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/oddshuffles.ll

1816 lines
87 KiB
LLVM
Raw Normal View History

[X86] Add baseline test for "odd" shuffles. NFC. Adds a baseline test for lowering shuffles where the width of the output vector is not twice the size of the input vectors. Many of those sequences are suboptimal, and will hopefully be improved in follow-up patches. llvm-svn: 279888
2016-08-27 08:10:24 +08:00
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
[X86][SSE] Regenerate odd shuffle tests with common prefixes llvm-svn: 280661
2016-09-05 22:15:38 +08:00
; RUN: llc < %s -mtriple=x86_64-pc-linux -mattr=+sse2 | FileCheck %s --check-prefix=SSE --check-prefix=SSE2
; RUN: llc < %s -mtriple=x86_64-pc-linux -mattr=+sse4.2 | FileCheck %s --check-prefix=SSE --check-prefix=SSE42
; RUN: llc < %s -mtriple=x86_64-pc-linux -mattr=+avx | FileCheck %s --check-prefix=AVX --check-prefix=AVX1
X86 Tests: Update more isel tests with FastVariableShuffle feature Summary: Added the FastVariableShuffle feature to cases that resembled processors for which this fearure is on. For AVX2 there are processors with and w/o this fearue enable. For AVX512 only KNL does enable this feature so cases which only have +avx512f were left without the FastVariableShuffle enabled. Reviewers: RKSimon, craig.topper Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D41851 llvm-svn: 322090
2018-01-10 00:26:06 +08:00
; RUN: llc < %s -mtriple=x86_64-pc-linux -mattr=+avx2 | FileCheck %s --check-prefixes=AVX,AVX2,AVX2-SLOW
; RUN: llc < %s -mtriple=x86_64-pc-linux -mattr=+avx2,+fast-variable-shuffle | FileCheck %s --check-prefixes=AVX,AVX2,AVX2-FAST
[X86][XOP] Add XOP oddshuffles tests XOP codegen is often different to generic AVX - thank you vpperm! llvm-svn: 315176
2017-10-08 20:58:15 +08:00
; RUN: llc < %s -mtriple=x86_64-pc-linux -mattr=+xop | FileCheck %s --check-prefix=XOP
[X86] Add baseline test for "odd" shuffles. NFC. Adds a baseline test for lowering shuffles where the width of the output vector is not twice the size of the input vectors. Many of those sequences are suboptimal, and will hopefully be improved in follow-up patches. llvm-svn: 279888
2016-08-27 08:10:24 +08:00
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
define void @v3i64(<2 x i64> %a, <2 x i64> %b, <3 x i64>* %p) nounwind {
[X86] Add baseline test for "odd" shuffles. NFC. Adds a baseline test for lowering shuffles where the width of the output vector is not twice the size of the input vectors. Many of those sequences are suboptimal, and will hopefully be improved in follow-up patches. llvm-svn: 279888
2016-08-27 08:10:24 +08:00
; SSE2-LABEL: v3i64:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/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] Add baseline test for "odd" shuffles. NFC. Adds a baseline test for lowering shuffles where the width of the output vector is not twice the size of the input vectors. Many of those sequences are suboptimal, and will hopefully be improved in follow-up patches. llvm-svn: 279888
2016-08-27 08:10:24 +08:00
; SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm0[2,3,0,1]
; SSE2-NEXT: punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
; SSE2-NEXT: movq %xmm2, 16(%rdi)
; SSE2-NEXT: movdqa %xmm0, (%rdi)
; SSE2-NEXT: retq
;
; SSE42-LABEL: v3i64:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; SSE42: # %bb.0:
[X86] Add baseline test for "odd" shuffles. NFC. Adds a baseline test for lowering shuffles where the width of the output vector is not twice the size of the input vectors. Many of those sequences are suboptimal, and will hopefully be improved in follow-up patches. llvm-svn: 279888
2016-08-27 08:10:24 +08:00
; SSE42-NEXT: pextrq $1, %xmm0, 16(%rdi)
; SSE42-NEXT: punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
; SSE42-NEXT: movdqa %xmm0, (%rdi)
; SSE42-NEXT: retq
;
[DAGCombiner] Add EXTRACT_SUBVECTOR to SimplifyDemandedVectorElts As discussed on PR37989, this patch adds EXTRACT_SUBVECTOR handling to TargetLowering::SimplifyDemandedVectorElts and calls it from DAGCombiner::visitEXTRACT_SUBVECTOR. Differential Revision: https://reviews.llvm.org/D48825 llvm-svn: 336490
2018-07-08 01:30:06 +08:00
; AVX-LABEL: v3i64:
; AVX: # %bb.0:
; AVX-NEXT: vpunpcklqdq {{.*#+}} xmm1 = xmm0[0],xmm1[0]
; AVX-NEXT: vpextrq $1, %xmm0, 16(%rdi)
; AVX-NEXT: vmovdqa %xmm1, (%rdi)
; AVX-NEXT: retq
[X86][XOP] Add XOP oddshuffles tests XOP codegen is often different to generic AVX - thank you vpperm! llvm-svn: 315176
2017-10-08 20:58:15 +08:00
;
; XOP-LABEL: v3i64:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; XOP: # %bb.0:
[X86][XOP] Add XOP oddshuffles tests XOP codegen is often different to generic AVX - thank you vpperm! llvm-svn: 315176
2017-10-08 20:58:15 +08:00
; XOP-NEXT: vpunpcklqdq {{.*#+}} xmm1 = xmm0[0],xmm1[0]
; XOP-NEXT: vpextrq $1, %xmm0, 16(%rdi)
; XOP-NEXT: vmovdqa %xmm1, (%rdi)
; XOP-NEXT: retq
[X86] Add baseline test for "odd" shuffles. NFC. Adds a baseline test for lowering shuffles where the width of the output vector is not twice the size of the input vectors. Many of those sequences are suboptimal, and will hopefully be improved in follow-up patches. llvm-svn: 279888
2016-08-27 08:10:24 +08:00
%r = shufflevector <2 x i64> %a, <2 x i64> %b, <3 x i32> <i32 0, i32 2, i32 1>
store <3 x i64> %r, <3 x i64>* %p
ret void
}
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
define void @v3f64(<2 x double> %a, <2 x double> %b, <3 x double>* %p) nounwind {
[X86][SSE] Regenerate odd shuffle tests with common prefixes llvm-svn: 280661
2016-09-05 22:15:38 +08:00
; SSE-LABEL: v3f64:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/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:
[X86][SSE] Regenerate odd shuffle tests with common prefixes llvm-svn: 280661
2016-09-05 22:15:38 +08:00
; SSE-NEXT: movhpd %xmm0, 16(%rdi)
; SSE-NEXT: unpcklpd {{.*#+}} xmm0 = xmm0[0],xmm1[0]
; SSE-NEXT: movapd %xmm0, (%rdi)
; SSE-NEXT: retq
;
[DAGCombiner] Add EXTRACT_SUBVECTOR to SimplifyDemandedVectorElts As discussed on PR37989, this patch adds EXTRACT_SUBVECTOR handling to TargetLowering::SimplifyDemandedVectorElts and calls it from DAGCombiner::visitEXTRACT_SUBVECTOR. Differential Revision: https://reviews.llvm.org/D48825 llvm-svn: 336490
2018-07-08 01:30:06 +08:00
; AVX-LABEL: v3f64:
; AVX: # %bb.0:
; AVX-NEXT: vunpcklpd {{.*#+}} xmm1 = xmm0[0],xmm1[0]
; AVX-NEXT: vmovhpd %xmm0, 16(%rdi)
; AVX-NEXT: vmovapd %xmm1, (%rdi)
; AVX-NEXT: retq
[X86][XOP] Add XOP oddshuffles tests XOP codegen is often different to generic AVX - thank you vpperm! llvm-svn: 315176
2017-10-08 20:58:15 +08:00
;
; XOP-LABEL: v3f64:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; XOP: # %bb.0:
[X86][XOP] Add XOP oddshuffles tests XOP codegen is often different to generic AVX - thank you vpperm! llvm-svn: 315176
2017-10-08 20:58:15 +08:00
; XOP-NEXT: vunpcklpd {{.*#+}} xmm1 = xmm0[0],xmm1[0]
; XOP-NEXT: vmovhpd %xmm0, 16(%rdi)
; XOP-NEXT: vmovapd %xmm1, (%rdi)
; XOP-NEXT: retq
[SelectionDAG] Generate vector_shuffle nodes for undersized result vector sizes Prior to this, we could generate a vector_shuffle from an IR shuffle when the size of the result was exactly the sum of the sizes of the input vectors. If the output vector was narrower - e.g. a <12 x i8> being formed by a shuffle with two <8 x i8> inputs - we would lower the shuffle to a sequence of extracts and inserts. Instead, we can form a larger vector_shuffle, and then extract a subvector of the right size - e.g. shuffle the two <8 x i8> inputs into a <16 x i8> and then extract a <12 x i8>. This also includes a target-specific X86 combine that in the presence of AVX2 combines: (vector_shuffle <mask> (concat_vectors t1, undef) (concat_vectors t2, undef)) into: (vector_shuffle <mask> (concat_vectors t1, t2), undef) in cases where this allows us to form VPERMD/VPERMQ. (This is not a separate commit, as that pattern does not appear without the DAGBuilder change.) llvm-svn: 280418
2016-09-02 05:32:09 +08:00
%r = shufflevector <2 x double> %a, <2 x double> %b, <3 x i32> <i32 0, i32 2, i32 1>
store <3 x double> %r, <3 x double>* %p
ret void
}
[X86] Add baseline test for "odd" shuffles. NFC. Adds a baseline test for lowering shuffles where the width of the output vector is not twice the size of the input vectors. Many of those sequences are suboptimal, and will hopefully be improved in follow-up patches. llvm-svn: 279888
2016-08-27 08:10:24 +08:00
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
define void @v3i32(<2 x i32> %a, <2 x i32> %b, <3 x i32>* %p) nounwind {
[X86] Add baseline test for "odd" shuffles. NFC. Adds a baseline test for lowering shuffles where the width of the output vector is not twice the size of the input vectors. Many of those sequences are suboptimal, and will hopefully be improved in follow-up patches. llvm-svn: 279888
2016-08-27 08:10:24 +08:00
; SSE2-LABEL: v3i32:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/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:
[DAGCombiner] Call SimplifyDemandedVectorElts from EXTRACT_VECTOR_ELT If we are only extracting vector elements via EXTRACT_VECTOR_ELT(s) we may be able to use SimplifyDemandedVectorElts to avoid unnecessary vector ops. Differential Revision: https://reviews.llvm.org/D49262 llvm-svn: 337258
2018-07-17 17:45:35 +08:00
; SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm0[2,3,0,1]
; SSE2-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; SSE2-NEXT: movd %xmm2, 8(%rdi)
; SSE2-NEXT: movq %xmm0, (%rdi)
[X86] Add baseline test for "odd" shuffles. NFC. Adds a baseline test for lowering shuffles where the width of the output vector is not twice the size of the input vectors. Many of those sequences are suboptimal, and will hopefully be improved in follow-up patches. llvm-svn: 279888
2016-08-27 08:10:24 +08:00
; SSE2-NEXT: retq
;
; SSE42-LABEL: v3i32:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; SSE42: # %bb.0:
[DAGCombiner] Call SimplifyDemandedVectorElts from EXTRACT_VECTOR_ELT If we are only extracting vector elements via EXTRACT_VECTOR_ELT(s) we may be able to use SimplifyDemandedVectorElts to avoid unnecessary vector ops. Differential Revision: https://reviews.llvm.org/D49262 llvm-svn: 337258
2018-07-17 17:45:35 +08:00
; SSE42-NEXT: extractps $2, %xmm0, 8(%rdi)
; SSE42-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; SSE42-NEXT: movlps %xmm0, (%rdi)
[X86] Add baseline test for "odd" shuffles. NFC. Adds a baseline test for lowering shuffles where the width of the output vector is not twice the size of the input vectors. Many of those sequences are suboptimal, and will hopefully be improved in follow-up patches. llvm-svn: 279888
2016-08-27 08:10:24 +08:00
; SSE42-NEXT: retq
;
[DAGCombiner] Call SimplifyDemandedVectorElts from EXTRACT_VECTOR_ELT If we are only extracting vector elements via EXTRACT_VECTOR_ELT(s) we may be able to use SimplifyDemandedVectorElts to avoid unnecessary vector ops. Differential Revision: https://reviews.llvm.org/D49262 llvm-svn: 337258
2018-07-17 17:45:35 +08:00
; AVX-LABEL: v3i32:
; AVX: # %bb.0:
; AVX-NEXT: vunpcklps {{.*#+}} xmm1 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; AVX-NEXT: vextractps $2, %xmm0, 8(%rdi)
; AVX-NEXT: vmovlps %xmm1, (%rdi)
; AVX-NEXT: retq
[X86][XOP] Add XOP oddshuffles tests XOP codegen is often different to generic AVX - thank you vpperm! llvm-svn: 315176
2017-10-08 20:58:15 +08:00
;
; XOP-LABEL: v3i32:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; XOP: # %bb.0:
[DAGCombiner] Call SimplifyDemandedVectorElts from EXTRACT_VECTOR_ELT If we are only extracting vector elements via EXTRACT_VECTOR_ELT(s) we may be able to use SimplifyDemandedVectorElts to avoid unnecessary vector ops. Differential Revision: https://reviews.llvm.org/D49262 llvm-svn: 337258
2018-07-17 17:45:35 +08:00
; XOP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
[X86][SSE] Add custom execution domain fixing for BLENDPD/BLENDPS/PBLENDD/PBLENDW (PR34873) Add support for custom execution domain fixing and implement support for BLENDPD/BLENDPS/PBLENDD/PBLENDW. Differential Revision: https://reviews.llvm.org/D42042 llvm-svn: 322524
2018-01-16 06:18:45 +08:00
; XOP-NEXT: vextractps $2, %xmm0, 8(%rdi)
; XOP-NEXT: vmovlps %xmm1, (%rdi)
[X86][XOP] Add XOP oddshuffles tests XOP codegen is often different to generic AVX - thank you vpperm! llvm-svn: 315176
2017-10-08 20:58:15 +08:00
; XOP-NEXT: retq
[X86] Add baseline test for "odd" shuffles. NFC. Adds a baseline test for lowering shuffles where the width of the output vector is not twice the size of the input vectors. Many of those sequences are suboptimal, and will hopefully be improved in follow-up patches. llvm-svn: 279888
2016-08-27 08:10:24 +08:00
%r = shufflevector <2 x i32> %a, <2 x i32> %b, <3 x i32> <i32 0, i32 2, i32 1>
store <3 x i32> %r, <3 x i32>* %p
ret void
}
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
define void @v5i16(<4 x i16> %a, <4 x i16> %b, <5 x i16>* %p) nounwind {
[X86] Add baseline test for "odd" shuffles. NFC. Adds a baseline test for lowering shuffles where the width of the output vector is not twice the size of the input vectors. Many of those sequences are suboptimal, and will hopefully be improved in follow-up patches. llvm-svn: 279888
2016-08-27 08:10:24 +08:00
; SSE2-LABEL: v5i16:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/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:
[DAGCombiner] Call SimplifyDemandedVectorElts from EXTRACT_VECTOR_ELT If we are only extracting vector elements via EXTRACT_VECTOR_ELT(s) we may be able to use SimplifyDemandedVectorElts to avoid unnecessary vector ops. Differential Revision: https://reviews.llvm.org/D49262 llvm-svn: 337258
2018-07-17 17:45:35 +08:00
; SSE2-NEXT: pshuflw {{.*#+}} xmm2 = xmm0[0,2,2,3,4,5,6,7]
[X86] Add baseline test for "odd" shuffles. NFC. Adds a baseline test for lowering shuffles where the width of the output vector is not twice the size of the input vectors. Many of those sequences are suboptimal, and will hopefully be improved in follow-up patches. llvm-svn: 279888
2016-08-27 08:10:24 +08:00
; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm1[2,1,2,3]
; SSE2-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[2,0,2,3,4,5,6,7]
; SSE2-NEXT: punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; SSE2-NEXT: pextrw $6, %xmm0, %eax
; SSE2-NEXT: movw %ax, 8(%rdi)
; SSE2-NEXT: movq %xmm2, (%rdi)
; SSE2-NEXT: retq
;
; SSE42-LABEL: v5i16:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; SSE42: # %bb.0:
[DAGCombiner] Call SimplifyDemandedVectorElts from EXTRACT_VECTOR_ELT If we are only extracting vector elements via EXTRACT_VECTOR_ELT(s) we may be able to use SimplifyDemandedVectorElts to avoid unnecessary vector ops. Differential Revision: https://reviews.llvm.org/D49262 llvm-svn: 337258
2018-07-17 17:45:35 +08:00
; SSE42-NEXT: pshuflw {{.*#+}} xmm2 = xmm0[0,2,2,3,4,5,6,7]
[X86] Add baseline test for "odd" shuffles. NFC. Adds a baseline test for lowering shuffles where the width of the output vector is not twice the size of the input vectors. Many of those sequences are suboptimal, and will hopefully be improved in follow-up patches. llvm-svn: 279888
2016-08-27 08:10:24 +08:00
; SSE42-NEXT: pshufd {{.*#+}} xmm1 = xmm1[2,1,2,3]
; SSE42-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[2,0,2,3,4,5,6,7]
; SSE42-NEXT: punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; SSE42-NEXT: pextrw $6, %xmm0, 8(%rdi)
; SSE42-NEXT: movq %xmm2, (%rdi)
; SSE42-NEXT: retq
;
X86 Tests: Update more isel tests with FastVariableShuffle feature Summary: Added the FastVariableShuffle feature to cases that resembled processors for which this fearure is on. For AVX2 there are processors with and w/o this fearue enable. For AVX512 only KNL does enable this feature so cases which only have +avx512f were left without the FastVariableShuffle enabled. Reviewers: RKSimon, craig.topper Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D41851 llvm-svn: 322090
2018-01-10 00:26:06 +08:00
; AVX1-LABEL: v5i16:
; AVX1: # %bb.0:
[DAGCombiner] Call SimplifyDemandedVectorElts from EXTRACT_VECTOR_ELT If we are only extracting vector elements via EXTRACT_VECTOR_ELT(s) we may be able to use SimplifyDemandedVectorElts to avoid unnecessary vector ops. Differential Revision: https://reviews.llvm.org/D49262 llvm-svn: 337258
2018-07-17 17:45:35 +08:00
; AVX1-NEXT: vpshuflw {{.*#+}} xmm2 = xmm0[0,2,2,3,4,5,6,7]
X86 Tests: Update more isel tests with FastVariableShuffle feature Summary: Added the FastVariableShuffle feature to cases that resembled processors for which this fearure is on. For AVX2 there are processors with and w/o this fearue enable. For AVX512 only KNL does enable this feature so cases which only have +avx512f were left without the FastVariableShuffle enabled. Reviewers: RKSimon, craig.topper Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D41851 llvm-svn: 322090
2018-01-10 00:26:06 +08:00
; AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[2,1,2,3]
; AVX1-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[2,0,2,3,4,5,6,7]
; AVX1-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX1-NEXT: vpextrw $6, %xmm0, 8(%rdi)
; AVX1-NEXT: vmovq %xmm1, (%rdi)
; AVX1-NEXT: retq
;
; AVX2-SLOW-LABEL: v5i16:
; AVX2-SLOW: # %bb.0:
[DAGCombiner] Call SimplifyDemandedVectorElts from EXTRACT_VECTOR_ELT If we are only extracting vector elements via EXTRACT_VECTOR_ELT(s) we may be able to use SimplifyDemandedVectorElts to avoid unnecessary vector ops. Differential Revision: https://reviews.llvm.org/D49262 llvm-svn: 337258
2018-07-17 17:45:35 +08:00
; AVX2-SLOW-NEXT: vpshuflw {{.*#+}} xmm2 = xmm0[0,2,2,3,4,5,6,7]
X86 Tests: Update more isel tests with FastVariableShuffle feature Summary: Added the FastVariableShuffle feature to cases that resembled processors for which this fearure is on. For AVX2 there are processors with and w/o this fearue enable. For AVX512 only KNL does enable this feature so cases which only have +avx512f were left without the FastVariableShuffle enabled. Reviewers: RKSimon, craig.topper Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D41851 llvm-svn: 322090
2018-01-10 00:26:06 +08:00
; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[2,1,2,3]
; AVX2-SLOW-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[2,0,2,3,4,5,6,7]
; AVX2-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX2-SLOW-NEXT: vpextrw $6, %xmm0, 8(%rdi)
; AVX2-SLOW-NEXT: vmovq %xmm1, (%rdi)
; AVX2-SLOW-NEXT: retq
;
; AVX2-FAST-LABEL: v5i16:
; AVX2-FAST: # %bb.0:
; AVX2-FAST-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[4,5,8,9,4,5,6,7,8,9,10,11,12,13,14,15]
[DAGCombiner] Call SimplifyDemandedVectorElts from EXTRACT_VECTOR_ELT If we are only extracting vector elements via EXTRACT_VECTOR_ELT(s) we may be able to use SimplifyDemandedVectorElts to avoid unnecessary vector ops. Differential Revision: https://reviews.llvm.org/D49262 llvm-svn: 337258
2018-07-17 17:45:35 +08:00
; AVX2-FAST-NEXT: vpshuflw {{.*#+}} xmm2 = xmm0[0,2,2,3,4,5,6,7]
X86 Tests: Update more isel tests with FastVariableShuffle feature Summary: Added the FastVariableShuffle feature to cases that resembled processors for which this fearure is on. For AVX2 there are processors with and w/o this fearue enable. For AVX512 only KNL does enable this feature so cases which only have +avx512f were left without the FastVariableShuffle enabled. Reviewers: RKSimon, craig.topper Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D41851 llvm-svn: 322090
2018-01-10 00:26:06 +08:00
; AVX2-FAST-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX2-FAST-NEXT: vpextrw $6, %xmm0, 8(%rdi)
; AVX2-FAST-NEXT: vmovq %xmm1, (%rdi)
; AVX2-FAST-NEXT: retq
[X86][XOP] Add XOP oddshuffles tests XOP codegen is often different to generic AVX - thank you vpperm! llvm-svn: 315176
2017-10-08 20:58:15 +08:00
;
; XOP-LABEL: v5i16:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; XOP: # %bb.0:
[DAGCombiner] Call SimplifyDemandedVectorElts from EXTRACT_VECTOR_ELT If we are only extracting vector elements via EXTRACT_VECTOR_ELT(s) we may be able to use SimplifyDemandedVectorElts to avoid unnecessary vector ops. Differential Revision: https://reviews.llvm.org/D49262 llvm-svn: 337258
2018-07-17 17:45:35 +08:00
; XOP-NEXT: vpperm {{.*#+}} xmm1 = xmm0[0,1],xmm1[4,5],xmm0[4,5],xmm1[8,9],xmm0[4,5],xmm1[4,5],xmm0[6,7],xmm1[6,7]
[X86][XOP] Add XOP oddshuffles tests XOP codegen is often different to generic AVX - thank you vpperm! llvm-svn: 315176
2017-10-08 20:58:15 +08:00
; XOP-NEXT: vpextrw $6, %xmm0, 8(%rdi)
; XOP-NEXT: vmovq %xmm1, (%rdi)
; XOP-NEXT: retq
[X86] Add baseline test for "odd" shuffles. NFC. Adds a baseline test for lowering shuffles where the width of the output vector is not twice the size of the input vectors. Many of those sequences are suboptimal, and will hopefully be improved in follow-up patches. llvm-svn: 279888
2016-08-27 08:10:24 +08:00
%r = shufflevector <4 x i16> %a, <4 x i16> %b, <5 x i32> <i32 0, i32 5, i32 1, i32 6, i32 3>
store <5 x i16> %r, <5 x i16>* %p
ret void
}
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
define void @v5i32(<4 x i32> %a, <4 x i32> %b, <5 x i32>* %p) nounwind {
[X86] Add baseline test for "odd" shuffles. NFC. Adds a baseline test for lowering shuffles where the width of the output vector is not twice the size of the input vectors. Many of those sequences are suboptimal, and will hopefully be improved in follow-up patches. llvm-svn: 279888
2016-08-27 08:10:24 +08:00
; SSE2-LABEL: v5i32:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/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] Add baseline test for "odd" shuffles. NFC. Adds a baseline test for lowering shuffles where the width of the output vector is not twice the size of the input vectors. Many of those sequences are suboptimal, and will hopefully be improved in follow-up patches. llvm-svn: 279888
2016-08-27 08:10:24 +08:00
; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,2,2,3]
; SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm0[3,1,2,3]
; SSE2-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; SSE2-NEXT: movd %xmm2, 16(%rdi)
; SSE2-NEXT: movdqa %xmm0, (%rdi)
; SSE2-NEXT: retq
;
; SSE42-LABEL: v5i32:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; SSE42: # %bb.0:
[X86] Add baseline test for "odd" shuffles. NFC. Adds a baseline test for lowering shuffles where the width of the output vector is not twice the size of the input vectors. Many of those sequences are suboptimal, and will hopefully be improved in follow-up patches. llvm-svn: 279888
2016-08-27 08:10:24 +08:00
; SSE42-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,1,2,2]
; SSE42-NEXT: pmovzxdq {{.*#+}} xmm2 = xmm0[0],zero,xmm0[1],zero
; SSE42-NEXT: pblendw {{.*#+}} xmm2 = xmm2[0,1],xmm1[2,3],xmm2[4,5],xmm1[6,7]
; SSE42-NEXT: pextrd $3, %xmm0, 16(%rdi)
; SSE42-NEXT: movdqa %xmm2, (%rdi)
; SSE42-NEXT: retq
;
[X86][SSE] Regenerate odd shuffle tests with common prefixes llvm-svn: 280661
2016-09-05 22:15:38 +08:00
; AVX1-LABEL: v5i32:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/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:
[DAGCombiner] Add EXTRACT_SUBVECTOR to SimplifyDemandedVectorElts As discussed on PR37989, this patch adds EXTRACT_SUBVECTOR handling to TargetLowering::SimplifyDemandedVectorElts and calls it from DAGCombiner::visitEXTRACT_SUBVECTOR. Differential Revision: https://reviews.llvm.org/D48825 llvm-svn: 336490
2018-07-08 01:30:06 +08:00
; AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[0,1,2,2]
; AVX1-NEXT: vpmovzxdq {{.*#+}} xmm2 = xmm0[0],zero,xmm0[1],zero
; AVX1-NEXT: vpblendw {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3],xmm2[4,5],xmm1[6,7]
; AVX1-NEXT: vpextrd $3, %xmm0, 16(%rdi)
; AVX1-NEXT: vmovdqa %xmm1, (%rdi)
[X86][SSE] Regenerate odd shuffle tests with common prefixes llvm-svn: 280661
2016-09-05 22:15:38 +08:00
; AVX1-NEXT: retq
[X86] Add baseline test for "odd" shuffles. NFC. Adds a baseline test for lowering shuffles where the width of the output vector is not twice the size of the input vectors. Many of those sequences are suboptimal, and will hopefully be improved in follow-up patches. llvm-svn: 279888
2016-08-27 08:10:24 +08:00
;
; AVX2-LABEL: v5i32:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/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:
[DAGCombiner] Add EXTRACT_SUBVECTOR to SimplifyDemandedVectorElts As discussed on PR37989, this patch adds EXTRACT_SUBVECTOR handling to TargetLowering::SimplifyDemandedVectorElts and calls it from DAGCombiner::visitEXTRACT_SUBVECTOR. Differential Revision: https://reviews.llvm.org/D48825 llvm-svn: 336490
2018-07-08 01:30:06 +08:00
; AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[0,1,2,2]
; AVX2-NEXT: vpmovzxdq {{.*#+}} xmm2 = xmm0[0],zero,xmm0[1],zero
; AVX2-NEXT: vpblendd {{.*#+}} xmm1 = xmm2[0],xmm1[1],xmm2[2],xmm1[3]
; AVX2-NEXT: vpextrd $3, %xmm0, 16(%rdi)
; AVX2-NEXT: vmovdqa %xmm1, (%rdi)
[X86] Add baseline test for "odd" shuffles. NFC. Adds a baseline test for lowering shuffles where the width of the output vector is not twice the size of the input vectors. Many of those sequences are suboptimal, and will hopefully be improved in follow-up patches. llvm-svn: 279888
2016-08-27 08:10:24 +08:00
; AVX2-NEXT: retq
[X86][XOP] Add XOP oddshuffles tests XOP codegen is often different to generic AVX - thank you vpperm! llvm-svn: 315176
2017-10-08 20:58:15 +08:00
;
; XOP-LABEL: v5i32:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; XOP: # %bb.0:
[DAGCombiner] Add EXTRACT_SUBVECTOR to SimplifyDemandedVectorElts As discussed on PR37989, this patch adds EXTRACT_SUBVECTOR handling to TargetLowering::SimplifyDemandedVectorElts and calls it from DAGCombiner::visitEXTRACT_SUBVECTOR. Differential Revision: https://reviews.llvm.org/D48825 llvm-svn: 336490
2018-07-08 01:30:06 +08:00
; XOP-NEXT: vpperm {{.*#+}} xmm1 = xmm0[0,1,2,3],xmm1[4,5,6,7],xmm0[4,5,6,7],xmm1[8,9,10,11]
; XOP-NEXT: vpextrd $3, %xmm0, 16(%rdi)
; XOP-NEXT: vmovdqa %xmm1, (%rdi)
[X86][XOP] Add XOP oddshuffles tests XOP codegen is often different to generic AVX - thank you vpperm! llvm-svn: 315176
2017-10-08 20:58:15 +08:00
; XOP-NEXT: retq
[X86] Add baseline test for "odd" shuffles. NFC. Adds a baseline test for lowering shuffles where the width of the output vector is not twice the size of the input vectors. Many of those sequences are suboptimal, and will hopefully be improved in follow-up patches. llvm-svn: 279888
2016-08-27 08:10:24 +08:00
%r = shufflevector <4 x i32> %a, <4 x i32> %b, <5 x i32> <i32 0, i32 5, i32 1, i32 6, i32 3>
store <5 x i32> %r, <5 x i32>* %p
ret void
}
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
define void @v5f32(<4 x float> %a, <4 x float> %b, <5 x float>* %p) nounwind {
[SelectionDAG] Generate vector_shuffle nodes for undersized result vector sizes Prior to this, we could generate a vector_shuffle from an IR shuffle when the size of the result was exactly the sum of the sizes of the input vectors. If the output vector was narrower - e.g. a <12 x i8> being formed by a shuffle with two <8 x i8> inputs - we would lower the shuffle to a sequence of extracts and inserts. Instead, we can form a larger vector_shuffle, and then extract a subvector of the right size - e.g. shuffle the two <8 x i8> inputs into a <16 x i8> and then extract a <12 x i8>. This also includes a target-specific X86 combine that in the presence of AVX2 combines: (vector_shuffle <mask> (concat_vectors t1, undef) (concat_vectors t2, undef)) into: (vector_shuffle <mask> (concat_vectors t1, t2), undef) in cases where this allows us to form VPERMD/VPERMQ. (This is not a separate commit, as that pattern does not appear without the DAGBuilder change.) llvm-svn: 280418
2016-09-02 05:32:09 +08:00
; SSE2-LABEL: v5f32:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/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:
[SelectionDAG] Generate vector_shuffle nodes for undersized result vector sizes Prior to this, we could generate a vector_shuffle from an IR shuffle when the size of the result was exactly the sum of the sizes of the input vectors. If the output vector was narrower - e.g. a <12 x i8> being formed by a shuffle with two <8 x i8> inputs - we would lower the shuffle to a sequence of extracts and inserts. Instead, we can form a larger vector_shuffle, and then extract a subvector of the right size - e.g. shuffle the two <8 x i8> inputs into a <16 x i8> and then extract a <12 x i8>. This also includes a target-specific X86 combine that in the presence of AVX2 combines: (vector_shuffle <mask> (concat_vectors t1, undef) (concat_vectors t2, undef)) into: (vector_shuffle <mask> (concat_vectors t1, t2), undef) in cases where this allows us to form VPERMD/VPERMQ. (This is not a separate commit, as that pattern does not appear without the DAGBuilder change.) llvm-svn: 280418
2016-09-02 05:32:09 +08:00
; SSE2-NEXT: movaps %xmm0, %xmm2
; SSE2-NEXT: shufps {{.*#+}} xmm2 = xmm2[0,1],xmm1[1,2]
; SSE2-NEXT: shufps {{.*#+}} xmm2 = xmm2[0,2,1,3]
; SSE2-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,1,2,3]
; SSE2-NEXT: movss %xmm0, 16(%rdi)
; SSE2-NEXT: movaps %xmm2, (%rdi)
; SSE2-NEXT: retq
;
; SSE42-LABEL: v5f32:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; SSE42: # %bb.0:
[SelectionDAG] Generate vector_shuffle nodes for undersized result vector sizes Prior to this, we could generate a vector_shuffle from an IR shuffle when the size of the result was exactly the sum of the sizes of the input vectors. If the output vector was narrower - e.g. a <12 x i8> being formed by a shuffle with two <8 x i8> inputs - we would lower the shuffle to a sequence of extracts and inserts. Instead, we can form a larger vector_shuffle, and then extract a subvector of the right size - e.g. shuffle the two <8 x i8> inputs into a <16 x i8> and then extract a <12 x i8>. This also includes a target-specific X86 combine that in the presence of AVX2 combines: (vector_shuffle <mask> (concat_vectors t1, undef) (concat_vectors t2, undef)) into: (vector_shuffle <mask> (concat_vectors t1, t2), undef) in cases where this allows us to form VPERMD/VPERMQ. (This is not a separate commit, as that pattern does not appear without the DAGBuilder change.) llvm-svn: 280418
2016-09-02 05:32:09 +08:00
; SSE42-NEXT: extractps $3, %xmm0, 16(%rdi)
; SSE42-NEXT: shufps {{.*#+}} xmm0 = xmm0[0,1],xmm1[1,2]
; SSE42-NEXT: shufps {{.*#+}} xmm0 = xmm0[0,2,1,3]
; SSE42-NEXT: movaps %xmm0, (%rdi)
; SSE42-NEXT: retq
;
[DAGCombiner] Add EXTRACT_SUBVECTOR to SimplifyDemandedVectorElts As discussed on PR37989, this patch adds EXTRACT_SUBVECTOR handling to TargetLowering::SimplifyDemandedVectorElts and calls it from DAGCombiner::visitEXTRACT_SUBVECTOR. Differential Revision: https://reviews.llvm.org/D48825 llvm-svn: 336490
2018-07-08 01:30:06 +08:00
; AVX-LABEL: v5f32:
; AVX: # %bb.0:
; AVX-NEXT: vshufps {{.*#+}} xmm1 = xmm0[0,1],xmm1[1,2]
; AVX-NEXT: vpermilps {{.*#+}} xmm1 = xmm1[0,2,1,3]
; AVX-NEXT: vextractps $3, %xmm0, 16(%rdi)
; AVX-NEXT: vmovaps %xmm1, (%rdi)
; AVX-NEXT: retq
[X86][XOP] Add XOP oddshuffles tests XOP codegen is often different to generic AVX - thank you vpperm! llvm-svn: 315176
2017-10-08 20:58:15 +08:00
;
; XOP-LABEL: v5f32:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; XOP: # %bb.0:
[X86][XOP] Add XOP oddshuffles tests XOP codegen is often different to generic AVX - thank you vpperm! llvm-svn: 315176
2017-10-08 20:58:15 +08:00
; XOP-NEXT: vshufps {{.*#+}} xmm1 = xmm0[0,1],xmm1[1,2]
; XOP-NEXT: vpermilps {{.*#+}} xmm1 = xmm1[0,2,1,3]
; XOP-NEXT: vextractps $3, %xmm0, 16(%rdi)
; XOP-NEXT: vmovaps %xmm1, (%rdi)
; XOP-NEXT: retq
[SelectionDAG] Generate vector_shuffle nodes for undersized result vector sizes Prior to this, we could generate a vector_shuffle from an IR shuffle when the size of the result was exactly the sum of the sizes of the input vectors. If the output vector was narrower - e.g. a <12 x i8> being formed by a shuffle with two <8 x i8> inputs - we would lower the shuffle to a sequence of extracts and inserts. Instead, we can form a larger vector_shuffle, and then extract a subvector of the right size - e.g. shuffle the two <8 x i8> inputs into a <16 x i8> and then extract a <12 x i8>. This also includes a target-specific X86 combine that in the presence of AVX2 combines: (vector_shuffle <mask> (concat_vectors t1, undef) (concat_vectors t2, undef)) into: (vector_shuffle <mask> (concat_vectors t1, t2), undef) in cases where this allows us to form VPERMD/VPERMQ. (This is not a separate commit, as that pattern does not appear without the DAGBuilder change.) llvm-svn: 280418
2016-09-02 05:32:09 +08:00
%r = shufflevector <4 x float> %a, <4 x float> %b, <5 x i32> <i32 0, i32 5, i32 1, i32 6, i32 3>
store <5 x float> %r, <5 x float>* %p
ret void
}
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
define void @v7i8(<4 x i8> %a, <4 x i8> %b, <7 x i8>* %p) nounwind {
[X86] Add baseline test for "odd" shuffles. NFC. Adds a baseline test for lowering shuffles where the width of the output vector is not twice the size of the input vectors. Many of those sequences are suboptimal, and will hopefully be improved in follow-up patches. llvm-svn: 279888
2016-08-27 08:10:24 +08:00
; SSE2-LABEL: v7i8:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/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:
Don't combine a shuffle of two BUILD_VECTORs with duplicate elements. Targets can't handle this case well in general; we often transform a shuffle of two cheap BUILD_VECTORs to element-by-element insertion, which is very inefficient. Fixes https://llvm.org/bugs/show_bug.cgi?id=31364 . Partially fixes https://llvm.org/bugs/show_bug.cgi?id=31301. Differential Revision: https://reviews.llvm.org/D27787 llvm-svn: 289874
2016-12-16 05:36:59 +08:00
; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,3,1,3]
; SSE2-NEXT: movdqa {{.*#+}} xmm2 = [0,65535,0,65535,0,65535,65535,65535]
; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm1[2,1,0,3]
; SSE2-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[0,0,2,0,4,5,6,7]
; SSE2-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,6,4,7]
; SSE2-NEXT: pand %xmm2, %xmm1
[X86] Add baseline test for "odd" shuffles. NFC. Adds a baseline test for lowering shuffles where the width of the output vector is not twice the size of the input vectors. Many of those sequences are suboptimal, and will hopefully be improved in follow-up patches. llvm-svn: 279888
2016-08-27 08:10:24 +08:00
; SSE2-NEXT: pandn %xmm0, %xmm2
Don't combine a shuffle of two BUILD_VECTORs with duplicate elements. Targets can't handle this case well in general; we often transform a shuffle of two cheap BUILD_VECTORs to element-by-element insertion, which is very inefficient. Fixes https://llvm.org/bugs/show_bug.cgi?id=31364 . Partially fixes https://llvm.org/bugs/show_bug.cgi?id=31301. Differential Revision: https://reviews.llvm.org/D27787 llvm-svn: 289874
2016-12-16 05:36:59 +08:00
; SSE2-NEXT: por %xmm1, %xmm2
; SSE2-NEXT: movdqa {{.*#+}} xmm0 = [255,255,255,255,255,255,255,255]
; SSE2-NEXT: pand %xmm2, %xmm0
; SSE2-NEXT: packuswb %xmm0, %xmm0
; SSE2-NEXT: movdqa %xmm2, -{{[0-9]+}}(%rsp)
[X86] Add baseline test for "odd" shuffles. NFC. Adds a baseline test for lowering shuffles where the width of the output vector is not twice the size of the input vectors. Many of those sequences are suboptimal, and will hopefully be improved in follow-up patches. llvm-svn: 279888
2016-08-27 08:10:24 +08:00
; SSE2-NEXT: movb -{{[0-9]+}}(%rsp), %al
; SSE2-NEXT: movb %al, 6(%rdi)
[X86][SSE] Attempt to extract vector elements through target shuffles DAGCombiner already supports peeking thorough shuffles to improve vector element extraction, but legalization often leaves us in situations where we need to extract vector elements after shuffles have already been lowered. This patch adds support for VECTOR_EXTRACT_ELEMENT/PEXTRW/PEXTRB instructions to attempt to handle target shuffles as well. I've covered some basic scenarios including handling shuffle mask scaling and the implicit zero-extension of PEXTRW/PEXTRB, there is more that could be done here (that I've mentioned in TODOs) but I haven't found many cases where its worth it. Differential Revision: https://reviews.llvm.org/D30176 llvm-svn: 296381
2017-02-28 05:01:57 +08:00
; SSE2-NEXT: movd %xmm0, (%rdi)
; SSE2-NEXT: pextrw $2, %xmm0, %eax
[X86] Add baseline test for "odd" shuffles. NFC. Adds a baseline test for lowering shuffles where the width of the output vector is not twice the size of the input vectors. Many of those sequences are suboptimal, and will hopefully be improved in follow-up patches. llvm-svn: 279888
2016-08-27 08:10:24 +08:00
; SSE2-NEXT: movw %ax, 4(%rdi)
; SSE2-NEXT: retq
;
; SSE42-LABEL: v7i8:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; SSE42: # %bb.0:
Don't combine a shuffle of two BUILD_VECTORs with duplicate elements. Targets can't handle this case well in general; we often transform a shuffle of two cheap BUILD_VECTORs to element-by-element insertion, which is very inefficient. Fixes https://llvm.org/bugs/show_bug.cgi?id=31364 . Partially fixes https://llvm.org/bugs/show_bug.cgi?id=31301. Differential Revision: https://reviews.llvm.org/D27787 llvm-svn: 289874
2016-12-16 05:36:59 +08:00
; SSE42-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,3,1,3]
[X86][SSE] Attempt to extract vector elements through target shuffles DAGCombiner already supports peeking thorough shuffles to improve vector element extraction, but legalization often leaves us in situations where we need to extract vector elements after shuffles have already been lowered. This patch adds support for VECTOR_EXTRACT_ELEMENT/PEXTRW/PEXTRB instructions to attempt to handle target shuffles as well. I've covered some basic scenarios including handling shuffle mask scaling and the implicit zero-extension of PEXTRW/PEXTRB, there is more that could be done here (that I've mentioned in TODOs) but I haven't found many cases where its worth it. Differential Revision: https://reviews.llvm.org/D30176 llvm-svn: 296381
2017-02-28 05:01:57 +08:00
; SSE42-NEXT: pextrb $0, %xmm1, 6(%rdi)
Don't combine a shuffle of two BUILD_VECTORs with duplicate elements. Targets can't handle this case well in general; we often transform a shuffle of two cheap BUILD_VECTORs to element-by-element insertion, which is very inefficient. Fixes https://llvm.org/bugs/show_bug.cgi?id=31364 . Partially fixes https://llvm.org/bugs/show_bug.cgi?id=31301. Differential Revision: https://reviews.llvm.org/D27787 llvm-svn: 289874
2016-12-16 05:36:59 +08:00
; SSE42-NEXT: pshufb {{.*#+}} xmm1 = xmm1[8,9,8,9,4,5,8,9,0,1,12,13,0,1,14,15]
; SSE42-NEXT: pblendw {{.*#+}} xmm1 = xmm0[0],xmm1[1],xmm0[2],xmm1[3],xmm0[4],xmm1[5,6,7]
[TargetLowering] Add vector BITCAST support to SimplifyDemandedVectorElts Notably helps cleanup after legalization of vector types Differential Revision: https://reviews.llvm.org/D43674 llvm-svn: 326838
2018-03-07 06:32:01 +08:00
; SSE42-NEXT: pshufb {{.*#+}} xmm1 = xmm1[0,2,4,6,8,10,12,u,u,u,u,u,u,u,u,u]
[X86][SSE] Attempt to extract vector elements through target shuffles DAGCombiner already supports peeking thorough shuffles to improve vector element extraction, but legalization often leaves us in situations where we need to extract vector elements after shuffles have already been lowered. This patch adds support for VECTOR_EXTRACT_ELEMENT/PEXTRW/PEXTRB instructions to attempt to handle target shuffles as well. I've covered some basic scenarios including handling shuffle mask scaling and the implicit zero-extension of PEXTRW/PEXTRB, there is more that could be done here (that I've mentioned in TODOs) but I haven't found many cases where its worth it. Differential Revision: https://reviews.llvm.org/D30176 llvm-svn: 296381
2017-02-28 05:01:57 +08:00
; SSE42-NEXT: pextrw $2, %xmm1, 4(%rdi)
; SSE42-NEXT: movd %xmm1, (%rdi)
[X86] Add baseline test for "odd" shuffles. NFC. Adds a baseline test for lowering shuffles where the width of the output vector is not twice the size of the input vectors. Many of those sequences are suboptimal, and will hopefully be improved in follow-up patches. llvm-svn: 279888
2016-08-27 08:10:24 +08:00
; SSE42-NEXT: retq
;
; AVX-LABEL: v7i8:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/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:
Don't combine a shuffle of two BUILD_VECTORs with duplicate elements. Targets can't handle this case well in general; we often transform a shuffle of two cheap BUILD_VECTORs to element-by-element insertion, which is very inefficient. Fixes https://llvm.org/bugs/show_bug.cgi?id=31364 . Partially fixes https://llvm.org/bugs/show_bug.cgi?id=31301. Differential Revision: https://reviews.llvm.org/D27787 llvm-svn: 289874
2016-12-16 05:36:59 +08:00
; AVX-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,3,1,3]
[X86][SSE] Attempt to extract vector elements through target shuffles DAGCombiner already supports peeking thorough shuffles to improve vector element extraction, but legalization often leaves us in situations where we need to extract vector elements after shuffles have already been lowered. This patch adds support for VECTOR_EXTRACT_ELEMENT/PEXTRW/PEXTRB instructions to attempt to handle target shuffles as well. I've covered some basic scenarios including handling shuffle mask scaling and the implicit zero-extension of PEXTRW/PEXTRB, there is more that could be done here (that I've mentioned in TODOs) but I haven't found many cases where its worth it. Differential Revision: https://reviews.llvm.org/D30176 llvm-svn: 296381
2017-02-28 05:01:57 +08:00
; AVX-NEXT: vpshufb {{.*#+}} xmm2 = xmm1[8,9,8,9,4,5,8,9,0,1,12,13,0,1,14,15]
; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0],xmm2[1],xmm0[2],xmm2[3],xmm0[4],xmm2[5,6,7]
[TargetLowering] Add vector BITCAST support to SimplifyDemandedVectorElts Notably helps cleanup after legalization of vector types Differential Revision: https://reviews.llvm.org/D43674 llvm-svn: 326838
2018-03-07 06:32:01 +08:00
; AVX-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,2,4,6,8,10,12,u,u,u,u,u,u,u,u,u]
[X86][SSE] Attempt to extract vector elements through target shuffles DAGCombiner already supports peeking thorough shuffles to improve vector element extraction, but legalization often leaves us in situations where we need to extract vector elements after shuffles have already been lowered. This patch adds support for VECTOR_EXTRACT_ELEMENT/PEXTRW/PEXTRB instructions to attempt to handle target shuffles as well. I've covered some basic scenarios including handling shuffle mask scaling and the implicit zero-extension of PEXTRW/PEXTRB, there is more that could be done here (that I've mentioned in TODOs) but I haven't found many cases where its worth it. Differential Revision: https://reviews.llvm.org/D30176 llvm-svn: 296381
2017-02-28 05:01:57 +08:00
; AVX-NEXT: vpextrb $0, %xmm1, 6(%rdi)
; AVX-NEXT: vpextrw $2, %xmm0, 4(%rdi)
; AVX-NEXT: vmovd %xmm0, (%rdi)
[X86] Add baseline test for "odd" shuffles. NFC. Adds a baseline test for lowering shuffles where the width of the output vector is not twice the size of the input vectors. Many of those sequences are suboptimal, and will hopefully be improved in follow-up patches. llvm-svn: 279888
2016-08-27 08:10:24 +08:00
; AVX-NEXT: retq
[X86][XOP] Add XOP oddshuffles tests XOP codegen is often different to generic AVX - thank you vpperm! llvm-svn: 315176
2017-10-08 20:58:15 +08:00
;
; XOP-LABEL: v7i8:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; XOP: # %bb.0:
[TargetLowering] Add vector BITCAST support to SimplifyDemandedVectorElts Notably helps cleanup after legalization of vector types Differential Revision: https://reviews.llvm.org/D43674 llvm-svn: 326838
2018-03-07 06:32:01 +08:00
; XOP-NEXT: vpperm {{.*#+}} xmm0 = xmm0[0],xmm1[8],xmm0[12],xmm1[8],xmm0[4],xmm1[12,0,u,u,u,u,u,u,u,u,u]
[X86][XOP] Add XOP oddshuffles tests XOP codegen is often different to generic AVX - thank you vpperm! llvm-svn: 315176
2017-10-08 20:58:15 +08:00
; XOP-NEXT: vpextrb $0, %xmm1, 6(%rdi)
; XOP-NEXT: vpextrw $2, %xmm0, 4(%rdi)
; XOP-NEXT: vmovd %xmm0, (%rdi)
; XOP-NEXT: retq
[X86] Add baseline test for "odd" shuffles. NFC. Adds a baseline test for lowering shuffles where the width of the output vector is not twice the size of the input vectors. Many of those sequences are suboptimal, and will hopefully be improved in follow-up patches. llvm-svn: 279888
2016-08-27 08:10:24 +08:00
%r = shufflevector <4 x i8> %a, <4 x i8> %b, <7 x i32> <i32 0, i32 6, i32 3, i32 6, i32 1, i32 7, i32 4>
store <7 x i8> %r, <7 x i8>* %p
ret void
}
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
define void @v7i16(<4 x i16> %a, <4 x i16> %b, <7 x i16>* %p) nounwind {
[X86] Add baseline test for "odd" shuffles. NFC. Adds a baseline test for lowering shuffles where the width of the output vector is not twice the size of the input vectors. Many of those sequences are suboptimal, and will hopefully be improved in follow-up patches. llvm-svn: 279888
2016-08-27 08:10:24 +08:00
; SSE2-LABEL: v7i16:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/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] Add baseline test for "odd" shuffles. NFC. Adds a baseline test for lowering shuffles where the width of the output vector is not twice the size of the input vectors. Many of those sequences are suboptimal, and will hopefully be improved in follow-up patches. llvm-svn: 279888
2016-08-27 08:10:24 +08:00
; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,3,1,3]
; SSE2-NEXT: movdqa {{.*#+}} xmm2 = [0,65535,0,65535,0,65535,65535,65535]
; SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm1[2,1,0,3]
; SSE2-NEXT: pshuflw {{.*#+}} xmm3 = xmm3[0,0,2,0,4,5,6,7]
; SSE2-NEXT: pshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,6,4,7]
; SSE2-NEXT: pand %xmm2, %xmm3
; SSE2-NEXT: pandn %xmm0, %xmm2
; SSE2-NEXT: por %xmm3, %xmm2
; SSE2-NEXT: movd %xmm1, %eax
; SSE2-NEXT: movw %ax, 12(%rdi)
; SSE2-NEXT: movq %xmm2, (%rdi)
; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm2[2,3,0,1]
; SSE2-NEXT: movd %xmm0, 8(%rdi)
; SSE2-NEXT: retq
;
; SSE42-LABEL: v7i16:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; SSE42: # %bb.0:
[X86] Add baseline test for "odd" shuffles. NFC. Adds a baseline test for lowering shuffles where the width of the output vector is not twice the size of the input vectors. Many of those sequences are suboptimal, and will hopefully be improved in follow-up patches. llvm-svn: 279888
2016-08-27 08:10:24 +08:00
; SSE42-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,3,1,3]
; SSE42-NEXT: pextrw $0, %xmm1, 12(%rdi)
; SSE42-NEXT: pshufb {{.*#+}} xmm1 = xmm1[8,9,8,9,4,5,8,9,0,1,12,13,0,1,14,15]
; SSE42-NEXT: pblendw {{.*#+}} xmm1 = xmm0[0],xmm1[1],xmm0[2],xmm1[3],xmm0[4],xmm1[5,6,7]
; SSE42-NEXT: pextrd $2, %xmm1, 8(%rdi)
; SSE42-NEXT: movq %xmm1, (%rdi)
; SSE42-NEXT: retq
;
; AVX-LABEL: v7i16:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/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] Add baseline test for "odd" shuffles. NFC. Adds a baseline test for lowering shuffles where the width of the output vector is not twice the size of the input vectors. Many of those sequences are suboptimal, and will hopefully be improved in follow-up patches. llvm-svn: 279888
2016-08-27 08:10:24 +08:00
; AVX-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,3,1,3]
; AVX-NEXT: vpshufb {{.*#+}} xmm2 = xmm1[8,9,8,9,4,5,8,9,0,1,12,13,0,1,14,15]
; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0],xmm2[1],xmm0[2],xmm2[3],xmm0[4],xmm2[5,6,7]
; AVX-NEXT: vpextrw $0, %xmm1, 12(%rdi)
; AVX-NEXT: vpextrd $2, %xmm0, 8(%rdi)
; AVX-NEXT: vmovq %xmm0, (%rdi)
; AVX-NEXT: retq
[X86][XOP] Add XOP oddshuffles tests XOP codegen is often different to generic AVX - thank you vpperm! llvm-svn: 315176
2017-10-08 20:58:15 +08:00
;
; XOP-LABEL: v7i16:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; XOP: # %bb.0:
[X86][XOP] Add XOP oddshuffles tests XOP codegen is often different to generic AVX - thank you vpperm! llvm-svn: 315176
2017-10-08 20:58:15 +08:00
; XOP-NEXT: vpperm {{.*#+}} xmm0 = xmm0[0,1],xmm1[8,9],xmm0[12,13],xmm1[8,9],xmm0[4,5],xmm1[12,13,0,1,14,15]
; XOP-NEXT: vpextrw $0, %xmm1, 12(%rdi)
; XOP-NEXT: vpextrd $2, %xmm0, 8(%rdi)
; XOP-NEXT: vmovq %xmm0, (%rdi)
; XOP-NEXT: retq
[X86] Add baseline test for "odd" shuffles. NFC. Adds a baseline test for lowering shuffles where the width of the output vector is not twice the size of the input vectors. Many of those sequences are suboptimal, and will hopefully be improved in follow-up patches. llvm-svn: 279888
2016-08-27 08:10:24 +08:00
%r = shufflevector <4 x i16> %a, <4 x i16> %b, <7 x i32> <i32 0, i32 6, i32 3, i32 6, i32 1, i32 7, i32 4>
store <7 x i16> %r, <7 x i16>* %p
ret void
}
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
define void @v7i32(<4 x i32> %a, <4 x i32> %b, <7 x i32>* %p) nounwind {
[X86] Add baseline test for "odd" shuffles. NFC. Adds a baseline test for lowering shuffles where the width of the output vector is not twice the size of the input vectors. Many of those sequences are suboptimal, and will hopefully be improved in follow-up patches. llvm-svn: 279888
2016-08-27 08:10:24 +08:00
; SSE2-LABEL: v7i32:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/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] Add baseline test for "odd" shuffles. NFC. Adds a baseline test for lowering shuffles where the width of the output vector is not twice the size of the input vectors. Many of those sequences are suboptimal, and will hopefully be improved in follow-up patches. llvm-svn: 279888
2016-08-27 08:10:24 +08:00
; SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm1[0,1,2,2]
; SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm0[0,1,0,3]
; SSE2-NEXT: punpckhdq {{.*#+}} xmm3 = xmm3[2],xmm2[2],xmm3[3],xmm2[3]
[DAGCombiner] Call SimplifyDemandedVectorElts from EXTRACT_VECTOR_ELT If we are only extracting vector elements via EXTRACT_VECTOR_ELT(s) we may be able to use SimplifyDemandedVectorElts to avoid unnecessary vector ops. Differential Revision: https://reviews.llvm.org/D49262 llvm-svn: 337258
2018-07-17 17:45:35 +08:00
; SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm1[3,1,2,3]
; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,2,3]
; SSE2-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]
[X86] Add baseline test for "odd" shuffles. NFC. Adds a baseline test for lowering shuffles where the width of the output vector is not twice the size of the input vectors. Many of those sequences are suboptimal, and will hopefully be improved in follow-up patches. llvm-svn: 279888
2016-08-27 08:10:24 +08:00
; SSE2-NEXT: movd %xmm1, 24(%rdi)
[DAGCombiner] Call SimplifyDemandedVectorElts from EXTRACT_VECTOR_ELT If we are only extracting vector elements via EXTRACT_VECTOR_ELT(s) we may be able to use SimplifyDemandedVectorElts to avoid unnecessary vector ops. Differential Revision: https://reviews.llvm.org/D49262 llvm-svn: 337258
2018-07-17 17:45:35 +08:00
; SSE2-NEXT: movq %xmm0, 16(%rdi)
[X86] Add baseline test for "odd" shuffles. NFC. Adds a baseline test for lowering shuffles where the width of the output vector is not twice the size of the input vectors. Many of those sequences are suboptimal, and will hopefully be improved in follow-up patches. llvm-svn: 279888
2016-08-27 08:10:24 +08:00
; SSE2-NEXT: movdqa %xmm3, (%rdi)
; SSE2-NEXT: retq
;
; SSE42-LABEL: v7i32:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; SSE42: # %bb.0:
[DAGCombiner] Call SimplifyDemandedVectorElts from EXTRACT_VECTOR_ELT If we are only extracting vector elements via EXTRACT_VECTOR_ELT(s) we may be able to use SimplifyDemandedVectorElts to avoid unnecessary vector ops. Differential Revision: https://reviews.llvm.org/D49262 llvm-svn: 337258
2018-07-17 17:45:35 +08:00
; SSE42-NEXT: movdqa %xmm0, %xmm2
; SSE42-NEXT: pblendw {{.*#+}} xmm2 = xmm2[0,1,2,3],xmm1[4,5],xmm2[6,7]
; SSE42-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,2,3,2]
; SSE42-NEXT: pblendw {{.*#+}} xmm0 = xmm0[0,1,2,3],xmm1[4,5,6,7]
; SSE42-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,3,2,3]
[X86] Add baseline test for "odd" shuffles. NFC. Adds a baseline test for lowering shuffles where the width of the output vector is not twice the size of the input vectors. Many of those sequences are suboptimal, and will hopefully be improved in follow-up patches. llvm-svn: 279888
2016-08-27 08:10:24 +08:00
; SSE42-NEXT: movd %xmm1, 24(%rdi)
[DAGCombiner] Call SimplifyDemandedVectorElts from EXTRACT_VECTOR_ELT If we are only extracting vector elements via EXTRACT_VECTOR_ELT(s) we may be able to use SimplifyDemandedVectorElts to avoid unnecessary vector ops. Differential Revision: https://reviews.llvm.org/D49262 llvm-svn: 337258
2018-07-17 17:45:35 +08:00
; SSE42-NEXT: movq %xmm0, 16(%rdi)
; SSE42-NEXT: movdqa %xmm2, (%rdi)
[X86] Add baseline test for "odd" shuffles. NFC. Adds a baseline test for lowering shuffles where the width of the output vector is not twice the size of the input vectors. Many of those sequences are suboptimal, and will hopefully be improved in follow-up patches. llvm-svn: 279888
2016-08-27 08:10:24 +08:00
; SSE42-NEXT: retq
;
[X86][SSE] Regenerate odd shuffle tests with common prefixes llvm-svn: 280661
2016-09-05 22:15:38 +08:00
; AVX1-LABEL: v7i32:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/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] Regenerate odd shuffle tests with common prefixes llvm-svn: 280661
2016-09-05 22:15:38 +08:00
; AVX1-NEXT: vblendps {{.*#+}} xmm2 = xmm0[0,1],xmm1[2],xmm0[3]
; AVX1-NEXT: vpermilps {{.*#+}} xmm2 = xmm2[0,2,3,2]
; AVX1-NEXT: vblendps {{.*#+}} xmm0 = xmm1[0],xmm0[1],xmm1[2,3]
; AVX1-NEXT: vpermilps {{.*#+}} xmm0 = xmm0[1,3,0,3]
[X86][SSE] Fix domains for scalar store instructions As discussed on D27692 llvm-svn: 289834
2016-12-16 01:09:24 +08:00
; AVX1-NEXT: vmovss %xmm1, 24(%rdi)
[X86][SSE] Regenerate odd shuffle tests with common prefixes llvm-svn: 280661
2016-09-05 22:15:38 +08:00
; AVX1-NEXT: vmovlps %xmm0, 16(%rdi)
; AVX1-NEXT: vmovaps %xmm2, (%rdi)
; AVX1-NEXT: retq
[X86] Add baseline test for "odd" shuffles. NFC. Adds a baseline test for lowering shuffles where the width of the output vector is not twice the size of the input vectors. Many of those sequences are suboptimal, and will hopefully be improved in follow-up patches. llvm-svn: 279888
2016-08-27 08:10:24 +08:00
;
; AVX2-LABEL: v7i32:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/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:
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
; AVX2-NEXT: # kill: def $xmm0 killed $xmm0 def $ymm0
[X86] Add VPERMPD/VPERMQ and VPERMPS/VPERMD to the execution domain fixing table. llvm-svn: 313610
2017-09-19 12:39:55 +08:00
; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
; AVX2-NEXT: vmovaps {{.*#+}} ymm2 = <0,6,3,6,1,7,4,u>
; AVX2-NEXT: vpermps %ymm0, %ymm2, %ymm0
; AVX2-NEXT: vmovss %xmm1, 24(%rdi)
; AVX2-NEXT: vextractf128 $1, %ymm0, %xmm1
; AVX2-NEXT: vmovlps %xmm1, 16(%rdi)
; AVX2-NEXT: vmovaps %xmm0, (%rdi)
[X86] Add baseline test for "odd" shuffles. NFC. Adds a baseline test for lowering shuffles where the width of the output vector is not twice the size of the input vectors. Many of those sequences are suboptimal, and will hopefully be improved in follow-up patches. llvm-svn: 279888
2016-08-27 08:10:24 +08:00
; AVX2-NEXT: vzeroupper
; AVX2-NEXT: retq
[X86][XOP] Add XOP oddshuffles tests XOP codegen is often different to generic AVX - thank you vpperm! llvm-svn: 315176
2017-10-08 20:58:15 +08:00
;
; XOP-LABEL: v7i32:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; XOP: # %bb.0:
[X86][XOP] Add XOP oddshuffles tests XOP codegen is often different to generic AVX - thank you vpperm! llvm-svn: 315176
2017-10-08 20:58:15 +08:00
; XOP-NEXT: vblendps {{.*#+}} xmm2 = xmm0[0,1],xmm1[2],xmm0[3]
; XOP-NEXT: vpermilps {{.*#+}} xmm2 = xmm2[0,2,3,2]
; XOP-NEXT: vblendps {{.*#+}} xmm0 = xmm1[0],xmm0[1],xmm1[2,3]
; XOP-NEXT: vpermilps {{.*#+}} xmm0 = xmm0[1,3,0,3]
; XOP-NEXT: vmovss %xmm1, 24(%rdi)
; XOP-NEXT: vmovlps %xmm0, 16(%rdi)
; XOP-NEXT: vmovaps %xmm2, (%rdi)
; XOP-NEXT: retq
[X86] Add baseline test for "odd" shuffles. NFC. Adds a baseline test for lowering shuffles where the width of the output vector is not twice the size of the input vectors. Many of those sequences are suboptimal, and will hopefully be improved in follow-up patches. llvm-svn: 279888
2016-08-27 08:10:24 +08:00
%r = shufflevector <4 x i32> %a, <4 x i32> %b, <7 x i32> <i32 0, i32 6, i32 3, i32 6, i32 1, i32 7, i32 4>
store <7 x i32> %r, <7 x i32>* %p
ret void
}
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
define void @v12i8(<8 x i8> %a, <8 x i8> %b, <12 x i8>* %p) nounwind {
[X86] Add baseline test for "odd" shuffles. NFC. Adds a baseline test for lowering shuffles where the width of the output vector is not twice the size of the input vectors. Many of those sequences are suboptimal, and will hopefully be improved in follow-up patches. llvm-svn: 279888
2016-08-27 08:10:24 +08:00
; SSE2-LABEL: v12i8:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/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] Add baseline test for "odd" shuffles. NFC. Adds a baseline test for lowering shuffles where the width of the output vector is not twice the size of the input vectors. Many of those sequences are suboptimal, and will hopefully be improved in follow-up patches. llvm-svn: 279888
2016-08-27 08:10:24 +08:00
; SSE2-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
; SSE2-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[0,0,2,3,4,5,6,7]
; SSE2-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,6,6,7]
; SSE2-NEXT: movdqa {{.*#+}} xmm2 = [255,255,0,255,255,0,255,255,0,255,255,0,255,255,255,255]
; SSE2-NEXT: pand {{.*}}(%rip), %xmm0
; SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm0[3,1,2,3]
; SSE2-NEXT: pshuflw {{.*#+}} xmm3 = xmm3[0,3,1,3,4,5,6,7]
; SSE2-NEXT: pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,6,5,4,7]
; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,3,2,1]
; SSE2-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[0,2,2,1,4,5,6,7]
; SSE2-NEXT: pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,5,5,6,4]
; SSE2-NEXT: packuswb %xmm3, %xmm0
; SSE2-NEXT: pand %xmm2, %xmm0
; SSE2-NEXT: pandn %xmm1, %xmm2
; SSE2-NEXT: por %xmm0, %xmm2
; SSE2-NEXT: movq %xmm2, (%rdi)
; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm2[2,3,0,1]
; SSE2-NEXT: movd %xmm0, 8(%rdi)
; SSE2-NEXT: retq
;
; SSE42-LABEL: v12i8:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; SSE42: # %bb.0:
[X86] Add baseline test for "odd" shuffles. NFC. Adds a baseline test for lowering shuffles where the width of the output vector is not twice the size of the input vectors. Many of those sequences are suboptimal, and will hopefully be improved in follow-up patches. llvm-svn: 279888
2016-08-27 08:10:24 +08:00
; SSE42-NEXT: pshufb {{.*#+}} xmm1 = zero,zero,xmm1[0],zero,zero,xmm1[2],zero,zero,xmm1[4],zero,zero,xmm1[6,u,u,u,u]
; SSE42-NEXT: pshufb {{.*#+}} xmm0 = xmm0[0,8],zero,xmm0[2,10],zero,xmm0[4,12],zero,xmm0[6,14],zero,xmm0[u,u,u,u]
; SSE42-NEXT: por %xmm1, %xmm0
; SSE42-NEXT: pextrd $2, %xmm0, 8(%rdi)
; SSE42-NEXT: movq %xmm0, (%rdi)
; SSE42-NEXT: retq
;
; AVX-LABEL: v12i8:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/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] Add baseline test for "odd" shuffles. NFC. Adds a baseline test for lowering shuffles where the width of the output vector is not twice the size of the input vectors. Many of those sequences are suboptimal, and will hopefully be improved in follow-up patches. llvm-svn: 279888
2016-08-27 08:10:24 +08:00
; AVX-NEXT: vpshufb {{.*#+}} xmm1 = zero,zero,xmm1[0],zero,zero,xmm1[2],zero,zero,xmm1[4],zero,zero,xmm1[6,u,u,u,u]
; AVX-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,8],zero,xmm0[2,10],zero,xmm0[4,12],zero,xmm0[6,14],zero,xmm0[u,u,u,u]
; AVX-NEXT: vpor %xmm1, %xmm0, %xmm0
; AVX-NEXT: vpextrd $2, %xmm0, 8(%rdi)
; AVX-NEXT: vmovq %xmm0, (%rdi)
; AVX-NEXT: retq
[X86][XOP] Add XOP oddshuffles tests XOP codegen is often different to generic AVX - thank you vpperm! llvm-svn: 315176
2017-10-08 20:58:15 +08:00
;
; XOP-LABEL: v12i8:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; XOP: # %bb.0:
[X86][XOP] Add XOP oddshuffles tests XOP codegen is often different to generic AVX - thank you vpperm! llvm-svn: 315176
2017-10-08 20:58:15 +08:00
; XOP-NEXT: vpshufb {{.*#+}} xmm1 = zero,zero,xmm1[0],zero,zero,xmm1[2],zero,zero,xmm1[4],zero,zero,xmm1[6,u,u,u,u]
; XOP-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,8],zero,xmm0[2,10],zero,xmm0[4,12],zero,xmm0[6,14],zero,xmm0[u,u,u,u]
; XOP-NEXT: vpor %xmm1, %xmm0, %xmm0
; XOP-NEXT: vpextrd $2, %xmm0, 8(%rdi)
; XOP-NEXT: vmovq %xmm0, (%rdi)
; XOP-NEXT: retq
[X86] Add baseline test for "odd" shuffles. NFC. Adds a baseline test for lowering shuffles where the width of the output vector is not twice the size of the input vectors. Many of those sequences are suboptimal, and will hopefully be improved in follow-up patches. llvm-svn: 279888
2016-08-27 08:10:24 +08:00
%r = shufflevector <8 x i8> %a, <8 x i8> %b, <12 x i32> <i32 0, i32 4, i32 8, i32 1, i32 5, i32 9, i32 2, i32 6, i32 10, i32 3, i32 7, i32 11>
store <12 x i8> %r, <12 x i8>* %p
ret void
}
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
define void @v12i16(<8 x i16> %a, <8 x i16> %b, <12 x i16>* %p) nounwind {
[X86] Add baseline test for "odd" shuffles. NFC. Adds a baseline test for lowering shuffles where the width of the output vector is not twice the size of the input vectors. Many of those sequences are suboptimal, and will hopefully be improved in follow-up patches. llvm-svn: 279888
2016-08-27 08:10:24 +08:00
; SSE2-LABEL: v12i16:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/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] Add baseline test for "odd" shuffles. NFC. Adds a baseline test for lowering shuffles where the width of the output vector is not twice the size of the input vectors. Many of those sequences are suboptimal, and will hopefully be improved in follow-up patches. llvm-svn: 279888
2016-08-27 08:10:24 +08:00
; SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm1[0,0,0,3]
; SSE2-NEXT: movdqa {{.*#+}} xmm3 = [65535,65535,0,65535,65535,0,65535,65535]
; SSE2-NEXT: pshufhw {{.*#+}} xmm4 = xmm0[0,1,2,3,6,5,4,7]
; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm4[0,3,2,1]
; SSE2-NEXT: pshuflw {{.*#+}} xmm4 = xmm4[0,2,2,1,4,5,6,7]
; SSE2-NEXT: pshufhw {{.*#+}} xmm4 = xmm4[0,1,2,3,5,5,6,4]
; SSE2-NEXT: pand %xmm3, %xmm4
; SSE2-NEXT: pandn %xmm2, %xmm3
; SSE2-NEXT: por %xmm4, %xmm3
; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,2,3]
; SSE2-NEXT: movdqa {{.*#+}} xmm2 = [65535,0,0,65535,65535,65535,65535,65535]
; SSE2-NEXT: pand %xmm2, %xmm1
; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[3,1,2,3]
; SSE2-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[0,3,1,3,4,5,6,7]
; SSE2-NEXT: pandn %xmm0, %xmm2
; SSE2-NEXT: por %xmm1, %xmm2
; SSE2-NEXT: movq %xmm2, 16(%rdi)
; SSE2-NEXT: movdqa %xmm3, (%rdi)
; SSE2-NEXT: retq
;
; SSE42-LABEL: v12i16:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; SSE42: # %bb.0:
[X86] Add baseline test for "odd" shuffles. NFC. Adds a baseline test for lowering shuffles where the width of the output vector is not twice the size of the input vectors. Many of those sequences are suboptimal, and will hopefully be improved in follow-up patches. llvm-svn: 279888
2016-08-27 08:10:24 +08:00
; SSE42-NEXT: pshufd {{.*#+}} xmm2 = xmm1[1,1,2,3]
; SSE42-NEXT: pshufd {{.*#+}} xmm3 = xmm0[3,1,2,3]
; SSE42-NEXT: pshuflw {{.*#+}} xmm3 = xmm3[0,3,1,3,4,5,6,7]
; SSE42-NEXT: pblendw {{.*#+}} xmm3 = xmm2[0],xmm3[1,2],xmm2[3,4,5,6,7]
; SSE42-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,0,0,3]
; SSE42-NEXT: pshufb {{.*#+}} xmm0 = xmm0[0,1,8,9,8,9,2,3,10,11,10,11,4,5,12,13]
; SSE42-NEXT: pblendw {{.*#+}} xmm0 = xmm0[0,1],xmm1[2],xmm0[3,4],xmm1[5],xmm0[6,7]
; SSE42-NEXT: movdqa %xmm0, (%rdi)
; SSE42-NEXT: movq %xmm3, 16(%rdi)
; SSE42-NEXT: retq
;
[X86][SSE] Regenerate odd shuffle tests with common prefixes llvm-svn: 280661
2016-09-05 22:15:38 +08:00
; AVX1-LABEL: v12i16:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/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] Regenerate odd shuffle tests with common prefixes llvm-svn: 280661
2016-09-05 22:15:38 +08:00
; AVX1-NEXT: vpshufd {{.*#+}} xmm2 = xmm1[1,1,2,3]
; AVX1-NEXT: vpshufd {{.*#+}} xmm3 = xmm0[3,1,2,3]
; AVX1-NEXT: vpshuflw {{.*#+}} xmm3 = xmm3[0,3,1,3,4,5,6,7]
; AVX1-NEXT: vpblendw {{.*#+}} xmm2 = xmm2[0],xmm3[1,2],xmm2[3],xmm3[4,5,6,7]
; AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[0,0,0,3]
; AVX1-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,1,8,9,8,9,2,3,10,11,10,11,4,5,12,13]
; AVX1-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0,1],xmm1[2],xmm0[3,4],xmm1[5],xmm0[6,7]
; AVX1-NEXT: vmovdqa %xmm0, (%rdi)
; AVX1-NEXT: vmovq %xmm2, 16(%rdi)
; AVX1-NEXT: retq
[X86] Add baseline test for "odd" shuffles. NFC. Adds a baseline test for lowering shuffles where the width of the output vector is not twice the size of the input vectors. Many of those sequences are suboptimal, and will hopefully be improved in follow-up patches. llvm-svn: 279888
2016-08-27 08:10:24 +08:00
;
X86 Tests: Update more isel tests with FastVariableShuffle feature Summary: Added the FastVariableShuffle feature to cases that resembled processors for which this fearure is on. For AVX2 there are processors with and w/o this fearue enable. For AVX512 only KNL does enable this feature so cases which only have +avx512f were left without the FastVariableShuffle enabled. Reviewers: RKSimon, craig.topper Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D41851 llvm-svn: 322090
2018-01-10 00:26:06 +08:00
; AVX2-SLOW-LABEL: v12i16:
; AVX2-SLOW: # %bb.0:
; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm2 = xmm1[1,1,2,3]
; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm3 = xmm0[3,1,2,3]
; AVX2-SLOW-NEXT: vpshuflw {{.*#+}} xmm3 = xmm3[0,3,1,3,4,5,6,7]
; AVX2-SLOW-NEXT: vpblendw {{.*#+}} xmm2 = xmm2[0],xmm3[1,2],xmm2[3],xmm3[4,5,6,7]
; AVX2-SLOW-NEXT: vpbroadcastd %xmm1, %xmm1
; AVX2-SLOW-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,1,8,9,8,9,2,3,10,11,10,11,4,5,12,13]
; AVX2-SLOW-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0,1],xmm1[2],xmm0[3,4],xmm1[5],xmm0[6,7]
; AVX2-SLOW-NEXT: vmovdqa %xmm0, (%rdi)
; AVX2-SLOW-NEXT: vmovq %xmm2, 16(%rdi)
; AVX2-SLOW-NEXT: retq
;
; AVX2-FAST-LABEL: v12i16:
; AVX2-FAST: # %bb.0:
; AVX2-FAST-NEXT: vpbroadcastd %xmm1, %xmm2
; AVX2-FAST-NEXT: vpshufb {{.*#+}} xmm3 = xmm0[0,1,8,9,8,9,2,3,10,11,10,11,4,5,12,13]
; AVX2-FAST-NEXT: vpblendw {{.*#+}} xmm2 = xmm3[0,1],xmm2[2],xmm3[3,4],xmm2[5],xmm3[6,7]
; AVX2-FAST-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[1,1,2,3]
; AVX2-FAST-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[12,13,6,7,14,15,6,7,8,9,10,11,12,13,14,15]
; AVX2-FAST-NEXT: vpblendw {{.*#+}} xmm0 = xmm1[0],xmm0[1,2],xmm1[3],xmm0[4,5,6,7]
; AVX2-FAST-NEXT: vmovq %xmm0, 16(%rdi)
; AVX2-FAST-NEXT: vmovdqa %xmm2, (%rdi)
; AVX2-FAST-NEXT: retq
[X86][XOP] Add XOP oddshuffles tests XOP codegen is often different to generic AVX - thank you vpperm! llvm-svn: 315176
2017-10-08 20:58:15 +08:00
;
; XOP-LABEL: v12i16:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; XOP: # %bb.0:
[X86][XOP] Add XOP oddshuffles tests XOP codegen is often different to generic AVX - thank you vpperm! llvm-svn: 315176
2017-10-08 20:58:15 +08:00
; XOP-NEXT: vpperm {{.*#+}} xmm2 = xmm0[0,1,8,9],xmm1[0,1],xmm0[2,3,10,11],xmm1[2,3],xmm0[4,5,12,13]
; XOP-NEXT: vpperm {{.*#+}} xmm0 = xmm1[4,5],xmm0[6,7,14,15],xmm1[6,7],xmm0[8,9,10,11,12,13,14,15]
; XOP-NEXT: vmovq %xmm0, 16(%rdi)
; XOP-NEXT: vmovdqa %xmm2, (%rdi)
; XOP-NEXT: retq
[X86] Add baseline test for "odd" shuffles. NFC. Adds a baseline test for lowering shuffles where the width of the output vector is not twice the size of the input vectors. Many of those sequences are suboptimal, and will hopefully be improved in follow-up patches. llvm-svn: 279888
2016-08-27 08:10:24 +08:00
%r = shufflevector <8 x i16> %a, <8 x i16> %b, <12 x i32> <i32 0, i32 4, i32 8, i32 1, i32 5, i32 9, i32 2, i32 6, i32 10, i32 3, i32 7, i32 11>
store <12 x i16> %r, <12 x i16>* %p
ret void
}
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
define void @v12i32(<8 x i32> %a, <8 x i32> %b, <12 x i32>* %p) nounwind {
[X86] Add baseline test for "odd" shuffles. NFC. Adds a baseline test for lowering shuffles where the width of the output vector is not twice the size of the input vectors. Many of those sequences are suboptimal, and will hopefully be improved in follow-up patches. llvm-svn: 279888
2016-08-27 08:10:24 +08:00
; SSE2-LABEL: v12i32:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/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] Combine (some) target shuffles with multiple uses As discussed on D41794, we have many cases where we fail to combine shuffles as the input operands have other uses. This patch permits these shuffles to be combined as long as they don't introduce additional variable shuffle masks, which should reduce instruction dependencies and allow the total number of shuffles to still drop without increasing the constant pool. However, this may mean that some memory folds may no longer occur, and on pre-AVX require the occasional extra register move. This also exposes some poor PMULDQ/PMULUDQ codegen which was doing unnecessary upper/lower calculations which will in fact fold to zero/undef - the fix will be added in a followup commit. Differential Revision: https://reviews.llvm.org/D50328 llvm-svn: 339335
2018-08-09 20:30:02 +08:00
; SSE2-NEXT: movaps %xmm2, %xmm3
; SSE2-NEXT: shufps {{.*#+}} xmm3 = xmm3[0,0],xmm0[1,0]
; SSE2-NEXT: movaps %xmm0, %xmm4
; SSE2-NEXT: unpcklps {{.*#+}} xmm4 = xmm4[0],xmm1[0],xmm4[1],xmm1[1]
; SSE2-NEXT: shufps {{.*#+}} xmm4 = xmm4[0,1],xmm3[0,2]
; SSE2-NEXT: movaps %xmm0, %xmm3
; SSE2-NEXT: shufps {{.*#+}} xmm3 = xmm3[2,0],xmm1[2,1]
; SSE2-NEXT: movaps %xmm2, %xmm5
; SSE2-NEXT: shufps {{.*#+}} xmm5 = xmm5[1,0],xmm1[1,0]
; SSE2-NEXT: shufps {{.*#+}} xmm5 = xmm5[2,0],xmm3[0,2]
; SSE2-NEXT: shufps {{.*#+}} xmm1 = xmm1[3,2],xmm2[3,2]
[X86][SSE] Allow float domain crossing if we are merging 2 or more shuffles and the root started as a float domain shuffle llvm-svn: 325349
2018-02-16 22:57:25 +08:00
; SSE2-NEXT: unpckhps {{.*#+}} xmm2 = xmm2[2],xmm0[2],xmm2[3],xmm0[3]
[X86][SSE] Combine (some) target shuffles with multiple uses As discussed on D41794, we have many cases where we fail to combine shuffles as the input operands have other uses. This patch permits these shuffles to be combined as long as they don't introduce additional variable shuffle masks, which should reduce instruction dependencies and allow the total number of shuffles to still drop without increasing the constant pool. However, this may mean that some memory folds may no longer occur, and on pre-AVX require the occasional extra register move. This also exposes some poor PMULDQ/PMULUDQ codegen which was doing unnecessary upper/lower calculations which will in fact fold to zero/undef - the fix will be added in a followup commit. Differential Revision: https://reviews.llvm.org/D50328 llvm-svn: 339335
2018-08-09 20:30:02 +08:00
; SSE2-NEXT: shufps {{.*#+}} xmm2 = xmm2[0,3],xmm1[0,2]
; SSE2-NEXT: movaps %xmm2, 32(%rdi)
; SSE2-NEXT: movaps %xmm5, 16(%rdi)
; SSE2-NEXT: movaps %xmm4, (%rdi)
[X86] Add baseline test for "odd" shuffles. NFC. Adds a baseline test for lowering shuffles where the width of the output vector is not twice the size of the input vectors. Many of those sequences are suboptimal, and will hopefully be improved in follow-up patches. llvm-svn: 279888
2016-08-27 08:10:24 +08:00
; SSE2-NEXT: retq
;
; SSE42-LABEL: v12i32:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; SSE42: # %bb.0:
[DAG] Generalize build_vector -> vector_shuffle combine for more than 2 inputs This generalizes the build_vector -> vector_shuffle combine to support any number of inputs. The idea is to create a binary tree of shuffles, where the first layer performs pairwise shuffles of the input vectors placing each input element into the correct lane, and the rest of the tree blends these shuffles together. This doesn't try to be smart and create any sort of "optimal" shuffles. The assumption is that even a "poor" shuffle sequence is better than extracting and inserting the elements one by one. Differential Revision: https://reviews.llvm.org/D24683 llvm-svn: 283480
2016-10-07 02:58:24 +08:00
; SSE42-NEXT: pshufd {{.*#+}} xmm3 = xmm1[0,0,1,1]
; SSE42-NEXT: pshufd {{.*#+}} xmm4 = xmm0[0,1,0,1]
; SSE42-NEXT: pblendw {{.*#+}} xmm4 = xmm4[0,1],xmm3[2,3],xmm4[4,5,6,7]
; SSE42-NEXT: pshufd {{.*#+}} xmm3 = xmm2[0,1,0,1]
; SSE42-NEXT: pblendw {{.*#+}} xmm3 = xmm4[0,1,2,3],xmm3[4,5],xmm4[6,7]
; SSE42-NEXT: pshufd {{.*#+}} xmm4 = xmm1[1,1,2,2]
; SSE42-NEXT: pblendw {{.*#+}} xmm4 = xmm4[0,1],xmm2[2,3],xmm4[4,5,6,7]
; SSE42-NEXT: pblendw {{.*#+}} xmm4 = xmm4[0,1,2,3],xmm0[4,5],xmm4[6,7]
; SSE42-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
; SSE42-NEXT: pshufd {{.*#+}} xmm2 = xmm2[2,3,2,3]
; SSE42-NEXT: pblendw {{.*#+}} xmm2 = xmm2[0,1],xmm0[2,3],xmm2[4,5,6,7]
; SSE42-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,2,3,3]
; SSE42-NEXT: pblendw {{.*#+}} xmm0 = xmm2[0,1,2,3],xmm0[4,5],xmm2[6,7]
; SSE42-NEXT: movdqa %xmm0, 32(%rdi)
; SSE42-NEXT: movdqa %xmm4, 16(%rdi)
; SSE42-NEXT: movdqa %xmm3, (%rdi)
[X86] Add baseline test for "odd" shuffles. NFC. Adds a baseline test for lowering shuffles where the width of the output vector is not twice the size of the input vectors. Many of those sequences are suboptimal, and will hopefully be improved in follow-up patches. llvm-svn: 279888
2016-08-27 08:10:24 +08:00
; SSE42-NEXT: retq
;
[X86][SSE] Regenerate odd shuffle tests with common prefixes llvm-svn: 280661
2016-09-05 22:15:38 +08:00
; AVX1-LABEL: v12i32:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/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] Regenerate odd shuffle tests with common prefixes llvm-svn: 280661
2016-09-05 22:15:38 +08:00
; AVX1-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm0[2,3,0,1]
; AVX1-NEXT: vmovsldup {{.*#+}} ymm2 = ymm2[0,0,2,2,4,4,6,6]
; AVX1-NEXT: vpermilps {{.*#+}} ymm3 = ymm0[0,u,u,1,5,u,u,6]
; AVX1-NEXT: vblendps {{.*#+}} ymm2 = ymm3[0],ymm2[1],ymm3[2,3,4,5],ymm2[6],ymm3[7]
; AVX1-NEXT: vmovddup {{.*#+}} xmm3 = xmm1[0,0]
; AVX1-NEXT: vinsertf128 $1, %xmm3, %ymm3, %ymm3
; AVX1-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1],ymm3[2],ymm2[3,4],ymm3[5],ymm2[6,7]
; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm3
; AVX1-NEXT: vshufps {{.*#+}} xmm0 = xmm0[0,3],xmm3[3,3]
; AVX1-NEXT: vpermilpd {{.*#+}} xmm1 = xmm1[1,1]
; AVX1-NEXT: vblendps {{.*#+}} xmm0 = xmm1[0],xmm0[1,2],xmm1[3]
[X86][SSE] Add custom execution domain fixing for BLENDPD/BLENDPS/PBLENDD/PBLENDW (PR34873) Add support for custom execution domain fixing and implement support for BLENDPD/BLENDPS/PBLENDD/PBLENDW. Differential Revision: https://reviews.llvm.org/D42042 llvm-svn: 322524
2018-01-16 06:18:45 +08:00
; AVX1-NEXT: vmovaps %xmm0, 32(%rdi)
[X86][SSE] Regenerate odd shuffle tests with common prefixes llvm-svn: 280661
2016-09-05 22:15:38 +08:00
; AVX1-NEXT: vmovaps %ymm2, (%rdi)
; AVX1-NEXT: vzeroupper
; AVX1-NEXT: retq
[X86] Add baseline test for "odd" shuffles. NFC. Adds a baseline test for lowering shuffles where the width of the output vector is not twice the size of the input vectors. Many of those sequences are suboptimal, and will hopefully be improved in follow-up patches. llvm-svn: 279888
2016-08-27 08:10:24 +08:00
;
X86 Tests: Update more isel tests with FastVariableShuffle feature Summary: Added the FastVariableShuffle feature to cases that resembled processors for which this fearure is on. For AVX2 there are processors with and w/o this fearue enable. For AVX512 only KNL does enable this feature so cases which only have +avx512f were left without the FastVariableShuffle enabled. Reviewers: RKSimon, craig.topper Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D41851 llvm-svn: 322090
2018-01-10 00:26:06 +08:00
; AVX2-SLOW-LABEL: v12i32:
; AVX2-SLOW: # %bb.0:
[x86] split more v8f32/v8i32 shuffles in lowering Similar to D57867 - this is a small patch with lots of test diffs. With half-vector-width narrowing potential, using an extract + 128-bit vshufps is a win because it replaces a 256-bit shuffle with a 128-bit shufle. This seems like it should be a win even for targets with 'fast-variable-shuffle', but we are intentionally deferring that to an independent change to make sure that is true. Differential Revision: https://reviews.llvm.org/D58181 llvm-svn: 354279
2019-02-19 00:46:12 +08:00
; AVX2-SLOW-NEXT: vmovaps {{.*#+}} ymm2 = <0,4,u,1,5,u,2,6>
; AVX2-SLOW-NEXT: vpermps %ymm0, %ymm2, %ymm2
; AVX2-SLOW-NEXT: vbroadcastsd %xmm1, %ymm3
; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1],ymm3[2],ymm2[3,4],ymm3[5],ymm2[6,7]
; AVX2-SLOW-NEXT: vextractf128 $1, %ymm0, %xmm3
; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm0 = xmm0[0,3],xmm3[3,3]
; AVX2-SLOW-NEXT: vpermilps {{.*#+}} xmm1 = xmm1[2,3,2,3]
; AVX2-SLOW-NEXT: vblendps {{.*#+}} xmm0 = xmm1[0],xmm0[1,2],xmm1[3]
; AVX2-SLOW-NEXT: vmovaps %xmm0, 32(%rdi)
; AVX2-SLOW-NEXT: vmovaps %ymm2, (%rdi)
X86 Tests: Update more isel tests with FastVariableShuffle feature Summary: Added the FastVariableShuffle feature to cases that resembled processors for which this fearure is on. For AVX2 there are processors with and w/o this fearue enable. For AVX512 only KNL does enable this feature so cases which only have +avx512f were left without the FastVariableShuffle enabled. Reviewers: RKSimon, craig.topper Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D41851 llvm-svn: 322090
2018-01-10 00:26:06 +08:00
; AVX2-SLOW-NEXT: vzeroupper
; AVX2-SLOW-NEXT: retq
;
; AVX2-FAST-LABEL: v12i32:
; AVX2-FAST: # %bb.0:
[X86] Fix a subtle bug in the custom execution domain fixing for blends. The code tried to find the immediate by using getNumOperands() on the MachineInstr, but there might be implicit-defs after the immediate that get counted. Instead use getNumOperands() from the instruction description which will only count the operands that are defined in the td file. llvm-svn: 337088
2018-07-14 14:30:30 +08:00
; AVX2-FAST-NEXT: vmovaps {{.*#+}} ymm2 = <0,4,u,1,5,u,2,6>
; AVX2-FAST-NEXT: vpermps %ymm0, %ymm2, %ymm2
; AVX2-FAST-NEXT: vbroadcastsd %xmm1, %ymm3
; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1],ymm3[2],ymm2[3,4],ymm3[5],ymm2[6,7]
; AVX2-FAST-NEXT: vmovaps {{.*#+}} ymm3 = [3,3,7,7,7,7,6,7]
; AVX2-FAST-NEXT: vpermps %ymm0, %ymm3, %ymm0
; AVX2-FAST-NEXT: vpermilps {{.*#+}} xmm1 = xmm1[2,3,2,3]
; AVX2-FAST-NEXT: vblendps {{.*#+}} xmm0 = xmm1[0],xmm0[1,2],xmm1[3]
; AVX2-FAST-NEXT: vmovaps %xmm0, 32(%rdi)
; AVX2-FAST-NEXT: vmovaps %ymm2, (%rdi)
X86 Tests: Update more isel tests with FastVariableShuffle feature Summary: Added the FastVariableShuffle feature to cases that resembled processors for which this fearure is on. For AVX2 there are processors with and w/o this fearue enable. For AVX512 only KNL does enable this feature so cases which only have +avx512f were left without the FastVariableShuffle enabled. Reviewers: RKSimon, craig.topper Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D41851 llvm-svn: 322090
2018-01-10 00:26:06 +08:00
; AVX2-FAST-NEXT: vzeroupper
; AVX2-FAST-NEXT: retq
[X86][XOP] Add XOP oddshuffles tests XOP codegen is often different to generic AVX - thank you vpperm! llvm-svn: 315176
2017-10-08 20:58:15 +08:00
;
; XOP-LABEL: v12i32:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; XOP: # %bb.0:
[X86][XOP] Add XOP oddshuffles tests XOP codegen is often different to generic AVX - thank you vpperm! llvm-svn: 315176
2017-10-08 20:58:15 +08:00
; XOP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm0[2,3,0,1]
; XOP-NEXT: vpermil2ps {{.*#+}} ymm2 = ymm0[0],ymm2[0],ymm0[u,1,5,u],ymm2[6],ymm0[6]
; XOP-NEXT: vmovddup {{.*#+}} xmm3 = xmm1[0,0]
; XOP-NEXT: vinsertf128 $1, %xmm3, %ymm3, %ymm3
; XOP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1],ymm3[2],ymm2[3,4],ymm3[5],ymm2[6,7]
; XOP-NEXT: vextractf128 $1, %ymm0, %xmm3
; XOP-NEXT: vshufps {{.*#+}} xmm0 = xmm0[0,3],xmm3[3,3]
; XOP-NEXT: vpermilpd {{.*#+}} xmm1 = xmm1[1,1]
; XOP-NEXT: vblendps {{.*#+}} xmm0 = xmm1[0],xmm0[1,2],xmm1[3]
[X86][SSE] Add custom execution domain fixing for BLENDPD/BLENDPS/PBLENDD/PBLENDW (PR34873) Add support for custom execution domain fixing and implement support for BLENDPD/BLENDPS/PBLENDD/PBLENDW. Differential Revision: https://reviews.llvm.org/D42042 llvm-svn: 322524
2018-01-16 06:18:45 +08:00
; XOP-NEXT: vmovaps %xmm0, 32(%rdi)
[X86][XOP] Add XOP oddshuffles tests XOP codegen is often different to generic AVX - thank you vpperm! llvm-svn: 315176
2017-10-08 20:58:15 +08:00
; XOP-NEXT: vmovaps %ymm2, (%rdi)
; XOP-NEXT: vzeroupper
; XOP-NEXT: retq
[X86] Add baseline test for "odd" shuffles. NFC. Adds a baseline test for lowering shuffles where the width of the output vector is not twice the size of the input vectors. Many of those sequences are suboptimal, and will hopefully be improved in follow-up patches. llvm-svn: 279888
2016-08-27 08:10:24 +08:00
%r = shufflevector <8 x i32> %a, <8 x i32> %b, <12 x i32> <i32 0, i32 4, i32 8, i32 1, i32 5, i32 9, i32 2, i32 6, i32 10, i32 3, i32 7, i32 11>
store <12 x i32> %r, <12 x i32>* %p
ret void
}
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
define void @pr29025(<4 x i8> %a, <4 x i8> %b, <4 x i8> %c, <12 x i8> *%p) nounwind {
[X86] Add baseline test for "odd" shuffles. NFC. Adds a baseline test for lowering shuffles where the width of the output vector is not twice the size of the input vectors. Many of those sequences are suboptimal, and will hopefully be improved in follow-up patches. llvm-svn: 279888
2016-08-27 08:10:24 +08:00
; SSE2-LABEL: pr29025:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/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:
[SelectionDAG] Generate vector_shuffle nodes for undersized result vector sizes Prior to this, we could generate a vector_shuffle from an IR shuffle when the size of the result was exactly the sum of the sizes of the input vectors. If the output vector was narrower - e.g. a <12 x i8> being formed by a shuffle with two <8 x i8> inputs - we would lower the shuffle to a sequence of extracts and inserts. Instead, we can form a larger vector_shuffle, and then extract a subvector of the right size - e.g. shuffle the two <8 x i8> inputs into a <16 x i8> and then extract a <12 x i8>. This also includes a target-specific X86 combine that in the presence of AVX2 combines: (vector_shuffle <mask> (concat_vectors t1, undef) (concat_vectors t2, undef)) into: (vector_shuffle <mask> (concat_vectors t1, t2), undef) in cases where this allows us to form VPERMD/VPERMQ. (This is not a separate commit, as that pattern does not appear without the DAGBuilder change.) llvm-svn: 280418
2016-09-02 05:32:09 +08:00
; SSE2-NEXT: movdqa {{.*#+}} xmm3 = [255,255,255,255]
; SSE2-NEXT: pand %xmm3, %xmm1
; SSE2-NEXT: pand %xmm3, %xmm0
; SSE2-NEXT: packuswb %xmm1, %xmm0
; SSE2-NEXT: packuswb %xmm0, %xmm0
; SSE2-NEXT: pxor %xmm1, %xmm1
[X86] Add baseline test for "odd" shuffles. NFC. Adds a baseline test for lowering shuffles where the width of the output vector is not twice the size of the input vectors. Many of those sequences are suboptimal, and will hopefully be improved in follow-up patches. llvm-svn: 279888
2016-08-27 08:10:24 +08:00
; SSE2-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3],xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
[SelectionDAG] Generate vector_shuffle nodes for undersized result vector sizes Prior to this, we could generate a vector_shuffle from an IR shuffle when the size of the result was exactly the sum of the sizes of the input vectors. If the output vector was narrower - e.g. a <12 x i8> being formed by a shuffle with two <8 x i8> inputs - we would lower the shuffle to a sequence of extracts and inserts. Instead, we can form a larger vector_shuffle, and then extract a subvector of the right size - e.g. shuffle the two <8 x i8> inputs into a <16 x i8> and then extract a <12 x i8>. This also includes a target-specific X86 combine that in the presence of AVX2 combines: (vector_shuffle <mask> (concat_vectors t1, undef) (concat_vectors t2, undef)) into: (vector_shuffle <mask> (concat_vectors t1, t2), undef) in cases where this allows us to form VPERMD/VPERMQ. (This is not a separate commit, as that pattern does not appear without the DAGBuilder change.) llvm-svn: 280418
2016-09-02 05:32:09 +08:00
; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[3,1,2,3]
; SSE2-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[0,3,1,3,4,5,6,7]
; SSE2-NEXT: pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,6,5,4,7]
; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,3,2,1]
; SSE2-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[0,2,2,1,4,5,6,7]
; SSE2-NEXT: pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,5,5,6,4]
; SSE2-NEXT: packuswb %xmm1, %xmm0
; SSE2-NEXT: movdqa {{.*#+}} xmm1 = [255,255,0,255,255,0,255,255,0,255,255,0,255,255,255,255]
; SSE2-NEXT: pand %xmm1, %xmm0
; SSE2-NEXT: pand %xmm3, %xmm2
; SSE2-NEXT: packuswb %xmm2, %xmm2
; SSE2-NEXT: packuswb %xmm2, %xmm2
; SSE2-NEXT: pshuflw {{.*#+}} xmm2 = xmm2[0,0,1,1,4,5,6,7]
[X86][SSE] Match PSHUFLW/PSHUFHW + PSHUFD vXi16 shuffle patterns (PR34686) As noted in PR34686, we are relying on a PSHUFD+PSHUFLW+PSHUFHW shuffle chain for most general vXi16 unary shuffles. This patch checks for simpler PSHUFLW+PSHUFD and PSHUFHW+PSHUFD cases beforehand, building on some existing code that just handled splat shuffles. By doing so we also prevent premature use of PSHUFB shuffles which can be slower and require the creation/loading of constant shuffle masks. We now have the 'fast-variable-shuffle' option for hardware that prefers combining 2 or more shuffles to VPSHUFB etc. Differential Revision: https://reviews.llvm.org/D38318 llvm-svn: 321553
2017-12-29 22:41:50 +08:00
; SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,0,1,3]
[SelectionDAG] Generate vector_shuffle nodes for undersized result vector sizes Prior to this, we could generate a vector_shuffle from an IR shuffle when the size of the result was exactly the sum of the sizes of the input vectors. If the output vector was narrower - e.g. a <12 x i8> being formed by a shuffle with two <8 x i8> inputs - we would lower the shuffle to a sequence of extracts and inserts. Instead, we can form a larger vector_shuffle, and then extract a subvector of the right size - e.g. shuffle the two <8 x i8> inputs into a <16 x i8> and then extract a <12 x i8>. This also includes a target-specific X86 combine that in the presence of AVX2 combines: (vector_shuffle <mask> (concat_vectors t1, undef) (concat_vectors t2, undef)) into: (vector_shuffle <mask> (concat_vectors t1, t2), undef) in cases where this allows us to form VPERMD/VPERMQ. (This is not a separate commit, as that pattern does not appear without the DAGBuilder change.) llvm-svn: 280418
2016-09-02 05:32:09 +08:00
; SSE2-NEXT: pandn %xmm2, %xmm1
; SSE2-NEXT: por %xmm0, %xmm1
; SSE2-NEXT: movq %xmm1, (%rdi)
; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,0,1]
[X86] Add baseline test for "odd" shuffles. NFC. Adds a baseline test for lowering shuffles where the width of the output vector is not twice the size of the input vectors. Many of those sequences are suboptimal, and will hopefully be improved in follow-up patches. llvm-svn: 279888
2016-08-27 08:10:24 +08:00
; SSE2-NEXT: movd %xmm0, 8(%rdi)
; SSE2-NEXT: retq
;
; SSE42-LABEL: pr29025:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; SSE42: # %bb.0:
[SelectionDAG] Generate vector_shuffle nodes for undersized result vector sizes Prior to this, we could generate a vector_shuffle from an IR shuffle when the size of the result was exactly the sum of the sizes of the input vectors. If the output vector was narrower - e.g. a <12 x i8> being formed by a shuffle with two <8 x i8> inputs - we would lower the shuffle to a sequence of extracts and inserts. Instead, we can form a larger vector_shuffle, and then extract a subvector of the right size - e.g. shuffle the two <8 x i8> inputs into a <16 x i8> and then extract a <12 x i8>. This also includes a target-specific X86 combine that in the presence of AVX2 combines: (vector_shuffle <mask> (concat_vectors t1, undef) (concat_vectors t2, undef)) into: (vector_shuffle <mask> (concat_vectors t1, t2), undef) in cases where this allows us to form VPERMD/VPERMQ. (This is not a separate commit, as that pattern does not appear without the DAGBuilder change.) llvm-svn: 280418
2016-09-02 05:32:09 +08:00
; SSE42-NEXT: movdqa {{.*#+}} xmm3 = <0,4,8,12,u,u,u,u,u,u,u,u,u,u,u,u>
; SSE42-NEXT: pshufb %xmm3, %xmm1
; SSE42-NEXT: pshufb %xmm3, %xmm0
; SSE42-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; SSE42-NEXT: pshufb %xmm3, %xmm2
; SSE42-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3],xmm0[4],xmm2[4],xmm0[5],xmm2[5],xmm0[6],xmm2[6],xmm0[7],xmm2[7]
; SSE42-NEXT: pshufb {{.*#+}} xmm0 = xmm0[0,8,1,2,10,3,4,12,5,6,14,7,u,u,u,u]
[X86] Add baseline test for "odd" shuffles. NFC. Adds a baseline test for lowering shuffles where the width of the output vector is not twice the size of the input vectors. Many of those sequences are suboptimal, and will hopefully be improved in follow-up patches. llvm-svn: 279888
2016-08-27 08:10:24 +08:00
; SSE42-NEXT: pextrd $2, %xmm0, 8(%rdi)
; SSE42-NEXT: movq %xmm0, (%rdi)
; SSE42-NEXT: retq
;
; AVX-LABEL: pr29025:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/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:
[SelectionDAG] Generate vector_shuffle nodes for undersized result vector sizes Prior to this, we could generate a vector_shuffle from an IR shuffle when the size of the result was exactly the sum of the sizes of the input vectors. If the output vector was narrower - e.g. a <12 x i8> being formed by a shuffle with two <8 x i8> inputs - we would lower the shuffle to a sequence of extracts and inserts. Instead, we can form a larger vector_shuffle, and then extract a subvector of the right size - e.g. shuffle the two <8 x i8> inputs into a <16 x i8> and then extract a <12 x i8>. This also includes a target-specific X86 combine that in the presence of AVX2 combines: (vector_shuffle <mask> (concat_vectors t1, undef) (concat_vectors t2, undef)) into: (vector_shuffle <mask> (concat_vectors t1, t2), undef) in cases where this allows us to form VPERMD/VPERMQ. (This is not a separate commit, as that pattern does not appear without the DAGBuilder change.) llvm-svn: 280418
2016-09-02 05:32:09 +08:00
; AVX-NEXT: vmovdqa {{.*#+}} xmm3 = <0,4,8,12,u,u,u,u,u,u,u,u,u,u,u,u>
; AVX-NEXT: vpshufb %xmm3, %xmm1, %xmm1
; AVX-NEXT: vpshufb %xmm3, %xmm0, %xmm0
; AVX-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; AVX-NEXT: vpshufb %xmm3, %xmm2, %xmm1
; AVX-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3],xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
; AVX-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,8,1,2,10,3,4,12,5,6,14,7,u,u,u,u]
[X86] Add baseline test for "odd" shuffles. NFC. Adds a baseline test for lowering shuffles where the width of the output vector is not twice the size of the input vectors. Many of those sequences are suboptimal, and will hopefully be improved in follow-up patches. llvm-svn: 279888
2016-08-27 08:10:24 +08:00
; AVX-NEXT: vpextrd $2, %xmm0, 8(%rdi)
; AVX-NEXT: vmovq %xmm0, (%rdi)
; AVX-NEXT: retq
[X86][XOP] Add XOP oddshuffles tests XOP codegen is often different to generic AVX - thank you vpperm! llvm-svn: 315176
2017-10-08 20:58:15 +08:00
;
; XOP-LABEL: pr29025:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; XOP: # %bb.0:
[X86][XOP] Add XOP oddshuffles tests XOP codegen is often different to generic AVX - thank you vpperm! llvm-svn: 315176
2017-10-08 20:58:15 +08:00
; XOP-NEXT: vpperm {{.*#+}} xmm0 = xmm0[0,4,8,12],xmm1[0,4,8,12],xmm0[u,u,u,u,u,u,u,u]
; XOP-NEXT: vpperm {{.*#+}} xmm0 = xmm0[0,4],xmm2[0],xmm0[1,5],xmm2[4],xmm0[2,6],xmm2[8],xmm0[3,7],xmm2[12],xmm0[u,u,u,u]
; XOP-NEXT: vpextrd $2, %xmm0, 8(%rdi)
; XOP-NEXT: vmovq %xmm0, (%rdi)
; XOP-NEXT: retq
[X86] Add baseline test for "odd" shuffles. NFC. Adds a baseline test for lowering shuffles where the width of the output vector is not twice the size of the input vectors. Many of those sequences are suboptimal, and will hopefully be improved in follow-up patches. llvm-svn: 279888
2016-08-27 08:10:24 +08:00
%s1 = shufflevector <4 x i8> %a, <4 x i8> %b, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
%s2 = shufflevector <4 x i8> %c, <4 x i8> undef, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef>
%r = shufflevector <8 x i8> %s1, <8 x i8> %s2, <12 x i32> <i32 0, i32 4, i32 8, i32 1, i32 5, i32 9, i32 2, i32 6, i32 10, i32 3, i32 7, i32 11>
store <12 x i8> %r, <12 x i8>* %p, align 1
ret void
}
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
define void @interleave_24i8_out(<24 x i8>* %p, <8 x i8>* %q1, <8 x i8>* %q2, <8 x i8>* %q3) nounwind {
; SSE2-LABEL: interleave_24i8_out:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/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] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
; SSE2-NEXT: movdqu (%rdi), %xmm0
; SSE2-NEXT: movq {{.*#+}} xmm1 = mem[0],zero
[X86][SSE] Add shuffle combine support for OR(PSHUFB,PSHUFB) style patterns. If each element is zero from one (or both) inputs then we can combine these into a single shuffle mask. llvm-svn: 339686
2018-08-15 00:00:05 +08:00
; SSE2-NEXT: movdqa {{.*#+}} xmm4 = [255,255,0,255,255,0,255,255,255,255,255,255,255,255,255,255]
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
; SSE2-NEXT: movdqa %xmm0, %xmm2
[X86][SSE] Add shuffle combine support for OR(PSHUFB,PSHUFB) style patterns. If each element is zero from one (or both) inputs then we can combine these into a single shuffle mask. llvm-svn: 339686
2018-08-15 00:00:05 +08:00
; SSE2-NEXT: pand %xmm4, %xmm2
; SSE2-NEXT: pandn %xmm1, %xmm4
; SSE2-NEXT: por %xmm2, %xmm4
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
; SSE2-NEXT: pxor %xmm2, %xmm2
[X86][SSE] Add shuffle combine support for OR(PSHUFB,PSHUFB) style patterns. If each element is zero from one (or both) inputs then we can combine these into a single shuffle mask. llvm-svn: 339686
2018-08-15 00:00:05 +08:00
; SSE2-NEXT: punpcklbw {{.*#+}} xmm4 = xmm4[0],xmm2[0],xmm4[1],xmm2[1],xmm4[2],xmm2[2],xmm4[3],xmm2[3],xmm4[4],xmm2[4],xmm4[5],xmm2[5],xmm4[6],xmm2[6],xmm4[7],xmm2[7]
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
; SSE2-NEXT: movdqa {{.*#+}} xmm5 = [65535,0,65535,65535,0,65535,65535,0]
[X86][SSE] Add shuffle combine support for OR(PSHUFB,PSHUFB) style patterns. If each element is zero from one (or both) inputs then we can combine these into a single shuffle mask. llvm-svn: 339686
2018-08-15 00:00:05 +08:00
; SSE2-NEXT: pand %xmm5, %xmm4
; SSE2-NEXT: movdqa %xmm0, %xmm3
; SSE2-NEXT: punpckhbw {{.*#+}} xmm3 = xmm3[8],xmm2[8],xmm3[9],xmm2[9],xmm3[10],xmm2[10],xmm3[11],xmm2[11],xmm3[12],xmm2[12],xmm3[13],xmm2[13],xmm3[14],xmm2[14],xmm3[15],xmm2[15]
; SSE2-NEXT: pandn %xmm3, %xmm5
; SSE2-NEXT: por %xmm4, %xmm5
; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm5[0,2,1,3]
; SSE2-NEXT: pshufhw {{.*#+}} xmm4 = xmm4[0,1,2,3,4,7,6,5]
; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm4[0,3,2,1]
; SSE2-NEXT: pshuflw {{.*#+}} xmm4 = xmm4[0,3,2,1,4,5,6,7]
; SSE2-NEXT: pshufhw {{.*#+}} xmm4 = xmm4[0,1,2,3,6,5,4,7]
; SSE2-NEXT: packuswb %xmm0, %xmm4
; SSE2-NEXT: movq %xmm4, (%rsi)
; SSE2-NEXT: movdqa {{.*#+}} xmm4 = [0,255,255,0,255,255,0,255,255,255,255,255,255,255,255,255]
; SSE2-NEXT: movdqa %xmm0, %xmm5
; SSE2-NEXT: pand %xmm4, %xmm5
; SSE2-NEXT: pandn %xmm1, %xmm4
; SSE2-NEXT: por %xmm5, %xmm4
; SSE2-NEXT: punpcklbw {{.*#+}} xmm4 = xmm4[0],xmm2[0],xmm4[1],xmm2[1],xmm4[2],xmm2[2],xmm4[3],xmm2[3],xmm4[4],xmm2[4],xmm4[5],xmm2[5],xmm4[6],xmm2[6],xmm4[7],xmm2[7]
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
; SSE2-NEXT: movdqa {{.*#+}} xmm5 = [65535,65535,0,65535,65535,0,65535,65535]
[X86][SSE] Add shuffle combine support for OR(PSHUFB,PSHUFB) style patterns. If each element is zero from one (or both) inputs then we can combine these into a single shuffle mask. llvm-svn: 339686
2018-08-15 00:00:05 +08:00
; SSE2-NEXT: pand %xmm5, %xmm4
; SSE2-NEXT: pandn %xmm3, %xmm5
; SSE2-NEXT: por %xmm4, %xmm5
; SSE2-NEXT: pshuflw {{.*#+}} xmm4 = xmm5[2,1,0,3,4,5,6,7]
; SSE2-NEXT: pshufhw {{.*#+}} xmm4 = xmm4[0,1,2,3,6,5,4,7]
; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm4[0,3,2,1]
; SSE2-NEXT: pshuflw {{.*#+}} xmm4 = xmm4[1,2,3,0,4,5,6,7]
; SSE2-NEXT: pshufhw {{.*#+}} xmm4 = xmm4[0,1,2,3,5,6,7,4]
; SSE2-NEXT: packuswb %xmm0, %xmm4
; SSE2-NEXT: movq %xmm4, (%rdx)
; SSE2-NEXT: movdqa {{.*#+}} xmm4 = [255,0,255,255,0,255,255,0,255,255,255,255,255,255,255,255]
; SSE2-NEXT: pand %xmm4, %xmm0
; SSE2-NEXT: pandn %xmm1, %xmm4
; SSE2-NEXT: por %xmm0, %xmm4
; SSE2-NEXT: punpcklbw {{.*#+}} xmm4 = xmm4[0],xmm2[0],xmm4[1],xmm2[1],xmm4[2],xmm2[2],xmm4[3],xmm2[3],xmm4[4],xmm2[4],xmm4[5],xmm2[5],xmm4[6],xmm2[6],xmm4[7],xmm2[7]
; SSE2-NEXT: movdqa {{.*#+}} xmm0 = [0,65535,65535,0,65535,65535,0,65535]
; SSE2-NEXT: pand %xmm0, %xmm4
; SSE2-NEXT: pandn %xmm3, %xmm0
; SSE2-NEXT: por %xmm4, %xmm0
; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[3,1,2,0]
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
; SSE2-NEXT: pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,7,6,5]
; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[3,1,2,0]
; SSE2-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[2,1,0,3,4,5,6,7]
; SSE2-NEXT: pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,6,5,4,7]
; SSE2-NEXT: packuswb %xmm0, %xmm0
; SSE2-NEXT: movq %xmm0, (%rcx)
; SSE2-NEXT: retq
;
; SSE42-LABEL: interleave_24i8_out:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; SSE42: # %bb.0:
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
; SSE42-NEXT: movdqu (%rdi), %xmm0
; SSE42-NEXT: movq {{.*#+}} xmm1 = mem[0],zero
; SSE42-NEXT: movdqa %xmm1, %xmm2
; SSE42-NEXT: pshufb {{.*#+}} xmm2 = zero,zero,zero,zero,zero,zero,xmm2[2,5,u,u,u,u,u,u,u,u]
; SSE42-NEXT: movdqa %xmm0, %xmm3
; SSE42-NEXT: pshufb {{.*#+}} xmm3 = xmm3[0,3,6,9,12,15],zero,zero,xmm3[u,u,u,u,u,u,u,u]
; SSE42-NEXT: por %xmm2, %xmm3
; SSE42-NEXT: movq %xmm3, (%rsi)
; SSE42-NEXT: movdqa %xmm1, %xmm2
; SSE42-NEXT: pshufb {{.*#+}} xmm2 = zero,zero,zero,zero,zero,xmm2[0,3,6,u,u,u,u,u,u,u,u]
; SSE42-NEXT: movdqa %xmm0, %xmm3
; SSE42-NEXT: pshufb {{.*#+}} xmm3 = xmm3[1,4,7,10,13],zero,zero,zero,xmm3[u,u,u,u,u,u,u,u]
; SSE42-NEXT: por %xmm2, %xmm3
; SSE42-NEXT: movq %xmm3, (%rdx)
; SSE42-NEXT: pshufb {{.*#+}} xmm1 = zero,zero,zero,zero,zero,xmm1[1,4,7,u,u,u,u,u,u,u,u]
; SSE42-NEXT: pshufb {{.*#+}} xmm0 = xmm0[2,5,8,11,14],zero,zero,zero,xmm0[u,u,u,u,u,u,u,u]
; SSE42-NEXT: por %xmm1, %xmm0
; SSE42-NEXT: movq %xmm0, (%rcx)
; SSE42-NEXT: retq
;
; AVX-LABEL: interleave_24i8_out:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/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] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
; AVX-NEXT: vmovdqu (%rdi), %xmm0
; AVX-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero
; AVX-NEXT: vpshufb {{.*#+}} xmm2 = zero,zero,zero,zero,zero,zero,xmm1[2,5,u,u,u,u,u,u,u,u]
; AVX-NEXT: vpshufb {{.*#+}} xmm3 = xmm0[0,3,6,9,12,15],zero,zero,xmm0[u,u,u,u,u,u,u,u]
; AVX-NEXT: vpor %xmm2, %xmm3, %xmm2
; AVX-NEXT: vmovq %xmm2, (%rsi)
; AVX-NEXT: vpshufb {{.*#+}} xmm2 = zero,zero,zero,zero,zero,xmm1[0,3,6,u,u,u,u,u,u,u,u]
; AVX-NEXT: vpshufb {{.*#+}} xmm3 = xmm0[1,4,7,10,13],zero,zero,zero,xmm0[u,u,u,u,u,u,u,u]
; AVX-NEXT: vpor %xmm2, %xmm3, %xmm2
; AVX-NEXT: vmovq %xmm2, (%rdx)
; AVX-NEXT: vpshufb {{.*#+}} xmm1 = zero,zero,zero,zero,zero,xmm1[1,4,7,u,u,u,u,u,u,u,u]
; AVX-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[2,5,8,11,14],zero,zero,zero,xmm0[u,u,u,u,u,u,u,u]
; AVX-NEXT: vpor %xmm1, %xmm0, %xmm0
; AVX-NEXT: vmovq %xmm0, (%rcx)
; AVX-NEXT: retq
[X86][XOP] Add XOP oddshuffles tests XOP codegen is often different to generic AVX - thank you vpperm! llvm-svn: 315176
2017-10-08 20:58:15 +08:00
;
; XOP-LABEL: interleave_24i8_out:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; XOP: # %bb.0:
[X86][XOP] Add XOP oddshuffles tests XOP codegen is often different to generic AVX - thank you vpperm! llvm-svn: 315176
2017-10-08 20:58:15 +08:00
; XOP-NEXT: vmovdqu (%rdi), %xmm0
; XOP-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero
; XOP-NEXT: vpshufb {{.*#+}} xmm2 = zero,zero,zero,zero,zero,zero,xmm1[2,5,u,u,u,u,u,u,u,u]
; XOP-NEXT: vpshufb {{.*#+}} xmm3 = xmm0[0,3,6,9,12,15],zero,zero,xmm0[u,u,u,u,u,u,u,u]
; XOP-NEXT: vpor %xmm2, %xmm3, %xmm2
; XOP-NEXT: vmovq %xmm2, (%rsi)
; XOP-NEXT: vpshufb {{.*#+}} xmm2 = zero,zero,zero,zero,zero,xmm1[0,3,6,u,u,u,u,u,u,u,u]
; XOP-NEXT: vpshufb {{.*#+}} xmm3 = xmm0[1,4,7,10,13],zero,zero,zero,xmm0[u,u,u,u,u,u,u,u]
; XOP-NEXT: vpor %xmm2, %xmm3, %xmm2
; XOP-NEXT: vmovq %xmm2, (%rdx)
; XOP-NEXT: vpshufb {{.*#+}} xmm1 = zero,zero,zero,zero,zero,xmm1[1,4,7,u,u,u,u,u,u,u,u]
; XOP-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[2,5,8,11,14],zero,zero,zero,xmm0[u,u,u,u,u,u,u,u]
; XOP-NEXT: vpor %xmm1, %xmm0, %xmm0
; XOP-NEXT: vmovq %xmm0, (%rcx)
; XOP-NEXT: retq
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
%wide.vec = load <24 x i8>, <24 x i8>* %p, align 4
%s1 = shufflevector <24 x i8> %wide.vec, <24 x i8> undef, <8 x i32> <i32 0, i32 3, i32 6, i32 9, i32 12, i32 15, i32 18, i32 21>
%s2 = shufflevector <24 x i8> %wide.vec, <24 x i8> undef, <8 x i32> <i32 1, i32 4, i32 7, i32 10, i32 13, i32 16, i32 19, i32 22>
%s3 = shufflevector <24 x i8> %wide.vec, <24 x i8> undef, <8 x i32> <i32 2, i32 5, i32 8, i32 11, i32 14, i32 17, i32 20, i32 23>
store <8 x i8> %s1, <8 x i8>* %q1, align 4
store <8 x i8> %s2, <8 x i8>* %q2, align 4
store <8 x i8> %s3, <8 x i8>* %q3, align 4
ret void
}
define void @interleave_24i8_in(<24 x i8>* %p, <8 x i8>* %q1, <8 x i8>* %q2, <8 x i8>* %q3) nounwind {
; SSE2-LABEL: interleave_24i8_in:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/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] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
; SSE2-NEXT: movq {{.*#+}} xmm1 = mem[0],zero
[X86][SSE] Combine (some) target shuffles with multiple uses As discussed on D41794, we have many cases where we fail to combine shuffles as the input operands have other uses. This patch permits these shuffles to be combined as long as they don't introduce additional variable shuffle masks, which should reduce instruction dependencies and allow the total number of shuffles to still drop without increasing the constant pool. However, this may mean that some memory folds may no longer occur, and on pre-AVX require the occasional extra register move. This also exposes some poor PMULDQ/PMULUDQ codegen which was doing unnecessary upper/lower calculations which will in fact fold to zero/undef - the fix will be added in a followup commit. Differential Revision: https://reviews.llvm.org/D50328 llvm-svn: 339335
2018-08-09 20:30:02 +08:00
; SSE2-NEXT: movq {{.*#+}} xmm2 = mem[0],zero
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
; SSE2-NEXT: movq {{.*#+}} xmm0 = mem[0],zero
[X86][SSE] Combine (some) target shuffles with multiple uses As discussed on D41794, we have many cases where we fail to combine shuffles as the input operands have other uses. This patch permits these shuffles to be combined as long as they don't introduce additional variable shuffle masks, which should reduce instruction dependencies and allow the total number of shuffles to still drop without increasing the constant pool. However, this may mean that some memory folds may no longer occur, and on pre-AVX require the occasional extra register move. This also exposes some poor PMULDQ/PMULUDQ codegen which was doing unnecessary upper/lower calculations which will in fact fold to zero/undef - the fix will be added in a followup commit. Differential Revision: https://reviews.llvm.org/D50328 llvm-svn: 339335
2018-08-09 20:30:02 +08:00
; SSE2-NEXT: pxor %xmm3, %xmm3
; SSE2-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1],xmm1[2],xmm3[2],xmm1[3],xmm3[3],xmm1[4],xmm3[4],xmm1[5],xmm3[5],xmm1[6],xmm3[6],xmm1[7],xmm3[7]
; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm1[1,1,2,2]
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
; SSE2-NEXT: movdqa {{.*#+}} xmm5 = [65535,65535,0,65535,65535,0,65535,65535]
; SSE2-NEXT: pand %xmm5, %xmm4
[X86][SSE] Combine (some) target shuffles with multiple uses As discussed on D41794, we have many cases where we fail to combine shuffles as the input operands have other uses. This patch permits these shuffles to be combined as long as they don't introduce additional variable shuffle masks, which should reduce instruction dependencies and allow the total number of shuffles to still drop without increasing the constant pool. However, this may mean that some memory folds may no longer occur, and on pre-AVX require the occasional extra register move. This also exposes some poor PMULDQ/PMULUDQ codegen which was doing unnecessary upper/lower calculations which will in fact fold to zero/undef - the fix will be added in a followup commit. Differential Revision: https://reviews.llvm.org/D50328 llvm-svn: 339335
2018-08-09 20:30:02 +08:00
; SSE2-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3],xmm2[4],xmm3[4],xmm2[5],xmm3[5],xmm2[6],xmm3[6],xmm2[7],xmm3[7]
; SSE2-NEXT: pshuflw {{.*#+}} xmm3 = xmm2[0,1,3,3,4,5,6,7]
; SSE2-NEXT: pshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,4,6,7]
; SSE2-NEXT: pandn %xmm3, %xmm5
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
; SSE2-NEXT: por %xmm4, %xmm5
[X86][SSE] Combine (some) target shuffles with multiple uses As discussed on D41794, we have many cases where we fail to combine shuffles as the input operands have other uses. This patch permits these shuffles to be combined as long as they don't introduce additional variable shuffle masks, which should reduce instruction dependencies and allow the total number of shuffles to still drop without increasing the constant pool. However, this may mean that some memory folds may no longer occur, and on pre-AVX require the occasional extra register move. This also exposes some poor PMULDQ/PMULUDQ codegen which was doing unnecessary upper/lower calculations which will in fact fold to zero/undef - the fix will be added in a followup commit. Differential Revision: https://reviews.llvm.org/D50328 llvm-svn: 339335
2018-08-09 20:30:02 +08:00
; SSE2-NEXT: movdqa %xmm1, %xmm3
; SSE2-NEXT: punpcklwd {{.*#+}} xmm3 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm3[0,1,2,1]
; SSE2-NEXT: pshuflw {{.*#+}} xmm3 = xmm3[0,1,2,2,4,5,6,7]
; SSE2-NEXT: pshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,7,5,4,5]
; SSE2-NEXT: packuswb %xmm5, %xmm3
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
; SSE2-NEXT: movdqa {{.*#+}} xmm4 = [255,255,0,255,255,0,255,255,0,255,255,0,255,255,0,255]
[X86][SSE] Combine (some) target shuffles with multiple uses As discussed on D41794, we have many cases where we fail to combine shuffles as the input operands have other uses. This patch permits these shuffles to be combined as long as they don't introduce additional variable shuffle masks, which should reduce instruction dependencies and allow the total number of shuffles to still drop without increasing the constant pool. However, this may mean that some memory folds may no longer occur, and on pre-AVX require the occasional extra register move. This also exposes some poor PMULDQ/PMULUDQ codegen which was doing unnecessary upper/lower calculations which will in fact fold to zero/undef - the fix will be added in a followup commit. Differential Revision: https://reviews.llvm.org/D50328 llvm-svn: 339335
2018-08-09 20:30:02 +08:00
; SSE2-NEXT: pand %xmm4, %xmm3
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
; SSE2-NEXT: pshufd {{.*#+}} xmm5 = xmm0[0,1,0,1]
; SSE2-NEXT: pshuflw {{.*#+}} xmm5 = xmm5[0,0,0,3,4,5,6,7]
; SSE2-NEXT: pshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,5,5,6,6]
; SSE2-NEXT: pandn %xmm5, %xmm4
[X86][SSE] Combine (some) target shuffles with multiple uses As discussed on D41794, we have many cases where we fail to combine shuffles as the input operands have other uses. This patch permits these shuffles to be combined as long as they don't introduce additional variable shuffle masks, which should reduce instruction dependencies and allow the total number of shuffles to still drop without increasing the constant pool. However, this may mean that some memory folds may no longer occur, and on pre-AVX require the occasional extra register move. This also exposes some poor PMULDQ/PMULUDQ codegen which was doing unnecessary upper/lower calculations which will in fact fold to zero/undef - the fix will be added in a followup commit. Differential Revision: https://reviews.llvm.org/D50328 llvm-svn: 339335
2018-08-09 20:30:02 +08:00
; SSE2-NEXT: por %xmm3, %xmm4
; SSE2-NEXT: punpckhwd {{.*#+}} xmm2 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm2[2,1,2,3]
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
; SSE2-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[2,1,1,0,4,5,6,7]
; SSE2-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,7,6,7]
; SSE2-NEXT: packuswb %xmm0, %xmm1
; SSE2-NEXT: movdqa {{.*#+}} xmm2 = [255,0,255,255,0,255,255,0,255,255,255,255,255,255,255,255]
; SSE2-NEXT: pand %xmm2, %xmm1
; SSE2-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[2,1,3,3,4,5,6,7]
; SSE2-NEXT: pandn %xmm0, %xmm2
; SSE2-NEXT: por %xmm1, %xmm2
; SSE2-NEXT: movq %xmm2, 16(%rdi)
; SSE2-NEXT: movdqu %xmm4, (%rdi)
; SSE2-NEXT: retq
;
; SSE42-LABEL: interleave_24i8_in:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; SSE42: # %bb.0:
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
; SSE42-NEXT: movq {{.*#+}} xmm0 = mem[0],zero
; SSE42-NEXT: movq {{.*#+}} xmm1 = mem[0],zero
; SSE42-NEXT: punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
Add LiveRangeShrink pass to shrink live range within BB. Summary: LiveRangeShrink pass moves instruction right after the definition with the same BB if the instruction and its operands all have more than one use. This pass is inexpensive and guarantees optimal live-range within BB. Reviewers: davidxl, wmi, hfinkel, MatzeB, andreadb Reviewed By: MatzeB, andreadb Subscribers: hiraditya, jyknight, sanjoy, skatkov, gberry, jholewinski, qcolombet, javed.absar, krytarowski, atrick, spatel, RKSimon, andreadb, MatzeB, mehdi_amini, mgorny, efriedma, davide, dberlin, llvm-commits Differential Revision: https://reviews.llvm.org/D32563 llvm-svn: 304371
2017-06-01 07:25:25 +08:00
; SSE42-NEXT: movq {{.*#+}} xmm1 = mem[0],zero
; SSE42-NEXT: movdqa %xmm0, %xmm2
; SSE42-NEXT: pshufb {{.*#+}} xmm2 = xmm2[0,8],zero,xmm2[1,9],zero,xmm2[2,10],zero,xmm2[3,11],zero,xmm2[4,12],zero,xmm2[5]
; SSE42-NEXT: movdqa %xmm1, %xmm3
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
; SSE42-NEXT: pshufb {{.*#+}} xmm3 = zero,zero,xmm3[0],zero,zero,xmm3[1],zero,zero,xmm3[2],zero,zero,xmm3[3],zero,zero,xmm3[4],zero
Add LiveRangeShrink pass to shrink live range within BB. Summary: LiveRangeShrink pass moves instruction right after the definition with the same BB if the instruction and its operands all have more than one use. This pass is inexpensive and guarantees optimal live-range within BB. Reviewers: davidxl, wmi, hfinkel, MatzeB, andreadb Reviewed By: MatzeB, andreadb Subscribers: hiraditya, jyknight, sanjoy, skatkov, gberry, jholewinski, qcolombet, javed.absar, krytarowski, atrick, spatel, RKSimon, andreadb, MatzeB, mehdi_amini, mgorny, efriedma, davide, dberlin, llvm-commits Differential Revision: https://reviews.llvm.org/D32563 llvm-svn: 304371
2017-06-01 07:25:25 +08:00
; SSE42-NEXT: por %xmm2, %xmm3
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
; SSE42-NEXT: pshufb {{.*#+}} xmm0 = xmm0[13],zero,xmm0[6,14],zero,xmm0[7,15],zero,xmm0[u,u,u,u,u,u,u,u]
Add LiveRangeShrink pass to shrink live range within BB. Summary: LiveRangeShrink pass moves instruction right after the definition with the same BB if the instruction and its operands all have more than one use. This pass is inexpensive and guarantees optimal live-range within BB. Reviewers: davidxl, wmi, hfinkel, MatzeB, andreadb Reviewed By: MatzeB, andreadb Subscribers: hiraditya, jyknight, sanjoy, skatkov, gberry, jholewinski, qcolombet, javed.absar, krytarowski, atrick, spatel, RKSimon, andreadb, MatzeB, mehdi_amini, mgorny, efriedma, davide, dberlin, llvm-commits Differential Revision: https://reviews.llvm.org/D32563 llvm-svn: 304371
2017-06-01 07:25:25 +08:00
; SSE42-NEXT: pshufb {{.*#+}} xmm1 = zero,xmm1[5],zero,zero,xmm1[6],zero,zero,xmm1[7,u,u,u,u,u,u,u,u]
; SSE42-NEXT: por %xmm0, %xmm1
; SSE42-NEXT: movq %xmm1, 16(%rdi)
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
; SSE42-NEXT: movdqu %xmm3, (%rdi)
; SSE42-NEXT: retq
;
; AVX-LABEL: interleave_24i8_in:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/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] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
; AVX-NEXT: vmovq {{.*#+}} xmm0 = mem[0],zero
; AVX-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero
; AVX-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
Add LiveRangeShrink pass to shrink live range within BB. Summary: LiveRangeShrink pass moves instruction right after the definition with the same BB if the instruction and its operands all have more than one use. This pass is inexpensive and guarantees optimal live-range within BB. Reviewers: davidxl, wmi, hfinkel, MatzeB, andreadb Reviewed By: MatzeB, andreadb Subscribers: hiraditya, jyknight, sanjoy, skatkov, gberry, jholewinski, qcolombet, javed.absar, krytarowski, atrick, spatel, RKSimon, andreadb, MatzeB, mehdi_amini, mgorny, efriedma, davide, dberlin, llvm-commits Differential Revision: https://reviews.llvm.org/D32563 llvm-svn: 304371
2017-06-01 07:25:25 +08:00
; AVX-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero
; AVX-NEXT: vpshufb {{.*#+}} xmm2 = xmm0[0,8],zero,xmm0[1,9],zero,xmm0[2,10],zero,xmm0[3,11],zero,xmm0[4,12],zero,xmm0[5]
; AVX-NEXT: vpshufb {{.*#+}} xmm3 = zero,zero,xmm1[0],zero,zero,xmm1[1],zero,zero,xmm1[2],zero,zero,xmm1[3],zero,zero,xmm1[4],zero
; AVX-NEXT: vpor %xmm3, %xmm2, %xmm2
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
; AVX-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[13],zero,xmm0[6,14],zero,xmm0[7,15],zero,xmm0[u,u,u,u,u,u,u,u]
Add LiveRangeShrink pass to shrink live range within BB. Summary: LiveRangeShrink pass moves instruction right after the definition with the same BB if the instruction and its operands all have more than one use. This pass is inexpensive and guarantees optimal live-range within BB. Reviewers: davidxl, wmi, hfinkel, MatzeB, andreadb Reviewed By: MatzeB, andreadb Subscribers: hiraditya, jyknight, sanjoy, skatkov, gberry, jholewinski, qcolombet, javed.absar, krytarowski, atrick, spatel, RKSimon, andreadb, MatzeB, mehdi_amini, mgorny, efriedma, davide, dberlin, llvm-commits Differential Revision: https://reviews.llvm.org/D32563 llvm-svn: 304371
2017-06-01 07:25:25 +08:00
; AVX-NEXT: vpshufb {{.*#+}} xmm1 = zero,xmm1[5],zero,zero,xmm1[6],zero,zero,xmm1[7,u,u,u,u,u,u,u,u]
; AVX-NEXT: vpor %xmm1, %xmm0, %xmm0
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
; AVX-NEXT: vmovq %xmm0, 16(%rdi)
Add LiveRangeShrink pass to shrink live range within BB. Summary: LiveRangeShrink pass moves instruction right after the definition with the same BB if the instruction and its operands all have more than one use. This pass is inexpensive and guarantees optimal live-range within BB. Reviewers: davidxl, wmi, hfinkel, MatzeB, andreadb Reviewed By: MatzeB, andreadb Subscribers: hiraditya, jyknight, sanjoy, skatkov, gberry, jholewinski, qcolombet, javed.absar, krytarowski, atrick, spatel, RKSimon, andreadb, MatzeB, mehdi_amini, mgorny, efriedma, davide, dberlin, llvm-commits Differential Revision: https://reviews.llvm.org/D32563 llvm-svn: 304371
2017-06-01 07:25:25 +08:00
; AVX-NEXT: vmovdqu %xmm2, (%rdi)
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
; AVX-NEXT: retq
[X86][XOP] Add XOP oddshuffles tests XOP codegen is often different to generic AVX - thank you vpperm! llvm-svn: 315176
2017-10-08 20:58:15 +08:00
;
; XOP-LABEL: interleave_24i8_in:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; XOP: # %bb.0:
[X86][XOP] Add XOP oddshuffles tests XOP codegen is often different to generic AVX - thank you vpperm! llvm-svn: 315176
2017-10-08 20:58:15 +08:00
; XOP-NEXT: vmovq {{.*#+}} xmm0 = mem[0],zero
; XOP-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero
; XOP-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
; XOP-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero
; XOP-NEXT: vpshufb {{.*#+}} xmm2 = xmm0[0,8],zero,xmm0[1,9],zero,xmm0[2,10],zero,xmm0[3,11],zero,xmm0[4,12],zero,xmm0[5]
; XOP-NEXT: vpshufb {{.*#+}} xmm3 = zero,zero,xmm1[0],zero,zero,xmm1[1],zero,zero,xmm1[2],zero,zero,xmm1[3],zero,zero,xmm1[4],zero
; XOP-NEXT: vpor %xmm3, %xmm2, %xmm2
; XOP-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[13],zero,xmm0[6,14],zero,xmm0[7,15],zero,xmm0[u,u,u,u,u,u,u,u]
; XOP-NEXT: vpshufb {{.*#+}} xmm1 = zero,xmm1[5],zero,zero,xmm1[6],zero,zero,xmm1[7,u,u,u,u,u,u,u,u]
; XOP-NEXT: vpor %xmm1, %xmm0, %xmm0
; XOP-NEXT: vmovq %xmm0, 16(%rdi)
; XOP-NEXT: vmovdqu %xmm2, (%rdi)
; XOP-NEXT: retq
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
%s1 = load <8 x i8>, <8 x i8>* %q1, align 4
%s2 = load <8 x i8>, <8 x i8>* %q2, align 4
%s3 = load <8 x i8>, <8 x i8>* %q3, align 4
%t1 = shufflevector <8 x i8> %s1, <8 x i8> %s2, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
%t2 = shufflevector <8 x i8> %s3, <8 x i8> undef, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
%interleaved = shufflevector <16 x i8> %t1, <16 x i8> %t2, <24 x i32> <i32 0, i32 8, i32 16, i32 1, i32 9, i32 17, i32 2, i32 10, i32 18, i32 3, i32 11, i32 19, i32 4, i32 12, i32 20, i32 5, i32 13, i32 21, i32 6, i32 14, i32 22, i32 7, i32 15, i32 23>
store <24 x i8> %interleaved, <24 x i8>* %p, align 4
ret void
}
define void @interleave_24i16_out(<24 x i16>* %p, <8 x i16>* %q1, <8 x i16>* %q2, <8 x i16>* %q3) nounwind {
; SSE2-LABEL: interleave_24i16_out:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/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:
[DAG] Generalize build_vector -> vector_shuffle combine for more than 2 inputs This generalizes the build_vector -> vector_shuffle combine to support any number of inputs. The idea is to create a binary tree of shuffles, where the first layer performs pairwise shuffles of the input vectors placing each input element into the correct lane, and the rest of the tree blends these shuffles together. This doesn't try to be smart and create any sort of "optimal" shuffles. The assumption is that even a "poor" shuffle sequence is better than extracting and inserting the elements one by one. Differential Revision: https://reviews.llvm.org/D24683 llvm-svn: 283480
2016-10-07 02:58:24 +08:00
; SSE2-NEXT: movdqu (%rdi), %xmm3
; SSE2-NEXT: movdqu 16(%rdi), %xmm2
; SSE2-NEXT: movdqu 32(%rdi), %xmm8
; SSE2-NEXT: movdqa {{.*#+}} xmm1 = [65535,0,65535,65535,0,65535,65535,0]
; SSE2-NEXT: movdqa %xmm3, %xmm4
; SSE2-NEXT: pand %xmm1, %xmm4
; SSE2-NEXT: pandn %xmm2, %xmm1
; SSE2-NEXT: por %xmm4, %xmm1
; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,2,1,3]
; SSE2-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,6,5,6,7]
; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,2,1,3]
; SSE2-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[0,3,2,1,4,5,6,7]
[X86][SSE] Combine (some) target shuffles with multiple uses As discussed on D41794, we have many cases where we fail to combine shuffles as the input operands have other uses. This patch permits these shuffles to be combined as long as they don't introduce additional variable shuffle masks, which should reduce instruction dependencies and allow the total number of shuffles to still drop without increasing the constant pool. However, this may mean that some memory folds may no longer occur, and on pre-AVX require the occasional extra register move. This also exposes some poor PMULDQ/PMULUDQ codegen which was doing unnecessary upper/lower calculations which will in fact fold to zero/undef - the fix will be added in a followup commit. Differential Revision: https://reviews.llvm.org/D50328 llvm-svn: 339335
2018-08-09 20:30:02 +08:00
; SSE2-NEXT: pshufhw {{.*#+}} xmm4 = xmm1[0,1,2,3,4,7,6,7]
; SSE2-NEXT: pshufd {{.*#+}} xmm5 = xmm8[0,1,2,1]
; SSE2-NEXT: pshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,4,5,6,5]
; SSE2-NEXT: shufps {{.*#+}} xmm5 = xmm5[3,0],xmm4[2,0]
; SSE2-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,1],xmm5[2,0]
[DAG] Generalize build_vector -> vector_shuffle combine for more than 2 inputs This generalizes the build_vector -> vector_shuffle combine to support any number of inputs. The idea is to create a binary tree of shuffles, where the first layer performs pairwise shuffles of the input vectors placing each input element into the correct lane, and the rest of the tree blends these shuffles together. This doesn't try to be smart and create any sort of "optimal" shuffles. The assumption is that even a "poor" shuffle sequence is better than extracting and inserting the elements one by one. Differential Revision: https://reviews.llvm.org/D24683 llvm-svn: 283480
2016-10-07 02:58:24 +08:00
; SSE2-NEXT: movdqa {{.*#+}} xmm4 = [65535,65535,0,65535,65535,0,65535,65535]
; SSE2-NEXT: movdqa %xmm4, %xmm5
; SSE2-NEXT: pandn %xmm2, %xmm5
; SSE2-NEXT: movdqa %xmm3, %xmm6
; SSE2-NEXT: pand %xmm4, %xmm6
; SSE2-NEXT: por %xmm5, %xmm6
; SSE2-NEXT: pshuflw {{.*#+}} xmm5 = xmm6[2,1,2,3,4,5,6,7]
; SSE2-NEXT: pshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,4,5,4,7]
; SSE2-NEXT: pshufd {{.*#+}} xmm5 = xmm5[0,3,2,3]
; SSE2-NEXT: pshuflw {{.*#+}} xmm5 = xmm5[1,2,3,0,4,5,6,7]
; SSE2-NEXT: pshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,5,5,6,7]
; SSE2-NEXT: movdqa {{.*#+}} xmm6 = [65535,65535,65535,65535,65535,0,0,0]
; SSE2-NEXT: pand %xmm6, %xmm5
; SSE2-NEXT: pshuflw {{.*#+}} xmm7 = xmm8[0,3,2,3,4,5,6,7]
; SSE2-NEXT: pshufd {{.*#+}} xmm7 = xmm7[0,1,0,3]
; SSE2-NEXT: pshufhw {{.*#+}} xmm7 = xmm7[0,1,2,3,4,4,5,6]
; SSE2-NEXT: movdqa %xmm6, %xmm0
; SSE2-NEXT: pandn %xmm7, %xmm0
; SSE2-NEXT: por %xmm5, %xmm0
; SSE2-NEXT: pand %xmm4, %xmm2
; SSE2-NEXT: pandn %xmm3, %xmm4
; SSE2-NEXT: por %xmm2, %xmm4
; SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm4[3,1,2,0]
; SSE2-NEXT: pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,6,5,6,7]
; SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm2[2,1,0,3]
; SSE2-NEXT: pshuflw {{.*#+}} xmm2 = xmm2[2,1,0,3,4,5,6,7]
; SSE2-NEXT: pand %xmm6, %xmm2
; SSE2-NEXT: pshufhw {{.*#+}} xmm3 = xmm8[0,1,2,3,4,7,6,7]
; SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm3[0,1,2,0]
; SSE2-NEXT: pshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,7,4,5]
; SSE2-NEXT: pandn %xmm3, %xmm6
; SSE2-NEXT: por %xmm2, %xmm6
; SSE2-NEXT: movups %xmm1, (%rsi)
; SSE2-NEXT: movdqu %xmm0, (%rdx)
; SSE2-NEXT: movdqu %xmm6, (%rcx)
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
; SSE2-NEXT: retq
;
; SSE42-LABEL: interleave_24i16_out:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; SSE42: # %bb.0:
[DAG] Generalize build_vector -> vector_shuffle combine for more than 2 inputs This generalizes the build_vector -> vector_shuffle combine to support any number of inputs. The idea is to create a binary tree of shuffles, where the first layer performs pairwise shuffles of the input vectors placing each input element into the correct lane, and the rest of the tree blends these shuffles together. This doesn't try to be smart and create any sort of "optimal" shuffles. The assumption is that even a "poor" shuffle sequence is better than extracting and inserting the elements one by one. Differential Revision: https://reviews.llvm.org/D24683 llvm-svn: 283480
2016-10-07 02:58:24 +08:00
; SSE42-NEXT: movdqu (%rdi), %xmm0
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
; SSE42-NEXT: movdqu 16(%rdi), %xmm1
[DAG] Generalize build_vector -> vector_shuffle combine for more than 2 inputs This generalizes the build_vector -> vector_shuffle combine to support any number of inputs. The idea is to create a binary tree of shuffles, where the first layer performs pairwise shuffles of the input vectors placing each input element into the correct lane, and the rest of the tree blends these shuffles together. This doesn't try to be smart and create any sort of "optimal" shuffles. The assumption is that even a "poor" shuffle sequence is better than extracting and inserting the elements one by one. Differential Revision: https://reviews.llvm.org/D24683 llvm-svn: 283480
2016-10-07 02:58:24 +08:00
; SSE42-NEXT: movdqu 32(%rdi), %xmm2
; SSE42-NEXT: pshufd {{.*#+}} xmm3 = xmm2[0,1,2,1]
; SSE42-NEXT: pshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,5,6,5]
; SSE42-NEXT: movdqa %xmm0, %xmm4
; SSE42-NEXT: pblendw {{.*#+}} xmm4 = xmm4[0],xmm1[1],xmm4[2,3],xmm1[4],xmm4[5,6],xmm1[7]
Recommit r354363 "[X86][SSE] Generalize X86ISD::BLENDI support to more value types" And its follow ups r354511, r354640. A follow patch will fix the issue that caused it to be reverted. llvm-svn: 354737
2019-02-24 05:41:42 +08:00
; SSE42-NEXT: pshufb {{.*#+}} xmm4 = xmm4[0,1,6,7,12,13,2,3,8,9,14,15,12,13,14,15]
[DAG] Generalize build_vector -> vector_shuffle combine for more than 2 inputs This generalizes the build_vector -> vector_shuffle combine to support any number of inputs. The idea is to create a binary tree of shuffles, where the first layer performs pairwise shuffles of the input vectors placing each input element into the correct lane, and the rest of the tree blends these shuffles together. This doesn't try to be smart and create any sort of "optimal" shuffles. The assumption is that even a "poor" shuffle sequence is better than extracting and inserting the elements one by one. Differential Revision: https://reviews.llvm.org/D24683 llvm-svn: 283480
2016-10-07 02:58:24 +08:00
; SSE42-NEXT: pblendw {{.*#+}} xmm4 = xmm4[0,1,2,3,4,5],xmm3[6,7]
; SSE42-NEXT: movdqa %xmm2, %xmm3
; SSE42-NEXT: pshufb {{.*#+}} xmm3 = xmm3[0,1,6,7,4,5,6,7,0,1,0,1,6,7,12,13]
; SSE42-NEXT: movdqa %xmm0, %xmm5
; SSE42-NEXT: pblendw {{.*#+}} xmm5 = xmm5[0,1],xmm1[2],xmm5[3,4],xmm1[5],xmm5[6,7]
; SSE42-NEXT: pshufb {{.*#+}} xmm5 = xmm5[2,3,8,9,14,15,4,5,10,11,10,11,8,9,14,15]
; SSE42-NEXT: pblendw {{.*#+}} xmm5 = xmm5[0,1,2,3,4],xmm3[5,6,7]
; SSE42-NEXT: pshufb {{.*#+}} xmm2 = xmm2[0,1,2,3,4,5,6,7,8,9,2,3,8,9,14,15]
; SSE42-NEXT: pblendw {{.*#+}} xmm1 = xmm1[0,1],xmm0[2],xmm1[3,4],xmm0[5],xmm1[6,7]
; SSE42-NEXT: pshufb {{.*#+}} xmm1 = xmm1[4,5,10,11,0,1,6,7,12,13,14,15,0,1,2,3]
; SSE42-NEXT: pblendw {{.*#+}} xmm1 = xmm1[0,1,2,3,4],xmm2[5,6,7]
; SSE42-NEXT: movdqu %xmm4, (%rsi)
; SSE42-NEXT: movdqu %xmm5, (%rdx)
; SSE42-NEXT: movdqu %xmm1, (%rcx)
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
; SSE42-NEXT: retq
;
; AVX1-LABEL: interleave_24i16_out:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/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] allow vector load narrowing with multi-use values This is a long-awaited follow-up suggested in D33578. Since then, we've picked up even more opportunities for vector narrowing from changes like D53784, so there are a lot of test diffs. Apart from 2-3 strange cases, these are all wins. I've structured this to be no-functional-change-intended for any target except for x86 because I couldn't tell if AArch64, ARM, and AMDGPU would improve or not. All of those targets have existing regression tests (4, 4, 10 files respectively) that would be affected. Also, Hexagon overrides the shouldReduceLoadWidth() hook, but doesn't show any regression test diffs. The trade-off is deciding if an extra vector load is better than a single wide load + extract_subvector. For x86, this is almost always better (on paper at least) because we often can fold loads into subsequent ops and not increase the official instruction count. There's also some unknown -- but potentially large -- benefit from using narrower vector ops if wide ops are implemented with multiple uops and/or frequency throttling is avoided. Differential Revision: https://reviews.llvm.org/D54073 llvm-svn: 346595
2018-11-11 04:05:31 +08:00
; AVX1-NEXT: vmovdqu (%rdi), %xmm0
; AVX1-NEXT: vmovdqu 16(%rdi), %xmm1
; AVX1-NEXT: vmovdqu 32(%rdi), %xmm2
; AVX1-NEXT: vpshufd {{.*#+}} xmm3 = xmm2[0,1,2,1]
; AVX1-NEXT: vpshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,5,6,5]
; AVX1-NEXT: vpblendw {{.*#+}} xmm4 = xmm0[0],xmm1[1],xmm0[2,3],xmm1[4],xmm0[5,6],xmm1[7]
Revert r356996 "[DAG] Avoid smart constructor-based dangling nodes." This patch appears to trigger very large compile time increases in halide builds. llvm-svn: 357116
2019-03-28 03:54:41 +08:00
; AVX1-NEXT: vpshufb {{.*#+}} xmm4 = xmm4[0,1,6,7,12,13,2,3,8,9,14,15,12,13,14,15]
[x86] allow vector load narrowing with multi-use values This is a long-awaited follow-up suggested in D33578. Since then, we've picked up even more opportunities for vector narrowing from changes like D53784, so there are a lot of test diffs. Apart from 2-3 strange cases, these are all wins. I've structured this to be no-functional-change-intended for any target except for x86 because I couldn't tell if AArch64, ARM, and AMDGPU would improve or not. All of those targets have existing regression tests (4, 4, 10 files respectively) that would be affected. Also, Hexagon overrides the shouldReduceLoadWidth() hook, but doesn't show any regression test diffs. The trade-off is deciding if an extra vector load is better than a single wide load + extract_subvector. For x86, this is almost always better (on paper at least) because we often can fold loads into subsequent ops and not increase the official instruction count. There's also some unknown -- but potentially large -- benefit from using narrower vector ops if wide ops are implemented with multiple uops and/or frequency throttling is avoided. Differential Revision: https://reviews.llvm.org/D54073 llvm-svn: 346595
2018-11-11 04:05:31 +08:00
; AVX1-NEXT: vpblendw {{.*#+}} xmm3 = xmm4[0,1,2,3,4,5],xmm3[6,7]
; AVX1-NEXT: vpshufb {{.*#+}} xmm4 = xmm2[0,1,6,7,4,5,6,7,0,1,0,1,6,7,12,13]
; AVX1-NEXT: vpblendw {{.*#+}} xmm5 = xmm0[0,1],xmm1[2],xmm0[3,4],xmm1[5],xmm0[6,7]
; AVX1-NEXT: vpshufb {{.*#+}} xmm5 = xmm5[2,3,8,9,14,15,4,5,10,11,u,u,u,u,u,u]
; AVX1-NEXT: vpblendw {{.*#+}} xmm4 = xmm5[0,1,2,3,4],xmm4[5,6,7]
; AVX1-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[0,1,2,3,4,5,6,7,8,9,2,3,8,9,14,15]
; AVX1-NEXT: vpblendw {{.*#+}} xmm0 = xmm1[0,1],xmm0[2],xmm1[3,4],xmm0[5],xmm1[6,7]
; AVX1-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[4,5,10,11,0,1,6,7,12,13,14,15,0,1,2,3]
; AVX1-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0,1,2,3,4],xmm2[5,6,7]
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
; AVX1-NEXT: vmovdqu %xmm3, (%rsi)
; AVX1-NEXT: vmovdqu %xmm4, (%rdx)
; AVX1-NEXT: vmovdqu %xmm0, (%rcx)
; AVX1-NEXT: retq
;
; AVX2-LABEL: interleave_24i16_out:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/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:
Revert rL311242 "Extension of shuffle vector pattern detection, updating post rebase." Summary: This reverts commit rL311242. Differential Revision: https://reviews.llvm.org/D36924 llvm-svn: 311246
2017-08-20 00:40:06 +08:00
; AVX2-NEXT: vmovdqu (%rdi), %ymm0
; AVX2-NEXT: vmovdqu 32(%rdi), %xmm1
; AVX2-NEXT: vpblendw {{.*#+}} ymm2 = ymm0[0,1],ymm1[2],ymm0[3,4],ymm1[5],ymm0[6,7,8,9],ymm1[10],ymm0[11,12],ymm1[13],ymm0[14,15]
[DAG] Allow build-to-shuffle combine to combine builds from two wide vectors. This allows us to, in some cases, create a vector_shuffle out of a build_vector, when the inputs to the build are extract_elements from two different vectors, at least one of which is wider than the output. (E.g. a <8 x i16> being constructed out of elements from a <16 x i16> and a <8 x i16>). Differential Revision: https://reviews.llvm.org/D24491 llvm-svn: 281402
2016-09-14 05:53:32 +08:00
; AVX2-NEXT: vextracti128 $1, %ymm2, %xmm3
; AVX2-NEXT: vpblendw {{.*#+}} xmm2 = xmm2[0],xmm3[1],xmm2[2,3],xmm3[4],xmm2[5,6],xmm3[7]
; AVX2-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[0,1,6,7,12,13,2,3,8,9,14,15,4,5,10,11]
Revert rL311242 "Extension of shuffle vector pattern detection, updating post rebase." Summary: This reverts commit rL311242. Differential Revision: https://reviews.llvm.org/D36924 llvm-svn: 311246
2017-08-20 00:40:06 +08:00
; AVX2-NEXT: vpblendw {{.*#+}} ymm3 = ymm1[0],ymm0[1,2],ymm1[3],ymm0[4,5],ymm1[6],ymm0[7],ymm1[8],ymm0[9,10],ymm1[11],ymm0[12,13],ymm1[14],ymm0[15]
[DAG] Allow build-to-shuffle combine to combine builds from two wide vectors. This allows us to, in some cases, create a vector_shuffle out of a build_vector, when the inputs to the build are extract_elements from two different vectors, at least one of which is wider than the output. (E.g. a <8 x i16> being constructed out of elements from a <16 x i16> and a <8 x i16>). Differential Revision: https://reviews.llvm.org/D24491 llvm-svn: 281402
2016-09-14 05:53:32 +08:00
; AVX2-NEXT: vextracti128 $1, %ymm3, %xmm4
; AVX2-NEXT: vpblendw {{.*#+}} xmm3 = xmm3[0,1],xmm4[2],xmm3[3,4],xmm4[5],xmm3[6,7]
; AVX2-NEXT: vpshufb {{.*#+}} xmm3 = xmm3[2,3,8,9,14,15,4,5,10,11,0,1,6,7,12,13]
Revert rL311242 "Extension of shuffle vector pattern detection, updating post rebase." Summary: This reverts commit rL311242. Differential Revision: https://reviews.llvm.org/D36924 llvm-svn: 311246
2017-08-20 00:40:06 +08:00
; AVX2-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2,3],ymm1[4],ymm0[5,6],ymm1[7],ymm0[8],ymm1[9],ymm0[10,11],ymm1[12],ymm0[13,14],ymm1[15]
[DAG] Allow build-to-shuffle combine to combine builds from two wide vectors. This allows us to, in some cases, create a vector_shuffle out of a build_vector, when the inputs to the build are extract_elements from two different vectors, at least one of which is wider than the output. (E.g. a <8 x i16> being constructed out of elements from a <16 x i16> and a <8 x i16>). Differential Revision: https://reviews.llvm.org/D24491 llvm-svn: 281402
2016-09-14 05:53:32 +08:00
; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
; AVX2-NEXT: vpblendw {{.*#+}} xmm0 = xmm1[0],xmm0[1,2],xmm1[3],xmm0[4,5],xmm1[6],xmm0[7]
; AVX2-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[4,5,10,11,0,1,6,7,12,13,2,3,8,9,14,15]
; AVX2-NEXT: vmovdqu %xmm2, (%rsi)
; AVX2-NEXT: vmovdqu %xmm3, (%rdx)
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
; AVX2-NEXT: vmovdqu %xmm0, (%rcx)
; AVX2-NEXT: vzeroupper
; AVX2-NEXT: retq
[X86][XOP] Add XOP oddshuffles tests XOP codegen is often different to generic AVX - thank you vpperm! llvm-svn: 315176
2017-10-08 20:58:15 +08:00
;
; XOP-LABEL: interleave_24i16_out:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; XOP: # %bb.0:
[x86] allow vector load narrowing with multi-use values This is a long-awaited follow-up suggested in D33578. Since then, we've picked up even more opportunities for vector narrowing from changes like D53784, so there are a lot of test diffs. Apart from 2-3 strange cases, these are all wins. I've structured this to be no-functional-change-intended for any target except for x86 because I couldn't tell if AArch64, ARM, and AMDGPU would improve or not. All of those targets have existing regression tests (4, 4, 10 files respectively) that would be affected. Also, Hexagon overrides the shouldReduceLoadWidth() hook, but doesn't show any regression test diffs. The trade-off is deciding if an extra vector load is better than a single wide load + extract_subvector. For x86, this is almost always better (on paper at least) because we often can fold loads into subsequent ops and not increase the official instruction count. There's also some unknown -- but potentially large -- benefit from using narrower vector ops if wide ops are implemented with multiple uops and/or frequency throttling is avoided. Differential Revision: https://reviews.llvm.org/D54073 llvm-svn: 346595
2018-11-11 04:05:31 +08:00
; XOP-NEXT: vmovdqu (%rdi), %xmm0
; XOP-NEXT: vmovdqu 16(%rdi), %xmm1
; XOP-NEXT: vmovdqu 32(%rdi), %xmm2
; XOP-NEXT: vpblendw {{.*#+}} xmm3 = xmm0[0],xmm1[1],xmm0[2,3],xmm1[4],xmm0[5,6],xmm1[7]
; XOP-NEXT: vpperm {{.*#+}} xmm3 = xmm3[0,1,6,7,12,13,2,3,8,9,14,15],xmm2[4,5,10,11]
; XOP-NEXT: vpblendw {{.*#+}} xmm4 = xmm0[0,1],xmm1[2],xmm0[3,4],xmm1[5],xmm0[6,7]
; XOP-NEXT: vpperm {{.*#+}} xmm4 = xmm4[2,3,8,9,14,15,4,5,10,11],xmm2[0,1,6,7,12,13]
; XOP-NEXT: vpperm {{.*#+}} xmm0 = xmm0[4,5,10,11],xmm1[0,1,6,7,12,13,14,15,0,1,2,3]
; XOP-NEXT: vpperm {{.*#+}} xmm0 = xmm0[0,1,2,3,4,5,6,7,8,9],xmm2[2,3,8,9,14,15]
[X86][XOP] Add XOP oddshuffles tests XOP codegen is often different to generic AVX - thank you vpperm! llvm-svn: 315176
2017-10-08 20:58:15 +08:00
; XOP-NEXT: vmovdqu %xmm3, (%rsi)
; XOP-NEXT: vmovdqu %xmm4, (%rdx)
; XOP-NEXT: vmovdqu %xmm0, (%rcx)
; XOP-NEXT: retq
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
%wide.vec = load <24 x i16>, <24 x i16>* %p, align 4
%s1 = shufflevector <24 x i16> %wide.vec, <24 x i16> undef, <8 x i32> <i32 0, i32 3, i32 6, i32 9, i32 12, i32 15, i32 18, i32 21>
%s2 = shufflevector <24 x i16> %wide.vec, <24 x i16> undef, <8 x i32> <i32 1, i32 4, i32 7, i32 10, i32 13, i32 16, i32 19, i32 22>
%s3 = shufflevector <24 x i16> %wide.vec, <24 x i16> undef, <8 x i32> <i32 2, i32 5, i32 8, i32 11, i32 14, i32 17, i32 20, i32 23>
store <8 x i16> %s1, <8 x i16>* %q1, align 4
store <8 x i16> %s2, <8 x i16>* %q2, align 4
store <8 x i16> %s3, <8 x i16>* %q3, align 4
ret void
}
define void @interleave_24i16_in(<24 x i16>* %p, <8 x i16>* %q1, <8 x i16>* %q2, <8 x i16>* %q3) nounwind {
; SSE2-LABEL: interleave_24i16_in:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/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:
[DAG] Generalize build_vector -> vector_shuffle combine for more than 2 inputs This generalizes the build_vector -> vector_shuffle combine to support any number of inputs. The idea is to create a binary tree of shuffles, where the first layer performs pairwise shuffles of the input vectors placing each input element into the correct lane, and the rest of the tree blends these shuffles together. This doesn't try to be smart and create any sort of "optimal" shuffles. The assumption is that even a "poor" shuffle sequence is better than extracting and inserting the elements one by one. Differential Revision: https://reviews.llvm.org/D24683 llvm-svn: 283480
2016-10-07 02:58:24 +08:00
; SSE2-NEXT: movdqu (%rsi), %xmm3
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
; SSE2-NEXT: movdqu (%rdx), %xmm2
; SSE2-NEXT: movdqu (%rcx), %xmm1
[DAG] Generalize build_vector -> vector_shuffle combine for more than 2 inputs This generalizes the build_vector -> vector_shuffle combine to support any number of inputs. The idea is to create a binary tree of shuffles, where the first layer performs pairwise shuffles of the input vectors placing each input element into the correct lane, and the rest of the tree blends these shuffles together. This doesn't try to be smart and create any sort of "optimal" shuffles. The assumption is that even a "poor" shuffle sequence is better than extracting and inserting the elements one by one. Differential Revision: https://reviews.llvm.org/D24683 llvm-svn: 283480
2016-10-07 02:58:24 +08:00
; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm1[0,0,0,3]
; SSE2-NEXT: movdqa {{.*#+}} xmm0 = [65535,65535,0,65535,65535,0,65535,65535]
; SSE2-NEXT: movdqa %xmm0, %xmm5
; SSE2-NEXT: pandn %xmm4, %xmm5
; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm3[0,3,3,3]
; SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm3[1,1,2,2]
; SSE2-NEXT: punpcklwd {{.*#+}} xmm3 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm3[0,1,2,1]
; SSE2-NEXT: pshuflw {{.*#+}} xmm3 = xmm3[0,1,2,2,4,5,6,7]
; SSE2-NEXT: pshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,7,5,4,5]
; SSE2-NEXT: pand %xmm0, %xmm3
; SSE2-NEXT: por %xmm5, %xmm3
; SSE2-NEXT: movdqa %xmm0, %xmm5
; SSE2-NEXT: pandn %xmm4, %xmm5
; SSE2-NEXT: pshuflw {{.*#+}} xmm4 = xmm2[0,1,3,3,4,5,6,7]
; SSE2-NEXT: punpckhwd {{.*#+}} xmm2 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
; SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm2[2,1,2,3]
; SSE2-NEXT: pshuflw {{.*#+}} xmm2 = xmm2[2,3,2,0,4,5,6,7]
; SSE2-NEXT: pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,5,5,6,7]
; SSE2-NEXT: pand %xmm0, %xmm2
; SSE2-NEXT: por %xmm5, %xmm2
; SSE2-NEXT: movdqa {{.*#+}} xmm5 = [65535,0,65535,65535,0,65535,65535,0]
; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,2,2]
; SSE2-NEXT: pand %xmm5, %xmm1
; SSE2-NEXT: pandn %xmm6, %xmm5
; SSE2-NEXT: por %xmm1, %xmm5
; SSE2-NEXT: pand %xmm0, %xmm5
; SSE2-NEXT: pshufhw {{.*#+}} xmm1 = xmm4[0,1,2,3,4,4,6,7]
; SSE2-NEXT: pandn %xmm1, %xmm0
; SSE2-NEXT: por %xmm5, %xmm0
; SSE2-NEXT: movdqu %xmm0, 16(%rdi)
; SSE2-NEXT: movdqu %xmm2, 32(%rdi)
; SSE2-NEXT: movdqu %xmm3, (%rdi)
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
; SSE2-NEXT: retq
;
; SSE42-LABEL: interleave_24i16_in:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; SSE42: # %bb.0:
[DAG] Generalize build_vector -> vector_shuffle combine for more than 2 inputs This generalizes the build_vector -> vector_shuffle combine to support any number of inputs. The idea is to create a binary tree of shuffles, where the first layer performs pairwise shuffles of the input vectors placing each input element into the correct lane, and the rest of the tree blends these shuffles together. This doesn't try to be smart and create any sort of "optimal" shuffles. The assumption is that even a "poor" shuffle sequence is better than extracting and inserting the elements one by one. Differential Revision: https://reviews.llvm.org/D24683 llvm-svn: 283480
2016-10-07 02:58:24 +08:00
; SSE42-NEXT: movdqu (%rsi), %xmm0
; SSE42-NEXT: movdqu (%rdx), %xmm1
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
; SSE42-NEXT: movdqu (%rcx), %xmm2
[DAG] Generalize build_vector -> vector_shuffle combine for more than 2 inputs This generalizes the build_vector -> vector_shuffle combine to support any number of inputs. The idea is to create a binary tree of shuffles, where the first layer performs pairwise shuffles of the input vectors placing each input element into the correct lane, and the rest of the tree blends these shuffles together. This doesn't try to be smart and create any sort of "optimal" shuffles. The assumption is that even a "poor" shuffle sequence is better than extracting and inserting the elements one by one. Differential Revision: https://reviews.llvm.org/D24683 llvm-svn: 283480
2016-10-07 02:58:24 +08:00
; SSE42-NEXT: pshufd {{.*#+}} xmm3 = xmm0[1,1,2,2]
; SSE42-NEXT: pshufd {{.*#+}} xmm4 = xmm0[0,3,3,3]
; SSE42-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
; SSE42-NEXT: pshufb {{.*#+}} xmm0 = xmm0[0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
; SSE42-NEXT: pshufd {{.*#+}} xmm5 = xmm2[0,0,0,3]
; SSE42-NEXT: pblendw {{.*#+}} xmm5 = xmm0[0,1],xmm5[2],xmm0[3,4],xmm5[5],xmm0[6,7]
; SSE42-NEXT: pshufd {{.*#+}} xmm0 = xmm2[1,1,2,2]
; SSE42-NEXT: pblendw {{.*#+}} xmm0 = xmm0[0],xmm3[1],xmm0[2,3],xmm3[4],xmm0[5,6],xmm3[7]
; SSE42-NEXT: pshuflw {{.*#+}} xmm3 = xmm1[0,1,3,3,4,5,6,7]
; SSE42-NEXT: pshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,4,6,7]
; SSE42-NEXT: pblendw {{.*#+}} xmm3 = xmm0[0,1],xmm3[2],xmm0[3,4],xmm3[5],xmm0[6,7]
; SSE42-NEXT: punpckhwd {{.*#+}} xmm1 = xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7]
; SSE42-NEXT: pshufb {{.*#+}} xmm1 = xmm1[4,5,6,7,4,5,8,9,10,11,10,11,12,13,14,15]
; SSE42-NEXT: pblendw {{.*#+}} xmm4 = xmm1[0,1],xmm4[2],xmm1[3,4],xmm4[5],xmm1[6,7]
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
; SSE42-NEXT: movdqu %xmm4, 32(%rdi)
[DAG] Generalize build_vector -> vector_shuffle combine for more than 2 inputs This generalizes the build_vector -> vector_shuffle combine to support any number of inputs. The idea is to create a binary tree of shuffles, where the first layer performs pairwise shuffles of the input vectors placing each input element into the correct lane, and the rest of the tree blends these shuffles together. This doesn't try to be smart and create any sort of "optimal" shuffles. The assumption is that even a "poor" shuffle sequence is better than extracting and inserting the elements one by one. Differential Revision: https://reviews.llvm.org/D24683 llvm-svn: 283480
2016-10-07 02:58:24 +08:00
; SSE42-NEXT: movdqu %xmm3, 16(%rdi)
; SSE42-NEXT: movdqu %xmm5, (%rdi)
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
; SSE42-NEXT: retq
;
; AVX1-LABEL: interleave_24i16_in:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/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] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
; AVX1-NEXT: vmovdqu (%rsi), %xmm0
; AVX1-NEXT: vmovdqu (%rdx), %xmm1
[X86][AVX] Add combineConcatVectors support to improve subvector handling Attempt to combine CONCAT_VECTORS nodes, which we only really have pre-legalization. This encourages a lot of X86ISD::SUBV_BROADCAST generation, so I've added SimplifyDemandedVectorEltsForTargetNode handling for this at the same time. The X86ISD::VTRUNC regression in shuffle-vs-trunc-256-widen.ll will be handled in a future commit. llvm-svn: 356064
2019-03-14 00:37:30 +08:00
; AVX1-NEXT: vpshufd {{.*#+}} xmm2 = xmm0[1,1,2,2]
; AVX1-NEXT: vpshuflw {{.*#+}} xmm3 = xmm1[0,1,3,3,4,5,6,7]
; AVX1-NEXT: vpshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,4,6,7]
; AVX1-NEXT: vpblendw {{.*#+}} xmm2 = xmm2[0,1],xmm3[2],xmm2[3,4],xmm3[5],xmm2[6,7]
; AVX1-NEXT: vmovdqu (%rcx), %xmm3
; AVX1-NEXT: vpshufd {{.*#+}} xmm4 = xmm3[1,1,2,2]
; AVX1-NEXT: vpblendw {{.*#+}} xmm2 = xmm4[0],xmm2[1,2],xmm4[3],xmm2[4,5],xmm4[6],xmm2[7]
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
; AVX1-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
[X86][SSE] SimplifyDemandedVectorEltsForTargetNode - simplify PSHUFB masks Attempt to simplify PSHUFB masks (even non-constant ones) - we should probably be able to simplify other variable shuffles as well as the need arises. llvm-svn: 343919
2018-10-06 21:49:31 +08:00
; AVX1-NEXT: vpshufb {{.*#+}} xmm4 = xmm4[0,1,2,3,u,u,4,5,6,7,u,u,8,9,10,11]
[X86][AVX] Add combineConcatVectors support to improve subvector handling Attempt to combine CONCAT_VECTORS nodes, which we only really have pre-legalization. This encourages a lot of X86ISD::SUBV_BROADCAST generation, so I've added SimplifyDemandedVectorEltsForTargetNode handling for this at the same time. The X86ISD::VTRUNC regression in shuffle-vs-trunc-256-widen.ll will be handled in a future commit. llvm-svn: 356064
2019-03-14 00:37:30 +08:00
; AVX1-NEXT: vpshufd {{.*#+}} xmm5 = xmm3[0,0,0,3]
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
; AVX1-NEXT: vpblendw {{.*#+}} xmm4 = xmm4[0,1],xmm5[2],xmm4[3,4],xmm5[5],xmm4[6,7]
[X86][AVX] Add combineConcatVectors support to improve subvector handling Attempt to combine CONCAT_VECTORS nodes, which we only really have pre-legalization. This encourages a lot of X86ISD::SUBV_BROADCAST generation, so I've added SimplifyDemandedVectorEltsForTargetNode handling for this at the same time. The X86ISD::VTRUNC regression in shuffle-vs-trunc-256-widen.ll will be handled in a future commit. llvm-svn: 356064
2019-03-14 00:37:30 +08:00
; AVX1-NEXT: vinsertf128 $1, %xmm2, %ymm4, %ymm2
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
; AVX1-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
; AVX1-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[4,5,10,11,10,11,8,9,8,9,14,15,12,13,14,15]
[X86][AVX] Add combineConcatVectors support to improve subvector handling Attempt to combine CONCAT_VECTORS nodes, which we only really have pre-legalization. This encourages a lot of X86ISD::SUBV_BROADCAST generation, so I've added SimplifyDemandedVectorEltsForTargetNode handling for this at the same time. The X86ISD::VTRUNC regression in shuffle-vs-trunc-256-widen.ll will be handled in a future commit. llvm-svn: 356064
2019-03-14 00:37:30 +08:00
; AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm3[2,2,3,3]
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
; AVX1-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2,3],xmm1[4],xmm0[5,6],xmm1[7]
; AVX1-NEXT: vmovdqu %xmm0, 32(%rdi)
[X86][AVX] Add combineConcatVectors support to improve subvector handling Attempt to combine CONCAT_VECTORS nodes, which we only really have pre-legalization. This encourages a lot of X86ISD::SUBV_BROADCAST generation, so I've added SimplifyDemandedVectorEltsForTargetNode handling for this at the same time. The X86ISD::VTRUNC regression in shuffle-vs-trunc-256-widen.ll will be handled in a future commit. llvm-svn: 356064
2019-03-14 00:37:30 +08:00
; AVX1-NEXT: vmovups %ymm2, (%rdi)
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
; AVX1-NEXT: vzeroupper
; AVX1-NEXT: retq
;
; AVX2-LABEL: interleave_24i16_in:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/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] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
; AVX2-NEXT: vmovdqu (%rsi), %xmm0
; AVX2-NEXT: vmovdqu (%rdx), %xmm1
[X86][AVX] Add combineConcatVectors support to improve subvector handling Attempt to combine CONCAT_VECTORS nodes, which we only really have pre-legalization. This encourages a lot of X86ISD::SUBV_BROADCAST generation, so I've added SimplifyDemandedVectorEltsForTargetNode handling for this at the same time. The X86ISD::VTRUNC regression in shuffle-vs-trunc-256-widen.ll will be handled in a future commit. llvm-svn: 356064
2019-03-14 00:37:30 +08:00
; AVX2-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm2
; AVX2-NEXT: vbroadcasti128 {{.*#+}} ymm3 = mem[0,1,0,1]
Revert r356996 "[DAG] Avoid smart constructor-based dangling nodes." This patch appears to trigger very large compile time increases in halide builds. llvm-svn: 357116
2019-03-28 03:54:41 +08:00
; AVX2-NEXT: vpshufb {{.*#+}} ymm4 = ymm2[0,1,2,3,6,7,2,3,8,9,8,9,4,5,6,7,16,17,18,19,22,23,18,19,24,25,24,25,20,21,22,23]
[X86][AVX] Add combineConcatVectors support to improve subvector handling Attempt to combine CONCAT_VECTORS nodes, which we only really have pre-legalization. This encourages a lot of X86ISD::SUBV_BROADCAST generation, so I've added SimplifyDemandedVectorEltsForTargetNode handling for this at the same time. The X86ISD::VTRUNC regression in shuffle-vs-trunc-256-widen.ll will be handled in a future commit. llvm-svn: 356064
2019-03-14 00:37:30 +08:00
; AVX2-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,3,0,1]
; AVX2-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[u,u,0,1,u,u,u,u,2,3,u,u,u,u,4,5,u,u,22,23,u,u,u,u,24,25,u,u,u,u,26,27]
; AVX2-NEXT: vpblendw {{.*#+}} ymm2 = ymm4[0],ymm2[1],ymm4[2,3],ymm2[4],ymm4[5,6],ymm2[7],ymm4[8],ymm2[9],ymm4[10,11],ymm2[12],ymm4[13,14],ymm2[15]
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
; AVX2-NEXT: vmovdqa {{.*#+}} ymm4 = <u,0,0,u,1,1,u,2>
[X86][AVX] Add combineConcatVectors support to improve subvector handling Attempt to combine CONCAT_VECTORS nodes, which we only really have pre-legalization. This encourages a lot of X86ISD::SUBV_BROADCAST generation, so I've added SimplifyDemandedVectorEltsForTargetNode handling for this at the same time. The X86ISD::VTRUNC regression in shuffle-vs-trunc-256-widen.ll will be handled in a future commit. llvm-svn: 356064
2019-03-14 00:37:30 +08:00
; AVX2-NEXT: vpermd %ymm3, %ymm4, %ymm4
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
; AVX2-NEXT: vmovdqa {{.*#+}} ymm5 = [255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255]
[X86][AVX] Add combineConcatVectors support to improve subvector handling Attempt to combine CONCAT_VECTORS nodes, which we only really have pre-legalization. This encourages a lot of X86ISD::SUBV_BROADCAST generation, so I've added SimplifyDemandedVectorEltsForTargetNode handling for this at the same time. The X86ISD::VTRUNC regression in shuffle-vs-trunc-256-widen.ll will be handled in a future commit. llvm-svn: 356064
2019-03-14 00:37:30 +08:00
; AVX2-NEXT: vpblendvb %ymm5, %ymm2, %ymm4, %ymm2
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
; AVX2-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
; AVX2-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[4,5,10,11,10,11,8,9,8,9,14,15,12,13,14,15]
[X86][AVX] Add combineConcatVectors support to improve subvector handling Attempt to combine CONCAT_VECTORS nodes, which we only really have pre-legalization. This encourages a lot of X86ISD::SUBV_BROADCAST generation, so I've added SimplifyDemandedVectorEltsForTargetNode handling for this at the same time. The X86ISD::VTRUNC regression in shuffle-vs-trunc-256-widen.ll will be handled in a future commit. llvm-svn: 356064
2019-03-14 00:37:30 +08:00
; AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm3[2,2,3,3]
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
; AVX2-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2,3],xmm1[4],xmm0[5,6],xmm1[7]
; AVX2-NEXT: vmovdqu %xmm0, 32(%rdi)
[X86][AVX] Add combineConcatVectors support to improve subvector handling Attempt to combine CONCAT_VECTORS nodes, which we only really have pre-legalization. This encourages a lot of X86ISD::SUBV_BROADCAST generation, so I've added SimplifyDemandedVectorEltsForTargetNode handling for this at the same time. The X86ISD::VTRUNC regression in shuffle-vs-trunc-256-widen.ll will be handled in a future commit. llvm-svn: 356064
2019-03-14 00:37:30 +08:00
; AVX2-NEXT: vmovdqu %ymm2, (%rdi)
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
; AVX2-NEXT: vzeroupper
; AVX2-NEXT: retq
[X86][XOP] Add XOP oddshuffles tests XOP codegen is often different to generic AVX - thank you vpperm! llvm-svn: 315176
2017-10-08 20:58:15 +08:00
;
; XOP-LABEL: interleave_24i16_in:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; XOP: # %bb.0:
[X86][XOP] Add XOP oddshuffles tests XOP codegen is often different to generic AVX - thank you vpperm! llvm-svn: 315176
2017-10-08 20:58:15 +08:00
; XOP-NEXT: vmovdqu (%rsi), %xmm0
; XOP-NEXT: vmovdqu (%rdx), %xmm1
; XOP-NEXT: vmovdqu (%rcx), %xmm2
[X86] SimplifyDemandedVectorElts - add VPPERM support We need to add support for all variable shuffle mask ops, but VPPERM is the only one that already has test coverage. llvm-svn: 358165
2019-04-11 21:30:38 +08:00
; XOP-NEXT: vpperm {{.*#+}} xmm3 = xmm0[u,u,6,7],xmm1[6,7],xmm0[u,u,8,9],xmm1[8,9],xmm0[u,u,10,11]
[X86][XOP] Add XOP oddshuffles tests XOP codegen is often different to generic AVX - thank you vpperm! llvm-svn: 315176
2017-10-08 20:58:15 +08:00
; XOP-NEXT: vpshufd {{.*#+}} xmm4 = xmm2[1,1,2,2]
; XOP-NEXT: vpblendw {{.*#+}} xmm3 = xmm4[0],xmm3[1,2],xmm4[3],xmm3[4,5],xmm4[6],xmm3[7]
; XOP-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
; XOP-NEXT: vpperm {{.*#+}} xmm4 = xmm4[0,1,2,3],xmm2[0,1],xmm4[4,5,6,7],xmm2[2,3],xmm4[8,9,10,11]
; XOP-NEXT: vinsertf128 $1, %xmm3, %ymm4, %ymm3
; XOP-NEXT: vpperm {{.*#+}} xmm0 = xmm1[10,11],xmm0[12,13,12,13],xmm1[12,13,12,13],xmm0[14,15],xmm1[14,15],xmm0[14,15]
; XOP-NEXT: vpshufd {{.*#+}} xmm1 = xmm2[2,2,3,3]
; XOP-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2,3],xmm1[4],xmm0[5,6],xmm1[7]
; XOP-NEXT: vmovdqu %xmm0, 32(%rdi)
; XOP-NEXT: vmovups %ymm3, (%rdi)
; XOP-NEXT: vzeroupper
; XOP-NEXT: retq
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
%s1 = load <8 x i16>, <8 x i16>* %q1, align 4
%s2 = load <8 x i16>, <8 x i16>* %q2, align 4
%s3 = load <8 x i16>, <8 x i16>* %q3, align 4
%t1 = shufflevector <8 x i16> %s1, <8 x i16> %s2, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
%t2 = shufflevector <8 x i16> %s3, <8 x i16> undef, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
%interleaved = shufflevector <16 x i16> %t1, <16 x i16> %t2, <24 x i32> <i32 0, i32 8, i32 16, i32 1, i32 9, i32 17, i32 2, i32 10, i32 18, i32 3, i32 11, i32 19, i32 4, i32 12, i32 20, i32 5, i32 13, i32 21, i32 6, i32 14, i32 22, i32 7, i32 15, i32 23>
store <24 x i16> %interleaved, <24 x i16>* %p, align 4
ret void
}
define void @interleave_24i32_out(<24 x i32>* %p, <8 x i32>* %q1, <8 x i32>* %q2, <8 x i32>* %q3) nounwind {
; SSE2-LABEL: interleave_24i32_out:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/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] Add initial SimplifyDemandedVectorEltsForTargetNode support This patch adds an initial x86 SimplifyDemandedVectorEltsForTargetNode implementation to handle target shuffles. Currently the patch only decodes a target shuffle, calls SimplifyDemandedVectorElts on its input operands and removes any shuffle that reduces to undef/zero/identity. Future work will need to integrate this with combineX86ShufflesRecursively, add support for other x86 ops, etc. NOTE: There is a minor regression that appears to be affecting further (extractelement?) combines which I haven't been able to solve yet - possibly something to do with how nodes are added to the worklist after simplification. Differential Revision: https://reviews.llvm.org/D52140 llvm-svn: 342564
2018-09-20 02:11:34 +08:00
; SSE2-NEXT: movdqu 64(%rdi), %xmm10
[X86][SSE] Combine (some) target shuffles with multiple uses As discussed on D41794, we have many cases where we fail to combine shuffles as the input operands have other uses. This patch permits these shuffles to be combined as long as they don't introduce additional variable shuffle masks, which should reduce instruction dependencies and allow the total number of shuffles to still drop without increasing the constant pool. However, this may mean that some memory folds may no longer occur, and on pre-AVX require the occasional extra register move. This also exposes some poor PMULDQ/PMULUDQ codegen which was doing unnecessary upper/lower calculations which will in fact fold to zero/undef - the fix will be added in a followup commit. Differential Revision: https://reviews.llvm.org/D50328 llvm-svn: 339335
2018-08-09 20:30:02 +08:00
; SSE2-NEXT: movups 80(%rdi), %xmm8
[X86] Add initial SimplifyDemandedVectorEltsForTargetNode support This patch adds an initial x86 SimplifyDemandedVectorEltsForTargetNode implementation to handle target shuffles. Currently the patch only decodes a target shuffle, calls SimplifyDemandedVectorElts on its input operands and removes any shuffle that reduces to undef/zero/identity. Future work will need to integrate this with combineX86ShufflesRecursively, add support for other x86 ops, etc. NOTE: There is a minor regression that appears to be affecting further (extractelement?) combines which I haven't been able to solve yet - possibly something to do with how nodes are added to the worklist after simplification. Differential Revision: https://reviews.llvm.org/D52140 llvm-svn: 342564
2018-09-20 02:11:34 +08:00
; SSE2-NEXT: movdqu (%rdi), %xmm0
; SSE2-NEXT: movdqu 16(%rdi), %xmm11
; SSE2-NEXT: movups 32(%rdi), %xmm5
; SSE2-NEXT: movdqu 48(%rdi), %xmm9
; SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm0[2,3,0,1]
; SSE2-NEXT: movaps %xmm5, %xmm7
; SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm11[1,1,2,3]
; SSE2-NEXT: punpckldq {{.*#+}} xmm6 = xmm6[0],xmm3[0],xmm6[1],xmm3[1]
; SSE2-NEXT: shufps {{.*#+}} xmm6 = xmm6[0,1],xmm5[0,3]
[X86] SimplifyDemandedVectorEltsForTargetNode - remove identity target shuffles before simplifying inputs By removing demanded target shuffles that simplify to zero/undef/identity before simplifying its inputs we improve chances of further simplification, as only the immediate parent user of the combined is added back to the work list - this still doesn't help us if its passed through other ops though (bitcasts....). llvm-svn: 343390
2018-09-30 02:15:26 +08:00
; SSE2-NEXT: shufps {{.*#+}} xmm5 = xmm5[1,0],xmm11[2,0]
[X86] Add initial SimplifyDemandedVectorEltsForTargetNode support This patch adds an initial x86 SimplifyDemandedVectorEltsForTargetNode implementation to handle target shuffles. Currently the patch only decodes a target shuffle, calls SimplifyDemandedVectorElts on its input operands and removes any shuffle that reduces to undef/zero/identity. Future work will need to integrate this with combineX86ShufflesRecursively, add support for other x86 ops, etc. NOTE: There is a minor regression that appears to be affecting further (extractelement?) combines which I haven't been able to solve yet - possibly something to do with how nodes are added to the worklist after simplification. Differential Revision: https://reviews.llvm.org/D52140 llvm-svn: 342564
2018-09-20 02:11:34 +08:00
; SSE2-NEXT: movdqa %xmm0, %xmm3
; SSE2-NEXT: shufps {{.*#+}} xmm3 = xmm3[0,3],xmm5[2,0]
[X86][SSE] Combine (some) target shuffles with multiple uses As discussed on D41794, we have many cases where we fail to combine shuffles as the input operands have other uses. This patch permits these shuffles to be combined as long as they don't introduce additional variable shuffle masks, which should reduce instruction dependencies and allow the total number of shuffles to still drop without increasing the constant pool. However, this may mean that some memory folds may no longer occur, and on pre-AVX require the occasional extra register move. This also exposes some poor PMULDQ/PMULUDQ codegen which was doing unnecessary upper/lower calculations which will in fact fold to zero/undef - the fix will be added in a followup commit. Differential Revision: https://reviews.llvm.org/D50328 llvm-svn: 339335
2018-08-09 20:30:02 +08:00
; SSE2-NEXT: movaps %xmm8, %xmm5
[X86] Add initial SimplifyDemandedVectorEltsForTargetNode support This patch adds an initial x86 SimplifyDemandedVectorEltsForTargetNode implementation to handle target shuffles. Currently the patch only decodes a target shuffle, calls SimplifyDemandedVectorElts on its input operands and removes any shuffle that reduces to undef/zero/identity. Future work will need to integrate this with combineX86ShufflesRecursively, add support for other x86 ops, etc. NOTE: There is a minor regression that appears to be affecting further (extractelement?) combines which I haven't been able to solve yet - possibly something to do with how nodes are added to the worklist after simplification. Differential Revision: https://reviews.llvm.org/D52140 llvm-svn: 342564
2018-09-20 02:11:34 +08:00
; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm10[1,1,2,3]
; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm9[2,3,0,1]
; SSE2-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm4[0],xmm1[1],xmm4[1]
; SSE2-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,1],xmm8[0,3]
[X86] SimplifyDemandedVectorEltsForTargetNode - remove identity target shuffles before simplifying inputs By removing demanded target shuffles that simplify to zero/undef/identity before simplifying its inputs we improve chances of further simplification, as only the immediate parent user of the combined is added back to the work list - this still doesn't help us if its passed through other ops though (bitcasts....). llvm-svn: 343390
2018-09-30 02:15:26 +08:00
; SSE2-NEXT: shufps {{.*#+}} xmm8 = xmm8[1,0],xmm10[2,0]
[X86] Add initial SimplifyDemandedVectorEltsForTargetNode support This patch adds an initial x86 SimplifyDemandedVectorEltsForTargetNode implementation to handle target shuffles. Currently the patch only decodes a target shuffle, calls SimplifyDemandedVectorElts on its input operands and removes any shuffle that reduces to undef/zero/identity. Future work will need to integrate this with combineX86ShufflesRecursively, add support for other x86 ops, etc. NOTE: There is a minor regression that appears to be affecting further (extractelement?) combines which I haven't been able to solve yet - possibly something to do with how nodes are added to the worklist after simplification. Differential Revision: https://reviews.llvm.org/D52140 llvm-svn: 342564
2018-09-20 02:11:34 +08:00
; SSE2-NEXT: movdqa %xmm9, %xmm2
; SSE2-NEXT: shufps {{.*#+}} xmm9 = xmm9[0,3],xmm8[2,0]
; SSE2-NEXT: shufps {{.*#+}} xmm2 = xmm2[1,0],xmm10[0,0]
; SSE2-NEXT: movaps %xmm2, %xmm4
; SSE2-NEXT: shufps {{.*#+}} xmm4 = xmm4[0,2],xmm10[3,3]
; SSE2-NEXT: shufps {{.*#+}} xmm5 = xmm5[2,0],xmm4[2,0]
; SSE2-NEXT: shufps {{.*#+}} xmm2 = xmm2[0,2],xmm5[2,0]
; SSE2-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,0],xmm11[0,0]
; SSE2-NEXT: movaps %xmm0, %xmm4
; SSE2-NEXT: shufps {{.*#+}} xmm4 = xmm4[0,2],xmm11[3,3]
; SSE2-NEXT: shufps {{.*#+}} xmm7 = xmm7[2,0],xmm4[2,0]
; SSE2-NEXT: shufps {{.*#+}} xmm0 = xmm0[0,2],xmm7[2,0]
; SSE2-NEXT: movups %xmm9, 16(%rsi)
[X86][SSE] Combine (some) target shuffles with multiple uses As discussed on D41794, we have many cases where we fail to combine shuffles as the input operands have other uses. This patch permits these shuffles to be combined as long as they don't introduce additional variable shuffle masks, which should reduce instruction dependencies and allow the total number of shuffles to still drop without increasing the constant pool. However, this may mean that some memory folds may no longer occur, and on pre-AVX require the occasional extra register move. This also exposes some poor PMULDQ/PMULUDQ codegen which was doing unnecessary upper/lower calculations which will in fact fold to zero/undef - the fix will be added in a followup commit. Differential Revision: https://reviews.llvm.org/D50328 llvm-svn: 339335
2018-08-09 20:30:02 +08:00
; SSE2-NEXT: movups %xmm3, (%rsi)
[X86] Add initial SimplifyDemandedVectorEltsForTargetNode support This patch adds an initial x86 SimplifyDemandedVectorEltsForTargetNode implementation to handle target shuffles. Currently the patch only decodes a target shuffle, calls SimplifyDemandedVectorElts on its input operands and removes any shuffle that reduces to undef/zero/identity. Future work will need to integrate this with combineX86ShufflesRecursively, add support for other x86 ops, etc. NOTE: There is a minor regression that appears to be affecting further (extractelement?) combines which I haven't been able to solve yet - possibly something to do with how nodes are added to the worklist after simplification. Differential Revision: https://reviews.llvm.org/D52140 llvm-svn: 342564
2018-09-20 02:11:34 +08:00
; SSE2-NEXT: movups %xmm2, 16(%rdx)
[DAG] Generalize build_vector -> vector_shuffle combine for more than 2 inputs This generalizes the build_vector -> vector_shuffle combine to support any number of inputs. The idea is to create a binary tree of shuffles, where the first layer performs pairwise shuffles of the input vectors placing each input element into the correct lane, and the rest of the tree blends these shuffles together. This doesn't try to be smart and create any sort of "optimal" shuffles. The assumption is that even a "poor" shuffle sequence is better than extracting and inserting the elements one by one. Differential Revision: https://reviews.llvm.org/D24683 llvm-svn: 283480
2016-10-07 02:58:24 +08:00
; SSE2-NEXT: movups %xmm0, (%rdx)
[X86] Add initial SimplifyDemandedVectorEltsForTargetNode support This patch adds an initial x86 SimplifyDemandedVectorEltsForTargetNode implementation to handle target shuffles. Currently the patch only decodes a target shuffle, calls SimplifyDemandedVectorElts on its input operands and removes any shuffle that reduces to undef/zero/identity. Future work will need to integrate this with combineX86ShufflesRecursively, add support for other x86 ops, etc. NOTE: There is a minor regression that appears to be affecting further (extractelement?) combines which I haven't been able to solve yet - possibly something to do with how nodes are added to the worklist after simplification. Differential Revision: https://reviews.llvm.org/D52140 llvm-svn: 342564
2018-09-20 02:11:34 +08:00
; SSE2-NEXT: movups %xmm1, 16(%rcx)
; SSE2-NEXT: movups %xmm6, (%rcx)
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
; SSE2-NEXT: retq
;
; SSE42-LABEL: interleave_24i32_out:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; SSE42: # %bb.0:
[x86] use a single shufps when it can save instructions This is a tiny patch with a big pile of test changes. This partially fixes PR27885: https://llvm.org/bugs/show_bug.cgi?id=27885 My motivating case looks like this: - vpshufd {{.*#+}} xmm1 = xmm1[0,1,0,2] - vpshufd {{.*#+}} xmm0 = xmm0[0,2,2,3] - vpblendw {{.*#+}} xmm0 = xmm0[0,1,2,3],xmm1[4,5,6,7] + vshufps {{.*#+}} xmm0 = xmm0[0,2],xmm1[0,2] And this happens several times in the diffs. For chips with domain-crossing penalties, the instruction count and size reduction should usually overcome any potential domain-crossing penalty due to using an FP op in a sequence of int ops. For chips such as recent Intel big cores and Atom, there is no domain-crossing penalty for shufps, so using shufps is a pure win. So the test case diffs all appear to be improvements except one test in vector-shuffle-combining.ll where we miss an opportunity to use a shift to generate zero elements and one test in combine-sra.ll where multiple uses prevent the expected shuffle combining. Differential Revision: https://reviews.llvm.org/D27692 llvm-svn: 289837
2016-12-16 02:03:38 +08:00
; SSE42-NEXT: movdqu 80(%rdi), %xmm9
; SSE42-NEXT: movdqu 64(%rdi), %xmm10
; SSE42-NEXT: movdqu (%rdi), %xmm4
; SSE42-NEXT: movdqu 16(%rdi), %xmm2
; SSE42-NEXT: movdqu 32(%rdi), %xmm11
; SSE42-NEXT: movdqu 48(%rdi), %xmm5
; SSE42-NEXT: pshufd {{.*#+}} xmm8 = xmm11[0,1,0,1]
; SSE42-NEXT: movdqa %xmm2, %xmm7
; SSE42-NEXT: pblendw {{.*#+}} xmm7 = xmm7[0,1],xmm4[2,3],xmm7[4,5,6,7]
[DAG] Generalize build_vector -> vector_shuffle combine for more than 2 inputs This generalizes the build_vector -> vector_shuffle combine to support any number of inputs. The idea is to create a binary tree of shuffles, where the first layer performs pairwise shuffles of the input vectors placing each input element into the correct lane, and the rest of the tree blends these shuffles together. This doesn't try to be smart and create any sort of "optimal" shuffles. The assumption is that even a "poor" shuffle sequence is better than extracting and inserting the elements one by one. Differential Revision: https://reviews.llvm.org/D24683 llvm-svn: 283480
2016-10-07 02:58:24 +08:00
; SSE42-NEXT: pshufd {{.*#+}} xmm6 = xmm4[2,3,0,1]
[x86] use a single shufps when it can save instructions This is a tiny patch with a big pile of test changes. This partially fixes PR27885: https://llvm.org/bugs/show_bug.cgi?id=27885 My motivating case looks like this: - vpshufd {{.*#+}} xmm1 = xmm1[0,1,0,2] - vpshufd {{.*#+}} xmm0 = xmm0[0,2,2,3] - vpblendw {{.*#+}} xmm0 = xmm0[0,1,2,3],xmm1[4,5,6,7] + vshufps {{.*#+}} xmm0 = xmm0[0,2],xmm1[0,2] And this happens several times in the diffs. For chips with domain-crossing penalties, the instruction count and size reduction should usually overcome any potential domain-crossing penalty due to using an FP op in a sequence of int ops. For chips such as recent Intel big cores and Atom, there is no domain-crossing penalty for shufps, so using shufps is a pure win. So the test case diffs all appear to be improvements except one test in vector-shuffle-combining.ll where we miss an opportunity to use a shift to generate zero elements and one test in combine-sra.ll where multiple uses prevent the expected shuffle combining. Differential Revision: https://reviews.llvm.org/D27692 llvm-svn: 289837
2016-12-16 02:03:38 +08:00
; SSE42-NEXT: shufps {{.*#+}} xmm4 = xmm4[0,3],xmm2[2,3]
[X86][SSE] Add custom execution domain fixing for BLENDPD/BLENDPS/PBLENDD/PBLENDW (PR34873) Add support for custom execution domain fixing and implement support for BLENDPD/BLENDPS/PBLENDD/PBLENDW. Differential Revision: https://reviews.llvm.org/D42042 llvm-svn: 322524
2018-01-16 06:18:45 +08:00
; SSE42-NEXT: blendps {{.*#+}} xmm4 = xmm4[0,1,2],xmm8[3]
[x86] use a single shufps when it can save instructions This is a tiny patch with a big pile of test changes. This partially fixes PR27885: https://llvm.org/bugs/show_bug.cgi?id=27885 My motivating case looks like this: - vpshufd {{.*#+}} xmm1 = xmm1[0,1,0,2] - vpshufd {{.*#+}} xmm0 = xmm0[0,2,2,3] - vpblendw {{.*#+}} xmm0 = xmm0[0,1,2,3],xmm1[4,5,6,7] + vshufps {{.*#+}} xmm0 = xmm0[0,2],xmm1[0,2] And this happens several times in the diffs. For chips with domain-crossing penalties, the instruction count and size reduction should usually overcome any potential domain-crossing penalty due to using an FP op in a sequence of int ops. For chips such as recent Intel big cores and Atom, there is no domain-crossing penalty for shufps, so using shufps is a pure win. So the test case diffs all appear to be improvements except one test in vector-shuffle-combining.ll where we miss an opportunity to use a shift to generate zero elements and one test in combine-sra.ll where multiple uses prevent the expected shuffle combining. Differential Revision: https://reviews.llvm.org/D27692 llvm-svn: 289837
2016-12-16 02:03:38 +08:00
; SSE42-NEXT: movdqa %xmm10, %xmm1
; SSE42-NEXT: pblendw {{.*#+}} xmm1 = xmm1[0,1],xmm5[2,3],xmm1[4,5,6,7]
; SSE42-NEXT: pshufd {{.*#+}} xmm0 = xmm5[2,3,0,1]
; SSE42-NEXT: shufps {{.*#+}} xmm5 = xmm5[0,3],xmm10[2,3]
; SSE42-NEXT: pshufd {{.*#+}} xmm3 = xmm9[0,1,0,1]
; SSE42-NEXT: pblendw {{.*#+}} xmm3 = xmm5[0,1,2,3,4,5],xmm3[6,7]
; SSE42-NEXT: pshufd {{.*#+}} xmm5 = xmm11[0,1,2,2]
; SSE42-NEXT: pshufd {{.*#+}} xmm7 = xmm7[1,0,3,3]
; SSE42-NEXT: pblendw {{.*#+}} xmm7 = xmm7[0,1,2,3,4,5],xmm5[6,7]
[DAG] Generalize build_vector -> vector_shuffle combine for more than 2 inputs This generalizes the build_vector -> vector_shuffle combine to support any number of inputs. The idea is to create a binary tree of shuffles, where the first layer performs pairwise shuffles of the input vectors placing each input element into the correct lane, and the rest of the tree blends these shuffles together. This doesn't try to be smart and create any sort of "optimal" shuffles. The assumption is that even a "poor" shuffle sequence is better than extracting and inserting the elements one by one. Differential Revision: https://reviews.llvm.org/D24683 llvm-svn: 283480
2016-10-07 02:58:24 +08:00
; SSE42-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,0,3,3]
[x86] use a single shufps when it can save instructions This is a tiny patch with a big pile of test changes. This partially fixes PR27885: https://llvm.org/bugs/show_bug.cgi?id=27885 My motivating case looks like this: - vpshufd {{.*#+}} xmm1 = xmm1[0,1,0,2] - vpshufd {{.*#+}} xmm0 = xmm0[0,2,2,3] - vpblendw {{.*#+}} xmm0 = xmm0[0,1,2,3],xmm1[4,5,6,7] + vshufps {{.*#+}} xmm0 = xmm0[0,2],xmm1[0,2] And this happens several times in the diffs. For chips with domain-crossing penalties, the instruction count and size reduction should usually overcome any potential domain-crossing penalty due to using an FP op in a sequence of int ops. For chips such as recent Intel big cores and Atom, there is no domain-crossing penalty for shufps, so using shufps is a pure win. So the test case diffs all appear to be improvements except one test in vector-shuffle-combining.ll where we miss an opportunity to use a shift to generate zero elements and one test in combine-sra.ll where multiple uses prevent the expected shuffle combining. Differential Revision: https://reviews.llvm.org/D27692 llvm-svn: 289837
2016-12-16 02:03:38 +08:00
; SSE42-NEXT: pshufd {{.*#+}} xmm5 = xmm9[0,1,2,2]
; SSE42-NEXT: pblendw {{.*#+}} xmm5 = xmm1[0,1,2,3,4,5],xmm5[6,7]
; SSE42-NEXT: pblendw {{.*#+}} xmm6 = xmm6[0,1],xmm2[2,3],xmm6[4,5,6,7]
; SSE42-NEXT: pshufd {{.*#+}} xmm1 = xmm11[0,1,0,3]
; SSE42-NEXT: pblendw {{.*#+}} xmm1 = xmm6[0,1,2,3],xmm1[4,5,6,7]
; SSE42-NEXT: pblendw {{.*#+}} xmm0 = xmm0[0,1],xmm10[2,3],xmm0[4,5,6,7]
; SSE42-NEXT: pshufd {{.*#+}} xmm2 = xmm9[0,1,0,3]
; SSE42-NEXT: pblendw {{.*#+}} xmm2 = xmm0[0,1,2,3],xmm2[4,5,6,7]
; SSE42-NEXT: movdqu %xmm3, 16(%rsi)
[X86][SSE] Add custom execution domain fixing for BLENDPD/BLENDPS/PBLENDD/PBLENDW (PR34873) Add support for custom execution domain fixing and implement support for BLENDPD/BLENDPS/PBLENDD/PBLENDW. Differential Revision: https://reviews.llvm.org/D42042 llvm-svn: 322524
2018-01-16 06:18:45 +08:00
; SSE42-NEXT: movups %xmm4, (%rsi)
[x86] use a single shufps when it can save instructions This is a tiny patch with a big pile of test changes. This partially fixes PR27885: https://llvm.org/bugs/show_bug.cgi?id=27885 My motivating case looks like this: - vpshufd {{.*#+}} xmm1 = xmm1[0,1,0,2] - vpshufd {{.*#+}} xmm0 = xmm0[0,2,2,3] - vpblendw {{.*#+}} xmm0 = xmm0[0,1,2,3],xmm1[4,5,6,7] + vshufps {{.*#+}} xmm0 = xmm0[0,2],xmm1[0,2] And this happens several times in the diffs. For chips with domain-crossing penalties, the instruction count and size reduction should usually overcome any potential domain-crossing penalty due to using an FP op in a sequence of int ops. For chips such as recent Intel big cores and Atom, there is no domain-crossing penalty for shufps, so using shufps is a pure win. So the test case diffs all appear to be improvements except one test in vector-shuffle-combining.ll where we miss an opportunity to use a shift to generate zero elements and one test in combine-sra.ll where multiple uses prevent the expected shuffle combining. Differential Revision: https://reviews.llvm.org/D27692 llvm-svn: 289837
2016-12-16 02:03:38 +08:00
; SSE42-NEXT: movdqu %xmm5, 16(%rdx)
; SSE42-NEXT: movdqu %xmm7, (%rdx)
[DAG] Generalize build_vector -> vector_shuffle combine for more than 2 inputs This generalizes the build_vector -> vector_shuffle combine to support any number of inputs. The idea is to create a binary tree of shuffles, where the first layer performs pairwise shuffles of the input vectors placing each input element into the correct lane, and the rest of the tree blends these shuffles together. This doesn't try to be smart and create any sort of "optimal" shuffles. The assumption is that even a "poor" shuffle sequence is better than extracting and inserting the elements one by one. Differential Revision: https://reviews.llvm.org/D24683 llvm-svn: 283480
2016-10-07 02:58:24 +08:00
; SSE42-NEXT: movdqu %xmm2, 16(%rcx)
; SSE42-NEXT: movdqu %xmm1, (%rcx)
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
; SSE42-NEXT: retq
;
; AVX1-LABEL: interleave_24i32_out:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/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:
[DAG] Generalize build_vector -> vector_shuffle combine for more than 2 inputs This generalizes the build_vector -> vector_shuffle combine to support any number of inputs. The idea is to create a binary tree of shuffles, where the first layer performs pairwise shuffles of the input vectors placing each input element into the correct lane, and the rest of the tree blends these shuffles together. This doesn't try to be smart and create any sort of "optimal" shuffles. The assumption is that even a "poor" shuffle sequence is better than extracting and inserting the elements one by one. Differential Revision: https://reviews.llvm.org/D24683 llvm-svn: 283480
2016-10-07 02:58:24 +08:00
; AVX1-NEXT: vmovups (%rdi), %ymm0
; AVX1-NEXT: vmovups 32(%rdi), %ymm1
[x86] allow vector load narrowing with multi-use values This is a long-awaited follow-up suggested in D33578. Since then, we've picked up even more opportunities for vector narrowing from changes like D53784, so there are a lot of test diffs. Apart from 2-3 strange cases, these are all wins. I've structured this to be no-functional-change-intended for any target except for x86 because I couldn't tell if AArch64, ARM, and AMDGPU would improve or not. All of those targets have existing regression tests (4, 4, 10 files respectively) that would be affected. Also, Hexagon overrides the shouldReduceLoadWidth() hook, but doesn't show any regression test diffs. The trade-off is deciding if an extra vector load is better than a single wide load + extract_subvector. For x86, this is almost always better (on paper at least) because we often can fold loads into subsequent ops and not increase the official instruction count. There's also some unknown -- but potentially large -- benefit from using narrower vector ops if wide ops are implemented with multiple uops and/or frequency throttling is avoided. Differential Revision: https://reviews.llvm.org/D54073 llvm-svn: 346595
2018-11-11 04:05:31 +08:00
; AVX1-NEXT: vmovups 80(%rdi), %xmm2
; AVX1-NEXT: vmovups 64(%rdi), %xmm3
; AVX1-NEXT: vinsertps {{.*#+}} xmm4 = zero,zero,xmm3[2],xmm2[1]
[DAG] Generalize build_vector -> vector_shuffle combine for more than 2 inputs This generalizes the build_vector -> vector_shuffle combine to support any number of inputs. The idea is to create a binary tree of shuffles, where the first layer performs pairwise shuffles of the input vectors placing each input element into the correct lane, and the rest of the tree blends these shuffles together. This doesn't try to be smart and create any sort of "optimal" shuffles. The assumption is that even a "poor" shuffle sequence is better than extracting and inserting the elements one by one. Differential Revision: https://reviews.llvm.org/D24683 llvm-svn: 283480
2016-10-07 02:58:24 +08:00
; AVX1-NEXT: vinsertf128 $1, %xmm4, %ymm0, %ymm4
; AVX1-NEXT: vblendps {{.*#+}} ymm5 = ymm0[0],ymm1[1],ymm0[2,3],ymm1[4],ymm0[5,6],ymm1[7]
; AVX1-NEXT: vextractf128 $1, %ymm5, %xmm6
; AVX1-NEXT: vblendps {{.*#+}} xmm5 = xmm5[0,1],xmm6[2],xmm5[3]
; AVX1-NEXT: vpermilps {{.*#+}} xmm5 = xmm5[0,3,2,1]
; AVX1-NEXT: vpermilps {{.*#+}} xmm6 = xmm6[0,3,2,3]
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
; AVX1-NEXT: vinsertf128 $1, %xmm6, %ymm5, %ymm5
[X86][SSE] Add custom execution domain fixing for BLENDPD/BLENDPS/PBLENDD/PBLENDW (PR34873) Add support for custom execution domain fixing and implement support for BLENDPD/BLENDPS/PBLENDD/PBLENDW. Differential Revision: https://reviews.llvm.org/D42042 llvm-svn: 322524
2018-01-16 06:18:45 +08:00
; AVX1-NEXT: vblendps {{.*#+}} ymm4 = ymm5[0,1,2,3,4,5],ymm4[6,7]
[x86] allow vector load narrowing with multi-use values This is a long-awaited follow-up suggested in D33578. Since then, we've picked up even more opportunities for vector narrowing from changes like D53784, so there are a lot of test diffs. Apart from 2-3 strange cases, these are all wins. I've structured this to be no-functional-change-intended for any target except for x86 because I couldn't tell if AArch64, ARM, and AMDGPU would improve or not. All of those targets have existing regression tests (4, 4, 10 files respectively) that would be affected. Also, Hexagon overrides the shouldReduceLoadWidth() hook, but doesn't show any regression test diffs. The trade-off is deciding if an extra vector load is better than a single wide load + extract_subvector. For x86, this is almost always better (on paper at least) because we often can fold loads into subsequent ops and not increase the official instruction count. There's also some unknown -- but potentially large -- benefit from using narrower vector ops if wide ops are implemented with multiple uops and/or frequency throttling is avoided. Differential Revision: https://reviews.llvm.org/D54073 llvm-svn: 346595
2018-11-11 04:05:31 +08:00
; AVX1-NEXT: vblendps {{.*#+}} xmm5 = xmm3[0,1],xmm2[2],xmm3[3]
[DAG] Generalize build_vector -> vector_shuffle combine for more than 2 inputs This generalizes the build_vector -> vector_shuffle combine to support any number of inputs. The idea is to create a binary tree of shuffles, where the first layer performs pairwise shuffles of the input vectors placing each input element into the correct lane, and the rest of the tree blends these shuffles together. This doesn't try to be smart and create any sort of "optimal" shuffles. The assumption is that even a "poor" shuffle sequence is better than extracting and inserting the elements one by one. Differential Revision: https://reviews.llvm.org/D24683 llvm-svn: 283480
2016-10-07 02:58:24 +08:00
; AVX1-NEXT: vpermilps {{.*#+}} xmm5 = xmm5[0,0,3,2]
; AVX1-NEXT: vinsertf128 $1, %xmm5, %ymm0, %ymm5
; AVX1-NEXT: vblendps {{.*#+}} ymm6 = ymm0[0,1],ymm1[2],ymm0[3,4],ymm1[5],ymm0[6,7]
; AVX1-NEXT: vextractf128 $1, %ymm6, %xmm7
; AVX1-NEXT: vblendps {{.*#+}} xmm6 = xmm7[0],xmm6[1,2],xmm7[3]
; AVX1-NEXT: vpermilps {{.*#+}} xmm6 = xmm6[1,0,3,2]
; AVX1-NEXT: vmovshdup {{.*#+}} xmm7 = xmm7[1,1,3,3]
; AVX1-NEXT: vinsertf128 $1, %xmm7, %ymm6, %ymm6
; AVX1-NEXT: vblendps {{.*#+}} ymm5 = ymm6[0,1,2,3,4],ymm5[5,6,7]
[x86] allow vector load narrowing with multi-use values This is a long-awaited follow-up suggested in D33578. Since then, we've picked up even more opportunities for vector narrowing from changes like D53784, so there are a lot of test diffs. Apart from 2-3 strange cases, these are all wins. I've structured this to be no-functional-change-intended for any target except for x86 because I couldn't tell if AArch64, ARM, and AMDGPU would improve or not. All of those targets have existing regression tests (4, 4, 10 files respectively) that would be affected. Also, Hexagon overrides the shouldReduceLoadWidth() hook, but doesn't show any regression test diffs. The trade-off is deciding if an extra vector load is better than a single wide load + extract_subvector. For x86, this is almost always better (on paper at least) because we often can fold loads into subsequent ops and not increase the official instruction count. There's also some unknown -- but potentially large -- benefit from using narrower vector ops if wide ops are implemented with multiple uops and/or frequency throttling is avoided. Differential Revision: https://reviews.llvm.org/D54073 llvm-svn: 346595
2018-11-11 04:05:31 +08:00
; AVX1-NEXT: vshufps {{.*#+}} xmm2 = xmm3[0,1],xmm2[0,3]
[DAG] Generalize build_vector -> vector_shuffle combine for more than 2 inputs This generalizes the build_vector -> vector_shuffle combine to support any number of inputs. The idea is to create a binary tree of shuffles, where the first layer performs pairwise shuffles of the input vectors placing each input element into the correct lane, and the rest of the tree blends these shuffles together. This doesn't try to be smart and create any sort of "optimal" shuffles. The assumption is that even a "poor" shuffle sequence is better than extracting and inserting the elements one by one. Differential Revision: https://reviews.llvm.org/D24683 llvm-svn: 283480
2016-10-07 02:58:24 +08:00
; AVX1-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm2
; AVX1-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2],ymm1[3,4],ymm0[5],ymm1[6,7]
; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1
; AVX1-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2,3]
; AVX1-NEXT: vpermilps {{.*#+}} xmm0 = xmm0[2,1,0,3]
; AVX1-NEXT: vpermilpd {{.*#+}} xmm1 = xmm1[1,0]
; AVX1-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
; AVX1-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5,6,7]
[X86][SSE] Add custom execution domain fixing for BLENDPD/BLENDPS/PBLENDD/PBLENDW (PR34873) Add support for custom execution domain fixing and implement support for BLENDPD/BLENDPS/PBLENDD/PBLENDW. Differential Revision: https://reviews.llvm.org/D42042 llvm-svn: 322524
2018-01-16 06:18:45 +08:00
; AVX1-NEXT: vmovups %ymm4, (%rsi)
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
; AVX1-NEXT: vmovups %ymm5, (%rdx)
; AVX1-NEXT: vmovups %ymm0, (%rcx)
; AVX1-NEXT: vzeroupper
; AVX1-NEXT: retq
;
X86 Tests: Update more isel tests with FastVariableShuffle feature Summary: Added the FastVariableShuffle feature to cases that resembled processors for which this fearure is on. For AVX2 there are processors with and w/o this fearue enable. For AVX512 only KNL does enable this feature so cases which only have +avx512f were left without the FastVariableShuffle enabled. Reviewers: RKSimon, craig.topper Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D41851 llvm-svn: 322090
2018-01-10 00:26:06 +08:00
; AVX2-SLOW-LABEL: interleave_24i32_out:
; AVX2-SLOW: # %bb.0:
; AVX2-SLOW-NEXT: vmovups (%rdi), %ymm0
; AVX2-SLOW-NEXT: vmovups 32(%rdi), %ymm1
; AVX2-SLOW-NEXT: vmovups 64(%rdi), %ymm2
; AVX2-SLOW-NEXT: vmovaps {{.*#+}} ymm3 = <u,u,u,u,u,u,2,5>
; AVX2-SLOW-NEXT: vpermps %ymm2, %ymm3, %ymm3
; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm4 = ymm0[0],ymm1[1],ymm0[2,3],ymm1[4],ymm0[5,6],ymm1[7]
; AVX2-SLOW-NEXT: vmovaps {{.*#+}} ymm5 = <0,3,6,1,4,7,u,u>
; AVX2-SLOW-NEXT: vpermps %ymm4, %ymm5, %ymm4
; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm3 = ymm4[0,1,2,3,4,5],ymm3[6,7]
; AVX2-SLOW-NEXT: vmovaps {{.*#+}} ymm4 = <u,u,u,u,u,0,3,6>
; AVX2-SLOW-NEXT: vpermps %ymm2, %ymm4, %ymm4
; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm5 = ymm0[0,1],ymm1[2],ymm0[3,4],ymm1[5],ymm0[6,7]
; AVX2-SLOW-NEXT: vmovaps {{.*#+}} ymm6 = <1,4,7,2,5,u,u,u>
; AVX2-SLOW-NEXT: vpermps %ymm5, %ymm6, %ymm5
; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm4 = ymm5[0,1,2,3,4],ymm4[5,6,7]
; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2],ymm1[3,4],ymm0[5],ymm1[6,7]
; AVX2-SLOW-NEXT: vmovaps {{.*#+}} ymm1 = <2,5,0,3,6,u,u,u>
; AVX2-SLOW-NEXT: vpermps %ymm0, %ymm1, %ymm0
; AVX2-SLOW-NEXT: vpermilps {{.*#+}} ymm1 = ymm2[0,1,0,3,4,5,4,7]
; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,1,0,3]
; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5,6,7]
; AVX2-SLOW-NEXT: vmovups %ymm3, (%rsi)
; AVX2-SLOW-NEXT: vmovups %ymm4, (%rdx)
; AVX2-SLOW-NEXT: vmovups %ymm0, (%rcx)
; AVX2-SLOW-NEXT: vzeroupper
; AVX2-SLOW-NEXT: retq
;
; AVX2-FAST-LABEL: interleave_24i32_out:
; AVX2-FAST: # %bb.0:
; AVX2-FAST-NEXT: vmovups (%rdi), %ymm0
; AVX2-FAST-NEXT: vmovups 32(%rdi), %ymm1
; AVX2-FAST-NEXT: vmovups 64(%rdi), %ymm2
; AVX2-FAST-NEXT: vmovaps {{.*#+}} ymm3 = <u,u,u,u,u,u,2,5>
; AVX2-FAST-NEXT: vpermps %ymm2, %ymm3, %ymm3
; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm4 = ymm0[0],ymm1[1],ymm0[2,3],ymm1[4],ymm0[5,6],ymm1[7]
; AVX2-FAST-NEXT: vmovaps {{.*#+}} ymm5 = <0,3,6,1,4,7,u,u>
; AVX2-FAST-NEXT: vpermps %ymm4, %ymm5, %ymm4
; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm3 = ymm4[0,1,2,3,4,5],ymm3[6,7]
; AVX2-FAST-NEXT: vmovaps {{.*#+}} ymm4 = <u,u,u,u,u,0,3,6>
; AVX2-FAST-NEXT: vpermps %ymm2, %ymm4, %ymm4
; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm5 = ymm0[0,1],ymm1[2],ymm0[3,4],ymm1[5],ymm0[6,7]
; AVX2-FAST-NEXT: vmovaps {{.*#+}} ymm6 = <1,4,7,2,5,u,u,u>
; AVX2-FAST-NEXT: vpermps %ymm5, %ymm6, %ymm5
; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm4 = ymm5[0,1,2,3,4],ymm4[5,6,7]
; AVX2-FAST-NEXT: vmovaps {{.*#+}} ymm5 = [0,1,0,3,0,1,4,7]
; AVX2-FAST-NEXT: vpermps %ymm2, %ymm5, %ymm2
; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2],ymm1[3,4],ymm0[5],ymm1[6,7]
; AVX2-FAST-NEXT: vmovaps {{.*#+}} ymm1 = <2,5,0,3,6,u,u,u>
; AVX2-FAST-NEXT: vpermps %ymm0, %ymm1, %ymm0
; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5,6,7]
; AVX2-FAST-NEXT: vmovups %ymm3, (%rsi)
; AVX2-FAST-NEXT: vmovups %ymm4, (%rdx)
; AVX2-FAST-NEXT: vmovups %ymm0, (%rcx)
; AVX2-FAST-NEXT: vzeroupper
; AVX2-FAST-NEXT: retq
[X86][XOP] Add XOP oddshuffles tests XOP codegen is often different to generic AVX - thank you vpperm! llvm-svn: 315176
2017-10-08 20:58:15 +08:00
;
; XOP-LABEL: interleave_24i32_out:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; XOP: # %bb.0:
[X86][XOP] Add XOP oddshuffles tests XOP codegen is often different to generic AVX - thank you vpperm! llvm-svn: 315176
2017-10-08 20:58:15 +08:00
; XOP-NEXT: vmovups (%rdi), %ymm0
; XOP-NEXT: vmovups 32(%rdi), %ymm1
[x86] allow vector load narrowing with multi-use values This is a long-awaited follow-up suggested in D33578. Since then, we've picked up even more opportunities for vector narrowing from changes like D53784, so there are a lot of test diffs. Apart from 2-3 strange cases, these are all wins. I've structured this to be no-functional-change-intended for any target except for x86 because I couldn't tell if AArch64, ARM, and AMDGPU would improve or not. All of those targets have existing regression tests (4, 4, 10 files respectively) that would be affected. Also, Hexagon overrides the shouldReduceLoadWidth() hook, but doesn't show any regression test diffs. The trade-off is deciding if an extra vector load is better than a single wide load + extract_subvector. For x86, this is almost always better (on paper at least) because we often can fold loads into subsequent ops and not increase the official instruction count. There's also some unknown -- but potentially large -- benefit from using narrower vector ops if wide ops are implemented with multiple uops and/or frequency throttling is avoided. Differential Revision: https://reviews.llvm.org/D54073 llvm-svn: 346595
2018-11-11 04:05:31 +08:00
; XOP-NEXT: vmovups 80(%rdi), %xmm2
; XOP-NEXT: vmovups 64(%rdi), %xmm3
; XOP-NEXT: vinsertps {{.*#+}} xmm4 = zero,zero,xmm3[2],xmm2[1]
[X86][XOP] Add XOP oddshuffles tests XOP codegen is often different to generic AVX - thank you vpperm! llvm-svn: 315176
2017-10-08 20:58:15 +08:00
; XOP-NEXT: vinsertf128 $1, %xmm4, %ymm0, %ymm4
; XOP-NEXT: vblendps {{.*#+}} ymm5 = ymm0[0],ymm1[1],ymm0[2,3],ymm1[4],ymm0[5,6],ymm1[7]
; XOP-NEXT: vextractf128 $1, %ymm5, %xmm6
; XOP-NEXT: vblendps {{.*#+}} xmm5 = xmm5[0,1],xmm6[2],xmm5[3]
; XOP-NEXT: vpermilps {{.*#+}} xmm5 = xmm5[0,3,2,1]
; XOP-NEXT: vpermilps {{.*#+}} xmm6 = xmm6[0,3,2,3]
; XOP-NEXT: vinsertf128 $1, %xmm6, %ymm5, %ymm5
[X86][SSE] Add custom execution domain fixing for BLENDPD/BLENDPS/PBLENDD/PBLENDW (PR34873) Add support for custom execution domain fixing and implement support for BLENDPD/BLENDPS/PBLENDD/PBLENDW. Differential Revision: https://reviews.llvm.org/D42042 llvm-svn: 322524
2018-01-16 06:18:45 +08:00
; XOP-NEXT: vblendps {{.*#+}} ymm4 = ymm5[0,1,2,3,4,5],ymm4[6,7]
[x86] allow vector load narrowing with multi-use values This is a long-awaited follow-up suggested in D33578. Since then, we've picked up even more opportunities for vector narrowing from changes like D53784, so there are a lot of test diffs. Apart from 2-3 strange cases, these are all wins. I've structured this to be no-functional-change-intended for any target except for x86 because I couldn't tell if AArch64, ARM, and AMDGPU would improve or not. All of those targets have existing regression tests (4, 4, 10 files respectively) that would be affected. Also, Hexagon overrides the shouldReduceLoadWidth() hook, but doesn't show any regression test diffs. The trade-off is deciding if an extra vector load is better than a single wide load + extract_subvector. For x86, this is almost always better (on paper at least) because we often can fold loads into subsequent ops and not increase the official instruction count. There's also some unknown -- but potentially large -- benefit from using narrower vector ops if wide ops are implemented with multiple uops and/or frequency throttling is avoided. Differential Revision: https://reviews.llvm.org/D54073 llvm-svn: 346595
2018-11-11 04:05:31 +08:00
; XOP-NEXT: vblendps {{.*#+}} xmm5 = xmm3[0,1],xmm2[2],xmm3[3]
[X86][XOP] Add XOP oddshuffles tests XOP codegen is often different to generic AVX - thank you vpperm! llvm-svn: 315176
2017-10-08 20:58:15 +08:00
; XOP-NEXT: vpermilps {{.*#+}} xmm5 = xmm5[0,0,3,2]
; XOP-NEXT: vinsertf128 $1, %xmm5, %ymm0, %ymm5
; XOP-NEXT: vblendps {{.*#+}} ymm6 = ymm0[0,1],ymm1[2],ymm0[3,4],ymm1[5],ymm0[6,7]
; XOP-NEXT: vextractf128 $1, %ymm6, %xmm7
; XOP-NEXT: vblendps {{.*#+}} xmm6 = xmm7[0],xmm6[1,2],xmm7[3]
; XOP-NEXT: vpermilps {{.*#+}} xmm6 = xmm6[1,0,3,2]
; XOP-NEXT: vmovshdup {{.*#+}} xmm7 = xmm7[1,1,3,3]
; XOP-NEXT: vinsertf128 $1, %xmm7, %ymm6, %ymm6
; XOP-NEXT: vblendps {{.*#+}} ymm5 = ymm6[0,1,2,3,4],ymm5[5,6,7]
[x86] allow vector load narrowing with multi-use values This is a long-awaited follow-up suggested in D33578. Since then, we've picked up even more opportunities for vector narrowing from changes like D53784, so there are a lot of test diffs. Apart from 2-3 strange cases, these are all wins. I've structured this to be no-functional-change-intended for any target except for x86 because I couldn't tell if AArch64, ARM, and AMDGPU would improve or not. All of those targets have existing regression tests (4, 4, 10 files respectively) that would be affected. Also, Hexagon overrides the shouldReduceLoadWidth() hook, but doesn't show any regression test diffs. The trade-off is deciding if an extra vector load is better than a single wide load + extract_subvector. For x86, this is almost always better (on paper at least) because we often can fold loads into subsequent ops and not increase the official instruction count. There's also some unknown -- but potentially large -- benefit from using narrower vector ops if wide ops are implemented with multiple uops and/or frequency throttling is avoided. Differential Revision: https://reviews.llvm.org/D54073 llvm-svn: 346595
2018-11-11 04:05:31 +08:00
; XOP-NEXT: vshufps {{.*#+}} xmm2 = xmm3[0,1],xmm2[0,3]
[X86][XOP] Add XOP oddshuffles tests XOP codegen is often different to generic AVX - thank you vpperm! llvm-svn: 315176
2017-10-08 20:58:15 +08:00
; XOP-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm2
; XOP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2],ymm1[3,4],ymm0[5],ymm1[6,7]
; XOP-NEXT: vextractf128 $1, %ymm0, %xmm1
; XOP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2,3]
; XOP-NEXT: vpermilps {{.*#+}} xmm0 = xmm0[2,1,0,3]
; XOP-NEXT: vpermilpd {{.*#+}} xmm1 = xmm1[1,0]
; XOP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
; XOP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5,6,7]
[X86][SSE] Add custom execution domain fixing for BLENDPD/BLENDPS/PBLENDD/PBLENDW (PR34873) Add support for custom execution domain fixing and implement support for BLENDPD/BLENDPS/PBLENDD/PBLENDW. Differential Revision: https://reviews.llvm.org/D42042 llvm-svn: 322524
2018-01-16 06:18:45 +08:00
; XOP-NEXT: vmovups %ymm4, (%rsi)
[X86][XOP] Add XOP oddshuffles tests XOP codegen is often different to generic AVX - thank you vpperm! llvm-svn: 315176
2017-10-08 20:58:15 +08:00
; XOP-NEXT: vmovups %ymm5, (%rdx)
; XOP-NEXT: vmovups %ymm0, (%rcx)
; XOP-NEXT: vzeroupper
; XOP-NEXT: retq
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
%wide.vec = load <24 x i32>, <24 x i32>* %p, align 4
%s1 = shufflevector <24 x i32> %wide.vec, <24 x i32> undef, <8 x i32> <i32 0, i32 3, i32 6, i32 9, i32 12, i32 15, i32 18, i32 21>
%s2 = shufflevector <24 x i32> %wide.vec, <24 x i32> undef, <8 x i32> <i32 1, i32 4, i32 7, i32 10, i32 13, i32 16, i32 19, i32 22>
%s3 = shufflevector <24 x i32> %wide.vec, <24 x i32> undef, <8 x i32> <i32 2, i32 5, i32 8, i32 11, i32 14, i32 17, i32 20, i32 23>
store <8 x i32> %s1, <8 x i32>* %q1, align 4
store <8 x i32> %s2, <8 x i32>* %q2, align 4
store <8 x i32> %s3, <8 x i32>* %q3, align 4
ret void
}
define void @interleave_24i32_in(<24 x i32>* %p, <8 x i32>* %q1, <8 x i32>* %q2, <8 x i32>* %q3) nounwind {
; SSE2-LABEL: interleave_24i32_in:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/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:
[DAGCombiner] If a TokenFactor would be merged into its user, consider the user later. Summary: A number of optimizations are inhibited by single-use TokenFactors not being merged into the TokenFactor using it. This makes we consider if we can do the merge immediately. Most tests changes here are due to the change in visitation causing minor reorderings and associated reassociation of paired memory operations. CodeGen tests with non-reordering changes: X86/aligned-variadic.ll -- memory-based add folded into stored leaq value. X86/constant-combiners.ll -- Optimizes out overlap between stores. X86/pr40631_deadstore_elision -- folds constant byte store into preceding quad word constant store. Reviewers: RKSimon, craig.topper, spatel, efriedma, courbet Reviewed By: courbet Subscribers: dylanmckay, sdardis, nemanjai, jvesely, nhaehnle, javed.absar, eraman, hiraditya, kbarton, jrtc27, atanasyan, jsji, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D59260 llvm-svn: 356068
2019-03-14 01:07:09 +08:00
; SSE2-NEXT: movups (%rsi), %xmm1
; SSE2-NEXT: movups 16(%rsi), %xmm0
; SSE2-NEXT: movups (%rdx), %xmm8
; SSE2-NEXT: movups 16(%rdx), %xmm5
; SSE2-NEXT: movups (%rcx), %xmm3
; SSE2-NEXT: movups 16(%rcx), %xmm6
; SSE2-NEXT: movaps %xmm3, %xmm7
; SSE2-NEXT: shufps {{.*#+}} xmm7 = xmm7[0,0],xmm1[1,0]
; SSE2-NEXT: movaps %xmm1, %xmm9
; SSE2-NEXT: unpcklps {{.*#+}} xmm9 = xmm9[0],xmm8[0],xmm9[1],xmm8[1]
; SSE2-NEXT: shufps {{.*#+}} xmm9 = xmm9[0,1],xmm7[0,2]
[X86][SSE] Combine (some) target shuffles with multiple uses As discussed on D41794, we have many cases where we fail to combine shuffles as the input operands have other uses. This patch permits these shuffles to be combined as long as they don't introduce additional variable shuffle masks, which should reduce instruction dependencies and allow the total number of shuffles to still drop without increasing the constant pool. However, this may mean that some memory folds may no longer occur, and on pre-AVX require the occasional extra register move. This also exposes some poor PMULDQ/PMULUDQ codegen which was doing unnecessary upper/lower calculations which will in fact fold to zero/undef - the fix will be added in a followup commit. Differential Revision: https://reviews.llvm.org/D50328 llvm-svn: 339335
2018-08-09 20:30:02 +08:00
; SSE2-NEXT: movaps %xmm5, %xmm7
[DAGCombiner] If a TokenFactor would be merged into its user, consider the user later. Summary: A number of optimizations are inhibited by single-use TokenFactors not being merged into the TokenFactor using it. This makes we consider if we can do the merge immediately. Most tests changes here are due to the change in visitation causing minor reorderings and associated reassociation of paired memory operations. CodeGen tests with non-reordering changes: X86/aligned-variadic.ll -- memory-based add folded into stored leaq value. X86/constant-combiners.ll -- Optimizes out overlap between stores. X86/pr40631_deadstore_elision -- folds constant byte store into preceding quad word constant store. Reviewers: RKSimon, craig.topper, spatel, efriedma, courbet Reviewed By: courbet Subscribers: dylanmckay, sdardis, nemanjai, jvesely, nhaehnle, javed.absar, eraman, hiraditya, kbarton, jrtc27, atanasyan, jsji, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D59260 llvm-svn: 356068
2019-03-14 01:07:09 +08:00
; SSE2-NEXT: shufps {{.*#+}} xmm7 = xmm7[3,2],xmm6[3,2]
; SSE2-NEXT: movaps %xmm6, %xmm4
; SSE2-NEXT: unpckhps {{.*#+}} xmm4 = xmm4[2],xmm0[2],xmm4[3],xmm0[3]
; SSE2-NEXT: shufps {{.*#+}} xmm4 = xmm4[0,3],xmm7[0,2]
; SSE2-NEXT: movaps %xmm0, %xmm7
; SSE2-NEXT: shufps {{.*#+}} xmm7 = xmm7[2,0],xmm5[2,1]
; SSE2-NEXT: movaps %xmm6, %xmm2
; SSE2-NEXT: shufps {{.*#+}} xmm2 = xmm2[1,0],xmm5[1,0]
[X86][SSE] Combine (some) target shuffles with multiple uses As discussed on D41794, we have many cases where we fail to combine shuffles as the input operands have other uses. This patch permits these shuffles to be combined as long as they don't introduce additional variable shuffle masks, which should reduce instruction dependencies and allow the total number of shuffles to still drop without increasing the constant pool. However, this may mean that some memory folds may no longer occur, and on pre-AVX require the occasional extra register move. This also exposes some poor PMULDQ/PMULUDQ codegen which was doing unnecessary upper/lower calculations which will in fact fold to zero/undef - the fix will be added in a followup commit. Differential Revision: https://reviews.llvm.org/D50328 llvm-svn: 339335
2018-08-09 20:30:02 +08:00
; SSE2-NEXT: shufps {{.*#+}} xmm2 = xmm2[2,0],xmm7[0,2]
[DAGCombiner] If a TokenFactor would be merged into its user, consider the user later. Summary: A number of optimizations are inhibited by single-use TokenFactors not being merged into the TokenFactor using it. This makes we consider if we can do the merge immediately. Most tests changes here are due to the change in visitation causing minor reorderings and associated reassociation of paired memory operations. CodeGen tests with non-reordering changes: X86/aligned-variadic.ll -- memory-based add folded into stored leaq value. X86/constant-combiners.ll -- Optimizes out overlap between stores. X86/pr40631_deadstore_elision -- folds constant byte store into preceding quad word constant store. Reviewers: RKSimon, craig.topper, spatel, efriedma, courbet Reviewed By: courbet Subscribers: dylanmckay, sdardis, nemanjai, jvesely, nhaehnle, javed.absar, eraman, hiraditya, kbarton, jrtc27, atanasyan, jsji, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D59260 llvm-svn: 356068
2019-03-14 01:07:09 +08:00
; SSE2-NEXT: shufps {{.*#+}} xmm6 = xmm6[0,0],xmm0[1,0]
; SSE2-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm5[0],xmm0[1],xmm5[1]
; SSE2-NEXT: shufps {{.*#+}} xmm0 = xmm0[0,1],xmm6[0,2]
[X86][SSE] Combine (some) target shuffles with multiple uses As discussed on D41794, we have many cases where we fail to combine shuffles as the input operands have other uses. This patch permits these shuffles to be combined as long as they don't introduce additional variable shuffle masks, which should reduce instruction dependencies and allow the total number of shuffles to still drop without increasing the constant pool. However, this may mean that some memory folds may no longer occur, and on pre-AVX require the occasional extra register move. This also exposes some poor PMULDQ/PMULUDQ codegen which was doing unnecessary upper/lower calculations which will in fact fold to zero/undef - the fix will be added in a followup commit. Differential Revision: https://reviews.llvm.org/D50328 llvm-svn: 339335
2018-08-09 20:30:02 +08:00
; SSE2-NEXT: movaps %xmm8, %xmm5
[DAGCombiner] If a TokenFactor would be merged into its user, consider the user later. Summary: A number of optimizations are inhibited by single-use TokenFactors not being merged into the TokenFactor using it. This makes we consider if we can do the merge immediately. Most tests changes here are due to the change in visitation causing minor reorderings and associated reassociation of paired memory operations. CodeGen tests with non-reordering changes: X86/aligned-variadic.ll -- memory-based add folded into stored leaq value. X86/constant-combiners.ll -- Optimizes out overlap between stores. X86/pr40631_deadstore_elision -- folds constant byte store into preceding quad word constant store. Reviewers: RKSimon, craig.topper, spatel, efriedma, courbet Reviewed By: courbet Subscribers: dylanmckay, sdardis, nemanjai, jvesely, nhaehnle, javed.absar, eraman, hiraditya, kbarton, jrtc27, atanasyan, jsji, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D59260 llvm-svn: 356068
2019-03-14 01:07:09 +08:00
; SSE2-NEXT: shufps {{.*#+}} xmm5 = xmm5[3,2],xmm3[3,2]
; SSE2-NEXT: movaps %xmm3, %xmm6
; SSE2-NEXT: unpckhps {{.*#+}} xmm6 = xmm6[2],xmm1[2],xmm6[3],xmm1[3]
; SSE2-NEXT: shufps {{.*#+}} xmm6 = xmm6[0,3],xmm5[0,2]
; SSE2-NEXT: shufps {{.*#+}} xmm1 = xmm1[2,0],xmm8[2,1]
; SSE2-NEXT: shufps {{.*#+}} xmm3 = xmm3[1,0],xmm8[1,0]
; SSE2-NEXT: shufps {{.*#+}} xmm3 = xmm3[2,0],xmm1[0,2]
; SSE2-NEXT: movups %xmm3, 16(%rdi)
; SSE2-NEXT: movups %xmm6, 32(%rdi)
; SSE2-NEXT: movups %xmm0, 48(%rdi)
; SSE2-NEXT: movups %xmm2, 64(%rdi)
[X86][SSE] Combine (some) target shuffles with multiple uses As discussed on D41794, we have many cases where we fail to combine shuffles as the input operands have other uses. This patch permits these shuffles to be combined as long as they don't introduce additional variable shuffle masks, which should reduce instruction dependencies and allow the total number of shuffles to still drop without increasing the constant pool. However, this may mean that some memory folds may no longer occur, and on pre-AVX require the occasional extra register move. This also exposes some poor PMULDQ/PMULUDQ codegen which was doing unnecessary upper/lower calculations which will in fact fold to zero/undef - the fix will be added in a followup commit. Differential Revision: https://reviews.llvm.org/D50328 llvm-svn: 339335
2018-08-09 20:30:02 +08:00
; SSE2-NEXT: movups %xmm4, 80(%rdi)
[DAGCombiner] If a TokenFactor would be merged into its user, consider the user later. Summary: A number of optimizations are inhibited by single-use TokenFactors not being merged into the TokenFactor using it. This makes we consider if we can do the merge immediately. Most tests changes here are due to the change in visitation causing minor reorderings and associated reassociation of paired memory operations. CodeGen tests with non-reordering changes: X86/aligned-variadic.ll -- memory-based add folded into stored leaq value. X86/constant-combiners.ll -- Optimizes out overlap between stores. X86/pr40631_deadstore_elision -- folds constant byte store into preceding quad word constant store. Reviewers: RKSimon, craig.topper, spatel, efriedma, courbet Reviewed By: courbet Subscribers: dylanmckay, sdardis, nemanjai, jvesely, nhaehnle, javed.absar, eraman, hiraditya, kbarton, jrtc27, atanasyan, jsji, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D59260 llvm-svn: 356068
2019-03-14 01:07:09 +08:00
; SSE2-NEXT: movups %xmm9, (%rdi)
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
; SSE2-NEXT: retq
;
; SSE42-LABEL: interleave_24i32_in:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; SSE42: # %bb.0:
[DAGCombiner] If a TokenFactor would be merged into its user, consider the user later. Summary: A number of optimizations are inhibited by single-use TokenFactors not being merged into the TokenFactor using it. This makes we consider if we can do the merge immediately. Most tests changes here are due to the change in visitation causing minor reorderings and associated reassociation of paired memory operations. CodeGen tests with non-reordering changes: X86/aligned-variadic.ll -- memory-based add folded into stored leaq value. X86/constant-combiners.ll -- Optimizes out overlap between stores. X86/pr40631_deadstore_elision -- folds constant byte store into preceding quad word constant store. Reviewers: RKSimon, craig.topper, spatel, efriedma, courbet Reviewed By: courbet Subscribers: dylanmckay, sdardis, nemanjai, jvesely, nhaehnle, javed.absar, eraman, hiraditya, kbarton, jrtc27, atanasyan, jsji, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D59260 llvm-svn: 356068
2019-03-14 01:07:09 +08:00
; SSE42-NEXT: movdqu (%rsi), %xmm8
; SSE42-NEXT: movdqu 16(%rsi), %xmm4
; SSE42-NEXT: movdqu (%rdx), %xmm2
; SSE42-NEXT: movdqu 16(%rdx), %xmm5
; SSE42-NEXT: movdqu (%rcx), %xmm3
; SSE42-NEXT: movdqu 16(%rcx), %xmm6
; SSE42-NEXT: pshufd {{.*#+}} xmm1 = xmm2[0,0,1,1]
; SSE42-NEXT: pshufd {{.*#+}} xmm7 = xmm8[0,1,0,1]
; SSE42-NEXT: pblendw {{.*#+}} xmm7 = xmm7[0,1],xmm1[2,3],xmm7[4,5,6,7]
; SSE42-NEXT: pshufd {{.*#+}} xmm1 = xmm3[0,1,0,1]
; SSE42-NEXT: pblendw {{.*#+}} xmm1 = xmm7[0,1,2,3],xmm1[4,5],xmm7[6,7]
; SSE42-NEXT: pshufd {{.*#+}} xmm7 = xmm4[2,3,0,1]
; SSE42-NEXT: pshufd {{.*#+}} xmm0 = xmm6[2,3,2,3]
; SSE42-NEXT: pblendw {{.*#+}} xmm0 = xmm0[0,1],xmm7[2,3],xmm0[4,5,6,7]
; SSE42-NEXT: pshufd {{.*#+}} xmm7 = xmm5[2,2,3,3]
; SSE42-NEXT: pblendw {{.*#+}} xmm7 = xmm0[0,1,2,3],xmm7[4,5],xmm0[6,7]
; SSE42-NEXT: pshufd {{.*#+}} xmm0 = xmm5[1,1,2,2]
; SSE42-NEXT: pblendw {{.*#+}} xmm0 = xmm0[0,1],xmm6[2,3],xmm0[4,5,6,7]
; SSE42-NEXT: pblendw {{.*#+}} xmm0 = xmm0[0,1,2,3],xmm4[4,5],xmm0[6,7]
; SSE42-NEXT: pshufd {{.*#+}} xmm5 = xmm5[0,0,1,1]
; SSE42-NEXT: pshufd {{.*#+}} xmm4 = xmm4[0,1,0,1]
; SSE42-NEXT: pblendw {{.*#+}} xmm4 = xmm4[0,1],xmm5[2,3],xmm4[4,5,6,7]
; SSE42-NEXT: pshufd {{.*#+}} xmm5 = xmm6[0,1,0,1]
; SSE42-NEXT: pblendw {{.*#+}} xmm5 = xmm4[0,1,2,3],xmm5[4,5],xmm4[6,7]
; SSE42-NEXT: pshufd {{.*#+}} xmm4 = xmm8[2,3,0,1]
; SSE42-NEXT: pshufd {{.*#+}} xmm6 = xmm3[2,3,2,3]
; SSE42-NEXT: pblendw {{.*#+}} xmm6 = xmm6[0,1],xmm4[2,3],xmm6[4,5,6,7]
; SSE42-NEXT: pshufd {{.*#+}} xmm4 = xmm2[2,2,3,3]
; SSE42-NEXT: pblendw {{.*#+}} xmm4 = xmm6[0,1,2,3],xmm4[4,5],xmm6[6,7]
; SSE42-NEXT: pshufd {{.*#+}} xmm2 = xmm2[1,1,2,2]
; SSE42-NEXT: pblendw {{.*#+}} xmm2 = xmm2[0,1],xmm3[2,3],xmm2[4,5,6,7]
; SSE42-NEXT: pblendw {{.*#+}} xmm2 = xmm2[0,1,2,3],xmm8[4,5],xmm2[6,7]
; SSE42-NEXT: movdqu %xmm2, 16(%rdi)
; SSE42-NEXT: movdqu %xmm4, 32(%rdi)
; SSE42-NEXT: movdqu %xmm5, 48(%rdi)
; SSE42-NEXT: movdqu %xmm0, 64(%rdi)
; SSE42-NEXT: movdqu %xmm7, 80(%rdi)
; SSE42-NEXT: movdqu %xmm1, (%rdi)
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
; SSE42-NEXT: retq
;
; AVX1-LABEL: interleave_24i32_in:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/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] allow vector load narrowing with multi-use values This is a long-awaited follow-up suggested in D33578. Since then, we've picked up even more opportunities for vector narrowing from changes like D53784, so there are a lot of test diffs. Apart from 2-3 strange cases, these are all wins. I've structured this to be no-functional-change-intended for any target except for x86 because I couldn't tell if AArch64, ARM, and AMDGPU would improve or not. All of those targets have existing regression tests (4, 4, 10 files respectively) that would be affected. Also, Hexagon overrides the shouldReduceLoadWidth() hook, but doesn't show any regression test diffs. The trade-off is deciding if an extra vector load is better than a single wide load + extract_subvector. For x86, this is almost always better (on paper at least) because we often can fold loads into subsequent ops and not increase the official instruction count. There's also some unknown -- but potentially large -- benefit from using narrower vector ops if wide ops are implemented with multiple uops and/or frequency throttling is avoided. Differential Revision: https://reviews.llvm.org/D54073 llvm-svn: 346595
2018-11-11 04:05:31 +08:00
; AVX1-NEXT: vmovupd (%rsi), %ymm0
; AVX1-NEXT: vmovupd (%rcx), %ymm1
[DAGCombiner] If a TokenFactor would be merged into its user, consider the user later. Summary: A number of optimizations are inhibited by single-use TokenFactors not being merged into the TokenFactor using it. This makes we consider if we can do the merge immediately. Most tests changes here are due to the change in visitation causing minor reorderings and associated reassociation of paired memory operations. CodeGen tests with non-reordering changes: X86/aligned-variadic.ll -- memory-based add folded into stored leaq value. X86/constant-combiners.ll -- Optimizes out overlap between stores. X86/pr40631_deadstore_elision -- folds constant byte store into preceding quad word constant store. Reviewers: RKSimon, craig.topper, spatel, efriedma, courbet Reviewed By: courbet Subscribers: dylanmckay, sdardis, nemanjai, jvesely, nhaehnle, javed.absar, eraman, hiraditya, kbarton, jrtc27, atanasyan, jsji, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D59260 llvm-svn: 356068
2019-03-14 01:07:09 +08:00
; AVX1-NEXT: vmovups 16(%rcx), %xmm2
; AVX1-NEXT: vmovups (%rdx), %xmm3
; AVX1-NEXT: vmovups 16(%rdx), %xmm4
; AVX1-NEXT: vshufps {{.*#+}} xmm5 = xmm4[3,0],xmm2[3,0]
; AVX1-NEXT: vshufps {{.*#+}} xmm5 = xmm2[2,1],xmm5[0,2]
; AVX1-NEXT: vshufps {{.*#+}} xmm2 = xmm2[1,0],xmm4[1,0]
; AVX1-NEXT: vshufps {{.*#+}} xmm2 = xmm2[2,0],xmm4[2,2]
Recommit r354363 "[X86][SSE] Generalize X86ISD::BLENDI support to more value types" And its follow ups r354511, r354640. A follow patch will fix the issue that caused it to be reverted. llvm-svn: 354737
2019-02-24 05:41:42 +08:00
; AVX1-NEXT: vinsertf128 $1, %xmm5, %ymm2, %ymm2
; AVX1-NEXT: vpermilpd {{.*#+}} ymm4 = ymm0[1,1,3,3]
; AVX1-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3,2,3]
[DAGCombiner] If a TokenFactor would be merged into its user, consider the user later. Summary: A number of optimizations are inhibited by single-use TokenFactors not being merged into the TokenFactor using it. This makes we consider if we can do the merge immediately. Most tests changes here are due to the change in visitation causing minor reorderings and associated reassociation of paired memory operations. CodeGen tests with non-reordering changes: X86/aligned-variadic.ll -- memory-based add folded into stored leaq value. X86/constant-combiners.ll -- Optimizes out overlap between stores. X86/pr40631_deadstore_elision -- folds constant byte store into preceding quad word constant store. Reviewers: RKSimon, craig.topper, spatel, efriedma, courbet Reviewed By: courbet Subscribers: dylanmckay, sdardis, nemanjai, jvesely, nhaehnle, javed.absar, eraman, hiraditya, kbarton, jrtc27, atanasyan, jsji, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D59260 llvm-svn: 356068
2019-03-14 01:07:09 +08:00
; AVX1-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1],ymm4[2],ymm2[3,4],ymm4[5],ymm2[6,7]
; AVX1-NEXT: vmovups (%rsi), %xmm4
; AVX1-NEXT: vshufps {{.*#+}} xmm5 = xmm4[2,0],xmm3[2,0]
; AVX1-NEXT: vshufps {{.*#+}} xmm5 = xmm3[1,1],xmm5[0,2]
; AVX1-NEXT: vshufps {{.*#+}} xmm3 = xmm3[0,0],xmm4[0,0]
; AVX1-NEXT: vshufps {{.*#+}} xmm3 = xmm3[2,0],xmm4[2,1]
; AVX1-NEXT: vinsertf128 $1, %xmm5, %ymm3, %ymm3
; AVX1-NEXT: vmovddup {{.*#+}} xmm4 = xmm1[0,0]
; AVX1-NEXT: vinsertf128 $1, %xmm4, %ymm4, %ymm4
[DAG] Generalize build_vector -> vector_shuffle combine for more than 2 inputs This generalizes the build_vector -> vector_shuffle combine to support any number of inputs. The idea is to create a binary tree of shuffles, where the first layer performs pairwise shuffles of the input vectors placing each input element into the correct lane, and the rest of the tree blends these shuffles together. This doesn't try to be smart and create any sort of "optimal" shuffles. The assumption is that even a "poor" shuffle sequence is better than extracting and inserting the elements one by one. Differential Revision: https://reviews.llvm.org/D24683 llvm-svn: 283480
2016-10-07 02:58:24 +08:00
; AVX1-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1],ymm4[2],ymm3[3,4],ymm4[5],ymm3[6,7]
; AVX1-NEXT: vpermilpd {{.*#+}} ymm0 = ymm0[1,0,2,2]
[x86] allow vector load narrowing with multi-use values This is a long-awaited follow-up suggested in D33578. Since then, we've picked up even more opportunities for vector narrowing from changes like D53784, so there are a lot of test diffs. Apart from 2-3 strange cases, these are all wins. I've structured this to be no-functional-change-intended for any target except for x86 because I couldn't tell if AArch64, ARM, and AMDGPU would improve or not. All of those targets have existing regression tests (4, 4, 10 files respectively) that would be affected. Also, Hexagon overrides the shouldReduceLoadWidth() hook, but doesn't show any regression test diffs. The trade-off is deciding if an extra vector load is better than a single wide load + extract_subvector. For x86, this is almost always better (on paper at least) because we often can fold loads into subsequent ops and not increase the official instruction count. There's also some unknown -- but potentially large -- benefit from using narrower vector ops if wide ops are implemented with multiple uops and/or frequency throttling is avoided. Differential Revision: https://reviews.llvm.org/D54073 llvm-svn: 346595
2018-11-11 04:05:31 +08:00
; AVX1-NEXT: vpermilpd {{.*#+}} ymm1 = ymm1[1,1,2,2]
; AVX1-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2,3],ymm0[4],ymm1[5,6],ymm0[7]
; AVX1-NEXT: vpermilps {{.*#+}} ymm1 = mem[0,0,3,3,4,4,7,7]
[DAG] Generalize build_vector -> vector_shuffle combine for more than 2 inputs This generalizes the build_vector -> vector_shuffle combine to support any number of inputs. The idea is to create a binary tree of shuffles, where the first layer performs pairwise shuffles of the input vectors placing each input element into the correct lane, and the rest of the tree blends these shuffles together. This doesn't try to be smart and create any sort of "optimal" shuffles. The assumption is that even a "poor" shuffle sequence is better than extracting and inserting the elements one by one. Differential Revision: https://reviews.llvm.org/D24683 llvm-svn: 283480
2016-10-07 02:58:24 +08:00
; AVX1-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4],ymm1[5],ymm0[6,7]
[X86][SSE] Add custom execution domain fixing for BLENDPD/BLENDPS/PBLENDD/PBLENDW (PR34873) Add support for custom execution domain fixing and implement support for BLENDPD/BLENDPS/PBLENDD/PBLENDW. Differential Revision: https://reviews.llvm.org/D42042 llvm-svn: 322524
2018-01-16 06:18:45 +08:00
; AVX1-NEXT: vmovups %ymm0, 32(%rdi)
[DAGCombiner] If a TokenFactor would be merged into its user, consider the user later. Summary: A number of optimizations are inhibited by single-use TokenFactors not being merged into the TokenFactor using it. This makes we consider if we can do the merge immediately. Most tests changes here are due to the change in visitation causing minor reorderings and associated reassociation of paired memory operations. CodeGen tests with non-reordering changes: X86/aligned-variadic.ll -- memory-based add folded into stored leaq value. X86/constant-combiners.ll -- Optimizes out overlap between stores. X86/pr40631_deadstore_elision -- folds constant byte store into preceding quad word constant store. Reviewers: RKSimon, craig.topper, spatel, efriedma, courbet Reviewed By: courbet Subscribers: dylanmckay, sdardis, nemanjai, jvesely, nhaehnle, javed.absar, eraman, hiraditya, kbarton, jrtc27, atanasyan, jsji, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D59260 llvm-svn: 356068
2019-03-14 01:07:09 +08:00
; AVX1-NEXT: vmovups %ymm3, (%rdi)
; AVX1-NEXT: vmovups %ymm2, 64(%rdi)
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
; AVX1-NEXT: vzeroupper
; AVX1-NEXT: retq
;
X86 Tests: Update more isel tests with FastVariableShuffle feature Summary: Added the FastVariableShuffle feature to cases that resembled processors for which this fearure is on. For AVX2 there are processors with and w/o this fearue enable. For AVX512 only KNL does enable this feature so cases which only have +avx512f were left without the FastVariableShuffle enabled. Reviewers: RKSimon, craig.topper Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D41851 llvm-svn: 322090
2018-01-10 00:26:06 +08:00
; AVX2-SLOW-LABEL: interleave_24i32_in:
; AVX2-SLOW: # %bb.0:
; AVX2-SLOW-NEXT: vmovups (%rsi), %ymm0
; AVX2-SLOW-NEXT: vmovups (%rdx), %ymm1
; AVX2-SLOW-NEXT: vmovups (%rcx), %ymm2
[x86] allow vector load narrowing with multi-use values This is a long-awaited follow-up suggested in D33578. Since then, we've picked up even more opportunities for vector narrowing from changes like D53784, so there are a lot of test diffs. Apart from 2-3 strange cases, these are all wins. I've structured this to be no-functional-change-intended for any target except for x86 because I couldn't tell if AArch64, ARM, and AMDGPU would improve or not. All of those targets have existing regression tests (4, 4, 10 files respectively) that would be affected. Also, Hexagon overrides the shouldReduceLoadWidth() hook, but doesn't show any regression test diffs. The trade-off is deciding if an extra vector load is better than a single wide load + extract_subvector. For x86, this is almost always better (on paper at least) because we often can fold loads into subsequent ops and not increase the official instruction count. There's also some unknown -- but potentially large -- benefit from using narrower vector ops if wide ops are implemented with multiple uops and/or frequency throttling is avoided. Differential Revision: https://reviews.llvm.org/D54073 llvm-svn: 346595
2018-11-11 04:05:31 +08:00
; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm3 = ymm2[2,1,3,3]
; AVX2-SLOW-NEXT: vpermilps {{.*#+}} ymm4 = ymm1[1,2,3,3,5,6,7,7]
; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm4 = ymm4[2,2,2,3]
X86 Tests: Update more isel tests with FastVariableShuffle feature Summary: Added the FastVariableShuffle feature to cases that resembled processors for which this fearure is on. For AVX2 there are processors with and w/o this fearue enable. For AVX512 only KNL does enable this feature so cases which only have +avx512f were left without the FastVariableShuffle enabled. Reviewers: RKSimon, craig.topper Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D41851 llvm-svn: 322090
2018-01-10 00:26:06 +08:00
; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm3 = ymm4[0],ymm3[1],ymm4[2,3],ymm3[4],ymm4[5,6],ymm3[7]
[x86] allow vector load narrowing with multi-use values This is a long-awaited follow-up suggested in D33578. Since then, we've picked up even more opportunities for vector narrowing from changes like D53784, so there are a lot of test diffs. Apart from 2-3 strange cases, these are all wins. I've structured this to be no-functional-change-intended for any target except for x86 because I couldn't tell if AArch64, ARM, and AMDGPU would improve or not. All of those targets have existing regression tests (4, 4, 10 files respectively) that would be affected. Also, Hexagon overrides the shouldReduceLoadWidth() hook, but doesn't show any regression test diffs. The trade-off is deciding if an extra vector load is better than a single wide load + extract_subvector. For x86, this is almost always better (on paper at least) because we often can fold loads into subsequent ops and not increase the official instruction count. There's also some unknown -- but potentially large -- benefit from using narrower vector ops if wide ops are implemented with multiple uops and/or frequency throttling is avoided. Differential Revision: https://reviews.llvm.org/D54073 llvm-svn: 346595
2018-11-11 04:05:31 +08:00
; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm4 = ymm0[0,3,3,3]
X86 Tests: Update more isel tests with FastVariableShuffle feature Summary: Added the FastVariableShuffle feature to cases that resembled processors for which this fearure is on. For AVX2 there are processors with and w/o this fearue enable. For AVX512 only KNL does enable this feature so cases which only have +avx512f were left without the FastVariableShuffle enabled. Reviewers: RKSimon, craig.topper Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D41851 llvm-svn: 322090
2018-01-10 00:26:06 +08:00
; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1],ymm4[2],ymm3[3,4],ymm4[5],ymm3[6,7]
[x86] allow vector load narrowing with multi-use values This is a long-awaited follow-up suggested in D33578. Since then, we've picked up even more opportunities for vector narrowing from changes like D53784, so there are a lot of test diffs. Apart from 2-3 strange cases, these are all wins. I've structured this to be no-functional-change-intended for any target except for x86 because I couldn't tell if AArch64, ARM, and AMDGPU would improve or not. All of those targets have existing regression tests (4, 4, 10 files respectively) that would be affected. Also, Hexagon overrides the shouldReduceLoadWidth() hook, but doesn't show any regression test diffs. The trade-off is deciding if an extra vector load is better than a single wide load + extract_subvector. For x86, this is almost always better (on paper at least) because we often can fold loads into subsequent ops and not increase the official instruction count. There's also some unknown -- but potentially large -- benefit from using narrower vector ops if wide ops are implemented with multiple uops and/or frequency throttling is avoided. Differential Revision: https://reviews.llvm.org/D54073 llvm-svn: 346595
2018-11-11 04:05:31 +08:00
; AVX2-SLOW-NEXT: vpermilps {{.*#+}} xmm4 = mem[1,0,2,2]
; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm4 = ymm4[0,1,0,1]
; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm5 = ymm0[0,0,2,1]
X86 Tests: Update more isel tests with FastVariableShuffle feature Summary: Added the FastVariableShuffle feature to cases that resembled processors for which this fearure is on. For AVX2 there are processors with and w/o this fearue enable. For AVX512 only KNL does enable this feature so cases which only have +avx512f were left without the FastVariableShuffle enabled. Reviewers: RKSimon, craig.topper Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D41851 llvm-svn: 322090
2018-01-10 00:26:06 +08:00
; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm4 = ymm5[0],ymm4[1],ymm5[2,3],ymm4[4],ymm5[5,6],ymm4[7]
[x86] allow vector load narrowing with multi-use values This is a long-awaited follow-up suggested in D33578. Since then, we've picked up even more opportunities for vector narrowing from changes like D53784, so there are a lot of test diffs. Apart from 2-3 strange cases, these are all wins. I've structured this to be no-functional-change-intended for any target except for x86 because I couldn't tell if AArch64, ARM, and AMDGPU would improve or not. All of those targets have existing regression tests (4, 4, 10 files respectively) that would be affected. Also, Hexagon overrides the shouldReduceLoadWidth() hook, but doesn't show any regression test diffs. The trade-off is deciding if an extra vector load is better than a single wide load + extract_subvector. For x86, this is almost always better (on paper at least) because we often can fold loads into subsequent ops and not increase the official instruction count. There's also some unknown -- but potentially large -- benefit from using narrower vector ops if wide ops are implemented with multiple uops and/or frequency throttling is avoided. Differential Revision: https://reviews.llvm.org/D54073 llvm-svn: 346595
2018-11-11 04:05:31 +08:00
; AVX2-SLOW-NEXT: vbroadcastsd (%rcx), %ymm5
X86 Tests: Update more isel tests with FastVariableShuffle feature Summary: Added the FastVariableShuffle feature to cases that resembled processors for which this fearure is on. For AVX2 there are processors with and w/o this fearue enable. For AVX512 only KNL does enable this feature so cases which only have +avx512f were left without the FastVariableShuffle enabled. Reviewers: RKSimon, craig.topper Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D41851 llvm-svn: 322090
2018-01-10 00:26:06 +08:00
; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1],ymm5[2],ymm4[3,4],ymm5[5],ymm4[6,7]
; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[1,1,2,2]
; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[1,1,2,2]
; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0],ymm0[1],ymm2[2,3],ymm0[4],ymm2[5,6],ymm0[7]
; AVX2-SLOW-NEXT: vpermilps {{.*#+}} ymm1 = ymm1[0,0,3,3,4,4,7,7]
; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4],ymm1[5],ymm0[6,7]
; AVX2-SLOW-NEXT: vmovups %ymm0, 32(%rdi)
[x86] allow vector load narrowing with multi-use values This is a long-awaited follow-up suggested in D33578. Since then, we've picked up even more opportunities for vector narrowing from changes like D53784, so there are a lot of test diffs. Apart from 2-3 strange cases, these are all wins. I've structured this to be no-functional-change-intended for any target except for x86 because I couldn't tell if AArch64, ARM, and AMDGPU would improve or not. All of those targets have existing regression tests (4, 4, 10 files respectively) that would be affected. Also, Hexagon overrides the shouldReduceLoadWidth() hook, but doesn't show any regression test diffs. The trade-off is deciding if an extra vector load is better than a single wide load + extract_subvector. For x86, this is almost always better (on paper at least) because we often can fold loads into subsequent ops and not increase the official instruction count. There's also some unknown -- but potentially large -- benefit from using narrower vector ops if wide ops are implemented with multiple uops and/or frequency throttling is avoided. Differential Revision: https://reviews.llvm.org/D54073 llvm-svn: 346595
2018-11-11 04:05:31 +08:00
; AVX2-SLOW-NEXT: vmovups %ymm4, (%rdi)
; AVX2-SLOW-NEXT: vmovups %ymm3, 64(%rdi)
X86 Tests: Update more isel tests with FastVariableShuffle feature Summary: Added the FastVariableShuffle feature to cases that resembled processors for which this fearure is on. For AVX2 there are processors with and w/o this fearue enable. For AVX512 only KNL does enable this feature so cases which only have +avx512f were left without the FastVariableShuffle enabled. Reviewers: RKSimon, craig.topper Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D41851 llvm-svn: 322090
2018-01-10 00:26:06 +08:00
; AVX2-SLOW-NEXT: vzeroupper
; AVX2-SLOW-NEXT: retq
;
; AVX2-FAST-LABEL: interleave_24i32_in:
; AVX2-FAST: # %bb.0:
; AVX2-FAST-NEXT: vmovups (%rsi), %ymm0
; AVX2-FAST-NEXT: vmovups (%rdx), %ymm1
; AVX2-FAST-NEXT: vmovups (%rcx), %ymm2
[x86] allow vector load narrowing with multi-use values This is a long-awaited follow-up suggested in D33578. Since then, we've picked up even more opportunities for vector narrowing from changes like D53784, so there are a lot of test diffs. Apart from 2-3 strange cases, these are all wins. I've structured this to be no-functional-change-intended for any target except for x86 because I couldn't tell if AArch64, ARM, and AMDGPU would improve or not. All of those targets have existing regression tests (4, 4, 10 files respectively) that would be affected. Also, Hexagon overrides the shouldReduceLoadWidth() hook, but doesn't show any regression test diffs. The trade-off is deciding if an extra vector load is better than a single wide load + extract_subvector. For x86, this is almost always better (on paper at least) because we often can fold loads into subsequent ops and not increase the official instruction count. There's also some unknown -- but potentially large -- benefit from using narrower vector ops if wide ops are implemented with multiple uops and/or frequency throttling is avoided. Differential Revision: https://reviews.llvm.org/D54073 llvm-svn: 346595
2018-11-11 04:05:31 +08:00
; AVX2-FAST-NEXT: vmovaps {{.*#+}} ymm3 = [5,6,5,6,5,6,7,7]
[X86][AVX] getFauxShuffleMask - add support for INSERT_SUBVECTOR subvector shuffles Decode subvector shuffles from INSERT_SUBVECTOR(SRC0, SHUFFLE(EXTRACT_SUBVECTOR(SRC1)) This was found necessary while investigating PR39161 llvm-svn: 343853
2018-10-05 22:41:00 +08:00
; AVX2-FAST-NEXT: vpermps %ymm1, %ymm3, %ymm3
[x86] allow vector load narrowing with multi-use values This is a long-awaited follow-up suggested in D33578. Since then, we've picked up even more opportunities for vector narrowing from changes like D53784, so there are a lot of test diffs. Apart from 2-3 strange cases, these are all wins. I've structured this to be no-functional-change-intended for any target except for x86 because I couldn't tell if AArch64, ARM, and AMDGPU would improve or not. All of those targets have existing regression tests (4, 4, 10 files respectively) that would be affected. Also, Hexagon overrides the shouldReduceLoadWidth() hook, but doesn't show any regression test diffs. The trade-off is deciding if an extra vector load is better than a single wide load + extract_subvector. For x86, this is almost always better (on paper at least) because we often can fold loads into subsequent ops and not increase the official instruction count. There's also some unknown -- but potentially large -- benefit from using narrower vector ops if wide ops are implemented with multiple uops and/or frequency throttling is avoided. Differential Revision: https://reviews.llvm.org/D54073 llvm-svn: 346595
2018-11-11 04:05:31 +08:00
; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm4 = ymm2[2,1,3,3]
; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0],ymm4[1],ymm3[2,3],ymm4[4],ymm3[5,6],ymm4[7]
; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm4 = ymm0[0,3,3,3]
X86 Tests: Update more isel tests with FastVariableShuffle feature Summary: Added the FastVariableShuffle feature to cases that resembled processors for which this fearure is on. For AVX2 there are processors with and w/o this fearue enable. For AVX512 only KNL does enable this feature so cases which only have +avx512f were left without the FastVariableShuffle enabled. Reviewers: RKSimon, craig.topper Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D41851 llvm-svn: 322090
2018-01-10 00:26:06 +08:00
; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1],ymm4[2],ymm3[3,4],ymm4[5],ymm3[6,7]
[DAGCombiner] If a TokenFactor would be merged into its user, consider the user later. Summary: A number of optimizations are inhibited by single-use TokenFactors not being merged into the TokenFactor using it. This makes we consider if we can do the merge immediately. Most tests changes here are due to the change in visitation causing minor reorderings and associated reassociation of paired memory operations. CodeGen tests with non-reordering changes: X86/aligned-variadic.ll -- memory-based add folded into stored leaq value. X86/constant-combiners.ll -- Optimizes out overlap between stores. X86/pr40631_deadstore_elision -- folds constant byte store into preceding quad word constant store. Reviewers: RKSimon, craig.topper, spatel, efriedma, courbet Reviewed By: courbet Subscribers: dylanmckay, sdardis, nemanjai, jvesely, nhaehnle, javed.absar, eraman, hiraditya, kbarton, jrtc27, atanasyan, jsji, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D59260 llvm-svn: 356068
2019-03-14 01:07:09 +08:00
; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm4 = ymm0[1,1,2,2]
[x86] allow vector load narrowing with multi-use values This is a long-awaited follow-up suggested in D33578. Since then, we've picked up even more opportunities for vector narrowing from changes like D53784, so there are a lot of test diffs. Apart from 2-3 strange cases, these are all wins. I've structured this to be no-functional-change-intended for any target except for x86 because I couldn't tell if AArch64, ARM, and AMDGPU would improve or not. All of those targets have existing regression tests (4, 4, 10 files respectively) that would be affected. Also, Hexagon overrides the shouldReduceLoadWidth() hook, but doesn't show any regression test diffs. The trade-off is deciding if an extra vector load is better than a single wide load + extract_subvector. For x86, this is almost always better (on paper at least) because we often can fold loads into subsequent ops and not increase the official instruction count. There's also some unknown -- but potentially large -- benefit from using narrower vector ops if wide ops are implemented with multiple uops and/or frequency throttling is avoided. Differential Revision: https://reviews.llvm.org/D54073 llvm-svn: 346595
2018-11-11 04:05:31 +08:00
; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[1,1,2,2]
[DAGCombiner] If a TokenFactor would be merged into its user, consider the user later. Summary: A number of optimizations are inhibited by single-use TokenFactors not being merged into the TokenFactor using it. This makes we consider if we can do the merge immediately. Most tests changes here are due to the change in visitation causing minor reorderings and associated reassociation of paired memory operations. CodeGen tests with non-reordering changes: X86/aligned-variadic.ll -- memory-based add folded into stored leaq value. X86/constant-combiners.ll -- Optimizes out overlap between stores. X86/pr40631_deadstore_elision -- folds constant byte store into preceding quad word constant store. Reviewers: RKSimon, craig.topper, spatel, efriedma, courbet Reviewed By: courbet Subscribers: dylanmckay, sdardis, nemanjai, jvesely, nhaehnle, javed.absar, eraman, hiraditya, kbarton, jrtc27, atanasyan, jsji, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D59260 llvm-svn: 356068
2019-03-14 01:07:09 +08:00
; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0],ymm4[1],ymm2[2,3],ymm4[4],ymm2[5,6],ymm4[7]
; AVX2-FAST-NEXT: vpermilps {{.*#+}} ymm4 = ymm1[0,0,3,3,4,4,7,7]
; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1],ymm4[2],ymm2[3,4],ymm4[5],ymm2[6,7]
; AVX2-FAST-NEXT: vmovaps {{.*#+}} ymm4 = [1,0,2,2,1,0,2,2]
; AVX2-FAST-NEXT: vpermps %ymm1, %ymm4, %ymm1
; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,0,2,1]
; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2,3],ymm1[4],ymm0[5,6],ymm1[7]
; AVX2-FAST-NEXT: vbroadcastsd (%rcx), %ymm1
[x86] allow vector load narrowing with multi-use values This is a long-awaited follow-up suggested in D33578. Since then, we've picked up even more opportunities for vector narrowing from changes like D53784, so there are a lot of test diffs. Apart from 2-3 strange cases, these are all wins. I've structured this to be no-functional-change-intended for any target except for x86 because I couldn't tell if AArch64, ARM, and AMDGPU would improve or not. All of those targets have existing regression tests (4, 4, 10 files respectively) that would be affected. Also, Hexagon overrides the shouldReduceLoadWidth() hook, but doesn't show any regression test diffs. The trade-off is deciding if an extra vector load is better than a single wide load + extract_subvector. For x86, this is almost always better (on paper at least) because we often can fold loads into subsequent ops and not increase the official instruction count. There's also some unknown -- but potentially large -- benefit from using narrower vector ops if wide ops are implemented with multiple uops and/or frequency throttling is avoided. Differential Revision: https://reviews.llvm.org/D54073 llvm-svn: 346595
2018-11-11 04:05:31 +08:00
; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4],ymm1[5],ymm0[6,7]
[DAGCombiner] If a TokenFactor would be merged into its user, consider the user later. Summary: A number of optimizations are inhibited by single-use TokenFactors not being merged into the TokenFactor using it. This makes we consider if we can do the merge immediately. Most tests changes here are due to the change in visitation causing minor reorderings and associated reassociation of paired memory operations. CodeGen tests with non-reordering changes: X86/aligned-variadic.ll -- memory-based add folded into stored leaq value. X86/constant-combiners.ll -- Optimizes out overlap between stores. X86/pr40631_deadstore_elision -- folds constant byte store into preceding quad word constant store. Reviewers: RKSimon, craig.topper, spatel, efriedma, courbet Reviewed By: courbet Subscribers: dylanmckay, sdardis, nemanjai, jvesely, nhaehnle, javed.absar, eraman, hiraditya, kbarton, jrtc27, atanasyan, jsji, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D59260 llvm-svn: 356068
2019-03-14 01:07:09 +08:00
; AVX2-FAST-NEXT: vmovups %ymm0, (%rdi)
; AVX2-FAST-NEXT: vmovups %ymm2, 32(%rdi)
[x86] allow vector load narrowing with multi-use values This is a long-awaited follow-up suggested in D33578. Since then, we've picked up even more opportunities for vector narrowing from changes like D53784, so there are a lot of test diffs. Apart from 2-3 strange cases, these are all wins. I've structured this to be no-functional-change-intended for any target except for x86 because I couldn't tell if AArch64, ARM, and AMDGPU would improve or not. All of those targets have existing regression tests (4, 4, 10 files respectively) that would be affected. Also, Hexagon overrides the shouldReduceLoadWidth() hook, but doesn't show any regression test diffs. The trade-off is deciding if an extra vector load is better than a single wide load + extract_subvector. For x86, this is almost always better (on paper at least) because we often can fold loads into subsequent ops and not increase the official instruction count. There's also some unknown -- but potentially large -- benefit from using narrower vector ops if wide ops are implemented with multiple uops and/or frequency throttling is avoided. Differential Revision: https://reviews.llvm.org/D54073 llvm-svn: 346595
2018-11-11 04:05:31 +08:00
; AVX2-FAST-NEXT: vmovups %ymm3, 64(%rdi)
X86 Tests: Update more isel tests with FastVariableShuffle feature Summary: Added the FastVariableShuffle feature to cases that resembled processors for which this fearure is on. For AVX2 there are processors with and w/o this fearue enable. For AVX512 only KNL does enable this feature so cases which only have +avx512f were left without the FastVariableShuffle enabled. Reviewers: RKSimon, craig.topper Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D41851 llvm-svn: 322090
2018-01-10 00:26:06 +08:00
; AVX2-FAST-NEXT: vzeroupper
; AVX2-FAST-NEXT: retq
[X86][XOP] Add XOP oddshuffles tests XOP codegen is often different to generic AVX - thank you vpperm! llvm-svn: 315176
2017-10-08 20:58:15 +08:00
;
; XOP-LABEL: interleave_24i32_in:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; XOP: # %bb.0:
[x86] allow vector load narrowing with multi-use values This is a long-awaited follow-up suggested in D33578. Since then, we've picked up even more opportunities for vector narrowing from changes like D53784, so there are a lot of test diffs. Apart from 2-3 strange cases, these are all wins. I've structured this to be no-functional-change-intended for any target except for x86 because I couldn't tell if AArch64, ARM, and AMDGPU would improve or not. All of those targets have existing regression tests (4, 4, 10 files respectively) that would be affected. Also, Hexagon overrides the shouldReduceLoadWidth() hook, but doesn't show any regression test diffs. The trade-off is deciding if an extra vector load is better than a single wide load + extract_subvector. For x86, this is almost always better (on paper at least) because we often can fold loads into subsequent ops and not increase the official instruction count. There's also some unknown -- but potentially large -- benefit from using narrower vector ops if wide ops are implemented with multiple uops and/or frequency throttling is avoided. Differential Revision: https://reviews.llvm.org/D54073 llvm-svn: 346595
2018-11-11 04:05:31 +08:00
; XOP-NEXT: vmovupd (%rsi), %ymm0
[X86][AVX] Add VMOVDDUP-VPBROADCASTQ execution domain mapping Noticed in D57514. Differential Revision: https://reviews.llvm.org/D57519 llvm-svn: 352922
2019-02-02 05:41:30 +08:00
; XOP-NEXT: vmovups (%rcx), %ymm1
[DAGCombiner] If a TokenFactor would be merged into its user, consider the user later. Summary: A number of optimizations are inhibited by single-use TokenFactors not being merged into the TokenFactor using it. This makes we consider if we can do the merge immediately. Most tests changes here are due to the change in visitation causing minor reorderings and associated reassociation of paired memory operations. CodeGen tests with non-reordering changes: X86/aligned-variadic.ll -- memory-based add folded into stored leaq value. X86/constant-combiners.ll -- Optimizes out overlap between stores. X86/pr40631_deadstore_elision -- folds constant byte store into preceding quad word constant store. Reviewers: RKSimon, craig.topper, spatel, efriedma, courbet Reviewed By: courbet Subscribers: dylanmckay, sdardis, nemanjai, jvesely, nhaehnle, javed.absar, eraman, hiraditya, kbarton, jrtc27, atanasyan, jsji, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D59260 llvm-svn: 356068
2019-03-14 01:07:09 +08:00
; XOP-NEXT: vmovups 16(%rcx), %xmm2
; XOP-NEXT: vmovups (%rdx), %xmm3
; XOP-NEXT: vmovups 16(%rdx), %xmm4
; XOP-NEXT: vshufps {{.*#+}} xmm5 = xmm4[3,0],xmm2[3,0]
; XOP-NEXT: vshufps {{.*#+}} xmm5 = xmm2[2,1],xmm5[0,2]
; XOP-NEXT: vshufps {{.*#+}} xmm2 = xmm2[1,0],xmm4[1,0]
; XOP-NEXT: vshufps {{.*#+}} xmm2 = xmm2[2,0],xmm4[2,2]
Recommit r354363 "[X86][SSE] Generalize X86ISD::BLENDI support to more value types" And its follow ups r354511, r354640. A follow patch will fix the issue that caused it to be reverted. llvm-svn: 354737
2019-02-24 05:41:42 +08:00
; XOP-NEXT: vinsertf128 $1, %xmm5, %ymm2, %ymm2
; XOP-NEXT: vpermilpd {{.*#+}} ymm4 = ymm0[1,1,3,3]
; XOP-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3,2,3]
[DAGCombiner] If a TokenFactor would be merged into its user, consider the user later. Summary: A number of optimizations are inhibited by single-use TokenFactors not being merged into the TokenFactor using it. This makes we consider if we can do the merge immediately. Most tests changes here are due to the change in visitation causing minor reorderings and associated reassociation of paired memory operations. CodeGen tests with non-reordering changes: X86/aligned-variadic.ll -- memory-based add folded into stored leaq value. X86/constant-combiners.ll -- Optimizes out overlap between stores. X86/pr40631_deadstore_elision -- folds constant byte store into preceding quad word constant store. Reviewers: RKSimon, craig.topper, spatel, efriedma, courbet Reviewed By: courbet Subscribers: dylanmckay, sdardis, nemanjai, jvesely, nhaehnle, javed.absar, eraman, hiraditya, kbarton, jrtc27, atanasyan, jsji, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D59260 llvm-svn: 356068
2019-03-14 01:07:09 +08:00
; XOP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1],ymm4[2],ymm2[3,4],ymm4[5],ymm2[6,7]
; XOP-NEXT: vmovups (%rsi), %xmm4
; XOP-NEXT: vshufps {{.*#+}} xmm5 = xmm4[2,0],xmm3[2,0]
; XOP-NEXT: vshufps {{.*#+}} xmm5 = xmm3[1,1],xmm5[0,2]
; XOP-NEXT: vshufps {{.*#+}} xmm3 = xmm3[0,0],xmm4[0,0]
; XOP-NEXT: vshufps {{.*#+}} xmm3 = xmm3[2,0],xmm4[2,1]
; XOP-NEXT: vinsertf128 $1, %xmm5, %ymm3, %ymm3
; XOP-NEXT: vmovddup {{.*#+}} xmm4 = xmm1[0,0]
; XOP-NEXT: vinsertf128 $1, %xmm4, %ymm4, %ymm4
[X86][XOP] Add XOP oddshuffles tests XOP codegen is often different to generic AVX - thank you vpperm! llvm-svn: 315176
2017-10-08 20:58:15 +08:00
; XOP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1],ymm4[2],ymm3[3,4],ymm4[5],ymm3[6,7]
[x86] allow vector load narrowing with multi-use values This is a long-awaited follow-up suggested in D33578. Since then, we've picked up even more opportunities for vector narrowing from changes like D53784, so there are a lot of test diffs. Apart from 2-3 strange cases, these are all wins. I've structured this to be no-functional-change-intended for any target except for x86 because I couldn't tell if AArch64, ARM, and AMDGPU would improve or not. All of those targets have existing regression tests (4, 4, 10 files respectively) that would be affected. Also, Hexagon overrides the shouldReduceLoadWidth() hook, but doesn't show any regression test diffs. The trade-off is deciding if an extra vector load is better than a single wide load + extract_subvector. For x86, this is almost always better (on paper at least) because we often can fold loads into subsequent ops and not increase the official instruction count. There's also some unknown -- but potentially large -- benefit from using narrower vector ops if wide ops are implemented with multiple uops and/or frequency throttling is avoided. Differential Revision: https://reviews.llvm.org/D54073 llvm-svn: 346595
2018-11-11 04:05:31 +08:00
; XOP-NEXT: vpermil2ps {{.*#+}} ymm0 = ymm1[2],ymm0[3],ymm1[2,3],ymm0[4],ymm1[5,4],ymm0[5]
; XOP-NEXT: vpermilps {{.*#+}} ymm1 = mem[0,0,3,3,4,4,7,7]
[X86][XOP] Add XOP oddshuffles tests XOP codegen is often different to generic AVX - thank you vpperm! llvm-svn: 315176
2017-10-08 20:58:15 +08:00
; XOP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4],ymm1[5],ymm0[6,7]
; XOP-NEXT: vmovups %ymm0, 32(%rdi)
[DAGCombiner] If a TokenFactor would be merged into its user, consider the user later. Summary: A number of optimizations are inhibited by single-use TokenFactors not being merged into the TokenFactor using it. This makes we consider if we can do the merge immediately. Most tests changes here are due to the change in visitation causing minor reorderings and associated reassociation of paired memory operations. CodeGen tests with non-reordering changes: X86/aligned-variadic.ll -- memory-based add folded into stored leaq value. X86/constant-combiners.ll -- Optimizes out overlap between stores. X86/pr40631_deadstore_elision -- folds constant byte store into preceding quad word constant store. Reviewers: RKSimon, craig.topper, spatel, efriedma, courbet Reviewed By: courbet Subscribers: dylanmckay, sdardis, nemanjai, jvesely, nhaehnle, javed.absar, eraman, hiraditya, kbarton, jrtc27, atanasyan, jsji, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D59260 llvm-svn: 356068
2019-03-14 01:07:09 +08:00
; XOP-NEXT: vmovups %ymm3, (%rdi)
; XOP-NEXT: vmovups %ymm2, 64(%rdi)
[X86][XOP] Add XOP oddshuffles tests XOP codegen is often different to generic AVX - thank you vpperm! llvm-svn: 315176
2017-10-08 20:58:15 +08:00
; XOP-NEXT: vzeroupper
; XOP-NEXT: retq
[X86] Add more baseline tests for "irregular" shuffles. NFC. This adds more tests for shuffles where the output width does not match the input width and/or the output is generated from more than two inputs. llvm-svn: 281005
2016-09-09 08:49:29 +08:00
%s1 = load <8 x i32>, <8 x i32>* %q1, align 4
%s2 = load <8 x i32>, <8 x i32>* %q2, align 4
%s3 = load <8 x i32>, <8 x i32>* %q3, align 4
%t1 = shufflevector <8 x i32> %s1, <8 x i32> %s2, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
%t2 = shufflevector <8 x i32> %s3, <8 x i32> undef, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
%interleaved = shufflevector <16 x i32> %t1, <16 x i32> %t2, <24 x i32> <i32 0, i32 8, i32 16, i32 1, i32 9, i32 17, i32 2, i32 10, i32 18, i32 3, i32 11, i32 19, i32 4, i32 12, i32 20, i32 5, i32 13, i32 21, i32 6, i32 14, i32 22, i32 7, i32 15, i32 23>
store <24 x i32> %interleaved, <24 x i32>* %p, align 4
ret void
}
[DAG] Fix crash in build_vector -> vector_shuffle combine Fixes a crash in the build_vector -> vector_shuffle combine when the first vector input is twice as wide as the output, and the second input vector is even wider. llvm-svn: 283953
2016-10-12 06:44:31 +08:00
define <2 x double> @wrongorder(<4 x double> %A, <8 x double>* %P) #0 {
; SSE2-LABEL: wrongorder:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/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:
[DAG] Fix crash in build_vector -> vector_shuffle combine Fixes a crash in the build_vector -> vector_shuffle combine when the first vector input is twice as wide as the output, and the second input vector is even wider. llvm-svn: 283953
2016-10-12 06:44:31 +08:00
; SSE2-NEXT: movlhps {{.*#+}} xmm0 = xmm0[0,0]
; SSE2-NEXT: movaps %xmm0, 48(%rdi)
; SSE2-NEXT: movaps %xmm0, 32(%rdi)
; SSE2-NEXT: movaps %xmm0, 16(%rdi)
; SSE2-NEXT: movaps %xmm0, (%rdi)
; SSE2-NEXT: retq
;
; SSE42-LABEL: wrongorder:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; SSE42: # %bb.0:
[DAG] Fix crash in build_vector -> vector_shuffle combine Fixes a crash in the build_vector -> vector_shuffle combine when the first vector input is twice as wide as the output, and the second input vector is even wider. llvm-svn: 283953
2016-10-12 06:44:31 +08:00
; SSE42-NEXT: movddup {{.*#+}} xmm0 = xmm0[0,0]
; SSE42-NEXT: movapd %xmm0, 48(%rdi)
; SSE42-NEXT: movapd %xmm0, 32(%rdi)
; SSE42-NEXT: movapd %xmm0, 16(%rdi)
; SSE42-NEXT: movapd %xmm0, (%rdi)
; SSE42-NEXT: retq
;
; AVX1-LABEL: wrongorder:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/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:
[DAG] Fix crash in build_vector -> vector_shuffle combine Fixes a crash in the build_vector -> vector_shuffle combine when the first vector input is twice as wide as the output, and the second input vector is even wider. llvm-svn: 283953
2016-10-12 06:44:31 +08:00
; AVX1-NEXT: vmovddup {{.*#+}} xmm0 = xmm0[0,0]
; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm1
[X86] Don't base domain decisions on VEXTRACTF128/VINSERTF128 if only AVX1 is available. Seems the execution dependency pass likes to use FP instructions when most of the consuming code is integer if a vextractf128 instruction produced the register. Without AVX2 we don't have the corresponding integer instruction available. This patch suppresses the domain on these instructions to GenericDomain if AVX2 is not supported so that they are ignored by domain fixing. If AVX2 is supported we'll report the correct domain and allow them to switch between integer and fp. Overall I think this produces better results in the modified test cases. llvm-svn: 294824
2017-02-11 13:32:57 +08:00
; AVX1-NEXT: vmovaps %ymm1, 32(%rdi)
; AVX1-NEXT: vmovaps %ymm1, (%rdi)
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; AVX1-NEXT: # kill: def $xmm0 killed $xmm0 killed $ymm0
[DAG] Fix crash in build_vector -> vector_shuffle combine Fixes a crash in the build_vector -> vector_shuffle combine when the first vector input is twice as wide as the output, and the second input vector is even wider. llvm-svn: 283953
2016-10-12 06:44:31 +08:00
; AVX1-NEXT: vzeroupper
; AVX1-NEXT: retq
;
; AVX2-LABEL: wrongorder:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/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][AVX] Attempt to share broadcasts of different widths (PR39454) If we have broadcasts of different vector widths, keep the longest vector width and extract subvectors for the shorter vectors (which should be free). Differential Revision: https://reviews.llvm.org/D57663 llvm-svn: 353154
2019-02-05 18:58:43 +08:00
; AVX2-NEXT: vbroadcastsd %xmm0, %ymm0
; AVX2-NEXT: vmovaps %ymm0, 32(%rdi)
; AVX2-NEXT: vmovaps %ymm0, (%rdi)
; AVX2-NEXT: # kill: def $xmm0 killed $xmm0 killed $ymm0
[DAG] Fix crash in build_vector -> vector_shuffle combine Fixes a crash in the build_vector -> vector_shuffle combine when the first vector input is twice as wide as the output, and the second input vector is even wider. llvm-svn: 283953
2016-10-12 06:44:31 +08:00
; AVX2-NEXT: vzeroupper
; AVX2-NEXT: retq
[X86][XOP] Add XOP oddshuffles tests XOP codegen is often different to generic AVX - thank you vpperm! llvm-svn: 315176
2017-10-08 20:58:15 +08:00
;
; XOP-LABEL: wrongorder:
[CodeGen] Unify MBB reference format in both MIR and debug output As part of the unification of the debug format and the MIR format, print MBB references as '%bb.5'. The MIR printer prints the IR name of a MBB only for block definitions. * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g' * find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g' * find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g' * grep -nr 'BB#' and fix Differential Revision: https://reviews.llvm.org/D40422 llvm-svn: 319665
2017-12-05 01:18:51 +08:00
; XOP: # %bb.0:
[X86][XOP] Add XOP oddshuffles tests XOP codegen is often different to generic AVX - thank you vpperm! llvm-svn: 315176
2017-10-08 20:58:15 +08:00
; XOP-NEXT: vmovddup {{.*#+}} xmm0 = xmm0[0,0]
; XOP-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm1
; XOP-NEXT: vmovaps %ymm1, 32(%rdi)
; XOP-NEXT: vmovaps %ymm1, (%rdi)
Followup on Proposal to move MIR physical register namespace to '$' sigil. Discussed here: http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html In preparation for adding support for named vregs we are changing the sigil for physical registers in MIR to '$' from '%'. This will prevent name clashes of named physical register with named vregs. llvm-svn: 323922
2018-02-01 06:04:26 +08:00
; XOP-NEXT: # kill: def $xmm0 killed $xmm0 killed $ymm0
[X86][XOP] Add XOP oddshuffles tests XOP codegen is often different to generic AVX - thank you vpperm! llvm-svn: 315176
2017-10-08 20:58:15 +08:00
; XOP-NEXT: vzeroupper
; XOP-NEXT: retq
[DAG] Fix crash in build_vector -> vector_shuffle combine Fixes a crash in the build_vector -> vector_shuffle combine when the first vector input is twice as wide as the output, and the second input vector is even wider. llvm-svn: 283953
2016-10-12 06:44:31 +08:00
%shuffle = shufflevector <4 x double> %A, <4 x double> %A, <8 x i32> zeroinitializer
store <8 x double> %shuffle, <8 x double>* %P, align 64
%m2 = load <8 x double>, <8 x double>* %P, align 64
store <8 x double> %m2, <8 x double>* %P, align 64
%m3 = load <8 x double>, <8 x double>* %P, align 64
%m4 = shufflevector <8 x double> %m3, <8 x double> undef, <2 x i32> <i32 2, i32 0>
ret <2 x double> %m4
}
[DAGCombine] combineShuffleOfScalars - handle non-zero SCALAR_TO_VECTOR indices (PR41097) rL356292 reduces the size of scalar_to_vector if we know the upper bits are undef - which means that shuffles may find they are suddenly referencing scalar_to_vector elements other than zero - so make sure we handle this as undef. llvm-svn: 356327
2019-03-17 01:36:26 +08:00
define void @PR41097() {
; SSE2-LABEL: PR41097:
; SSE2: # %bb.0:
[DAGCombine] Prune unnused nodes. Summary: Nodes that have no uses are eventually pruned when they are selected from the worklist. Record nodes newly added to the worklist or DAG and perform pruning after every combine attempt. Reviewers: efriedma, RKSimon, craig.topper, spatel, jyknight Reviewed By: jyknight Subscribers: jdoerfert, jyknight, nemanjai, jvesely, nhaehnle, javed.absar, hiraditya, jsji, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D58070 llvm-svn: 357283
2019-03-30 01:35:56 +08:00
; SSE2-NEXT: movd {{.*#+}} xmm0 = mem[0],zero,zero,zero
[DAGCombine] combineShuffleOfScalars - handle non-zero SCALAR_TO_VECTOR indices (PR41097) rL356292 reduces the size of scalar_to_vector if we know the upper bits are undef - which means that shuffles may find they are suddenly referencing scalar_to_vector elements other than zero - so make sure we handle this as undef. llvm-svn: 356327
2019-03-17 01:36:26 +08:00
; SSE2-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
; SSE2-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[0,0,2,3,4,5,6,7]
; SSE2-NEXT: psrad $24, %xmm0
[DAGCombine] Prune unnused nodes. Summary: Nodes that have no uses are eventually pruned when they are selected from the worklist. Record nodes newly added to the worklist or DAG and perform pruning after every combine attempt. Reviewers: efriedma, RKSimon, craig.topper, spatel, jyknight Reviewed By: jyknight Subscribers: jdoerfert, jyknight, nemanjai, jvesely, nhaehnle, javed.absar, hiraditya, jsji, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D58070 llvm-svn: 357283
2019-03-30 01:35:56 +08:00
; SSE2-NEXT: pxor %xmm1, %xmm1
[DAGCombine] combineShuffleOfScalars - handle non-zero SCALAR_TO_VECTOR indices (PR41097) rL356292 reduces the size of scalar_to_vector if we know the upper bits are undef - which means that shuffles may find they are suddenly referencing scalar_to_vector elements other than zero - so make sure we handle this as undef. llvm-svn: 356327
2019-03-17 01:36:26 +08:00
; SSE2-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; SSE2-NEXT: movdqu %xmm0, (%rax)
; SSE2-NEXT: retq
;
; SSE42-LABEL: PR41097:
; SSE42: # %bb.0:
; SSE42-NEXT: movd {{.*#+}} xmm0 = mem[0],zero,zero,zero
; SSE42-NEXT: pshufb {{.*#+}} xmm0 = xmm0[0,3,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; SSE42-NEXT: pmovsxbd %xmm0, %xmm0
; SSE42-NEXT: pmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
; SSE42-NEXT: movdqu %xmm0, (%rax)
; SSE42-NEXT: retq
;
; AVX-LABEL: PR41097:
; AVX: # %bb.0:
; AVX-NEXT: vmovd {{.*#+}} xmm0 = mem[0],zero,zero,zero
; AVX-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,3,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX-NEXT: vpmovsxbd %xmm0, %xmm0
; AVX-NEXT: vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
; AVX-NEXT: vmovdqu %xmm0, (%rax)
; AVX-NEXT: retq
;
; XOP-LABEL: PR41097:
; XOP: # %bb.0:
; XOP-NEXT: vmovd {{.*#+}} xmm0 = mem[0],zero,zero,zero
; XOP-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,3,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; XOP-NEXT: vpmovsxbd %xmm0, %xmm0
; XOP-NEXT: vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
; XOP-NEXT: vmovdqu %xmm0, (%rax)
; XOP-NEXT: retq
%wide.vec = load <6 x i8>, <6 x i8>* undef, align 1
%strided.vec = shufflevector <6 x i8> %wide.vec, <6 x i8> undef, <2 x i32> <i32 0, i32 3>
%tmp = sext <2 x i8> %strided.vec to <2 x i32>
%tmp7 = zext <2 x i32> %tmp to <2 x i64>
store <2 x i64> %tmp7, <2 x i64>* undef, align 8
ret void
}