Skip to content

[InstCombine] Optimise x / sqrt(y / z) with fast-math pattern. #76737

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

Merged
merged 4 commits into from
Feb 9, 2024
Merged

[InstCombine] Optimise x / sqrt(y / z) with fast-math pattern. #76737

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
f106f58
[InstCombine] Optimise x / sqrt(y / z) with fast-math pattern.
zjaffal Jan 2, 2024
303ce4c
[InstCombine] Make sure all instructions have `arcp` and `reassoc` flags
zjaffal Jan 9, 2024
40ab484
Reorder checks
arsenm Feb 8, 2024
fca21cf
Fix duplicated condition and formatting
arsenm Feb 8, 2024
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
30 changes: 30 additions & 0 deletions llvm/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1709,6 +1709,33 @@ static Instruction *foldFDivPowDivisor(BinaryOperator &I,
return BinaryOperator::CreateFMulFMF(Op0, Pow, &I);
}

/// Convert div to mul if we have an sqrt divisor iff sqrt's operand is a fdiv
/// instruction.
static Instruction *foldFDivSqrtDivisor(BinaryOperator &I,
InstCombiner::BuilderTy &Builder) {
// X / sqrt(Y / Z) --> X * sqrt(Z / Y)
if (!I.hasAllowReassoc() || !I.hasAllowReciprocal())
return nullptr;
Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
auto *II = dyn_cast<IntrinsicInst>(Op1);
if (!II || II->getIntrinsicID() != Intrinsic::sqrt || !II->hasOneUse() ||
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Would be cleaner to use PatternMatch. Something like

if (match(I, m_Sqrt(m_OneUse(m_FDiv(m_Value(Op0), m_Value(Op1))) {

}

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I agree but then we won't be able to check if sqrt has one use or fdiv has the necessary flags

!II->hasAllowReassoc() || !II->hasAllowReciprocal())
return nullptr;

Value *Y, *Z;
auto *DivOp = dyn_cast<Instruction>(II->getOperand(0));
if (!DivOp || !DivOp->hasAllowReassoc() || !I.hasAllowReciprocal() ||
!DivOp->hasOneUse())
return nullptr;
if (match(DivOp, m_FDiv(m_Value(Y), m_Value(Z)))) {
Value *SwapDiv = Builder.CreateFDivFMF(Z, Y, DivOp);
Value *NewSqrt =
Builder.CreateUnaryIntrinsic(II->getIntrinsicID(), SwapDiv, II);
return BinaryOperator::CreateFMulFMF(Op0, NewSqrt, &I);
}
return nullptr;
}

Instruction *InstCombinerImpl::visitFDiv(BinaryOperator &I) {
Module *M = I.getModule();

Expand Down Expand Up @@ -1816,6 +1843,9 @@ Instruction *InstCombinerImpl::visitFDiv(BinaryOperator &I) {
if (Instruction *Mul = foldFDivPowDivisor(I, Builder))
return Mul;

if (Instruction *Mul = foldFDivSqrtDivisor(I, Builder))
return Mul;

// pow(X, Y) / X --> pow(X, Y-1)
if (I.hasAllowReassoc() &&
match(Op0, m_OneUse(m_Intrinsic<Intrinsic::pow>(m_Specific(Op1),
Expand Down
18 changes: 9 additions & 9 deletions llvm/test/Transforms/InstCombine/fdiv-sqrt.ll
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -6,9 +6,9 @@ declare double @llvm.sqrt.f64(double)
define double @sqrt_div_fast(double %x, double %y, double %z) {
; CHECK-LABEL: @sqrt_div_fast(
; CHECK-NEXT: entry:
; CHECK-NEXT: [[DIV:%.*]] = fdiv fast double [[Y:%.*]], [[Z:%.*]]
; CHECK-NEXT: [[SQRT:%.*]] = call fast double @llvm.sqrt.f64(double [[DIV]])
; CHECK-NEXT: [[DIV1:%.*]] = fdiv fast double [[X:%.*]], [[SQRT]]
; CHECK-NEXT: [[TMP0:%.*]] = fdiv fast double [[Z:%.*]], [[Y:%.*]]
; CHECK-NEXT: [[TMP1:%.*]] = call fast double @llvm.sqrt.f64(double [[TMP0]])
; CHECK-NEXT: [[DIV1:%.*]] = fmul fast double [[TMP1]], [[X:%.*]]
; CHECK-NEXT: ret double [[DIV1]]
;
entry:
Expand Down Expand Up @@ -36,9 +36,9 @@ entry:
define double @sqrt_div_reassoc_arcp(double %x, double %y, double %z) {
; CHECK-LABEL: @sqrt_div_reassoc_arcp(
; CHECK-NEXT: entry:
; CHECK-NEXT: [[DIV:%.*]] = fdiv reassoc arcp double [[Y:%.*]], [[Z:%.*]]
; CHECK-NEXT: [[SQRT:%.*]] = call reassoc arcp double @llvm.sqrt.f64(double [[DIV]])
; CHECK-NEXT: [[DIV1:%.*]] = fdiv reassoc arcp double [[X:%.*]], [[SQRT]]
; CHECK-NEXT: [[TMP0:%.*]] = fdiv reassoc arcp double [[Z:%.*]], [[Y:%.*]]
; CHECK-NEXT: [[TMP1:%.*]] = call reassoc arcp double @llvm.sqrt.f64(double [[TMP0]])
; CHECK-NEXT: [[DIV1:%.*]] = fmul reassoc arcp double [[TMP1]], [[X:%.*]]
; CHECK-NEXT: ret double [[DIV1]]
;
entry:
Expand Down Expand Up @@ -96,9 +96,9 @@ entry:
define double @sqrt_div_arcp_missing(double %x, double %y, double %z) {
; CHECK-LABEL: @sqrt_div_arcp_missing(
; CHECK-NEXT: entry:
; CHECK-NEXT: [[DIV:%.*]] = fdiv reassoc double [[Y:%.*]], [[Z:%.*]]
; CHECK-NEXT: [[SQRT:%.*]] = call reassoc arcp double @llvm.sqrt.f64(double [[DIV]])
; CHECK-NEXT: [[DIV1:%.*]] = fdiv reassoc arcp double [[X:%.*]], [[SQRT]]
; CHECK-NEXT: [[TMP0:%.*]] = fdiv reassoc double [[Z:%.*]], [[Y:%.*]]
; CHECK-NEXT: [[TMP1:%.*]] = call reassoc arcp double @llvm.sqrt.f64(double [[TMP0]])
; CHECK-NEXT: [[DIV1:%.*]] = fmul reassoc arcp double [[TMP1]], [[X:%.*]]
; CHECK-NEXT: ret double [[DIV1]]
;
entry:
Expand Down