maxjhandsome
/
llvm-project
forked from OSchip/llvm-project
1
0
Fork 0
Code Issues Pull Requests Packages Releases Wiki Activity
llvm-project/llvm/test/CodeGen/X86/vector-shuffle-combining-av...

158 lines
12 KiB
LLVM
Raw Normal View History

[X86][AVX512] Add some initial VBMI target shuffle combine tests llvm-svn: 287494
2016-11-20 22:45:46 +08:00
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
; RUN: llc < %s -mtriple=i686-unknown -mattr=+avx512vbmi,+avx512vl | FileCheck %s --check-prefix=X32
; RUN: llc < %s -mtriple=x86_64-unknown -mattr=+avx512vbmi,+avx512vl | FileCheck %s --check-prefix=X64
declare <16 x i8> @llvm.x86.avx512.mask.permvar.qi.128(<16 x i8>, <16 x i8>, <16 x i8>, i16)
declare <16 x i8> @llvm.x86.avx512.mask.vpermi2var.qi.128(<16 x i8>, <16 x i8>, <16 x i8>, i16)
declare <16 x i8> @llvm.x86.avx512.mask.vpermt2var.qi.128(<16 x i8>, <16 x i8>, <16 x i8>, i16)
declare <16 x i8> @llvm.x86.avx512.maskz.vpermt2var.qi.128(<16 x i8>, <16 x i8>, <16 x i8>, i16)
declare <32 x i8> @llvm.x86.avx512.mask.permvar.qi.256(<32 x i8>, <32 x i8>, <32 x i8>, i32)
declare <32 x i8> @llvm.x86.avx512.mask.vpermi2var.qi.256(<32 x i8>, <32 x i8>, <32 x i8>, i32)
declare <32 x i8> @llvm.x86.avx512.mask.vpermt2var.qi.256(<32 x i8>, <32 x i8>, <32 x i8>, i32)
declare <32 x i8> @llvm.x86.avx512.maskz.vpermt2var.qi.256(<32 x i8>, <32 x i8>, <32 x i8>, i32)
declare <64 x i8> @llvm.x86.avx512.mask.permvar.qi.512(<64 x i8>, <64 x i8>, <64 x i8>, i64)
declare <64 x i8> @llvm.x86.avx512.mask.vpermi2var.qi.512(<64 x i8>, <64 x i8>, <64 x i8>, i64)
declare <64 x i8> @llvm.x86.avx512.mask.vpermt2var.qi.512(<64 x i8>, <64 x i8>, <64 x i8>, i64)
declare <64 x i8> @llvm.x86.avx512.maskz.vpermt2var.qi.512(<64 x i8>, <64 x i8>, <64 x i8>, i64)
declare <16 x i8> @llvm.x86.ssse3.pshuf.b.128(<16 x i8>, <16 x i8>)
declare <32 x i8> @llvm.x86.avx2.pshuf.b(<32 x i8>, <32 x i8>)
declare <64 x i8> @llvm.x86.avx512.mask.pshuf.b.512(<64 x i8>, <64 x i8>, <64 x i8>, i64)
define <16 x i8> @combine_vpermt2var_16i8_identity(<16 x i8> %x0, <16 x i8> %x1) {
; X32-LABEL: combine_vpermt2var_16i8_identity:
; X32: # BB#0:
; X32-NEXT: retl
;
; X64-LABEL: combine_vpermt2var_16i8_identity:
; X64: # BB#0:
; X64-NEXT: retq
%res0 = call <16 x i8> @llvm.x86.avx512.maskz.vpermt2var.qi.128(<16 x i8> <i8 15, i8 14, i8 13, i8 12, i8 11, i8 10, i8 9, i8 8, i8 7, i8 6, i8 5, i8 4, i8 3, i8 2, i8 1, i8 0>, <16 x i8> %x0, <16 x i8> %x1, i16 -1)
%res1 = call <16 x i8> @llvm.x86.avx512.maskz.vpermt2var.qi.128(<16 x i8> <i8 15, i8 30, i8 13, i8 28, i8 11, i8 26, i8 9, i8 24, i8 7, i8 22, i8 5, i8 20, i8 3, i8 18, i8 1, i8 16>, <16 x i8> %res0, <16 x i8> %res0, i16 -1)
ret <16 x i8> %res1
}
define <16 x i8> @combine_vpermt2var_16i8_identity_mask(<16 x i8> %x0, <16 x i8> %x1, i16 %m) {
; X32-LABEL: combine_vpermt2var_16i8_identity_mask:
; X32: # BB#0:
; X32-NEXT: kmovw {{[0-9]+}}(%esp), %k1
[AVX-512] Remove patterns that select vmovdqu8/16 for unmasked loads. Prefer vmovdqa64/vmovdqu64 instead. These were taking priority over the aligned load instructions since there is no vmovda8/16. I don't think there is really a difference between aligned and unaligned on newer cpus so I don't think it matters which instructions we use. But with this change we reduce the size of the isel table a little and we allow the aligned information to pass through to the evex->vec pass and produce the same output has avx/avx2 in some cases. I also generally dislike patterns rooted in a bitcast which these were. Differential Revision: https://reviews.llvm.org/D35977 llvm-svn: 309589
2017-08-01 01:35:44 +08:00
; X32-NEXT: vmovdqa {{.*#+}} xmm2 = [15,14,13,12,11,10,9,8,7,6,5,4,3,2,1,0]
[AVX-512] Add support for commuting VPERMT2(B/W/D/Q/PS/PD) to/from VPERMI2(B/W/D/Q/PS/PD). Summary: The index and one of the table operands can be swapped by changing the opcode to the other version. Neither of these operands are the one that can load from memory so this can't be used to increase memory folding opportunities. We need to handle the unmasked forms and the kz forms. Since the load operand isn't being commuted we can commute the load and broadcast instructions too. Reviewers: igorb, delena, Ayal, Farhana, RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D25652 llvm-svn: 287621
2016-11-22 12:57:34 +08:00
; X32-NEXT: vpermi2b %xmm1, %xmm0, %xmm2 {%k1} {z}
[AVX-512] Remove patterns that select vmovdqu8/16 for unmasked loads. Prefer vmovdqa64/vmovdqu64 instead. These were taking priority over the aligned load instructions since there is no vmovda8/16. I don't think there is really a difference between aligned and unaligned on newer cpus so I don't think it matters which instructions we use. But with this change we reduce the size of the isel table a little and we allow the aligned information to pass through to the evex->vec pass and produce the same output has avx/avx2 in some cases. I also generally dislike patterns rooted in a bitcast which these were. Differential Revision: https://reviews.llvm.org/D35977 llvm-svn: 309589
2017-08-01 01:35:44 +08:00
; X32-NEXT: vmovdqa {{.*#+}} xmm0 = [15,30,13,28,11,26,9,24,7,22,5,20,3,18,1,16]
[AVX-512] Add support for commuting VPERMT2(B/W/D/Q/PS/PD) to/from VPERMI2(B/W/D/Q/PS/PD). Summary: The index and one of the table operands can be swapped by changing the opcode to the other version. Neither of these operands are the one that can load from memory so this can't be used to increase memory folding opportunities. We need to handle the unmasked forms and the kz forms. Since the load operand isn't being commuted we can commute the load and broadcast instructions too. Reviewers: igorb, delena, Ayal, Farhana, RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D25652 llvm-svn: 287621
2016-11-22 12:57:34 +08:00
; X32-NEXT: vpermi2b %xmm2, %xmm2, %xmm0 {%k1} {z}
[X86][AVX512] Add some initial VBMI target shuffle combine tests llvm-svn: 287494
2016-11-20 22:45:46 +08:00
; X32-NEXT: retl
;
; X64-LABEL: combine_vpermt2var_16i8_identity_mask:
; X64: # BB#0:
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; X64-NEXT: kmovd %edi, %k1
[AVX-512] Remove patterns that select vmovdqu8/16 for unmasked loads. Prefer vmovdqa64/vmovdqu64 instead. These were taking priority over the aligned load instructions since there is no vmovda8/16. I don't think there is really a difference between aligned and unaligned on newer cpus so I don't think it matters which instructions we use. But with this change we reduce the size of the isel table a little and we allow the aligned information to pass through to the evex->vec pass and produce the same output has avx/avx2 in some cases. I also generally dislike patterns rooted in a bitcast which these were. Differential Revision: https://reviews.llvm.org/D35977 llvm-svn: 309589
2017-08-01 01:35:44 +08:00
; X64-NEXT: vmovdqa {{.*#+}} xmm2 = [15,14,13,12,11,10,9,8,7,6,5,4,3,2,1,0]
[AVX-512] Add support for commuting VPERMT2(B/W/D/Q/PS/PD) to/from VPERMI2(B/W/D/Q/PS/PD). Summary: The index and one of the table operands can be swapped by changing the opcode to the other version. Neither of these operands are the one that can load from memory so this can't be used to increase memory folding opportunities. We need to handle the unmasked forms and the kz forms. Since the load operand isn't being commuted we can commute the load and broadcast instructions too. Reviewers: igorb, delena, Ayal, Farhana, RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D25652 llvm-svn: 287621
2016-11-22 12:57:34 +08:00
; X64-NEXT: vpermi2b %xmm1, %xmm0, %xmm2 {%k1} {z}
[AVX-512] Remove patterns that select vmovdqu8/16 for unmasked loads. Prefer vmovdqa64/vmovdqu64 instead. These were taking priority over the aligned load instructions since there is no vmovda8/16. I don't think there is really a difference between aligned and unaligned on newer cpus so I don't think it matters which instructions we use. But with this change we reduce the size of the isel table a little and we allow the aligned information to pass through to the evex->vec pass and produce the same output has avx/avx2 in some cases. I also generally dislike patterns rooted in a bitcast which these were. Differential Revision: https://reviews.llvm.org/D35977 llvm-svn: 309589
2017-08-01 01:35:44 +08:00
; X64-NEXT: vmovdqa {{.*#+}} xmm0 = [15,30,13,28,11,26,9,24,7,22,5,20,3,18,1,16]
[AVX-512] Add support for commuting VPERMT2(B/W/D/Q/PS/PD) to/from VPERMI2(B/W/D/Q/PS/PD). Summary: The index and one of the table operands can be swapped by changing the opcode to the other version. Neither of these operands are the one that can load from memory so this can't be used to increase memory folding opportunities. We need to handle the unmasked forms and the kz forms. Since the load operand isn't being commuted we can commute the load and broadcast instructions too. Reviewers: igorb, delena, Ayal, Farhana, RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D25652 llvm-svn: 287621
2016-11-22 12:57:34 +08:00
; X64-NEXT: vpermi2b %xmm2, %xmm2, %xmm0 {%k1} {z}
[X86][AVX512] Add some initial VBMI target shuffle combine tests llvm-svn: 287494
2016-11-20 22:45:46 +08:00
; X64-NEXT: retq
%res0 = call <16 x i8> @llvm.x86.avx512.maskz.vpermt2var.qi.128(<16 x i8> <i8 15, i8 14, i8 13, i8 12, i8 11, i8 10, i8 9, i8 8, i8 7, i8 6, i8 5, i8 4, i8 3, i8 2, i8 1, i8 0>, <16 x i8> %x0, <16 x i8> %x1, i16 %m)
%res1 = call <16 x i8> @llvm.x86.avx512.maskz.vpermt2var.qi.128(<16 x i8> <i8 15, i8 30, i8 13, i8 28, i8 11, i8 26, i8 9, i8 24, i8 7, i8 22, i8 5, i8 20, i8 3, i8 18, i8 1, i8 16>, <16 x i8> %res0, <16 x i8> %res0, i16 %m)
ret <16 x i8> %res1
}
define <16 x i8> @combine_vpermi2var_16i8_as_vpshufb(<16 x i8> %x0, <16 x i8> %x1) {
; X32-LABEL: combine_vpermi2var_16i8_as_vpshufb:
; X32: # BB#0:
; X32-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[15,0,14,1,13,2,12,3,11,4,10,5,9,6,8,7]
; X32-NEXT: retl
;
; X64-LABEL: combine_vpermi2var_16i8_as_vpshufb:
; X64: # BB#0:
; X64-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[15,0,14,1,13,2,12,3,11,4,10,5,9,6,8,7]
; X64-NEXT: retq
%res0 = call <16 x i8> @llvm.x86.avx512.mask.vpermi2var.qi.128(<16 x i8> %x0, <16 x i8> <i8 15, i8 14, i8 13, i8 12, i8 11, i8 10, i8 9, i8 8, i8 7, i8 6, i8 5, i8 4, i8 3, i8 2, i8 1, i8 0>, <16 x i8> %x1, i16 -1)
%res1 = call <16 x i8> @llvm.x86.avx512.mask.vpermi2var.qi.128(<16 x i8> %res0, <16 x i8> <i8 0, i8 15, i8 1, i8 14, i8 2, i8 13, i8 3, i8 12, i8 4, i8 11, i8 5, i8 10, i8 6, i8 9, i8 7, i8 8>, <16 x i8> %res0, i16 -1)
ret <16 x i8> %res1
}
define <32 x i8> @combine_vpermi2var_32i8_as_vpermb(<32 x i8> %x0, <32 x i8> %x1) {
; X32-LABEL: combine_vpermi2var_32i8_as_vpermb:
; X32: # BB#0:
[AVX-512] Remove patterns that select vmovdqu8/16 for unmasked loads. Prefer vmovdqa64/vmovdqu64 instead. These were taking priority over the aligned load instructions since there is no vmovda8/16. I don't think there is really a difference between aligned and unaligned on newer cpus so I don't think it matters which instructions we use. But with this change we reduce the size of the isel table a little and we allow the aligned information to pass through to the evex->vec pass and produce the same output has avx/avx2 in some cases. I also generally dislike patterns rooted in a bitcast which these were. Differential Revision: https://reviews.llvm.org/D35977 llvm-svn: 309589
2017-08-01 01:35:44 +08:00
; X32-NEXT: vmovdqa {{.*#+}} ymm1 = [0,0,1,23,2,22,3,21,4,22,5,21,6,20,7,19,0,0,1,23,2,22,3,21,4,22,5,21,6,20,7,19]
[X86][AVX512] Add support for VBMI VPERMV target shuffle combines llvm-svn: 287495
2016-11-20 23:05:45 +08:00
; X32-NEXT: vpermb %ymm0, %ymm1, %ymm0
[X86][AVX512] Add some initial VBMI target shuffle combine tests llvm-svn: 287494
2016-11-20 22:45:46 +08:00
; X32-NEXT: retl
;
; X64-LABEL: combine_vpermi2var_32i8_as_vpermb:
; X64: # BB#0:
[AVX-512] Remove patterns that select vmovdqu8/16 for unmasked loads. Prefer vmovdqa64/vmovdqu64 instead. These were taking priority over the aligned load instructions since there is no vmovda8/16. I don't think there is really a difference between aligned and unaligned on newer cpus so I don't think it matters which instructions we use. But with this change we reduce the size of the isel table a little and we allow the aligned information to pass through to the evex->vec pass and produce the same output has avx/avx2 in some cases. I also generally dislike patterns rooted in a bitcast which these were. Differential Revision: https://reviews.llvm.org/D35977 llvm-svn: 309589
2017-08-01 01:35:44 +08:00
; X64-NEXT: vmovdqa {{.*#+}} ymm1 = [0,0,1,23,2,22,3,21,4,22,5,21,6,20,7,19,0,0,1,23,2,22,3,21,4,22,5,21,6,20,7,19]
[X86][AVX512] Add support for VBMI VPERMV target shuffle combines llvm-svn: 287495
2016-11-20 23:05:45 +08:00
; X64-NEXT: vpermb %ymm0, %ymm1, %ymm0
[X86][AVX512] Add some initial VBMI target shuffle combine tests llvm-svn: 287494
2016-11-20 22:45:46 +08:00
; X64-NEXT: retq
%res0 = shufflevector <32 x i8> %x0, <32 x i8> %x1, <32 x i32> <i32 0, i32 32, i32 1, i32 33, i32 2, i32 34, i32 3, i32 35, i32 4, i32 36, i32 5, i32 37, i32 6, i32 38, i32 7, i32 39, i32 16, i32 48, i32 17, i32 49, i32 18, i32 50, i32 19, i32 51, i32 20, i32 52, i32 21, i32 53, i32 22, i32 54, i32 23, i32 55>
%res1 = call <32 x i8> @llvm.x86.avx512.mask.vpermi2var.qi.256(<32 x i8> %res0, <32 x i8> <i8 0, i8 32, i8 2, i8 30, i8 4, i8 28, i8 6, i8 26, i8 8, i8 28, i8 10, i8 26, i8 12, i8 24, i8 14, i8 22, i8 0, i8 32, i8 2, i8 30, i8 4, i8 28, i8 6, i8 26, i8 8, i8 28, i8 10, i8 26, i8 12, i8 24, i8 14, i8 22>, <32 x i8> %res0, i32 -1)
ret <32 x i8> %res1
}
define <64 x i8> @combine_vpermi2var_64i8_as_vpermb(<64 x i8> %x0, <64 x i8> %x1) {
; X32-LABEL: combine_vpermi2var_64i8_as_vpermb:
; X32: # BB#0:
[AVX-512] Remove patterns that select vmovdqu8/16 for unmasked loads. Prefer vmovdqa64/vmovdqu64 instead. These were taking priority over the aligned load instructions since there is no vmovda8/16. I don't think there is really a difference between aligned and unaligned on newer cpus so I don't think it matters which instructions we use. But with this change we reduce the size of the isel table a little and we allow the aligned information to pass through to the evex->vec pass and produce the same output has avx/avx2 in some cases. I also generally dislike patterns rooted in a bitcast which these were. Differential Revision: https://reviews.llvm.org/D35977 llvm-svn: 309589
2017-08-01 01:35:44 +08:00
; X32-NEXT: vmovdqa64 {{.*#+}} zmm1 = [0,32,1,23,2,22,3,21,4,22,5,21,6,20,7,19,0,32,1,23,2,22,3,21,4,22,5,21,6,20,7,19,0,32,1,23,2,22,3,21,4,22,5,21,6,20,7,19,0,32,1,23,2,22,3,21,4,22,5,21,6,20,7,19]
[X86][AVX512] Add support for VBMI VPERMV target shuffle combines llvm-svn: 287495
2016-11-20 23:05:45 +08:00
; X32-NEXT: vpermb %zmm0, %zmm1, %zmm0
[X86][AVX512] Add some initial VBMI target shuffle combine tests llvm-svn: 287494
2016-11-20 22:45:46 +08:00
; X32-NEXT: retl
;
; X64-LABEL: combine_vpermi2var_64i8_as_vpermb:
; X64: # BB#0:
[AVX-512] Remove patterns that select vmovdqu8/16 for unmasked loads. Prefer vmovdqa64/vmovdqu64 instead. These were taking priority over the aligned load instructions since there is no vmovda8/16. I don't think there is really a difference between aligned and unaligned on newer cpus so I don't think it matters which instructions we use. But with this change we reduce the size of the isel table a little and we allow the aligned information to pass through to the evex->vec pass and produce the same output has avx/avx2 in some cases. I also generally dislike patterns rooted in a bitcast which these were. Differential Revision: https://reviews.llvm.org/D35977 llvm-svn: 309589
2017-08-01 01:35:44 +08:00
; X64-NEXT: vmovdqa64 {{.*#+}} zmm1 = [0,32,1,23,2,22,3,21,4,22,5,21,6,20,7,19,0,32,1,23,2,22,3,21,4,22,5,21,6,20,7,19,0,32,1,23,2,22,3,21,4,22,5,21,6,20,7,19,0,32,1,23,2,22,3,21,4,22,5,21,6,20,7,19]
[X86][AVX512] Add support for VBMI VPERMV target shuffle combines llvm-svn: 287495
2016-11-20 23:05:45 +08:00
; X64-NEXT: vpermb %zmm0, %zmm1, %zmm0
[X86][AVX512] Add some initial VBMI target shuffle combine tests llvm-svn: 287494
2016-11-20 22:45:46 +08:00
; X64-NEXT: retq
%res0 = shufflevector <64 x i8> %x0, <64 x i8> %x1, <64 x i32> <i32 0, i32 64, i32 1, i32 65, i32 2, i32 66, i32 3, i32 67, i32 4, i32 68, i32 5, i32 69, i32 6, i32 70, i32 7, i32 71, i32 16, i32 80, i32 17, i32 81, i32 18, i32 82, i32 19, i32 83, i32 20, i32 84, i32 21, i32 85, i32 22, i32 86, i32 23, i32 87, i32 32, i32 96, i32 33, i32 97, i32 34, i32 98, i32 35, i32 99, i32 36, i32 100, i32 37, i32 101, i32 38, i32 102, i32 39, i32 103, i32 48, i32 112, i32 49, i32 113, i32 50, i32 114, i32 51, i32 115, i32 52, i32 116, i32 53, i32 117, i32 54, i32 118, i32 55, i32 119>
%res1 = call <64 x i8> @llvm.x86.avx512.mask.vpermi2var.qi.512(<64 x i8> %res0, <64 x i8> <i8 0, i8 32, i8 2, i8 30, i8 4, i8 28, i8 6, i8 26, i8 8, i8 28, i8 10, i8 26, i8 12, i8 24, i8 14, i8 22, i8 0, i8 32, i8 2, i8 30, i8 4, i8 28, i8 6, i8 26, i8 8, i8 28, i8 10, i8 26, i8 12, i8 24, i8 14, i8 22, i8 0, i8 32, i8 2, i8 30, i8 4, i8 28, i8 6, i8 26, i8 8, i8 28, i8 10, i8 26, i8 12, i8 24, i8 14, i8 22, i8 0, i8 32, i8 2, i8 30, i8 4, i8 28, i8 6, i8 26, i8 8, i8 28, i8 10, i8 26, i8 12, i8 24, i8 14, i8 22>, <64 x i8> %res0, i64 -1)
ret <64 x i8> %res1
}
define <16 x i8> @combine_vpermt2var_vpermi2var_16i8_as_vperm2(<16 x i8> %x0, <16 x i8> %x1) {
; X32-LABEL: combine_vpermt2var_vpermi2var_16i8_as_vperm2:
; X32: # BB#0:
[AVX-512] Remove patterns that select vmovdqu8/16 for unmasked loads. Prefer vmovdqa64/vmovdqu64 instead. These were taking priority over the aligned load instructions since there is no vmovda8/16. I don't think there is really a difference between aligned and unaligned on newer cpus so I don't think it matters which instructions we use. But with this change we reduce the size of the isel table a little and we allow the aligned information to pass through to the evex->vec pass and produce the same output has avx/avx2 in some cases. I also generally dislike patterns rooted in a bitcast which these were. Differential Revision: https://reviews.llvm.org/D35977 llvm-svn: 309589
2017-08-01 01:35:44 +08:00
; X32-NEXT: vmovdqa {{.*#+}} xmm2 = [0,31,2,29,4,27,6,25,8,23,10,21,12,19,14,17]
[X86][AVX512] Add some initial VBMI target shuffle combine tests llvm-svn: 287494
2016-11-20 22:45:46 +08:00
; X32-NEXT: vpermi2b %xmm1, %xmm0, %xmm2
[AVX-512] Remove patterns that select vmovdqu8/16 for unmasked loads. Prefer vmovdqa64/vmovdqu64 instead. These were taking priority over the aligned load instructions since there is no vmovda8/16. I don't think there is really a difference between aligned and unaligned on newer cpus so I don't think it matters which instructions we use. But with this change we reduce the size of the isel table a little and we allow the aligned information to pass through to the evex->vec pass and produce the same output has avx/avx2 in some cases. I also generally dislike patterns rooted in a bitcast which these were. Differential Revision: https://reviews.llvm.org/D35977 llvm-svn: 309589
2017-08-01 01:35:44 +08:00
; X32-NEXT: vmovdqa {{.*#+}} xmm0 = [0,17,2,18,4,19,6,21,8,23,10,25,12,27,14,29]
[AVX-512] Add support for commuting VPERMT2(B/W/D/Q/PS/PD) to/from VPERMI2(B/W/D/Q/PS/PD). Summary: The index and one of the table operands can be swapped by changing the opcode to the other version. Neither of these operands are the one that can load from memory so this can't be used to increase memory folding opportunities. We need to handle the unmasked forms and the kz forms. Since the load operand isn't being commuted we can commute the load and broadcast instructions too. Reviewers: igorb, delena, Ayal, Farhana, RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D25652 llvm-svn: 287621
2016-11-22 12:57:34 +08:00
; X32-NEXT: vpermi2b %xmm2, %xmm2, %xmm0
[X86][AVX512] Add some initial VBMI target shuffle combine tests llvm-svn: 287494
2016-11-20 22:45:46 +08:00
; X32-NEXT: retl
;
; X64-LABEL: combine_vpermt2var_vpermi2var_16i8_as_vperm2:
; X64: # BB#0:
[AVX-512] Remove patterns that select vmovdqu8/16 for unmasked loads. Prefer vmovdqa64/vmovdqu64 instead. These were taking priority over the aligned load instructions since there is no vmovda8/16. I don't think there is really a difference between aligned and unaligned on newer cpus so I don't think it matters which instructions we use. But with this change we reduce the size of the isel table a little and we allow the aligned information to pass through to the evex->vec pass and produce the same output has avx/avx2 in some cases. I also generally dislike patterns rooted in a bitcast which these were. Differential Revision: https://reviews.llvm.org/D35977 llvm-svn: 309589
2017-08-01 01:35:44 +08:00
; X64-NEXT: vmovdqa {{.*#+}} xmm2 = [0,31,2,29,4,27,6,25,8,23,10,21,12,19,14,17]
[X86][AVX512] Add some initial VBMI target shuffle combine tests llvm-svn: 287494
2016-11-20 22:45:46 +08:00
; X64-NEXT: vpermi2b %xmm1, %xmm0, %xmm2
[AVX-512] Remove patterns that select vmovdqu8/16 for unmasked loads. Prefer vmovdqa64/vmovdqu64 instead. These were taking priority over the aligned load instructions since there is no vmovda8/16. I don't think there is really a difference between aligned and unaligned on newer cpus so I don't think it matters which instructions we use. But with this change we reduce the size of the isel table a little and we allow the aligned information to pass through to the evex->vec pass and produce the same output has avx/avx2 in some cases. I also generally dislike patterns rooted in a bitcast which these were. Differential Revision: https://reviews.llvm.org/D35977 llvm-svn: 309589
2017-08-01 01:35:44 +08:00
; X64-NEXT: vmovdqa {{.*#+}} xmm0 = [0,17,2,18,4,19,6,21,8,23,10,25,12,27,14,29]
[AVX-512] Add support for commuting VPERMT2(B/W/D/Q/PS/PD) to/from VPERMI2(B/W/D/Q/PS/PD). Summary: The index and one of the table operands can be swapped by changing the opcode to the other version. Neither of these operands are the one that can load from memory so this can't be used to increase memory folding opportunities. We need to handle the unmasked forms and the kz forms. Since the load operand isn't being commuted we can commute the load and broadcast instructions too. Reviewers: igorb, delena, Ayal, Farhana, RKSimon Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D25652 llvm-svn: 287621
2016-11-22 12:57:34 +08:00
; X64-NEXT: vpermi2b %xmm2, %xmm2, %xmm0
[X86][AVX512] Add some initial VBMI target shuffle combine tests llvm-svn: 287494
2016-11-20 22:45:46 +08:00
; X64-NEXT: retq
%res0 = call <16 x i8> @llvm.x86.avx512.mask.vpermi2var.qi.128(<16 x i8> %x0, <16 x i8> <i8 0, i8 31, i8 2, i8 29, i8 4, i8 27, i8 6, i8 25, i8 8, i8 23, i8 10, i8 21, i8 12, i8 19, i8 14, i8 17>, <16 x i8> %x1, i16 -1)
%res1 = call <16 x i8> @llvm.x86.avx512.maskz.vpermt2var.qi.128(<16 x i8> <i8 0, i8 17, i8 2, i8 18, i8 4, i8 19, i8 6, i8 21, i8 8, i8 23, i8 10, i8 25, i8 12, i8 27, i8 14, i8 29>, <16 x i8> %res0, <16 x i8> %res0, i16 -1)
ret <16 x i8> %res1
}
[X86][AVX512] Add support for VBMI VPERMV3 target shuffle combines llvm-svn: 287496
2016-11-20 23:24:38 +08:00
define <32 x i8> @combine_vpermi2var_32i8_as_vperm2(<32 x i8> %x0, <32 x i8> %x1) {
; X32-LABEL: combine_vpermi2var_32i8_as_vperm2:
; X32: # BB#0:
[AVX-512] Remove patterns that select vmovdqu8/16 for unmasked loads. Prefer vmovdqa64/vmovdqu64 instead. These were taking priority over the aligned load instructions since there is no vmovda8/16. I don't think there is really a difference between aligned and unaligned on newer cpus so I don't think it matters which instructions we use. But with this change we reduce the size of the isel table a little and we allow the aligned information to pass through to the evex->vec pass and produce the same output has avx/avx2 in some cases. I also generally dislike patterns rooted in a bitcast which these were. Differential Revision: https://reviews.llvm.org/D35977 llvm-svn: 309589
2017-08-01 01:35:44 +08:00
; X32-NEXT: vmovdqa {{.*#+}} ymm2 = [0,32,1,23,2,22,3,21,4,22,5,21,6,20,7,19,0,32,1,23,2,22,3,21,4,22,5,21,6,20,7,19]
[X86][AVX512] Add support for VBMI VPERMV3 target shuffle combines llvm-svn: 287496
2016-11-20 23:24:38 +08:00
; X32-NEXT: vpermt2b %ymm1, %ymm2, %ymm0
; X32-NEXT: retl
;
; X64-LABEL: combine_vpermi2var_32i8_as_vperm2:
; X64: # BB#0:
[AVX-512] Remove patterns that select vmovdqu8/16 for unmasked loads. Prefer vmovdqa64/vmovdqu64 instead. These were taking priority over the aligned load instructions since there is no vmovda8/16. I don't think there is really a difference between aligned and unaligned on newer cpus so I don't think it matters which instructions we use. But with this change we reduce the size of the isel table a little and we allow the aligned information to pass through to the evex->vec pass and produce the same output has avx/avx2 in some cases. I also generally dislike patterns rooted in a bitcast which these were. Differential Revision: https://reviews.llvm.org/D35977 llvm-svn: 309589
2017-08-01 01:35:44 +08:00
; X64-NEXT: vmovdqa {{.*#+}} ymm2 = [0,32,1,23,2,22,3,21,4,22,5,21,6,20,7,19,0,32,1,23,2,22,3,21,4,22,5,21,6,20,7,19]
[X86][AVX512] Add support for VBMI VPERMV3 target shuffle combines llvm-svn: 287496
2016-11-20 23:24:38 +08:00
; X64-NEXT: vpermt2b %ymm1, %ymm2, %ymm0
; X64-NEXT: retq
%res0 = shufflevector <32 x i8> %x0, <32 x i8> %x1, <32 x i32> <i32 0, i32 32, i32 1, i32 33, i32 2, i32 34, i32 3, i32 35, i32 4, i32 36, i32 5, i32 37, i32 6, i32 38, i32 7, i32 39, i32 16, i32 48, i32 17, i32 49, i32 18, i32 50, i32 19, i32 51, i32 20, i32 52, i32 21, i32 53, i32 22, i32 54, i32 23, i32 55>
%res1 = call <32 x i8> @llvm.x86.avx512.mask.vpermi2var.qi.256(<32 x i8> %res0, <32 x i8> <i8 0, i8 32, i8 2, i8 30, i8 4, i8 28, i8 6, i8 26, i8 8, i8 28, i8 10, i8 26, i8 12, i8 24, i8 14, i8 22, i8 0, i8 32, i8 2, i8 30, i8 4, i8 28, i8 6, i8 26, i8 8, i8 28, i8 10, i8 26, i8 12, i8 24, i8 14, i8 22>, <32 x i8> %x1, i32 -1)
ret <32 x i8> %res1
}
define <64 x i8> @combine_vpermi2var_64i8_as_vperm2(<64 x i8> %x0, <64 x i8> %x1) {
; X32-LABEL: combine_vpermi2var_64i8_as_vperm2:
; X32: # BB#0:
[AVX-512] Remove patterns that select vmovdqu8/16 for unmasked loads. Prefer vmovdqa64/vmovdqu64 instead. These were taking priority over the aligned load instructions since there is no vmovda8/16. I don't think there is really a difference between aligned and unaligned on newer cpus so I don't think it matters which instructions we use. But with this change we reduce the size of the isel table a little and we allow the aligned information to pass through to the evex->vec pass and produce the same output has avx/avx2 in some cases. I also generally dislike patterns rooted in a bitcast which these were. Differential Revision: https://reviews.llvm.org/D35977 llvm-svn: 309589
2017-08-01 01:35:44 +08:00
; X32-NEXT: vmovdqa64 {{.*#+}} zmm2 = [0,80,1,70,2,54,3,49,4,36,5,23,6,18,7,5,0,90,1,100,2,110,3,120,4,22,5,21,6,20,7,19,0,32,1,23,2,22,3,21,4,22,5,21,6,20,7,19,0,32,1,23,2,22,3,21,4,22,5,21,6,20,7,19]
[X86][AVX512] Add support for VBMI VPERMV3 target shuffle combines llvm-svn: 287496
2016-11-20 23:24:38 +08:00
; X32-NEXT: vpermt2b %zmm1, %zmm2, %zmm0
; X32-NEXT: retl
;
; X64-LABEL: combine_vpermi2var_64i8_as_vperm2:
; X64: # BB#0:
[AVX-512] Remove patterns that select vmovdqu8/16 for unmasked loads. Prefer vmovdqa64/vmovdqu64 instead. These were taking priority over the aligned load instructions since there is no vmovda8/16. I don't think there is really a difference between aligned and unaligned on newer cpus so I don't think it matters which instructions we use. But with this change we reduce the size of the isel table a little and we allow the aligned information to pass through to the evex->vec pass and produce the same output has avx/avx2 in some cases. I also generally dislike patterns rooted in a bitcast which these were. Differential Revision: https://reviews.llvm.org/D35977 llvm-svn: 309589
2017-08-01 01:35:44 +08:00
; X64-NEXT: vmovdqa64 {{.*#+}} zmm2 = [0,80,1,70,2,54,3,49,4,36,5,23,6,18,7,5,0,90,1,100,2,110,3,120,4,22,5,21,6,20,7,19,0,32,1,23,2,22,3,21,4,22,5,21,6,20,7,19,0,32,1,23,2,22,3,21,4,22,5,21,6,20,7,19]
[X86][AVX512] Add support for VBMI VPERMV3 target shuffle combines llvm-svn: 287496
2016-11-20 23:24:38 +08:00
; X64-NEXT: vpermt2b %zmm1, %zmm2, %zmm0
; X64-NEXT: retq
%res0 = shufflevector <64 x i8> %x0, <64 x i8> %x1, <64 x i32> <i32 0, i32 64, i32 1, i32 65, i32 2, i32 66, i32 3, i32 67, i32 4, i32 68, i32 5, i32 69, i32 6, i32 70, i32 7, i32 71, i32 16, i32 80, i32 17, i32 81, i32 18, i32 82, i32 19, i32 83, i32 20, i32 84, i32 21, i32 85, i32 22, i32 86, i32 23, i32 87, i32 32, i32 96, i32 33, i32 97, i32 34, i32 98, i32 35, i32 99, i32 36, i32 100, i32 37, i32 101, i32 38, i32 102, i32 39, i32 103, i32 48, i32 112, i32 49, i32 113, i32 50, i32 114, i32 51, i32 115, i32 52, i32 116, i32 53, i32 117, i32 54, i32 118, i32 55, i32 119>
%res1 = call <64 x i8> @llvm.x86.avx512.mask.vpermi2var.qi.512(<64 x i8> %res0, <64 x i8> <i8 0, i8 80, i8 2, i8 70, i8 4, i8 60, i8 6, i8 50, i8 8, i8 40, i8 10, i8 30, i8 12, i8 20, i8 14, i8 10, i8 0, i8 90, i8 2, i8 100, i8 4, i8 110, i8 6, i8 120, i8 8, i8 28, i8 10, i8 26, i8 12, i8 24, i8 14, i8 22, i8 0, i8 32, i8 2, i8 30, i8 4, i8 28, i8 6, i8 26, i8 8, i8 28, i8 10, i8 26, i8 12, i8 24, i8 14, i8 22, i8 0, i8 32, i8 2, i8 30, i8 4, i8 28, i8 6, i8 26, i8 8, i8 28, i8 10, i8 26, i8 12, i8 24, i8 14, i8 22>, <64 x i8> %x1, i64 -1)
ret <64 x i8> %res1
}