Skip to content

[SLP] Include vectorized calls in spill cost #125650

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

Closed
wants to merge 2 commits into from
Closed

[SLP] Include vectorized calls in spill cost #125650

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
1a01b91
Precommit test
lukel97 Feb 4, 2025
2761386
[SLP] Include vectorized calls in spill cost
lukel97 Feb 4, 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
6 changes: 3 additions & 3 deletions llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -12248,10 +12248,10 @@ InstructionCost BoUpSLP::getSpillCost() {
};

// Debug information does not impact spill cost.
// Vectorized calls, represented as vector intrinsics, do not impact spill
// cost.
// Vectorized calls, represented as vector intrinsics, may still impact
// spill cost if scalarized in codegen.
if (const auto *CB = dyn_cast<CallBase>(&*PrevInstIt);
CB && !NoCallIntrinsic(CB) && !isVectorized(CB))
CB && !NoCallIntrinsic(CB))
NumCalls++;

++PrevInstIt;
Expand Down
46 changes: 46 additions & 0 deletions llvm/test/Transforms/SLPVectorizer/RISCV/math-function.ll
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -141,6 +141,52 @@ entry:
ret <4 x float> %vecins.3
}

; Don't vectorize @llvm.exp because it will be scalarized in codegen.
define void @exp_v2f64(ptr %a) {
; CHECK-LABEL: define void @exp_v2f64
; CHECK-SAME: (ptr [[A:%.*]]) #[[ATTR1]] {
; CHECK-NEXT: [[X:%.*]] = load double, ptr [[A]], align 8
; CHECK-NEXT: [[GEP:%.*]] = getelementptr double, ptr [[A]], i64 1
; CHECK-NEXT: [[Y:%.*]] = load double, ptr [[GEP]], align 8
; CHECK-NEXT: [[X_ADD:%.*]] = fadd double [[X]], 1.000000e+00
; CHECK-NEXT: [[Y_ADD:%.*]] = fadd double [[Y]], 1.000000e+00
; CHECK-NEXT: [[X_EXP:%.*]] = call double @llvm.exp.f64(double [[X_ADD]])
; CHECK-NEXT: [[Y_EXP:%.*]] = call double @llvm.exp.f64(double [[Y_ADD]])
; CHECK-NEXT: [[X_ADD2:%.*]] = fadd double [[X_EXP]], 1.000000e+00
; CHECK-NEXT: [[Y_ADD2:%.*]] = fadd double [[Y_EXP]], 1.000000e+00
; CHECK-NEXT: store double [[X_ADD2]], ptr [[A]], align 8
; CHECK-NEXT: store double [[Y_ADD2]], ptr [[GEP]], align 8
; CHECK-NEXT: ret void
;
; DEFAULT-LABEL: define void @exp_v2f64
; DEFAULT-SAME: (ptr [[A:%.*]]) #[[ATTR1]] {
; DEFAULT-NEXT: [[X:%.*]] = load double, ptr [[A]], align 8
; DEFAULT-NEXT: [[GEP:%.*]] = getelementptr double, ptr [[A]], i64 1
; DEFAULT-NEXT: [[Y:%.*]] = load double, ptr [[GEP]], align 8
; DEFAULT-NEXT: [[X_ADD:%.*]] = fadd double [[X]], 1.000000e+00
; DEFAULT-NEXT: [[Y_ADD:%.*]] = fadd double [[Y]], 1.000000e+00
; DEFAULT-NEXT: [[X_EXP:%.*]] = call double @llvm.exp.f64(double [[X_ADD]])
; DEFAULT-NEXT: [[Y_EXP:%.*]] = call double @llvm.exp.f64(double [[Y_ADD]])
; DEFAULT-NEXT: [[X_ADD2:%.*]] = fadd double [[X_EXP]], 1.000000e+00
; DEFAULT-NEXT: [[Y_ADD2:%.*]] = fadd double [[Y_EXP]], 1.000000e+00
; DEFAULT-NEXT: store double [[X_ADD2]], ptr [[A]], align 8
; DEFAULT-NEXT: store double [[Y_ADD2]], ptr [[GEP]], align 8
; DEFAULT-NEXT: ret void
;
%x = load double, ptr %a
%gep = getelementptr double, ptr %a, i64 1
%y = load double, ptr %gep
%x.add = fadd double %x, 1.0
%y.add = fadd double %y, 1.0
%x.exp = call double @llvm.exp(double %x.add)
%y.exp = call double @llvm.exp(double %y.add)
%x.add2 = fadd double %x.exp, 1.0
%y.add2 = fadd double %y.exp, 1.0
store double %x.add2, ptr %a
store double %y.add2, ptr %gep
ret void
}

declare float @expf(float) readonly nounwind willreturn

; We can not vectorized exp since RISCV has no such instruction.
Expand Down