Skip to content

[AArch64] - Improve costing for Identity shuffles for SVE targets. #165375

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
wants to merge 4 commits into
base: main
Choose a base branch
from
+90 −8
Open

[AArch64] - Improve costing for Identity shuffles for SVE targets. #165375

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
70bf624
[AArch64] - Improve costing for Identity shuffles for SVE targets.
Oct 28, 2025
eca25d4
Merge branch 'main' into identity-shuffle-costs
pawan-nirpal-031 Oct 28, 2025
d2f6cc3
Update identity-shuffle-sve.ll
pawan-nirpal-031 Nov 3, 2025
cae9a19
fix tests
Nov 6, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
17 changes: 9 additions & 8 deletions llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -5895,6 +5895,15 @@ AArch64TTIImpl::getShuffleCost(TTI::ShuffleKind Kind, VectorType *DstTy,
SrcTy = DstTy;
}

// Check for identity masks, which we can treat as free for both fixed and
// scalable vector paths.
if (!Mask.empty() && LT.second.isFixedLengthVector() &&
(Kind == TTI::SK_PermuteTwoSrc || Kind == TTI::SK_PermuteSingleSrc) &&
all_of(enumerate(Mask), [](const auto &M) {
return M.value() < 0 || M.value() == (int)M.index();
}))
return 0;

// Segmented shuffle matching.
if (Kind == TTI::SK_PermuteSingleSrc && isa<FixedVectorType>(SrcTy) &&
!Mask.empty() && SrcTy->getPrimitiveSizeInBits().isNonZero() &&
Expand Down Expand Up @@ -5942,14 +5951,6 @@ AArch64TTIImpl::getShuffleCost(TTI::ShuffleKind Kind, VectorType *DstTy,
all_of(Mask, [](int E) { return E < 8; }))
return getPerfectShuffleCost(Mask);

// Check for identity masks, which we can treat as free.
if (!Mask.empty() && LT.second.isFixedLengthVector() &&
(Kind == TTI::SK_PermuteTwoSrc || Kind == TTI::SK_PermuteSingleSrc) &&
all_of(enumerate(Mask), [](const auto &M) {
return M.value() < 0 || M.value() == (int)M.index();
}))
return 0;

// Check for other shuffles that are not SK_ kinds but we have native
// instructions for, for example ZIP and UZP.
unsigned Unused;
Expand Down
12 changes: 12 additions & 0 deletions llvm/test/Analysis/CostModel/AArch64/shuffle-other.ll
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -490,3 +490,15 @@ define void @vst4(ptr %p) {

ret void
}

define <16 x i8> @identity_shuffle_costs() #0 {
bb:
; CHECK-LABEL: 'identity_shuffle_costs'
; CHECK: Cost Model: Found costs of 0 for: %shufflevector142 = shufflevector <16 x i8> %trunc125, <16 x i8> %trunc133, <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>
%trunc125 = trunc <16 x i32> zeroinitializer to <16 x i8>
%trunc133 = trunc <16 x i32> zeroinitializer to <16 x i8>
%shufflevector142 = shufflevector <16 x i8> %trunc125, <16 x i8> %trunc133, <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>
ret <16 x i8> %shufflevector142
}

attributes #0 = { "target-features"="+sve,+neon" }
69 changes: 69 additions & 0 deletions llvm/test/Transforms/VectorCombine/AArch64/identity-shuffle-sve.ll
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,69 @@
; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
; RUN: opt -passes=vector-combine -mtriple=aarch64-unknown-linux-gnu -S %s | FileCheck %s
target triple = "aarch64-unknown-linux-gnu"

define i32 @ham(ptr %call12) local_unnamed_addr #0 {
; CHECK-LABEL: define i32 @ham(
; CHECK-SAME: ptr [[CALL12:%.*]]) local_unnamed_addr #[[ATTR0:[0-9]+]] {
; CHECK: [[TMP2:%.*]] = mul nuw nsw <32 x i32> [[TMP1:%.*]], <i32 19595, i32 19595, i32 19595, i32 19595, i32 19595, i32 19595, i32 19595, i32 19595, i32 19595, i32 19595, i32 19595, i32 19595, i32 19595, i32 19595, i32 19595, i32 19595, i32 32767, i32 32767, i32 32767, i32 32767, i32 32767, i32 32767, i32 32767, i32 32767, i32 32767, i32 32767, i32 32767, i32 32767, i32 32767, i32 32767, i32 32767, i32 32767>

; CHECK: [[TMP5:%.*]] = mul nuw <32 x i32> [[TMP4:%.*]], <i32 38470, i32 38470, i32 38470, i32 38470, i32 38470, i32 38470, i32 38470, i32 38470, i32 38470, i32 38470, i32 38470, i32 38470, i32 38470, i32 38470, i32 38470, i32 38470, i32 16762097, i32 16762097, i32 16762097, i32 16762097, i32 16762097, i32 16762097, i32 16762097, i32 16762097, i32 16762097, i32 16762097, i32 16762097, i32 16762097, i32 16762097, i32 16762097, i32 16762097, i32 16762097>

; CHECK: [[TMP8:%.*]] = mul nuw <32 x i32> [[TMP7:%.*]], <i32 7471, i32 7471, i32 7471, i32 7471, i32 7471, i32 7471, i32 7471, i32 7471, i32 7471, i32 7471, i32 7471, i32 7471, i32 7471, i32 7471, i32 7471, i32 7471, i32 16759568, i32 16759568, i32 16759568, i32 16759568, i32 16759568, i32 16759568, i32 16759568, i32 16759568, i32 16759568, i32 16759568, i32 16759568, i32 16759568, i32 16759568, i32 16759568, i32 16759568, i32 16759568>

bb:
br label %bb48

bb48: ; preds = %bb48, %bb
%phi49 = phi i64 [ 0, %bb ], [ %add86, %bb48 ]
%mul50 = mul i64 %phi49, 3
%getelementptr53 = getelementptr i8, ptr %call12, i64 %mul50
%load54 = load <48 x i8>, ptr %getelementptr53, align 1, !tbaa !10, !alias.scope !18
%shufflevector = shufflevector <48 x i8> %load54, <48 x i8> poison, <16 x i32> <i32 0, i32 3, i32 6, i32 9, i32 12, i32 15, i32 18, i32 21, i32 24, i32 27, i32 30, i32 33, i32 36, i32 39, i32 42, i32 45>
%zext57 = zext <16 x i8> %shufflevector to <16 x i32>
%mul58 = mul nuw nsw <16 x i32> %zext57, splat (i32 19595)
%zext59 = zext <16 x i8> %shufflevector to <16 x i32>
%mul60 = mul nuw nsw <16 x i32> %zext59, splat (i32 38470)
%zext61 = zext <16 x i8> %shufflevector to <16 x i32>
%mul62 = mul nuw nsw <16 x i32> %zext61, splat (i32 7471)
%add63 = add nuw nsw <16 x i32> %mul58, splat (i32 32768)
%add64 = add nuw nsw <16 x i32> %add63, %mul60
%add65 = add nuw nsw <16 x i32> %add64, %mul62
%lshr = lshr <16 x i32> %add65, splat (i32 16)
%trunc66 = trunc nuw <16 x i32> %lshr to <16 x i8>
%mul67 = mul nuw nsw <16 x i32> %zext57, splat (i32 32767)
%mul68 = mul nuw <16 x i32> %zext59, splat (i32 16762097)
%mul69 = mul nuw <16 x i32> %zext61, splat (i32 16759568)
%add70 = add nuw nsw <16 x i32> %mul67, splat (i32 32768)
%add71 = add nuw <16 x i32> %add70, %mul68
%add72 = add <16 x i32> %add71, %mul69
%lshr73 = lshr <16 x i32> %add72, splat (i32 16)
%trunc74 = trunc <16 x i32> %lshr73 to <16 x i8>
%mul75 = mul nuw nsw <16 x i32> %zext57, splat (i32 13282)
%mul76 = mul nuw <16 x i32> %zext59, splat (i32 16744449)
%mul77 = mul nuw nsw <16 x i32> %zext61, splat (i32 19485)
%add78 = add nuw nsw <16 x i32> %mul75, splat (i32 32768)
%add79 = add nuw <16 x i32> %add78, %mul76
%add80 = add nuw <16 x i32> %add79, %mul77
%lshr81 = lshr <16 x i32> %add80, splat (i32 16)
%trunc82 = trunc <16 x i32> %lshr81 to <16 x i8>
%shufflevector83 = shufflevector <16 x i8> %trunc66, <16 x i8> %trunc74, <32 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, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
%shufflevector84 = shufflevector <16 x i8> %trunc82, <16 x i8> poison, <32 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, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison>
store <32 x i8> %shufflevector83, ptr %getelementptr53, align 1, !tbaa !10, !noalias !18
%add86 = add nuw i64 %phi49, 16
%icmp87 = icmp eq i64 %add86, %mul50
br i1 %icmp87, label %bb205, label %bb48

bb205: ; preds = %bb48, %bb
ret i32 0
}

attributes #0 = { nounwind uwtable vscale_range(1,16) "target-cpu"="cortex-a57" "target-features"="+sve,+neon"}

!8 = !{!"omnipotent char", !9, i64 0}
!9 = !{!"Simple C/C++ TBAA"}
!10 = !{!8, !8, i64 0}
!12 = !{!"int", !8, i64 0}
!18 = !{!19}
!19 = distinct !{!19, !20}
!20 = distinct !{!20, !"LVerDomain"}