[InstCombine] Optimize redundant floating point comparisons in `or`/`and` inst's #158097

Rajveer100 · 2025-09-11T15:26:54Z

We can eliminate one of the fcmp when we have two same olt or ogt instructions matched in or/and simplification.

llvmbot · 2025-09-11T15:27:27Z

@llvm/pr-subscribers-llvm-analysis

@llvm/pr-subscribers-llvm-transforms

Author: Rajveer Singh Bharadwaj (Rajveer100)

Changes

Resolves #157371

We can eliminate one of the fcmp when we have two same olt or ogt instructions matched in or simplification.

Full diff: https://github.com/llvm/llvm-project/pull/158097.diff

3 Files Affected:

(modified) llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp (+24)
(modified) llvm/test/Transforms/InstCombine/or-fcmp.ll (+60)
(added) llvm/test/Transforms/InstCombine/redundant-fcmp.ll (+54)

diff --git a/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp b/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp
index 6e46898634070..0719932e1394d 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp
@@ -4514,6 +4514,30 @@ Instruction *InstCombinerImpl::visitOr(BinaryOperator &I) {
     if (Value *V = SimplifyAddWithRemainder(I))
       return replaceInstUsesWith(I, V);
 
+  Value *V0, *V1;
+  const APFloat *V0Op1, *V1Op1;
+  if (match(Op0, m_SpecificFCmp(FCmpInst::FCMP_OLT, m_Value(V0),
+                                m_APFloat(V0Op1))) &&
+      match(Op1, m_SpecificFCmp(FCmpInst::FCMP_OLT, m_Value(V1),
+                                m_APFloat(V1Op1)))) {
+    if (V0 == V1) {
+      if (V0Op1 > V1Op1)
+        replaceInstUsesWith(I, Op0);
+      else if (V1Op1 > V0Op1)
+        replaceInstUsesWith(I, Op1);
+    }
+  } else if (match(Op0, m_SpecificFCmp(FCmpInst::FCMP_OGT, m_Value(V0),
+                                m_APFloat(V0Op1))) &&
+      match(Op1, m_SpecificFCmp(FCmpInst::FCMP_OGT, m_Value(V1),
+                                m_APFloat(V1Op1)))) {
+    if (V0 == V1) {
+      if (V0Op1 < V1Op1)
+        replaceInstUsesWith(I, Op1);
+      else if (V1Op1 < V0Op1)
+        replaceInstUsesWith(I, Op0);
+    }
+  }
+
   return nullptr;
 }
 
diff --git a/llvm/test/Transforms/InstCombine/or-fcmp.ll b/llvm/test/Transforms/InstCombine/or-fcmp.ll
index 193fe4b5cc722..926f66f622914 100644
--- a/llvm/test/Transforms/InstCombine/or-fcmp.ll
+++ b/llvm/test/Transforms/InstCombine/or-fcmp.ll
@@ -4657,3 +4657,63 @@ define i1 @or_fcmp_reassoc4(i1 %x, double %a, double %b) {
   %retval = or i1 %cmp1, %or
   ret i1 %retval
 }
+
+define i1 @or_fcmp_redundant1(ptr %arg0) {
+; CHECK-LABEL: @or_fcmp_redundant1(
+; CHECK-NEXT:    [[V0:%.*]] = load double, ptr [[ARG0:%.*]], align 8
+; CHECK-NEXT:    [[V1:%.*]] = fcmp nsz olt double [[V0]], 1.000000e-02
+; CHECK-NEXT:    [[V2:%.*]] = fcmp nsz olt double [[V0]], 1.990000e+00
+; CHECK-NEXT:    [[V3:%.*]] = or i1 [[V1]], [[V2]]
+; CHECK-NEXT:    ret i1 [[V2]]
+;
+  %v0 = load double, ptr %arg0, align 8
+  %v1 = fcmp nsz olt double %v0, 1.000000e-02
+  %v2 = fcmp nsz olt double %v0, 1.990000e+00
+  %v3 = or i1 %v1, %v2
+  ret i1 %v3
+}
+
+define i1 @or_fcmp_redundant2(ptr %arg0) {
+; CHECK-LABEL: @or_fcmp_redundant2(
+; CHECK-NEXT:    [[V0:%.*]] = load double, ptr [[ARG0:%.*]], align 8
+; CHECK-NEXT:    [[V1:%.*]] = fcmp nsz olt double [[V0]], 2.300000e+00
+; CHECK-NEXT:    [[V2:%.*]] = fcmp nsz olt double [[V0]], 1.990000e+00
+; CHECK-NEXT:    [[V3:%.*]] = or i1 [[V1]], [[V2]]
+; CHECK-NEXT:    ret i1 [[V1]]
+;
+  %v0 = load double, ptr %arg0, align 8
+  %v1 = fcmp nsz olt double %v0, 2.300000e+00
+  %v2 = fcmp nsz olt double %v0, 1.990000e+00
+  %v3 = or i1 %v1, %v2
+  ret i1 %v3
+}
+
+define i1 @or_fcmp_redundant3(ptr %arg0) {
+; CHECK-LABEL: @or_fcmp_redundant3(
+; CHECK-NEXT:    [[V0:%.*]] = load double, ptr [[ARG0:%.*]], align 8
+; CHECK-NEXT:    [[V1:%.*]] = fcmp nsz ogt double [[V0]], 1.000000e-02
+; CHECK-NEXT:    [[V2:%.*]] = fcmp nsz olt double [[V0]], 1.990000e+00
+; CHECK-NEXT:    [[V3:%.*]] = or i1 [[V1]], [[V2]]
+; CHECK-NEXT:    ret i1 [[V3]]
+;
+  %v0 = load double, ptr %arg0, align 8
+  %v1 = fcmp nsz ogt double %v0, 1.000000e-02
+  %v2 = fcmp nsz olt double %v0, 1.990000e+00
+  %v3 = or i1 %v1, %v2
+  ret i1 %v3
+}
+
+define i1 @or_fcmp_redundant4(ptr %arg0) {
+; CHECK-LABEL: @or_fcmp_redundant4(
+; CHECK-NEXT:    [[V0:%.*]] = load double, ptr [[ARG0:%.*]], align 8
+; CHECK-NEXT:    [[V1:%.*]] = fcmp nsz ogt double [[V0]], 2.300000e+00
+; CHECK-NEXT:    [[V2:%.*]] = fcmp nsz ogt double [[V0]], 1.990000e+00
+; CHECK-NEXT:    [[V3:%.*]] = or i1 [[V1]], [[V2]]
+; CHECK-NEXT:    ret i1 [[V1]]
+;
+  %v0 = load double, ptr %arg0, align 8
+  %v1 = fcmp nsz ogt double %v0, 2.300000e+00
+  %v2 = fcmp nsz ogt double %v0, 1.990000e+00
+  %v3 = or i1 %v1, %v2
+  ret i1 %v3
+}
diff --git a/llvm/test/Transforms/InstCombine/redundant-fcmp.ll b/llvm/test/Transforms/InstCombine/redundant-fcmp.ll
new file mode 100644
index 0000000000000..934d9e6c46507
--- /dev/null
+++ b/llvm/test/Transforms/InstCombine/redundant-fcmp.ll
@@ -0,0 +1,54 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt < %s -O3 -S | FileCheck %s
+
+define i1 @or_fcmp_redundant1(ptr %arg0) {
+; CHECK-LABEL: @or_fcmp_redundant1(
+; CHECK-NEXT:    [[V0:%.*]] = load double, ptr [[ARG0:%.*]], align 8
+; CHECK-NEXT:    [[V2:%.*]] = fcmp nsz olt double [[V0]], 1.990000e+00
+; CHECK-NEXT:    ret i1 [[V2]]
+;
+  %v0 = load double, ptr %arg0, align 8
+  %v1 = fcmp nsz olt double %v0, 1.000000e-02
+  %v2 = fcmp nsz olt double %v0, 1.990000e+00
+  %v3 = or i1 %v1, %v2
+  ret i1 %v3
+}
+
+define i1 @or_fcmp_redundant2(ptr %arg0) {
+; CHECK-LABEL: @or_fcmp_redundant2(
+; CHECK-NEXT:    [[V0:%.*]] = load double, ptr [[ARG0:%.*]], align 8
+; CHECK-NEXT:    [[V1:%.*]] = fcmp nsz olt double [[V0]], 2.300000e+00
+; CHECK-NEXT:    ret i1 [[V1]]
+;
+  %v0 = load double, ptr %arg0, align 8
+  %v1 = fcmp nsz olt double %v0, 2.300000e+00
+  %v2 = fcmp nsz olt double %v0, 1.990000e+00
+  %v3 = or i1 %v1, %v2
+  ret i1 %v3
+}
+
+define i1 @or_fcmp_redundant3(ptr %arg0) {
+; CHECK-LABEL: @or_fcmp_redundant3(
+; CHECK-NEXT:    [[V0:%.*]] = load double, ptr [[ARG0:%.*]], align 8
+; CHECK-NEXT:    [[V3:%.*]] = fcmp nsz ogt double [[V0]], 1.990000e+00
+; CHECK-NEXT:    ret i1 [[V3]]
+;
+  %v0 = load double, ptr %arg0, align 8
+  %v1 = fcmp nsz ogt double %v0, 1.000000e-02
+  %v2 = fcmp nsz ogt double %v0, 1.990000e+00
+  %v3 = or i1 %v1, %v2
+  ret i1 %v3
+}
+
+define i1 @or_fcmp_redundant4(ptr %arg0) {
+; CHECK-LABEL: @or_fcmp_redundant4(
+; CHECK-NEXT:    [[V0:%.*]] = load double, ptr [[ARG0:%.*]], align 8
+; CHECK-NEXT:    [[V1:%.*]] = fcmp nsz ogt double [[V0]], 2.300000e+00
+; CHECK-NEXT:    ret i1 [[V1]]
+;
+  %v0 = load double, ptr %arg0, align 8
+  %v1 = fcmp nsz ogt double %v0, 2.300000e+00
+  %v2 = fcmp nsz ogt double %v0, 1.990000e+00
+  %v3 = or i1 %v1, %v2
+  ret i1 %v3
+}

dtcxzyw · 2025-09-11T15:30:30Z

@zyw-bot mfuzz

llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp

XChy

Should check !isAnd if you only fold the pattern or(fcmp, fcmp).

llvm/lib/Analysis/InstructionSimplify.cpp

llvm/test/Transforms/InstCombine/or-fcmp.ll

XChy · 2025-09-13T08:56:48Z

General proof: proof

Rajveer100 · 2025-09-13T11:03:09Z

Should check !isAnd if you only fold the pattern or(fcmp, fcmp).

Yeah, for the and case we just return the other op.

github-actions · 2025-09-13T11:33:17Z

✅ With the latest revision this PR passed the C/C++ code formatter.

…and` inst's Resolves llvm#157371 We can eliminate one of the `fcmp` when we have two same `olt` or `ogt` instructions matched in `or`/`and` simplification.

Rajveer100 requested a review from dtcxzyw September 11, 2025 15:26

Rajveer100 requested a review from nikic as a code owner September 11, 2025 15:26

llvmbot added llvm:instcombine Covers the InstCombine, InstSimplify and AggressiveInstCombine passes llvm:transforms labels Sep 11, 2025

Rajveer100 force-pushed the opt-redundant-fcmp branch from 45430d5 to 7a8a3b0 Compare September 11, 2025 15:27

Rajveer100 added the floating-point Floating-point math label Sep 11, 2025

zyw-bot mentioned this pull request Sep 11, 2025

Fuzz PR158097 dtcxzyw/llvm-mutation-based-fuzz-service#98

Open

dtcxzyw requested changes Sep 11, 2025

View reviewed changes

Rajveer100 force-pushed the opt-redundant-fcmp branch from 7a8a3b0 to 28d9dfa Compare September 13, 2025 08:20

llvmbot added the llvm:analysis Includes value tracking, cost tables and constant folding label Sep 13, 2025

Rajveer100 requested a review from dtcxzyw September 13, 2025 08:20

XChy reviewed Sep 13, 2025

View reviewed changes

llvm/lib/Analysis/InstructionSimplify.cpp Outdated Show resolved Hide resolved

XChy reviewed Sep 13, 2025

View reviewed changes

llvm/lib/Analysis/InstructionSimplify.cpp Show resolved Hide resolved

llvm/test/Transforms/InstCombine/or-fcmp.ll Show resolved Hide resolved

Rajveer100 force-pushed the opt-redundant-fcmp branch from 28d9dfa to d219143 Compare September 13, 2025 11:20

Rajveer100 requested a review from XChy September 13, 2025 11:21

Rajveer100 force-pushed the opt-redundant-fcmp branch from d219143 to a65787b Compare September 13, 2025 11:29

Rajveer100 changed the title ~~[InstCombine] Optimize redundant floating point comparisons in or inst~~ [InstCombine] Optimize redundant floating point comparisons in or/and inst's Sep 13, 2025

[InstCombine] Optimize redundant floating point comparisons in or/`…

24257f7

…and` inst's Resolves llvm#157371 We can eliminate one of the `fcmp` when we have two same `olt` or `ogt` instructions matched in `or`/`and` simplification.

Rajveer100 force-pushed the opt-redundant-fcmp branch from a65787b to 24257f7 Compare September 13, 2025 11:35

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[InstCombine] Optimize redundant floating point comparisons in `or`/`and` inst's #158097

[InstCombine] Optimize redundant floating point comparisons in `or`/`and` inst's #158097

Rajveer100 commented Sep 11, 2025 •

edited

Loading

Uh oh!

llvmbot commented Sep 11, 2025 •

edited

Loading

Uh oh!

dtcxzyw commented Sep 11, 2025

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

XChy left a comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

XChy commented Sep 13, 2025 •

edited

Loading

Uh oh!

Rajveer100 commented Sep 13, 2025

Uh oh!

github-actions bot commented Sep 13, 2025 •

edited

Loading

Uh oh!

Uh oh!

[InstCombine] Optimize redundant floating point comparisons in or/and inst's #158097

Are you sure you want to change the base?

[InstCombine] Optimize redundant floating point comparisons in or/and inst's #158097

Conversation

Rajveer100 commented Sep 11, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

llvmbot commented Sep 11, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

dtcxzyw commented Sep 11, 2025

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

XChy left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

XChy commented Sep 13, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Rajveer100 commented Sep 13, 2025

Uh oh!

github-actions bot commented Sep 13, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

[InstCombine] Optimize redundant floating point comparisons in `or`/`and` inst's #158097

[InstCombine] Optimize redundant floating point comparisons in `or`/`and` inst's #158097

Rajveer100 commented Sep 11, 2025 •

edited

Loading

llvmbot commented Sep 11, 2025 •

edited

Loading

XChy commented Sep 13, 2025 •

edited

Loading

github-actions bot commented Sep 13, 2025 •

edited

Loading