[llvm] [InstCombineCompares] Replace the sqrt in if-condition (PR #91707)
via llvm-commits
llvm-commits at lists.llvm.org
Fri May 10 00:41:17 PDT 2024
llvmbot wrote:
<!--LLVM PR SUMMARY COMMENT-->
@llvm/pr-subscribers-llvm-transforms
Author: kk_compiler (whokeke)
<details>
<summary>Changes</summary>
Adding the transform, fcmp sqrt(X), Y => fcmp X, copysign(Y*Y, Y) , to help sink the sqrt. And give a option `replace-sqrt-compare-by-square` to control it.
---
Full diff: https://github.com/llvm/llvm-project/pull/91707.diff
2 Files Affected:
- (modified) llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp (+49)
- (added) llvm/test/Transforms/InstCombine/fcmp-fsqrt.ll (+87)
``````````diff
diff --git a/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp b/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
index e1a3194a1beb7..ed0a11a56ab2c 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
@@ -38,6 +38,11 @@ using namespace PatternMatch;
// How many times is a select replaced by one of its operands?
STATISTIC(NumSel, "Number of select opts");
+static cl::opt<bool> ReplaceSqrtInIfCondition(
+ "replace-sqrt-compare-by-square", cl::init(false), cl::Hidden,
+ cl::desc(
+ "Try to replcae the sqrt, in the comparison condition, by the square."
+ "Like that fcmp sqrt(X), Y => fcmp X, copysing(Y*Y, Y)"));
/// Compute Result = In1+In2, returning true if the result overflowed for this
/// type.
@@ -8068,6 +8073,50 @@ Instruction *InstCombinerImpl::visitFCmpInst(FCmpInst &I) {
return new FCmpInst(I.getSwappedPredicate(), X, NegC, "", &I);
}
+ /// Try to sink the sqrt in this following case:
+ /// \code
+ /// tmp = sqrt(input)
+ /// if (tmp > cond)
+ /// branch_without_tmp
+ /// else
+ /// branch_with_tmp
+ /// ...
+ /// \endcode
+ /// Is optimized to:
+ /// \code
+ /// if (input > copysign(cond*cond, cond))
+ /// branch_without_tmp
+ /// else
+ /// tmp = sqrt(input)
+ /// branch_with_tmp
+ /// ...
+ /// \endcode
+ /// Only call sqrt in the branch that needs the sqrt result,
+ /// not if it is not needed, reducing the number of calls to sqrt.
+ if (I.isFast() && ReplaceSqrtInIfCondition) {
+ // fcmp sqrt(X), sqrt(Y) => fcmp X, Y
+ if (match(Op0, m_Intrinsic<Intrinsic::sqrt>(m_Value(X))) &&
+ match(Op1, m_Intrinsic<Intrinsic::sqrt>(m_Value(Y)))) {
+ auto CIX = cast<CallInst>(Op0);
+ auto CIY = cast<CallInst>(Op1);
+ return new FCmpInst(Pred, CIX->getOperand(0), CIY->getOperand(0), ""
+ } else if (match(Op0, m_Intrinsic<Intrinsic::sqrt>(m_Value(X)))) {
+ // fcmp sqrt(X), Y => fcmp X, copysing(Y*Y, Y)
+ auto CI = cast<CallInst>(Op0);
+ Value *YY = Builder.CreateFMulFMF(Op1, Op1, &I);
+ Value *SYY =
+ Builder.CreateBinaryIntrinsic(Intrinsic::copysign, YY, Op1, &I);
+ return new FCmpInst(Pred, CI->getOperand(0), SYY, "", &I);
+ } else if (match(Op1, m_Intrinsic<Intrinsic::sqrt>(m_Value(Y)))) {
+ // fcmp X, sqrt(Y) => fcmp copysign(X*X, X), Y
+ auto CI = cast<CallInst>(Op1);
+ Value *XX = Builder.CreateFMulFMF(Op0, Op0, &I);
+ Value *SXX =
+ Builder.CreateBinaryIntrinsic(Intrinsic::copysign, XX, Op0, &I);
+ return new FCmpInst(Pred, SXX, CI->getOperand(0), "", &I);
+ }
+ }
+
// fcmp (fadd X, 0.0), Y --> fcmp X, Y
if (match(Op0, m_FAdd(m_Value(X), m_AnyZeroFP())))
return new FCmpInst(Pred, X, Op1, "", &I);
diff --git a/llvm/test/Transforms/InstCombine/fcmp-fsqrt.ll b/llvm/test/Transforms/InstCombine/fcmp-fsqrt.ll
new file mode 100644
index 0000000000000..919e2ecfff962
--- /dev/null
+++ b/llvm/test/Transforms/InstCombine/fcmp-fsqrt.ll
@@ -0,0 +1,87 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt < %s -instcombine -S -replace-sqrt-compare-by-square=true | FileCheck %s -check-prefix=OPT-TRUE
+; RUN: opt < %s -instcombine -S | FileCheck %s -check-prefix=OPT-FALSE
+
+define i1 @foo1_fast(float %a, float %b) {
+; OPT-TRUE-LABEL: @foo1_fast(
+; OPT-TRUE-NEXT: [[FCMP:%.*]] = fcmp fast ogt float %a, %b
+; OPT-TRUE-NEXT: ret i1 [[FCMP]]
+;
+; OPT-FALSE-LABEL: @foo1_fast(
+; OPT-FALSE-NEXT: %c = call fast float @llvm.sqrt.f32(float %a)
+; OPT-FALSE-NEXT: %d = call fast float @llvm.sqrt.f32(float %b)
+; OPT-FALSE-NEXT: [[FCMP:%.*]] = fcmp fast ogt float %c, %d
+; OPT-FALSE-NEXT: ret i1 [[FCMP]]
+;
+ %c = call fast float @llvm.sqrt.f32(float %a)
+ %d = call fast float @llvm.sqrt.f32(float %b)
+ %e = fcmp fast ogt float %c, %d
+ ret i1 %e
+}
+
+define i1 @foo2_fast(float %a, float %b) {
+; OPT-TRUE-LABEL: @foo2_fast(
+; OPT-TRUE-NEXT: [[FMUL:%.*]] = fmul fast float %b, %b
+; OPT-TRUE-NEXT: [[SIGN:%.*]] = call fast float @llvm.copysign.f32(float [[FMUL]], float %b)
+; OPT-TRUE-NEXT: [[FCMP:%.*]] = fcmp fast ogt float [[SIGN]], %a
+; OPT-TRUE-NEXT: ret i1 [[FCMP]]
+;
+; OPT-FALSE-LABEL: @foo2_fast(
+; OPT-FALSE-NEXT: [[SQRT:%.*]] = call fast float @llvm.sqrt.f32(float %a)
+; OPT-FALSE-NEXT: [[FCMP:%.*]] = fcmp fast olt float [[SQRT]], %b
+; OPT-FALSE-NEXT: ret i1 [[FCMP]]
+;
+ %c = call fast float @llvm.sqrt.f32(float %a)
+ %d = fcmp fast ogt float %b, %c
+ ret i1 %d
+}
+
+define i1 @foo3_fast(float %a, float %b) {
+; OPT-TRUE-LABEL: @foo3_fast(
+; OPT-TRUE-NEXT: [[FMUL:%.*]] = fmul fast float %b, %b
+; OPT-TRUE-NEXT: [[SIGN:%.*]] = call fast float @llvm.copysign.f32(float [[FMUL]], float %b)
+; OPT-TRUE-NEXT: [[FCMP:%.*]] = fcmp fast olt float [[SIGN]], %a
+; OPT-TRUE-NEXT: ret i1 [[FCMP]]
+;
+; OPT-FALSE-LABEL: @foo3_fast(
+; OPT-FALSE-NEXT: [[SQRT:%.*]] = call fast float @llvm.sqrt.f32(float %a)
+; OPT-FALSE-NEXT: [[FCMP:%.*]] = fcmp fast ogt float [[SQRT]], %b
+; OPT-FALSE-NEXT: ret i1 [[FCMP]]
+;
+ %c = call fast float @llvm.sqrt.f32(float %a)
+ %d = fcmp fast ogt float %c, %b
+ ret i1 %d
+}
+
+define i1 @foo1_no_fast(float %a, float %b) {
+; OPT-TRUE-LABEL: @foo1_no_fast(
+; OPT-TRUE: [[FCMP:%.*]] = fcmp ogt float %c, %d
+; OPT-TRUE-NEXT: ret i1 [[FCMP]]
+;
+ %c = call float @llvm.sqrt.f32(float %a)
+ %d = call float @llvm.sqrt.f32(float %b)
+ %e = fcmp ogt float %c, %d
+ ret i1 %e
+}
+
+define i1 @foo2_no_fast(float %a, float %b) {
+; OPT-TRUE-LABEL: @foo2_no_fast(
+; OPT-TRUE: [[FCMP:%.*]] = fcmp olt float %c, %b
+; OPT-TRUE-NEXT: ret i1 [[FCMP]]
+;
+ %c = call float @llvm.sqrt.f32(float %a)
+ %d = fcmp ogt float %b, %c
+ ret i1 %d
+}
+
+define i1 @foo3_no_fast(float %a, float %b) {
+; OPT-TRUE-LABEL: @foo3_no_fast(
+; OPT-TRUE: [[FCMP:%.*]] = fcmp ogt float %c, %b
+; OPT-TRUE-NEXT: ret i1 [[FCMP]]
+;
+ %c = call float @llvm.sqrt.f32(float %a)
+ %d = fcmp ogt float %c, %b
+ ret i1 %d
+}
+
+declare float @llvm.sqrt.f32(float)
``````````
</details>
https://github.com/llvm/llvm-project/pull/91707
More information about the llvm-commits
mailing list