[llvm] [InstCombine] Support gep nuw in icmp folds (PR #118472)
Nikita Popov via llvm-commits
llvm-commits at lists.llvm.org
Tue Dec 3 03:32:17 PST 2024
https://github.com/nikic created https://github.com/llvm/llvm-project/pull/118472
Unsigned icmp of gep nuw folds to unsigned icmp of offsets. Unsigned icmp of gep nusw nuw folds to unsigned samesign icmp of offsets.
Proofs: https://alive2.llvm.org/ce/z/VEwQY8
>From bbd4be4a6c50706c00e01d8878ceed65bd46cec8 Mon Sep 17 00:00:00 2001
From: Nikita Popov <npopov at redhat.com>
Date: Tue, 3 Dec 2024 12:30:33 +0100
Subject: [PATCH] [InstCombine] Support gep nuw in icmp folds
Unsigned icmp of gep nuw folds to unsigned icmp of offsets.
Unsigned icmp of gep nusw nuw folds to unsigned samesign icmp of
offsets.
Proofs: https://alive2.llvm.org/ce/z/VEwQY8
---
.../InstCombine/InstCombineCompares.cpp | 36 ++++++---
llvm/test/Transforms/InstCombine/icmp-gep.ll | 75 +++++++++++++++++++
2 files changed, 102 insertions(+), 9 deletions(-)
diff --git a/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp b/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
index 579214c28fc304..ffc0b33171b8fc 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
@@ -690,13 +690,32 @@ Instruction *InstCombinerImpl::foldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
if (!isa<GetElementPtrInst>(RHS))
RHS = RHS->stripPointerCasts();
+ auto CanFold = [Cond](GEPNoWrapFlags NW) {
+ if (ICmpInst::isEquality(Cond))
+ return true;
+
+ // Unsigned predicates can be folded if the GEPs have *any* nowrap flags.
+ assert(ICmpInst::isUnsigned(Cond));
+ return NW != GEPNoWrapFlags::none();
+ };
+
+ auto NewICmp = [Cond](GEPNoWrapFlags NW, Value *Op1, Value *Op2) {
+ if (!NW.hasNoUnsignedWrap()) {
+ // Convert signed to unsigned comparison.
+ return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Op1, Op2);
+ }
+
+ auto *I = new ICmpInst(Cond, Op1, Op2);
+ I->setSameSign(NW.hasNoUnsignedSignedWrap());
+ return I;
+ };
+
Value *PtrBase = GEPLHS->getOperand(0);
- if (PtrBase == RHS &&
- (GEPLHS->hasNoUnsignedSignedWrap() || ICmpInst::isEquality(Cond))) {
+ if (PtrBase == RHS && CanFold(GEPLHS->getNoWrapFlags())) {
// ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
Value *Offset = EmitGEPOffset(GEPLHS);
- return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
- Constant::getNullValue(Offset->getType()));
+ return NewICmp(GEPLHS->getNoWrapFlags(), Offset,
+ Constant::getNullValue(Offset->getType()));
}
if (GEPLHS->isInBounds() && ICmpInst::isEquality(Cond) &&
@@ -814,19 +833,18 @@ Instruction *InstCombinerImpl::foldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
return replaceInstUsesWith(I, // No comparison is needed here.
ConstantInt::get(I.getType(), ICmpInst::isTrueWhenEqual(Cond)));
- else if (NumDifferences == 1 && NW.hasNoUnsignedSignedWrap()) {
+ else if (NumDifferences == 1 && CanFold(NW)) {
Value *LHSV = GEPLHS->getOperand(DiffOperand);
Value *RHSV = GEPRHS->getOperand(DiffOperand);
- // Make sure we do a signed comparison here.
- return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
+ return NewICmp(NW, LHSV, RHSV);
}
}
- if (NW.hasNoUnsignedSignedWrap() || CmpInst::isEquality(Cond)) {
+ if (CanFold(NW)) {
// ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
Value *L = EmitGEPOffset(GEPLHS, /*RewriteGEP=*/true);
Value *R = EmitGEPOffset(GEPRHS, /*RewriteGEP=*/true);
- return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
+ return NewICmp(NW, L, R);
}
}
diff --git a/llvm/test/Transforms/InstCombine/icmp-gep.ll b/llvm/test/Transforms/InstCombine/icmp-gep.ll
index 1545d034b2ac31..1bc000cd6ebf1e 100644
--- a/llvm/test/Transforms/InstCombine/icmp-gep.ll
+++ b/llvm/test/Transforms/InstCombine/icmp-gep.ll
@@ -143,6 +143,44 @@ define i1 @ult_base_nusw(ptr %x, i64 %y) {
ret i1 %r
}
+define i1 @ugt_base_nuw(ptr %x, i64 %y) {
+; CHECK-LABEL: @ugt_base_nuw(
+; CHECK-NEXT: [[R:%.*]] = icmp ne i64 [[Y:%.*]], 0
+; CHECK-NEXT: ret i1 [[R]]
+;
+ %g = getelementptr nuw i8, ptr %x, i64 %y
+ %r = icmp ugt ptr %g, %x
+ ret i1 %r
+}
+
+define i1 @ugt_base_nusw_nuw(ptr %x, i64 %y) {
+; CHECK-LABEL: @ugt_base_nusw_nuw(
+; CHECK-NEXT: [[R:%.*]] = icmp ne i64 [[Y:%.*]], 0
+; CHECK-NEXT: ret i1 [[R]]
+;
+ %g = getelementptr nusw nuw i8, ptr %x, i64 %y
+ %r = icmp ugt ptr %g, %x
+ ret i1 %r
+}
+
+define i1 @uge_base_nuw(ptr %x, i64 %y) {
+; CHECK-LABEL: @uge_base_nuw(
+; CHECK-NEXT: ret i1 true
+;
+ %g = getelementptr nuw i8, ptr %x, i64 %y
+ %r = icmp uge ptr %g, %x
+ ret i1 %r
+}
+
+define i1 @uge_base_nusw_nuw(ptr %x, i64 %y) {
+; CHECK-LABEL: @uge_base_nusw_nuw(
+; CHECK-NEXT: ret i1 true
+;
+ %g = getelementptr nusw nuw i8, ptr %x, i64 %y
+ %r = icmp uge ptr %g, %x
+ ret i1 %r
+}
+
define i1 @ugt_base_inbounds_commute(i64 %y) {
; CHECK-LABEL: @ugt_base_inbounds_commute(
; CHECK-NEXT: [[X:%.*]] = call ptr @getptr()
@@ -319,6 +357,43 @@ define i1 @test60_nusw_inbounds(ptr %foo, i64 %i, i64 %j) {
ret i1 %cmp
}
+define i1 @test60_nuw(ptr %foo, i64 %i, i64 %j) {
+; CHECK-LABEL: @test60_nuw(
+; CHECK-NEXT: [[GEP1_IDX:%.*]] = shl nuw i64 [[I:%.*]], 2
+; CHECK-NEXT: [[CMP:%.*]] = icmp ult i64 [[GEP1_IDX]], [[J:%.*]]
+; CHECK-NEXT: ret i1 [[CMP]]
+;
+ %gep1 = getelementptr nuw i32, ptr %foo, i64 %i
+ %gep2 = getelementptr nuw i8, ptr %foo, i64 %j
+ %cmp = icmp ult ptr %gep1, %gep2
+ ret i1 %cmp
+}
+
+define i1 @test60_nusw_nuw(ptr %foo, i64 %i, i64 %j) {
+; CHECK-LABEL: @test60_nusw_nuw(
+; CHECK-NEXT: [[GEP1_IDX:%.*]] = shl nuw nsw i64 [[I:%.*]], 2
+; CHECK-NEXT: [[CMP:%.*]] = icmp samesign ult i64 [[GEP1_IDX]], [[J:%.*]]
+; CHECK-NEXT: ret i1 [[CMP]]
+;
+ %gep1 = getelementptr nusw nuw i32, ptr %foo, i64 %i
+ %gep2 = getelementptr nusw nuw i8, ptr %foo, i64 %j
+ %cmp = icmp ult ptr %gep1, %gep2
+ ret i1 %cmp
+}
+
+define i1 @test60_nusw_nuw_mix(ptr %foo, i64 %i, i64 %j) {
+; CHECK-LABEL: @test60_nusw_nuw_mix(
+; CHECK-NEXT: [[GEP1:%.*]] = getelementptr nuw i32, ptr [[FOO:%.*]], i64 [[I:%.*]]
+; CHECK-NEXT: [[GEP2:%.*]] = getelementptr nusw i8, ptr [[FOO]], i64 [[J:%.*]]
+; CHECK-NEXT: [[CMP:%.*]] = icmp ult ptr [[GEP1]], [[GEP2]]
+; CHECK-NEXT: ret i1 [[CMP]]
+;
+ %gep1 = getelementptr nuw i32, ptr %foo, i64 %i
+ %gep2 = getelementptr nusw i8, ptr %foo, i64 %j
+ %cmp = icmp ult ptr %gep1, %gep2
+ ret i1 %cmp
+}
+
define i1 @test_gep_ult_no_inbounds(ptr %foo, i64 %i, i64 %j) {
; CHECK-LABEL: @test_gep_ult_no_inbounds(
; CHECK-NEXT: [[GEP1:%.*]] = getelementptr i32, ptr [[FOO:%.*]], i64 [[I:%.*]]
More information about the llvm-commits
mailing list