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/avx512dq-intrinsics-upgrade.ll

159 lines
7.8 KiB
LLVM
Raw Normal View History

[AVX-512] Remove vextract intrinsics and autoupgrade to native shufflevectors. This unfortunately generates some really terrible code without VLX support due to v2i1 and v4i1 not being legal. Hopefully we can improve that in future patches. llvm-svn: 290863
2017-01-03 13:45:46 +08:00
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
; RUN: llc < %s -mtriple=x86_64-apple-darwin -mcpu=knl -mattr=+avx512dq | FileCheck %s
declare <2 x double> @llvm.x86.avx512.mask.vextractf64x2.512(<8 x double>, i32, <2 x double>, i8)
define <2 x double>@test_int_x86_avx512_mask_vextractf64x2_512(<8 x double> %x0, <2 x double> %x2, i8 %x3) {
; CHECK-LABEL: test_int_x86_avx512_mask_vextractf64x2_512:
; CHECK: ## BB#0:
; CHECK-NEXT: vextractf64x2 $1, %zmm0, %xmm0
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovw %edi, %k0
[AVX-512] Remove vextract intrinsics and autoupgrade to native shufflevectors. This unfortunately generates some really terrible code without VLX support due to v2i1 and v4i1 not being legal. Hopefully we can improve that in future patches. llvm-svn: 290863
2017-01-03 13:45:46 +08:00
; CHECK-NEXT: kshiftlb $7, %k0, %k1
; CHECK-NEXT: kshiftrb $7, %k1, %k1
; CHECK-NEXT: kshiftlb $6, %k0, %k0
; CHECK-NEXT: kshiftrb $7, %k0, %k0
; CHECK-NEXT: kmovw %k0, %eax
; CHECK-NEXT: vmovq %rax, %xmm2
; CHECK-NEXT: kmovw %k1, %eax
; CHECK-NEXT: vmovq %rax, %xmm3
; CHECK-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm3[0],xmm2[0]
; CHECK-NEXT: vpsllq $63, %xmm2, %xmm2
[X86][AVX512] Add support for ASHR v2i64/v4i64 support without VLX Use v8i64 ASHR instructions if we don't have VLX. Differential Revision: https://reviews.llvm.org/D28537 llvm-svn: 295656
2017-02-20 20:16:38 +08:00
; CHECK-NEXT: vpsraq $63, %zmm2, %zmm2
[AVX-512] Remove vextract intrinsics and autoupgrade to native shufflevectors. This unfortunately generates some really terrible code without VLX support due to v2i1 and v4i1 not being legal. Hopefully we can improve that in future patches. llvm-svn: 290863
2017-01-03 13:45:46 +08:00
; CHECK-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm1
; CHECK-NEXT: vandpd %xmm0, %xmm2, %xmm2
; CHECK-NEXT: vaddpd %xmm0, %xmm1, %xmm0
; CHECK-NEXT: vaddpd %xmm0, %xmm2, %xmm0
; CHECK-NEXT: retq
%res = call <2 x double> @llvm.x86.avx512.mask.vextractf64x2.512(<8 x double> %x0,i32 1, <2 x double> %x2, i8 %x3)
%res2 = call <2 x double> @llvm.x86.avx512.mask.vextractf64x2.512(<8 x double> %x0,i32 1, <2 x double> zeroinitializer, i8 %x3)
%res1 = call <2 x double> @llvm.x86.avx512.mask.vextractf64x2.512(<8 x double> %x0,i32 1, <2 x double> zeroinitializer, i8 -1)
%res3 = fadd <2 x double> %res, %res1
%res4 = fadd <2 x double> %res2, %res3
ret <2 x double> %res4
}
declare <8 x float> @llvm.x86.avx512.mask.vextractf32x8.512(<16 x float>, i32, <8 x float>, i8)
define <8 x float>@test_int_x86_avx512_mask_vextractf32x8(<16 x float> %x0, <8 x float> %x2, i8 %x3) {
; CHECK-LABEL: test_int_x86_avx512_mask_vextractf32x8:
; CHECK: ## BB#0:
; CHECK-NEXT: vextractf32x8 $1, %zmm0, %ymm2
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovw %edi, %k1
[AVX-512] Remove vextract intrinsics and autoupgrade to native shufflevectors. This unfortunately generates some really terrible code without VLX support due to v2i1 and v4i1 not being legal. Hopefully we can improve that in future patches. llvm-svn: 290863
2017-01-03 13:45:46 +08:00
; CHECK-NEXT: vextractf32x8 $1, %zmm0, %ymm1 {%k1}
; CHECK-NEXT: vaddps %ymm2, %ymm1, %ymm1
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: 302938
2017-05-13 03:29:27 +08:00
; CHECK-NEXT: vextractf32x8 $1, %zmm0, %ymm0 {%k1} {z}
[AVX-512] Remove vextract intrinsics and autoupgrade to native shufflevectors. This unfortunately generates some really terrible code without VLX support due to v2i1 and v4i1 not being legal. Hopefully we can improve that in future patches. llvm-svn: 290863
2017-01-03 13:45:46 +08:00
; CHECK-NEXT: vaddps %ymm1, %ymm0, %ymm0
; CHECK-NEXT: retq
%res = call <8 x float> @llvm.x86.avx512.mask.vextractf32x8.512(<16 x float> %x0,i32 1, <8 x float> %x2, i8 %x3)
%res2 = call <8 x float> @llvm.x86.avx512.mask.vextractf32x8.512(<16 x float> %x0,i32 1, <8 x float> zeroinitializer, i8 %x3)
%res1 = call <8 x float> @llvm.x86.avx512.mask.vextractf32x8.512(<16 x float> %x0,i32 1, <8 x float> zeroinitializer, i8 -1)
%res3 = fadd <8 x float> %res, %res1
%res4 = fadd <8 x float> %res2, %res3
ret <8 x float> %res4
}
[AVX-512] Remove vinsert intrinsics and autoupgrade to native shufflevectors. There are some codegen problems here that I'll try to fix in future commits. llvm-svn: 290864
2017-01-03 13:45:57 +08:00
declare <16 x float> @llvm.x86.avx512.mask.insertf32x8.512(<16 x float>, <8 x float>, i32, <16 x float>, i16)
define <16 x float>@test_int_x86_avx512_mask_insertf32x8_512(<16 x float> %x0, <8 x float> %x1, <16 x float> %x3, i16 %x4) {
; CHECK-LABEL: test_int_x86_avx512_mask_insertf32x8_512:
; CHECK: ## BB#0:
[AVX-512] Teach shuffle lowering to use vinsert instructions for shuffles corresponding to 256-bit subvector inserts. llvm-svn: 290870
2017-01-03 15:00:40 +08:00
; CHECK-NEXT: vinsertf32x8 $1, %ymm1, %zmm0, %zmm3
[AVX-512] Remove vinsert intrinsics and autoupgrade to native shufflevectors. There are some codegen problems here that I'll try to fix in future commits. llvm-svn: 290864
2017-01-03 13:45:57 +08:00
; CHECK-NEXT: kmovw %edi, %k1
[AVX-512] Teach shuffle lowering to use vinsert instructions for shuffles corresponding to 256-bit subvector inserts. llvm-svn: 290870
2017-01-03 15:00:40 +08:00
; CHECK-NEXT: vinsertf32x8 $1, %ymm1, %zmm0, %zmm2 {%k1}
; CHECK-NEXT: vinsertf32x8 $1, %ymm1, %zmm0, %zmm0 {%k1} {z}
[AVX-512] Remove vinsert intrinsics and autoupgrade to native shufflevectors. There are some codegen problems here that I'll try to fix in future commits. llvm-svn: 290864
2017-01-03 13:45:57 +08:00
; CHECK-NEXT: vaddps %zmm0, %zmm2, %zmm0
; CHECK-NEXT: vaddps %zmm0, %zmm3, %zmm0
; CHECK-NEXT: retq
%res = call <16 x float> @llvm.x86.avx512.mask.insertf32x8.512(<16 x float> %x0, <8 x float> %x1, i32 1, <16 x float> %x3, i16 %x4)
%res1 = call <16 x float> @llvm.x86.avx512.mask.insertf32x8.512(<16 x float> %x0, <8 x float> %x1, i32 1, <16 x float> zeroinitializer, i16 %x4)
%res2 = call <16 x float> @llvm.x86.avx512.mask.insertf32x8.512(<16 x float> %x0, <8 x float> %x1, i32 1, <16 x float> %x3, i16 -1)
%res3 = fadd <16 x float> %res, %res1
%res4 = fadd <16 x float> %res2, %res3
ret <16 x float> %res4
}
declare <8 x double> @llvm.x86.avx512.mask.insertf64x2.512(<8 x double>, <2 x double>, i32, <8 x double>, i8)
define <8 x double>@test_int_x86_avx512_mask_insertf64x2_512(<8 x double> %x0, <2 x double> %x1,<8 x double> %x3, i8 %x4) {
; CHECK-LABEL: test_int_x86_avx512_mask_insertf64x2_512:
; CHECK: ## BB#0:
[AVX-512] Add support for detecting 512-bit shuffles that contain a 128-bit subvector insertion from the lowest subvector of one of the sources. These are best handled with a vinsert32x4 or vinsert64x2 instruction. llvm-svn: 290946
2017-01-04 15:32:03 +08:00
; CHECK-NEXT: vinsertf64x2 $1, %xmm1, %zmm0, %zmm3
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovw %edi, %k1
[AVX-512] Add support for detecting 512-bit shuffles that contain a 128-bit subvector insertion from the lowest subvector of one of the sources. These are best handled with a vinsert32x4 or vinsert64x2 instruction. llvm-svn: 290946
2017-01-04 15:32:03 +08:00
; CHECK-NEXT: vinsertf64x2 $1, %xmm1, %zmm0, %zmm2 {%k1}
; CHECK-NEXT: vinsertf64x2 $1, %xmm1, %zmm0, %zmm0 {%k1} {z}
[AVX-512] Remove vinsert intrinsics and autoupgrade to native shufflevectors. There are some codegen problems here that I'll try to fix in future commits. llvm-svn: 290864
2017-01-03 13:45:57 +08:00
; CHECK-NEXT: vaddpd %zmm0, %zmm2, %zmm0
[AVX-512] Add support for detecting 512-bit shuffles that contain a 128-bit subvector insertion from the lowest subvector of one of the sources. These are best handled with a vinsert32x4 or vinsert64x2 instruction. llvm-svn: 290946
2017-01-04 15:32:03 +08:00
; CHECK-NEXT: vaddpd %zmm3, %zmm0, %zmm0
[AVX-512] Remove vinsert intrinsics and autoupgrade to native shufflevectors. There are some codegen problems here that I'll try to fix in future commits. llvm-svn: 290864
2017-01-03 13:45:57 +08:00
; CHECK-NEXT: retq
%res = call <8 x double> @llvm.x86.avx512.mask.insertf64x2.512(<8 x double> %x0, <2 x double> %x1, i32 1, <8 x double> %x3, i8 %x4)
%res1 = call <8 x double> @llvm.x86.avx512.mask.insertf64x2.512(<8 x double> %x0, <2 x double> %x1, i32 1, <8 x double> zeroinitializer, i8 %x4)
%res2 = call <8 x double> @llvm.x86.avx512.mask.insertf64x2.512(<8 x double> %x0, <2 x double> %x1, i32 1, <8 x double> %x3, i8 -1)
%res3 = fadd <8 x double> %res, %res1
%res4 = fadd <8 x double> %res3, %res2
ret <8 x double> %res4
}
declare <16 x i32> @llvm.x86.avx512.mask.inserti32x8.512(<16 x i32>, <8 x i32>, i32, <16 x i32>, i16)
define <16 x i32>@test_int_x86_avx512_mask_inserti32x8_512(<16 x i32> %x0, <8 x i32> %x1, <16 x i32> %x3, i16 %x4) {
; CHECK-LABEL: test_int_x86_avx512_mask_inserti32x8_512:
; CHECK: ## BB#0:
[AVX-512] Teach shuffle lowering to use vinsert instructions for shuffles corresponding to 256-bit subvector inserts. llvm-svn: 290870
2017-01-03 15:00:40 +08:00
; CHECK-NEXT: vinserti32x8 $1, %ymm1, %zmm0, %zmm3
[AVX-512] Remove vinsert intrinsics and autoupgrade to native shufflevectors. There are some codegen problems here that I'll try to fix in future commits. llvm-svn: 290864
2017-01-03 13:45:57 +08:00
; CHECK-NEXT: kmovw %edi, %k1
[AVX-512] Teach shuffle lowering to use vinsert instructions for shuffles corresponding to 256-bit subvector inserts. llvm-svn: 290870
2017-01-03 15:00:40 +08:00
; CHECK-NEXT: vinserti32x8 $1, %ymm1, %zmm0, %zmm2 {%k1}
; CHECK-NEXT: vinserti32x8 $1, %ymm1, %zmm0, %zmm0 {%k1} {z}
[AVX-512] Remove vinsert intrinsics and autoupgrade to native shufflevectors. There are some codegen problems here that I'll try to fix in future commits. llvm-svn: 290864
2017-01-03 13:45:57 +08:00
; CHECK-NEXT: vpaddd %zmm0, %zmm2, %zmm0
; CHECK-NEXT: vpaddd %zmm3, %zmm0, %zmm0
; CHECK-NEXT: retq
%res = call <16 x i32> @llvm.x86.avx512.mask.inserti32x8.512(<16 x i32> %x0, <8 x i32> %x1, i32 1, <16 x i32> %x3, i16 %x4)
%res1 = call <16 x i32> @llvm.x86.avx512.mask.inserti32x8.512(<16 x i32> %x0, <8 x i32> %x1, i32 1, <16 x i32> zeroinitializer, i16 %x4)
%res2 = call <16 x i32> @llvm.x86.avx512.mask.inserti32x8.512(<16 x i32> %x0, <8 x i32> %x1, i32 1, <16 x i32> %x3, i16 -1)
%res3 = add <16 x i32> %res, %res1
%res4 = add <16 x i32> %res3, %res2
ret <16 x i32> %res4
}
declare <8 x i64> @llvm.x86.avx512.mask.inserti64x2.512(<8 x i64>, <2 x i64>, i32, <8 x i64>, i8)
define <8 x i64>@test_int_x86_avx512_mask_inserti64x2_512(<8 x i64> %x0, <2 x i64> %x1, <8 x i64> %x3, i8 %x4) {
; CHECK-LABEL: test_int_x86_avx512_mask_inserti64x2_512:
; CHECK: ## BB#0:
[AVX-512] Add support for detecting 512-bit shuffles that contain a 128-bit subvector insertion from the lowest subvector of one of the sources. These are best handled with a vinsert32x4 or vinsert64x2 instruction. llvm-svn: 290946
2017-01-04 15:32:03 +08:00
; CHECK-NEXT: vinserti64x2 $1, %xmm1, %zmm0, %zmm3
[AVX-512] Fix accidental uses of AH/BH/CH/DH after copies to/from mask registers We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register. This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8. I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa. Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition. This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI. Differential Revision: https://reviews.llvm.org/D30968 llvm-svn: 298928
2017-03-29 00:35:29 +08:00
; CHECK-NEXT: kmovw %edi, %k1
[AVX-512] Add support for detecting 512-bit shuffles that contain a 128-bit subvector insertion from the lowest subvector of one of the sources. These are best handled with a vinsert32x4 or vinsert64x2 instruction. llvm-svn: 290946
2017-01-04 15:32:03 +08:00
; CHECK-NEXT: vinserti64x2 $1, %xmm1, %zmm0, %zmm2 {%k1}
; CHECK-NEXT: vinserti64x2 $1, %xmm1, %zmm0, %zmm0 {%k1} {z}
[AVX-512] Remove vinsert intrinsics and autoupgrade to native shufflevectors. There are some codegen problems here that I'll try to fix in future commits. llvm-svn: 290864
2017-01-03 13:45:57 +08:00
; CHECK-NEXT: vpaddq %zmm0, %zmm2, %zmm0
[AVX-512] Add support for detecting 512-bit shuffles that contain a 128-bit subvector insertion from the lowest subvector of one of the sources. These are best handled with a vinsert32x4 or vinsert64x2 instruction. llvm-svn: 290946
2017-01-04 15:32:03 +08:00
; CHECK-NEXT: vpaddq %zmm0, %zmm3, %zmm0
[AVX-512] Remove vinsert intrinsics and autoupgrade to native shufflevectors. There are some codegen problems here that I'll try to fix in future commits. llvm-svn: 290864
2017-01-03 13:45:57 +08:00
; CHECK-NEXT: retq
%res = call <8 x i64> @llvm.x86.avx512.mask.inserti64x2.512(<8 x i64> %x0, <2 x i64> %x1, i32 1, <8 x i64> %x3, i8 %x4)
%res1 = call <8 x i64> @llvm.x86.avx512.mask.inserti64x2.512(<8 x i64> %x0, <2 x i64> %x1, i32 1, <8 x i64> zeroinitializer, i8 %x4)
%res2 = call <8 x i64> @llvm.x86.avx512.mask.inserti64x2.512(<8 x i64> %x0, <2 x i64> %x1, i32 1, <8 x i64> %x3, i8 -1)
%res3 = add <8 x i64> %res, %res1
%res4 = add <8 x i64> %res2, %res3
ret <8 x i64> %res4
}
[X86][LLVM] Converting __mm{|256|512}_movm_epi{8|16|32|64} LLVMIR call into generic intrinsics. This patch is a part one of two reviews, one for the clang and the other for LLVM. The patch deletes the back-end intrinsics and adds support for them in the auto upgrade. Differential Revision: https://reviews.llvm.org/D31393 llvm-svn: 299432
2017-04-04 21:32:14 +08:00
declare <16 x i32> @llvm.x86.avx512.cvtmask2d.512(i16)
define <16 x i32>@test_int_x86_avx512_cvtmask2d_512(i16 %x0) {
; CHECK-LABEL: test_int_x86_avx512_cvtmask2d_512:
; CHECK: ## BB#0:
; CHECK-NEXT: kmovw %edi, %k0
; CHECK-NEXT: vpmovm2d %k0, %zmm0
; CHECK-NEXT: retq
%res = call <16 x i32> @llvm.x86.avx512.cvtmask2d.512(i16 %x0)
ret <16 x i32> %res
}
declare <8 x i64> @llvm.x86.avx512.cvtmask2q.512(i8)
define <8 x i64>@test_int_x86_avx512_cvtmask2q_512(i8 %x0) {
; CHECK-LABEL: test_int_x86_avx512_cvtmask2q_512:
; CHECK: ## BB#0:
; CHECK-NEXT: kmovw %edi, %k0
; CHECK-NEXT: vpmovm2q %k0, %zmm0
; CHECK-NEXT: retq
%res = call <8 x i64> @llvm.x86.avx512.cvtmask2q.512(i8 %x0)
ret <8 x i64> %res
}