[llvm] [InstCombine] optimize unnecessary sext instruction with add + cmp (PR #152291)

[via llvm-commits]

llvmbot wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-llvm-transforms

Author: Gaurav Dhingra (gxyd)

<details>
<summary>Changes</summary>

Fixes https://github.com/llvm/llvm-project/issues/130088

---
Full diff: https://github.com/llvm/llvm-project/pull/152291.diff


4 Files Affected:

- (modified) llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp (+34) 
- (modified) llvm/lib/Transforms/InstCombine/InstCombineInternal.h (+1) 
- (modified) llvm/test/Transforms/InstCombine/icmp-add.ll (+2-3) 
- (added) llvm/test/Transforms/InstCombine/sext_for_add_cmp.ll (+16) 


``````````diff
diff --git a/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp b/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
index cf94d28100488..ae329e9aaec4c 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
@@ -6381,6 +6381,36 @@ Instruction *InstCombinerImpl::foldICmpWithZextOrSext(ICmpInst &ICmp) {
   return new ICmpInst(CmpInst::ICMP_SLT, X, Constant::getNullValue(SrcTy));
 }
 
+Instruction *InstCombinerImpl::foldICmpWithSextAndAdd(ICmpInst &ICmp) {
+  Value *X;
+  ConstantInt *Y, *Z;
+  // Match the pattern: icmp ult (add (sext X), Y), Z
+  // where X is a value, Y and Z are integer constants
+  // icmp ult (add(sext(X), Y)), Z  -> icmp ult (add(X, Y)), Z
+  if (match(&ICmp, m_SpecificICmp(CmpInst::ICMP_ULT,
+                            m_Add(m_SExt(m_Value(X)), m_ConstantInt(Y)),
+                            m_ConstantInt(Z)))) {
+    Type *XType = X->getType();
+    if (!XType->isIntegerTy())
+      return nullptr;
+
+    unsigned XBitWidth = XType->getIntegerBitWidth();
+    auto YValue = Y->getSExtValue();
+    auto ZValue = Z->getSExtValue();
+
+    auto MinValue = -(1LL << (XBitWidth - 1));
+    auto MaxValue = (1LL << (XBitWidth - 1)) - 1;
+
+    // // Check if Y and Z fit within X's type without wrapping
+    if (YValue < MinValue || YValue > MaxValue || ZValue < MinValue || ZValue > MaxValue)
+      return nullptr; // Cannot optimize if Y or Z would wrap in X's type
+
+    Value *NewAdd = Builder.CreateAdd(X, ConstantInt::get(XType, YValue));
+    return new ICmpInst(CmpInst::ICMP_ULT, NewAdd, ConstantInt::get(XType, ZValue));
+  }
+  return nullptr;
+}
+
 /// Handle icmp (cast x), (cast or constant).
 Instruction *InstCombinerImpl::foldICmpWithCastOp(ICmpInst &ICmp) {
   // If any operand of ICmp is a inttoptr roundtrip cast then remove it as
@@ -7727,6 +7757,10 @@ Instruction *InstCombinerImpl::visitICmpInst(ICmpInst &I) {
     }
   }
 
+  if (Instruction *Res = foldICmpWithSextAndAdd(I)) {
+    return Res;
+  }
+
   // In case of a comparison with two select instructions having the same
   // condition, check whether one of the resulting branches can be simplified.
   // If so, just compare the other branch and select the appropriate result.
diff --git a/llvm/lib/Transforms/InstCombine/InstCombineInternal.h b/llvm/lib/Transforms/InstCombine/InstCombineInternal.h
index c67e27e5b3e7c..ed652720f0dac 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineInternal.h
+++ b/llvm/lib/Transforms/InstCombine/InstCombineInternal.h
@@ -744,6 +744,7 @@ class LLVM_LIBRARY_VISIBILITY InstCombinerImpl final
                                      const APInt &C);
   Instruction *foldICmpTruncWithTruncOrExt(ICmpInst &Cmp,
                                            const SimplifyQuery &Q);
+  Instruction *foldICmpWithSextAndAdd(ICmpInst &ICmp);
   Instruction *foldICmpAndConstant(ICmpInst &Cmp, BinaryOperator *And,
                                    const APInt &C);
   Instruction *foldICmpXorConstant(ICmpInst &Cmp, BinaryOperator *Xor,
diff --git a/llvm/test/Transforms/InstCombine/icmp-add.ll b/llvm/test/Transforms/InstCombine/icmp-add.ll
index 1a41c1f3e1045..df85f5a3caa65 100644
--- a/llvm/test/Transforms/InstCombine/icmp-add.ll
+++ b/llvm/test/Transforms/InstCombine/icmp-add.ll
@@ -3259,9 +3259,8 @@ define i1 @zext_range_check_mergable(i8 %x) {
 define i1 @sext_range_check_ult(i8 %x) {
 ; CHECK-LABEL: @sext_range_check_ult(
 ; CHECK-NEXT:  entry:
-; CHECK-NEXT:    [[CONV:%.*]] = sext i8 [[X:%.*]] to i32
-; CHECK-NEXT:    [[ADD:%.*]] = add nsw i32 [[CONV]], -4
-; CHECK-NEXT:    [[CMP:%.*]] = icmp ult i32 [[ADD]], 3
+; CHECK-NEXT:    [[TMP0:%.*]] = add i8 [[X:%.*]], -4
+; CHECK-NEXT:    [[CMP:%.*]] = icmp ult i8 [[TMP0]], 3
 ; CHECK-NEXT:    ret i1 [[CMP]]
 ;
 entry:
diff --git a/llvm/test/Transforms/InstCombine/sext_for_add_cmp.ll b/llvm/test/Transforms/InstCombine/sext_for_add_cmp.ll
new file mode 100644
index 0000000000000..b8617df750abb
--- /dev/null
+++ b/llvm/test/Transforms/InstCombine/sext_for_add_cmp.ll
@@ -0,0 +1,16 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
+; RUN: opt < %s -passes=instcombine -S | FileCheck %s
+define i1 @sext_add_cmp_to_add_cmp(i16 noundef %x) {
+; CHECK-LABEL: define i1 @sext_add_cmp_to_add_cmp(
+; CHECK-SAME: i16 noundef [[X:%.*]]) {
+; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    [[TMP0:%.*]] = add i16 [[X]], 128
+; CHECK-NEXT:    [[OR_COND_I:%.*]] = icmp ult i16 [[TMP0]], 256
+; CHECK-NEXT:    ret i1 [[OR_COND_I]]
+;
+entry:
+  %conv = sext i16 %x to i128
+  %add = add i128 %conv, 128
+  %or.cond.i = icmp ult i128 %add, 256
+  ret i1 %or.cond.i
+}

``````````

</details>


https://github.com/llvm/llvm-project/pull/152291