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/vector-tzcnt-128.ll

1906 lines
73 KiB
LLVM
Raw Normal View History

[X86] Updated test checks script to generalise LCPI symbol refs The script now replace '.LCPI888_8' style asm symbols with the {{\.LCPI.*}} re pattern - this helps stop hardcoded symbols in 32-bit x86 tests changing with every edit of the file Refreshed some tests to demonstrate the new check llvm-svn: 272488
2016-06-12 04:39:21 +08:00
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
[X86] Updated vector tzcnt tests. Added vec512 tests. llvm-svn: 246922
2015-09-05 18:19:07 +08:00
; RUN: llc < %s -mtriple=x86_64-unknown-unknown | FileCheck %s --check-prefix=ALL --check-prefix=SSE --check-prefix=SSE2
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+sse3 | FileCheck %s --check-prefix=ALL --check-prefix=SSE --check-prefix=SSE3
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+ssse3 | FileCheck %s --check-prefix=ALL --check-prefix=SSE --check-prefix=SSSE3
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+sse4.1 | FileCheck %s --check-prefix=ALL --check-prefix=SSE --check-prefix=SSE41
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx | FileCheck %s --check-prefix=ALL --check-prefix=AVX --check-prefix=AVX1
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx2 | FileCheck %s --check-prefix=ALL --check-prefix=AVX --check-prefix=AVX2
[AVX512] Add avx512cd+vl runs to vector-tzcnt-128/256 tests to show using the vplzcntd/q instructions. llvm-svn: 266860
2016-04-20 13:19:01 +08:00
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512cd,+avx512vl | FileCheck %s --check-prefix=ALL --check-prefix=AVX --check-prefix=AVX512CDVL
[AVX512] Teach lowering to use vplzcntd/q to implement 128/256-bit CTTZ_ZERO_UNDEF even without VLX support. We can just extend to 512-bits and extract like we do for CTLZ. llvm-svn: 267100
2016-04-22 11:22:38 +08:00
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512cd,-avx512vl | FileCheck %s --check-prefix=ALL --check-prefix=AVX --check-prefix=AVX512CD
[X86] Adding vpopcntd and vpopcntq instructions AVX512_VPOPCNTDQ is a new feature set that was published by Intel. The patch represents the LLVM side of the addition of two new intrinsic based instructions (vpopcntd and vpopcntq). Differential Revision: https://reviews.llvm.org/D33169 llvm-svn: 303858
2017-05-25 21:45:23 +08:00
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512vpopcntdq | FileCheck %s --check-prefix=ALL --check-prefix=AVX512VPOPCNTDQ
[X86] Add 128 and 256-bit VPOPCNTDQ instructions. Adjust some tablegen classes LZCNT/POPCNT. I think when this instruction was first published it was only for a Knights CPU and thus VLX version was missing. llvm-svn: 320910
2017-12-16 10:40:28 +08:00
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512vpopcntdq,+avx512vl | FileCheck %s --check-prefix=ALL --check-prefix=AVX512VPOPCNTDQVL
[x86][icelake]BITALG vpopcnt{b,w} Differential Revision: https://reviews.llvm.org/D40213 llvm-svn: 318748
2017-11-21 18:32:42 +08:00
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512bitalg | FileCheck %s --check-prefix=ALL --check-prefix=BITALG_NOVLX
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512bitalg,+avx512vl | FileCheck %s --check-prefix=ALL --check-prefix=BITALG
[X86][SSE] Added 32-bit tests for vector lzcnt/tzcnt tests v2i64 tests are particularly bad on 32-bit targets. llvm-svn: 265789
2016-04-08 23:01:31 +08:00
;
; Just one 32-bit run to make sure we do reasonable things for i64 tzcnt.
; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=+sse4.1 | FileCheck %s --check-prefix=ALL --check-prefix=X32-SSE --check-prefix=X32-SSE41
[X86] Added tzcnt vector tests. llvm-svn: 239264
2015-06-08 05:01:34 +08:00
[X86][SSE] Tidied up vector CTLZ/CTTZ. NFCI. llvm-svn: 242645
2015-07-20 01:09:43 +08:00
define <2 x i64> @testv2i64(<2 x i64> %in) nounwind {
[X86] Added tzcnt vector tests. llvm-svn: 239264
2015-06-08 05:01:34 +08:00
; SSE2-LABEL: testv2i64:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; SSE2: # %bb.0:
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; SSE2-NEXT: pcmpeqd %xmm1, %xmm1
; SSE2-NEXT: paddq %xmm0, %xmm1
; SSE2-NEXT: pandn %xmm1, %xmm0
; SSE2-NEXT: movdqa %xmm0, %xmm1
; SSE2-NEXT: psrlw $1, %xmm1
; SSE2-NEXT: pand {{.*}}(%rip), %xmm1
; SSE2-NEXT: psubb %xmm1, %xmm0
; SSE2-NEXT: movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
; SSE2-NEXT: movdqa %xmm0, %xmm2
; SSE2-NEXT: pand %xmm1, %xmm2
; SSE2-NEXT: psrlw $2, %xmm0
; SSE2-NEXT: pand %xmm1, %xmm0
; SSE2-NEXT: paddb %xmm2, %xmm0
; SSE2-NEXT: movdqa %xmm0, %xmm1
; SSE2-NEXT: psrlw $4, %xmm1
; SSE2-NEXT: paddb %xmm0, %xmm1
; SSE2-NEXT: pand {{.*}}(%rip), %xmm1
; SSE2-NEXT: pxor %xmm0, %xmm0
; SSE2-NEXT: psadbw %xmm0, %xmm1
; SSE2-NEXT: movdqa %xmm1, %xmm0
[X86] Added tzcnt vector tests. llvm-svn: 239264
2015-06-08 05:01:34 +08:00
; SSE2-NEXT: retq
;
; SSE3-LABEL: testv2i64:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; SSE3: # %bb.0:
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; SSE3-NEXT: pcmpeqd %xmm1, %xmm1
; SSE3-NEXT: paddq %xmm0, %xmm1
; SSE3-NEXT: pandn %xmm1, %xmm0
; SSE3-NEXT: movdqa %xmm0, %xmm1
; SSE3-NEXT: psrlw $1, %xmm1
; SSE3-NEXT: pand {{.*}}(%rip), %xmm1
; SSE3-NEXT: psubb %xmm1, %xmm0
; SSE3-NEXT: movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
; SSE3-NEXT: movdqa %xmm0, %xmm2
; SSE3-NEXT: pand %xmm1, %xmm2
; SSE3-NEXT: psrlw $2, %xmm0
; SSE3-NEXT: pand %xmm1, %xmm0
; SSE3-NEXT: paddb %xmm2, %xmm0
; SSE3-NEXT: movdqa %xmm0, %xmm1
; SSE3-NEXT: psrlw $4, %xmm1
; SSE3-NEXT: paddb %xmm0, %xmm1
; SSE3-NEXT: pand {{.*}}(%rip), %xmm1
; SSE3-NEXT: pxor %xmm0, %xmm0
; SSE3-NEXT: psadbw %xmm0, %xmm1
; SSE3-NEXT: movdqa %xmm1, %xmm0
[X86] Added tzcnt vector tests. llvm-svn: 239264
2015-06-08 05:01:34 +08:00
; SSE3-NEXT: retq
;
; SSSE3-LABEL: testv2i64:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; SSSE3: # %bb.0:
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; SSSE3-NEXT: pcmpeqd %xmm1, %xmm1
; SSSE3-NEXT: paddq %xmm0, %xmm1
; SSSE3-NEXT: pandn %xmm1, %xmm0
; SSSE3-NEXT: movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
; SSSE3-NEXT: movdqa %xmm0, %xmm2
; SSSE3-NEXT: pand %xmm1, %xmm2
; SSSE3-NEXT: movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
[x86] transform vector inc/dec to use -1 constant (PR33483) Convert vector increment or decrement to sub/add with an all-ones constant: add X, <1, 1...> --> sub X, <-1, -1...> sub X, <1, 1...> --> add X, <-1, -1...> The all-ones vector constant can be materialized using a pcmpeq instruction that is commonly recognized as an idiom (has no register dependency), so that's better than loading a splat 1 constant. AVX512 uses 'vpternlogd' for 512-bit vectors because there is apparently no better way to produce 512 one-bits. The general advantages of this lowering are: 1. pcmpeq has lower latency than a memop on every uarch I looked at in Agner's tables, so in theory, this could be better for perf, but... 2. That seems unlikely to affect any OOO implementation, and I can't measure any real perf difference from this transform on Haswell or Jaguar, but... 3. It doesn't look like it from the diffs, but this is an overall size win because we eliminate 16 - 64 constant bytes in the case of a vector load. If we're broadcasting a scalar load (which might itself be a bug), then we're replacing a scalar constant load + broadcast with a single cheap op, so that should always be smaller/better too. 4. This makes the DAG/isel output more consistent - we use pcmpeq already for padd x, -1 and psub x, -1, so we should use that form for +1 too because we can. If there's some reason to favor a constant load on some CPU, let's make the reverse transform for all of these cases (either here in the DAG or in a later machine pass). This should fix: https://bugs.llvm.org/show_bug.cgi?id=33483 Differential Revision: https://reviews.llvm.org/D34336 llvm-svn: 306289
2017-06-26 22:19:26 +08:00
; SSSE3-NEXT: movdqa %xmm3, %xmm4
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; SSSE3-NEXT: pshufb %xmm2, %xmm4
; SSSE3-NEXT: psrlw $4, %xmm0
; SSSE3-NEXT: pand %xmm1, %xmm0
; SSSE3-NEXT: pshufb %xmm0, %xmm3
; SSSE3-NEXT: paddb %xmm4, %xmm3
; SSSE3-NEXT: pxor %xmm0, %xmm0
; SSSE3-NEXT: psadbw %xmm3, %xmm0
[X86] Added tzcnt vector tests. llvm-svn: 239264
2015-06-08 05:01:34 +08:00
; SSSE3-NEXT: retq
;
; SSE41-LABEL: testv2i64:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; SSE41: # %bb.0:
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; SSE41-NEXT: pcmpeqd %xmm1, %xmm1
; SSE41-NEXT: paddq %xmm0, %xmm1
; SSE41-NEXT: pandn %xmm1, %xmm0
; SSE41-NEXT: movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
; SSE41-NEXT: movdqa %xmm0, %xmm2
; SSE41-NEXT: pand %xmm1, %xmm2
; SSE41-NEXT: movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
[x86] transform vector inc/dec to use -1 constant (PR33483) Convert vector increment or decrement to sub/add with an all-ones constant: add X, <1, 1...> --> sub X, <-1, -1...> sub X, <1, 1...> --> add X, <-1, -1...> The all-ones vector constant can be materialized using a pcmpeq instruction that is commonly recognized as an idiom (has no register dependency), so that's better than loading a splat 1 constant. AVX512 uses 'vpternlogd' for 512-bit vectors because there is apparently no better way to produce 512 one-bits. The general advantages of this lowering are: 1. pcmpeq has lower latency than a memop on every uarch I looked at in Agner's tables, so in theory, this could be better for perf, but... 2. That seems unlikely to affect any OOO implementation, and I can't measure any real perf difference from this transform on Haswell or Jaguar, but... 3. It doesn't look like it from the diffs, but this is an overall size win because we eliminate 16 - 64 constant bytes in the case of a vector load. If we're broadcasting a scalar load (which might itself be a bug), then we're replacing a scalar constant load + broadcast with a single cheap op, so that should always be smaller/better too. 4. This makes the DAG/isel output more consistent - we use pcmpeq already for padd x, -1 and psub x, -1, so we should use that form for +1 too because we can. If there's some reason to favor a constant load on some CPU, let's make the reverse transform for all of these cases (either here in the DAG or in a later machine pass). This should fix: https://bugs.llvm.org/show_bug.cgi?id=33483 Differential Revision: https://reviews.llvm.org/D34336 llvm-svn: 306289
2017-06-26 22:19:26 +08:00
; SSE41-NEXT: movdqa %xmm3, %xmm4
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; SSE41-NEXT: pshufb %xmm2, %xmm4
; SSE41-NEXT: psrlw $4, %xmm0
; SSE41-NEXT: pand %xmm1, %xmm0
; SSE41-NEXT: pshufb %xmm0, %xmm3
; SSE41-NEXT: paddb %xmm4, %xmm3
; SSE41-NEXT: pxor %xmm0, %xmm0
; SSE41-NEXT: psadbw %xmm3, %xmm0
[X86] Added tzcnt vector tests. llvm-svn: 239264
2015-06-08 05:01:34 +08:00
; SSE41-NEXT: retq
;
[X86][SSE] Improve CTTZ lowering when CTLZ is legal If we have better CTLZ support than CTPOP, then use cttz(x) = width - ctlz(~x & (x - 1)) - and remove the CTTZ_ZERO_UNDEF handling as it no longer gives better codegen. Similar to rL344447, this is also closer to LegalizeDAG's approach llvm-svn: 344448
2018-10-13 21:05:19 +08:00
; AVX1-LABEL: testv2i64:
; AVX1: # %bb.0:
; AVX1-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
; AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm1
; AVX1-NEXT: vpandn %xmm1, %xmm0, %xmm0
; AVX1-NEXT: vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
; AVX1-NEXT: vpand %xmm1, %xmm0, %xmm2
; AVX1-NEXT: vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
; AVX1-NEXT: vpshufb %xmm2, %xmm3, %xmm2
; AVX1-NEXT: vpsrlw $4, %xmm0, %xmm0
; AVX1-NEXT: vpand %xmm1, %xmm0, %xmm0
; AVX1-NEXT: vpshufb %xmm0, %xmm3, %xmm0
; AVX1-NEXT: vpaddb %xmm2, %xmm0, %xmm0
; AVX1-NEXT: vpxor %xmm1, %xmm1, %xmm1
; AVX1-NEXT: vpsadbw %xmm1, %xmm0, %xmm0
; AVX1-NEXT: retq
;
; AVX2-LABEL: testv2i64:
; AVX2: # %bb.0:
; AVX2-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
; AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm1
; AVX2-NEXT: vpandn %xmm1, %xmm0, %xmm0
; AVX2-NEXT: vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
; AVX2-NEXT: vpand %xmm1, %xmm0, %xmm2
; AVX2-NEXT: vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
; AVX2-NEXT: vpshufb %xmm2, %xmm3, %xmm2
; AVX2-NEXT: vpsrlw $4, %xmm0, %xmm0
; AVX2-NEXT: vpand %xmm1, %xmm0, %xmm0
; AVX2-NEXT: vpshufb %xmm0, %xmm3, %xmm0
; AVX2-NEXT: vpaddb %xmm2, %xmm0, %xmm0
; AVX2-NEXT: vpxor %xmm1, %xmm1, %xmm1
; AVX2-NEXT: vpsadbw %xmm1, %xmm0, %xmm0
; AVX2-NEXT: retq
;
; AVX512CDVL-LABEL: testv2i64:
; AVX512CDVL: # %bb.0:
; AVX512CDVL-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
; AVX512CDVL-NEXT: vpaddq %xmm1, %xmm0, %xmm1
; AVX512CDVL-NEXT: vpandn %xmm1, %xmm0, %xmm0
; AVX512CDVL-NEXT: vplzcntq %xmm0, %xmm0
; AVX512CDVL-NEXT: vmovdqa {{.*#+}} xmm1 = [64,64]
; AVX512CDVL-NEXT: vpsubq %xmm0, %xmm1, %xmm0
; AVX512CDVL-NEXT: retq
;
; AVX512CD-LABEL: testv2i64:
; AVX512CD: # %bb.0:
; AVX512CD-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
; AVX512CD-NEXT: vpaddq %xmm1, %xmm0, %xmm1
; AVX512CD-NEXT: vpandn %xmm1, %xmm0, %xmm0
; AVX512CD-NEXT: vplzcntq %zmm0, %zmm0
; AVX512CD-NEXT: vmovdqa {{.*#+}} xmm1 = [64,64]
; AVX512CD-NEXT: vpsubq %xmm0, %xmm1, %xmm0
; AVX512CD-NEXT: vzeroupper
; AVX512CD-NEXT: retq
[X86][SSE] Added 32-bit tests for vector lzcnt/tzcnt tests v2i64 tests are particularly bad on 32-bit targets. llvm-svn: 265789
2016-04-08 23:01:31 +08:00
;
[X86] Adding vpopcntd and vpopcntq instructions AVX512_VPOPCNTDQ is a new feature set that was published by Intel. The patch represents the LLVM side of the addition of two new intrinsic based instructions (vpopcntd and vpopcntq). Differential Revision: https://reviews.llvm.org/D33169 llvm-svn: 303858
2017-05-25 21:45:23 +08:00
; AVX512VPOPCNTDQ-LABEL: testv2i64:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; AVX512VPOPCNTDQ: # %bb.0:
[x86] transform vector inc/dec to use -1 constant (PR33483) Convert vector increment or decrement to sub/add with an all-ones constant: add X, <1, 1...> --> sub X, <-1, -1...> sub X, <1, 1...> --> add X, <-1, -1...> The all-ones vector constant can be materialized using a pcmpeq instruction that is commonly recognized as an idiom (has no register dependency), so that's better than loading a splat 1 constant. AVX512 uses 'vpternlogd' for 512-bit vectors because there is apparently no better way to produce 512 one-bits. The general advantages of this lowering are: 1. pcmpeq has lower latency than a memop on every uarch I looked at in Agner's tables, so in theory, this could be better for perf, but... 2. That seems unlikely to affect any OOO implementation, and I can't measure any real perf difference from this transform on Haswell or Jaguar, but... 3. It doesn't look like it from the diffs, but this is an overall size win because we eliminate 16 - 64 constant bytes in the case of a vector load. If we're broadcasting a scalar load (which might itself be a bug), then we're replacing a scalar constant load + broadcast with a single cheap op, so that should always be smaller/better too. 4. This makes the DAG/isel output more consistent - we use pcmpeq already for padd x, -1 and psub x, -1, so we should use that form for +1 too because we can. If there's some reason to favor a constant load on some CPU, let's make the reverse transform for all of these cases (either here in the DAG or in a later machine pass). This should fix: https://bugs.llvm.org/show_bug.cgi?id=33483 Differential Revision: https://reviews.llvm.org/D34336 llvm-svn: 306289
2017-06-26 22:19:26 +08:00
; AVX512VPOPCNTDQ-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; AVX512VPOPCNTDQ-NEXT: vpaddq %xmm1, %xmm0, %xmm1
; AVX512VPOPCNTDQ-NEXT: vpandn %xmm1, %xmm0, %xmm0
[X86] Adding vpopcntd and vpopcntq instructions AVX512_VPOPCNTDQ is a new feature set that was published by Intel. The patch represents the LLVM side of the addition of two new intrinsic based instructions (vpopcntd and vpopcntq). Differential Revision: https://reviews.llvm.org/D33169 llvm-svn: 303858
2017-05-25 21:45:23 +08:00
; AVX512VPOPCNTDQ-NEXT: vpopcntq %zmm0, %zmm0
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
; AVX512VPOPCNTDQ-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
[X86] Adding vpopcntd and vpopcntq instructions AVX512_VPOPCNTDQ is a new feature set that was published by Intel. The patch represents the LLVM side of the addition of two new intrinsic based instructions (vpopcntd and vpopcntq). Differential Revision: https://reviews.llvm.org/D33169 llvm-svn: 303858
2017-05-25 21:45:23 +08:00
; AVX512VPOPCNTDQ-NEXT: vzeroupper
; AVX512VPOPCNTDQ-NEXT: retq
;
[X86] Add 128 and 256-bit VPOPCNTDQ instructions. Adjust some tablegen classes LZCNT/POPCNT. I think when this instruction was first published it was only for a Knights CPU and thus VLX version was missing. llvm-svn: 320910
2017-12-16 10:40:28 +08:00
; AVX512VPOPCNTDQVL-LABEL: testv2i64:
; AVX512VPOPCNTDQVL: # %bb.0:
; AVX512VPOPCNTDQVL-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; AVX512VPOPCNTDQVL-NEXT: vpaddq %xmm1, %xmm0, %xmm1
; AVX512VPOPCNTDQVL-NEXT: vpandn %xmm1, %xmm0, %xmm0
[X86] Add 128 and 256-bit VPOPCNTDQ instructions. Adjust some tablegen classes LZCNT/POPCNT. I think when this instruction was first published it was only for a Knights CPU and thus VLX version was missing. llvm-svn: 320910
2017-12-16 10:40:28 +08:00
; AVX512VPOPCNTDQVL-NEXT: vpopcntq %xmm0, %xmm0
; AVX512VPOPCNTDQVL-NEXT: retq
;
[X86] Regenerate the vector-popcnt and vector-tzcnt tests to get BITALG CHECK linse on all functions not just the vXi16/vXi8. llvm-svn: 318885
2017-11-23 07:35:12 +08:00
; BITALG_NOVLX-LABEL: testv2i64:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; BITALG_NOVLX: # %bb.0:
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; BITALG_NOVLX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
; BITALG_NOVLX-NEXT: vpaddq %xmm1, %xmm0, %xmm1
; BITALG_NOVLX-NEXT: vpandn %xmm1, %xmm0, %xmm0
[X86][SSE] LowerVectorCTPOP - pull out repeated byte sum stage. Pull out repeated byte sum stage for popcount of vector elements > 8bits. This allows us to simplify the LUT/BITMATH popcnt code to always assume vXi8 vectors, and also improves avx512bitalg codegen which only has access to vpopcntb/vpopcntw. llvm-svn: 344348
2018-10-12 22:18:47 +08:00
; BITALG_NOVLX-NEXT: vpopcntb %zmm0, %zmm0
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; BITALG_NOVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1
[X86] Regenerate the vector-popcnt and vector-tzcnt tests to get BITALG CHECK linse on all functions not just the vXi16/vXi8. llvm-svn: 318885
2017-11-23 07:35:12 +08:00
; BITALG_NOVLX-NEXT: vpsadbw %xmm1, %xmm0, %xmm0
[X86][SSE] LowerVectorCTPOP - pull out repeated byte sum stage. Pull out repeated byte sum stage for popcount of vector elements > 8bits. This allows us to simplify the LUT/BITMATH popcnt code to always assume vXi8 vectors, and also improves avx512bitalg codegen which only has access to vpopcntb/vpopcntw. llvm-svn: 344348
2018-10-12 22:18:47 +08:00
; BITALG_NOVLX-NEXT: vzeroupper
[X86] Regenerate the vector-popcnt and vector-tzcnt tests to get BITALG CHECK linse on all functions not just the vXi16/vXi8. llvm-svn: 318885
2017-11-23 07:35:12 +08:00
; BITALG_NOVLX-NEXT: retq
;
; BITALG-LABEL: testv2i64:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; BITALG: # %bb.0:
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; BITALG-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
; BITALG-NEXT: vpaddq %xmm1, %xmm0, %xmm1
; BITALG-NEXT: vpandn %xmm1, %xmm0, %xmm0
[X86][SSE] LowerVectorCTPOP - pull out repeated byte sum stage. Pull out repeated byte sum stage for popcount of vector elements > 8bits. This allows us to simplify the LUT/BITMATH popcnt code to always assume vXi8 vectors, and also improves avx512bitalg codegen which only has access to vpopcntb/vpopcntw. llvm-svn: 344348
2018-10-12 22:18:47 +08:00
; BITALG-NEXT: vpopcntb %xmm0, %xmm0
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; BITALG-NEXT: vpxor %xmm1, %xmm1, %xmm1
[X86] Regenerate the vector-popcnt and vector-tzcnt tests to get BITALG CHECK linse on all functions not just the vXi16/vXi8. llvm-svn: 318885
2017-11-23 07:35:12 +08:00
; BITALG-NEXT: vpsadbw %xmm1, %xmm0, %xmm0
; BITALG-NEXT: retq
;
[X86][SSE] Added 32-bit tests for vector lzcnt/tzcnt tests v2i64 tests are particularly bad on 32-bit targets. llvm-svn: 265789
2016-04-08 23:01:31 +08:00
; X32-SSE-LABEL: testv2i64:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; X32-SSE: # %bb.0:
[X86][SSE] combineIncDecVector - use isConstantSplat Use isConstantSplat instead of ISD::isConstantSplatVector to let us us peek through to illegal types (in this case for i686 targets to recognise i64 constants) llvm-svn: 344452
2018-10-13 22:45:44 +08:00
; X32-SSE-NEXT: pcmpeqd %xmm1, %xmm1
; X32-SSE-NEXT: paddq %xmm0, %xmm1
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; X32-SSE-NEXT: pandn %xmm1, %xmm0
; X32-SSE-NEXT: movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
; X32-SSE-NEXT: movdqa %xmm0, %xmm2
; X32-SSE-NEXT: pand %xmm1, %xmm2
; X32-SSE-NEXT: movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
; X32-SSE-NEXT: movdqa %xmm3, %xmm4
; X32-SSE-NEXT: pshufb %xmm2, %xmm4
; X32-SSE-NEXT: psrlw $4, %xmm0
; X32-SSE-NEXT: pand %xmm1, %xmm0
; X32-SSE-NEXT: pshufb %xmm0, %xmm3
; X32-SSE-NEXT: paddb %xmm4, %xmm3
; X32-SSE-NEXT: pxor %xmm0, %xmm0
; X32-SSE-NEXT: psadbw %xmm3, %xmm0
[X86][SSE] Added 32-bit tests for vector lzcnt/tzcnt tests v2i64 tests are particularly bad on 32-bit targets. llvm-svn: 265789
2016-04-08 23:01:31 +08:00
; X32-SSE-NEXT: retl
[X86] Added tzcnt vector tests. llvm-svn: 239264
2015-06-08 05:01:34 +08:00
%out = call <2 x i64> @llvm.cttz.v2i64(<2 x i64> %in, i1 0)
ret <2 x i64> %out
}
[X86][SSE] Tidied up vector CTLZ/CTTZ. NFCI. llvm-svn: 242645
2015-07-20 01:09:43 +08:00
define <2 x i64> @testv2i64u(<2 x i64> %in) nounwind {
[X86] Added tzcnt vector tests. llvm-svn: 239264
2015-06-08 05:01:34 +08:00
; SSE2-LABEL: testv2i64u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; SSE2: # %bb.0:
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; SSE2-NEXT: pcmpeqd %xmm1, %xmm1
; SSE2-NEXT: paddq %xmm0, %xmm1
; SSE2-NEXT: pandn %xmm1, %xmm0
; SSE2-NEXT: movdqa %xmm0, %xmm1
; SSE2-NEXT: psrlw $1, %xmm1
; SSE2-NEXT: pand {{.*}}(%rip), %xmm1
; SSE2-NEXT: psubb %xmm1, %xmm0
; SSE2-NEXT: movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
; SSE2-NEXT: movdqa %xmm0, %xmm2
; SSE2-NEXT: pand %xmm1, %xmm2
; SSE2-NEXT: psrlw $2, %xmm0
; SSE2-NEXT: pand %xmm1, %xmm0
; SSE2-NEXT: paddb %xmm2, %xmm0
; SSE2-NEXT: movdqa %xmm0, %xmm1
; SSE2-NEXT: psrlw $4, %xmm1
; SSE2-NEXT: paddb %xmm0, %xmm1
; SSE2-NEXT: pand {{.*}}(%rip), %xmm1
; SSE2-NEXT: pxor %xmm0, %xmm0
; SSE2-NEXT: psadbw %xmm0, %xmm1
; SSE2-NEXT: movdqa %xmm1, %xmm0
[X86] Added tzcnt vector tests. llvm-svn: 239264
2015-06-08 05:01:34 +08:00
; SSE2-NEXT: retq
;
; SSE3-LABEL: testv2i64u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; SSE3: # %bb.0:
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; SSE3-NEXT: pcmpeqd %xmm1, %xmm1
; SSE3-NEXT: paddq %xmm0, %xmm1
; SSE3-NEXT: pandn %xmm1, %xmm0
; SSE3-NEXT: movdqa %xmm0, %xmm1
; SSE3-NEXT: psrlw $1, %xmm1
; SSE3-NEXT: pand {{.*}}(%rip), %xmm1
; SSE3-NEXT: psubb %xmm1, %xmm0
; SSE3-NEXT: movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
; SSE3-NEXT: movdqa %xmm0, %xmm2
; SSE3-NEXT: pand %xmm1, %xmm2
; SSE3-NEXT: psrlw $2, %xmm0
; SSE3-NEXT: pand %xmm1, %xmm0
; SSE3-NEXT: paddb %xmm2, %xmm0
; SSE3-NEXT: movdqa %xmm0, %xmm1
; SSE3-NEXT: psrlw $4, %xmm1
; SSE3-NEXT: paddb %xmm0, %xmm1
; SSE3-NEXT: pand {{.*}}(%rip), %xmm1
; SSE3-NEXT: pxor %xmm0, %xmm0
; SSE3-NEXT: psadbw %xmm0, %xmm1
; SSE3-NEXT: movdqa %xmm1, %xmm0
[X86] Added tzcnt vector tests. llvm-svn: 239264
2015-06-08 05:01:34 +08:00
; SSE3-NEXT: retq
;
; SSSE3-LABEL: testv2i64u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; SSSE3: # %bb.0:
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; SSSE3-NEXT: pcmpeqd %xmm1, %xmm1
; SSSE3-NEXT: paddq %xmm0, %xmm1
; SSSE3-NEXT: pandn %xmm1, %xmm0
; SSSE3-NEXT: movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
; SSSE3-NEXT: movdqa %xmm0, %xmm2
; SSSE3-NEXT: pand %xmm1, %xmm2
; SSSE3-NEXT: movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
[x86] transform vector inc/dec to use -1 constant (PR33483) Convert vector increment or decrement to sub/add with an all-ones constant: add X, <1, 1...> --> sub X, <-1, -1...> sub X, <1, 1...> --> add X, <-1, -1...> The all-ones vector constant can be materialized using a pcmpeq instruction that is commonly recognized as an idiom (has no register dependency), so that's better than loading a splat 1 constant. AVX512 uses 'vpternlogd' for 512-bit vectors because there is apparently no better way to produce 512 one-bits. The general advantages of this lowering are: 1. pcmpeq has lower latency than a memop on every uarch I looked at in Agner's tables, so in theory, this could be better for perf, but... 2. That seems unlikely to affect any OOO implementation, and I can't measure any real perf difference from this transform on Haswell or Jaguar, but... 3. It doesn't look like it from the diffs, but this is an overall size win because we eliminate 16 - 64 constant bytes in the case of a vector load. If we're broadcasting a scalar load (which might itself be a bug), then we're replacing a scalar constant load + broadcast with a single cheap op, so that should always be smaller/better too. 4. This makes the DAG/isel output more consistent - we use pcmpeq already for padd x, -1 and psub x, -1, so we should use that form for +1 too because we can. If there's some reason to favor a constant load on some CPU, let's make the reverse transform for all of these cases (either here in the DAG or in a later machine pass). This should fix: https://bugs.llvm.org/show_bug.cgi?id=33483 Differential Revision: https://reviews.llvm.org/D34336 llvm-svn: 306289
2017-06-26 22:19:26 +08:00
; SSSE3-NEXT: movdqa %xmm3, %xmm4
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; SSSE3-NEXT: pshufb %xmm2, %xmm4
; SSSE3-NEXT: psrlw $4, %xmm0
; SSSE3-NEXT: pand %xmm1, %xmm0
; SSSE3-NEXT: pshufb %xmm0, %xmm3
; SSSE3-NEXT: paddb %xmm4, %xmm3
; SSSE3-NEXT: pxor %xmm0, %xmm0
; SSSE3-NEXT: psadbw %xmm3, %xmm0
[X86] Added tzcnt vector tests. llvm-svn: 239264
2015-06-08 05:01:34 +08:00
; SSSE3-NEXT: retq
;
; SSE41-LABEL: testv2i64u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; SSE41: # %bb.0:
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; SSE41-NEXT: pcmpeqd %xmm1, %xmm1
; SSE41-NEXT: paddq %xmm0, %xmm1
; SSE41-NEXT: pandn %xmm1, %xmm0
; SSE41-NEXT: movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
; SSE41-NEXT: movdqa %xmm0, %xmm2
; SSE41-NEXT: pand %xmm1, %xmm2
; SSE41-NEXT: movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
[x86] transform vector inc/dec to use -1 constant (PR33483) Convert vector increment or decrement to sub/add with an all-ones constant: add X, <1, 1...> --> sub X, <-1, -1...> sub X, <1, 1...> --> add X, <-1, -1...> The all-ones vector constant can be materialized using a pcmpeq instruction that is commonly recognized as an idiom (has no register dependency), so that's better than loading a splat 1 constant. AVX512 uses 'vpternlogd' for 512-bit vectors because there is apparently no better way to produce 512 one-bits. The general advantages of this lowering are: 1. pcmpeq has lower latency than a memop on every uarch I looked at in Agner's tables, so in theory, this could be better for perf, but... 2. That seems unlikely to affect any OOO implementation, and I can't measure any real perf difference from this transform on Haswell or Jaguar, but... 3. It doesn't look like it from the diffs, but this is an overall size win because we eliminate 16 - 64 constant bytes in the case of a vector load. If we're broadcasting a scalar load (which might itself be a bug), then we're replacing a scalar constant load + broadcast with a single cheap op, so that should always be smaller/better too. 4. This makes the DAG/isel output more consistent - we use pcmpeq already for padd x, -1 and psub x, -1, so we should use that form for +1 too because we can. If there's some reason to favor a constant load on some CPU, let's make the reverse transform for all of these cases (either here in the DAG or in a later machine pass). This should fix: https://bugs.llvm.org/show_bug.cgi?id=33483 Differential Revision: https://reviews.llvm.org/D34336 llvm-svn: 306289
2017-06-26 22:19:26 +08:00
; SSE41-NEXT: movdqa %xmm3, %xmm4
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; SSE41-NEXT: pshufb %xmm2, %xmm4
; SSE41-NEXT: psrlw $4, %xmm0
; SSE41-NEXT: pand %xmm1, %xmm0
; SSE41-NEXT: pshufb %xmm0, %xmm3
; SSE41-NEXT: paddb %xmm4, %xmm3
; SSE41-NEXT: pxor %xmm0, %xmm0
; SSE41-NEXT: psadbw %xmm3, %xmm0
[X86] Added tzcnt vector tests. llvm-svn: 239264
2015-06-08 05:01:34 +08:00
; SSE41-NEXT: retq
;
[AVX512] Add avx512cd+vl runs to vector-tzcnt-128/256 tests to show using the vplzcntd/q instructions. llvm-svn: 266860
2016-04-20 13:19:01 +08:00
; AVX1-LABEL: testv2i64u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; AVX1: # %bb.0:
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; AVX1-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
; AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm1
; AVX1-NEXT: vpandn %xmm1, %xmm0, %xmm0
; AVX1-NEXT: vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
; AVX1-NEXT: vpand %xmm1, %xmm0, %xmm2
; AVX1-NEXT: vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
; AVX1-NEXT: vpshufb %xmm2, %xmm3, %xmm2
[X86][SSE] Don't decide when to scalarize CTTZ/CTLZ for performance at lowering - this is what cost models are for Improved CTTZ/CTLZ costings will be added shortly llvm-svn: 277713
2016-08-04 18:14:39 +08:00
; AVX1-NEXT: vpsrlw $4, %xmm0, %xmm0
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; AVX1-NEXT: vpand %xmm1, %xmm0, %xmm0
; AVX1-NEXT: vpshufb %xmm0, %xmm3, %xmm0
; AVX1-NEXT: vpaddb %xmm2, %xmm0, %xmm0
; AVX1-NEXT: vpxor %xmm1, %xmm1, %xmm1
[X86][SSE] Don't decide when to scalarize CTTZ/CTLZ for performance at lowering - this is what cost models are for Improved CTTZ/CTLZ costings will be added shortly llvm-svn: 277713
2016-08-04 18:14:39 +08:00
; AVX1-NEXT: vpsadbw %xmm1, %xmm0, %xmm0
[AVX512] Add avx512cd+vl runs to vector-tzcnt-128/256 tests to show using the vplzcntd/q instructions. llvm-svn: 266860
2016-04-20 13:19:01 +08:00
; AVX1-NEXT: retq
;
; AVX2-LABEL: testv2i64u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; AVX2: # %bb.0:
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; AVX2-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
; AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm1
; AVX2-NEXT: vpandn %xmm1, %xmm0, %xmm0
; AVX2-NEXT: vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
; AVX2-NEXT: vpand %xmm1, %xmm0, %xmm2
; AVX2-NEXT: vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
; AVX2-NEXT: vpshufb %xmm2, %xmm3, %xmm2
[X86][SSE] Don't decide when to scalarize CTTZ/CTLZ for performance at lowering - this is what cost models are for Improved CTTZ/CTLZ costings will be added shortly llvm-svn: 277713
2016-08-04 18:14:39 +08:00
; AVX2-NEXT: vpsrlw $4, %xmm0, %xmm0
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; AVX2-NEXT: vpand %xmm1, %xmm0, %xmm0
; AVX2-NEXT: vpshufb %xmm0, %xmm3, %xmm0
; AVX2-NEXT: vpaddb %xmm2, %xmm0, %xmm0
; AVX2-NEXT: vpxor %xmm1, %xmm1, %xmm1
[X86][SSE] Don't decide when to scalarize CTTZ/CTLZ for performance at lowering - this is what cost models are for Improved CTTZ/CTLZ costings will be added shortly llvm-svn: 277713
2016-08-04 18:14:39 +08:00
; AVX2-NEXT: vpsadbw %xmm1, %xmm0, %xmm0
[AVX512] Add avx512cd+vl runs to vector-tzcnt-128/256 tests to show using the vplzcntd/q instructions. llvm-svn: 266860
2016-04-20 13:19:01 +08:00
; AVX2-NEXT: retq
;
; AVX512CDVL-LABEL: testv2i64u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; AVX512CDVL: # %bb.0:
[X86][SSE] Improve CTTZ lowering when CTLZ is legal If we have better CTLZ support than CTPOP, then use cttz(x) = width - ctlz(~x & (x - 1)) - and remove the CTTZ_ZERO_UNDEF handling as it no longer gives better codegen. Similar to rL344447, this is also closer to LegalizeDAG's approach llvm-svn: 344448
2018-10-13 21:05:19 +08:00
; AVX512CDVL-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
; AVX512CDVL-NEXT: vpaddq %xmm1, %xmm0, %xmm1
; AVX512CDVL-NEXT: vpandn %xmm1, %xmm0, %xmm0
[AVX512] Add avx512cd+vl runs to vector-tzcnt-128/256 tests to show using the vplzcntd/q instructions. llvm-svn: 266860
2016-04-20 13:19:01 +08:00
; AVX512CDVL-NEXT: vplzcntq %xmm0, %xmm0
[X86][SSE] Improve CTTZ lowering when CTLZ is legal If we have better CTLZ support than CTPOP, then use cttz(x) = width - ctlz(~x & (x - 1)) - and remove the CTTZ_ZERO_UNDEF handling as it no longer gives better codegen. Similar to rL344447, this is also closer to LegalizeDAG's approach llvm-svn: 344448
2018-10-13 21:05:19 +08:00
; AVX512CDVL-NEXT: vmovdqa {{.*#+}} xmm1 = [64,64]
[AVX512] Add avx512cd+vl runs to vector-tzcnt-128/256 tests to show using the vplzcntd/q instructions. llvm-svn: 266860
2016-04-20 13:19:01 +08:00
; AVX512CDVL-NEXT: vpsubq %xmm0, %xmm1, %xmm0
; AVX512CDVL-NEXT: retq
[X86][SSE] Added 32-bit tests for vector lzcnt/tzcnt tests v2i64 tests are particularly bad on 32-bit targets. llvm-svn: 265789
2016-04-08 23:01:31 +08:00
;
[AVX512] Teach lowering to use vplzcntd/q to implement 128/256-bit CTTZ_ZERO_UNDEF even without VLX support. We can just extend to 512-bits and extract like we do for CTLZ. llvm-svn: 267100
2016-04-22 11:22:38 +08:00
; AVX512CD-LABEL: testv2i64u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; AVX512CD: # %bb.0:
[X86][SSE] Improve CTTZ lowering when CTLZ is legal If we have better CTLZ support than CTPOP, then use cttz(x) = width - ctlz(~x & (x - 1)) - and remove the CTTZ_ZERO_UNDEF handling as it no longer gives better codegen. Similar to rL344447, this is also closer to LegalizeDAG's approach llvm-svn: 344448
2018-10-13 21:05:19 +08:00
; AVX512CD-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
; AVX512CD-NEXT: vpaddq %xmm1, %xmm0, %xmm1
; AVX512CD-NEXT: vpandn %xmm1, %xmm0, %xmm0
[AVX512] Teach lowering to use vplzcntd/q to implement 128/256-bit CTTZ_ZERO_UNDEF even without VLX support. We can just extend to 512-bits and extract like we do for CTLZ. llvm-svn: 267100
2016-04-22 11:22:38 +08:00
; AVX512CD-NEXT: vplzcntq %zmm0, %zmm0
[X86][SSE] Improve CTTZ lowering when CTLZ is legal If we have better CTLZ support than CTPOP, then use cttz(x) = width - ctlz(~x & (x - 1)) - and remove the CTTZ_ZERO_UNDEF handling as it no longer gives better codegen. Similar to rL344447, this is also closer to LegalizeDAG's approach llvm-svn: 344448
2018-10-13 21:05:19 +08:00
; AVX512CD-NEXT: vmovdqa {{.*#+}} xmm1 = [64,64]
[AVX512] Teach lowering to use vplzcntd/q to implement 128/256-bit CTTZ_ZERO_UNDEF even without VLX support. We can just extend to 512-bits and extract like we do for CTLZ. llvm-svn: 267100
2016-04-22 11:22:38 +08:00
; AVX512CD-NEXT: vpsubq %xmm0, %xmm1, %xmm0
[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
; AVX512CD-NEXT: vzeroupper
[AVX512] Teach lowering to use vplzcntd/q to implement 128/256-bit CTTZ_ZERO_UNDEF even without VLX support. We can just extend to 512-bits and extract like we do for CTLZ. llvm-svn: 267100
2016-04-22 11:22:38 +08:00
; AVX512CD-NEXT: retq
;
[X86] Adding vpopcntd and vpopcntq instructions AVX512_VPOPCNTDQ is a new feature set that was published by Intel. The patch represents the LLVM side of the addition of two new intrinsic based instructions (vpopcntd and vpopcntq). Differential Revision: https://reviews.llvm.org/D33169 llvm-svn: 303858
2017-05-25 21:45:23 +08:00
; AVX512VPOPCNTDQ-LABEL: testv2i64u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; AVX512VPOPCNTDQ: # %bb.0:
[x86] transform vector inc/dec to use -1 constant (PR33483) Convert vector increment or decrement to sub/add with an all-ones constant: add X, <1, 1...> --> sub X, <-1, -1...> sub X, <1, 1...> --> add X, <-1, -1...> The all-ones vector constant can be materialized using a pcmpeq instruction that is commonly recognized as an idiom (has no register dependency), so that's better than loading a splat 1 constant. AVX512 uses 'vpternlogd' for 512-bit vectors because there is apparently no better way to produce 512 one-bits. The general advantages of this lowering are: 1. pcmpeq has lower latency than a memop on every uarch I looked at in Agner's tables, so in theory, this could be better for perf, but... 2. That seems unlikely to affect any OOO implementation, and I can't measure any real perf difference from this transform on Haswell or Jaguar, but... 3. It doesn't look like it from the diffs, but this is an overall size win because we eliminate 16 - 64 constant bytes in the case of a vector load. If we're broadcasting a scalar load (which might itself be a bug), then we're replacing a scalar constant load + broadcast with a single cheap op, so that should always be smaller/better too. 4. This makes the DAG/isel output more consistent - we use pcmpeq already for padd x, -1 and psub x, -1, so we should use that form for +1 too because we can. If there's some reason to favor a constant load on some CPU, let's make the reverse transform for all of these cases (either here in the DAG or in a later machine pass). This should fix: https://bugs.llvm.org/show_bug.cgi?id=33483 Differential Revision: https://reviews.llvm.org/D34336 llvm-svn: 306289
2017-06-26 22:19:26 +08:00
; AVX512VPOPCNTDQ-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; AVX512VPOPCNTDQ-NEXT: vpaddq %xmm1, %xmm0, %xmm1
; AVX512VPOPCNTDQ-NEXT: vpandn %xmm1, %xmm0, %xmm0
[X86] Adding vpopcntd and vpopcntq instructions AVX512_VPOPCNTDQ is a new feature set that was published by Intel. The patch represents the LLVM side of the addition of two new intrinsic based instructions (vpopcntd and vpopcntq). Differential Revision: https://reviews.llvm.org/D33169 llvm-svn: 303858
2017-05-25 21:45:23 +08:00
; AVX512VPOPCNTDQ-NEXT: vpopcntq %zmm0, %zmm0
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
; AVX512VPOPCNTDQ-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
[X86] Adding vpopcntd and vpopcntq instructions AVX512_VPOPCNTDQ is a new feature set that was published by Intel. The patch represents the LLVM side of the addition of two new intrinsic based instructions (vpopcntd and vpopcntq). Differential Revision: https://reviews.llvm.org/D33169 llvm-svn: 303858
2017-05-25 21:45:23 +08:00
; AVX512VPOPCNTDQ-NEXT: vzeroupper
; AVX512VPOPCNTDQ-NEXT: retq
;
[X86] Add 128 and 256-bit VPOPCNTDQ instructions. Adjust some tablegen classes LZCNT/POPCNT. I think when this instruction was first published it was only for a Knights CPU and thus VLX version was missing. llvm-svn: 320910
2017-12-16 10:40:28 +08:00
; AVX512VPOPCNTDQVL-LABEL: testv2i64u:
; AVX512VPOPCNTDQVL: # %bb.0:
; AVX512VPOPCNTDQVL-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; AVX512VPOPCNTDQVL-NEXT: vpaddq %xmm1, %xmm0, %xmm1
; AVX512VPOPCNTDQVL-NEXT: vpandn %xmm1, %xmm0, %xmm0
[X86] Add 128 and 256-bit VPOPCNTDQ instructions. Adjust some tablegen classes LZCNT/POPCNT. I think when this instruction was first published it was only for a Knights CPU and thus VLX version was missing. llvm-svn: 320910
2017-12-16 10:40:28 +08:00
; AVX512VPOPCNTDQVL-NEXT: vpopcntq %xmm0, %xmm0
; AVX512VPOPCNTDQVL-NEXT: retq
;
[X86] Regenerate the vector-popcnt and vector-tzcnt tests to get BITALG CHECK linse on all functions not just the vXi16/vXi8. llvm-svn: 318885
2017-11-23 07:35:12 +08:00
; BITALG_NOVLX-LABEL: testv2i64u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; BITALG_NOVLX: # %bb.0:
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; BITALG_NOVLX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
; BITALG_NOVLX-NEXT: vpaddq %xmm1, %xmm0, %xmm1
; BITALG_NOVLX-NEXT: vpandn %xmm1, %xmm0, %xmm0
[X86][SSE] LowerVectorCTPOP - pull out repeated byte sum stage. Pull out repeated byte sum stage for popcount of vector elements > 8bits. This allows us to simplify the LUT/BITMATH popcnt code to always assume vXi8 vectors, and also improves avx512bitalg codegen which only has access to vpopcntb/vpopcntw. llvm-svn: 344348
2018-10-12 22:18:47 +08:00
; BITALG_NOVLX-NEXT: vpopcntb %zmm0, %zmm0
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; BITALG_NOVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1
[X86] Regenerate the vector-popcnt and vector-tzcnt tests to get BITALG CHECK linse on all functions not just the vXi16/vXi8. llvm-svn: 318885
2017-11-23 07:35:12 +08:00
; BITALG_NOVLX-NEXT: vpsadbw %xmm1, %xmm0, %xmm0
[X86][SSE] LowerVectorCTPOP - pull out repeated byte sum stage. Pull out repeated byte sum stage for popcount of vector elements > 8bits. This allows us to simplify the LUT/BITMATH popcnt code to always assume vXi8 vectors, and also improves avx512bitalg codegen which only has access to vpopcntb/vpopcntw. llvm-svn: 344348
2018-10-12 22:18:47 +08:00
; BITALG_NOVLX-NEXT: vzeroupper
[X86] Regenerate the vector-popcnt and vector-tzcnt tests to get BITALG CHECK linse on all functions not just the vXi16/vXi8. llvm-svn: 318885
2017-11-23 07:35:12 +08:00
; BITALG_NOVLX-NEXT: retq
;
; BITALG-LABEL: testv2i64u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; BITALG: # %bb.0:
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; BITALG-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
; BITALG-NEXT: vpaddq %xmm1, %xmm0, %xmm1
; BITALG-NEXT: vpandn %xmm1, %xmm0, %xmm0
[X86][SSE] LowerVectorCTPOP - pull out repeated byte sum stage. Pull out repeated byte sum stage for popcount of vector elements > 8bits. This allows us to simplify the LUT/BITMATH popcnt code to always assume vXi8 vectors, and also improves avx512bitalg codegen which only has access to vpopcntb/vpopcntw. llvm-svn: 344348
2018-10-12 22:18:47 +08:00
; BITALG-NEXT: vpopcntb %xmm0, %xmm0
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; BITALG-NEXT: vpxor %xmm1, %xmm1, %xmm1
[X86] Regenerate the vector-popcnt and vector-tzcnt tests to get BITALG CHECK linse on all functions not just the vXi16/vXi8. llvm-svn: 318885
2017-11-23 07:35:12 +08:00
; BITALG-NEXT: vpsadbw %xmm1, %xmm0, %xmm0
; BITALG-NEXT: retq
;
[X86][SSE] Added 32-bit tests for vector lzcnt/tzcnt tests v2i64 tests are particularly bad on 32-bit targets. llvm-svn: 265789
2016-04-08 23:01:31 +08:00
; X32-SSE-LABEL: testv2i64u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; X32-SSE: # %bb.0:
[X86][SSE] combineIncDecVector - use isConstantSplat Use isConstantSplat instead of ISD::isConstantSplatVector to let us us peek through to illegal types (in this case for i686 targets to recognise i64 constants) llvm-svn: 344452
2018-10-13 22:45:44 +08:00
; X32-SSE-NEXT: pcmpeqd %xmm1, %xmm1
; X32-SSE-NEXT: paddq %xmm0, %xmm1
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; X32-SSE-NEXT: pandn %xmm1, %xmm0
; X32-SSE-NEXT: movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
; X32-SSE-NEXT: movdqa %xmm0, %xmm2
; X32-SSE-NEXT: pand %xmm1, %xmm2
; X32-SSE-NEXT: movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
; X32-SSE-NEXT: movdqa %xmm3, %xmm4
; X32-SSE-NEXT: pshufb %xmm2, %xmm4
; X32-SSE-NEXT: psrlw $4, %xmm0
; X32-SSE-NEXT: pand %xmm1, %xmm0
; X32-SSE-NEXT: pshufb %xmm0, %xmm3
; X32-SSE-NEXT: paddb %xmm4, %xmm3
; X32-SSE-NEXT: pxor %xmm0, %xmm0
; X32-SSE-NEXT: psadbw %xmm3, %xmm0
[X86][SSE] Added 32-bit tests for vector lzcnt/tzcnt tests v2i64 tests are particularly bad on 32-bit targets. llvm-svn: 265789
2016-04-08 23:01:31 +08:00
; X32-SSE-NEXT: retl
[X86] Added tzcnt vector tests. llvm-svn: 239264
2015-06-08 05:01:34 +08:00
%out = call <2 x i64> @llvm.cttz.v2i64(<2 x i64> %in, i1 -1)
ret <2 x i64> %out
}
[X86][SSE] Tidied up vector CTLZ/CTTZ. NFCI. llvm-svn: 242645
2015-07-20 01:09:43 +08:00
define <4 x i32> @testv4i32(<4 x i32> %in) nounwind {
[X86] Added tzcnt vector tests. llvm-svn: 239264
2015-06-08 05:01:34 +08:00
; SSE2-LABEL: testv4i32:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; SSE2: # %bb.0:
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; SSE2-NEXT: pcmpeqd %xmm1, %xmm1
; SSE2-NEXT: paddd %xmm0, %xmm1
; SSE2-NEXT: pandn %xmm1, %xmm0
; SSE2-NEXT: movdqa %xmm0, %xmm1
; SSE2-NEXT: psrlw $1, %xmm1
; SSE2-NEXT: pand {{.*}}(%rip), %xmm1
; SSE2-NEXT: psubb %xmm1, %xmm0
; SSE2-NEXT: movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
[X86][SSE] Vectorize CTTZ + CTTZ_ZERO_UNDEF Now that we have fast vector CTPOP implementations we can use this to speed up vector CTTZ using the pattern (cttz(x) = ctpop((x & -x) - 1)) Additionally, for AVX512CD that provides lzcnt instructions we can use the pattern (cttz_undef(x) = (width - 1) - ctlz(x & -x)) Differential Revision: http://reviews.llvm.org/D12663 llvm-svn: 248091
2015-09-19 21:22:57 +08:00
; SSE2-NEXT: movdqa %xmm0, %xmm2
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; SSE2-NEXT: pand %xmm1, %xmm2
; SSE2-NEXT: psrlw $2, %xmm0
; SSE2-NEXT: pand %xmm1, %xmm0
; SSE2-NEXT: paddb %xmm2, %xmm0
; SSE2-NEXT: movdqa %xmm0, %xmm1
; SSE2-NEXT: psrlw $4, %xmm1
; SSE2-NEXT: paddb %xmm0, %xmm1
; SSE2-NEXT: pand {{.*}}(%rip), %xmm1
; SSE2-NEXT: pxor %xmm0, %xmm0
; SSE2-NEXT: movdqa %xmm1, %xmm2
; SSE2-NEXT: punpckhdq {{.*#+}} xmm2 = xmm2[2],xmm0[2],xmm2[3],xmm0[3]
; SSE2-NEXT: psadbw %xmm0, %xmm2
; SSE2-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
; SSE2-NEXT: psadbw %xmm0, %xmm1
; SSE2-NEXT: packuswb %xmm2, %xmm1
; SSE2-NEXT: movdqa %xmm1, %xmm0
[X86] Added tzcnt vector tests. llvm-svn: 239264
2015-06-08 05:01:34 +08:00
; SSE2-NEXT: retq
;
; SSE3-LABEL: testv4i32:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; SSE3: # %bb.0:
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; SSE3-NEXT: pcmpeqd %xmm1, %xmm1
; SSE3-NEXT: paddd %xmm0, %xmm1
; SSE3-NEXT: pandn %xmm1, %xmm0
; SSE3-NEXT: movdqa %xmm0, %xmm1
; SSE3-NEXT: psrlw $1, %xmm1
; SSE3-NEXT: pand {{.*}}(%rip), %xmm1
; SSE3-NEXT: psubb %xmm1, %xmm0
; SSE3-NEXT: movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
[X86][SSE] Vectorize CTTZ + CTTZ_ZERO_UNDEF Now that we have fast vector CTPOP implementations we can use this to speed up vector CTTZ using the pattern (cttz(x) = ctpop((x & -x) - 1)) Additionally, for AVX512CD that provides lzcnt instructions we can use the pattern (cttz_undef(x) = (width - 1) - ctlz(x & -x)) Differential Revision: http://reviews.llvm.org/D12663 llvm-svn: 248091
2015-09-19 21:22:57 +08:00
; SSE3-NEXT: movdqa %xmm0, %xmm2
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; SSE3-NEXT: pand %xmm1, %xmm2
; SSE3-NEXT: psrlw $2, %xmm0
; SSE3-NEXT: pand %xmm1, %xmm0
; SSE3-NEXT: paddb %xmm2, %xmm0
; SSE3-NEXT: movdqa %xmm0, %xmm1
; SSE3-NEXT: psrlw $4, %xmm1
; SSE3-NEXT: paddb %xmm0, %xmm1
; SSE3-NEXT: pand {{.*}}(%rip), %xmm1
; SSE3-NEXT: pxor %xmm0, %xmm0
; SSE3-NEXT: movdqa %xmm1, %xmm2
; SSE3-NEXT: punpckhdq {{.*#+}} xmm2 = xmm2[2],xmm0[2],xmm2[3],xmm0[3]
; SSE3-NEXT: psadbw %xmm0, %xmm2
; SSE3-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
; SSE3-NEXT: psadbw %xmm0, %xmm1
; SSE3-NEXT: packuswb %xmm2, %xmm1
; SSE3-NEXT: movdqa %xmm1, %xmm0
[X86] Added tzcnt vector tests. llvm-svn: 239264
2015-06-08 05:01:34 +08:00
; SSE3-NEXT: retq
;
; SSSE3-LABEL: testv4i32:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; SSSE3: # %bb.0:
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; SSSE3-NEXT: pcmpeqd %xmm1, %xmm1
; SSSE3-NEXT: paddd %xmm0, %xmm1
; SSSE3-NEXT: pandn %xmm1, %xmm0
[x86] transform vector inc/dec to use -1 constant (PR33483) Convert vector increment or decrement to sub/add with an all-ones constant: add X, <1, 1...> --> sub X, <-1, -1...> sub X, <1, 1...> --> add X, <-1, -1...> The all-ones vector constant can be materialized using a pcmpeq instruction that is commonly recognized as an idiom (has no register dependency), so that's better than loading a splat 1 constant. AVX512 uses 'vpternlogd' for 512-bit vectors because there is apparently no better way to produce 512 one-bits. The general advantages of this lowering are: 1. pcmpeq has lower latency than a memop on every uarch I looked at in Agner's tables, so in theory, this could be better for perf, but... 2. That seems unlikely to affect any OOO implementation, and I can't measure any real perf difference from this transform on Haswell or Jaguar, but... 3. It doesn't look like it from the diffs, but this is an overall size win because we eliminate 16 - 64 constant bytes in the case of a vector load. If we're broadcasting a scalar load (which might itself be a bug), then we're replacing a scalar constant load + broadcast with a single cheap op, so that should always be smaller/better too. 4. This makes the DAG/isel output more consistent - we use pcmpeq already for padd x, -1 and psub x, -1, so we should use that form for +1 too because we can. If there's some reason to favor a constant load on some CPU, let's make the reverse transform for all of these cases (either here in the DAG or in a later machine pass). This should fix: https://bugs.llvm.org/show_bug.cgi?id=33483 Differential Revision: https://reviews.llvm.org/D34336 llvm-svn: 306289
2017-06-26 22:19:26 +08:00
; SSSE3-NEXT: movdqa {{.*#+}} xmm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; SSSE3-NEXT: movdqa %xmm0, %xmm3
[x86] transform vector inc/dec to use -1 constant (PR33483) Convert vector increment or decrement to sub/add with an all-ones constant: add X, <1, 1...> --> sub X, <-1, -1...> sub X, <1, 1...> --> add X, <-1, -1...> The all-ones vector constant can be materialized using a pcmpeq instruction that is commonly recognized as an idiom (has no register dependency), so that's better than loading a splat 1 constant. AVX512 uses 'vpternlogd' for 512-bit vectors because there is apparently no better way to produce 512 one-bits. The general advantages of this lowering are: 1. pcmpeq has lower latency than a memop on every uarch I looked at in Agner's tables, so in theory, this could be better for perf, but... 2. That seems unlikely to affect any OOO implementation, and I can't measure any real perf difference from this transform on Haswell or Jaguar, but... 3. It doesn't look like it from the diffs, but this is an overall size win because we eliminate 16 - 64 constant bytes in the case of a vector load. If we're broadcasting a scalar load (which might itself be a bug), then we're replacing a scalar constant load + broadcast with a single cheap op, so that should always be smaller/better too. 4. This makes the DAG/isel output more consistent - we use pcmpeq already for padd x, -1 and psub x, -1, so we should use that form for +1 too because we can. If there's some reason to favor a constant load on some CPU, let's make the reverse transform for all of these cases (either here in the DAG or in a later machine pass). This should fix: https://bugs.llvm.org/show_bug.cgi?id=33483 Differential Revision: https://reviews.llvm.org/D34336 llvm-svn: 306289
2017-06-26 22:19:26 +08:00
; SSSE3-NEXT: pand %xmm2, %xmm3
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; SSSE3-NEXT: movdqa {{.*#+}} xmm1 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
; SSSE3-NEXT: movdqa %xmm1, %xmm4
; SSSE3-NEXT: pshufb %xmm3, %xmm4
; SSSE3-NEXT: psrlw $4, %xmm0
; SSSE3-NEXT: pand %xmm2, %xmm0
; SSSE3-NEXT: pshufb %xmm0, %xmm1
; SSSE3-NEXT: paddb %xmm4, %xmm1
; SSSE3-NEXT: pxor %xmm0, %xmm0
; SSSE3-NEXT: movdqa %xmm1, %xmm2
; SSSE3-NEXT: punpckhdq {{.*#+}} xmm2 = xmm2[2],xmm0[2],xmm2[3],xmm0[3]
; SSSE3-NEXT: psadbw %xmm0, %xmm2
; SSSE3-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
; SSSE3-NEXT: psadbw %xmm0, %xmm1
; SSSE3-NEXT: packuswb %xmm2, %xmm1
; SSSE3-NEXT: movdqa %xmm1, %xmm0
[X86] Added tzcnt vector tests. llvm-svn: 239264
2015-06-08 05:01:34 +08:00
; SSSE3-NEXT: retq
;
; SSE41-LABEL: testv4i32:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; SSE41: # %bb.0:
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; SSE41-NEXT: pcmpeqd %xmm1, %xmm1
; SSE41-NEXT: paddd %xmm0, %xmm1
; SSE41-NEXT: pandn %xmm1, %xmm0
; SSE41-NEXT: movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
; SSE41-NEXT: movdqa %xmm0, %xmm2
; SSE41-NEXT: pand %xmm1, %xmm2
; SSE41-NEXT: movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
; SSE41-NEXT: movdqa %xmm3, %xmm4
; SSE41-NEXT: pshufb %xmm2, %xmm4
[x86] transform vector inc/dec to use -1 constant (PR33483) Convert vector increment or decrement to sub/add with an all-ones constant: add X, <1, 1...> --> sub X, <-1, -1...> sub X, <1, 1...> --> add X, <-1, -1...> The all-ones vector constant can be materialized using a pcmpeq instruction that is commonly recognized as an idiom (has no register dependency), so that's better than loading a splat 1 constant. AVX512 uses 'vpternlogd' for 512-bit vectors because there is apparently no better way to produce 512 one-bits. The general advantages of this lowering are: 1. pcmpeq has lower latency than a memop on every uarch I looked at in Agner's tables, so in theory, this could be better for perf, but... 2. That seems unlikely to affect any OOO implementation, and I can't measure any real perf difference from this transform on Haswell or Jaguar, but... 3. It doesn't look like it from the diffs, but this is an overall size win because we eliminate 16 - 64 constant bytes in the case of a vector load. If we're broadcasting a scalar load (which might itself be a bug), then we're replacing a scalar constant load + broadcast with a single cheap op, so that should always be smaller/better too. 4. This makes the DAG/isel output more consistent - we use pcmpeq already for padd x, -1 and psub x, -1, so we should use that form for +1 too because we can. If there's some reason to favor a constant load on some CPU, let's make the reverse transform for all of these cases (either here in the DAG or in a later machine pass). This should fix: https://bugs.llvm.org/show_bug.cgi?id=33483 Differential Revision: https://reviews.llvm.org/D34336 llvm-svn: 306289
2017-06-26 22:19:26 +08:00
; SSE41-NEXT: psrlw $4, %xmm0
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; SSE41-NEXT: pand %xmm1, %xmm0
; SSE41-NEXT: pshufb %xmm0, %xmm3
; SSE41-NEXT: paddb %xmm4, %xmm3
; SSE41-NEXT: pxor %xmm1, %xmm1
; SSE41-NEXT: pmovzxdq {{.*#+}} xmm0 = xmm3[0],zero,xmm3[1],zero
; SSE41-NEXT: punpckhdq {{.*#+}} xmm3 = xmm3[2],xmm1[2],xmm3[3],xmm1[3]
; SSE41-NEXT: psadbw %xmm1, %xmm3
[X86][SSE] Vectorize CTTZ + CTTZ_ZERO_UNDEF Now that we have fast vector CTPOP implementations we can use this to speed up vector CTTZ using the pattern (cttz(x) = ctpop((x & -x) - 1)) Additionally, for AVX512CD that provides lzcnt instructions we can use the pattern (cttz_undef(x) = (width - 1) - ctlz(x & -x)) Differential Revision: http://reviews.llvm.org/D12663 llvm-svn: 248091
2015-09-19 21:22:57 +08:00
; SSE41-NEXT: psadbw %xmm1, %xmm0
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; SSE41-NEXT: packuswb %xmm3, %xmm0
[X86] Added tzcnt vector tests. llvm-svn: 239264
2015-06-08 05:01:34 +08:00
; SSE41-NEXT: retq
;
[X86][SSE] Vectorize CTTZ + CTTZ_ZERO_UNDEF Now that we have fast vector CTPOP implementations we can use this to speed up vector CTTZ using the pattern (cttz(x) = ctpop((x & -x) - 1)) Additionally, for AVX512CD that provides lzcnt instructions we can use the pattern (cttz_undef(x) = (width - 1) - ctlz(x & -x)) Differential Revision: http://reviews.llvm.org/D12663 llvm-svn: 248091
2015-09-19 21:22:57 +08:00
; AVX1-LABEL: testv4i32:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; AVX1: # %bb.0:
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; AVX1-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
; AVX1-NEXT: vpaddd %xmm1, %xmm0, %xmm1
; AVX1-NEXT: vpandn %xmm1, %xmm0, %xmm0
; AVX1-NEXT: vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
; AVX1-NEXT: vpand %xmm1, %xmm0, %xmm2
; AVX1-NEXT: vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
; AVX1-NEXT: vpshufb %xmm2, %xmm3, %xmm2
[X86][SSE] Vectorize CTTZ + CTTZ_ZERO_UNDEF Now that we have fast vector CTPOP implementations we can use this to speed up vector CTTZ using the pattern (cttz(x) = ctpop((x & -x) - 1)) Additionally, for AVX512CD that provides lzcnt instructions we can use the pattern (cttz_undef(x) = (width - 1) - ctlz(x & -x)) Differential Revision: http://reviews.llvm.org/D12663 llvm-svn: 248091
2015-09-19 21:22:57 +08:00
; AVX1-NEXT: vpsrlw $4, %xmm0, %xmm0
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; AVX1-NEXT: vpand %xmm1, %xmm0, %xmm0
; AVX1-NEXT: vpshufb %xmm0, %xmm3, %xmm0
; AVX1-NEXT: vpaddb %xmm2, %xmm0, %xmm0
; AVX1-NEXT: vpxor %xmm1, %xmm1, %xmm1
[X86][SSE] Vectorize CTTZ + CTTZ_ZERO_UNDEF Now that we have fast vector CTPOP implementations we can use this to speed up vector CTTZ using the pattern (cttz(x) = ctpop((x & -x) - 1)) Additionally, for AVX512CD that provides lzcnt instructions we can use the pattern (cttz_undef(x) = (width - 1) - ctlz(x & -x)) Differential Revision: http://reviews.llvm.org/D12663 llvm-svn: 248091
2015-09-19 21:22:57 +08:00
; AVX1-NEXT: vpunpckhdq {{.*#+}} xmm2 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]
[X86] Fix several issues related to X86's psadbw instruction. This patch fixes the following issues: 1. Fix the return type of X86psadbw: it should not be the same type of inputs. For vNi8 inputs the output should be vMi64, where M = N/8. 2. Fix the return type of int_x86_avx512_psad_bw_512 accordingly. 3. Fix the definiton of PSADBW, VPSADBW, and VPSADBWY accordingly. 4. Adjust the return type when building a DAG node of X86ISD::PSADBW type. 5. Update related tests. Differential revision: http://reviews.llvm.org/D14897 llvm-svn: 254010
2015-11-25 03:51:26 +08:00
; AVX1-NEXT: vpsadbw %xmm1, %xmm2, %xmm2
[X86][SSE] Add initial support for combining target shuffles to (V)PMOVZX. We can only handle 128-bit vectors until we support target shuffle inputs of different size to the output. llvm-svn: 288140
2016-11-29 22:18:51 +08:00
; AVX1-NEXT: vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
[X86] Fix several issues related to X86's psadbw instruction. This patch fixes the following issues: 1. Fix the return type of X86psadbw: it should not be the same type of inputs. For vNi8 inputs the output should be vMi64, where M = N/8. 2. Fix the return type of int_x86_avx512_psad_bw_512 accordingly. 3. Fix the definiton of PSADBW, VPSADBW, and VPSADBWY accordingly. 4. Adjust the return type when building a DAG node of X86ISD::PSADBW type. 5. Update related tests. Differential revision: http://reviews.llvm.org/D14897 llvm-svn: 254010
2015-11-25 03:51:26 +08:00
; AVX1-NEXT: vpsadbw %xmm1, %xmm0, %xmm0
[X86][SSE] Vectorize CTTZ + CTTZ_ZERO_UNDEF Now that we have fast vector CTPOP implementations we can use this to speed up vector CTTZ using the pattern (cttz(x) = ctpop((x & -x) - 1)) Additionally, for AVX512CD that provides lzcnt instructions we can use the pattern (cttz_undef(x) = (width - 1) - ctlz(x & -x)) Differential Revision: http://reviews.llvm.org/D12663 llvm-svn: 248091
2015-09-19 21:22:57 +08:00
; AVX1-NEXT: vpackuswb %xmm2, %xmm0, %xmm0
; AVX1-NEXT: retq
;
; AVX2-LABEL: testv4i32:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; AVX2: # %bb.0:
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; AVX2-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
; AVX2-NEXT: vpaddd %xmm1, %xmm0, %xmm1
; AVX2-NEXT: vpandn %xmm1, %xmm0, %xmm0
; AVX2-NEXT: vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
; AVX2-NEXT: vpand %xmm1, %xmm0, %xmm2
; AVX2-NEXT: vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
; AVX2-NEXT: vpshufb %xmm2, %xmm3, %xmm2
[X86][SSE] Vectorize CTTZ + CTTZ_ZERO_UNDEF Now that we have fast vector CTPOP implementations we can use this to speed up vector CTTZ using the pattern (cttz(x) = ctpop((x & -x) - 1)) Additionally, for AVX512CD that provides lzcnt instructions we can use the pattern (cttz_undef(x) = (width - 1) - ctlz(x & -x)) Differential Revision: http://reviews.llvm.org/D12663 llvm-svn: 248091
2015-09-19 21:22:57 +08:00
; AVX2-NEXT: vpsrlw $4, %xmm0, %xmm0
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; AVX2-NEXT: vpand %xmm1, %xmm0, %xmm0
; AVX2-NEXT: vpshufb %xmm0, %xmm3, %xmm0
; AVX2-NEXT: vpaddb %xmm2, %xmm0, %xmm0
; AVX2-NEXT: vpxor %xmm1, %xmm1, %xmm1
[X86][SSE] Vectorize CTTZ + CTTZ_ZERO_UNDEF Now that we have fast vector CTPOP implementations we can use this to speed up vector CTTZ using the pattern (cttz(x) = ctpop((x & -x) - 1)) Additionally, for AVX512CD that provides lzcnt instructions we can use the pattern (cttz_undef(x) = (width - 1) - ctlz(x & -x)) Differential Revision: http://reviews.llvm.org/D12663 llvm-svn: 248091
2015-09-19 21:22:57 +08:00
; AVX2-NEXT: vpunpckhdq {{.*#+}} xmm2 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]
[X86] Fix several issues related to X86's psadbw instruction. This patch fixes the following issues: 1. Fix the return type of X86psadbw: it should not be the same type of inputs. For vNi8 inputs the output should be vMi64, where M = N/8. 2. Fix the return type of int_x86_avx512_psad_bw_512 accordingly. 3. Fix the definiton of PSADBW, VPSADBW, and VPSADBWY accordingly. 4. Adjust the return type when building a DAG node of X86ISD::PSADBW type. 5. Update related tests. Differential revision: http://reviews.llvm.org/D14897 llvm-svn: 254010
2015-11-25 03:51:26 +08:00
; AVX2-NEXT: vpsadbw %xmm1, %xmm2, %xmm2
[X86][SSE] Add initial support for combining target shuffles to (V)PMOVZX. We can only handle 128-bit vectors until we support target shuffle inputs of different size to the output. llvm-svn: 288140
2016-11-29 22:18:51 +08:00
; AVX2-NEXT: vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
[X86] Fix several issues related to X86's psadbw instruction. This patch fixes the following issues: 1. Fix the return type of X86psadbw: it should not be the same type of inputs. For vNi8 inputs the output should be vMi64, where M = N/8. 2. Fix the return type of int_x86_avx512_psad_bw_512 accordingly. 3. Fix the definiton of PSADBW, VPSADBW, and VPSADBWY accordingly. 4. Adjust the return type when building a DAG node of X86ISD::PSADBW type. 5. Update related tests. Differential revision: http://reviews.llvm.org/D14897 llvm-svn: 254010
2015-11-25 03:51:26 +08:00
; AVX2-NEXT: vpsadbw %xmm1, %xmm0, %xmm0
[X86][SSE] Vectorize CTTZ + CTTZ_ZERO_UNDEF Now that we have fast vector CTPOP implementations we can use this to speed up vector CTTZ using the pattern (cttz(x) = ctpop((x & -x) - 1)) Additionally, for AVX512CD that provides lzcnt instructions we can use the pattern (cttz_undef(x) = (width - 1) - ctlz(x & -x)) Differential Revision: http://reviews.llvm.org/D12663 llvm-svn: 248091
2015-09-19 21:22:57 +08:00
; AVX2-NEXT: vpackuswb %xmm2, %xmm0, %xmm0
; AVX2-NEXT: retq
[X86][SSE] Added 32-bit tests for vector lzcnt/tzcnt tests v2i64 tests are particularly bad on 32-bit targets. llvm-svn: 265789
2016-04-08 23:01:31 +08:00
;
[AVX512] Add avx512cd+vl runs to vector-tzcnt-128/256 tests to show using the vplzcntd/q instructions. llvm-svn: 266860
2016-04-20 13:19:01 +08:00
; AVX512CDVL-LABEL: testv4i32:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; AVX512CDVL: # %bb.0:
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; AVX512CDVL-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
; AVX512CDVL-NEXT: vpaddd %xmm1, %xmm0, %xmm1
; AVX512CDVL-NEXT: vpandn %xmm1, %xmm0, %xmm0
[X86][SSE] Improve CTTZ lowering when CTLZ is legal If we have better CTLZ support than CTPOP, then use cttz(x) = width - ctlz(~x & (x - 1)) - and remove the CTTZ_ZERO_UNDEF handling as it no longer gives better codegen. Similar to rL344447, this is also closer to LegalizeDAG's approach llvm-svn: 344448
2018-10-13 21:05:19 +08:00
; AVX512CDVL-NEXT: vplzcntd %xmm0, %xmm0
; AVX512CDVL-NEXT: vpbroadcastd {{.*#+}} xmm1 = [32,32,32,32]
; AVX512CDVL-NEXT: vpsubd %xmm0, %xmm1, %xmm0
[AVX512] Add avx512cd+vl runs to vector-tzcnt-128/256 tests to show using the vplzcntd/q instructions. llvm-svn: 266860
2016-04-20 13:19:01 +08:00
; AVX512CDVL-NEXT: retq
;
[AVX512] Teach lowering to use vplzcntd/q to implement 128/256-bit CTTZ_ZERO_UNDEF even without VLX support. We can just extend to 512-bits and extract like we do for CTLZ. llvm-svn: 267100
2016-04-22 11:22:38 +08:00
; AVX512CD-LABEL: testv4i32:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; AVX512CD: # %bb.0:
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; AVX512CD-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
; AVX512CD-NEXT: vpaddd %xmm1, %xmm0, %xmm1
; AVX512CD-NEXT: vpandn %xmm1, %xmm0, %xmm0
[X86][SSE] Improve CTTZ lowering when CTLZ is legal If we have better CTLZ support than CTPOP, then use cttz(x) = width - ctlz(~x & (x - 1)) - and remove the CTTZ_ZERO_UNDEF handling as it no longer gives better codegen. Similar to rL344447, this is also closer to LegalizeDAG's approach llvm-svn: 344448
2018-10-13 21:05:19 +08:00
; AVX512CD-NEXT: vplzcntd %zmm0, %zmm0
; AVX512CD-NEXT: vpbroadcastd {{.*#+}} xmm1 = [32,32,32,32]
; AVX512CD-NEXT: vpsubd %xmm0, %xmm1, %xmm0
; AVX512CD-NEXT: vzeroupper
[AVX512] Teach lowering to use vplzcntd/q to implement 128/256-bit CTTZ_ZERO_UNDEF even without VLX support. We can just extend to 512-bits and extract like we do for CTLZ. llvm-svn: 267100
2016-04-22 11:22:38 +08:00
; AVX512CD-NEXT: retq
;
[X86] Adding vpopcntd and vpopcntq instructions AVX512_VPOPCNTDQ is a new feature set that was published by Intel. The patch represents the LLVM side of the addition of two new intrinsic based instructions (vpopcntd and vpopcntq). Differential Revision: https://reviews.llvm.org/D33169 llvm-svn: 303858
2017-05-25 21:45:23 +08:00
; AVX512VPOPCNTDQ-LABEL: testv4i32:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; AVX512VPOPCNTDQ: # %bb.0:
[x86] transform vector inc/dec to use -1 constant (PR33483) Convert vector increment or decrement to sub/add with an all-ones constant: add X, <1, 1...> --> sub X, <-1, -1...> sub X, <1, 1...> --> add X, <-1, -1...> The all-ones vector constant can be materialized using a pcmpeq instruction that is commonly recognized as an idiom (has no register dependency), so that's better than loading a splat 1 constant. AVX512 uses 'vpternlogd' for 512-bit vectors because there is apparently no better way to produce 512 one-bits. The general advantages of this lowering are: 1. pcmpeq has lower latency than a memop on every uarch I looked at in Agner's tables, so in theory, this could be better for perf, but... 2. That seems unlikely to affect any OOO implementation, and I can't measure any real perf difference from this transform on Haswell or Jaguar, but... 3. It doesn't look like it from the diffs, but this is an overall size win because we eliminate 16 - 64 constant bytes in the case of a vector load. If we're broadcasting a scalar load (which might itself be a bug), then we're replacing a scalar constant load + broadcast with a single cheap op, so that should always be smaller/better too. 4. This makes the DAG/isel output more consistent - we use pcmpeq already for padd x, -1 and psub x, -1, so we should use that form for +1 too because we can. If there's some reason to favor a constant load on some CPU, let's make the reverse transform for all of these cases (either here in the DAG or in a later machine pass). This should fix: https://bugs.llvm.org/show_bug.cgi?id=33483 Differential Revision: https://reviews.llvm.org/D34336 llvm-svn: 306289
2017-06-26 22:19:26 +08:00
; AVX512VPOPCNTDQ-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; AVX512VPOPCNTDQ-NEXT: vpaddd %xmm1, %xmm0, %xmm1
; AVX512VPOPCNTDQ-NEXT: vpandn %xmm1, %xmm0, %xmm0
[X86] Adding vpopcntd and vpopcntq instructions AVX512_VPOPCNTDQ is a new feature set that was published by Intel. The patch represents the LLVM side of the addition of two new intrinsic based instructions (vpopcntd and vpopcntq). Differential Revision: https://reviews.llvm.org/D33169 llvm-svn: 303858
2017-05-25 21:45:23 +08:00
; AVX512VPOPCNTDQ-NEXT: vpopcntd %zmm0, %zmm0
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
; AVX512VPOPCNTDQ-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
[X86] Adding vpopcntd and vpopcntq instructions AVX512_VPOPCNTDQ is a new feature set that was published by Intel. The patch represents the LLVM side of the addition of two new intrinsic based instructions (vpopcntd and vpopcntq). Differential Revision: https://reviews.llvm.org/D33169 llvm-svn: 303858
2017-05-25 21:45:23 +08:00
; AVX512VPOPCNTDQ-NEXT: vzeroupper
; AVX512VPOPCNTDQ-NEXT: retq
;
[X86] Add 128 and 256-bit VPOPCNTDQ instructions. Adjust some tablegen classes LZCNT/POPCNT. I think when this instruction was first published it was only for a Knights CPU and thus VLX version was missing. llvm-svn: 320910
2017-12-16 10:40:28 +08:00
; AVX512VPOPCNTDQVL-LABEL: testv4i32:
; AVX512VPOPCNTDQVL: # %bb.0:
; AVX512VPOPCNTDQVL-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; AVX512VPOPCNTDQVL-NEXT: vpaddd %xmm1, %xmm0, %xmm1
; AVX512VPOPCNTDQVL-NEXT: vpandn %xmm1, %xmm0, %xmm0
[X86] Add 128 and 256-bit VPOPCNTDQ instructions. Adjust some tablegen classes LZCNT/POPCNT. I think when this instruction was first published it was only for a Knights CPU and thus VLX version was missing. llvm-svn: 320910
2017-12-16 10:40:28 +08:00
; AVX512VPOPCNTDQVL-NEXT: vpopcntd %xmm0, %xmm0
; AVX512VPOPCNTDQVL-NEXT: retq
;
[X86] Regenerate the vector-popcnt and vector-tzcnt tests to get BITALG CHECK linse on all functions not just the vXi16/vXi8. llvm-svn: 318885
2017-11-23 07:35:12 +08:00
; BITALG_NOVLX-LABEL: testv4i32:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; BITALG_NOVLX: # %bb.0:
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; BITALG_NOVLX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
; BITALG_NOVLX-NEXT: vpaddd %xmm1, %xmm0, %xmm1
; BITALG_NOVLX-NEXT: vpandn %xmm1, %xmm0, %xmm0
[X86][SSE] LowerVectorCTPOP - pull out repeated byte sum stage. Pull out repeated byte sum stage for popcount of vector elements > 8bits. This allows us to simplify the LUT/BITMATH popcnt code to always assume vXi8 vectors, and also improves avx512bitalg codegen which only has access to vpopcntb/vpopcntw. llvm-svn: 344348
2018-10-12 22:18:47 +08:00
; BITALG_NOVLX-NEXT: vpopcntb %zmm0, %zmm0
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; BITALG_NOVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1
[X86] Regenerate the vector-popcnt and vector-tzcnt tests to get BITALG CHECK linse on all functions not just the vXi16/vXi8. llvm-svn: 318885
2017-11-23 07:35:12 +08:00
; BITALG_NOVLX-NEXT: vpunpckhdq {{.*#+}} xmm2 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]
; BITALG_NOVLX-NEXT: vpsadbw %xmm1, %xmm2, %xmm2
; BITALG_NOVLX-NEXT: vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
; BITALG_NOVLX-NEXT: vpsadbw %xmm1, %xmm0, %xmm0
; BITALG_NOVLX-NEXT: vpackuswb %xmm2, %xmm0, %xmm0
[X86][SSE] LowerVectorCTPOP - pull out repeated byte sum stage. Pull out repeated byte sum stage for popcount of vector elements > 8bits. This allows us to simplify the LUT/BITMATH popcnt code to always assume vXi8 vectors, and also improves avx512bitalg codegen which only has access to vpopcntb/vpopcntw. llvm-svn: 344348
2018-10-12 22:18:47 +08:00
; BITALG_NOVLX-NEXT: vzeroupper
[X86] Regenerate the vector-popcnt and vector-tzcnt tests to get BITALG CHECK linse on all functions not just the vXi16/vXi8. llvm-svn: 318885
2017-11-23 07:35:12 +08:00
; BITALG_NOVLX-NEXT: retq
;
; BITALG-LABEL: testv4i32:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; BITALG: # %bb.0:
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; BITALG-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
; BITALG-NEXT: vpaddd %xmm1, %xmm0, %xmm1
; BITALG-NEXT: vpandn %xmm1, %xmm0, %xmm0
[X86][SSE] LowerVectorCTPOP - pull out repeated byte sum stage. Pull out repeated byte sum stage for popcount of vector elements > 8bits. This allows us to simplify the LUT/BITMATH popcnt code to always assume vXi8 vectors, and also improves avx512bitalg codegen which only has access to vpopcntb/vpopcntw. llvm-svn: 344348
2018-10-12 22:18:47 +08:00
; BITALG-NEXT: vpopcntb %xmm0, %xmm0
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; BITALG-NEXT: vpxor %xmm1, %xmm1, %xmm1
[X86] Regenerate the vector-popcnt and vector-tzcnt tests to get BITALG CHECK linse on all functions not just the vXi16/vXi8. llvm-svn: 318885
2017-11-23 07:35:12 +08:00
; BITALG-NEXT: vpunpckhdq {{.*#+}} xmm2 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]
; BITALG-NEXT: vpsadbw %xmm1, %xmm2, %xmm2
[X86] In LowerHorizontalByteSum, emit vector_shuffle nodes instead of directly using X86ISD::UNPCKL/X86ISD::UNPCKH. This gives shuffle lowering the freedom to use zero_extend_vector_inreg for the unpckl shuffle. Shuffle combining usually makes this swap later, but not when AVX512 is enabled it seems. While there also use DAG.getConstant to create a 0 vector instead of using the helper the forces a specific BUILD_VECTOR. I don't think that helper is usually needed. We're basically free to create a constant build_vector anytime and it will be legalized on its own. llvm-svn: 346574
2018-11-10 08:26:42 +08:00
; BITALG-NEXT: vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
[X86] Regenerate the vector-popcnt and vector-tzcnt tests to get BITALG CHECK linse on all functions not just the vXi16/vXi8. llvm-svn: 318885
2017-11-23 07:35:12 +08:00
; BITALG-NEXT: vpsadbw %xmm1, %xmm0, %xmm0
; BITALG-NEXT: vpackuswb %xmm2, %xmm0, %xmm0
; BITALG-NEXT: retq
;
[X86][SSE] Added 32-bit tests for vector lzcnt/tzcnt tests v2i64 tests are particularly bad on 32-bit targets. llvm-svn: 265789
2016-04-08 23:01:31 +08:00
; X32-SSE-LABEL: testv4i32:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; X32-SSE: # %bb.0:
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; X32-SSE-NEXT: pcmpeqd %xmm1, %xmm1
; X32-SSE-NEXT: paddd %xmm0, %xmm1
; X32-SSE-NEXT: pandn %xmm1, %xmm0
; X32-SSE-NEXT: movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
; X32-SSE-NEXT: movdqa %xmm0, %xmm2
; X32-SSE-NEXT: pand %xmm1, %xmm2
; X32-SSE-NEXT: movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
; X32-SSE-NEXT: movdqa %xmm3, %xmm4
; X32-SSE-NEXT: pshufb %xmm2, %xmm4
[x86] transform vector inc/dec to use -1 constant (PR33483) Convert vector increment or decrement to sub/add with an all-ones constant: add X, <1, 1...> --> sub X, <-1, -1...> sub X, <1, 1...> --> add X, <-1, -1...> The all-ones vector constant can be materialized using a pcmpeq instruction that is commonly recognized as an idiom (has no register dependency), so that's better than loading a splat 1 constant. AVX512 uses 'vpternlogd' for 512-bit vectors because there is apparently no better way to produce 512 one-bits. The general advantages of this lowering are: 1. pcmpeq has lower latency than a memop on every uarch I looked at in Agner's tables, so in theory, this could be better for perf, but... 2. That seems unlikely to affect any OOO implementation, and I can't measure any real perf difference from this transform on Haswell or Jaguar, but... 3. It doesn't look like it from the diffs, but this is an overall size win because we eliminate 16 - 64 constant bytes in the case of a vector load. If we're broadcasting a scalar load (which might itself be a bug), then we're replacing a scalar constant load + broadcast with a single cheap op, so that should always be smaller/better too. 4. This makes the DAG/isel output more consistent - we use pcmpeq already for padd x, -1 and psub x, -1, so we should use that form for +1 too because we can. If there's some reason to favor a constant load on some CPU, let's make the reverse transform for all of these cases (either here in the DAG or in a later machine pass). This should fix: https://bugs.llvm.org/show_bug.cgi?id=33483 Differential Revision: https://reviews.llvm.org/D34336 llvm-svn: 306289
2017-06-26 22:19:26 +08:00
; X32-SSE-NEXT: psrlw $4, %xmm0
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; X32-SSE-NEXT: pand %xmm1, %xmm0
; X32-SSE-NEXT: pshufb %xmm0, %xmm3
; X32-SSE-NEXT: paddb %xmm4, %xmm3
; X32-SSE-NEXT: pxor %xmm1, %xmm1
; X32-SSE-NEXT: pmovzxdq {{.*#+}} xmm0 = xmm3[0],zero,xmm3[1],zero
; X32-SSE-NEXT: punpckhdq {{.*#+}} xmm3 = xmm3[2],xmm1[2],xmm3[3],xmm1[3]
; X32-SSE-NEXT: psadbw %xmm1, %xmm3
[X86][SSE] Added 32-bit tests for vector lzcnt/tzcnt tests v2i64 tests are particularly bad on 32-bit targets. llvm-svn: 265789
2016-04-08 23:01:31 +08:00
; X32-SSE-NEXT: psadbw %xmm1, %xmm0
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; X32-SSE-NEXT: packuswb %xmm3, %xmm0
[X86][SSE] Added 32-bit tests for vector lzcnt/tzcnt tests v2i64 tests are particularly bad on 32-bit targets. llvm-svn: 265789
2016-04-08 23:01:31 +08:00
; X32-SSE-NEXT: retl
[X86] Added tzcnt vector tests. llvm-svn: 239264
2015-06-08 05:01:34 +08:00
%out = call <4 x i32> @llvm.cttz.v4i32(<4 x i32> %in, i1 0)
ret <4 x i32> %out
}
[X86][SSE] Tidied up vector CTLZ/CTTZ. NFCI. llvm-svn: 242645
2015-07-20 01:09:43 +08:00
define <4 x i32> @testv4i32u(<4 x i32> %in) nounwind {
[X86] Added tzcnt vector tests. llvm-svn: 239264
2015-06-08 05:01:34 +08:00
; SSE2-LABEL: testv4i32u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; SSE2: # %bb.0:
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; SSE2-NEXT: pcmpeqd %xmm1, %xmm1
; SSE2-NEXT: paddd %xmm0, %xmm1
; SSE2-NEXT: pandn %xmm1, %xmm0
; SSE2-NEXT: movdqa %xmm0, %xmm1
; SSE2-NEXT: psrlw $1, %xmm1
; SSE2-NEXT: pand {{.*}}(%rip), %xmm1
; SSE2-NEXT: psubb %xmm1, %xmm0
; SSE2-NEXT: movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
[X86][SSE] Vectorize CTTZ + CTTZ_ZERO_UNDEF Now that we have fast vector CTPOP implementations we can use this to speed up vector CTTZ using the pattern (cttz(x) = ctpop((x & -x) - 1)) Additionally, for AVX512CD that provides lzcnt instructions we can use the pattern (cttz_undef(x) = (width - 1) - ctlz(x & -x)) Differential Revision: http://reviews.llvm.org/D12663 llvm-svn: 248091
2015-09-19 21:22:57 +08:00
; SSE2-NEXT: movdqa %xmm0, %xmm2
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; SSE2-NEXT: pand %xmm1, %xmm2
; SSE2-NEXT: psrlw $2, %xmm0
; SSE2-NEXT: pand %xmm1, %xmm0
; SSE2-NEXT: paddb %xmm2, %xmm0
; SSE2-NEXT: movdqa %xmm0, %xmm1
; SSE2-NEXT: psrlw $4, %xmm1
; SSE2-NEXT: paddb %xmm0, %xmm1
; SSE2-NEXT: pand {{.*}}(%rip), %xmm1
; SSE2-NEXT: pxor %xmm0, %xmm0
; SSE2-NEXT: movdqa %xmm1, %xmm2
; SSE2-NEXT: punpckhdq {{.*#+}} xmm2 = xmm2[2],xmm0[2],xmm2[3],xmm0[3]
; SSE2-NEXT: psadbw %xmm0, %xmm2
; SSE2-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
; SSE2-NEXT: psadbw %xmm0, %xmm1
; SSE2-NEXT: packuswb %xmm2, %xmm1
; SSE2-NEXT: movdqa %xmm1, %xmm0
[X86] Added tzcnt vector tests. llvm-svn: 239264
2015-06-08 05:01:34 +08:00
; SSE2-NEXT: retq
;
; SSE3-LABEL: testv4i32u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; SSE3: # %bb.0:
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; SSE3-NEXT: pcmpeqd %xmm1, %xmm1
; SSE3-NEXT: paddd %xmm0, %xmm1
; SSE3-NEXT: pandn %xmm1, %xmm0
; SSE3-NEXT: movdqa %xmm0, %xmm1
; SSE3-NEXT: psrlw $1, %xmm1
; SSE3-NEXT: pand {{.*}}(%rip), %xmm1
; SSE3-NEXT: psubb %xmm1, %xmm0
; SSE3-NEXT: movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
[X86][SSE] Vectorize CTTZ + CTTZ_ZERO_UNDEF Now that we have fast vector CTPOP implementations we can use this to speed up vector CTTZ using the pattern (cttz(x) = ctpop((x & -x) - 1)) Additionally, for AVX512CD that provides lzcnt instructions we can use the pattern (cttz_undef(x) = (width - 1) - ctlz(x & -x)) Differential Revision: http://reviews.llvm.org/D12663 llvm-svn: 248091
2015-09-19 21:22:57 +08:00
; SSE3-NEXT: movdqa %xmm0, %xmm2
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; SSE3-NEXT: pand %xmm1, %xmm2
; SSE3-NEXT: psrlw $2, %xmm0
; SSE3-NEXT: pand %xmm1, %xmm0
; SSE3-NEXT: paddb %xmm2, %xmm0
; SSE3-NEXT: movdqa %xmm0, %xmm1
; SSE3-NEXT: psrlw $4, %xmm1
; SSE3-NEXT: paddb %xmm0, %xmm1
; SSE3-NEXT: pand {{.*}}(%rip), %xmm1
; SSE3-NEXT: pxor %xmm0, %xmm0
; SSE3-NEXT: movdqa %xmm1, %xmm2
; SSE3-NEXT: punpckhdq {{.*#+}} xmm2 = xmm2[2],xmm0[2],xmm2[3],xmm0[3]
; SSE3-NEXT: psadbw %xmm0, %xmm2
; SSE3-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
; SSE3-NEXT: psadbw %xmm0, %xmm1
; SSE3-NEXT: packuswb %xmm2, %xmm1
; SSE3-NEXT: movdqa %xmm1, %xmm0
[X86] Added tzcnt vector tests. llvm-svn: 239264
2015-06-08 05:01:34 +08:00
; SSE3-NEXT: retq
;
; SSSE3-LABEL: testv4i32u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; SSSE3: # %bb.0:
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; SSSE3-NEXT: pcmpeqd %xmm1, %xmm1
; SSSE3-NEXT: paddd %xmm0, %xmm1
; SSSE3-NEXT: pandn %xmm1, %xmm0
[x86] transform vector inc/dec to use -1 constant (PR33483) Convert vector increment or decrement to sub/add with an all-ones constant: add X, <1, 1...> --> sub X, <-1, -1...> sub X, <1, 1...> --> add X, <-1, -1...> The all-ones vector constant can be materialized using a pcmpeq instruction that is commonly recognized as an idiom (has no register dependency), so that's better than loading a splat 1 constant. AVX512 uses 'vpternlogd' for 512-bit vectors because there is apparently no better way to produce 512 one-bits. The general advantages of this lowering are: 1. pcmpeq has lower latency than a memop on every uarch I looked at in Agner's tables, so in theory, this could be better for perf, but... 2. That seems unlikely to affect any OOO implementation, and I can't measure any real perf difference from this transform on Haswell or Jaguar, but... 3. It doesn't look like it from the diffs, but this is an overall size win because we eliminate 16 - 64 constant bytes in the case of a vector load. If we're broadcasting a scalar load (which might itself be a bug), then we're replacing a scalar constant load + broadcast with a single cheap op, so that should always be smaller/better too. 4. This makes the DAG/isel output more consistent - we use pcmpeq already for padd x, -1 and psub x, -1, so we should use that form for +1 too because we can. If there's some reason to favor a constant load on some CPU, let's make the reverse transform for all of these cases (either here in the DAG or in a later machine pass). This should fix: https://bugs.llvm.org/show_bug.cgi?id=33483 Differential Revision: https://reviews.llvm.org/D34336 llvm-svn: 306289
2017-06-26 22:19:26 +08:00
; SSSE3-NEXT: movdqa {{.*#+}} xmm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; SSSE3-NEXT: movdqa %xmm0, %xmm3
[x86] transform vector inc/dec to use -1 constant (PR33483) Convert vector increment or decrement to sub/add with an all-ones constant: add X, <1, 1...> --> sub X, <-1, -1...> sub X, <1, 1...> --> add X, <-1, -1...> The all-ones vector constant can be materialized using a pcmpeq instruction that is commonly recognized as an idiom (has no register dependency), so that's better than loading a splat 1 constant. AVX512 uses 'vpternlogd' for 512-bit vectors because there is apparently no better way to produce 512 one-bits. The general advantages of this lowering are: 1. pcmpeq has lower latency than a memop on every uarch I looked at in Agner's tables, so in theory, this could be better for perf, but... 2. That seems unlikely to affect any OOO implementation, and I can't measure any real perf difference from this transform on Haswell or Jaguar, but... 3. It doesn't look like it from the diffs, but this is an overall size win because we eliminate 16 - 64 constant bytes in the case of a vector load. If we're broadcasting a scalar load (which might itself be a bug), then we're replacing a scalar constant load + broadcast with a single cheap op, so that should always be smaller/better too. 4. This makes the DAG/isel output more consistent - we use pcmpeq already for padd x, -1 and psub x, -1, so we should use that form for +1 too because we can. If there's some reason to favor a constant load on some CPU, let's make the reverse transform for all of these cases (either here in the DAG or in a later machine pass). This should fix: https://bugs.llvm.org/show_bug.cgi?id=33483 Differential Revision: https://reviews.llvm.org/D34336 llvm-svn: 306289
2017-06-26 22:19:26 +08:00
; SSSE3-NEXT: pand %xmm2, %xmm3
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; SSSE3-NEXT: movdqa {{.*#+}} xmm1 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
; SSSE3-NEXT: movdqa %xmm1, %xmm4
; SSSE3-NEXT: pshufb %xmm3, %xmm4
; SSSE3-NEXT: psrlw $4, %xmm0
; SSSE3-NEXT: pand %xmm2, %xmm0
; SSSE3-NEXT: pshufb %xmm0, %xmm1
; SSSE3-NEXT: paddb %xmm4, %xmm1
; SSSE3-NEXT: pxor %xmm0, %xmm0
; SSSE3-NEXT: movdqa %xmm1, %xmm2
; SSSE3-NEXT: punpckhdq {{.*#+}} xmm2 = xmm2[2],xmm0[2],xmm2[3],xmm0[3]
; SSSE3-NEXT: psadbw %xmm0, %xmm2
; SSSE3-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
; SSSE3-NEXT: psadbw %xmm0, %xmm1
; SSSE3-NEXT: packuswb %xmm2, %xmm1
; SSSE3-NEXT: movdqa %xmm1, %xmm0
[X86] Added tzcnt vector tests. llvm-svn: 239264
2015-06-08 05:01:34 +08:00
; SSSE3-NEXT: retq
;
; SSE41-LABEL: testv4i32u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; SSE41: # %bb.0:
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; SSE41-NEXT: pcmpeqd %xmm1, %xmm1
; SSE41-NEXT: paddd %xmm0, %xmm1
; SSE41-NEXT: pandn %xmm1, %xmm0
; SSE41-NEXT: movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
; SSE41-NEXT: movdqa %xmm0, %xmm2
; SSE41-NEXT: pand %xmm1, %xmm2
; SSE41-NEXT: movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
; SSE41-NEXT: movdqa %xmm3, %xmm4
; SSE41-NEXT: pshufb %xmm2, %xmm4
[x86] transform vector inc/dec to use -1 constant (PR33483) Convert vector increment or decrement to sub/add with an all-ones constant: add X, <1, 1...> --> sub X, <-1, -1...> sub X, <1, 1...> --> add X, <-1, -1...> The all-ones vector constant can be materialized using a pcmpeq instruction that is commonly recognized as an idiom (has no register dependency), so that's better than loading a splat 1 constant. AVX512 uses 'vpternlogd' for 512-bit vectors because there is apparently no better way to produce 512 one-bits. The general advantages of this lowering are: 1. pcmpeq has lower latency than a memop on every uarch I looked at in Agner's tables, so in theory, this could be better for perf, but... 2. That seems unlikely to affect any OOO implementation, and I can't measure any real perf difference from this transform on Haswell or Jaguar, but... 3. It doesn't look like it from the diffs, but this is an overall size win because we eliminate 16 - 64 constant bytes in the case of a vector load. If we're broadcasting a scalar load (which might itself be a bug), then we're replacing a scalar constant load + broadcast with a single cheap op, so that should always be smaller/better too. 4. This makes the DAG/isel output more consistent - we use pcmpeq already for padd x, -1 and psub x, -1, so we should use that form for +1 too because we can. If there's some reason to favor a constant load on some CPU, let's make the reverse transform for all of these cases (either here in the DAG or in a later machine pass). This should fix: https://bugs.llvm.org/show_bug.cgi?id=33483 Differential Revision: https://reviews.llvm.org/D34336 llvm-svn: 306289
2017-06-26 22:19:26 +08:00
; SSE41-NEXT: psrlw $4, %xmm0
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; SSE41-NEXT: pand %xmm1, %xmm0
; SSE41-NEXT: pshufb %xmm0, %xmm3
; SSE41-NEXT: paddb %xmm4, %xmm3
; SSE41-NEXT: pxor %xmm1, %xmm1
; SSE41-NEXT: pmovzxdq {{.*#+}} xmm0 = xmm3[0],zero,xmm3[1],zero
; SSE41-NEXT: punpckhdq {{.*#+}} xmm3 = xmm3[2],xmm1[2],xmm3[3],xmm1[3]
; SSE41-NEXT: psadbw %xmm1, %xmm3
[X86][SSE] Vectorize CTTZ + CTTZ_ZERO_UNDEF Now that we have fast vector CTPOP implementations we can use this to speed up vector CTTZ using the pattern (cttz(x) = ctpop((x & -x) - 1)) Additionally, for AVX512CD that provides lzcnt instructions we can use the pattern (cttz_undef(x) = (width - 1) - ctlz(x & -x)) Differential Revision: http://reviews.llvm.org/D12663 llvm-svn: 248091
2015-09-19 21:22:57 +08:00
; SSE41-NEXT: psadbw %xmm1, %xmm0
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; SSE41-NEXT: packuswb %xmm3, %xmm0
[X86] Added tzcnt vector tests. llvm-svn: 239264
2015-06-08 05:01:34 +08:00
; SSE41-NEXT: retq
;
[X86][SSE] Vectorize CTTZ + CTTZ_ZERO_UNDEF Now that we have fast vector CTPOP implementations we can use this to speed up vector CTTZ using the pattern (cttz(x) = ctpop((x & -x) - 1)) Additionally, for AVX512CD that provides lzcnt instructions we can use the pattern (cttz_undef(x) = (width - 1) - ctlz(x & -x)) Differential Revision: http://reviews.llvm.org/D12663 llvm-svn: 248091
2015-09-19 21:22:57 +08:00
; AVX1-LABEL: testv4i32u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; AVX1: # %bb.0:
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; AVX1-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
; AVX1-NEXT: vpaddd %xmm1, %xmm0, %xmm1
; AVX1-NEXT: vpandn %xmm1, %xmm0, %xmm0
; AVX1-NEXT: vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
; AVX1-NEXT: vpand %xmm1, %xmm0, %xmm2
; AVX1-NEXT: vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
; AVX1-NEXT: vpshufb %xmm2, %xmm3, %xmm2
[X86][SSE] Vectorize CTTZ + CTTZ_ZERO_UNDEF Now that we have fast vector CTPOP implementations we can use this to speed up vector CTTZ using the pattern (cttz(x) = ctpop((x & -x) - 1)) Additionally, for AVX512CD that provides lzcnt instructions we can use the pattern (cttz_undef(x) = (width - 1) - ctlz(x & -x)) Differential Revision: http://reviews.llvm.org/D12663 llvm-svn: 248091
2015-09-19 21:22:57 +08:00
; AVX1-NEXT: vpsrlw $4, %xmm0, %xmm0
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; AVX1-NEXT: vpand %xmm1, %xmm0, %xmm0
; AVX1-NEXT: vpshufb %xmm0, %xmm3, %xmm0
; AVX1-NEXT: vpaddb %xmm2, %xmm0, %xmm0
; AVX1-NEXT: vpxor %xmm1, %xmm1, %xmm1
[X86][SSE] Vectorize CTTZ + CTTZ_ZERO_UNDEF Now that we have fast vector CTPOP implementations we can use this to speed up vector CTTZ using the pattern (cttz(x) = ctpop((x & -x) - 1)) Additionally, for AVX512CD that provides lzcnt instructions we can use the pattern (cttz_undef(x) = (width - 1) - ctlz(x & -x)) Differential Revision: http://reviews.llvm.org/D12663 llvm-svn: 248091
2015-09-19 21:22:57 +08:00
; AVX1-NEXT: vpunpckhdq {{.*#+}} xmm2 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]
[X86] Fix several issues related to X86's psadbw instruction. This patch fixes the following issues: 1. Fix the return type of X86psadbw: it should not be the same type of inputs. For vNi8 inputs the output should be vMi64, where M = N/8. 2. Fix the return type of int_x86_avx512_psad_bw_512 accordingly. 3. Fix the definiton of PSADBW, VPSADBW, and VPSADBWY accordingly. 4. Adjust the return type when building a DAG node of X86ISD::PSADBW type. 5. Update related tests. Differential revision: http://reviews.llvm.org/D14897 llvm-svn: 254010
2015-11-25 03:51:26 +08:00
; AVX1-NEXT: vpsadbw %xmm1, %xmm2, %xmm2
[X86][SSE] Add initial support for combining target shuffles to (V)PMOVZX. We can only handle 128-bit vectors until we support target shuffle inputs of different size to the output. llvm-svn: 288140
2016-11-29 22:18:51 +08:00
; AVX1-NEXT: vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
[X86] Fix several issues related to X86's psadbw instruction. This patch fixes the following issues: 1. Fix the return type of X86psadbw: it should not be the same type of inputs. For vNi8 inputs the output should be vMi64, where M = N/8. 2. Fix the return type of int_x86_avx512_psad_bw_512 accordingly. 3. Fix the definiton of PSADBW, VPSADBW, and VPSADBWY accordingly. 4. Adjust the return type when building a DAG node of X86ISD::PSADBW type. 5. Update related tests. Differential revision: http://reviews.llvm.org/D14897 llvm-svn: 254010
2015-11-25 03:51:26 +08:00
; AVX1-NEXT: vpsadbw %xmm1, %xmm0, %xmm0
[X86][SSE] Vectorize CTTZ + CTTZ_ZERO_UNDEF Now that we have fast vector CTPOP implementations we can use this to speed up vector CTTZ using the pattern (cttz(x) = ctpop((x & -x) - 1)) Additionally, for AVX512CD that provides lzcnt instructions we can use the pattern (cttz_undef(x) = (width - 1) - ctlz(x & -x)) Differential Revision: http://reviews.llvm.org/D12663 llvm-svn: 248091
2015-09-19 21:22:57 +08:00
; AVX1-NEXT: vpackuswb %xmm2, %xmm0, %xmm0
; AVX1-NEXT: retq
;
; AVX2-LABEL: testv4i32u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; AVX2: # %bb.0:
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; AVX2-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
; AVX2-NEXT: vpaddd %xmm1, %xmm0, %xmm1
; AVX2-NEXT: vpandn %xmm1, %xmm0, %xmm0
; AVX2-NEXT: vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
; AVX2-NEXT: vpand %xmm1, %xmm0, %xmm2
; AVX2-NEXT: vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
; AVX2-NEXT: vpshufb %xmm2, %xmm3, %xmm2
[X86][SSE] Vectorize CTTZ + CTTZ_ZERO_UNDEF Now that we have fast vector CTPOP implementations we can use this to speed up vector CTTZ using the pattern (cttz(x) = ctpop((x & -x) - 1)) Additionally, for AVX512CD that provides lzcnt instructions we can use the pattern (cttz_undef(x) = (width - 1) - ctlz(x & -x)) Differential Revision: http://reviews.llvm.org/D12663 llvm-svn: 248091
2015-09-19 21:22:57 +08:00
; AVX2-NEXT: vpsrlw $4, %xmm0, %xmm0
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; AVX2-NEXT: vpand %xmm1, %xmm0, %xmm0
; AVX2-NEXT: vpshufb %xmm0, %xmm3, %xmm0
; AVX2-NEXT: vpaddb %xmm2, %xmm0, %xmm0
; AVX2-NEXT: vpxor %xmm1, %xmm1, %xmm1
[X86][SSE] Vectorize CTTZ + CTTZ_ZERO_UNDEF Now that we have fast vector CTPOP implementations we can use this to speed up vector CTTZ using the pattern (cttz(x) = ctpop((x & -x) - 1)) Additionally, for AVX512CD that provides lzcnt instructions we can use the pattern (cttz_undef(x) = (width - 1) - ctlz(x & -x)) Differential Revision: http://reviews.llvm.org/D12663 llvm-svn: 248091
2015-09-19 21:22:57 +08:00
; AVX2-NEXT: vpunpckhdq {{.*#+}} xmm2 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]
[X86] Fix several issues related to X86's psadbw instruction. This patch fixes the following issues: 1. Fix the return type of X86psadbw: it should not be the same type of inputs. For vNi8 inputs the output should be vMi64, where M = N/8. 2. Fix the return type of int_x86_avx512_psad_bw_512 accordingly. 3. Fix the definiton of PSADBW, VPSADBW, and VPSADBWY accordingly. 4. Adjust the return type when building a DAG node of X86ISD::PSADBW type. 5. Update related tests. Differential revision: http://reviews.llvm.org/D14897 llvm-svn: 254010
2015-11-25 03:51:26 +08:00
; AVX2-NEXT: vpsadbw %xmm1, %xmm2, %xmm2
[X86][SSE] Add initial support for combining target shuffles to (V)PMOVZX. We can only handle 128-bit vectors until we support target shuffle inputs of different size to the output. llvm-svn: 288140
2016-11-29 22:18:51 +08:00
; AVX2-NEXT: vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
[X86] Fix several issues related to X86's psadbw instruction. This patch fixes the following issues: 1. Fix the return type of X86psadbw: it should not be the same type of inputs. For vNi8 inputs the output should be vMi64, where M = N/8. 2. Fix the return type of int_x86_avx512_psad_bw_512 accordingly. 3. Fix the definiton of PSADBW, VPSADBW, and VPSADBWY accordingly. 4. Adjust the return type when building a DAG node of X86ISD::PSADBW type. 5. Update related tests. Differential revision: http://reviews.llvm.org/D14897 llvm-svn: 254010
2015-11-25 03:51:26 +08:00
; AVX2-NEXT: vpsadbw %xmm1, %xmm0, %xmm0
[X86][SSE] Vectorize CTTZ + CTTZ_ZERO_UNDEF Now that we have fast vector CTPOP implementations we can use this to speed up vector CTTZ using the pattern (cttz(x) = ctpop((x & -x) - 1)) Additionally, for AVX512CD that provides lzcnt instructions we can use the pattern (cttz_undef(x) = (width - 1) - ctlz(x & -x)) Differential Revision: http://reviews.llvm.org/D12663 llvm-svn: 248091
2015-09-19 21:22:57 +08:00
; AVX2-NEXT: vpackuswb %xmm2, %xmm0, %xmm0
; AVX2-NEXT: retq
[X86][SSE] Added 32-bit tests for vector lzcnt/tzcnt tests v2i64 tests are particularly bad on 32-bit targets. llvm-svn: 265789
2016-04-08 23:01:31 +08:00
;
[AVX512] Add avx512cd+vl runs to vector-tzcnt-128/256 tests to show using the vplzcntd/q instructions. llvm-svn: 266860
2016-04-20 13:19:01 +08:00
; AVX512CDVL-LABEL: testv4i32u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; AVX512CDVL: # %bb.0:
[X86][SSE] Improve CTTZ lowering when CTLZ is legal If we have better CTLZ support than CTPOP, then use cttz(x) = width - ctlz(~x & (x - 1)) - and remove the CTTZ_ZERO_UNDEF handling as it no longer gives better codegen. Similar to rL344447, this is also closer to LegalizeDAG's approach llvm-svn: 344448
2018-10-13 21:05:19 +08:00
; AVX512CDVL-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
; AVX512CDVL-NEXT: vpaddd %xmm1, %xmm0, %xmm1
; AVX512CDVL-NEXT: vpandn %xmm1, %xmm0, %xmm0
[AVX512] Add avx512cd+vl runs to vector-tzcnt-128/256 tests to show using the vplzcntd/q instructions. llvm-svn: 266860
2016-04-20 13:19:01 +08:00
; AVX512CDVL-NEXT: vplzcntd %xmm0, %xmm0
[X86][SSE] Improve CTTZ lowering when CTLZ is legal If we have better CTLZ support than CTPOP, then use cttz(x) = width - ctlz(~x & (x - 1)) - and remove the CTTZ_ZERO_UNDEF handling as it no longer gives better codegen. Similar to rL344447, this is also closer to LegalizeDAG's approach llvm-svn: 344448
2018-10-13 21:05:19 +08:00
; AVX512CDVL-NEXT: vpbroadcastd {{.*#+}} xmm1 = [32,32,32,32]
[AVX512] Add avx512cd+vl runs to vector-tzcnt-128/256 tests to show using the vplzcntd/q instructions. llvm-svn: 266860
2016-04-20 13:19:01 +08:00
; AVX512CDVL-NEXT: vpsubd %xmm0, %xmm1, %xmm0
; AVX512CDVL-NEXT: retq
;
[AVX512] Teach lowering to use vplzcntd/q to implement 128/256-bit CTTZ_ZERO_UNDEF even without VLX support. We can just extend to 512-bits and extract like we do for CTLZ. llvm-svn: 267100
2016-04-22 11:22:38 +08:00
; AVX512CD-LABEL: testv4i32u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; AVX512CD: # %bb.0:
[X86][SSE] Improve CTTZ lowering when CTLZ is legal If we have better CTLZ support than CTPOP, then use cttz(x) = width - ctlz(~x & (x - 1)) - and remove the CTTZ_ZERO_UNDEF handling as it no longer gives better codegen. Similar to rL344447, this is also closer to LegalizeDAG's approach llvm-svn: 344448
2018-10-13 21:05:19 +08:00
; AVX512CD-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
; AVX512CD-NEXT: vpaddd %xmm1, %xmm0, %xmm1
; AVX512CD-NEXT: vpandn %xmm1, %xmm0, %xmm0
[AVX512] Teach lowering to use vplzcntd/q to implement 128/256-bit CTTZ_ZERO_UNDEF even without VLX support. We can just extend to 512-bits and extract like we do for CTLZ. llvm-svn: 267100
2016-04-22 11:22:38 +08:00
; AVX512CD-NEXT: vplzcntd %zmm0, %zmm0
[X86][SSE] Improve CTTZ lowering when CTLZ is legal If we have better CTLZ support than CTPOP, then use cttz(x) = width - ctlz(~x & (x - 1)) - and remove the CTTZ_ZERO_UNDEF handling as it no longer gives better codegen. Similar to rL344447, this is also closer to LegalizeDAG's approach llvm-svn: 344448
2018-10-13 21:05:19 +08:00
; AVX512CD-NEXT: vpbroadcastd {{.*#+}} xmm1 = [32,32,32,32]
[AVX512] Teach lowering to use vplzcntd/q to implement 128/256-bit CTTZ_ZERO_UNDEF even without VLX support. We can just extend to 512-bits and extract like we do for CTLZ. llvm-svn: 267100
2016-04-22 11:22:38 +08:00
; AVX512CD-NEXT: vpsubd %xmm0, %xmm1, %xmm0
[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
; AVX512CD-NEXT: vzeroupper
[AVX512] Teach lowering to use vplzcntd/q to implement 128/256-bit CTTZ_ZERO_UNDEF even without VLX support. We can just extend to 512-bits and extract like we do for CTLZ. llvm-svn: 267100
2016-04-22 11:22:38 +08:00
; AVX512CD-NEXT: retq
;
[X86] Adding vpopcntd and vpopcntq instructions AVX512_VPOPCNTDQ is a new feature set that was published by Intel. The patch represents the LLVM side of the addition of two new intrinsic based instructions (vpopcntd and vpopcntq). Differential Revision: https://reviews.llvm.org/D33169 llvm-svn: 303858
2017-05-25 21:45:23 +08:00
; AVX512VPOPCNTDQ-LABEL: testv4i32u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; AVX512VPOPCNTDQ: # %bb.0:
[x86] transform vector inc/dec to use -1 constant (PR33483) Convert vector increment or decrement to sub/add with an all-ones constant: add X, <1, 1...> --> sub X, <-1, -1...> sub X, <1, 1...> --> add X, <-1, -1...> The all-ones vector constant can be materialized using a pcmpeq instruction that is commonly recognized as an idiom (has no register dependency), so that's better than loading a splat 1 constant. AVX512 uses 'vpternlogd' for 512-bit vectors because there is apparently no better way to produce 512 one-bits. The general advantages of this lowering are: 1. pcmpeq has lower latency than a memop on every uarch I looked at in Agner's tables, so in theory, this could be better for perf, but... 2. That seems unlikely to affect any OOO implementation, and I can't measure any real perf difference from this transform on Haswell or Jaguar, but... 3. It doesn't look like it from the diffs, but this is an overall size win because we eliminate 16 - 64 constant bytes in the case of a vector load. If we're broadcasting a scalar load (which might itself be a bug), then we're replacing a scalar constant load + broadcast with a single cheap op, so that should always be smaller/better too. 4. This makes the DAG/isel output more consistent - we use pcmpeq already for padd x, -1 and psub x, -1, so we should use that form for +1 too because we can. If there's some reason to favor a constant load on some CPU, let's make the reverse transform for all of these cases (either here in the DAG or in a later machine pass). This should fix: https://bugs.llvm.org/show_bug.cgi?id=33483 Differential Revision: https://reviews.llvm.org/D34336 llvm-svn: 306289
2017-06-26 22:19:26 +08:00
; AVX512VPOPCNTDQ-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; AVX512VPOPCNTDQ-NEXT: vpaddd %xmm1, %xmm0, %xmm1
; AVX512VPOPCNTDQ-NEXT: vpandn %xmm1, %xmm0, %xmm0
[X86] Adding vpopcntd and vpopcntq instructions AVX512_VPOPCNTDQ is a new feature set that was published by Intel. The patch represents the LLVM side of the addition of two new intrinsic based instructions (vpopcntd and vpopcntq). Differential Revision: https://reviews.llvm.org/D33169 llvm-svn: 303858
2017-05-25 21:45:23 +08:00
; AVX512VPOPCNTDQ-NEXT: vpopcntd %zmm0, %zmm0
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
; AVX512VPOPCNTDQ-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
[X86] Adding vpopcntd and vpopcntq instructions AVX512_VPOPCNTDQ is a new feature set that was published by Intel. The patch represents the LLVM side of the addition of two new intrinsic based instructions (vpopcntd and vpopcntq). Differential Revision: https://reviews.llvm.org/D33169 llvm-svn: 303858
2017-05-25 21:45:23 +08:00
; AVX512VPOPCNTDQ-NEXT: vzeroupper
; AVX512VPOPCNTDQ-NEXT: retq
;
[X86] Add 128 and 256-bit VPOPCNTDQ instructions. Adjust some tablegen classes LZCNT/POPCNT. I think when this instruction was first published it was only for a Knights CPU and thus VLX version was missing. llvm-svn: 320910
2017-12-16 10:40:28 +08:00
; AVX512VPOPCNTDQVL-LABEL: testv4i32u:
; AVX512VPOPCNTDQVL: # %bb.0:
; AVX512VPOPCNTDQVL-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; AVX512VPOPCNTDQVL-NEXT: vpaddd %xmm1, %xmm0, %xmm1
; AVX512VPOPCNTDQVL-NEXT: vpandn %xmm1, %xmm0, %xmm0
[X86] Add 128 and 256-bit VPOPCNTDQ instructions. Adjust some tablegen classes LZCNT/POPCNT. I think when this instruction was first published it was only for a Knights CPU and thus VLX version was missing. llvm-svn: 320910
2017-12-16 10:40:28 +08:00
; AVX512VPOPCNTDQVL-NEXT: vpopcntd %xmm0, %xmm0
; AVX512VPOPCNTDQVL-NEXT: retq
;
[X86] Regenerate the vector-popcnt and vector-tzcnt tests to get BITALG CHECK linse on all functions not just the vXi16/vXi8. llvm-svn: 318885
2017-11-23 07:35:12 +08:00
; BITALG_NOVLX-LABEL: testv4i32u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; BITALG_NOVLX: # %bb.0:
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; BITALG_NOVLX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
; BITALG_NOVLX-NEXT: vpaddd %xmm1, %xmm0, %xmm1
; BITALG_NOVLX-NEXT: vpandn %xmm1, %xmm0, %xmm0
[X86][SSE] LowerVectorCTPOP - pull out repeated byte sum stage. Pull out repeated byte sum stage for popcount of vector elements > 8bits. This allows us to simplify the LUT/BITMATH popcnt code to always assume vXi8 vectors, and also improves avx512bitalg codegen which only has access to vpopcntb/vpopcntw. llvm-svn: 344348
2018-10-12 22:18:47 +08:00
; BITALG_NOVLX-NEXT: vpopcntb %zmm0, %zmm0
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; BITALG_NOVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1
[X86] Regenerate the vector-popcnt and vector-tzcnt tests to get BITALG CHECK linse on all functions not just the vXi16/vXi8. llvm-svn: 318885
2017-11-23 07:35:12 +08:00
; BITALG_NOVLX-NEXT: vpunpckhdq {{.*#+}} xmm2 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]
; BITALG_NOVLX-NEXT: vpsadbw %xmm1, %xmm2, %xmm2
; BITALG_NOVLX-NEXT: vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
; BITALG_NOVLX-NEXT: vpsadbw %xmm1, %xmm0, %xmm0
; BITALG_NOVLX-NEXT: vpackuswb %xmm2, %xmm0, %xmm0
[X86][SSE] LowerVectorCTPOP - pull out repeated byte sum stage. Pull out repeated byte sum stage for popcount of vector elements > 8bits. This allows us to simplify the LUT/BITMATH popcnt code to always assume vXi8 vectors, and also improves avx512bitalg codegen which only has access to vpopcntb/vpopcntw. llvm-svn: 344348
2018-10-12 22:18:47 +08:00
; BITALG_NOVLX-NEXT: vzeroupper
[X86] Regenerate the vector-popcnt and vector-tzcnt tests to get BITALG CHECK linse on all functions not just the vXi16/vXi8. llvm-svn: 318885
2017-11-23 07:35:12 +08:00
; BITALG_NOVLX-NEXT: retq
;
; BITALG-LABEL: testv4i32u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; BITALG: # %bb.0:
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; BITALG-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
; BITALG-NEXT: vpaddd %xmm1, %xmm0, %xmm1
; BITALG-NEXT: vpandn %xmm1, %xmm0, %xmm0
[X86][SSE] LowerVectorCTPOP - pull out repeated byte sum stage. Pull out repeated byte sum stage for popcount of vector elements > 8bits. This allows us to simplify the LUT/BITMATH popcnt code to always assume vXi8 vectors, and also improves avx512bitalg codegen which only has access to vpopcntb/vpopcntw. llvm-svn: 344348
2018-10-12 22:18:47 +08:00
; BITALG-NEXT: vpopcntb %xmm0, %xmm0
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; BITALG-NEXT: vpxor %xmm1, %xmm1, %xmm1
[X86] Regenerate the vector-popcnt and vector-tzcnt tests to get BITALG CHECK linse on all functions not just the vXi16/vXi8. llvm-svn: 318885
2017-11-23 07:35:12 +08:00
; BITALG-NEXT: vpunpckhdq {{.*#+}} xmm2 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]
; BITALG-NEXT: vpsadbw %xmm1, %xmm2, %xmm2
[X86] In LowerHorizontalByteSum, emit vector_shuffle nodes instead of directly using X86ISD::UNPCKL/X86ISD::UNPCKH. This gives shuffle lowering the freedom to use zero_extend_vector_inreg for the unpckl shuffle. Shuffle combining usually makes this swap later, but not when AVX512 is enabled it seems. While there also use DAG.getConstant to create a 0 vector instead of using the helper the forces a specific BUILD_VECTOR. I don't think that helper is usually needed. We're basically free to create a constant build_vector anytime and it will be legalized on its own. llvm-svn: 346574
2018-11-10 08:26:42 +08:00
; BITALG-NEXT: vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
[X86] Regenerate the vector-popcnt and vector-tzcnt tests to get BITALG CHECK linse on all functions not just the vXi16/vXi8. llvm-svn: 318885
2017-11-23 07:35:12 +08:00
; BITALG-NEXT: vpsadbw %xmm1, %xmm0, %xmm0
; BITALG-NEXT: vpackuswb %xmm2, %xmm0, %xmm0
; BITALG-NEXT: retq
;
[X86][SSE] Added 32-bit tests for vector lzcnt/tzcnt tests v2i64 tests are particularly bad on 32-bit targets. llvm-svn: 265789
2016-04-08 23:01:31 +08:00
; X32-SSE-LABEL: testv4i32u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; X32-SSE: # %bb.0:
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; X32-SSE-NEXT: pcmpeqd %xmm1, %xmm1
; X32-SSE-NEXT: paddd %xmm0, %xmm1
; X32-SSE-NEXT: pandn %xmm1, %xmm0
; X32-SSE-NEXT: movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
; X32-SSE-NEXT: movdqa %xmm0, %xmm2
; X32-SSE-NEXT: pand %xmm1, %xmm2
; X32-SSE-NEXT: movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
; X32-SSE-NEXT: movdqa %xmm3, %xmm4
; X32-SSE-NEXT: pshufb %xmm2, %xmm4
[x86] transform vector inc/dec to use -1 constant (PR33483) Convert vector increment or decrement to sub/add with an all-ones constant: add X, <1, 1...> --> sub X, <-1, -1...> sub X, <1, 1...> --> add X, <-1, -1...> The all-ones vector constant can be materialized using a pcmpeq instruction that is commonly recognized as an idiom (has no register dependency), so that's better than loading a splat 1 constant. AVX512 uses 'vpternlogd' for 512-bit vectors because there is apparently no better way to produce 512 one-bits. The general advantages of this lowering are: 1. pcmpeq has lower latency than a memop on every uarch I looked at in Agner's tables, so in theory, this could be better for perf, but... 2. That seems unlikely to affect any OOO implementation, and I can't measure any real perf difference from this transform on Haswell or Jaguar, but... 3. It doesn't look like it from the diffs, but this is an overall size win because we eliminate 16 - 64 constant bytes in the case of a vector load. If we're broadcasting a scalar load (which might itself be a bug), then we're replacing a scalar constant load + broadcast with a single cheap op, so that should always be smaller/better too. 4. This makes the DAG/isel output more consistent - we use pcmpeq already for padd x, -1 and psub x, -1, so we should use that form for +1 too because we can. If there's some reason to favor a constant load on some CPU, let's make the reverse transform for all of these cases (either here in the DAG or in a later machine pass). This should fix: https://bugs.llvm.org/show_bug.cgi?id=33483 Differential Revision: https://reviews.llvm.org/D34336 llvm-svn: 306289
2017-06-26 22:19:26 +08:00
; X32-SSE-NEXT: psrlw $4, %xmm0
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; X32-SSE-NEXT: pand %xmm1, %xmm0
; X32-SSE-NEXT: pshufb %xmm0, %xmm3
; X32-SSE-NEXT: paddb %xmm4, %xmm3
; X32-SSE-NEXT: pxor %xmm1, %xmm1
; X32-SSE-NEXT: pmovzxdq {{.*#+}} xmm0 = xmm3[0],zero,xmm3[1],zero
; X32-SSE-NEXT: punpckhdq {{.*#+}} xmm3 = xmm3[2],xmm1[2],xmm3[3],xmm1[3]
; X32-SSE-NEXT: psadbw %xmm1, %xmm3
[X86][SSE] Added 32-bit tests for vector lzcnt/tzcnt tests v2i64 tests are particularly bad on 32-bit targets. llvm-svn: 265789
2016-04-08 23:01:31 +08:00
; X32-SSE-NEXT: psadbw %xmm1, %xmm0
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; X32-SSE-NEXT: packuswb %xmm3, %xmm0
[X86][SSE] Added 32-bit tests for vector lzcnt/tzcnt tests v2i64 tests are particularly bad on 32-bit targets. llvm-svn: 265789
2016-04-08 23:01:31 +08:00
; X32-SSE-NEXT: retl
[X86] Added tzcnt vector tests. llvm-svn: 239264
2015-06-08 05:01:34 +08:00
%out = call <4 x i32> @llvm.cttz.v4i32(<4 x i32> %in, i1 -1)
ret <4 x i32> %out
}
[X86][SSE] Tidied up vector CTLZ/CTTZ. NFCI. llvm-svn: 242645
2015-07-20 01:09:43 +08:00
define <8 x i16> @testv8i16(<8 x i16> %in) nounwind {
[X86] Added tzcnt vector tests. llvm-svn: 239264
2015-06-08 05:01:34 +08:00
; SSE2-LABEL: testv8i16:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; SSE2: # %bb.0:
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; SSE2-NEXT: pcmpeqd %xmm1, %xmm1
; SSE2-NEXT: paddw %xmm0, %xmm1
; SSE2-NEXT: pandn %xmm1, %xmm0
[x86] transform vector inc/dec to use -1 constant (PR33483) Convert vector increment or decrement to sub/add with an all-ones constant: add X, <1, 1...> --> sub X, <-1, -1...> sub X, <1, 1...> --> add X, <-1, -1...> The all-ones vector constant can be materialized using a pcmpeq instruction that is commonly recognized as an idiom (has no register dependency), so that's better than loading a splat 1 constant. AVX512 uses 'vpternlogd' for 512-bit vectors because there is apparently no better way to produce 512 one-bits. The general advantages of this lowering are: 1. pcmpeq has lower latency than a memop on every uarch I looked at in Agner's tables, so in theory, this could be better for perf, but... 2. That seems unlikely to affect any OOO implementation, and I can't measure any real perf difference from this transform on Haswell or Jaguar, but... 3. It doesn't look like it from the diffs, but this is an overall size win because we eliminate 16 - 64 constant bytes in the case of a vector load. If we're broadcasting a scalar load (which might itself be a bug), then we're replacing a scalar constant load + broadcast with a single cheap op, so that should always be smaller/better too. 4. This makes the DAG/isel output more consistent - we use pcmpeq already for padd x, -1 and psub x, -1, so we should use that form for +1 too because we can. If there's some reason to favor a constant load on some CPU, let's make the reverse transform for all of these cases (either here in the DAG or in a later machine pass). This should fix: https://bugs.llvm.org/show_bug.cgi?id=33483 Differential Revision: https://reviews.llvm.org/D34336 llvm-svn: 306289
2017-06-26 22:19:26 +08:00
; SSE2-NEXT: movdqa %xmm0, %xmm1
; SSE2-NEXT: psrlw $1, %xmm1
; SSE2-NEXT: pand {{.*}}(%rip), %xmm1
[X86][SSE] LowerVectorCTPOP - pull out repeated byte sum stage. Pull out repeated byte sum stage for popcount of vector elements > 8bits. This allows us to simplify the LUT/BITMATH popcnt code to always assume vXi8 vectors, and also improves avx512bitalg codegen which only has access to vpopcntb/vpopcntw. llvm-svn: 344348
2018-10-12 22:18:47 +08:00
; SSE2-NEXT: psubb %xmm1, %xmm0
; SSE2-NEXT: movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
[x86] transform vector inc/dec to use -1 constant (PR33483) Convert vector increment or decrement to sub/add with an all-ones constant: add X, <1, 1...> --> sub X, <-1, -1...> sub X, <1, 1...> --> add X, <-1, -1...> The all-ones vector constant can be materialized using a pcmpeq instruction that is commonly recognized as an idiom (has no register dependency), so that's better than loading a splat 1 constant. AVX512 uses 'vpternlogd' for 512-bit vectors because there is apparently no better way to produce 512 one-bits. The general advantages of this lowering are: 1. pcmpeq has lower latency than a memop on every uarch I looked at in Agner's tables, so in theory, this could be better for perf, but... 2. That seems unlikely to affect any OOO implementation, and I can't measure any real perf difference from this transform on Haswell or Jaguar, but... 3. It doesn't look like it from the diffs, but this is an overall size win because we eliminate 16 - 64 constant bytes in the case of a vector load. If we're broadcasting a scalar load (which might itself be a bug), then we're replacing a scalar constant load + broadcast with a single cheap op, so that should always be smaller/better too. 4. This makes the DAG/isel output more consistent - we use pcmpeq already for padd x, -1 and psub x, -1, so we should use that form for +1 too because we can. If there's some reason to favor a constant load on some CPU, let's make the reverse transform for all of these cases (either here in the DAG or in a later machine pass). This should fix: https://bugs.llvm.org/show_bug.cgi?id=33483 Differential Revision: https://reviews.llvm.org/D34336 llvm-svn: 306289
2017-06-26 22:19:26 +08:00
; SSE2-NEXT: movdqa %xmm0, %xmm2
; SSE2-NEXT: pand %xmm1, %xmm2
; SSE2-NEXT: psrlw $2, %xmm0
; SSE2-NEXT: pand %xmm1, %xmm0
[X86][SSE] LowerVectorCTPOP - pull out repeated byte sum stage. Pull out repeated byte sum stage for popcount of vector elements > 8bits. This allows us to simplify the LUT/BITMATH popcnt code to always assume vXi8 vectors, and also improves avx512bitalg codegen which only has access to vpopcntb/vpopcntw. llvm-svn: 344348
2018-10-12 22:18:47 +08:00
; SSE2-NEXT: paddb %xmm2, %xmm0
[x86] transform vector inc/dec to use -1 constant (PR33483) Convert vector increment or decrement to sub/add with an all-ones constant: add X, <1, 1...> --> sub X, <-1, -1...> sub X, <1, 1...> --> add X, <-1, -1...> The all-ones vector constant can be materialized using a pcmpeq instruction that is commonly recognized as an idiom (has no register dependency), so that's better than loading a splat 1 constant. AVX512 uses 'vpternlogd' for 512-bit vectors because there is apparently no better way to produce 512 one-bits. The general advantages of this lowering are: 1. pcmpeq has lower latency than a memop on every uarch I looked at in Agner's tables, so in theory, this could be better for perf, but... 2. That seems unlikely to affect any OOO implementation, and I can't measure any real perf difference from this transform on Haswell or Jaguar, but... 3. It doesn't look like it from the diffs, but this is an overall size win because we eliminate 16 - 64 constant bytes in the case of a vector load. If we're broadcasting a scalar load (which might itself be a bug), then we're replacing a scalar constant load + broadcast with a single cheap op, so that should always be smaller/better too. 4. This makes the DAG/isel output more consistent - we use pcmpeq already for padd x, -1 and psub x, -1, so we should use that form for +1 too because we can. If there's some reason to favor a constant load on some CPU, let's make the reverse transform for all of these cases (either here in the DAG or in a later machine pass). This should fix: https://bugs.llvm.org/show_bug.cgi?id=33483 Differential Revision: https://reviews.llvm.org/D34336 llvm-svn: 306289
2017-06-26 22:19:26 +08:00
; SSE2-NEXT: movdqa %xmm0, %xmm1
; SSE2-NEXT: psrlw $4, %xmm1
[X86][SSE] LowerVectorCTPOP - pull out repeated byte sum stage. Pull out repeated byte sum stage for popcount of vector elements > 8bits. This allows us to simplify the LUT/BITMATH popcnt code to always assume vXi8 vectors, and also improves avx512bitalg codegen which only has access to vpopcntb/vpopcntw. llvm-svn: 344348
2018-10-12 22:18:47 +08:00
; SSE2-NEXT: paddb %xmm0, %xmm1
[x86] transform vector inc/dec to use -1 constant (PR33483) Convert vector increment or decrement to sub/add with an all-ones constant: add X, <1, 1...> --> sub X, <-1, -1...> sub X, <1, 1...> --> add X, <-1, -1...> The all-ones vector constant can be materialized using a pcmpeq instruction that is commonly recognized as an idiom (has no register dependency), so that's better than loading a splat 1 constant. AVX512 uses 'vpternlogd' for 512-bit vectors because there is apparently no better way to produce 512 one-bits. The general advantages of this lowering are: 1. pcmpeq has lower latency than a memop on every uarch I looked at in Agner's tables, so in theory, this could be better for perf, but... 2. That seems unlikely to affect any OOO implementation, and I can't measure any real perf difference from this transform on Haswell or Jaguar, but... 3. It doesn't look like it from the diffs, but this is an overall size win because we eliminate 16 - 64 constant bytes in the case of a vector load. If we're broadcasting a scalar load (which might itself be a bug), then we're replacing a scalar constant load + broadcast with a single cheap op, so that should always be smaller/better too. 4. This makes the DAG/isel output more consistent - we use pcmpeq already for padd x, -1 and psub x, -1, so we should use that form for +1 too because we can. If there's some reason to favor a constant load on some CPU, let's make the reverse transform for all of these cases (either here in the DAG or in a later machine pass). This should fix: https://bugs.llvm.org/show_bug.cgi?id=33483 Differential Revision: https://reviews.llvm.org/D34336 llvm-svn: 306289
2017-06-26 22:19:26 +08:00
; SSE2-NEXT: pand {{.*}}(%rip), %xmm1
[X86][SSE] Vectorize CTTZ + CTTZ_ZERO_UNDEF Now that we have fast vector CTPOP implementations we can use this to speed up vector CTTZ using the pattern (cttz(x) = ctpop((x & -x) - 1)) Additionally, for AVX512CD that provides lzcnt instructions we can use the pattern (cttz_undef(x) = (width - 1) - ctlz(x & -x)) Differential Revision: http://reviews.llvm.org/D12663 llvm-svn: 248091
2015-09-19 21:22:57 +08:00
; SSE2-NEXT: movdqa %xmm1, %xmm0
; SSE2-NEXT: psllw $8, %xmm0
[x86] transform vector inc/dec to use -1 constant (PR33483) Convert vector increment or decrement to sub/add with an all-ones constant: add X, <1, 1...> --> sub X, <-1, -1...> sub X, <1, 1...> --> add X, <-1, -1...> The all-ones vector constant can be materialized using a pcmpeq instruction that is commonly recognized as an idiom (has no register dependency), so that's better than loading a splat 1 constant. AVX512 uses 'vpternlogd' for 512-bit vectors because there is apparently no better way to produce 512 one-bits. The general advantages of this lowering are: 1. pcmpeq has lower latency than a memop on every uarch I looked at in Agner's tables, so in theory, this could be better for perf, but... 2. That seems unlikely to affect any OOO implementation, and I can't measure any real perf difference from this transform on Haswell or Jaguar, but... 3. It doesn't look like it from the diffs, but this is an overall size win because we eliminate 16 - 64 constant bytes in the case of a vector load. If we're broadcasting a scalar load (which might itself be a bug), then we're replacing a scalar constant load + broadcast with a single cheap op, so that should always be smaller/better too. 4. This makes the DAG/isel output more consistent - we use pcmpeq already for padd x, -1 and psub x, -1, so we should use that form for +1 too because we can. If there's some reason to favor a constant load on some CPU, let's make the reverse transform for all of these cases (either here in the DAG or in a later machine pass). This should fix: https://bugs.llvm.org/show_bug.cgi?id=33483 Differential Revision: https://reviews.llvm.org/D34336 llvm-svn: 306289
2017-06-26 22:19:26 +08:00
; SSE2-NEXT: paddb %xmm1, %xmm0
[X86][SSE] Vectorize CTTZ + CTTZ_ZERO_UNDEF Now that we have fast vector CTPOP implementations we can use this to speed up vector CTTZ using the pattern (cttz(x) = ctpop((x & -x) - 1)) Additionally, for AVX512CD that provides lzcnt instructions we can use the pattern (cttz_undef(x) = (width - 1) - ctlz(x & -x)) Differential Revision: http://reviews.llvm.org/D12663 llvm-svn: 248091
2015-09-19 21:22:57 +08:00
; SSE2-NEXT: psrlw $8, %xmm0
[X86] Added tzcnt vector tests. llvm-svn: 239264
2015-06-08 05:01:34 +08:00
; SSE2-NEXT: retq
;
; SSE3-LABEL: testv8i16:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; SSE3: # %bb.0:
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; SSE3-NEXT: pcmpeqd %xmm1, %xmm1
; SSE3-NEXT: paddw %xmm0, %xmm1
; SSE3-NEXT: pandn %xmm1, %xmm0
[x86] transform vector inc/dec to use -1 constant (PR33483) Convert vector increment or decrement to sub/add with an all-ones constant: add X, <1, 1...> --> sub X, <-1, -1...> sub X, <1, 1...> --> add X, <-1, -1...> The all-ones vector constant can be materialized using a pcmpeq instruction that is commonly recognized as an idiom (has no register dependency), so that's better than loading a splat 1 constant. AVX512 uses 'vpternlogd' for 512-bit vectors because there is apparently no better way to produce 512 one-bits. The general advantages of this lowering are: 1. pcmpeq has lower latency than a memop on every uarch I looked at in Agner's tables, so in theory, this could be better for perf, but... 2. That seems unlikely to affect any OOO implementation, and I can't measure any real perf difference from this transform on Haswell or Jaguar, but... 3. It doesn't look like it from the diffs, but this is an overall size win because we eliminate 16 - 64 constant bytes in the case of a vector load. If we're broadcasting a scalar load (which might itself be a bug), then we're replacing a scalar constant load + broadcast with a single cheap op, so that should always be smaller/better too. 4. This makes the DAG/isel output more consistent - we use pcmpeq already for padd x, -1 and psub x, -1, so we should use that form for +1 too because we can. If there's some reason to favor a constant load on some CPU, let's make the reverse transform for all of these cases (either here in the DAG or in a later machine pass). This should fix: https://bugs.llvm.org/show_bug.cgi?id=33483 Differential Revision: https://reviews.llvm.org/D34336 llvm-svn: 306289
2017-06-26 22:19:26 +08:00
; SSE3-NEXT: movdqa %xmm0, %xmm1
; SSE3-NEXT: psrlw $1, %xmm1
; SSE3-NEXT: pand {{.*}}(%rip), %xmm1
[X86][SSE] LowerVectorCTPOP - pull out repeated byte sum stage. Pull out repeated byte sum stage for popcount of vector elements > 8bits. This allows us to simplify the LUT/BITMATH popcnt code to always assume vXi8 vectors, and also improves avx512bitalg codegen which only has access to vpopcntb/vpopcntw. llvm-svn: 344348
2018-10-12 22:18:47 +08:00
; SSE3-NEXT: psubb %xmm1, %xmm0
; SSE3-NEXT: movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
[x86] transform vector inc/dec to use -1 constant (PR33483) Convert vector increment or decrement to sub/add with an all-ones constant: add X, <1, 1...> --> sub X, <-1, -1...> sub X, <1, 1...> --> add X, <-1, -1...> The all-ones vector constant can be materialized using a pcmpeq instruction that is commonly recognized as an idiom (has no register dependency), so that's better than loading a splat 1 constant. AVX512 uses 'vpternlogd' for 512-bit vectors because there is apparently no better way to produce 512 one-bits. The general advantages of this lowering are: 1. pcmpeq has lower latency than a memop on every uarch I looked at in Agner's tables, so in theory, this could be better for perf, but... 2. That seems unlikely to affect any OOO implementation, and I can't measure any real perf difference from this transform on Haswell or Jaguar, but... 3. It doesn't look like it from the diffs, but this is an overall size win because we eliminate 16 - 64 constant bytes in the case of a vector load. If we're broadcasting a scalar load (which might itself be a bug), then we're replacing a scalar constant load + broadcast with a single cheap op, so that should always be smaller/better too. 4. This makes the DAG/isel output more consistent - we use pcmpeq already for padd x, -1 and psub x, -1, so we should use that form for +1 too because we can. If there's some reason to favor a constant load on some CPU, let's make the reverse transform for all of these cases (either here in the DAG or in a later machine pass). This should fix: https://bugs.llvm.org/show_bug.cgi?id=33483 Differential Revision: https://reviews.llvm.org/D34336 llvm-svn: 306289
2017-06-26 22:19:26 +08:00
; SSE3-NEXT: movdqa %xmm0, %xmm2
; SSE3-NEXT: pand %xmm1, %xmm2
; SSE3-NEXT: psrlw $2, %xmm0
; SSE3-NEXT: pand %xmm1, %xmm0
[X86][SSE] LowerVectorCTPOP - pull out repeated byte sum stage. Pull out repeated byte sum stage for popcount of vector elements > 8bits. This allows us to simplify the LUT/BITMATH popcnt code to always assume vXi8 vectors, and also improves avx512bitalg codegen which only has access to vpopcntb/vpopcntw. llvm-svn: 344348
2018-10-12 22:18:47 +08:00
; SSE3-NEXT: paddb %xmm2, %xmm0
[x86] transform vector inc/dec to use -1 constant (PR33483) Convert vector increment or decrement to sub/add with an all-ones constant: add X, <1, 1...> --> sub X, <-1, -1...> sub X, <1, 1...> --> add X, <-1, -1...> The all-ones vector constant can be materialized using a pcmpeq instruction that is commonly recognized as an idiom (has no register dependency), so that's better than loading a splat 1 constant. AVX512 uses 'vpternlogd' for 512-bit vectors because there is apparently no better way to produce 512 one-bits. The general advantages of this lowering are: 1. pcmpeq has lower latency than a memop on every uarch I looked at in Agner's tables, so in theory, this could be better for perf, but... 2. That seems unlikely to affect any OOO implementation, and I can't measure any real perf difference from this transform on Haswell or Jaguar, but... 3. It doesn't look like it from the diffs, but this is an overall size win because we eliminate 16 - 64 constant bytes in the case of a vector load. If we're broadcasting a scalar load (which might itself be a bug), then we're replacing a scalar constant load + broadcast with a single cheap op, so that should always be smaller/better too. 4. This makes the DAG/isel output more consistent - we use pcmpeq already for padd x, -1 and psub x, -1, so we should use that form for +1 too because we can. If there's some reason to favor a constant load on some CPU, let's make the reverse transform for all of these cases (either here in the DAG or in a later machine pass). This should fix: https://bugs.llvm.org/show_bug.cgi?id=33483 Differential Revision: https://reviews.llvm.org/D34336 llvm-svn: 306289
2017-06-26 22:19:26 +08:00
; SSE3-NEXT: movdqa %xmm0, %xmm1
; SSE3-NEXT: psrlw $4, %xmm1
[X86][SSE] LowerVectorCTPOP - pull out repeated byte sum stage. Pull out repeated byte sum stage for popcount of vector elements > 8bits. This allows us to simplify the LUT/BITMATH popcnt code to always assume vXi8 vectors, and also improves avx512bitalg codegen which only has access to vpopcntb/vpopcntw. llvm-svn: 344348
2018-10-12 22:18:47 +08:00
; SSE3-NEXT: paddb %xmm0, %xmm1
[x86] transform vector inc/dec to use -1 constant (PR33483) Convert vector increment or decrement to sub/add with an all-ones constant: add X, <1, 1...> --> sub X, <-1, -1...> sub X, <1, 1...> --> add X, <-1, -1...> The all-ones vector constant can be materialized using a pcmpeq instruction that is commonly recognized as an idiom (has no register dependency), so that's better than loading a splat 1 constant. AVX512 uses 'vpternlogd' for 512-bit vectors because there is apparently no better way to produce 512 one-bits. The general advantages of this lowering are: 1. pcmpeq has lower latency than a memop on every uarch I looked at in Agner's tables, so in theory, this could be better for perf, but... 2. That seems unlikely to affect any OOO implementation, and I can't measure any real perf difference from this transform on Haswell or Jaguar, but... 3. It doesn't look like it from the diffs, but this is an overall size win because we eliminate 16 - 64 constant bytes in the case of a vector load. If we're broadcasting a scalar load (which might itself be a bug), then we're replacing a scalar constant load + broadcast with a single cheap op, so that should always be smaller/better too. 4. This makes the DAG/isel output more consistent - we use pcmpeq already for padd x, -1 and psub x, -1, so we should use that form for +1 too because we can. If there's some reason to favor a constant load on some CPU, let's make the reverse transform for all of these cases (either here in the DAG or in a later machine pass). This should fix: https://bugs.llvm.org/show_bug.cgi?id=33483 Differential Revision: https://reviews.llvm.org/D34336 llvm-svn: 306289
2017-06-26 22:19:26 +08:00
; SSE3-NEXT: pand {{.*}}(%rip), %xmm1
[X86][SSE] Vectorize CTTZ + CTTZ_ZERO_UNDEF Now that we have fast vector CTPOP implementations we can use this to speed up vector CTTZ using the pattern (cttz(x) = ctpop((x & -x) - 1)) Additionally, for AVX512CD that provides lzcnt instructions we can use the pattern (cttz_undef(x) = (width - 1) - ctlz(x & -x)) Differential Revision: http://reviews.llvm.org/D12663 llvm-svn: 248091
2015-09-19 21:22:57 +08:00
; SSE3-NEXT: movdqa %xmm1, %xmm0
; SSE3-NEXT: psllw $8, %xmm0
[x86] transform vector inc/dec to use -1 constant (PR33483) Convert vector increment or decrement to sub/add with an all-ones constant: add X, <1, 1...> --> sub X, <-1, -1...> sub X, <1, 1...> --> add X, <-1, -1...> The all-ones vector constant can be materialized using a pcmpeq instruction that is commonly recognized as an idiom (has no register dependency), so that's better than loading a splat 1 constant. AVX512 uses 'vpternlogd' for 512-bit vectors because there is apparently no better way to produce 512 one-bits. The general advantages of this lowering are: 1. pcmpeq has lower latency than a memop on every uarch I looked at in Agner's tables, so in theory, this could be better for perf, but... 2. That seems unlikely to affect any OOO implementation, and I can't measure any real perf difference from this transform on Haswell or Jaguar, but... 3. It doesn't look like it from the diffs, but this is an overall size win because we eliminate 16 - 64 constant bytes in the case of a vector load. If we're broadcasting a scalar load (which might itself be a bug), then we're replacing a scalar constant load + broadcast with a single cheap op, so that should always be smaller/better too. 4. This makes the DAG/isel output more consistent - we use pcmpeq already for padd x, -1 and psub x, -1, so we should use that form for +1 too because we can. If there's some reason to favor a constant load on some CPU, let's make the reverse transform for all of these cases (either here in the DAG or in a later machine pass). This should fix: https://bugs.llvm.org/show_bug.cgi?id=33483 Differential Revision: https://reviews.llvm.org/D34336 llvm-svn: 306289
2017-06-26 22:19:26 +08:00
; SSE3-NEXT: paddb %xmm1, %xmm0
[X86][SSE] Vectorize CTTZ + CTTZ_ZERO_UNDEF Now that we have fast vector CTPOP implementations we can use this to speed up vector CTTZ using the pattern (cttz(x) = ctpop((x & -x) - 1)) Additionally, for AVX512CD that provides lzcnt instructions we can use the pattern (cttz_undef(x) = (width - 1) - ctlz(x & -x)) Differential Revision: http://reviews.llvm.org/D12663 llvm-svn: 248091
2015-09-19 21:22:57 +08:00
; SSE3-NEXT: psrlw $8, %xmm0
[X86] Added tzcnt vector tests. llvm-svn: 239264
2015-06-08 05:01:34 +08:00
; SSE3-NEXT: retq
;
; SSSE3-LABEL: testv8i16:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; SSSE3: # %bb.0:
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; SSSE3-NEXT: pcmpeqd %xmm1, %xmm1
; SSSE3-NEXT: paddw %xmm0, %xmm1
; SSSE3-NEXT: pandn %xmm1, %xmm0
[x86] transform vector inc/dec to use -1 constant (PR33483) Convert vector increment or decrement to sub/add with an all-ones constant: add X, <1, 1...> --> sub X, <-1, -1...> sub X, <1, 1...> --> add X, <-1, -1...> The all-ones vector constant can be materialized using a pcmpeq instruction that is commonly recognized as an idiom (has no register dependency), so that's better than loading a splat 1 constant. AVX512 uses 'vpternlogd' for 512-bit vectors because there is apparently no better way to produce 512 one-bits. The general advantages of this lowering are: 1. pcmpeq has lower latency than a memop on every uarch I looked at in Agner's tables, so in theory, this could be better for perf, but... 2. That seems unlikely to affect any OOO implementation, and I can't measure any real perf difference from this transform on Haswell or Jaguar, but... 3. It doesn't look like it from the diffs, but this is an overall size win because we eliminate 16 - 64 constant bytes in the case of a vector load. If we're broadcasting a scalar load (which might itself be a bug), then we're replacing a scalar constant load + broadcast with a single cheap op, so that should always be smaller/better too. 4. This makes the DAG/isel output more consistent - we use pcmpeq already for padd x, -1 and psub x, -1, so we should use that form for +1 too because we can. If there's some reason to favor a constant load on some CPU, let's make the reverse transform for all of these cases (either here in the DAG or in a later machine pass). This should fix: https://bugs.llvm.org/show_bug.cgi?id=33483 Differential Revision: https://reviews.llvm.org/D34336 llvm-svn: 306289
2017-06-26 22:19:26 +08:00
; SSSE3-NEXT: movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
; SSSE3-NEXT: movdqa %xmm0, %xmm2
; SSSE3-NEXT: pand %xmm1, %xmm2
[X86][SSE] Vectorize CTTZ + CTTZ_ZERO_UNDEF Now that we have fast vector CTPOP implementations we can use this to speed up vector CTTZ using the pattern (cttz(x) = ctpop((x & -x) - 1)) Additionally, for AVX512CD that provides lzcnt instructions we can use the pattern (cttz_undef(x) = (width - 1) - ctlz(x & -x)) Differential Revision: http://reviews.llvm.org/D12663 llvm-svn: 248091
2015-09-19 21:22:57 +08:00
; SSSE3-NEXT: movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
; SSSE3-NEXT: movdqa %xmm3, %xmm4
; SSSE3-NEXT: pshufb %xmm2, %xmm4
[x86] transform vector inc/dec to use -1 constant (PR33483) Convert vector increment or decrement to sub/add with an all-ones constant: add X, <1, 1...> --> sub X, <-1, -1...> sub X, <1, 1...> --> add X, <-1, -1...> The all-ones vector constant can be materialized using a pcmpeq instruction that is commonly recognized as an idiom (has no register dependency), so that's better than loading a splat 1 constant. AVX512 uses 'vpternlogd' for 512-bit vectors because there is apparently no better way to produce 512 one-bits. The general advantages of this lowering are: 1. pcmpeq has lower latency than a memop on every uarch I looked at in Agner's tables, so in theory, this could be better for perf, but... 2. That seems unlikely to affect any OOO implementation, and I can't measure any real perf difference from this transform on Haswell or Jaguar, but... 3. It doesn't look like it from the diffs, but this is an overall size win because we eliminate 16 - 64 constant bytes in the case of a vector load. If we're broadcasting a scalar load (which might itself be a bug), then we're replacing a scalar constant load + broadcast with a single cheap op, so that should always be smaller/better too. 4. This makes the DAG/isel output more consistent - we use pcmpeq already for padd x, -1 and psub x, -1, so we should use that form for +1 too because we can. If there's some reason to favor a constant load on some CPU, let's make the reverse transform for all of these cases (either here in the DAG or in a later machine pass). This should fix: https://bugs.llvm.org/show_bug.cgi?id=33483 Differential Revision: https://reviews.llvm.org/D34336 llvm-svn: 306289
2017-06-26 22:19:26 +08:00
; SSSE3-NEXT: psrlw $4, %xmm0
; SSSE3-NEXT: pand %xmm1, %xmm0
; SSSE3-NEXT: pshufb %xmm0, %xmm3
[X86][SSE] Vectorize CTTZ + CTTZ_ZERO_UNDEF Now that we have fast vector CTPOP implementations we can use this to speed up vector CTTZ using the pattern (cttz(x) = ctpop((x & -x) - 1)) Additionally, for AVX512CD that provides lzcnt instructions we can use the pattern (cttz_undef(x) = (width - 1) - ctlz(x & -x)) Differential Revision: http://reviews.llvm.org/D12663 llvm-svn: 248091
2015-09-19 21:22:57 +08:00
; SSSE3-NEXT: paddb %xmm4, %xmm3
; SSSE3-NEXT: movdqa %xmm3, %xmm0
; SSSE3-NEXT: psllw $8, %xmm0
; SSSE3-NEXT: paddb %xmm3, %xmm0
; SSSE3-NEXT: psrlw $8, %xmm0
[X86] Added tzcnt vector tests. llvm-svn: 239264
2015-06-08 05:01:34 +08:00
; SSSE3-NEXT: retq
;
; SSE41-LABEL: testv8i16:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; SSE41: # %bb.0:
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; SSE41-NEXT: pcmpeqd %xmm1, %xmm1
; SSE41-NEXT: paddw %xmm0, %xmm1
; SSE41-NEXT: pandn %xmm1, %xmm0
[x86] transform vector inc/dec to use -1 constant (PR33483) Convert vector increment or decrement to sub/add with an all-ones constant: add X, <1, 1...> --> sub X, <-1, -1...> sub X, <1, 1...> --> add X, <-1, -1...> The all-ones vector constant can be materialized using a pcmpeq instruction that is commonly recognized as an idiom (has no register dependency), so that's better than loading a splat 1 constant. AVX512 uses 'vpternlogd' for 512-bit vectors because there is apparently no better way to produce 512 one-bits. The general advantages of this lowering are: 1. pcmpeq has lower latency than a memop on every uarch I looked at in Agner's tables, so in theory, this could be better for perf, but... 2. That seems unlikely to affect any OOO implementation, and I can't measure any real perf difference from this transform on Haswell or Jaguar, but... 3. It doesn't look like it from the diffs, but this is an overall size win because we eliminate 16 - 64 constant bytes in the case of a vector load. If we're broadcasting a scalar load (which might itself be a bug), then we're replacing a scalar constant load + broadcast with a single cheap op, so that should always be smaller/better too. 4. This makes the DAG/isel output more consistent - we use pcmpeq already for padd x, -1 and psub x, -1, so we should use that form for +1 too because we can. If there's some reason to favor a constant load on some CPU, let's make the reverse transform for all of these cases (either here in the DAG or in a later machine pass). This should fix: https://bugs.llvm.org/show_bug.cgi?id=33483 Differential Revision: https://reviews.llvm.org/D34336 llvm-svn: 306289
2017-06-26 22:19:26 +08:00
; SSE41-NEXT: movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
; SSE41-NEXT: movdqa %xmm0, %xmm2
; SSE41-NEXT: pand %xmm1, %xmm2
[X86][SSE] Vectorize CTTZ + CTTZ_ZERO_UNDEF Now that we have fast vector CTPOP implementations we can use this to speed up vector CTTZ using the pattern (cttz(x) = ctpop((x & -x) - 1)) Additionally, for AVX512CD that provides lzcnt instructions we can use the pattern (cttz_undef(x) = (width - 1) - ctlz(x & -x)) Differential Revision: http://reviews.llvm.org/D12663 llvm-svn: 248091
2015-09-19 21:22:57 +08:00
; SSE41-NEXT: movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
; SSE41-NEXT: movdqa %xmm3, %xmm4
; SSE41-NEXT: pshufb %xmm2, %xmm4
[x86] transform vector inc/dec to use -1 constant (PR33483) Convert vector increment or decrement to sub/add with an all-ones constant: add X, <1, 1...> --> sub X, <-1, -1...> sub X, <1, 1...> --> add X, <-1, -1...> The all-ones vector constant can be materialized using a pcmpeq instruction that is commonly recognized as an idiom (has no register dependency), so that's better than loading a splat 1 constant. AVX512 uses 'vpternlogd' for 512-bit vectors because there is apparently no better way to produce 512 one-bits. The general advantages of this lowering are: 1. pcmpeq has lower latency than a memop on every uarch I looked at in Agner's tables, so in theory, this could be better for perf, but... 2. That seems unlikely to affect any OOO implementation, and I can't measure any real perf difference from this transform on Haswell or Jaguar, but... 3. It doesn't look like it from the diffs, but this is an overall size win because we eliminate 16 - 64 constant bytes in the case of a vector load. If we're broadcasting a scalar load (which might itself be a bug), then we're replacing a scalar constant load + broadcast with a single cheap op, so that should always be smaller/better too. 4. This makes the DAG/isel output more consistent - we use pcmpeq already for padd x, -1 and psub x, -1, so we should use that form for +1 too because we can. If there's some reason to favor a constant load on some CPU, let's make the reverse transform for all of these cases (either here in the DAG or in a later machine pass). This should fix: https://bugs.llvm.org/show_bug.cgi?id=33483 Differential Revision: https://reviews.llvm.org/D34336 llvm-svn: 306289
2017-06-26 22:19:26 +08:00
; SSE41-NEXT: psrlw $4, %xmm0
; SSE41-NEXT: pand %xmm1, %xmm0
; SSE41-NEXT: pshufb %xmm0, %xmm3
[X86][SSE] Vectorize CTTZ + CTTZ_ZERO_UNDEF Now that we have fast vector CTPOP implementations we can use this to speed up vector CTTZ using the pattern (cttz(x) = ctpop((x & -x) - 1)) Additionally, for AVX512CD that provides lzcnt instructions we can use the pattern (cttz_undef(x) = (width - 1) - ctlz(x & -x)) Differential Revision: http://reviews.llvm.org/D12663 llvm-svn: 248091
2015-09-19 21:22:57 +08:00
; SSE41-NEXT: paddb %xmm4, %xmm3
; SSE41-NEXT: movdqa %xmm3, %xmm0
; SSE41-NEXT: psllw $8, %xmm0
; SSE41-NEXT: paddb %xmm3, %xmm0
; SSE41-NEXT: psrlw $8, %xmm0
[X86] Added tzcnt vector tests. llvm-svn: 239264
2015-06-08 05:01:34 +08:00
; SSE41-NEXT: retq
;
[X86][AVX] Regenerate tzcnt tests. The EVEX -> VEX fix means that AVX/AVX512 code is more likely the same now. llvm-svn: 291241
2017-01-06 23:54:23 +08:00
; AVX-LABEL: testv8i16:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; AVX: # %bb.0:
[x86] transform vector inc/dec to use -1 constant (PR33483) Convert vector increment or decrement to sub/add with an all-ones constant: add X, <1, 1...> --> sub X, <-1, -1...> sub X, <1, 1...> --> add X, <-1, -1...> The all-ones vector constant can be materialized using a pcmpeq instruction that is commonly recognized as an idiom (has no register dependency), so that's better than loading a splat 1 constant. AVX512 uses 'vpternlogd' for 512-bit vectors because there is apparently no better way to produce 512 one-bits. The general advantages of this lowering are: 1. pcmpeq has lower latency than a memop on every uarch I looked at in Agner's tables, so in theory, this could be better for perf, but... 2. That seems unlikely to affect any OOO implementation, and I can't measure any real perf difference from this transform on Haswell or Jaguar, but... 3. It doesn't look like it from the diffs, but this is an overall size win because we eliminate 16 - 64 constant bytes in the case of a vector load. If we're broadcasting a scalar load (which might itself be a bug), then we're replacing a scalar constant load + broadcast with a single cheap op, so that should always be smaller/better too. 4. This makes the DAG/isel output more consistent - we use pcmpeq already for padd x, -1 and psub x, -1, so we should use that form for +1 too because we can. If there's some reason to favor a constant load on some CPU, let's make the reverse transform for all of these cases (either here in the DAG or in a later machine pass). This should fix: https://bugs.llvm.org/show_bug.cgi?id=33483 Differential Revision: https://reviews.llvm.org/D34336 llvm-svn: 306289
2017-06-26 22:19:26 +08:00
; AVX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; AVX-NEXT: vpaddw %xmm1, %xmm0, %xmm1
; AVX-NEXT: vpandn %xmm1, %xmm0, %xmm0
[X86][AVX] Regenerate tzcnt tests. The EVEX -> VEX fix means that AVX/AVX512 code is more likely the same now. llvm-svn: 291241
2017-01-06 23:54:23 +08:00
; AVX-NEXT: vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
; AVX-NEXT: vpand %xmm1, %xmm0, %xmm2
; AVX-NEXT: vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
; AVX-NEXT: vpshufb %xmm2, %xmm3, %xmm2
; AVX-NEXT: vpsrlw $4, %xmm0, %xmm0
; AVX-NEXT: vpand %xmm1, %xmm0, %xmm0
; AVX-NEXT: vpshufb %xmm0, %xmm3, %xmm0
; AVX-NEXT: vpaddb %xmm2, %xmm0, %xmm0
; AVX-NEXT: vpsllw $8, %xmm0, %xmm1
; AVX-NEXT: vpaddb %xmm0, %xmm1, %xmm0
; AVX-NEXT: vpsrlw $8, %xmm0, %xmm0
; AVX-NEXT: retq
[AVX512] Teach lowering to use vplzcntd/q to implement 128/256-bit CTTZ_ZERO_UNDEF even without VLX support. We can just extend to 512-bits and extract like we do for CTLZ. llvm-svn: 267100
2016-04-22 11:22:38 +08:00
;
[X86] Adding vpopcntd and vpopcntq instructions AVX512_VPOPCNTDQ is a new feature set that was published by Intel. The patch represents the LLVM side of the addition of two new intrinsic based instructions (vpopcntd and vpopcntq). Differential Revision: https://reviews.llvm.org/D33169 llvm-svn: 303858
2017-05-25 21:45:23 +08:00
; AVX512VPOPCNTDQ-LABEL: testv8i16:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; AVX512VPOPCNTDQ: # %bb.0:
[x86] transform vector inc/dec to use -1 constant (PR33483) Convert vector increment or decrement to sub/add with an all-ones constant: add X, <1, 1...> --> sub X, <-1, -1...> sub X, <1, 1...> --> add X, <-1, -1...> The all-ones vector constant can be materialized using a pcmpeq instruction that is commonly recognized as an idiom (has no register dependency), so that's better than loading a splat 1 constant. AVX512 uses 'vpternlogd' for 512-bit vectors because there is apparently no better way to produce 512 one-bits. The general advantages of this lowering are: 1. pcmpeq has lower latency than a memop on every uarch I looked at in Agner's tables, so in theory, this could be better for perf, but... 2. That seems unlikely to affect any OOO implementation, and I can't measure any real perf difference from this transform on Haswell or Jaguar, but... 3. It doesn't look like it from the diffs, but this is an overall size win because we eliminate 16 - 64 constant bytes in the case of a vector load. If we're broadcasting a scalar load (which might itself be a bug), then we're replacing a scalar constant load + broadcast with a single cheap op, so that should always be smaller/better too. 4. This makes the DAG/isel output more consistent - we use pcmpeq already for padd x, -1 and psub x, -1, so we should use that form for +1 too because we can. If there's some reason to favor a constant load on some CPU, let's make the reverse transform for all of these cases (either here in the DAG or in a later machine pass). This should fix: https://bugs.llvm.org/show_bug.cgi?id=33483 Differential Revision: https://reviews.llvm.org/D34336 llvm-svn: 306289
2017-06-26 22:19:26 +08:00
; AVX512VPOPCNTDQ-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; AVX512VPOPCNTDQ-NEXT: vpaddw %xmm1, %xmm0, %xmm1
; AVX512VPOPCNTDQ-NEXT: vpandn %xmm1, %xmm0, %xmm0
[X86] When using vpopcntdq for ctpop of v8i16 vectors, only promote to v8i32. Previously we promoted to v8i64, but we don't need to go all the way to 512-bits. If we have VLX we can use the 256-bit instruction. And even if we don't have VLX we can widen v8i32 to v16i32 and drop the upper half. llvm-svn: 320926
2017-12-17 03:31:36 +08:00
; AVX512VPOPCNTDQ-NEXT: vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
; AVX512VPOPCNTDQ-NEXT: vpopcntd %zmm0, %zmm0
; AVX512VPOPCNTDQ-NEXT: vpmovdw %zmm0, %ymm0
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
; AVX512VPOPCNTDQ-NEXT: # kill: def $xmm0 killed $xmm0 killed $ymm0
[X86][AVX512VPOPCNTDQ] Improve support for v16i8/v8i16/v16i16/ CTPOP Zero extend to v16i32/v8i64, use VPOPCNTDQ instructions and truncate back. llvm-svn: 306990
2017-07-03 03:32:37 +08:00
; AVX512VPOPCNTDQ-NEXT: vzeroupper
[X86] Adding vpopcntd and vpopcntq instructions AVX512_VPOPCNTDQ is a new feature set that was published by Intel. The patch represents the LLVM side of the addition of two new intrinsic based instructions (vpopcntd and vpopcntq). Differential Revision: https://reviews.llvm.org/D33169 llvm-svn: 303858
2017-05-25 21:45:23 +08:00
; AVX512VPOPCNTDQ-NEXT: retq
;
[X86] Add 128 and 256-bit VPOPCNTDQ instructions. Adjust some tablegen classes LZCNT/POPCNT. I think when this instruction was first published it was only for a Knights CPU and thus VLX version was missing. llvm-svn: 320910
2017-12-16 10:40:28 +08:00
; AVX512VPOPCNTDQVL-LABEL: testv8i16:
; AVX512VPOPCNTDQVL: # %bb.0:
; AVX512VPOPCNTDQVL-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; AVX512VPOPCNTDQVL-NEXT: vpaddw %xmm1, %xmm0, %xmm1
; AVX512VPOPCNTDQVL-NEXT: vpandn %xmm1, %xmm0, %xmm0
[X86] When using vpopcntdq for ctpop of v8i16 vectors, only promote to v8i32. Previously we promoted to v8i64, but we don't need to go all the way to 512-bits. If we have VLX we can use the 256-bit instruction. And even if we don't have VLX we can widen v8i32 to v16i32 and drop the upper half. llvm-svn: 320926
2017-12-17 03:31:36 +08:00
; AVX512VPOPCNTDQVL-NEXT: vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
; AVX512VPOPCNTDQVL-NEXT: vpopcntd %ymm0, %ymm0
; AVX512VPOPCNTDQVL-NEXT: vpmovdw %ymm0, %xmm0
[X86] Add 128 and 256-bit VPOPCNTDQ instructions. Adjust some tablegen classes LZCNT/POPCNT. I think when this instruction was first published it was only for a Knights CPU and thus VLX version was missing. llvm-svn: 320910
2017-12-16 10:40:28 +08:00
; AVX512VPOPCNTDQVL-NEXT: vzeroupper
; AVX512VPOPCNTDQVL-NEXT: retq
;
[x86][icelake]BITALG vpopcnt{b,w} Differential Revision: https://reviews.llvm.org/D40213 llvm-svn: 318748
2017-11-21 18:32:42 +08:00
; BITALG_NOVLX-LABEL: testv8i16:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; BITALG_NOVLX: # %bb.0:
[x86][icelake]BITALG vpopcnt{b,w} Differential Revision: https://reviews.llvm.org/D40213 llvm-svn: 318748
2017-11-21 18:32:42 +08:00
; BITALG_NOVLX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; BITALG_NOVLX-NEXT: vpaddw %xmm1, %xmm0, %xmm1
; BITALG_NOVLX-NEXT: vpandn %xmm1, %xmm0, %xmm0
[x86][icelake]BITALG vpopcnt{b,w} Differential Revision: https://reviews.llvm.org/D40213 llvm-svn: 318748
2017-11-21 18:32:42 +08:00
; BITALG_NOVLX-NEXT: vpopcntw %zmm0, %zmm0
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
; BITALG_NOVLX-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
[x86][icelake]BITALG vpopcnt{b,w} Differential Revision: https://reviews.llvm.org/D40213 llvm-svn: 318748
2017-11-21 18:32:42 +08:00
; BITALG_NOVLX-NEXT: vzeroupper
; BITALG_NOVLX-NEXT: retq
;
; BITALG-LABEL: testv8i16:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; BITALG: # %bb.0:
[x86][icelake]BITALG vpopcnt{b,w} Differential Revision: https://reviews.llvm.org/D40213 llvm-svn: 318748
2017-11-21 18:32:42 +08:00
; BITALG-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; BITALG-NEXT: vpaddw %xmm1, %xmm0, %xmm1
; BITALG-NEXT: vpandn %xmm1, %xmm0, %xmm0
[x86][icelake]BITALG vpopcnt{b,w} Differential Revision: https://reviews.llvm.org/D40213 llvm-svn: 318748
2017-11-21 18:32:42 +08:00
; BITALG-NEXT: vpopcntw %xmm0, %xmm0
; BITALG-NEXT: retq
;
[X86][SSE] Added 32-bit tests for vector lzcnt/tzcnt tests v2i64 tests are particularly bad on 32-bit targets. llvm-svn: 265789
2016-04-08 23:01:31 +08:00
; X32-SSE-LABEL: testv8i16:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; X32-SSE: # %bb.0:
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; X32-SSE-NEXT: pcmpeqd %xmm1, %xmm1
; X32-SSE-NEXT: paddw %xmm0, %xmm1
; X32-SSE-NEXT: pandn %xmm1, %xmm0
[x86] transform vector inc/dec to use -1 constant (PR33483) Convert vector increment or decrement to sub/add with an all-ones constant: add X, <1, 1...> --> sub X, <-1, -1...> sub X, <1, 1...> --> add X, <-1, -1...> The all-ones vector constant can be materialized using a pcmpeq instruction that is commonly recognized as an idiom (has no register dependency), so that's better than loading a splat 1 constant. AVX512 uses 'vpternlogd' for 512-bit vectors because there is apparently no better way to produce 512 one-bits. The general advantages of this lowering are: 1. pcmpeq has lower latency than a memop on every uarch I looked at in Agner's tables, so in theory, this could be better for perf, but... 2. That seems unlikely to affect any OOO implementation, and I can't measure any real perf difference from this transform on Haswell or Jaguar, but... 3. It doesn't look like it from the diffs, but this is an overall size win because we eliminate 16 - 64 constant bytes in the case of a vector load. If we're broadcasting a scalar load (which might itself be a bug), then we're replacing a scalar constant load + broadcast with a single cheap op, so that should always be smaller/better too. 4. This makes the DAG/isel output more consistent - we use pcmpeq already for padd x, -1 and psub x, -1, so we should use that form for +1 too because we can. If there's some reason to favor a constant load on some CPU, let's make the reverse transform for all of these cases (either here in the DAG or in a later machine pass). This should fix: https://bugs.llvm.org/show_bug.cgi?id=33483 Differential Revision: https://reviews.llvm.org/D34336 llvm-svn: 306289
2017-06-26 22:19:26 +08:00
; X32-SSE-NEXT: movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
; X32-SSE-NEXT: movdqa %xmm0, %xmm2
; X32-SSE-NEXT: pand %xmm1, %xmm2
[X86][SSE] Added 32-bit tests for vector lzcnt/tzcnt tests v2i64 tests are particularly bad on 32-bit targets. llvm-svn: 265789
2016-04-08 23:01:31 +08:00
; X32-SSE-NEXT: movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
; X32-SSE-NEXT: movdqa %xmm3, %xmm4
; X32-SSE-NEXT: pshufb %xmm2, %xmm4
[x86] transform vector inc/dec to use -1 constant (PR33483) Convert vector increment or decrement to sub/add with an all-ones constant: add X, <1, 1...> --> sub X, <-1, -1...> sub X, <1, 1...> --> add X, <-1, -1...> The all-ones vector constant can be materialized using a pcmpeq instruction that is commonly recognized as an idiom (has no register dependency), so that's better than loading a splat 1 constant. AVX512 uses 'vpternlogd' for 512-bit vectors because there is apparently no better way to produce 512 one-bits. The general advantages of this lowering are: 1. pcmpeq has lower latency than a memop on every uarch I looked at in Agner's tables, so in theory, this could be better for perf, but... 2. That seems unlikely to affect any OOO implementation, and I can't measure any real perf difference from this transform on Haswell or Jaguar, but... 3. It doesn't look like it from the diffs, but this is an overall size win because we eliminate 16 - 64 constant bytes in the case of a vector load. If we're broadcasting a scalar load (which might itself be a bug), then we're replacing a scalar constant load + broadcast with a single cheap op, so that should always be smaller/better too. 4. This makes the DAG/isel output more consistent - we use pcmpeq already for padd x, -1 and psub x, -1, so we should use that form for +1 too because we can. If there's some reason to favor a constant load on some CPU, let's make the reverse transform for all of these cases (either here in the DAG or in a later machine pass). This should fix: https://bugs.llvm.org/show_bug.cgi?id=33483 Differential Revision: https://reviews.llvm.org/D34336 llvm-svn: 306289
2017-06-26 22:19:26 +08:00
; X32-SSE-NEXT: psrlw $4, %xmm0
; X32-SSE-NEXT: pand %xmm1, %xmm0
; X32-SSE-NEXT: pshufb %xmm0, %xmm3
[X86][SSE] Added 32-bit tests for vector lzcnt/tzcnt tests v2i64 tests are particularly bad on 32-bit targets. llvm-svn: 265789
2016-04-08 23:01:31 +08:00
; X32-SSE-NEXT: paddb %xmm4, %xmm3
; X32-SSE-NEXT: movdqa %xmm3, %xmm0
; X32-SSE-NEXT: psllw $8, %xmm0
; X32-SSE-NEXT: paddb %xmm3, %xmm0
; X32-SSE-NEXT: psrlw $8, %xmm0
; X32-SSE-NEXT: retl
[X86] Added tzcnt vector tests. llvm-svn: 239264
2015-06-08 05:01:34 +08:00
%out = call <8 x i16> @llvm.cttz.v8i16(<8 x i16> %in, i1 0)
ret <8 x i16> %out
}
[X86][SSE] Tidied up vector CTLZ/CTTZ. NFCI. llvm-svn: 242645
2015-07-20 01:09:43 +08:00
define <8 x i16> @testv8i16u(<8 x i16> %in) nounwind {
[X86] Added tzcnt vector tests. llvm-svn: 239264
2015-06-08 05:01:34 +08:00
; SSE2-LABEL: testv8i16u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; SSE2: # %bb.0:
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; SSE2-NEXT: pcmpeqd %xmm1, %xmm1
; SSE2-NEXT: paddw %xmm0, %xmm1
; SSE2-NEXT: pandn %xmm1, %xmm0
[x86] transform vector inc/dec to use -1 constant (PR33483) Convert vector increment or decrement to sub/add with an all-ones constant: add X, <1, 1...> --> sub X, <-1, -1...> sub X, <1, 1...> --> add X, <-1, -1...> The all-ones vector constant can be materialized using a pcmpeq instruction that is commonly recognized as an idiom (has no register dependency), so that's better than loading a splat 1 constant. AVX512 uses 'vpternlogd' for 512-bit vectors because there is apparently no better way to produce 512 one-bits. The general advantages of this lowering are: 1. pcmpeq has lower latency than a memop on every uarch I looked at in Agner's tables, so in theory, this could be better for perf, but... 2. That seems unlikely to affect any OOO implementation, and I can't measure any real perf difference from this transform on Haswell or Jaguar, but... 3. It doesn't look like it from the diffs, but this is an overall size win because we eliminate 16 - 64 constant bytes in the case of a vector load. If we're broadcasting a scalar load (which might itself be a bug), then we're replacing a scalar constant load + broadcast with a single cheap op, so that should always be smaller/better too. 4. This makes the DAG/isel output more consistent - we use pcmpeq already for padd x, -1 and psub x, -1, so we should use that form for +1 too because we can. If there's some reason to favor a constant load on some CPU, let's make the reverse transform for all of these cases (either here in the DAG or in a later machine pass). This should fix: https://bugs.llvm.org/show_bug.cgi?id=33483 Differential Revision: https://reviews.llvm.org/D34336 llvm-svn: 306289
2017-06-26 22:19:26 +08:00
; SSE2-NEXT: movdqa %xmm0, %xmm1
; SSE2-NEXT: psrlw $1, %xmm1
; SSE2-NEXT: pand {{.*}}(%rip), %xmm1
[X86][SSE] LowerVectorCTPOP - pull out repeated byte sum stage. Pull out repeated byte sum stage for popcount of vector elements > 8bits. This allows us to simplify the LUT/BITMATH popcnt code to always assume vXi8 vectors, and also improves avx512bitalg codegen which only has access to vpopcntb/vpopcntw. llvm-svn: 344348
2018-10-12 22:18:47 +08:00
; SSE2-NEXT: psubb %xmm1, %xmm0
; SSE2-NEXT: movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
[x86] transform vector inc/dec to use -1 constant (PR33483) Convert vector increment or decrement to sub/add with an all-ones constant: add X, <1, 1...> --> sub X, <-1, -1...> sub X, <1, 1...> --> add X, <-1, -1...> The all-ones vector constant can be materialized using a pcmpeq instruction that is commonly recognized as an idiom (has no register dependency), so that's better than loading a splat 1 constant. AVX512 uses 'vpternlogd' for 512-bit vectors because there is apparently no better way to produce 512 one-bits. The general advantages of this lowering are: 1. pcmpeq has lower latency than a memop on every uarch I looked at in Agner's tables, so in theory, this could be better for perf, but... 2. That seems unlikely to affect any OOO implementation, and I can't measure any real perf difference from this transform on Haswell or Jaguar, but... 3. It doesn't look like it from the diffs, but this is an overall size win because we eliminate 16 - 64 constant bytes in the case of a vector load. If we're broadcasting a scalar load (which might itself be a bug), then we're replacing a scalar constant load + broadcast with a single cheap op, so that should always be smaller/better too. 4. This makes the DAG/isel output more consistent - we use pcmpeq already for padd x, -1 and psub x, -1, so we should use that form for +1 too because we can. If there's some reason to favor a constant load on some CPU, let's make the reverse transform for all of these cases (either here in the DAG or in a later machine pass). This should fix: https://bugs.llvm.org/show_bug.cgi?id=33483 Differential Revision: https://reviews.llvm.org/D34336 llvm-svn: 306289
2017-06-26 22:19:26 +08:00
; SSE2-NEXT: movdqa %xmm0, %xmm2
; SSE2-NEXT: pand %xmm1, %xmm2
; SSE2-NEXT: psrlw $2, %xmm0
; SSE2-NEXT: pand %xmm1, %xmm0
[X86][SSE] LowerVectorCTPOP - pull out repeated byte sum stage. Pull out repeated byte sum stage for popcount of vector elements > 8bits. This allows us to simplify the LUT/BITMATH popcnt code to always assume vXi8 vectors, and also improves avx512bitalg codegen which only has access to vpopcntb/vpopcntw. llvm-svn: 344348
2018-10-12 22:18:47 +08:00
; SSE2-NEXT: paddb %xmm2, %xmm0
[x86] transform vector inc/dec to use -1 constant (PR33483) Convert vector increment or decrement to sub/add with an all-ones constant: add X, <1, 1...> --> sub X, <-1, -1...> sub X, <1, 1...> --> add X, <-1, -1...> The all-ones vector constant can be materialized using a pcmpeq instruction that is commonly recognized as an idiom (has no register dependency), so that's better than loading a splat 1 constant. AVX512 uses 'vpternlogd' for 512-bit vectors because there is apparently no better way to produce 512 one-bits. The general advantages of this lowering are: 1. pcmpeq has lower latency than a memop on every uarch I looked at in Agner's tables, so in theory, this could be better for perf, but... 2. That seems unlikely to affect any OOO implementation, and I can't measure any real perf difference from this transform on Haswell or Jaguar, but... 3. It doesn't look like it from the diffs, but this is an overall size win because we eliminate 16 - 64 constant bytes in the case of a vector load. If we're broadcasting a scalar load (which might itself be a bug), then we're replacing a scalar constant load + broadcast with a single cheap op, so that should always be smaller/better too. 4. This makes the DAG/isel output more consistent - we use pcmpeq already for padd x, -1 and psub x, -1, so we should use that form for +1 too because we can. If there's some reason to favor a constant load on some CPU, let's make the reverse transform for all of these cases (either here in the DAG or in a later machine pass). This should fix: https://bugs.llvm.org/show_bug.cgi?id=33483 Differential Revision: https://reviews.llvm.org/D34336 llvm-svn: 306289
2017-06-26 22:19:26 +08:00
; SSE2-NEXT: movdqa %xmm0, %xmm1
; SSE2-NEXT: psrlw $4, %xmm1
[X86][SSE] LowerVectorCTPOP - pull out repeated byte sum stage. Pull out repeated byte sum stage for popcount of vector elements > 8bits. This allows us to simplify the LUT/BITMATH popcnt code to always assume vXi8 vectors, and also improves avx512bitalg codegen which only has access to vpopcntb/vpopcntw. llvm-svn: 344348
2018-10-12 22:18:47 +08:00
; SSE2-NEXT: paddb %xmm0, %xmm1
[x86] transform vector inc/dec to use -1 constant (PR33483) Convert vector increment or decrement to sub/add with an all-ones constant: add X, <1, 1...> --> sub X, <-1, -1...> sub X, <1, 1...> --> add X, <-1, -1...> The all-ones vector constant can be materialized using a pcmpeq instruction that is commonly recognized as an idiom (has no register dependency), so that's better than loading a splat 1 constant. AVX512 uses 'vpternlogd' for 512-bit vectors because there is apparently no better way to produce 512 one-bits. The general advantages of this lowering are: 1. pcmpeq has lower latency than a memop on every uarch I looked at in Agner's tables, so in theory, this could be better for perf, but... 2. That seems unlikely to affect any OOO implementation, and I can't measure any real perf difference from this transform on Haswell or Jaguar, but... 3. It doesn't look like it from the diffs, but this is an overall size win because we eliminate 16 - 64 constant bytes in the case of a vector load. If we're broadcasting a scalar load (which might itself be a bug), then we're replacing a scalar constant load + broadcast with a single cheap op, so that should always be smaller/better too. 4. This makes the DAG/isel output more consistent - we use pcmpeq already for padd x, -1 and psub x, -1, so we should use that form for +1 too because we can. If there's some reason to favor a constant load on some CPU, let's make the reverse transform for all of these cases (either here in the DAG or in a later machine pass). This should fix: https://bugs.llvm.org/show_bug.cgi?id=33483 Differential Revision: https://reviews.llvm.org/D34336 llvm-svn: 306289
2017-06-26 22:19:26 +08:00
; SSE2-NEXT: pand {{.*}}(%rip), %xmm1
[X86][SSE] Vectorize CTTZ + CTTZ_ZERO_UNDEF Now that we have fast vector CTPOP implementations we can use this to speed up vector CTTZ using the pattern (cttz(x) = ctpop((x & -x) - 1)) Additionally, for AVX512CD that provides lzcnt instructions we can use the pattern (cttz_undef(x) = (width - 1) - ctlz(x & -x)) Differential Revision: http://reviews.llvm.org/D12663 llvm-svn: 248091
2015-09-19 21:22:57 +08:00
; SSE2-NEXT: movdqa %xmm1, %xmm0
; SSE2-NEXT: psllw $8, %xmm0
[x86] transform vector inc/dec to use -1 constant (PR33483) Convert vector increment or decrement to sub/add with an all-ones constant: add X, <1, 1...> --> sub X, <-1, -1...> sub X, <1, 1...> --> add X, <-1, -1...> The all-ones vector constant can be materialized using a pcmpeq instruction that is commonly recognized as an idiom (has no register dependency), so that's better than loading a splat 1 constant. AVX512 uses 'vpternlogd' for 512-bit vectors because there is apparently no better way to produce 512 one-bits. The general advantages of this lowering are: 1. pcmpeq has lower latency than a memop on every uarch I looked at in Agner's tables, so in theory, this could be better for perf, but... 2. That seems unlikely to affect any OOO implementation, and I can't measure any real perf difference from this transform on Haswell or Jaguar, but... 3. It doesn't look like it from the diffs, but this is an overall size win because we eliminate 16 - 64 constant bytes in the case of a vector load. If we're broadcasting a scalar load (which might itself be a bug), then we're replacing a scalar constant load + broadcast with a single cheap op, so that should always be smaller/better too. 4. This makes the DAG/isel output more consistent - we use pcmpeq already for padd x, -1 and psub x, -1, so we should use that form for +1 too because we can. If there's some reason to favor a constant load on some CPU, let's make the reverse transform for all of these cases (either here in the DAG or in a later machine pass). This should fix: https://bugs.llvm.org/show_bug.cgi?id=33483 Differential Revision: https://reviews.llvm.org/D34336 llvm-svn: 306289
2017-06-26 22:19:26 +08:00
; SSE2-NEXT: paddb %xmm1, %xmm0
[X86][SSE] Vectorize CTTZ + CTTZ_ZERO_UNDEF Now that we have fast vector CTPOP implementations we can use this to speed up vector CTTZ using the pattern (cttz(x) = ctpop((x & -x) - 1)) Additionally, for AVX512CD that provides lzcnt instructions we can use the pattern (cttz_undef(x) = (width - 1) - ctlz(x & -x)) Differential Revision: http://reviews.llvm.org/D12663 llvm-svn: 248091
2015-09-19 21:22:57 +08:00
; SSE2-NEXT: psrlw $8, %xmm0
[X86] Added tzcnt vector tests. llvm-svn: 239264
2015-06-08 05:01:34 +08:00
; SSE2-NEXT: retq
;
; SSE3-LABEL: testv8i16u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; SSE3: # %bb.0:
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; SSE3-NEXT: pcmpeqd %xmm1, %xmm1
; SSE3-NEXT: paddw %xmm0, %xmm1
; SSE3-NEXT: pandn %xmm1, %xmm0
[x86] transform vector inc/dec to use -1 constant (PR33483) Convert vector increment or decrement to sub/add with an all-ones constant: add X, <1, 1...> --> sub X, <-1, -1...> sub X, <1, 1...> --> add X, <-1, -1...> The all-ones vector constant can be materialized using a pcmpeq instruction that is commonly recognized as an idiom (has no register dependency), so that's better than loading a splat 1 constant. AVX512 uses 'vpternlogd' for 512-bit vectors because there is apparently no better way to produce 512 one-bits. The general advantages of this lowering are: 1. pcmpeq has lower latency than a memop on every uarch I looked at in Agner's tables, so in theory, this could be better for perf, but... 2. That seems unlikely to affect any OOO implementation, and I can't measure any real perf difference from this transform on Haswell or Jaguar, but... 3. It doesn't look like it from the diffs, but this is an overall size win because we eliminate 16 - 64 constant bytes in the case of a vector load. If we're broadcasting a scalar load (which might itself be a bug), then we're replacing a scalar constant load + broadcast with a single cheap op, so that should always be smaller/better too. 4. This makes the DAG/isel output more consistent - we use pcmpeq already for padd x, -1 and psub x, -1, so we should use that form for +1 too because we can. If there's some reason to favor a constant load on some CPU, let's make the reverse transform for all of these cases (either here in the DAG or in a later machine pass). This should fix: https://bugs.llvm.org/show_bug.cgi?id=33483 Differential Revision: https://reviews.llvm.org/D34336 llvm-svn: 306289
2017-06-26 22:19:26 +08:00
; SSE3-NEXT: movdqa %xmm0, %xmm1
; SSE3-NEXT: psrlw $1, %xmm1
; SSE3-NEXT: pand {{.*}}(%rip), %xmm1
[X86][SSE] LowerVectorCTPOP - pull out repeated byte sum stage. Pull out repeated byte sum stage for popcount of vector elements > 8bits. This allows us to simplify the LUT/BITMATH popcnt code to always assume vXi8 vectors, and also improves avx512bitalg codegen which only has access to vpopcntb/vpopcntw. llvm-svn: 344348
2018-10-12 22:18:47 +08:00
; SSE3-NEXT: psubb %xmm1, %xmm0
; SSE3-NEXT: movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
[x86] transform vector inc/dec to use -1 constant (PR33483) Convert vector increment or decrement to sub/add with an all-ones constant: add X, <1, 1...> --> sub X, <-1, -1...> sub X, <1, 1...> --> add X, <-1, -1...> The all-ones vector constant can be materialized using a pcmpeq instruction that is commonly recognized as an idiom (has no register dependency), so that's better than loading a splat 1 constant. AVX512 uses 'vpternlogd' for 512-bit vectors because there is apparently no better way to produce 512 one-bits. The general advantages of this lowering are: 1. pcmpeq has lower latency than a memop on every uarch I looked at in Agner's tables, so in theory, this could be better for perf, but... 2. That seems unlikely to affect any OOO implementation, and I can't measure any real perf difference from this transform on Haswell or Jaguar, but... 3. It doesn't look like it from the diffs, but this is an overall size win because we eliminate 16 - 64 constant bytes in the case of a vector load. If we're broadcasting a scalar load (which might itself be a bug), then we're replacing a scalar constant load + broadcast with a single cheap op, so that should always be smaller/better too. 4. This makes the DAG/isel output more consistent - we use pcmpeq already for padd x, -1 and psub x, -1, so we should use that form for +1 too because we can. If there's some reason to favor a constant load on some CPU, let's make the reverse transform for all of these cases (either here in the DAG or in a later machine pass). This should fix: https://bugs.llvm.org/show_bug.cgi?id=33483 Differential Revision: https://reviews.llvm.org/D34336 llvm-svn: 306289
2017-06-26 22:19:26 +08:00
; SSE3-NEXT: movdqa %xmm0, %xmm2
; SSE3-NEXT: pand %xmm1, %xmm2
; SSE3-NEXT: psrlw $2, %xmm0
; SSE3-NEXT: pand %xmm1, %xmm0
[X86][SSE] LowerVectorCTPOP - pull out repeated byte sum stage. Pull out repeated byte sum stage for popcount of vector elements > 8bits. This allows us to simplify the LUT/BITMATH popcnt code to always assume vXi8 vectors, and also improves avx512bitalg codegen which only has access to vpopcntb/vpopcntw. llvm-svn: 344348
2018-10-12 22:18:47 +08:00
; SSE3-NEXT: paddb %xmm2, %xmm0
[x86] transform vector inc/dec to use -1 constant (PR33483) Convert vector increment or decrement to sub/add with an all-ones constant: add X, <1, 1...> --> sub X, <-1, -1...> sub X, <1, 1...> --> add X, <-1, -1...> The all-ones vector constant can be materialized using a pcmpeq instruction that is commonly recognized as an idiom (has no register dependency), so that's better than loading a splat 1 constant. AVX512 uses 'vpternlogd' for 512-bit vectors because there is apparently no better way to produce 512 one-bits. The general advantages of this lowering are: 1. pcmpeq has lower latency than a memop on every uarch I looked at in Agner's tables, so in theory, this could be better for perf, but... 2. That seems unlikely to affect any OOO implementation, and I can't measure any real perf difference from this transform on Haswell or Jaguar, but... 3. It doesn't look like it from the diffs, but this is an overall size win because we eliminate 16 - 64 constant bytes in the case of a vector load. If we're broadcasting a scalar load (which might itself be a bug), then we're replacing a scalar constant load + broadcast with a single cheap op, so that should always be smaller/better too. 4. This makes the DAG/isel output more consistent - we use pcmpeq already for padd x, -1 and psub x, -1, so we should use that form for +1 too because we can. If there's some reason to favor a constant load on some CPU, let's make the reverse transform for all of these cases (either here in the DAG or in a later machine pass). This should fix: https://bugs.llvm.org/show_bug.cgi?id=33483 Differential Revision: https://reviews.llvm.org/D34336 llvm-svn: 306289
2017-06-26 22:19:26 +08:00
; SSE3-NEXT: movdqa %xmm0, %xmm1
; SSE3-NEXT: psrlw $4, %xmm1
[X86][SSE] LowerVectorCTPOP - pull out repeated byte sum stage. Pull out repeated byte sum stage for popcount of vector elements > 8bits. This allows us to simplify the LUT/BITMATH popcnt code to always assume vXi8 vectors, and also improves avx512bitalg codegen which only has access to vpopcntb/vpopcntw. llvm-svn: 344348
2018-10-12 22:18:47 +08:00
; SSE3-NEXT: paddb %xmm0, %xmm1
[x86] transform vector inc/dec to use -1 constant (PR33483) Convert vector increment or decrement to sub/add with an all-ones constant: add X, <1, 1...> --> sub X, <-1, -1...> sub X, <1, 1...> --> add X, <-1, -1...> The all-ones vector constant can be materialized using a pcmpeq instruction that is commonly recognized as an idiom (has no register dependency), so that's better than loading a splat 1 constant. AVX512 uses 'vpternlogd' for 512-bit vectors because there is apparently no better way to produce 512 one-bits. The general advantages of this lowering are: 1. pcmpeq has lower latency than a memop on every uarch I looked at in Agner's tables, so in theory, this could be better for perf, but... 2. That seems unlikely to affect any OOO implementation, and I can't measure any real perf difference from this transform on Haswell or Jaguar, but... 3. It doesn't look like it from the diffs, but this is an overall size win because we eliminate 16 - 64 constant bytes in the case of a vector load. If we're broadcasting a scalar load (which might itself be a bug), then we're replacing a scalar constant load + broadcast with a single cheap op, so that should always be smaller/better too. 4. This makes the DAG/isel output more consistent - we use pcmpeq already for padd x, -1 and psub x, -1, so we should use that form for +1 too because we can. If there's some reason to favor a constant load on some CPU, let's make the reverse transform for all of these cases (either here in the DAG or in a later machine pass). This should fix: https://bugs.llvm.org/show_bug.cgi?id=33483 Differential Revision: https://reviews.llvm.org/D34336 llvm-svn: 306289
2017-06-26 22:19:26 +08:00
; SSE3-NEXT: pand {{.*}}(%rip), %xmm1
[X86][SSE] Vectorize CTTZ + CTTZ_ZERO_UNDEF Now that we have fast vector CTPOP implementations we can use this to speed up vector CTTZ using the pattern (cttz(x) = ctpop((x & -x) - 1)) Additionally, for AVX512CD that provides lzcnt instructions we can use the pattern (cttz_undef(x) = (width - 1) - ctlz(x & -x)) Differential Revision: http://reviews.llvm.org/D12663 llvm-svn: 248091
2015-09-19 21:22:57 +08:00
; SSE3-NEXT: movdqa %xmm1, %xmm0
; SSE3-NEXT: psllw $8, %xmm0
[x86] transform vector inc/dec to use -1 constant (PR33483) Convert vector increment or decrement to sub/add with an all-ones constant: add X, <1, 1...> --> sub X, <-1, -1...> sub X, <1, 1...> --> add X, <-1, -1...> The all-ones vector constant can be materialized using a pcmpeq instruction that is commonly recognized as an idiom (has no register dependency), so that's better than loading a splat 1 constant. AVX512 uses 'vpternlogd' for 512-bit vectors because there is apparently no better way to produce 512 one-bits. The general advantages of this lowering are: 1. pcmpeq has lower latency than a memop on every uarch I looked at in Agner's tables, so in theory, this could be better for perf, but... 2. That seems unlikely to affect any OOO implementation, and I can't measure any real perf difference from this transform on Haswell or Jaguar, but... 3. It doesn't look like it from the diffs, but this is an overall size win because we eliminate 16 - 64 constant bytes in the case of a vector load. If we're broadcasting a scalar load (which might itself be a bug), then we're replacing a scalar constant load + broadcast with a single cheap op, so that should always be smaller/better too. 4. This makes the DAG/isel output more consistent - we use pcmpeq already for padd x, -1 and psub x, -1, so we should use that form for +1 too because we can. If there's some reason to favor a constant load on some CPU, let's make the reverse transform for all of these cases (either here in the DAG or in a later machine pass). This should fix: https://bugs.llvm.org/show_bug.cgi?id=33483 Differential Revision: https://reviews.llvm.org/D34336 llvm-svn: 306289
2017-06-26 22:19:26 +08:00
; SSE3-NEXT: paddb %xmm1, %xmm0
[X86][SSE] Vectorize CTTZ + CTTZ_ZERO_UNDEF Now that we have fast vector CTPOP implementations we can use this to speed up vector CTTZ using the pattern (cttz(x) = ctpop((x & -x) - 1)) Additionally, for AVX512CD that provides lzcnt instructions we can use the pattern (cttz_undef(x) = (width - 1) - ctlz(x & -x)) Differential Revision: http://reviews.llvm.org/D12663 llvm-svn: 248091
2015-09-19 21:22:57 +08:00
; SSE3-NEXT: psrlw $8, %xmm0
[X86] Added tzcnt vector tests. llvm-svn: 239264
2015-06-08 05:01:34 +08:00
; SSE3-NEXT: retq
;
; SSSE3-LABEL: testv8i16u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; SSSE3: # %bb.0:
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; SSSE3-NEXT: pcmpeqd %xmm1, %xmm1
; SSSE3-NEXT: paddw %xmm0, %xmm1
; SSSE3-NEXT: pandn %xmm1, %xmm0
[x86] transform vector inc/dec to use -1 constant (PR33483) Convert vector increment or decrement to sub/add with an all-ones constant: add X, <1, 1...> --> sub X, <-1, -1...> sub X, <1, 1...> --> add X, <-1, -1...> The all-ones vector constant can be materialized using a pcmpeq instruction that is commonly recognized as an idiom (has no register dependency), so that's better than loading a splat 1 constant. AVX512 uses 'vpternlogd' for 512-bit vectors because there is apparently no better way to produce 512 one-bits. The general advantages of this lowering are: 1. pcmpeq has lower latency than a memop on every uarch I looked at in Agner's tables, so in theory, this could be better for perf, but... 2. That seems unlikely to affect any OOO implementation, and I can't measure any real perf difference from this transform on Haswell or Jaguar, but... 3. It doesn't look like it from the diffs, but this is an overall size win because we eliminate 16 - 64 constant bytes in the case of a vector load. If we're broadcasting a scalar load (which might itself be a bug), then we're replacing a scalar constant load + broadcast with a single cheap op, so that should always be smaller/better too. 4. This makes the DAG/isel output more consistent - we use pcmpeq already for padd x, -1 and psub x, -1, so we should use that form for +1 too because we can. If there's some reason to favor a constant load on some CPU, let's make the reverse transform for all of these cases (either here in the DAG or in a later machine pass). This should fix: https://bugs.llvm.org/show_bug.cgi?id=33483 Differential Revision: https://reviews.llvm.org/D34336 llvm-svn: 306289
2017-06-26 22:19:26 +08:00
; SSSE3-NEXT: movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
; SSSE3-NEXT: movdqa %xmm0, %xmm2
; SSSE3-NEXT: pand %xmm1, %xmm2
[X86][SSE] Vectorize CTTZ + CTTZ_ZERO_UNDEF Now that we have fast vector CTPOP implementations we can use this to speed up vector CTTZ using the pattern (cttz(x) = ctpop((x & -x) - 1)) Additionally, for AVX512CD that provides lzcnt instructions we can use the pattern (cttz_undef(x) = (width - 1) - ctlz(x & -x)) Differential Revision: http://reviews.llvm.org/D12663 llvm-svn: 248091
2015-09-19 21:22:57 +08:00
; SSSE3-NEXT: movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
; SSSE3-NEXT: movdqa %xmm3, %xmm4
; SSSE3-NEXT: pshufb %xmm2, %xmm4
[x86] transform vector inc/dec to use -1 constant (PR33483) Convert vector increment or decrement to sub/add with an all-ones constant: add X, <1, 1...> --> sub X, <-1, -1...> sub X, <1, 1...> --> add X, <-1, -1...> The all-ones vector constant can be materialized using a pcmpeq instruction that is commonly recognized as an idiom (has no register dependency), so that's better than loading a splat 1 constant. AVX512 uses 'vpternlogd' for 512-bit vectors because there is apparently no better way to produce 512 one-bits. The general advantages of this lowering are: 1. pcmpeq has lower latency than a memop on every uarch I looked at in Agner's tables, so in theory, this could be better for perf, but... 2. That seems unlikely to affect any OOO implementation, and I can't measure any real perf difference from this transform on Haswell or Jaguar, but... 3. It doesn't look like it from the diffs, but this is an overall size win because we eliminate 16 - 64 constant bytes in the case of a vector load. If we're broadcasting a scalar load (which might itself be a bug), then we're replacing a scalar constant load + broadcast with a single cheap op, so that should always be smaller/better too. 4. This makes the DAG/isel output more consistent - we use pcmpeq already for padd x, -1 and psub x, -1, so we should use that form for +1 too because we can. If there's some reason to favor a constant load on some CPU, let's make the reverse transform for all of these cases (either here in the DAG or in a later machine pass). This should fix: https://bugs.llvm.org/show_bug.cgi?id=33483 Differential Revision: https://reviews.llvm.org/D34336 llvm-svn: 306289
2017-06-26 22:19:26 +08:00
; SSSE3-NEXT: psrlw $4, %xmm0
; SSSE3-NEXT: pand %xmm1, %xmm0
; SSSE3-NEXT: pshufb %xmm0, %xmm3
[X86][SSE] Vectorize CTTZ + CTTZ_ZERO_UNDEF Now that we have fast vector CTPOP implementations we can use this to speed up vector CTTZ using the pattern (cttz(x) = ctpop((x & -x) - 1)) Additionally, for AVX512CD that provides lzcnt instructions we can use the pattern (cttz_undef(x) = (width - 1) - ctlz(x & -x)) Differential Revision: http://reviews.llvm.org/D12663 llvm-svn: 248091
2015-09-19 21:22:57 +08:00
; SSSE3-NEXT: paddb %xmm4, %xmm3
; SSSE3-NEXT: movdqa %xmm3, %xmm0
; SSSE3-NEXT: psllw $8, %xmm0
; SSSE3-NEXT: paddb %xmm3, %xmm0
; SSSE3-NEXT: psrlw $8, %xmm0
[X86] Added tzcnt vector tests. llvm-svn: 239264
2015-06-08 05:01:34 +08:00
; SSSE3-NEXT: retq
;
; SSE41-LABEL: testv8i16u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; SSE41: # %bb.0:
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; SSE41-NEXT: pcmpeqd %xmm1, %xmm1
; SSE41-NEXT: paddw %xmm0, %xmm1
; SSE41-NEXT: pandn %xmm1, %xmm0
[x86] transform vector inc/dec to use -1 constant (PR33483) Convert vector increment or decrement to sub/add with an all-ones constant: add X, <1, 1...> --> sub X, <-1, -1...> sub X, <1, 1...> --> add X, <-1, -1...> The all-ones vector constant can be materialized using a pcmpeq instruction that is commonly recognized as an idiom (has no register dependency), so that's better than loading a splat 1 constant. AVX512 uses 'vpternlogd' for 512-bit vectors because there is apparently no better way to produce 512 one-bits. The general advantages of this lowering are: 1. pcmpeq has lower latency than a memop on every uarch I looked at in Agner's tables, so in theory, this could be better for perf, but... 2. That seems unlikely to affect any OOO implementation, and I can't measure any real perf difference from this transform on Haswell or Jaguar, but... 3. It doesn't look like it from the diffs, but this is an overall size win because we eliminate 16 - 64 constant bytes in the case of a vector load. If we're broadcasting a scalar load (which might itself be a bug), then we're replacing a scalar constant load + broadcast with a single cheap op, so that should always be smaller/better too. 4. This makes the DAG/isel output more consistent - we use pcmpeq already for padd x, -1 and psub x, -1, so we should use that form for +1 too because we can. If there's some reason to favor a constant load on some CPU, let's make the reverse transform for all of these cases (either here in the DAG or in a later machine pass). This should fix: https://bugs.llvm.org/show_bug.cgi?id=33483 Differential Revision: https://reviews.llvm.org/D34336 llvm-svn: 306289
2017-06-26 22:19:26 +08:00
; SSE41-NEXT: movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
; SSE41-NEXT: movdqa %xmm0, %xmm2
; SSE41-NEXT: pand %xmm1, %xmm2
[X86][SSE] Vectorize CTTZ + CTTZ_ZERO_UNDEF Now that we have fast vector CTPOP implementations we can use this to speed up vector CTTZ using the pattern (cttz(x) = ctpop((x & -x) - 1)) Additionally, for AVX512CD that provides lzcnt instructions we can use the pattern (cttz_undef(x) = (width - 1) - ctlz(x & -x)) Differential Revision: http://reviews.llvm.org/D12663 llvm-svn: 248091
2015-09-19 21:22:57 +08:00
; SSE41-NEXT: movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
; SSE41-NEXT: movdqa %xmm3, %xmm4
; SSE41-NEXT: pshufb %xmm2, %xmm4
[x86] transform vector inc/dec to use -1 constant (PR33483) Convert vector increment or decrement to sub/add with an all-ones constant: add X, <1, 1...> --> sub X, <-1, -1...> sub X, <1, 1...> --> add X, <-1, -1...> The all-ones vector constant can be materialized using a pcmpeq instruction that is commonly recognized as an idiom (has no register dependency), so that's better than loading a splat 1 constant. AVX512 uses 'vpternlogd' for 512-bit vectors because there is apparently no better way to produce 512 one-bits. The general advantages of this lowering are: 1. pcmpeq has lower latency than a memop on every uarch I looked at in Agner's tables, so in theory, this could be better for perf, but... 2. That seems unlikely to affect any OOO implementation, and I can't measure any real perf difference from this transform on Haswell or Jaguar, but... 3. It doesn't look like it from the diffs, but this is an overall size win because we eliminate 16 - 64 constant bytes in the case of a vector load. If we're broadcasting a scalar load (which might itself be a bug), then we're replacing a scalar constant load + broadcast with a single cheap op, so that should always be smaller/better too. 4. This makes the DAG/isel output more consistent - we use pcmpeq already for padd x, -1 and psub x, -1, so we should use that form for +1 too because we can. If there's some reason to favor a constant load on some CPU, let's make the reverse transform for all of these cases (either here in the DAG or in a later machine pass). This should fix: https://bugs.llvm.org/show_bug.cgi?id=33483 Differential Revision: https://reviews.llvm.org/D34336 llvm-svn: 306289
2017-06-26 22:19:26 +08:00
; SSE41-NEXT: psrlw $4, %xmm0
; SSE41-NEXT: pand %xmm1, %xmm0
; SSE41-NEXT: pshufb %xmm0, %xmm3
[X86][SSE] Vectorize CTTZ + CTTZ_ZERO_UNDEF Now that we have fast vector CTPOP implementations we can use this to speed up vector CTTZ using the pattern (cttz(x) = ctpop((x & -x) - 1)) Additionally, for AVX512CD that provides lzcnt instructions we can use the pattern (cttz_undef(x) = (width - 1) - ctlz(x & -x)) Differential Revision: http://reviews.llvm.org/D12663 llvm-svn: 248091
2015-09-19 21:22:57 +08:00
; SSE41-NEXT: paddb %xmm4, %xmm3
; SSE41-NEXT: movdqa %xmm3, %xmm0
; SSE41-NEXT: psllw $8, %xmm0
; SSE41-NEXT: paddb %xmm3, %xmm0
; SSE41-NEXT: psrlw $8, %xmm0
[X86][SSE] Tidied up vector CTLZ/CTTZ. NFCI. llvm-svn: 242645
2015-07-20 01:09:43 +08:00
; SSE41-NEXT: retq
[X86] Added tzcnt vector tests. llvm-svn: 239264
2015-06-08 05:01:34 +08:00
;
[X86][AVX] Regenerate tzcnt tests. The EVEX -> VEX fix means that AVX/AVX512 code is more likely the same now. llvm-svn: 291241
2017-01-06 23:54:23 +08:00
; AVX-LABEL: testv8i16u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; AVX: # %bb.0:
[x86] transform vector inc/dec to use -1 constant (PR33483) Convert vector increment or decrement to sub/add with an all-ones constant: add X, <1, 1...> --> sub X, <-1, -1...> sub X, <1, 1...> --> add X, <-1, -1...> The all-ones vector constant can be materialized using a pcmpeq instruction that is commonly recognized as an idiom (has no register dependency), so that's better than loading a splat 1 constant. AVX512 uses 'vpternlogd' for 512-bit vectors because there is apparently no better way to produce 512 one-bits. The general advantages of this lowering are: 1. pcmpeq has lower latency than a memop on every uarch I looked at in Agner's tables, so in theory, this could be better for perf, but... 2. That seems unlikely to affect any OOO implementation, and I can't measure any real perf difference from this transform on Haswell or Jaguar, but... 3. It doesn't look like it from the diffs, but this is an overall size win because we eliminate 16 - 64 constant bytes in the case of a vector load. If we're broadcasting a scalar load (which might itself be a bug), then we're replacing a scalar constant load + broadcast with a single cheap op, so that should always be smaller/better too. 4. This makes the DAG/isel output more consistent - we use pcmpeq already for padd x, -1 and psub x, -1, so we should use that form for +1 too because we can. If there's some reason to favor a constant load on some CPU, let's make the reverse transform for all of these cases (either here in the DAG or in a later machine pass). This should fix: https://bugs.llvm.org/show_bug.cgi?id=33483 Differential Revision: https://reviews.llvm.org/D34336 llvm-svn: 306289
2017-06-26 22:19:26 +08:00
; AVX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; AVX-NEXT: vpaddw %xmm1, %xmm0, %xmm1
; AVX-NEXT: vpandn %xmm1, %xmm0, %xmm0
[X86][AVX] Regenerate tzcnt tests. The EVEX -> VEX fix means that AVX/AVX512 code is more likely the same now. llvm-svn: 291241
2017-01-06 23:54:23 +08:00
; AVX-NEXT: vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
; AVX-NEXT: vpand %xmm1, %xmm0, %xmm2
; AVX-NEXT: vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
; AVX-NEXT: vpshufb %xmm2, %xmm3, %xmm2
; AVX-NEXT: vpsrlw $4, %xmm0, %xmm0
; AVX-NEXT: vpand %xmm1, %xmm0, %xmm0
; AVX-NEXT: vpshufb %xmm0, %xmm3, %xmm0
; AVX-NEXT: vpaddb %xmm2, %xmm0, %xmm0
; AVX-NEXT: vpsllw $8, %xmm0, %xmm1
; AVX-NEXT: vpaddb %xmm0, %xmm1, %xmm0
; AVX-NEXT: vpsrlw $8, %xmm0, %xmm0
; AVX-NEXT: retq
[AVX512] Teach lowering to use vplzcntd/q to implement 128/256-bit CTTZ_ZERO_UNDEF even without VLX support. We can just extend to 512-bits and extract like we do for CTLZ. llvm-svn: 267100
2016-04-22 11:22:38 +08:00
;
[X86] Adding vpopcntd and vpopcntq instructions AVX512_VPOPCNTDQ is a new feature set that was published by Intel. The patch represents the LLVM side of the addition of two new intrinsic based instructions (vpopcntd and vpopcntq). Differential Revision: https://reviews.llvm.org/D33169 llvm-svn: 303858
2017-05-25 21:45:23 +08:00
; AVX512VPOPCNTDQ-LABEL: testv8i16u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; AVX512VPOPCNTDQ: # %bb.0:
[x86] transform vector inc/dec to use -1 constant (PR33483) Convert vector increment or decrement to sub/add with an all-ones constant: add X, <1, 1...> --> sub X, <-1, -1...> sub X, <1, 1...> --> add X, <-1, -1...> The all-ones vector constant can be materialized using a pcmpeq instruction that is commonly recognized as an idiom (has no register dependency), so that's better than loading a splat 1 constant. AVX512 uses 'vpternlogd' for 512-bit vectors because there is apparently no better way to produce 512 one-bits. The general advantages of this lowering are: 1. pcmpeq has lower latency than a memop on every uarch I looked at in Agner's tables, so in theory, this could be better for perf, but... 2. That seems unlikely to affect any OOO implementation, and I can't measure any real perf difference from this transform on Haswell or Jaguar, but... 3. It doesn't look like it from the diffs, but this is an overall size win because we eliminate 16 - 64 constant bytes in the case of a vector load. If we're broadcasting a scalar load (which might itself be a bug), then we're replacing a scalar constant load + broadcast with a single cheap op, so that should always be smaller/better too. 4. This makes the DAG/isel output more consistent - we use pcmpeq already for padd x, -1 and psub x, -1, so we should use that form for +1 too because we can. If there's some reason to favor a constant load on some CPU, let's make the reverse transform for all of these cases (either here in the DAG or in a later machine pass). This should fix: https://bugs.llvm.org/show_bug.cgi?id=33483 Differential Revision: https://reviews.llvm.org/D34336 llvm-svn: 306289
2017-06-26 22:19:26 +08:00
; AVX512VPOPCNTDQ-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; AVX512VPOPCNTDQ-NEXT: vpaddw %xmm1, %xmm0, %xmm1
; AVX512VPOPCNTDQ-NEXT: vpandn %xmm1, %xmm0, %xmm0
[X86] When using vpopcntdq for ctpop of v8i16 vectors, only promote to v8i32. Previously we promoted to v8i64, but we don't need to go all the way to 512-bits. If we have VLX we can use the 256-bit instruction. And even if we don't have VLX we can widen v8i32 to v16i32 and drop the upper half. llvm-svn: 320926
2017-12-17 03:31:36 +08:00
; AVX512VPOPCNTDQ-NEXT: vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
; AVX512VPOPCNTDQ-NEXT: vpopcntd %zmm0, %zmm0
; AVX512VPOPCNTDQ-NEXT: vpmovdw %zmm0, %ymm0
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
; AVX512VPOPCNTDQ-NEXT: # kill: def $xmm0 killed $xmm0 killed $ymm0
[X86][AVX512VPOPCNTDQ] Improve support for v16i8/v8i16/v16i16/ CTPOP Zero extend to v16i32/v8i64, use VPOPCNTDQ instructions and truncate back. llvm-svn: 306990
2017-07-03 03:32:37 +08:00
; AVX512VPOPCNTDQ-NEXT: vzeroupper
[X86] Adding vpopcntd and vpopcntq instructions AVX512_VPOPCNTDQ is a new feature set that was published by Intel. The patch represents the LLVM side of the addition of two new intrinsic based instructions (vpopcntd and vpopcntq). Differential Revision: https://reviews.llvm.org/D33169 llvm-svn: 303858
2017-05-25 21:45:23 +08:00
; AVX512VPOPCNTDQ-NEXT: retq
;
[X86] Add 128 and 256-bit VPOPCNTDQ instructions. Adjust some tablegen classes LZCNT/POPCNT. I think when this instruction was first published it was only for a Knights CPU and thus VLX version was missing. llvm-svn: 320910
2017-12-16 10:40:28 +08:00
; AVX512VPOPCNTDQVL-LABEL: testv8i16u:
; AVX512VPOPCNTDQVL: # %bb.0:
; AVX512VPOPCNTDQVL-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; AVX512VPOPCNTDQVL-NEXT: vpaddw %xmm1, %xmm0, %xmm1
; AVX512VPOPCNTDQVL-NEXT: vpandn %xmm1, %xmm0, %xmm0
[X86] When using vpopcntdq for ctpop of v8i16 vectors, only promote to v8i32. Previously we promoted to v8i64, but we don't need to go all the way to 512-bits. If we have VLX we can use the 256-bit instruction. And even if we don't have VLX we can widen v8i32 to v16i32 and drop the upper half. llvm-svn: 320926
2017-12-17 03:31:36 +08:00
; AVX512VPOPCNTDQVL-NEXT: vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
; AVX512VPOPCNTDQVL-NEXT: vpopcntd %ymm0, %ymm0
; AVX512VPOPCNTDQVL-NEXT: vpmovdw %ymm0, %xmm0
[X86] Add 128 and 256-bit VPOPCNTDQ instructions. Adjust some tablegen classes LZCNT/POPCNT. I think when this instruction was first published it was only for a Knights CPU and thus VLX version was missing. llvm-svn: 320910
2017-12-16 10:40:28 +08:00
; AVX512VPOPCNTDQVL-NEXT: vzeroupper
; AVX512VPOPCNTDQVL-NEXT: retq
;
[X86] Regenerate the vector-popcnt and vector-tzcnt tests to get BITALG CHECK linse on all functions not just the vXi16/vXi8. llvm-svn: 318885
2017-11-23 07:35:12 +08:00
; BITALG_NOVLX-LABEL: testv8i16u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; BITALG_NOVLX: # %bb.0:
[X86] Regenerate the vector-popcnt and vector-tzcnt tests to get BITALG CHECK linse on all functions not just the vXi16/vXi8. llvm-svn: 318885
2017-11-23 07:35:12 +08:00
; BITALG_NOVLX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; BITALG_NOVLX-NEXT: vpaddw %xmm1, %xmm0, %xmm1
; BITALG_NOVLX-NEXT: vpandn %xmm1, %xmm0, %xmm0
[X86] Regenerate the vector-popcnt and vector-tzcnt tests to get BITALG CHECK linse on all functions not just the vXi16/vXi8. llvm-svn: 318885
2017-11-23 07:35:12 +08:00
; BITALG_NOVLX-NEXT: vpopcntw %zmm0, %zmm0
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
; BITALG_NOVLX-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
[X86] Regenerate the vector-popcnt and vector-tzcnt tests to get BITALG CHECK linse on all functions not just the vXi16/vXi8. llvm-svn: 318885
2017-11-23 07:35:12 +08:00
; BITALG_NOVLX-NEXT: vzeroupper
; BITALG_NOVLX-NEXT: retq
;
; BITALG-LABEL: testv8i16u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; BITALG: # %bb.0:
[X86] Regenerate the vector-popcnt and vector-tzcnt tests to get BITALG CHECK linse on all functions not just the vXi16/vXi8. llvm-svn: 318885
2017-11-23 07:35:12 +08:00
; BITALG-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; BITALG-NEXT: vpaddw %xmm1, %xmm0, %xmm1
; BITALG-NEXT: vpandn %xmm1, %xmm0, %xmm0
[X86] Regenerate the vector-popcnt and vector-tzcnt tests to get BITALG CHECK linse on all functions not just the vXi16/vXi8. llvm-svn: 318885
2017-11-23 07:35:12 +08:00
; BITALG-NEXT: vpopcntw %xmm0, %xmm0
; BITALG-NEXT: retq
;
[X86][SSE] Added 32-bit tests for vector lzcnt/tzcnt tests v2i64 tests are particularly bad on 32-bit targets. llvm-svn: 265789
2016-04-08 23:01:31 +08:00
; X32-SSE-LABEL: testv8i16u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; X32-SSE: # %bb.0:
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; X32-SSE-NEXT: pcmpeqd %xmm1, %xmm1
; X32-SSE-NEXT: paddw %xmm0, %xmm1
; X32-SSE-NEXT: pandn %xmm1, %xmm0
[x86] transform vector inc/dec to use -1 constant (PR33483) Convert vector increment or decrement to sub/add with an all-ones constant: add X, <1, 1...> --> sub X, <-1, -1...> sub X, <1, 1...> --> add X, <-1, -1...> The all-ones vector constant can be materialized using a pcmpeq instruction that is commonly recognized as an idiom (has no register dependency), so that's better than loading a splat 1 constant. AVX512 uses 'vpternlogd' for 512-bit vectors because there is apparently no better way to produce 512 one-bits. The general advantages of this lowering are: 1. pcmpeq has lower latency than a memop on every uarch I looked at in Agner's tables, so in theory, this could be better for perf, but... 2. That seems unlikely to affect any OOO implementation, and I can't measure any real perf difference from this transform on Haswell or Jaguar, but... 3. It doesn't look like it from the diffs, but this is an overall size win because we eliminate 16 - 64 constant bytes in the case of a vector load. If we're broadcasting a scalar load (which might itself be a bug), then we're replacing a scalar constant load + broadcast with a single cheap op, so that should always be smaller/better too. 4. This makes the DAG/isel output more consistent - we use pcmpeq already for padd x, -1 and psub x, -1, so we should use that form for +1 too because we can. If there's some reason to favor a constant load on some CPU, let's make the reverse transform for all of these cases (either here in the DAG or in a later machine pass). This should fix: https://bugs.llvm.org/show_bug.cgi?id=33483 Differential Revision: https://reviews.llvm.org/D34336 llvm-svn: 306289
2017-06-26 22:19:26 +08:00
; X32-SSE-NEXT: movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
; X32-SSE-NEXT: movdqa %xmm0, %xmm2
; X32-SSE-NEXT: pand %xmm1, %xmm2
[X86][SSE] Added 32-bit tests for vector lzcnt/tzcnt tests v2i64 tests are particularly bad on 32-bit targets. llvm-svn: 265789
2016-04-08 23:01:31 +08:00
; X32-SSE-NEXT: movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
; X32-SSE-NEXT: movdqa %xmm3, %xmm4
; X32-SSE-NEXT: pshufb %xmm2, %xmm4
[x86] transform vector inc/dec to use -1 constant (PR33483) Convert vector increment or decrement to sub/add with an all-ones constant: add X, <1, 1...> --> sub X, <-1, -1...> sub X, <1, 1...> --> add X, <-1, -1...> The all-ones vector constant can be materialized using a pcmpeq instruction that is commonly recognized as an idiom (has no register dependency), so that's better than loading a splat 1 constant. AVX512 uses 'vpternlogd' for 512-bit vectors because there is apparently no better way to produce 512 one-bits. The general advantages of this lowering are: 1. pcmpeq has lower latency than a memop on every uarch I looked at in Agner's tables, so in theory, this could be better for perf, but... 2. That seems unlikely to affect any OOO implementation, and I can't measure any real perf difference from this transform on Haswell or Jaguar, but... 3. It doesn't look like it from the diffs, but this is an overall size win because we eliminate 16 - 64 constant bytes in the case of a vector load. If we're broadcasting a scalar load (which might itself be a bug), then we're replacing a scalar constant load + broadcast with a single cheap op, so that should always be smaller/better too. 4. This makes the DAG/isel output more consistent - we use pcmpeq already for padd x, -1 and psub x, -1, so we should use that form for +1 too because we can. If there's some reason to favor a constant load on some CPU, let's make the reverse transform for all of these cases (either here in the DAG or in a later machine pass). This should fix: https://bugs.llvm.org/show_bug.cgi?id=33483 Differential Revision: https://reviews.llvm.org/D34336 llvm-svn: 306289
2017-06-26 22:19:26 +08:00
; X32-SSE-NEXT: psrlw $4, %xmm0
; X32-SSE-NEXT: pand %xmm1, %xmm0
; X32-SSE-NEXT: pshufb %xmm0, %xmm3
[X86][SSE] Added 32-bit tests for vector lzcnt/tzcnt tests v2i64 tests are particularly bad on 32-bit targets. llvm-svn: 265789
2016-04-08 23:01:31 +08:00
; X32-SSE-NEXT: paddb %xmm4, %xmm3
; X32-SSE-NEXT: movdqa %xmm3, %xmm0
; X32-SSE-NEXT: psllw $8, %xmm0
; X32-SSE-NEXT: paddb %xmm3, %xmm0
; X32-SSE-NEXT: psrlw $8, %xmm0
; X32-SSE-NEXT: retl
[X86] Added tzcnt vector tests. llvm-svn: 239264
2015-06-08 05:01:34 +08:00
%out = call <8 x i16> @llvm.cttz.v8i16(<8 x i16> %in, i1 -1)
ret <8 x i16> %out
}
[X86][SSE] Tidied up vector CTLZ/CTTZ. NFCI. llvm-svn: 242645
2015-07-20 01:09:43 +08:00
define <16 x i8> @testv16i8(<16 x i8> %in) nounwind {
[X86] Added tzcnt vector tests. llvm-svn: 239264
2015-06-08 05:01:34 +08:00
; SSE2-LABEL: testv16i8:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; SSE2: # %bb.0:
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; SSE2-NEXT: pcmpeqd %xmm1, %xmm1
; SSE2-NEXT: paddb %xmm0, %xmm1
; SSE2-NEXT: pandn %xmm1, %xmm0
; SSE2-NEXT: movdqa %xmm0, %xmm1
; SSE2-NEXT: psrlw $1, %xmm1
; SSE2-NEXT: pand {{.*}}(%rip), %xmm1
; SSE2-NEXT: psubb %xmm1, %xmm0
; SSE2-NEXT: movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
; SSE2-NEXT: movdqa %xmm0, %xmm2
; SSE2-NEXT: pand %xmm1, %xmm2
; SSE2-NEXT: psrlw $2, %xmm0
; SSE2-NEXT: pand %xmm1, %xmm0
[x86] transform vector inc/dec to use -1 constant (PR33483) Convert vector increment or decrement to sub/add with an all-ones constant: add X, <1, 1...> --> sub X, <-1, -1...> sub X, <1, 1...> --> add X, <-1, -1...> The all-ones vector constant can be materialized using a pcmpeq instruction that is commonly recognized as an idiom (has no register dependency), so that's better than loading a splat 1 constant. AVX512 uses 'vpternlogd' for 512-bit vectors because there is apparently no better way to produce 512 one-bits. The general advantages of this lowering are: 1. pcmpeq has lower latency than a memop on every uarch I looked at in Agner's tables, so in theory, this could be better for perf, but... 2. That seems unlikely to affect any OOO implementation, and I can't measure any real perf difference from this transform on Haswell or Jaguar, but... 3. It doesn't look like it from the diffs, but this is an overall size win because we eliminate 16 - 64 constant bytes in the case of a vector load. If we're broadcasting a scalar load (which might itself be a bug), then we're replacing a scalar constant load + broadcast with a single cheap op, so that should always be smaller/better too. 4. This makes the DAG/isel output more consistent - we use pcmpeq already for padd x, -1 and psub x, -1, so we should use that form for +1 too because we can. If there's some reason to favor a constant load on some CPU, let's make the reverse transform for all of these cases (either here in the DAG or in a later machine pass). This should fix: https://bugs.llvm.org/show_bug.cgi?id=33483 Differential Revision: https://reviews.llvm.org/D34336 llvm-svn: 306289
2017-06-26 22:19:26 +08:00
; SSE2-NEXT: paddb %xmm2, %xmm0
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; SSE2-NEXT: movdqa %xmm0, %xmm1
; SSE2-NEXT: psrlw $4, %xmm1
; SSE2-NEXT: paddb %xmm0, %xmm1
; SSE2-NEXT: pand {{.*}}(%rip), %xmm1
; SSE2-NEXT: movdqa %xmm1, %xmm0
[X86] Added tzcnt vector tests. llvm-svn: 239264
2015-06-08 05:01:34 +08:00
; SSE2-NEXT: retq
;
; SSE3-LABEL: testv16i8:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; SSE3: # %bb.0:
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; SSE3-NEXT: pcmpeqd %xmm1, %xmm1
; SSE3-NEXT: paddb %xmm0, %xmm1
; SSE3-NEXT: pandn %xmm1, %xmm0
; SSE3-NEXT: movdqa %xmm0, %xmm1
; SSE3-NEXT: psrlw $1, %xmm1
; SSE3-NEXT: pand {{.*}}(%rip), %xmm1
; SSE3-NEXT: psubb %xmm1, %xmm0
; SSE3-NEXT: movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
; SSE3-NEXT: movdqa %xmm0, %xmm2
; SSE3-NEXT: pand %xmm1, %xmm2
; SSE3-NEXT: psrlw $2, %xmm0
; SSE3-NEXT: pand %xmm1, %xmm0
[x86] transform vector inc/dec to use -1 constant (PR33483) Convert vector increment or decrement to sub/add with an all-ones constant: add X, <1, 1...> --> sub X, <-1, -1...> sub X, <1, 1...> --> add X, <-1, -1...> The all-ones vector constant can be materialized using a pcmpeq instruction that is commonly recognized as an idiom (has no register dependency), so that's better than loading a splat 1 constant. AVX512 uses 'vpternlogd' for 512-bit vectors because there is apparently no better way to produce 512 one-bits. The general advantages of this lowering are: 1. pcmpeq has lower latency than a memop on every uarch I looked at in Agner's tables, so in theory, this could be better for perf, but... 2. That seems unlikely to affect any OOO implementation, and I can't measure any real perf difference from this transform on Haswell or Jaguar, but... 3. It doesn't look like it from the diffs, but this is an overall size win because we eliminate 16 - 64 constant bytes in the case of a vector load. If we're broadcasting a scalar load (which might itself be a bug), then we're replacing a scalar constant load + broadcast with a single cheap op, so that should always be smaller/better too. 4. This makes the DAG/isel output more consistent - we use pcmpeq already for padd x, -1 and psub x, -1, so we should use that form for +1 too because we can. If there's some reason to favor a constant load on some CPU, let's make the reverse transform for all of these cases (either here in the DAG or in a later machine pass). This should fix: https://bugs.llvm.org/show_bug.cgi?id=33483 Differential Revision: https://reviews.llvm.org/D34336 llvm-svn: 306289
2017-06-26 22:19:26 +08:00
; SSE3-NEXT: paddb %xmm2, %xmm0
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; SSE3-NEXT: movdqa %xmm0, %xmm1
; SSE3-NEXT: psrlw $4, %xmm1
; SSE3-NEXT: paddb %xmm0, %xmm1
; SSE3-NEXT: pand {{.*}}(%rip), %xmm1
; SSE3-NEXT: movdqa %xmm1, %xmm0
[X86] Added tzcnt vector tests. llvm-svn: 239264
2015-06-08 05:01:34 +08:00
; SSE3-NEXT: retq
;
; SSSE3-LABEL: testv16i8:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; SSSE3: # %bb.0:
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; SSSE3-NEXT: pcmpeqd %xmm1, %xmm1
; SSSE3-NEXT: paddb %xmm0, %xmm1
; SSSE3-NEXT: pandn %xmm1, %xmm0
; SSSE3-NEXT: movdqa {{.*#+}} xmm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
; SSSE3-NEXT: movdqa %xmm0, %xmm3
; SSSE3-NEXT: pand %xmm2, %xmm3
; SSSE3-NEXT: movdqa {{.*#+}} xmm1 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
; SSSE3-NEXT: movdqa %xmm1, %xmm4
[X86][SSE] Vectorize CTTZ + CTTZ_ZERO_UNDEF Now that we have fast vector CTPOP implementations we can use this to speed up vector CTTZ using the pattern (cttz(x) = ctpop((x & -x) - 1)) Additionally, for AVX512CD that provides lzcnt instructions we can use the pattern (cttz_undef(x) = (width - 1) - ctlz(x & -x)) Differential Revision: http://reviews.llvm.org/D12663 llvm-svn: 248091
2015-09-19 21:22:57 +08:00
; SSSE3-NEXT: pshufb %xmm3, %xmm4
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; SSSE3-NEXT: psrlw $4, %xmm0
; SSSE3-NEXT: pand %xmm2, %xmm0
; SSSE3-NEXT: pshufb %xmm0, %xmm1
; SSSE3-NEXT: paddb %xmm4, %xmm1
; SSSE3-NEXT: movdqa %xmm1, %xmm0
[X86] Added tzcnt vector tests. llvm-svn: 239264
2015-06-08 05:01:34 +08:00
; SSSE3-NEXT: retq
;
; SSE41-LABEL: testv16i8:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; SSE41: # %bb.0:
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; SSE41-NEXT: pcmpeqd %xmm1, %xmm1
; SSE41-NEXT: paddb %xmm0, %xmm1
; SSE41-NEXT: pandn %xmm1, %xmm0
; SSE41-NEXT: movdqa {{.*#+}} xmm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
; SSE41-NEXT: movdqa %xmm0, %xmm3
; SSE41-NEXT: pand %xmm2, %xmm3
; SSE41-NEXT: movdqa {{.*#+}} xmm1 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
; SSE41-NEXT: movdqa %xmm1, %xmm4
[X86][SSE] Vectorize CTTZ + CTTZ_ZERO_UNDEF Now that we have fast vector CTPOP implementations we can use this to speed up vector CTTZ using the pattern (cttz(x) = ctpop((x & -x) - 1)) Additionally, for AVX512CD that provides lzcnt instructions we can use the pattern (cttz_undef(x) = (width - 1) - ctlz(x & -x)) Differential Revision: http://reviews.llvm.org/D12663 llvm-svn: 248091
2015-09-19 21:22:57 +08:00
; SSE41-NEXT: pshufb %xmm3, %xmm4
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; SSE41-NEXT: psrlw $4, %xmm0
; SSE41-NEXT: pand %xmm2, %xmm0
; SSE41-NEXT: pshufb %xmm0, %xmm1
; SSE41-NEXT: paddb %xmm4, %xmm1
; SSE41-NEXT: movdqa %xmm1, %xmm0
[X86] Added tzcnt vector tests. llvm-svn: 239264
2015-06-08 05:01:34 +08:00
; SSE41-NEXT: retq
;
[X86][AVX] Regenerate tzcnt tests. The EVEX -> VEX fix means that AVX/AVX512 code is more likely the same now. llvm-svn: 291241
2017-01-06 23:54:23 +08:00
; AVX-LABEL: testv16i8:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; AVX: # %bb.0:
[x86] transform vector inc/dec to use -1 constant (PR33483) Convert vector increment or decrement to sub/add with an all-ones constant: add X, <1, 1...> --> sub X, <-1, -1...> sub X, <1, 1...> --> add X, <-1, -1...> The all-ones vector constant can be materialized using a pcmpeq instruction that is commonly recognized as an idiom (has no register dependency), so that's better than loading a splat 1 constant. AVX512 uses 'vpternlogd' for 512-bit vectors because there is apparently no better way to produce 512 one-bits. The general advantages of this lowering are: 1. pcmpeq has lower latency than a memop on every uarch I looked at in Agner's tables, so in theory, this could be better for perf, but... 2. That seems unlikely to affect any OOO implementation, and I can't measure any real perf difference from this transform on Haswell or Jaguar, but... 3. It doesn't look like it from the diffs, but this is an overall size win because we eliminate 16 - 64 constant bytes in the case of a vector load. If we're broadcasting a scalar load (which might itself be a bug), then we're replacing a scalar constant load + broadcast with a single cheap op, so that should always be smaller/better too. 4. This makes the DAG/isel output more consistent - we use pcmpeq already for padd x, -1 and psub x, -1, so we should use that form for +1 too because we can. If there's some reason to favor a constant load on some CPU, let's make the reverse transform for all of these cases (either here in the DAG or in a later machine pass). This should fix: https://bugs.llvm.org/show_bug.cgi?id=33483 Differential Revision: https://reviews.llvm.org/D34336 llvm-svn: 306289
2017-06-26 22:19:26 +08:00
; AVX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; AVX-NEXT: vpaddb %xmm1, %xmm0, %xmm1
; AVX-NEXT: vpandn %xmm1, %xmm0, %xmm0
[X86][AVX] Regenerate tzcnt tests. The EVEX -> VEX fix means that AVX/AVX512 code is more likely the same now. llvm-svn: 291241
2017-01-06 23:54:23 +08:00
; AVX-NEXT: vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
; AVX-NEXT: vpand %xmm1, %xmm0, %xmm2
; AVX-NEXT: vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
; AVX-NEXT: vpshufb %xmm2, %xmm3, %xmm2
; AVX-NEXT: vpsrlw $4, %xmm0, %xmm0
; AVX-NEXT: vpand %xmm1, %xmm0, %xmm0
; AVX-NEXT: vpshufb %xmm0, %xmm3, %xmm0
; AVX-NEXT: vpaddb %xmm2, %xmm0, %xmm0
; AVX-NEXT: retq
[AVX512] Teach lowering to use vplzcntd/q to implement 128/256-bit CTTZ_ZERO_UNDEF even without VLX support. We can just extend to 512-bits and extract like we do for CTLZ. llvm-svn: 267100
2016-04-22 11:22:38 +08:00
;
[X86] Adding vpopcntd and vpopcntq instructions AVX512_VPOPCNTDQ is a new feature set that was published by Intel. The patch represents the LLVM side of the addition of two new intrinsic based instructions (vpopcntd and vpopcntq). Differential Revision: https://reviews.llvm.org/D33169 llvm-svn: 303858
2017-05-25 21:45:23 +08:00
; AVX512VPOPCNTDQ-LABEL: testv16i8:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; AVX512VPOPCNTDQ: # %bb.0:
[x86] transform vector inc/dec to use -1 constant (PR33483) Convert vector increment or decrement to sub/add with an all-ones constant: add X, <1, 1...> --> sub X, <-1, -1...> sub X, <1, 1...> --> add X, <-1, -1...> The all-ones vector constant can be materialized using a pcmpeq instruction that is commonly recognized as an idiom (has no register dependency), so that's better than loading a splat 1 constant. AVX512 uses 'vpternlogd' for 512-bit vectors because there is apparently no better way to produce 512 one-bits. The general advantages of this lowering are: 1. pcmpeq has lower latency than a memop on every uarch I looked at in Agner's tables, so in theory, this could be better for perf, but... 2. That seems unlikely to affect any OOO implementation, and I can't measure any real perf difference from this transform on Haswell or Jaguar, but... 3. It doesn't look like it from the diffs, but this is an overall size win because we eliminate 16 - 64 constant bytes in the case of a vector load. If we're broadcasting a scalar load (which might itself be a bug), then we're replacing a scalar constant load + broadcast with a single cheap op, so that should always be smaller/better too. 4. This makes the DAG/isel output more consistent - we use pcmpeq already for padd x, -1 and psub x, -1, so we should use that form for +1 too because we can. If there's some reason to favor a constant load on some CPU, let's make the reverse transform for all of these cases (either here in the DAG or in a later machine pass). This should fix: https://bugs.llvm.org/show_bug.cgi?id=33483 Differential Revision: https://reviews.llvm.org/D34336 llvm-svn: 306289
2017-06-26 22:19:26 +08:00
; AVX512VPOPCNTDQ-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; AVX512VPOPCNTDQ-NEXT: vpaddb %xmm1, %xmm0, %xmm1
; AVX512VPOPCNTDQ-NEXT: vpandn %xmm1, %xmm0, %xmm0
[X86][AVX512VPOPCNTDQ] Improve support for v16i8/v8i16/v16i16/ CTPOP Zero extend to v16i32/v8i64, use VPOPCNTDQ instructions and truncate back. llvm-svn: 306990
2017-07-03 03:32:37 +08:00
; AVX512VPOPCNTDQ-NEXT: vpmovzxbd {{.*#+}} zmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4],zero,zero,zero,xmm0[5],zero,zero,zero,xmm0[6],zero,zero,zero,xmm0[7],zero,zero,zero,xmm0[8],zero,zero,zero,xmm0[9],zero,zero,zero,xmm0[10],zero,zero,zero,xmm0[11],zero,zero,zero,xmm0[12],zero,zero,zero,xmm0[13],zero,zero,zero,xmm0[14],zero,zero,zero,xmm0[15],zero,zero,zero
; AVX512VPOPCNTDQ-NEXT: vpopcntd %zmm0, %zmm0
; AVX512VPOPCNTDQ-NEXT: vpmovdb %zmm0, %xmm0
; AVX512VPOPCNTDQ-NEXT: vzeroupper
[X86] Adding vpopcntd and vpopcntq instructions AVX512_VPOPCNTDQ is a new feature set that was published by Intel. The patch represents the LLVM side of the addition of two new intrinsic based instructions (vpopcntd and vpopcntq). Differential Revision: https://reviews.llvm.org/D33169 llvm-svn: 303858
2017-05-25 21:45:23 +08:00
; AVX512VPOPCNTDQ-NEXT: retq
;
[X86] Add 128 and 256-bit VPOPCNTDQ instructions. Adjust some tablegen classes LZCNT/POPCNT. I think when this instruction was first published it was only for a Knights CPU and thus VLX version was missing. llvm-svn: 320910
2017-12-16 10:40:28 +08:00
; AVX512VPOPCNTDQVL-LABEL: testv16i8:
; AVX512VPOPCNTDQVL: # %bb.0:
; AVX512VPOPCNTDQVL-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; AVX512VPOPCNTDQVL-NEXT: vpaddb %xmm1, %xmm0, %xmm1
; AVX512VPOPCNTDQVL-NEXT: vpandn %xmm1, %xmm0, %xmm0
[X86] Add 128 and 256-bit VPOPCNTDQ instructions. Adjust some tablegen classes LZCNT/POPCNT. I think when this instruction was first published it was only for a Knights CPU and thus VLX version was missing. llvm-svn: 320910
2017-12-16 10:40:28 +08:00
; AVX512VPOPCNTDQVL-NEXT: vpmovzxbd {{.*#+}} zmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4],zero,zero,zero,xmm0[5],zero,zero,zero,xmm0[6],zero,zero,zero,xmm0[7],zero,zero,zero,xmm0[8],zero,zero,zero,xmm0[9],zero,zero,zero,xmm0[10],zero,zero,zero,xmm0[11],zero,zero,zero,xmm0[12],zero,zero,zero,xmm0[13],zero,zero,zero,xmm0[14],zero,zero,zero,xmm0[15],zero,zero,zero
; AVX512VPOPCNTDQVL-NEXT: vpopcntd %zmm0, %zmm0
; AVX512VPOPCNTDQVL-NEXT: vpmovdb %zmm0, %xmm0
; AVX512VPOPCNTDQVL-NEXT: vzeroupper
; AVX512VPOPCNTDQVL-NEXT: retq
;
[x86][icelake]BITALG vpopcnt{b,w} Differential Revision: https://reviews.llvm.org/D40213 llvm-svn: 318748
2017-11-21 18:32:42 +08:00
; BITALG_NOVLX-LABEL: testv16i8:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; BITALG_NOVLX: # %bb.0:
[x86][icelake]BITALG vpopcnt{b,w} Differential Revision: https://reviews.llvm.org/D40213 llvm-svn: 318748
2017-11-21 18:32:42 +08:00
; BITALG_NOVLX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; BITALG_NOVLX-NEXT: vpaddb %xmm1, %xmm0, %xmm1
; BITALG_NOVLX-NEXT: vpandn %xmm1, %xmm0, %xmm0
[x86][icelake]BITALG vpopcnt{b,w} Differential Revision: https://reviews.llvm.org/D40213 llvm-svn: 318748
2017-11-21 18:32:42 +08:00
; BITALG_NOVLX-NEXT: vpopcntb %zmm0, %zmm0
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
; BITALG_NOVLX-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
[x86][icelake]BITALG vpopcnt{b,w} Differential Revision: https://reviews.llvm.org/D40213 llvm-svn: 318748
2017-11-21 18:32:42 +08:00
; BITALG_NOVLX-NEXT: vzeroupper
; BITALG_NOVLX-NEXT: retq
;
; BITALG-LABEL: testv16i8:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; BITALG: # %bb.0:
[x86][icelake]BITALG vpopcnt{b,w} Differential Revision: https://reviews.llvm.org/D40213 llvm-svn: 318748
2017-11-21 18:32:42 +08:00
; BITALG-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; BITALG-NEXT: vpaddb %xmm1, %xmm0, %xmm1
; BITALG-NEXT: vpandn %xmm1, %xmm0, %xmm0
[x86][icelake]BITALG vpopcnt{b,w} Differential Revision: https://reviews.llvm.org/D40213 llvm-svn: 318748
2017-11-21 18:32:42 +08:00
; BITALG-NEXT: vpopcntb %xmm0, %xmm0
; BITALG-NEXT: retq
;
[X86][SSE] Added 32-bit tests for vector lzcnt/tzcnt tests v2i64 tests are particularly bad on 32-bit targets. llvm-svn: 265789
2016-04-08 23:01:31 +08:00
; X32-SSE-LABEL: testv16i8:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; X32-SSE: # %bb.0:
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; X32-SSE-NEXT: pcmpeqd %xmm1, %xmm1
; X32-SSE-NEXT: paddb %xmm0, %xmm1
; X32-SSE-NEXT: pandn %xmm1, %xmm0
; X32-SSE-NEXT: movdqa {{.*#+}} xmm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
; X32-SSE-NEXT: movdqa %xmm0, %xmm3
; X32-SSE-NEXT: pand %xmm2, %xmm3
; X32-SSE-NEXT: movdqa {{.*#+}} xmm1 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
; X32-SSE-NEXT: movdqa %xmm1, %xmm4
[X86][SSE] Added 32-bit tests for vector lzcnt/tzcnt tests v2i64 tests are particularly bad on 32-bit targets. llvm-svn: 265789
2016-04-08 23:01:31 +08:00
; X32-SSE-NEXT: pshufb %xmm3, %xmm4
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; X32-SSE-NEXT: psrlw $4, %xmm0
; X32-SSE-NEXT: pand %xmm2, %xmm0
; X32-SSE-NEXT: pshufb %xmm0, %xmm1
; X32-SSE-NEXT: paddb %xmm4, %xmm1
; X32-SSE-NEXT: movdqa %xmm1, %xmm0
[X86][SSE] Added 32-bit tests for vector lzcnt/tzcnt tests v2i64 tests are particularly bad on 32-bit targets. llvm-svn: 265789
2016-04-08 23:01:31 +08:00
; X32-SSE-NEXT: retl
[X86] Added tzcnt vector tests. llvm-svn: 239264
2015-06-08 05:01:34 +08:00
%out = call <16 x i8> @llvm.cttz.v16i8(<16 x i8> %in, i1 0)
ret <16 x i8> %out
}
[X86][SSE] Tidied up vector CTLZ/CTTZ. NFCI. llvm-svn: 242645
2015-07-20 01:09:43 +08:00
define <16 x i8> @testv16i8u(<16 x i8> %in) nounwind {
[X86] Added tzcnt vector tests. llvm-svn: 239264
2015-06-08 05:01:34 +08:00
; SSE2-LABEL: testv16i8u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; SSE2: # %bb.0:
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; SSE2-NEXT: pcmpeqd %xmm1, %xmm1
; SSE2-NEXT: paddb %xmm0, %xmm1
; SSE2-NEXT: pandn %xmm1, %xmm0
; SSE2-NEXT: movdqa %xmm0, %xmm1
; SSE2-NEXT: psrlw $1, %xmm1
; SSE2-NEXT: pand {{.*}}(%rip), %xmm1
; SSE2-NEXT: psubb %xmm1, %xmm0
; SSE2-NEXT: movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
; SSE2-NEXT: movdqa %xmm0, %xmm2
; SSE2-NEXT: pand %xmm1, %xmm2
; SSE2-NEXT: psrlw $2, %xmm0
; SSE2-NEXT: pand %xmm1, %xmm0
[x86] transform vector inc/dec to use -1 constant (PR33483) Convert vector increment or decrement to sub/add with an all-ones constant: add X, <1, 1...> --> sub X, <-1, -1...> sub X, <1, 1...> --> add X, <-1, -1...> The all-ones vector constant can be materialized using a pcmpeq instruction that is commonly recognized as an idiom (has no register dependency), so that's better than loading a splat 1 constant. AVX512 uses 'vpternlogd' for 512-bit vectors because there is apparently no better way to produce 512 one-bits. The general advantages of this lowering are: 1. pcmpeq has lower latency than a memop on every uarch I looked at in Agner's tables, so in theory, this could be better for perf, but... 2. That seems unlikely to affect any OOO implementation, and I can't measure any real perf difference from this transform on Haswell or Jaguar, but... 3. It doesn't look like it from the diffs, but this is an overall size win because we eliminate 16 - 64 constant bytes in the case of a vector load. If we're broadcasting a scalar load (which might itself be a bug), then we're replacing a scalar constant load + broadcast with a single cheap op, so that should always be smaller/better too. 4. This makes the DAG/isel output more consistent - we use pcmpeq already for padd x, -1 and psub x, -1, so we should use that form for +1 too because we can. If there's some reason to favor a constant load on some CPU, let's make the reverse transform for all of these cases (either here in the DAG or in a later machine pass). This should fix: https://bugs.llvm.org/show_bug.cgi?id=33483 Differential Revision: https://reviews.llvm.org/D34336 llvm-svn: 306289
2017-06-26 22:19:26 +08:00
; SSE2-NEXT: paddb %xmm2, %xmm0
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; SSE2-NEXT: movdqa %xmm0, %xmm1
; SSE2-NEXT: psrlw $4, %xmm1
; SSE2-NEXT: paddb %xmm0, %xmm1
; SSE2-NEXT: pand {{.*}}(%rip), %xmm1
; SSE2-NEXT: movdqa %xmm1, %xmm0
[X86] Added tzcnt vector tests. llvm-svn: 239264
2015-06-08 05:01:34 +08:00
; SSE2-NEXT: retq
;
; SSE3-LABEL: testv16i8u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; SSE3: # %bb.0:
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; SSE3-NEXT: pcmpeqd %xmm1, %xmm1
; SSE3-NEXT: paddb %xmm0, %xmm1
; SSE3-NEXT: pandn %xmm1, %xmm0
; SSE3-NEXT: movdqa %xmm0, %xmm1
; SSE3-NEXT: psrlw $1, %xmm1
; SSE3-NEXT: pand {{.*}}(%rip), %xmm1
; SSE3-NEXT: psubb %xmm1, %xmm0
; SSE3-NEXT: movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
; SSE3-NEXT: movdqa %xmm0, %xmm2
; SSE3-NEXT: pand %xmm1, %xmm2
; SSE3-NEXT: psrlw $2, %xmm0
; SSE3-NEXT: pand %xmm1, %xmm0
[x86] transform vector inc/dec to use -1 constant (PR33483) Convert vector increment or decrement to sub/add with an all-ones constant: add X, <1, 1...> --> sub X, <-1, -1...> sub X, <1, 1...> --> add X, <-1, -1...> The all-ones vector constant can be materialized using a pcmpeq instruction that is commonly recognized as an idiom (has no register dependency), so that's better than loading a splat 1 constant. AVX512 uses 'vpternlogd' for 512-bit vectors because there is apparently no better way to produce 512 one-bits. The general advantages of this lowering are: 1. pcmpeq has lower latency than a memop on every uarch I looked at in Agner's tables, so in theory, this could be better for perf, but... 2. That seems unlikely to affect any OOO implementation, and I can't measure any real perf difference from this transform on Haswell or Jaguar, but... 3. It doesn't look like it from the diffs, but this is an overall size win because we eliminate 16 - 64 constant bytes in the case of a vector load. If we're broadcasting a scalar load (which might itself be a bug), then we're replacing a scalar constant load + broadcast with a single cheap op, so that should always be smaller/better too. 4. This makes the DAG/isel output more consistent - we use pcmpeq already for padd x, -1 and psub x, -1, so we should use that form for +1 too because we can. If there's some reason to favor a constant load on some CPU, let's make the reverse transform for all of these cases (either here in the DAG or in a later machine pass). This should fix: https://bugs.llvm.org/show_bug.cgi?id=33483 Differential Revision: https://reviews.llvm.org/D34336 llvm-svn: 306289
2017-06-26 22:19:26 +08:00
; SSE3-NEXT: paddb %xmm2, %xmm0
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; SSE3-NEXT: movdqa %xmm0, %xmm1
; SSE3-NEXT: psrlw $4, %xmm1
; SSE3-NEXT: paddb %xmm0, %xmm1
; SSE3-NEXT: pand {{.*}}(%rip), %xmm1
; SSE3-NEXT: movdqa %xmm1, %xmm0
[X86] Added tzcnt vector tests. llvm-svn: 239264
2015-06-08 05:01:34 +08:00
; SSE3-NEXT: retq
;
; SSSE3-LABEL: testv16i8u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; SSSE3: # %bb.0:
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; SSSE3-NEXT: pcmpeqd %xmm1, %xmm1
; SSSE3-NEXT: paddb %xmm0, %xmm1
; SSSE3-NEXT: pandn %xmm1, %xmm0
; SSSE3-NEXT: movdqa {{.*#+}} xmm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
; SSSE3-NEXT: movdqa %xmm0, %xmm3
; SSSE3-NEXT: pand %xmm2, %xmm3
; SSSE3-NEXT: movdqa {{.*#+}} xmm1 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
; SSSE3-NEXT: movdqa %xmm1, %xmm4
[X86][SSE] Vectorize CTTZ + CTTZ_ZERO_UNDEF Now that we have fast vector CTPOP implementations we can use this to speed up vector CTTZ using the pattern (cttz(x) = ctpop((x & -x) - 1)) Additionally, for AVX512CD that provides lzcnt instructions we can use the pattern (cttz_undef(x) = (width - 1) - ctlz(x & -x)) Differential Revision: http://reviews.llvm.org/D12663 llvm-svn: 248091
2015-09-19 21:22:57 +08:00
; SSSE3-NEXT: pshufb %xmm3, %xmm4
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; SSSE3-NEXT: psrlw $4, %xmm0
; SSSE3-NEXT: pand %xmm2, %xmm0
; SSSE3-NEXT: pshufb %xmm0, %xmm1
; SSSE3-NEXT: paddb %xmm4, %xmm1
; SSSE3-NEXT: movdqa %xmm1, %xmm0
[X86] Added tzcnt vector tests. llvm-svn: 239264
2015-06-08 05:01:34 +08:00
; SSSE3-NEXT: retq
;
; SSE41-LABEL: testv16i8u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; SSE41: # %bb.0:
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; SSE41-NEXT: pcmpeqd %xmm1, %xmm1
; SSE41-NEXT: paddb %xmm0, %xmm1
; SSE41-NEXT: pandn %xmm1, %xmm0
; SSE41-NEXT: movdqa {{.*#+}} xmm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
; SSE41-NEXT: movdqa %xmm0, %xmm3
; SSE41-NEXT: pand %xmm2, %xmm3
; SSE41-NEXT: movdqa {{.*#+}} xmm1 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
; SSE41-NEXT: movdqa %xmm1, %xmm4
[X86][SSE] Vectorize CTTZ + CTTZ_ZERO_UNDEF Now that we have fast vector CTPOP implementations we can use this to speed up vector CTTZ using the pattern (cttz(x) = ctpop((x & -x) - 1)) Additionally, for AVX512CD that provides lzcnt instructions we can use the pattern (cttz_undef(x) = (width - 1) - ctlz(x & -x)) Differential Revision: http://reviews.llvm.org/D12663 llvm-svn: 248091
2015-09-19 21:22:57 +08:00
; SSE41-NEXT: pshufb %xmm3, %xmm4
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; SSE41-NEXT: psrlw $4, %xmm0
; SSE41-NEXT: pand %xmm2, %xmm0
; SSE41-NEXT: pshufb %xmm0, %xmm1
; SSE41-NEXT: paddb %xmm4, %xmm1
; SSE41-NEXT: movdqa %xmm1, %xmm0
[X86] Added tzcnt vector tests. llvm-svn: 239264
2015-06-08 05:01:34 +08:00
; SSE41-NEXT: retq
;
[X86][AVX] Regenerate tzcnt tests. The EVEX -> VEX fix means that AVX/AVX512 code is more likely the same now. llvm-svn: 291241
2017-01-06 23:54:23 +08:00
; AVX-LABEL: testv16i8u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; AVX: # %bb.0:
[x86] transform vector inc/dec to use -1 constant (PR33483) Convert vector increment or decrement to sub/add with an all-ones constant: add X, <1, 1...> --> sub X, <-1, -1...> sub X, <1, 1...> --> add X, <-1, -1...> The all-ones vector constant can be materialized using a pcmpeq instruction that is commonly recognized as an idiom (has no register dependency), so that's better than loading a splat 1 constant. AVX512 uses 'vpternlogd' for 512-bit vectors because there is apparently no better way to produce 512 one-bits. The general advantages of this lowering are: 1. pcmpeq has lower latency than a memop on every uarch I looked at in Agner's tables, so in theory, this could be better for perf, but... 2. That seems unlikely to affect any OOO implementation, and I can't measure any real perf difference from this transform on Haswell or Jaguar, but... 3. It doesn't look like it from the diffs, but this is an overall size win because we eliminate 16 - 64 constant bytes in the case of a vector load. If we're broadcasting a scalar load (which might itself be a bug), then we're replacing a scalar constant load + broadcast with a single cheap op, so that should always be smaller/better too. 4. This makes the DAG/isel output more consistent - we use pcmpeq already for padd x, -1 and psub x, -1, so we should use that form for +1 too because we can. If there's some reason to favor a constant load on some CPU, let's make the reverse transform for all of these cases (either here in the DAG or in a later machine pass). This should fix: https://bugs.llvm.org/show_bug.cgi?id=33483 Differential Revision: https://reviews.llvm.org/D34336 llvm-svn: 306289
2017-06-26 22:19:26 +08:00
; AVX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; AVX-NEXT: vpaddb %xmm1, %xmm0, %xmm1
; AVX-NEXT: vpandn %xmm1, %xmm0, %xmm0
[X86][AVX] Regenerate tzcnt tests. The EVEX -> VEX fix means that AVX/AVX512 code is more likely the same now. llvm-svn: 291241
2017-01-06 23:54:23 +08:00
; AVX-NEXT: vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
; AVX-NEXT: vpand %xmm1, %xmm0, %xmm2
; AVX-NEXT: vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
; AVX-NEXT: vpshufb %xmm2, %xmm3, %xmm2
; AVX-NEXT: vpsrlw $4, %xmm0, %xmm0
; AVX-NEXT: vpand %xmm1, %xmm0, %xmm0
; AVX-NEXT: vpshufb %xmm0, %xmm3, %xmm0
; AVX-NEXT: vpaddb %xmm2, %xmm0, %xmm0
; AVX-NEXT: retq
[AVX512] Teach lowering to use vplzcntd/q to implement 128/256-bit CTTZ_ZERO_UNDEF even without VLX support. We can just extend to 512-bits and extract like we do for CTLZ. llvm-svn: 267100
2016-04-22 11:22:38 +08:00
;
[X86] Adding vpopcntd and vpopcntq instructions AVX512_VPOPCNTDQ is a new feature set that was published by Intel. The patch represents the LLVM side of the addition of two new intrinsic based instructions (vpopcntd and vpopcntq). Differential Revision: https://reviews.llvm.org/D33169 llvm-svn: 303858
2017-05-25 21:45:23 +08:00
; AVX512VPOPCNTDQ-LABEL: testv16i8u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; AVX512VPOPCNTDQ: # %bb.0:
[x86] transform vector inc/dec to use -1 constant (PR33483) Convert vector increment or decrement to sub/add with an all-ones constant: add X, <1, 1...> --> sub X, <-1, -1...> sub X, <1, 1...> --> add X, <-1, -1...> The all-ones vector constant can be materialized using a pcmpeq instruction that is commonly recognized as an idiom (has no register dependency), so that's better than loading a splat 1 constant. AVX512 uses 'vpternlogd' for 512-bit vectors because there is apparently no better way to produce 512 one-bits. The general advantages of this lowering are: 1. pcmpeq has lower latency than a memop on every uarch I looked at in Agner's tables, so in theory, this could be better for perf, but... 2. That seems unlikely to affect any OOO implementation, and I can't measure any real perf difference from this transform on Haswell or Jaguar, but... 3. It doesn't look like it from the diffs, but this is an overall size win because we eliminate 16 - 64 constant bytes in the case of a vector load. If we're broadcasting a scalar load (which might itself be a bug), then we're replacing a scalar constant load + broadcast with a single cheap op, so that should always be smaller/better too. 4. This makes the DAG/isel output more consistent - we use pcmpeq already for padd x, -1 and psub x, -1, so we should use that form for +1 too because we can. If there's some reason to favor a constant load on some CPU, let's make the reverse transform for all of these cases (either here in the DAG or in a later machine pass). This should fix: https://bugs.llvm.org/show_bug.cgi?id=33483 Differential Revision: https://reviews.llvm.org/D34336 llvm-svn: 306289
2017-06-26 22:19:26 +08:00
; AVX512VPOPCNTDQ-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; AVX512VPOPCNTDQ-NEXT: vpaddb %xmm1, %xmm0, %xmm1
; AVX512VPOPCNTDQ-NEXT: vpandn %xmm1, %xmm0, %xmm0
[X86][AVX512VPOPCNTDQ] Improve support for v16i8/v8i16/v16i16/ CTPOP Zero extend to v16i32/v8i64, use VPOPCNTDQ instructions and truncate back. llvm-svn: 306990
2017-07-03 03:32:37 +08:00
; AVX512VPOPCNTDQ-NEXT: vpmovzxbd {{.*#+}} zmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4],zero,zero,zero,xmm0[5],zero,zero,zero,xmm0[6],zero,zero,zero,xmm0[7],zero,zero,zero,xmm0[8],zero,zero,zero,xmm0[9],zero,zero,zero,xmm0[10],zero,zero,zero,xmm0[11],zero,zero,zero,xmm0[12],zero,zero,zero,xmm0[13],zero,zero,zero,xmm0[14],zero,zero,zero,xmm0[15],zero,zero,zero
; AVX512VPOPCNTDQ-NEXT: vpopcntd %zmm0, %zmm0
; AVX512VPOPCNTDQ-NEXT: vpmovdb %zmm0, %xmm0
; AVX512VPOPCNTDQ-NEXT: vzeroupper
[X86] Adding vpopcntd and vpopcntq instructions AVX512_VPOPCNTDQ is a new feature set that was published by Intel. The patch represents the LLVM side of the addition of two new intrinsic based instructions (vpopcntd and vpopcntq). Differential Revision: https://reviews.llvm.org/D33169 llvm-svn: 303858
2017-05-25 21:45:23 +08:00
; AVX512VPOPCNTDQ-NEXT: retq
;
[X86] Add 128 and 256-bit VPOPCNTDQ instructions. Adjust some tablegen classes LZCNT/POPCNT. I think when this instruction was first published it was only for a Knights CPU and thus VLX version was missing. llvm-svn: 320910
2017-12-16 10:40:28 +08:00
; AVX512VPOPCNTDQVL-LABEL: testv16i8u:
; AVX512VPOPCNTDQVL: # %bb.0:
; AVX512VPOPCNTDQVL-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; AVX512VPOPCNTDQVL-NEXT: vpaddb %xmm1, %xmm0, %xmm1
; AVX512VPOPCNTDQVL-NEXT: vpandn %xmm1, %xmm0, %xmm0
[X86] Add 128 and 256-bit VPOPCNTDQ instructions. Adjust some tablegen classes LZCNT/POPCNT. I think when this instruction was first published it was only for a Knights CPU and thus VLX version was missing. llvm-svn: 320910
2017-12-16 10:40:28 +08:00
; AVX512VPOPCNTDQVL-NEXT: vpmovzxbd {{.*#+}} zmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4],zero,zero,zero,xmm0[5],zero,zero,zero,xmm0[6],zero,zero,zero,xmm0[7],zero,zero,zero,xmm0[8],zero,zero,zero,xmm0[9],zero,zero,zero,xmm0[10],zero,zero,zero,xmm0[11],zero,zero,zero,xmm0[12],zero,zero,zero,xmm0[13],zero,zero,zero,xmm0[14],zero,zero,zero,xmm0[15],zero,zero,zero
; AVX512VPOPCNTDQVL-NEXT: vpopcntd %zmm0, %zmm0
; AVX512VPOPCNTDQVL-NEXT: vpmovdb %zmm0, %xmm0
; AVX512VPOPCNTDQVL-NEXT: vzeroupper
; AVX512VPOPCNTDQVL-NEXT: retq
;
[X86] Regenerate the vector-popcnt and vector-tzcnt tests to get BITALG CHECK linse on all functions not just the vXi16/vXi8. llvm-svn: 318885
2017-11-23 07:35:12 +08:00
; BITALG_NOVLX-LABEL: testv16i8u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; BITALG_NOVLX: # %bb.0:
[X86] Regenerate the vector-popcnt and vector-tzcnt tests to get BITALG CHECK linse on all functions not just the vXi16/vXi8. llvm-svn: 318885
2017-11-23 07:35:12 +08:00
; BITALG_NOVLX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; BITALG_NOVLX-NEXT: vpaddb %xmm1, %xmm0, %xmm1
; BITALG_NOVLX-NEXT: vpandn %xmm1, %xmm0, %xmm0
[X86] Regenerate the vector-popcnt and vector-tzcnt tests to get BITALG CHECK linse on all functions not just the vXi16/vXi8. llvm-svn: 318885
2017-11-23 07:35:12 +08:00
; BITALG_NOVLX-NEXT: vpopcntb %zmm0, %zmm0
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
; BITALG_NOVLX-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
[X86] Regenerate the vector-popcnt and vector-tzcnt tests to get BITALG CHECK linse on all functions not just the vXi16/vXi8. llvm-svn: 318885
2017-11-23 07:35:12 +08:00
; BITALG_NOVLX-NEXT: vzeroupper
; BITALG_NOVLX-NEXT: retq
;
; BITALG-LABEL: testv16i8u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; BITALG: # %bb.0:
[X86] Regenerate the vector-popcnt and vector-tzcnt tests to get BITALG CHECK linse on all functions not just the vXi16/vXi8. llvm-svn: 318885
2017-11-23 07:35:12 +08:00
; BITALG-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; BITALG-NEXT: vpaddb %xmm1, %xmm0, %xmm1
; BITALG-NEXT: vpandn %xmm1, %xmm0, %xmm0
[X86] Regenerate the vector-popcnt and vector-tzcnt tests to get BITALG CHECK linse on all functions not just the vXi16/vXi8. llvm-svn: 318885
2017-11-23 07:35:12 +08:00
; BITALG-NEXT: vpopcntb %xmm0, %xmm0
; BITALG-NEXT: retq
;
[X86][SSE] Added 32-bit tests for vector lzcnt/tzcnt tests v2i64 tests are particularly bad on 32-bit targets. llvm-svn: 265789
2016-04-08 23:01:31 +08:00
; X32-SSE-LABEL: testv16i8u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; X32-SSE: # %bb.0:
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; X32-SSE-NEXT: pcmpeqd %xmm1, %xmm1
; X32-SSE-NEXT: paddb %xmm0, %xmm1
; X32-SSE-NEXT: pandn %xmm1, %xmm0
; X32-SSE-NEXT: movdqa {{.*#+}} xmm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
; X32-SSE-NEXT: movdqa %xmm0, %xmm3
; X32-SSE-NEXT: pand %xmm2, %xmm3
; X32-SSE-NEXT: movdqa {{.*#+}} xmm1 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
; X32-SSE-NEXT: movdqa %xmm1, %xmm4
[X86][SSE] Added 32-bit tests for vector lzcnt/tzcnt tests v2i64 tests are particularly bad on 32-bit targets. llvm-svn: 265789
2016-04-08 23:01:31 +08:00
; X32-SSE-NEXT: pshufb %xmm3, %xmm4
[X86][SSE] Change CTTZ vector lowering to cttz(x) = ctpop(~x & (x - 1)) This patch changes the vector CTTZ lowering from: cttz(x) = ctpop((x & -x) - 1) to: cttz(x) = ctpop(~x & (x - 1)) Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first). Differential Revision: https://reviews.llvm.org/D53214 llvm-svn: 344447
2018-10-13 20:12:06 +08:00
; X32-SSE-NEXT: psrlw $4, %xmm0
; X32-SSE-NEXT: pand %xmm2, %xmm0
; X32-SSE-NEXT: pshufb %xmm0, %xmm1
; X32-SSE-NEXT: paddb %xmm4, %xmm1
; X32-SSE-NEXT: movdqa %xmm1, %xmm0
[X86][SSE] Added 32-bit tests for vector lzcnt/tzcnt tests v2i64 tests are particularly bad on 32-bit targets. llvm-svn: 265789
2016-04-08 23:01:31 +08:00
; X32-SSE-NEXT: retl
[X86] Added tzcnt vector tests. llvm-svn: 239264
2015-06-08 05:01:34 +08:00
%out = call <16 x i8> @llvm.cttz.v16i8(<16 x i8> %in, i1 -1)
ret <16 x i8> %out
}
[X86][SSE] Tidied up vector CTLZ/CTTZ. NFCI. llvm-svn: 242645
2015-07-20 01:09:43 +08:00
define <2 x i64> @foldv2i64() nounwind {
[DAGCombiner] Added CTTZ vector constant folding support. llvm-svn: 239293
2015-06-08 17:57:09 +08:00
; SSE-LABEL: foldv2i64:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; SSE: # %bb.0:
[DAGCombiner] Added CTTZ vector constant folding support. llvm-svn: 239293
2015-06-08 17:57:09 +08:00
; SSE-NEXT: movl $8, %eax
[X86][SSE2] Fix asm string for movq (Move Quadword) instruction. Replace "mov{d|q}" with "movq". Differential Revision: https://reviews.llvm.org/D32220 llvm-svn: 301386
2017-04-26 15:08:44 +08:00
; SSE-NEXT: movq %rax, %xmm0
[DAGCombiner] Added CTTZ vector constant folding support. llvm-svn: 239293
2015-06-08 17:57:09 +08:00
; SSE-NEXT: retq
;
; AVX-LABEL: foldv2i64:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; AVX: # %bb.0:
[DAGCombiner] Added CTTZ vector constant folding support. llvm-svn: 239293
2015-06-08 17:57:09 +08:00
; AVX-NEXT: movl $8, %eax
; AVX-NEXT: vmovq %rax, %xmm0
; AVX-NEXT: retq
[X86][SSE] Added 32-bit tests for vector lzcnt/tzcnt tests v2i64 tests are particularly bad on 32-bit targets. llvm-svn: 265789
2016-04-08 23:01:31 +08:00
;
[X86] Adding vpopcntd and vpopcntq instructions AVX512_VPOPCNTDQ is a new feature set that was published by Intel. The patch represents the LLVM side of the addition of two new intrinsic based instructions (vpopcntd and vpopcntq). Differential Revision: https://reviews.llvm.org/D33169 llvm-svn: 303858
2017-05-25 21:45:23 +08:00
; AVX512VPOPCNTDQ-LABEL: foldv2i64:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; AVX512VPOPCNTDQ: # %bb.0:
[X86] Adding vpopcntd and vpopcntq instructions AVX512_VPOPCNTDQ is a new feature set that was published by Intel. The patch represents the LLVM side of the addition of two new intrinsic based instructions (vpopcntd and vpopcntq). Differential Revision: https://reviews.llvm.org/D33169 llvm-svn: 303858
2017-05-25 21:45:23 +08:00
; AVX512VPOPCNTDQ-NEXT: movl $8, %eax
; AVX512VPOPCNTDQ-NEXT: vmovq %rax, %xmm0
; AVX512VPOPCNTDQ-NEXT: retq
;
[X86] Add 128 and 256-bit VPOPCNTDQ instructions. Adjust some tablegen classes LZCNT/POPCNT. I think when this instruction was first published it was only for a Knights CPU and thus VLX version was missing. llvm-svn: 320910
2017-12-16 10:40:28 +08:00
; AVX512VPOPCNTDQVL-LABEL: foldv2i64:
; AVX512VPOPCNTDQVL: # %bb.0:
; AVX512VPOPCNTDQVL-NEXT: movl $8, %eax
; AVX512VPOPCNTDQVL-NEXT: vmovq %rax, %xmm0
; AVX512VPOPCNTDQVL-NEXT: retq
;
[X86] Regenerate the vector-popcnt and vector-tzcnt tests to get BITALG CHECK linse on all functions not just the vXi16/vXi8. llvm-svn: 318885
2017-11-23 07:35:12 +08:00
; BITALG_NOVLX-LABEL: foldv2i64:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; BITALG_NOVLX: # %bb.0:
[X86] Regenerate the vector-popcnt and vector-tzcnt tests to get BITALG CHECK linse on all functions not just the vXi16/vXi8. llvm-svn: 318885
2017-11-23 07:35:12 +08:00
; BITALG_NOVLX-NEXT: movl $8, %eax
; BITALG_NOVLX-NEXT: vmovq %rax, %xmm0
; BITALG_NOVLX-NEXT: retq
;
; BITALG-LABEL: foldv2i64:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; BITALG: # %bb.0:
[X86] Regenerate the vector-popcnt and vector-tzcnt tests to get BITALG CHECK linse on all functions not just the vXi16/vXi8. llvm-svn: 318885
2017-11-23 07:35:12 +08:00
; BITALG-NEXT: movl $8, %eax
; BITALG-NEXT: vmovq %rax, %xmm0
; BITALG-NEXT: retq
;
[X86][SSE] Added 32-bit tests for vector lzcnt/tzcnt tests v2i64 tests are particularly bad on 32-bit targets. llvm-svn: 265789
2016-04-08 23:01:31 +08:00
; X32-SSE-LABEL: foldv2i64:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; X32-SSE: # %bb.0:
[DAGCombine] Allow vector constant folding of any value type before type legalization The patch comes in 2 parts: 1 - it makes use of the SelectionDAG::NewNodesMustHaveLegalTypes flag to tell when it can safely constant fold illegal types. 2 - it correctly resets SelectionDAG::NewNodesMustHaveLegalTypes at the start of each call to SelectionDAGISel::CodeGenAndEmitDAG so all the pre-legalization stages can make use of it - not just the first basic block that gets handled. Fix for PR30760 Differential Revision: https://reviews.llvm.org/D29568 llvm-svn: 294749
2017-02-10 22:37:25 +08:00
; X32-SSE-NEXT: movl $8, %eax
; X32-SSE-NEXT: movd %eax, %xmm0
[X86][SSE] Added 32-bit tests for vector lzcnt/tzcnt tests v2i64 tests are particularly bad on 32-bit targets. llvm-svn: 265789
2016-04-08 23:01:31 +08:00
; X32-SSE-NEXT: retl
[DAGCombiner] Added CTTZ vector constant folding support. llvm-svn: 239293
2015-06-08 17:57:09 +08:00
%out = call <2 x i64> @llvm.cttz.v2i64(<2 x i64> <i64 256, i64 -1>, i1 0)
ret <2 x i64> %out
}
[X86][SSE] Tidied up vector CTLZ/CTTZ. NFCI. llvm-svn: 242645
2015-07-20 01:09:43 +08:00
define <2 x i64> @foldv2i64u() nounwind {
[DAGCombiner] Added CTTZ vector constant folding support. llvm-svn: 239293
2015-06-08 17:57:09 +08:00
; SSE-LABEL: foldv2i64u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; SSE: # %bb.0:
[DAGCombiner] Added CTTZ vector constant folding support. llvm-svn: 239293
2015-06-08 17:57:09 +08:00
; SSE-NEXT: movl $8, %eax
[X86][SSE2] Fix asm string for movq (Move Quadword) instruction. Replace "mov{d|q}" with "movq". Differential Revision: https://reviews.llvm.org/D32220 llvm-svn: 301386
2017-04-26 15:08:44 +08:00
; SSE-NEXT: movq %rax, %xmm0
[DAGCombiner] Added CTTZ vector constant folding support. llvm-svn: 239293
2015-06-08 17:57:09 +08:00
; SSE-NEXT: retq
;
; AVX-LABEL: foldv2i64u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; AVX: # %bb.0:
[DAGCombiner] Added CTTZ vector constant folding support. llvm-svn: 239293
2015-06-08 17:57:09 +08:00
; AVX-NEXT: movl $8, %eax
; AVX-NEXT: vmovq %rax, %xmm0
; AVX-NEXT: retq
[X86][SSE] Added 32-bit tests for vector lzcnt/tzcnt tests v2i64 tests are particularly bad on 32-bit targets. llvm-svn: 265789
2016-04-08 23:01:31 +08:00
;
[X86] Adding vpopcntd and vpopcntq instructions AVX512_VPOPCNTDQ is a new feature set that was published by Intel. The patch represents the LLVM side of the addition of two new intrinsic based instructions (vpopcntd and vpopcntq). Differential Revision: https://reviews.llvm.org/D33169 llvm-svn: 303858
2017-05-25 21:45:23 +08:00
; AVX512VPOPCNTDQ-LABEL: foldv2i64u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; AVX512VPOPCNTDQ: # %bb.0:
[X86] Adding vpopcntd and vpopcntq instructions AVX512_VPOPCNTDQ is a new feature set that was published by Intel. The patch represents the LLVM side of the addition of two new intrinsic based instructions (vpopcntd and vpopcntq). Differential Revision: https://reviews.llvm.org/D33169 llvm-svn: 303858
2017-05-25 21:45:23 +08:00
; AVX512VPOPCNTDQ-NEXT: movl $8, %eax
; AVX512VPOPCNTDQ-NEXT: vmovq %rax, %xmm0
; AVX512VPOPCNTDQ-NEXT: retq
;
[X86] Add 128 and 256-bit VPOPCNTDQ instructions. Adjust some tablegen classes LZCNT/POPCNT. I think when this instruction was first published it was only for a Knights CPU and thus VLX version was missing. llvm-svn: 320910
2017-12-16 10:40:28 +08:00
; AVX512VPOPCNTDQVL-LABEL: foldv2i64u:
; AVX512VPOPCNTDQVL: # %bb.0:
; AVX512VPOPCNTDQVL-NEXT: movl $8, %eax
; AVX512VPOPCNTDQVL-NEXT: vmovq %rax, %xmm0
; AVX512VPOPCNTDQVL-NEXT: retq
;
[X86] Regenerate the vector-popcnt and vector-tzcnt tests to get BITALG CHECK linse on all functions not just the vXi16/vXi8. llvm-svn: 318885
2017-11-23 07:35:12 +08:00
; BITALG_NOVLX-LABEL: foldv2i64u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; BITALG_NOVLX: # %bb.0:
[X86] Regenerate the vector-popcnt and vector-tzcnt tests to get BITALG CHECK linse on all functions not just the vXi16/vXi8. llvm-svn: 318885
2017-11-23 07:35:12 +08:00
; BITALG_NOVLX-NEXT: movl $8, %eax
; BITALG_NOVLX-NEXT: vmovq %rax, %xmm0
; BITALG_NOVLX-NEXT: retq
;
; BITALG-LABEL: foldv2i64u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; BITALG: # %bb.0:
[X86] Regenerate the vector-popcnt and vector-tzcnt tests to get BITALG CHECK linse on all functions not just the vXi16/vXi8. llvm-svn: 318885
2017-11-23 07:35:12 +08:00
; BITALG-NEXT: movl $8, %eax
; BITALG-NEXT: vmovq %rax, %xmm0
; BITALG-NEXT: retq
;
[X86][SSE] Added 32-bit tests for vector lzcnt/tzcnt tests v2i64 tests are particularly bad on 32-bit targets. llvm-svn: 265789
2016-04-08 23:01:31 +08:00
; X32-SSE-LABEL: foldv2i64u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; X32-SSE: # %bb.0:
[DAGCombine] Allow vector constant folding of any value type before type legalization The patch comes in 2 parts: 1 - it makes use of the SelectionDAG::NewNodesMustHaveLegalTypes flag to tell when it can safely constant fold illegal types. 2 - it correctly resets SelectionDAG::NewNodesMustHaveLegalTypes at the start of each call to SelectionDAGISel::CodeGenAndEmitDAG so all the pre-legalization stages can make use of it - not just the first basic block that gets handled. Fix for PR30760 Differential Revision: https://reviews.llvm.org/D29568 llvm-svn: 294749
2017-02-10 22:37:25 +08:00
; X32-SSE-NEXT: movl $8, %eax
; X32-SSE-NEXT: movd %eax, %xmm0
[X86][SSE] Added 32-bit tests for vector lzcnt/tzcnt tests v2i64 tests are particularly bad on 32-bit targets. llvm-svn: 265789
2016-04-08 23:01:31 +08:00
; X32-SSE-NEXT: retl
[DAGCombiner] Added CTTZ vector constant folding support. llvm-svn: 239293
2015-06-08 17:57:09 +08:00
%out = call <2 x i64> @llvm.cttz.v2i64(<2 x i64> <i64 256, i64 -1>, i1 -1)
ret <2 x i64> %out
}
[X86][SSE] Tidied up vector CTLZ/CTTZ. NFCI. llvm-svn: 242645
2015-07-20 01:09:43 +08:00
define <4 x i32> @foldv4i32() nounwind {
[DAGCombiner] Added CTTZ vector constant folding support. llvm-svn: 239293
2015-06-08 17:57:09 +08:00
; SSE-LABEL: foldv4i32:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; SSE: # %bb.0:
[DAGCombiner] Added CTTZ vector constant folding support. llvm-svn: 239293
2015-06-08 17:57:09 +08:00
; SSE-NEXT: movaps {{.*#+}} xmm0 = [8,0,32,0]
; SSE-NEXT: retq
;
[AVX-512] Fix duplicate column in AVX512 execution dependency table that was preventing VMOVDQU32/VMOVDQA32 from being recognized. Fix a bug in the code that stops execution dependency fix from turning operations on 32-bit integer element types into operations on 64-bit integer element types. llvm-svn: 277327
2016-08-01 15:55:33 +08:00
; AVX-LABEL: foldv4i32:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; AVX: # %bb.0:
[AVX-512] Fix duplicate column in AVX512 execution dependency table that was preventing VMOVDQU32/VMOVDQA32 from being recognized. Fix a bug in the code that stops execution dependency fix from turning operations on 32-bit integer element types into operations on 64-bit integer element types. llvm-svn: 277327
2016-08-01 15:55:33 +08:00
; AVX-NEXT: vmovaps {{.*#+}} xmm0 = [8,0,32,0]
; AVX-NEXT: retq
[AVX512] Teach lowering to use vplzcntd/q to implement 128/256-bit CTTZ_ZERO_UNDEF even without VLX support. We can just extend to 512-bits and extract like we do for CTLZ. llvm-svn: 267100
2016-04-22 11:22:38 +08:00
;
[X86] Adding vpopcntd and vpopcntq instructions AVX512_VPOPCNTDQ is a new feature set that was published by Intel. The patch represents the LLVM side of the addition of two new intrinsic based instructions (vpopcntd and vpopcntq). Differential Revision: https://reviews.llvm.org/D33169 llvm-svn: 303858
2017-05-25 21:45:23 +08:00
; AVX512VPOPCNTDQ-LABEL: foldv4i32:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; AVX512VPOPCNTDQ: # %bb.0:
[X86] Adding vpopcntd and vpopcntq instructions AVX512_VPOPCNTDQ is a new feature set that was published by Intel. The patch represents the LLVM side of the addition of two new intrinsic based instructions (vpopcntd and vpopcntq). Differential Revision: https://reviews.llvm.org/D33169 llvm-svn: 303858
2017-05-25 21:45:23 +08:00
; AVX512VPOPCNTDQ-NEXT: vmovaps {{.*#+}} xmm0 = [8,0,32,0]
; AVX512VPOPCNTDQ-NEXT: retq
;
[X86] Add 128 and 256-bit VPOPCNTDQ instructions. Adjust some tablegen classes LZCNT/POPCNT. I think when this instruction was first published it was only for a Knights CPU and thus VLX version was missing. llvm-svn: 320910
2017-12-16 10:40:28 +08:00
; AVX512VPOPCNTDQVL-LABEL: foldv4i32:
; AVX512VPOPCNTDQVL: # %bb.0:
; AVX512VPOPCNTDQVL-NEXT: vmovaps {{.*#+}} xmm0 = [8,0,32,0]
; AVX512VPOPCNTDQVL-NEXT: retq
;
[X86] Regenerate the vector-popcnt and vector-tzcnt tests to get BITALG CHECK linse on all functions not just the vXi16/vXi8. llvm-svn: 318885
2017-11-23 07:35:12 +08:00
; BITALG_NOVLX-LABEL: foldv4i32:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; BITALG_NOVLX: # %bb.0:
[X86] Regenerate the vector-popcnt and vector-tzcnt tests to get BITALG CHECK linse on all functions not just the vXi16/vXi8. llvm-svn: 318885
2017-11-23 07:35:12 +08:00
; BITALG_NOVLX-NEXT: vmovaps {{.*#+}} xmm0 = [8,0,32,0]
; BITALG_NOVLX-NEXT: retq
;
; BITALG-LABEL: foldv4i32:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; BITALG: # %bb.0:
[X86] Regenerate the vector-popcnt and vector-tzcnt tests to get BITALG CHECK linse on all functions not just the vXi16/vXi8. llvm-svn: 318885
2017-11-23 07:35:12 +08:00
; BITALG-NEXT: vmovaps {{.*#+}} xmm0 = [8,0,32,0]
; BITALG-NEXT: retq
;
[X86][SSE] Added 32-bit tests for vector lzcnt/tzcnt tests v2i64 tests are particularly bad on 32-bit targets. llvm-svn: 265789
2016-04-08 23:01:31 +08:00
; X32-SSE-LABEL: foldv4i32:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; X32-SSE: # %bb.0:
[X86][SSE] Added 32-bit tests for vector lzcnt/tzcnt tests v2i64 tests are particularly bad on 32-bit targets. llvm-svn: 265789
2016-04-08 23:01:31 +08:00
; X32-SSE-NEXT: movaps {{.*#+}} xmm0 = [8,0,32,0]
; X32-SSE-NEXT: retl
[DAGCombiner] Added CTTZ vector constant folding support. llvm-svn: 239293
2015-06-08 17:57:09 +08:00
%out = call <4 x i32> @llvm.cttz.v4i32(<4 x i32> <i32 256, i32 -1, i32 0, i32 255>, i1 0)
ret <4 x i32> %out
}
[X86][SSE] Tidied up vector CTLZ/CTTZ. NFCI. llvm-svn: 242645
2015-07-20 01:09:43 +08:00
define <4 x i32> @foldv4i32u() nounwind {
[DAGCombiner] Added CTTZ vector constant folding support. llvm-svn: 239293
2015-06-08 17:57:09 +08:00
; SSE-LABEL: foldv4i32u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; SSE: # %bb.0:
[DAGCombiner] Added CTTZ vector constant folding support. llvm-svn: 239293
2015-06-08 17:57:09 +08:00
; SSE-NEXT: movaps {{.*#+}} xmm0 = [8,0,32,0]
; SSE-NEXT: retq
;
[AVX-512] Fix duplicate column in AVX512 execution dependency table that was preventing VMOVDQU32/VMOVDQA32 from being recognized. Fix a bug in the code that stops execution dependency fix from turning operations on 32-bit integer element types into operations on 64-bit integer element types. llvm-svn: 277327
2016-08-01 15:55:33 +08:00
; AVX-LABEL: foldv4i32u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; AVX: # %bb.0:
[AVX-512] Fix duplicate column in AVX512 execution dependency table that was preventing VMOVDQU32/VMOVDQA32 from being recognized. Fix a bug in the code that stops execution dependency fix from turning operations on 32-bit integer element types into operations on 64-bit integer element types. llvm-svn: 277327
2016-08-01 15:55:33 +08:00
; AVX-NEXT: vmovaps {{.*#+}} xmm0 = [8,0,32,0]
; AVX-NEXT: retq
[AVX512] Teach lowering to use vplzcntd/q to implement 128/256-bit CTTZ_ZERO_UNDEF even without VLX support. We can just extend to 512-bits and extract like we do for CTLZ. llvm-svn: 267100
2016-04-22 11:22:38 +08:00
;
[X86] Adding vpopcntd and vpopcntq instructions AVX512_VPOPCNTDQ is a new feature set that was published by Intel. The patch represents the LLVM side of the addition of two new intrinsic based instructions (vpopcntd and vpopcntq). Differential Revision: https://reviews.llvm.org/D33169 llvm-svn: 303858
2017-05-25 21:45:23 +08:00
; AVX512VPOPCNTDQ-LABEL: foldv4i32u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; AVX512VPOPCNTDQ: # %bb.0:
[X86] Adding vpopcntd and vpopcntq instructions AVX512_VPOPCNTDQ is a new feature set that was published by Intel. The patch represents the LLVM side of the addition of two new intrinsic based instructions (vpopcntd and vpopcntq). Differential Revision: https://reviews.llvm.org/D33169 llvm-svn: 303858
2017-05-25 21:45:23 +08:00
; AVX512VPOPCNTDQ-NEXT: vmovaps {{.*#+}} xmm0 = [8,0,32,0]
; AVX512VPOPCNTDQ-NEXT: retq
;
[X86] Add 128 and 256-bit VPOPCNTDQ instructions. Adjust some tablegen classes LZCNT/POPCNT. I think when this instruction was first published it was only for a Knights CPU and thus VLX version was missing. llvm-svn: 320910
2017-12-16 10:40:28 +08:00
; AVX512VPOPCNTDQVL-LABEL: foldv4i32u:
; AVX512VPOPCNTDQVL: # %bb.0:
; AVX512VPOPCNTDQVL-NEXT: vmovaps {{.*#+}} xmm0 = [8,0,32,0]
; AVX512VPOPCNTDQVL-NEXT: retq
;
[X86] Regenerate the vector-popcnt and vector-tzcnt tests to get BITALG CHECK linse on all functions not just the vXi16/vXi8. llvm-svn: 318885
2017-11-23 07:35:12 +08:00
; BITALG_NOVLX-LABEL: foldv4i32u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; BITALG_NOVLX: # %bb.0:
[X86] Regenerate the vector-popcnt and vector-tzcnt tests to get BITALG CHECK linse on all functions not just the vXi16/vXi8. llvm-svn: 318885
2017-11-23 07:35:12 +08:00
; BITALG_NOVLX-NEXT: vmovaps {{.*#+}} xmm0 = [8,0,32,0]
; BITALG_NOVLX-NEXT: retq
;
; BITALG-LABEL: foldv4i32u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; BITALG: # %bb.0:
[X86] Regenerate the vector-popcnt and vector-tzcnt tests to get BITALG CHECK linse on all functions not just the vXi16/vXi8. llvm-svn: 318885
2017-11-23 07:35:12 +08:00
; BITALG-NEXT: vmovaps {{.*#+}} xmm0 = [8,0,32,0]
; BITALG-NEXT: retq
;
[X86][SSE] Added 32-bit tests for vector lzcnt/tzcnt tests v2i64 tests are particularly bad on 32-bit targets. llvm-svn: 265789
2016-04-08 23:01:31 +08:00
; X32-SSE-LABEL: foldv4i32u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; X32-SSE: # %bb.0:
[X86][SSE] Added 32-bit tests for vector lzcnt/tzcnt tests v2i64 tests are particularly bad on 32-bit targets. llvm-svn: 265789
2016-04-08 23:01:31 +08:00
; X32-SSE-NEXT: movaps {{.*#+}} xmm0 = [8,0,32,0]
; X32-SSE-NEXT: retl
[DAGCombiner] Added CTTZ vector constant folding support. llvm-svn: 239293
2015-06-08 17:57:09 +08:00
%out = call <4 x i32> @llvm.cttz.v4i32(<4 x i32> <i32 256, i32 -1, i32 0, i32 255>, i1 -1)
ret <4 x i32> %out
}
[X86][SSE] Tidied up vector CTLZ/CTTZ. NFCI. llvm-svn: 242645
2015-07-20 01:09:43 +08:00
define <8 x i16> @foldv8i16() nounwind {
[DAGCombiner] Added CTTZ vector constant folding support. llvm-svn: 239293
2015-06-08 17:57:09 +08:00
; SSE-LABEL: foldv8i16:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; SSE: # %bb.0:
[DAGCombiner] Added CTTZ vector constant folding support. llvm-svn: 239293
2015-06-08 17:57:09 +08:00
; SSE-NEXT: movaps {{.*#+}} xmm0 = [8,0,16,0,16,0,3,3]
; SSE-NEXT: retq
;
[AVX-512] Fix duplicate column in AVX512 execution dependency table that was preventing VMOVDQU32/VMOVDQA32 from being recognized. Fix a bug in the code that stops execution dependency fix from turning operations on 32-bit integer element types into operations on 64-bit integer element types. llvm-svn: 277327
2016-08-01 15:55:33 +08:00
; AVX-LABEL: foldv8i16:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; AVX: # %bb.0:
[AVX-512] Fix duplicate column in AVX512 execution dependency table that was preventing VMOVDQU32/VMOVDQA32 from being recognized. Fix a bug in the code that stops execution dependency fix from turning operations on 32-bit integer element types into operations on 64-bit integer element types. llvm-svn: 277327
2016-08-01 15:55:33 +08:00
; AVX-NEXT: vmovaps {{.*#+}} xmm0 = [8,0,16,0,16,0,3,3]
; AVX-NEXT: retq
[AVX512] Teach lowering to use vplzcntd/q to implement 128/256-bit CTTZ_ZERO_UNDEF even without VLX support. We can just extend to 512-bits and extract like we do for CTLZ. llvm-svn: 267100
2016-04-22 11:22:38 +08:00
;
[X86] Adding vpopcntd and vpopcntq instructions AVX512_VPOPCNTDQ is a new feature set that was published by Intel. The patch represents the LLVM side of the addition of two new intrinsic based instructions (vpopcntd and vpopcntq). Differential Revision: https://reviews.llvm.org/D33169 llvm-svn: 303858
2017-05-25 21:45:23 +08:00
; AVX512VPOPCNTDQ-LABEL: foldv8i16:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; AVX512VPOPCNTDQ: # %bb.0:
[X86] Adding vpopcntd and vpopcntq instructions AVX512_VPOPCNTDQ is a new feature set that was published by Intel. The patch represents the LLVM side of the addition of two new intrinsic based instructions (vpopcntd and vpopcntq). Differential Revision: https://reviews.llvm.org/D33169 llvm-svn: 303858
2017-05-25 21:45:23 +08:00
; AVX512VPOPCNTDQ-NEXT: vmovaps {{.*#+}} xmm0 = [8,0,16,0,16,0,3,3]
; AVX512VPOPCNTDQ-NEXT: retq
;
[X86] Add 128 and 256-bit VPOPCNTDQ instructions. Adjust some tablegen classes LZCNT/POPCNT. I think when this instruction was first published it was only for a Knights CPU and thus VLX version was missing. llvm-svn: 320910
2017-12-16 10:40:28 +08:00
; AVX512VPOPCNTDQVL-LABEL: foldv8i16:
; AVX512VPOPCNTDQVL: # %bb.0:
; AVX512VPOPCNTDQVL-NEXT: vmovaps {{.*#+}} xmm0 = [8,0,16,0,16,0,3,3]
; AVX512VPOPCNTDQVL-NEXT: retq
;
[X86] Regenerate the vector-popcnt and vector-tzcnt tests to get BITALG CHECK linse on all functions not just the vXi16/vXi8. llvm-svn: 318885
2017-11-23 07:35:12 +08:00
; BITALG_NOVLX-LABEL: foldv8i16:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; BITALG_NOVLX: # %bb.0:
[X86] Regenerate the vector-popcnt and vector-tzcnt tests to get BITALG CHECK linse on all functions not just the vXi16/vXi8. llvm-svn: 318885
2017-11-23 07:35:12 +08:00
; BITALG_NOVLX-NEXT: vmovaps {{.*#+}} xmm0 = [8,0,16,0,16,0,3,3]
; BITALG_NOVLX-NEXT: retq
;
; BITALG-LABEL: foldv8i16:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; BITALG: # %bb.0:
[X86] Regenerate the vector-popcnt and vector-tzcnt tests to get BITALG CHECK linse on all functions not just the vXi16/vXi8. llvm-svn: 318885
2017-11-23 07:35:12 +08:00
; BITALG-NEXT: vmovaps {{.*#+}} xmm0 = [8,0,16,0,16,0,3,3]
; BITALG-NEXT: retq
;
[X86][SSE] Added 32-bit tests for vector lzcnt/tzcnt tests v2i64 tests are particularly bad on 32-bit targets. llvm-svn: 265789
2016-04-08 23:01:31 +08:00
; X32-SSE-LABEL: foldv8i16:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; X32-SSE: # %bb.0:
[X86][SSE] Added 32-bit tests for vector lzcnt/tzcnt tests v2i64 tests are particularly bad on 32-bit targets. llvm-svn: 265789
2016-04-08 23:01:31 +08:00
; X32-SSE-NEXT: movaps {{.*#+}} xmm0 = [8,0,16,0,16,0,3,3]
; X32-SSE-NEXT: retl
[DAGCombiner] Added CTTZ vector constant folding support. llvm-svn: 239293
2015-06-08 17:57:09 +08:00
%out = call <8 x i16> @llvm.cttz.v8i16(<8 x i16> <i16 256, i16 -1, i16 0, i16 255, i16 -65536, i16 7, i16 24, i16 88>, i1 0)
ret <8 x i16> %out
}
[X86][SSE] Tidied up vector CTLZ/CTTZ. NFCI. llvm-svn: 242645
2015-07-20 01:09:43 +08:00
define <8 x i16> @foldv8i16u() nounwind {
[DAGCombiner] Added CTTZ vector constant folding support. llvm-svn: 239293
2015-06-08 17:57:09 +08:00
; SSE-LABEL: foldv8i16u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; SSE: # %bb.0:
[DAGCombiner] Added CTTZ vector constant folding support. llvm-svn: 239293
2015-06-08 17:57:09 +08:00
; SSE-NEXT: movaps {{.*#+}} xmm0 = [8,0,16,0,16,0,3,3]
; SSE-NEXT: retq
;
[AVX-512] Fix duplicate column in AVX512 execution dependency table that was preventing VMOVDQU32/VMOVDQA32 from being recognized. Fix a bug in the code that stops execution dependency fix from turning operations on 32-bit integer element types into operations on 64-bit integer element types. llvm-svn: 277327
2016-08-01 15:55:33 +08:00
; AVX-LABEL: foldv8i16u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; AVX: # %bb.0:
[AVX-512] Fix duplicate column in AVX512 execution dependency table that was preventing VMOVDQU32/VMOVDQA32 from being recognized. Fix a bug in the code that stops execution dependency fix from turning operations on 32-bit integer element types into operations on 64-bit integer element types. llvm-svn: 277327
2016-08-01 15:55:33 +08:00
; AVX-NEXT: vmovaps {{.*#+}} xmm0 = [8,0,16,0,16,0,3,3]
; AVX-NEXT: retq
[AVX512] Teach lowering to use vplzcntd/q to implement 128/256-bit CTTZ_ZERO_UNDEF even without VLX support. We can just extend to 512-bits and extract like we do for CTLZ. llvm-svn: 267100
2016-04-22 11:22:38 +08:00
;
[X86] Adding vpopcntd and vpopcntq instructions AVX512_VPOPCNTDQ is a new feature set that was published by Intel. The patch represents the LLVM side of the addition of two new intrinsic based instructions (vpopcntd and vpopcntq). Differential Revision: https://reviews.llvm.org/D33169 llvm-svn: 303858
2017-05-25 21:45:23 +08:00
; AVX512VPOPCNTDQ-LABEL: foldv8i16u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; AVX512VPOPCNTDQ: # %bb.0:
[X86] Adding vpopcntd and vpopcntq instructions AVX512_VPOPCNTDQ is a new feature set that was published by Intel. The patch represents the LLVM side of the addition of two new intrinsic based instructions (vpopcntd and vpopcntq). Differential Revision: https://reviews.llvm.org/D33169 llvm-svn: 303858
2017-05-25 21:45:23 +08:00
; AVX512VPOPCNTDQ-NEXT: vmovaps {{.*#+}} xmm0 = [8,0,16,0,16,0,3,3]
; AVX512VPOPCNTDQ-NEXT: retq
;
[X86] Add 128 and 256-bit VPOPCNTDQ instructions. Adjust some tablegen classes LZCNT/POPCNT. I think when this instruction was first published it was only for a Knights CPU and thus VLX version was missing. llvm-svn: 320910
2017-12-16 10:40:28 +08:00
; AVX512VPOPCNTDQVL-LABEL: foldv8i16u:
; AVX512VPOPCNTDQVL: # %bb.0:
; AVX512VPOPCNTDQVL-NEXT: vmovaps {{.*#+}} xmm0 = [8,0,16,0,16,0,3,3]
; AVX512VPOPCNTDQVL-NEXT: retq
;
[X86] Regenerate the vector-popcnt and vector-tzcnt tests to get BITALG CHECK linse on all functions not just the vXi16/vXi8. llvm-svn: 318885
2017-11-23 07:35:12 +08:00
; BITALG_NOVLX-LABEL: foldv8i16u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; BITALG_NOVLX: # %bb.0:
[X86] Regenerate the vector-popcnt and vector-tzcnt tests to get BITALG CHECK linse on all functions not just the vXi16/vXi8. llvm-svn: 318885
2017-11-23 07:35:12 +08:00
; BITALG_NOVLX-NEXT: vmovaps {{.*#+}} xmm0 = [8,0,16,0,16,0,3,3]
; BITALG_NOVLX-NEXT: retq
;
; BITALG-LABEL: foldv8i16u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; BITALG: # %bb.0:
[X86] Regenerate the vector-popcnt and vector-tzcnt tests to get BITALG CHECK linse on all functions not just the vXi16/vXi8. llvm-svn: 318885
2017-11-23 07:35:12 +08:00
; BITALG-NEXT: vmovaps {{.*#+}} xmm0 = [8,0,16,0,16,0,3,3]
; BITALG-NEXT: retq
;
[X86][SSE] Added 32-bit tests for vector lzcnt/tzcnt tests v2i64 tests are particularly bad on 32-bit targets. llvm-svn: 265789
2016-04-08 23:01:31 +08:00
; X32-SSE-LABEL: foldv8i16u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; X32-SSE: # %bb.0:
[X86][SSE] Added 32-bit tests for vector lzcnt/tzcnt tests v2i64 tests are particularly bad on 32-bit targets. llvm-svn: 265789
2016-04-08 23:01:31 +08:00
; X32-SSE-NEXT: movaps {{.*#+}} xmm0 = [8,0,16,0,16,0,3,3]
; X32-SSE-NEXT: retl
[DAGCombiner] Added CTTZ vector constant folding support. llvm-svn: 239293
2015-06-08 17:57:09 +08:00
%out = call <8 x i16> @llvm.cttz.v8i16(<8 x i16> <i16 256, i16 -1, i16 0, i16 255, i16 -65536, i16 7, i16 24, i16 88>, i1 -1)
ret <8 x i16> %out
}
[X86][SSE] Tidied up vector CTLZ/CTTZ. NFCI. llvm-svn: 242645
2015-07-20 01:09:43 +08:00
define <16 x i8> @foldv16i8() nounwind {
[DAGCombiner] Added CTTZ vector constant folding support. llvm-svn: 239293
2015-06-08 17:57:09 +08:00
; SSE-LABEL: foldv16i8:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; SSE: # %bb.0:
[DAGCombiner] Added CTTZ vector constant folding support. llvm-svn: 239293
2015-06-08 17:57:09 +08:00
; SSE-NEXT: movaps {{.*#+}} xmm0 = [8,0,8,0,8,0,3,3,1,1,0,1,2,3,4,5]
; SSE-NEXT: retq
;
[AVX-512] Fix duplicate column in AVX512 execution dependency table that was preventing VMOVDQU32/VMOVDQA32 from being recognized. Fix a bug in the code that stops execution dependency fix from turning operations on 32-bit integer element types into operations on 64-bit integer element types. llvm-svn: 277327
2016-08-01 15:55:33 +08:00
; AVX-LABEL: foldv16i8:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; AVX: # %bb.0:
[AVX-512] Fix duplicate column in AVX512 execution dependency table that was preventing VMOVDQU32/VMOVDQA32 from being recognized. Fix a bug in the code that stops execution dependency fix from turning operations on 32-bit integer element types into operations on 64-bit integer element types. llvm-svn: 277327
2016-08-01 15:55:33 +08:00
; AVX-NEXT: vmovaps {{.*#+}} xmm0 = [8,0,8,0,8,0,3,3,1,1,0,1,2,3,4,5]
; AVX-NEXT: retq
[AVX512] Teach lowering to use vplzcntd/q to implement 128/256-bit CTTZ_ZERO_UNDEF even without VLX support. We can just extend to 512-bits and extract like we do for CTLZ. llvm-svn: 267100
2016-04-22 11:22:38 +08:00
;
[X86] Adding vpopcntd and vpopcntq instructions AVX512_VPOPCNTDQ is a new feature set that was published by Intel. The patch represents the LLVM side of the addition of two new intrinsic based instructions (vpopcntd and vpopcntq). Differential Revision: https://reviews.llvm.org/D33169 llvm-svn: 303858
2017-05-25 21:45:23 +08:00
; AVX512VPOPCNTDQ-LABEL: foldv16i8:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; AVX512VPOPCNTDQ: # %bb.0:
[X86] Adding vpopcntd and vpopcntq instructions AVX512_VPOPCNTDQ is a new feature set that was published by Intel. The patch represents the LLVM side of the addition of two new intrinsic based instructions (vpopcntd and vpopcntq). Differential Revision: https://reviews.llvm.org/D33169 llvm-svn: 303858
2017-05-25 21:45:23 +08:00
; AVX512VPOPCNTDQ-NEXT: vmovaps {{.*#+}} xmm0 = [8,0,8,0,8,0,3,3,1,1,0,1,2,3,4,5]
; AVX512VPOPCNTDQ-NEXT: retq
;
[X86] Add 128 and 256-bit VPOPCNTDQ instructions. Adjust some tablegen classes LZCNT/POPCNT. I think when this instruction was first published it was only for a Knights CPU and thus VLX version was missing. llvm-svn: 320910
2017-12-16 10:40:28 +08:00
; AVX512VPOPCNTDQVL-LABEL: foldv16i8:
; AVX512VPOPCNTDQVL: # %bb.0:
; AVX512VPOPCNTDQVL-NEXT: vmovaps {{.*#+}} xmm0 = [8,0,8,0,8,0,3,3,1,1,0,1,2,3,4,5]
; AVX512VPOPCNTDQVL-NEXT: retq
;
[X86] Regenerate the vector-popcnt and vector-tzcnt tests to get BITALG CHECK linse on all functions not just the vXi16/vXi8. llvm-svn: 318885
2017-11-23 07:35:12 +08:00
; BITALG_NOVLX-LABEL: foldv16i8:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; BITALG_NOVLX: # %bb.0:
[X86] Regenerate the vector-popcnt and vector-tzcnt tests to get BITALG CHECK linse on all functions not just the vXi16/vXi8. llvm-svn: 318885
2017-11-23 07:35:12 +08:00
; BITALG_NOVLX-NEXT: vmovaps {{.*#+}} xmm0 = [8,0,8,0,8,0,3,3,1,1,0,1,2,3,4,5]
; BITALG_NOVLX-NEXT: retq
;
; BITALG-LABEL: foldv16i8:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; BITALG: # %bb.0:
[X86] Regenerate the vector-popcnt and vector-tzcnt tests to get BITALG CHECK linse on all functions not just the vXi16/vXi8. llvm-svn: 318885
2017-11-23 07:35:12 +08:00
; BITALG-NEXT: vmovaps {{.*#+}} xmm0 = [8,0,8,0,8,0,3,3,1,1,0,1,2,3,4,5]
; BITALG-NEXT: retq
;
[X86][SSE] Added 32-bit tests for vector lzcnt/tzcnt tests v2i64 tests are particularly bad on 32-bit targets. llvm-svn: 265789
2016-04-08 23:01:31 +08:00
; X32-SSE-LABEL: foldv16i8:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; X32-SSE: # %bb.0:
[X86][SSE] Added 32-bit tests for vector lzcnt/tzcnt tests v2i64 tests are particularly bad on 32-bit targets. llvm-svn: 265789
2016-04-08 23:01:31 +08:00
; X32-SSE-NEXT: movaps {{.*#+}} xmm0 = [8,0,8,0,8,0,3,3,1,1,0,1,2,3,4,5]
; X32-SSE-NEXT: retl
[DAGCombiner] Added CTTZ vector constant folding support. llvm-svn: 239293
2015-06-08 17:57:09 +08:00
%out = call <16 x i8> @llvm.cttz.v16i8(<16 x i8> <i8 256, i8 -1, i8 0, i8 255, i8 -65536, i8 7, i8 24, i8 88, i8 -2, i8 254, i8 1, i8 2, i8 4, i8 8, i8 16, i8 32>, i1 0)
ret <16 x i8> %out
}
[X86][SSE] Tidied up vector CTLZ/CTTZ. NFCI. llvm-svn: 242645
2015-07-20 01:09:43 +08:00
define <16 x i8> @foldv16i8u() nounwind {
[DAGCombiner] Added CTTZ vector constant folding support. llvm-svn: 239293
2015-06-08 17:57:09 +08:00
; SSE-LABEL: foldv16i8u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; SSE: # %bb.0:
[DAGCombiner] Added CTTZ vector constant folding support. llvm-svn: 239293
2015-06-08 17:57:09 +08:00
; SSE-NEXT: movaps {{.*#+}} xmm0 = [8,0,8,0,8,0,3,3,1,1,0,1,2,3,4,5]
; SSE-NEXT: retq
;
[AVX-512] Fix duplicate column in AVX512 execution dependency table that was preventing VMOVDQU32/VMOVDQA32 from being recognized. Fix a bug in the code that stops execution dependency fix from turning operations on 32-bit integer element types into operations on 64-bit integer element types. llvm-svn: 277327
2016-08-01 15:55:33 +08:00
; AVX-LABEL: foldv16i8u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; AVX: # %bb.0:
[AVX-512] Fix duplicate column in AVX512 execution dependency table that was preventing VMOVDQU32/VMOVDQA32 from being recognized. Fix a bug in the code that stops execution dependency fix from turning operations on 32-bit integer element types into operations on 64-bit integer element types. llvm-svn: 277327
2016-08-01 15:55:33 +08:00
; AVX-NEXT: vmovaps {{.*#+}} xmm0 = [8,0,8,0,8,0,3,3,1,1,0,1,2,3,4,5]
; AVX-NEXT: retq
[AVX512] Teach lowering to use vplzcntd/q to implement 128/256-bit CTTZ_ZERO_UNDEF even without VLX support. We can just extend to 512-bits and extract like we do for CTLZ. llvm-svn: 267100
2016-04-22 11:22:38 +08:00
;
[X86] Adding vpopcntd and vpopcntq instructions AVX512_VPOPCNTDQ is a new feature set that was published by Intel. The patch represents the LLVM side of the addition of two new intrinsic based instructions (vpopcntd and vpopcntq). Differential Revision: https://reviews.llvm.org/D33169 llvm-svn: 303858
2017-05-25 21:45:23 +08:00
; AVX512VPOPCNTDQ-LABEL: foldv16i8u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; AVX512VPOPCNTDQ: # %bb.0:
[X86] Adding vpopcntd and vpopcntq instructions AVX512_VPOPCNTDQ is a new feature set that was published by Intel. The patch represents the LLVM side of the addition of two new intrinsic based instructions (vpopcntd and vpopcntq). Differential Revision: https://reviews.llvm.org/D33169 llvm-svn: 303858
2017-05-25 21:45:23 +08:00
; AVX512VPOPCNTDQ-NEXT: vmovaps {{.*#+}} xmm0 = [8,0,8,0,8,0,3,3,1,1,0,1,2,3,4,5]
; AVX512VPOPCNTDQ-NEXT: retq
;
[X86] Add 128 and 256-bit VPOPCNTDQ instructions. Adjust some tablegen classes LZCNT/POPCNT. I think when this instruction was first published it was only for a Knights CPU and thus VLX version was missing. llvm-svn: 320910
2017-12-16 10:40:28 +08:00
; AVX512VPOPCNTDQVL-LABEL: foldv16i8u:
; AVX512VPOPCNTDQVL: # %bb.0:
; AVX512VPOPCNTDQVL-NEXT: vmovaps {{.*#+}} xmm0 = [8,0,8,0,8,0,3,3,1,1,0,1,2,3,4,5]
; AVX512VPOPCNTDQVL-NEXT: retq
;
[X86] Regenerate the vector-popcnt and vector-tzcnt tests to get BITALG CHECK linse on all functions not just the vXi16/vXi8. llvm-svn: 318885
2017-11-23 07:35:12 +08:00
; BITALG_NOVLX-LABEL: foldv16i8u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; BITALG_NOVLX: # %bb.0:
[X86] Regenerate the vector-popcnt and vector-tzcnt tests to get BITALG CHECK linse on all functions not just the vXi16/vXi8. llvm-svn: 318885
2017-11-23 07:35:12 +08:00
; BITALG_NOVLX-NEXT: vmovaps {{.*#+}} xmm0 = [8,0,8,0,8,0,3,3,1,1,0,1,2,3,4,5]
; BITALG_NOVLX-NEXT: retq
;
; BITALG-LABEL: foldv16i8u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; BITALG: # %bb.0:
[X86] Regenerate the vector-popcnt and vector-tzcnt tests to get BITALG CHECK linse on all functions not just the vXi16/vXi8. llvm-svn: 318885
2017-11-23 07:35:12 +08:00
; BITALG-NEXT: vmovaps {{.*#+}} xmm0 = [8,0,8,0,8,0,3,3,1,1,0,1,2,3,4,5]
; BITALG-NEXT: retq
;
[X86][SSE] Added 32-bit tests for vector lzcnt/tzcnt tests v2i64 tests are particularly bad on 32-bit targets. llvm-svn: 265789
2016-04-08 23:01:31 +08:00
; X32-SSE-LABEL: foldv16i8u:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; X32-SSE: # %bb.0:
[X86][SSE] Added 32-bit tests for vector lzcnt/tzcnt tests v2i64 tests are particularly bad on 32-bit targets. llvm-svn: 265789
2016-04-08 23:01:31 +08:00
; X32-SSE-NEXT: movaps {{.*#+}} xmm0 = [8,0,8,0,8,0,3,3,1,1,0,1,2,3,4,5]
; X32-SSE-NEXT: retl
[DAGCombiner] Added CTTZ vector constant folding support. llvm-svn: 239293
2015-06-08 17:57:09 +08:00
%out = call <16 x i8> @llvm.cttz.v16i8(<16 x i8> <i8 256, i8 -1, i8 0, i8 255, i8 -65536, i8 7, i8 24, i8 88, i8 -2, i8 254, i8 1, i8 2, i8 4, i8 8, i8 16, i8 32>, i1 -1)
ret <16 x i8> %out
}
[X86] Added tzcnt vector tests. llvm-svn: 239264
2015-06-08 05:01:34 +08:00
declare <2 x i64> @llvm.cttz.v2i64(<2 x i64>, i1)
declare <4 x i32> @llvm.cttz.v4i32(<4 x i32>, i1)
declare <8 x i16> @llvm.cttz.v8i16(<8 x i16>, i1)
declare <16 x i8> @llvm.cttz.v16i8(<16 x i8>, i1)