This is an archive of the discontinued LLVM Phabricator instance.

Differential D144442

[X86] Add tests for replacing `{v}unpck{l|h}pd` -> `{v}shufps`; NFC
ClosedPublic

Authored by goldstein.w.n on Feb 20 2023, 6:09 PM.

Details

Reviewers
pengfei
RKSimon
Commits
rG6957a8cc6c92: Add tests for replacing `{v}unpck{l|h}pd` -> `{v}shufps`; NFC

Diff Detail

Unit TestsFailed

TimeTest
3,890 msx64 debian > LLVM.CodeGen/X86::combine-and.ll
3,640 msx64 debian > LLVM.CodeGen/X86::copy-low-subvec-elt-to-high-subvec-elt.ll
4,560 msx64 debian > LLVM.CodeGen/X86::haddsub-shuf.ll
4,460 msx64 debian > LLVM.CodeGen/X86::haddsub-undef.ll
3,480 msx64 debian > LLVM.CodeGen/X86::matrix-multiply.ll
View Full Test Results (59 Failed)

Event Timeline

goldstein.w.n created this revision.Feb 20 2023, 6:09 PM
Herald added a project: Restricted Project. · View Herald TranscriptFeb 20 2023, 6:09 PM
Herald added a subscriber: pengfei. · View Herald Transcript
goldstein.w.n requested review of this revision.Feb 20 2023, 6:09 PM
Herald added a project: Restricted Project. · View Herald TranscriptFeb 20 2023, 6:09 PM
Herald added a subscriber: llvm-commits. · View Herald Transcript
goldstein.w.n added reviewers: pengfei, RKSimon.Feb 20 2023, 6:11 PM
goldstein.w.n added a parent revision: D143859: [X86] Adding tuning flags for int <-> fp domain switching penalties; NFC.
goldstein.w.n added a child revision: D143787: [X86] Add new pass `X86FixupInstTuning` for fixing up machine-instruction selection..
goldstein.w.n updated this revision to Diff 499024.Feb 20 2023, 7:52 PM

Rebase

Harbormaster completed remote builds in B214913: Diff 499024.Feb 20 2023, 10:20 PM
RKSimon added a comment.Feb 21 2023, 8:16 PM

We might want to consider a naming convention to accommodate future cpu-tuning tests - how about tuning-shuffle-unpckpd.ll?

llvm/test/CodeGen/X86/unpckpd-to-faster-alternative.ll
43 ↗(On Diff #499024)

Cleanup?

goldstein.w.n added a comment.Feb 21 2023, 11:33 PM
In D144442#4143398, @RKSimon wrote:

We might want to consider a naming convention to accommodate future cpu-tuning tests - how about tuning-shuffle-unpckpd.ll?

Agreed, changed in v2.

llvm/test/CodeGen/X86/unpckpd-to-faster-alternative.ll
43 ↗(On Diff #499024)

What do you mean? Thats the output from

/home/noah/programs/opensource/llvm-dev/src/llvm-project/llvm/utils/update_llc_test_checks.py --llc-binary /home/noah/programs/opensource/llvm-dev/src/llvm-project/build/bin/llc /home/noah/programs/opensource/llvm-dev/src/llvm-project/llvm/test/CodeGen/X86/unpckpd-to-faster-alternative.ll

What command should I use?

goldstein.w.n updated this revision to Diff 499383.Feb 21 2023, 11:39 PM

Rename test

Harbormaster completed remote builds in B215172: Diff 499383.Feb 22 2023, 12:31 AM
RKSimon added inline comments.Feb 22 2023, 4:00 AM
llvm/test/CodeGen/X86/tuning-shuffle-unpckpd.ll
46

These will need to be manually editted out and then the update script called to regenerate

goldstein.w.n updated this revision to Diff 499543.Feb 22 2023, 9:12 AM

Rebase

Harbormaster completed remote builds in B215283: Diff 499543.Feb 22 2023, 11:35 AM
RKSimon requested changes to this revision.Feb 22 2023, 2:53 PM
RKSimon added inline comments.
llvm/test/CodeGen/X86/tuning-shuffle-unpckpd.ll
46

Still broken :(

This revision now requires changes to proceed.Feb 22 2023, 2:53 PM
goldstein.w.n marked an inline comment as done.Feb 22 2023, 5:42 PM
goldstein.w.n added inline comments.
llvm/test/CodeGen/X86/tuning-shuffle-unpckpd.ll
46

Bah sorry, didn't see your comment this morning. Fixed.

goldstein.w.n updated this revision to Diff 499694.Feb 22 2023, 5:42 PM
goldstein.w.n marked an inline comment as done.

Cleanup checks

Harbormaster completed remote builds in B215402: Diff 499694.Feb 22 2023, 7:47 PM
RKSimon accepted this revision.Feb 23 2023, 5:52 PM

LGTM - cheers

This revision is now accepted and ready to land.Feb 23 2023, 5:52 PM
goldstein.w.n updated this revision to Diff 500061.Feb 23 2023, 8:09 PM

Rebase

Harbormaster completed remote builds in B215662: Diff 500061.Feb 23 2023, 10:21 PM
goldstein.w.n updated this revision to Diff 500305.Feb 24 2023, 3:09 PM

Rebase

goldstein.w.n updated this revision to Diff 500373.Feb 24 2023, 11:28 PM

Make tests complete

goldstein.w.n added a child revision: D144779: [X86] Add tests for replacing `{v}permilps` -> `{v}shufps/{v}pshufd`; NFC.Feb 24 2023, 11:37 PM
goldstein.w.n removed a child revision: D143787: [X86] Add new pass `X86FixupInstTuning` for fixing up machine-instruction selection..
Harbormaster completed remote builds in B215892: Diff 500373.Feb 25 2023, 4:32 AM
goldstein.w.n updated this revision to Diff 500474.Feb 25 2023, 4:59 PM

Rebase

Harbormaster completed remote builds in B215976: Diff 500474.Feb 25 2023, 6:54 PM
RKSimon requested changes to this revision.Feb 26 2023, 4:10 AM

Similar to D144779 - we should probably split the avx512 cases off and add evex-vex testing - whatever we decide we should try to keep to the same style for all these tuning-shuffle-* files.

This revision now requires changes to proceed.Feb 26 2023, 4:10 AM
RKSimon added a comment.Feb 26 2023, 4:14 AM

Also, this should block D144570 and not D144779/D144570 - ideally we're going to want D144570 to be in trunk for a few days before any follow ups to shake out any unexpected regressions to prevent a nasty long chain of reversions, which is why I'm so keen to get accepted patches in early instead of you stacking them all up.

goldstein.w.n added a comment.Feb 26 2023, 10:47 AM
In D144442#4153406, @RKSimon wrote:

Also, this should block D144570 and not D144779/D144570 - ideally we're going to want D144570 to be in trunk for a few days before any follow ups to shake out any unexpected regressions to prevent a nasty long chain of reversions, which is why I'm so keen to get accepted patches in early instead of you stacking them all up.

Hmm? You repeat D144570 for block/non-block.

D143786 -> D143859 -> D144779 -> D143787 -> D144442 -> D144570 seems like the order to me. We can wait after D143787 or D144570. Whichever you prefer.

goldstein.w.n removed a parent revision: D143859: [X86] Adding tuning flags for int <-> fp domain switching penalties; NFC.Feb 26 2023, 10:49 AM
goldstein.w.n edited child revisions, added: D144570: [X86] Add support for using Sched/Codesize information to `X86FixupInstTuning` Pass.; removed: D144779: [X86] Add tests for replacing `{v}permilps` -> `{v}shufps/{v}pshufd`; NFC.
goldstein.w.n added a parent revision: D143787: [X86] Add new pass `X86FixupInstTuning` for fixing up machine-instruction selection..Feb 26 2023, 10:52 AM
goldstein.w.n updated this revision to Diff 500615.Feb 26 2023, 12:33 PM

Split avx/avx512 tests

Harbormaster completed remote builds in B216099: Diff 500615.Feb 26 2023, 1:52 PM
RKSimon added inline comments.Feb 26 2023, 1:55 PM
llvm/test/CodeGen/X86/tuning-shuffle-unpckpd.ll
24

Drop the 512-bit vector test cases - they should be handled by the -avx512.ll file

goldstein.w.n updated this revision to Diff 500627.Feb 26 2023, 2:52 PM

Removed avx512 tests for avx file

goldstein.w.n marked an inline comment as done.Feb 26 2023, 2:54 PM
Harbormaster completed remote builds in B216109: Diff 500627.Feb 26 2023, 3:42 PM
RKSimon accepted this revision.Feb 27 2023, 7:51 AM

LGTM

This revision is now accepted and ready to land.Feb 27 2023, 7:51 AM
goldstein.w.n mentioned this in D143786: [X86] Add `TuningPreferShiftShuffle` for when Shifts are preferable to shuffles..Feb 27 2023, 10:34 AM
goldstein.w.n updated this revision to Diff 500963.Feb 27 2023, 4:19 PM

Rebase (after D144832, no dep on D143786)

Harbormaster completed remote builds in B216353: Diff 500963.Feb 27 2023, 4:20 PM
This revision was landed with ongoing or failed builds.Feb 27 2023, 4:54 PM
Closed by commit rG6957a8cc6c92: Add tests for replacing `{v}unpck{l|h}pd` -> `{v}shufps`; NFC (authored by goldstein.w.n). · Explain Why
This revision was automatically updated to reflect the committed changes.
goldstein.w.n added a commit: rG6957a8cc6c92: Add tests for replacing `{v}unpck{l|h}pd` -> `{v}shufps`; NFC.
goldstein.w.n removed a child revision: D144570: [X86] Add support for using Sched/Codesize information to `X86FixupInstTuning` Pass..Mar 7 2023, 4:26 PM

Revision Contents

PathSize
llvm/
test/
CodeGen/
X86/
93 lines

Diff 499383

llvm/test/CodeGen/X86/tuning-shuffle-unpckpd.ll

  • This file was added.
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mcpu=skylake-avx512 | FileCheck %s --check-prefixes=CHECK,CHECK-512,CHECK-SKX
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mcpu=skylake | FileCheck %s --check-prefixes=CHECK,CHECK-SKL
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mcpu=icelake-server | FileCheck %s --check-prefixes=CHECK,CHECK-512,CHECK-ICX
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mcpu=x86-64-v4 | FileCheck %s --check-prefixes=CHECK,CHECK-512,CHECK-V4
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mcpu=znver4 | FileCheck %s --check-prefixes=CHECK,CHECK-512,CHECK-ZNVER4
define <16 x float> @transform_VUNPCKLPDZrr(<16 x float> %a, <16 x float> %b) nounwind {
; CHECK-512-LABEL: transform_VUNPCKLPDZrr:
; CHECK-512: # %bb.0:
; CHECK-512-NEXT: vunpcklpd {{.*#+}} zmm0 = zmm0[0],zmm1[0],zmm0[2],zmm1[2],zmm0[4],zmm1[4],zmm0[6],zmm1[6]
; CHECK-512-NEXT: retq
;
; CHECK-SKL-LABEL: transform_VUNPCKLPDZrr:
; CHECK-SKL: # %bb.0:
; CHECK-SKL-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm2[0],ymm0[2],ymm2[2]
; CHECK-SKL-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm3[0],ymm1[2],ymm3[2]
; CHECK-SKL-NEXT: retq
%shufp = shufflevector <16 x float> %a, <16 x float> %b, <16 x i32> <i32 0, i32 1, i32 16, i32 17, i32 4, i32 5, i32 20, i32 21, i32 8, i32 9, i32 24, i32 25, i32 12, i32 13, i32 28, i32 29>
ret <16 x float> %shufp
}
define <16 x float> @transform_VUNPCKHPDZrr(<16 x float> %a, <16 x float> %b) nounwind {
; CHECK-512-LABEL: transform_VUNPCKHPDZrr:
RKSimonUnsubmitted
Done
ReplyInline Actions

Drop the 512-bit vector test cases - they should be handled by the -avx512.ll file

RKSimon: Drop the 512-bit vector test cases - they should be handled by the -avx512.ll file
; CHECK-512: # %bb.0:
; CHECK-512-NEXT: vunpckhpd {{.*#+}} zmm0 = zmm0[1],zmm1[1],zmm0[3],zmm1[3],zmm0[5],zmm1[5],zmm0[7],zmm1[7]
; CHECK-512-NEXT: retq
;
; CHECK-SKL-LABEL: transform_VUNPCKHPDZrr:
; CHECK-SKL: # %bb.0:
; CHECK-SKL-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm2[1],ymm0[3],ymm2[3]
; CHECK-SKL-NEXT: vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm3[1],ymm1[3],ymm3[3]
; CHECK-SKL-NEXT: retq
%shufp = shufflevector <16 x float> %a, <16 x float> %b, <16 x i32> <i32 2, i32 3, i32 18, i32 19, i32 6, i32 7, i32 22, i32 23, i32 10, i32 11, i32 26, i32 27, i32 14, i32 15, i32 30, i32 31>
ret <16 x float> %shufp
}
define <8 x float> @transform_VUNPCKLPDYrr(<8 x float> %a, <8 x float> %b) nounwind {
; CHECK-LABEL: transform_VUNPCKLPDYrr:
; CHECK: # %bb.0:
; CHECK-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; CHECK-NEXT: retq
; -LABEL: transform_VUNPCKLPDYrr:
; : # %bb.0:
; -NEXT: vshufps {{.*#+}} ymm0 = ymm0[0,1],ymm1[0,1],ymm0[4,5],ymm1[4,5]
; -NEXT: retq
RKSimonUnsubmitted
Not Done
ReplyInline Actions

These will need to be manually editted out and then the update script called to regenerate

RKSimon: These will need to be manually editted out and then the update script called to regenerate
RKSimonUnsubmitted
Done
ReplyInline Actions

Still broken :(

RKSimon: Still broken :(
goldstein.w.nAuthorUnsubmitted
Done
ReplyInline Actions

Bah sorry, didn't see your comment this morning. Fixed.

goldstein.w.n: Bah sorry, didn't see your comment this morning. Fixed.
%shufp = shufflevector <8 x float> %a, <8 x float> %b, <8 x i32> <i32 0, i32 1, i32 8, i32 9, i32 4, i32 5, i32 12, i32 13>
ret <8 x float> %shufp
}
define <8 x float> @transform_VUNPCKHPDYrr(<8 x float> %a, <8 x float> %b) nounwind {
; CHECK-LABEL: transform_VUNPCKHPDYrr:
; CHECK: # %bb.0:
; CHECK-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
; CHECK-NEXT: retq
; -LABEL: transform_VUNPCKHPDYrr:
; : # %bb.0:
; -NEXT: vshufps {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3],ymm0[6,7],ymm1[6,7]
; -NEXT: retq
%shufp = shufflevector <8 x float> %a, <8 x float> %b, <8 x i32> <i32 2, i32 3, i32 10, i32 11, i32 6, i32 7, i32 14, i32 15>
ret <8 x float> %shufp
}
define <4 x float> @transform_VUNPCKLPDrr(<4 x float> %a, <4 x float> %b) nounwind {
; CHECK-LABEL: transform_VUNPCKLPDrr:
; CHECK: # %bb.0:
; CHECK-NEXT: vmovlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0]
; CHECK-NEXT: retq
; -LABEL: transform_VUNPCKLPDrr:
; : # %bb.0:
; -NEXT: vshufps {{.*#+}} xmm0 = xmm0[0,1],xmm1[0,1]
; -NEXT: retq
%shufp = shufflevector <4 x float> %a, <4 x float> %b, <4 x i32> <i32 0, i32 1, i32 4, i32 5>
ret <4 x float> %shufp
}
define <4 x float> @transform_VUNPCKHPDrr(<4 x float> %a, <4 x float> %b) nounwind {
; CHECK-LABEL: transform_VUNPCKHPDrr:
; CHECK: # %bb.0:
; CHECK-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
; CHECK-NEXT: retq
; -LABEL: transform_VUNPCKHPDrr:
; : # %bb.0:
; -NEXT: vshufps {{.*#+}} xmm0 = xmm0[2,3],xmm1[2,3]
; -NEXT: retq
%shufp = shufflevector <4 x float> %a, <4 x float> %b, <4 x i32> <i32 2, i32 3, i32 6, i32 7>
ret <4 x float> %shufp
}
;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
; CHECK-ICX: {{.*}}
; CHECK-SKX: {{.*}}
; CHECK-V4: {{.*}}
; CHECK-ZNVER4: {{.*}}