm-svn: 326495
2018-03-02 06:15:39 +08:00
; SKX-NEXT: vptestmb %zmm0, %zmm0, %k0
; SKX-NEXT: vptestmb %zmm1, %zmm1, %k1
[SelectionDAG][X86] Fix insert_vector_elt lowering for v32i1/v64i1 with non-constant index Summary: Currently we don't handle v32i1/v64i1 insert_vector_elt correctly as we fail to look at the number of elements closely and assume it can only be v16i1 or v8i1. We also can't type legalize v64i1 insert_vector_elt correctly on KNL due to the type not being byte addressable as required by the legalizing through memory accesses path requires. For the first issue, the patch now tries to pick a 512-bit register with the correct number of elements and promotes to that. For the second issue, we now extend the vector to a byte addressable type, do the stores to memory, load the two halves, and then truncate the halves back to the original type. Technically since we changed the type, we may not need two loads, but actually checking that is more work and for the v64i1 case we do need them. Reviewers: RKSimon, delena, spatel, zvi Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D40942 llvm-svn: 320849
2017-12-16 03:35:22 +08:00
; SKX-NEXT: andl $127, %esi
; SKX-NEXT: testb %dil, %dil
[X86] Use SIGN_EXTEND to implement ANY_EXTEND from vXi1. llvm-svn: 321334
2017-12-22 10:30:26 +08:00
; SKX-NEXT: vpmovm2b %k1, %zmm0
[X86] Use vmovdqu64/vmovdqa64 for unmasked integer vector stores for consistency with loads. Previously we used 64 for vXi64 stores and 32 for everything else. This change uses 64 for everything just like do for loads. llvm-svn: 322820
2018-01-18 15:44:09 +08:00
; SKX-NEXT: vmovdqa64 %zmm0, {{[0-9]+}}(%rsp)
[X86] Use SIGN_EXTEND to implement ANY_EXTEND from vXi1. llvm-svn: 321334
2017-12-22 10:30:26 +08:00
; SKX-NEXT: vpmovm2b %k0, %zmm0
[X86] Use vmovdqu64/vmovdqa64 for unmasked integer vector stores for consistency with loads. Previously we used 64 for vXi64 stores and 32 for everything else. This change uses 64 for everything just like do for loads. llvm-svn: 322820
2018-01-18 15:44:09 +08:00
; SKX-NEXT: vmovdqa64 %zmm0, (%rsp)
[SelectionDAG] Reverse the order of operands in the ISD::ADD created by TargetLowering::getVectorElementPointer so that the FrameIndex is on the left. This seems to improve X86's ability to match this into an address computation. Otherwise the other operand gets assigned to the base register and the stack pointer + frame index ends up in the index register. But index registers can't encode ESP/RSP so we end up having to move it into another register to meet the constraint. I could try to improve the address matcher in X86, but swapping the producer seemed easier. Several other places already have the operands in this order so this is at least consistent. llvm-svn: 321370
2017-12-23 01:18:13 +08:00
; SKX-NEXT: setne (%rsp,%rsi)
[SelectionDAG][X86] Fix insert_vector_elt lowering for v32i1/v64i1 with non-constant index Summary: Currently we don't handle v32i1/v64i1 insert_vector_elt correctly as we fail to look at the number of elements closely and assume it can only be v16i1 or v8i1. We also can't type legalize v64i1 insert_vector_elt correctly on KNL due to the type not being byte addressable as required by the legalizing through memory accesses path requires. For the first issue, the patch now tries to pick a 512-bit register with the correct number of elements and promotes to that. For the second issue, we now extend the vector to a byte addressable type, do the stores to memory, load the two halves, and then truncate the halves back to the original type. Technically since we changed the type, we may not need two loads, but actually checking that is more work and for the v64i1 case we do need them. Reviewers: RKSimon, delena, spatel, zvi Reviewed By: RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D40942 llvm-svn: 320849
2017-12-16 03:35:22 +08:00
; SKX-NEXT: vpsllw $7, {{[0-9]+}}(%rsp), %zmm0
; SKX-NEXT: vpmovb2m %zmm0, %k0
; SKX-NEXT: vpsllw $7, (%rsp), %zmm0
; SKX-NEXT: vpmovb2m %zmm0, %k1
; SKX-NEXT: kmovq %k1, %rax
; SKX-NEXT: kmovq %k0, %rdx
; SKX-NEXT: movq %rbp, %rsp
; SKX-NEXT: popq %rbp
; SKX-NEXT: vzeroupper
; SKX-NEXT: retq
%t1 = icmp ugt <128 x i8> %a, zeroinitializer
%t2 = icmp ugt i8 %b, 0
%t3 = insertelement <128 x i1> %t1, i1 %t2, i32 %index
%t4 = bitcast <128 x i1> %t3 to i128
ret i128 %t4
}