[llvm] [InstCombine] Canonicalize the fcmp range check idiom into `fabs + fcmp` (PR #76367)
via llvm-commits
llvm-commits at lists.llvm.org
Mon Dec 25 11:54:23 PST 2023
llvmbot wrote:
<!--LLVM PR SUMMARY COMMENT-->
@llvm/pr-subscribers-llvm-transforms
Author: Yingwei Zheng (dtcxzyw)
<details>
<summary>Changes</summary>
This patch canonicalizes the fcmp range check idiom into `fabs + fcmp` since the canonicalized form is better than the original form for the backends.
Godbolt: https://godbolt.org/z/x3eqPb1fz
```
and (fcmp olt/ole/ult/ule x, C), (fcmp ogt/oge/ugt/uge x, -C) --> fabs(x) olt/ole/ult/ule C
or (fcmp ogt/oge/ugt/uge x, C), (fcmp olt/ole/ult/ule x, -C) --> fabs(x) ogt/oge/ugt/uge C
```
Alive2: https://alive2.llvm.org/ce/z/MRtoYq
---
Full diff: https://github.com/llvm/llvm-project/pull/76367.diff
2 Files Affected:
- (modified) llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp (+26)
- (added) llvm/test/Transforms/InstCombine/fcmp-range-check-idiom.ll (+279)
``````````diff
diff --git a/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp b/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp
index 5e362f4117d051..689ee6eaabb4b9 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp
@@ -1457,6 +1457,32 @@ Value *InstCombinerImpl::foldLogicOfFCmps(FCmpInst *LHS, FCmpInst *RHS,
}
}
+ // Canonicalize the range check idiom:
+ // and (fcmp olt/ole/ult/ule x, C), (fcmp ogt/oge/ugt/uge x, -C)
+ // --> fabs(x) olt/ole/ult/ule C
+ // or (fcmp ogt/oge/ugt/uge x, C), (fcmp olt/ole/ult/ule x, -C)
+ // --> fabs(x) ogt/oge/ugt/uge C
+ // TODO: Generalize to handle a negated variable operand?
+ const APFloat *LHSC, *RHSC;
+ if (LHS->hasOneUse() && RHS->hasOneUse() && LHS0 == RHS0 &&
+ FCmpInst::getSwappedPredicate(PredL) == PredR &&
+ match(LHS1, m_APFloatAllowUndef(LHSC)) &&
+ match(RHS1, m_APFloatAllowUndef(RHSC)) &&
+ LHSC->bitwiseIsEqual(neg(*RHSC))) {
+ auto IsLessThanOrLessEqual = [](FCmpInst::Predicate Pred) {
+ return (getFCmpCode(Pred) & 0b0110) == 0b0100;
+ };
+ if (IsLessThanOrLessEqual(IsAnd ? PredR : PredL)) {
+ std::swap(LHSC, RHSC);
+ std::swap(PredL, PredR);
+ }
+ if (IsLessThanOrLessEqual(IsAnd ? PredL : PredR)) {
+ Value *FAbs = Builder.CreateUnaryIntrinsic(Intrinsic::fabs, LHS0);
+ return Builder.CreateFCmp(PredL, FAbs,
+ ConstantFP::get(LHS0->getType(), *LHSC));
+ }
+ }
+
return nullptr;
}
diff --git a/llvm/test/Transforms/InstCombine/fcmp-range-check-idiom.ll b/llvm/test/Transforms/InstCombine/fcmp-range-check-idiom.ll
new file mode 100644
index 00000000000000..06e5ca05ff1501
--- /dev/null
+++ b/llvm/test/Transforms/InstCombine/fcmp-range-check-idiom.ll
@@ -0,0 +1,279 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 4
+; RUN: opt -S -passes=instcombine %s | FileCheck %s
+
+declare void @use(i1)
+
+define i1 @test_and_olt(float %x) {
+; CHECK-LABEL: define i1 @test_and_olt(
+; CHECK-SAME: float [[X:%.*]]) {
+; CHECK-NEXT: [[TMP1:%.*]] = call float @llvm.fabs.f32(float [[X]])
+; CHECK-NEXT: [[COND:%.*]] = fcmp olt float [[TMP1]], 0x3C00000000000000
+; CHECK-NEXT: ret i1 [[COND]]
+;
+ %cmp1 = fcmp olt float %x, 0x3C00000000000000
+ %cmp2 = fcmp ogt float %x, 0xBC00000000000000
+ %cond = and i1 %cmp1, %cmp2
+ ret i1 %cond
+}
+define i1 @test_and_ole(float %x) {
+; CHECK-LABEL: define i1 @test_and_ole(
+; CHECK-SAME: float [[X:%.*]]) {
+; CHECK-NEXT: [[TMP1:%.*]] = call float @llvm.fabs.f32(float [[X]])
+; CHECK-NEXT: [[COND:%.*]] = fcmp ole float [[TMP1]], 0x3C00000000000000
+; CHECK-NEXT: ret i1 [[COND]]
+;
+ %cmp1 = fcmp ole float %x, 0x3C00000000000000
+ %cmp2 = fcmp oge float %x, 0xBC00000000000000
+ %cond = and i1 %cmp1, %cmp2
+ ret i1 %cond
+}
+define i1 @test_or_ogt(float %x) {
+; CHECK-LABEL: define i1 @test_or_ogt(
+; CHECK-SAME: float [[X:%.*]]) {
+; CHECK-NEXT: [[TMP1:%.*]] = call float @llvm.fabs.f32(float [[X]])
+; CHECK-NEXT: [[COND:%.*]] = fcmp ogt float [[TMP1]], 0x3C00000000000000
+; CHECK-NEXT: ret i1 [[COND]]
+;
+ %cmp1 = fcmp ogt float %x, 0x3C00000000000000
+ %cmp2 = fcmp olt float %x, 0xBC00000000000000
+ %cond = or i1 %cmp1, %cmp2
+ ret i1 %cond
+}
+define i1 @test_or_oge(float %x) {
+; CHECK-LABEL: define i1 @test_or_oge(
+; CHECK-SAME: float [[X:%.*]]) {
+; CHECK-NEXT: [[TMP1:%.*]] = call float @llvm.fabs.f32(float [[X]])
+; CHECK-NEXT: [[COND:%.*]] = fcmp oge float [[TMP1]], 0x3C00000000000000
+; CHECK-NEXT: ret i1 [[COND]]
+;
+ %cmp1 = fcmp oge float %x, 0x3C00000000000000
+ %cmp2 = fcmp ole float %x, 0xBC00000000000000
+ %cond = or i1 %cmp1, %cmp2
+ ret i1 %cond
+}
+define i1 @test_and_ult(float %x) {
+; CHECK-LABEL: define i1 @test_and_ult(
+; CHECK-SAME: float [[X:%.*]]) {
+; CHECK-NEXT: [[TMP1:%.*]] = call float @llvm.fabs.f32(float [[X]])
+; CHECK-NEXT: [[COND:%.*]] = fcmp ult float [[TMP1]], 0x3C00000000000000
+; CHECK-NEXT: ret i1 [[COND]]
+;
+ %cmp1 = fcmp ult float %x, 0x3C00000000000000
+ %cmp2 = fcmp ugt float %x, 0xBC00000000000000
+ %cond = and i1 %cmp1, %cmp2
+ ret i1 %cond
+}
+define i1 @test_and_ule(float %x) {
+; CHECK-LABEL: define i1 @test_and_ule(
+; CHECK-SAME: float [[X:%.*]]) {
+; CHECK-NEXT: [[TMP1:%.*]] = call float @llvm.fabs.f32(float [[X]])
+; CHECK-NEXT: [[COND:%.*]] = fcmp ule float [[TMP1]], 0x3C00000000000000
+; CHECK-NEXT: ret i1 [[COND]]
+;
+ %cmp1 = fcmp ule float %x, 0x3C00000000000000
+ %cmp2 = fcmp uge float %x, 0xBC00000000000000
+ %cond = and i1 %cmp1, %cmp2
+ ret i1 %cond
+}
+define i1 @test_or_ugt(float %x) {
+; CHECK-LABEL: define i1 @test_or_ugt(
+; CHECK-SAME: float [[X:%.*]]) {
+; CHECK-NEXT: [[TMP1:%.*]] = call float @llvm.fabs.f32(float [[X]])
+; CHECK-NEXT: [[COND:%.*]] = fcmp ugt float [[TMP1]], 0x3C00000000000000
+; CHECK-NEXT: ret i1 [[COND]]
+;
+ %cmp1 = fcmp ugt float %x, 0x3C00000000000000
+ %cmp2 = fcmp ult float %x, 0xBC00000000000000
+ %cond = or i1 %cmp1, %cmp2
+ ret i1 %cond
+}
+define i1 @test_or_uge(float %x) {
+; CHECK-LABEL: define i1 @test_or_uge(
+; CHECK-SAME: float [[X:%.*]]) {
+; CHECK-NEXT: [[TMP1:%.*]] = call float @llvm.fabs.f32(float [[X]])
+; CHECK-NEXT: [[COND:%.*]] = fcmp uge float [[TMP1]], 0x3C00000000000000
+; CHECK-NEXT: ret i1 [[COND]]
+;
+ %cmp1 = fcmp uge float %x, 0x3C00000000000000
+ %cmp2 = fcmp ule float %x, 0xBC00000000000000
+ %cond = or i1 %cmp1, %cmp2
+ ret i1 %cond
+}
+define i1 @test_and_olt_commuted(float %x) {
+; CHECK-LABEL: define i1 @test_and_olt_commuted(
+; CHECK-SAME: float [[X:%.*]]) {
+; CHECK-NEXT: [[TMP1:%.*]] = call float @llvm.fabs.f32(float [[X]])
+; CHECK-NEXT: [[COND:%.*]] = fcmp olt float [[TMP1]], 0x3C00000000000000
+; CHECK-NEXT: ret i1 [[COND]]
+;
+ %cmp1 = fcmp olt float %x, 0x3C00000000000000
+ %cmp2 = fcmp ogt float %x, 0xBC00000000000000
+ %cond = and i1 %cmp2, %cmp1
+ ret i1 %cond
+}
+define i1 @test_and_olt_subnormal(float %x) {
+; CHECK-LABEL: define i1 @test_and_olt_subnormal(
+; CHECK-SAME: float [[X:%.*]]) {
+; CHECK-NEXT: [[TMP1:%.*]] = call float @llvm.fabs.f32(float [[X]])
+; CHECK-NEXT: [[COND:%.*]] = fcmp olt float [[TMP1]], 0x36A0000000000000
+; CHECK-NEXT: ret i1 [[COND]]
+;
+ %cmp1 = fcmp olt float %x, 0x36A0000000000000
+ %cmp2 = fcmp ogt float %x, 0xB6A0000000000000
+ %cond = and i1 %cmp1, %cmp2
+ ret i1 %cond
+}
+define i1 @test_and_olt_infinity(float %x) {
+; CHECK-LABEL: define i1 @test_and_olt_infinity(
+; CHECK-SAME: float [[X:%.*]]) {
+; CHECK-NEXT: [[TMP1:%.*]] = call float @llvm.fabs.f32(float [[X]])
+; CHECK-NEXT: [[COND:%.*]] = fcmp one float [[TMP1]], 0x7FF0000000000000
+; CHECK-NEXT: ret i1 [[COND]]
+;
+ %cmp1 = fcmp olt float %x, 0x7FF0000000000000
+ %cmp2 = fcmp ogt float %x, 0xFFF0000000000000
+ %cond = and i1 %cmp1, %cmp2
+ ret i1 %cond
+}
+define i1 @test_and_olt_zero(float %x) {
+; CHECK-LABEL: define i1 @test_and_olt_zero(
+; CHECK-SAME: float [[X:%.*]]) {
+; CHECK-NEXT: ret i1 false
+;
+ %cmp1 = fcmp olt float %x, 0x0000000000000000
+ %cmp2 = fcmp ogt float %x, 0x8000000000000000
+ %cond = and i1 %cmp1, %cmp2
+ ret i1 %cond
+}
+define i1 @test_and_ole_zero(float %x) {
+; CHECK-LABEL: define i1 @test_and_ole_zero(
+; CHECK-SAME: float [[X:%.*]]) {
+; CHECK-NEXT: [[COND:%.*]] = fcmp oeq float [[X]], 0.000000e+00
+; CHECK-NEXT: ret i1 [[COND]]
+;
+ %cmp1 = fcmp ole float %x, 0x0000000000000000
+ %cmp2 = fcmp oge float %x, 0x8000000000000000
+ %cond = and i1 %cmp1, %cmp2
+ ret i1 %cond
+}
+define i1 @test_and_olt_logical(float %x) {
+; CHECK-LABEL: define i1 @test_and_olt_logical(
+; CHECK-SAME: float [[X:%.*]]) {
+; CHECK-NEXT: [[TMP1:%.*]] = call float @llvm.fabs.f32(float [[X]])
+; CHECK-NEXT: [[COND:%.*]] = fcmp olt float [[TMP1]], 0x3C00000000000000
+; CHECK-NEXT: ret i1 [[COND]]
+;
+ %cmp1 = fcmp olt float %x, 0x3C00000000000000
+ %cmp2 = fcmp ogt float %x, 0xBC00000000000000
+ %cond = select i1 %cmp1, i1 %cmp2, i1 false
+ ret i1 %cond
+}
+define <2 x i1> @test_and_olt_undef(<2 x float> %x) {
+; CHECK-LABEL: define <2 x i1> @test_and_olt_undef(
+; CHECK-SAME: <2 x float> [[X:%.*]]) {
+; CHECK-NEXT: [[TMP1:%.*]] = call <2 x float> @llvm.fabs.v2f32(<2 x float> [[X]])
+; CHECK-NEXT: [[COND:%.*]] = fcmp olt <2 x float> [[TMP1]], <float 0x3C00000000000000, float 0x3C00000000000000>
+; CHECK-NEXT: ret <2 x i1> [[COND]]
+;
+ %cmp1 = fcmp olt <2 x float> %x, <float 0x3C00000000000000, float undef>
+ %cmp2 = fcmp ogt <2 x float> %x, <float 0xBC00000000000000, float undef>
+ %cond = and <2 x i1> %cmp1, %cmp2
+ ret <2 x i1> %cond
+}
+define i1 @test_and_olt_nan(float %x) {
+; CHECK-LABEL: define i1 @test_and_olt_nan(
+; CHECK-SAME: float [[X:%.*]]) {
+; CHECK-NEXT: ret i1 false
+;
+ %cmp1 = fcmp olt float %x, 0x7FF8000000000000
+ %cmp2 = fcmp ogt float %x, 0xFFF8000000000000
+ %cond = and i1 %cmp1, %cmp2
+ ret i1 %cond
+}
+define i1 @test_and_ogt(float %x) {
+; CHECK-LABEL: define i1 @test_and_ogt(
+; CHECK-SAME: float [[X:%.*]]) {
+; CHECK-NEXT: ret i1 false
+;
+ %cmp1 = fcmp ogt float %x, 0x3C00000000000000
+ %cmp2 = fcmp olt float %x, 0xBC00000000000000
+ %cond = and i1 %cmp1, %cmp2
+ ret i1 %cond
+}
+define i1 @test_or_olt(float %x) {
+; CHECK-LABEL: define i1 @test_or_olt(
+; CHECK-SAME: float [[X:%.*]]) {
+; CHECK-NEXT: [[TMP1:%.*]] = call float @llvm.fabs.f32(float [[X]])
+; CHECK-NEXT: [[COND:%.*]] = fcmp ogt float [[TMP1]], 0xBC00000000000000
+; CHECK-NEXT: ret i1 [[COND]]
+;
+ %cmp1 = fcmp olt float %x, 0x3C00000000000000
+ %cmp2 = fcmp ogt float %x, 0xBC00000000000000
+ %cond = or i1 %cmp1, %cmp2
+ ret i1 %cond
+}
+; Negative tests
+define i1 @test_and_olt_multiuse(float %x) {
+; CHECK-LABEL: define i1 @test_and_olt_multiuse(
+; CHECK-SAME: float [[X:%.*]]) {
+; CHECK-NEXT: [[CMP1:%.*]] = fcmp olt float [[X]], 0x3C00000000000000
+; CHECK-NEXT: call void @use(i1 [[CMP1]])
+; CHECK-NEXT: [[CMP2:%.*]] = fcmp ogt float [[X]], 0xBC00000000000000
+; CHECK-NEXT: [[COND:%.*]] = and i1 [[CMP1]], [[CMP2]]
+; CHECK-NEXT: ret i1 [[COND]]
+;
+ %cmp1 = fcmp olt float %x, 0x3C00000000000000
+ call void @use(i1 %cmp1)
+ %cmp2 = fcmp ogt float %x, 0xBC00000000000000
+ %cond = and i1 %cmp1, %cmp2
+ ret i1 %cond
+}
+define i1 @test_and_olt_mismatched_lhs(float %x, float %y) {
+; CHECK-LABEL: define i1 @test_and_olt_mismatched_lhs(
+; CHECK-SAME: float [[X:%.*]], float [[Y:%.*]]) {
+; CHECK-NEXT: [[CMP1:%.*]] = fcmp olt float [[X]], 0x3C00000000000000
+; CHECK-NEXT: [[CMP2:%.*]] = fcmp ogt float [[Y]], 0xBC00000000000000
+; CHECK-NEXT: [[COND:%.*]] = and i1 [[CMP1]], [[CMP2]]
+; CHECK-NEXT: ret i1 [[COND]]
+;
+ %cmp1 = fcmp olt float %x, 0x3C00000000000000
+ %cmp2 = fcmp ogt float %y, 0xBC00000000000000
+ %cond = and i1 %cmp1, %cmp2
+ ret i1 %cond
+}
+define i1 @test_and_olt_same_sign(float %x) {
+; CHECK-LABEL: define i1 @test_and_olt_same_sign(
+; CHECK-SAME: float [[X:%.*]]) {
+; CHECK-NEXT: ret i1 false
+;
+ %cmp1 = fcmp olt float %x, 0x3C00000000000000
+ %cmp2 = fcmp ogt float %x, 0x3C00000000000000
+ %cond = and i1 %cmp1, %cmp2
+ ret i1 %cond
+}
+define i1 @test_and_olt_mismatched_mag(float %x) {
+; CHECK-LABEL: define i1 @test_and_olt_mismatched_mag(
+; CHECK-SAME: float [[X:%.*]]) {
+; CHECK-NEXT: [[CMP1:%.*]] = fcmp olt float [[X]], 0x3C80000000000000
+; CHECK-NEXT: [[CMP2:%.*]] = fcmp ogt float [[X]], 0xBC00000000000000
+; CHECK-NEXT: [[COND:%.*]] = and i1 [[CMP1]], [[CMP2]]
+; CHECK-NEXT: ret i1 [[COND]]
+;
+ %cmp1 = fcmp olt float %x, 0x3C80000000000000
+ %cmp2 = fcmp ogt float %x, 0xBC00000000000000
+ %cond = and i1 %cmp1, %cmp2
+ ret i1 %cond
+}
+define i1 @test_and_olt_wrong_pred2(float %x) {
+; CHECK-LABEL: define i1 @test_and_olt_wrong_pred2(
+; CHECK-SAME: float [[X:%.*]]) {
+; CHECK-NEXT: [[CMP1:%.*]] = fcmp olt float [[X]], 0x3C00000000000000
+; CHECK-NEXT: [[CMP2:%.*]] = fcmp oge float [[X]], 0xBC00000000000000
+; CHECK-NEXT: [[COND:%.*]] = and i1 [[CMP1]], [[CMP2]]
+; CHECK-NEXT: ret i1 [[COND]]
+;
+ %cmp1 = fcmp olt float %x, 0x3C00000000000000
+ %cmp2 = fcmp oge float %x, 0xBC00000000000000
+ %cond = and i1 %cmp1, %cmp2
+ ret i1 %cond
+}
``````````
</details>
https://github.com/llvm/llvm-project/pull/76367
More information about the llvm-commits
mailing list