[llvm] Instcombine always rewrite gep (PR #142787)
Nikita Popov via llvm-commits
llvm-commits at lists.llvm.org
Wed Jun 4 08:14:30 PDT 2025
https://github.com/nikic created https://github.com/llvm/llvm-project/pull/142787
Resolve the TODO in OptimizePointerDifference: Always rewrite (non-trivial, multi-use) GEPs to reuse the offset arithmetic, even if there is only one GEP involved in the subtraction.
I plan to extend this code to deeper GEP chains, in which case always rewriting seems like a good idea.
>From 6820867a3574084a2b4fef9f02e7cb2d829c8217 Mon Sep 17 00:00:00 2001
From: Nikita Popov <npopov at redhat.com>
Date: Wed, 4 Jun 2025 17:08:01 +0200
Subject: [PATCH 1/2] [InstCombine] Add test for sub gep with multi-use
---
llvm/test/Transforms/InstCombine/sub-gep.ll | 32 +++++++++++++++++++++
1 file changed, 32 insertions(+)
diff --git a/llvm/test/Transforms/InstCombine/sub-gep.ll b/llvm/test/Transforms/InstCombine/sub-gep.ll
index c86a1a37bd7ad..8c0b57228d9ea 100644
--- a/llvm/test/Transforms/InstCombine/sub-gep.ll
+++ b/llvm/test/Transforms/InstCombine/sub-gep.ll
@@ -760,6 +760,38 @@ entry:
ret i64 %ret
}
+declare void @use(ptr)
+
+define i64 @sub_multi_use(ptr %base, i64 %idx) {
+; CHECK-LABEL: @sub_multi_use(
+; CHECK-NEXT: [[P2:%.*]] = getelementptr inbounds [0 x i32], ptr [[BASE:%.*]], i64 0, i64 [[IDX:%.*]]
+; CHECK-NEXT: call void @use(ptr [[P2]])
+; CHECK-NEXT: [[P2_IDX:%.*]] = shl nsw i64 [[IDX]], 2
+; CHECK-NEXT: ret i64 [[P2_IDX]]
+;
+ %p2 = getelementptr inbounds [0 x i32], ptr %base, i64 0, i64 %idx
+ call void @use(ptr %p2)
+ %i1 = ptrtoint ptr %base to i64
+ %i2 = ptrtoint ptr %p2 to i64
+ %d = sub i64 %i2, %i1
+ ret i64 %d
+}
+
+define i64 @sub_multi_use_nuw(ptr %base, i64 %idx) {
+; CHECK-LABEL: @sub_multi_use_nuw(
+; CHECK-NEXT: [[P2:%.*]] = getelementptr inbounds [0 x i32], ptr [[BASE:%.*]], i64 0, i64 [[IDX:%.*]]
+; CHECK-NEXT: call void @use(ptr [[P2]])
+; CHECK-NEXT: [[P2_IDX:%.*]] = shl nuw nsw i64 [[IDX]], 2
+; CHECK-NEXT: ret i64 [[P2_IDX]]
+;
+ %p2 = getelementptr inbounds [0 x i32], ptr %base, i64 0, i64 %idx
+ call void @use(ptr %p2)
+ %i1 = ptrtoint ptr %base to i64
+ %i2 = ptrtoint ptr %p2 to i64
+ %d = sub nuw i64 %i2, %i1
+ ret i64 %d
+}
+
define i1 @_gep_phi1(ptr %str1) {
; CHECK-LABEL: @_gep_phi1(
; CHECK-NEXT: entry:
>From 1583757a80e948e2b12ef208c42add32dd200f18 Mon Sep 17 00:00:00 2001
From: Nikita Popov <npopov at redhat.com>
Date: Wed, 4 Jun 2025 17:10:15 +0200
Subject: [PATCH 2/2] [InstCombine] Always rewrite multi-use GEP for pointer
difference
Resolve the TODO in OptimizePointerDifference: Always rewrite
(non-trivial, multi-use) GEPs to reuse the offset arithmetic,
even if there is only one GEP.
---
llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp | 11 ++++-------
llvm/test/Transforms/InstCombine/sub-gep.ll | 8 ++++----
2 files changed, 8 insertions(+), 11 deletions(-)
diff --git a/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp b/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp
index a9ac5ff9b9c89..fda56c93451e8 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp
@@ -2101,21 +2101,18 @@ Value *InstCombinerImpl::OptimizePointerDifference(Value *LHS, Value *RHS,
if (!GEP1)
return nullptr;
+ // Emit the offset of the GEP as an intptr_t.
// To avoid duplicating the offset arithmetic, rewrite the GEP to use the
// computed offset. This may erase the original GEP, so be sure to cache the
// nowrap flags before emitting the offset.
- // TODO: We should probably do this even if there is only one GEP.
- bool RewriteGEPs = GEP2 != nullptr;
-
- // Emit the offset of the GEP and an intptr_t.
GEPNoWrapFlags GEP1NW = GEP1->getNoWrapFlags();
- Value *Result = EmitGEPOffset(GEP1, RewriteGEPs);
+ Value *Result = EmitGEPOffset(GEP1, /*RewriteGEP=*/true);
// If this is a single inbounds GEP and the original sub was nuw,
// then the final multiplication is also nuw.
if (auto *I = dyn_cast<Instruction>(Result))
if (IsNUW && !GEP2 && !Swapped && GEP1NW.isInBounds() &&
- I->getOpcode() == Instruction::Mul)
+ I->getOpcode() == Instruction::Mul && I->use_empty())
I->setHasNoUnsignedWrap();
// If we have a 2nd GEP of the same base pointer, subtract the offsets.
@@ -2123,7 +2120,7 @@ Value *InstCombinerImpl::OptimizePointerDifference(Value *LHS, Value *RHS,
// If both GEPs are nuw and the original sub is nuw, the new sub is also nuw.
if (GEP2) {
GEPNoWrapFlags GEP2NW = GEP2->getNoWrapFlags();
- Value *Offset = EmitGEPOffset(GEP2, RewriteGEPs);
+ Value *Offset = EmitGEPOffset(GEP2, /*RewriteGEP=*/true);
Result = Builder.CreateSub(Result, Offset, "gepdiff",
IsNUW && GEP1NW.hasNoUnsignedWrap() &&
GEP2NW.hasNoUnsignedWrap(),
diff --git a/llvm/test/Transforms/InstCombine/sub-gep.ll b/llvm/test/Transforms/InstCombine/sub-gep.ll
index 8c0b57228d9ea..e6f6498e23389 100644
--- a/llvm/test/Transforms/InstCombine/sub-gep.ll
+++ b/llvm/test/Transforms/InstCombine/sub-gep.ll
@@ -764,9 +764,9 @@ declare void @use(ptr)
define i64 @sub_multi_use(ptr %base, i64 %idx) {
; CHECK-LABEL: @sub_multi_use(
-; CHECK-NEXT: [[P2:%.*]] = getelementptr inbounds [0 x i32], ptr [[BASE:%.*]], i64 0, i64 [[IDX:%.*]]
+; CHECK-NEXT: [[P2_IDX:%.*]] = shl nsw i64 [[IDX:%.*]], 2
+; CHECK-NEXT: [[P2:%.*]] = getelementptr inbounds i8, ptr [[BASE:%.*]], i64 [[P2_IDX]]
; CHECK-NEXT: call void @use(ptr [[P2]])
-; CHECK-NEXT: [[P2_IDX:%.*]] = shl nsw i64 [[IDX]], 2
; CHECK-NEXT: ret i64 [[P2_IDX]]
;
%p2 = getelementptr inbounds [0 x i32], ptr %base, i64 0, i64 %idx
@@ -779,9 +779,9 @@ define i64 @sub_multi_use(ptr %base, i64 %idx) {
define i64 @sub_multi_use_nuw(ptr %base, i64 %idx) {
; CHECK-LABEL: @sub_multi_use_nuw(
-; CHECK-NEXT: [[P2:%.*]] = getelementptr inbounds [0 x i32], ptr [[BASE:%.*]], i64 0, i64 [[IDX:%.*]]
+; CHECK-NEXT: [[P2_IDX:%.*]] = shl nsw i64 [[IDX:%.*]], 2
+; CHECK-NEXT: [[P2:%.*]] = getelementptr inbounds i8, ptr [[BASE:%.*]], i64 [[P2_IDX]]
; CHECK-NEXT: call void @use(ptr [[P2]])
-; CHECK-NEXT: [[P2_IDX:%.*]] = shl nuw nsw i64 [[IDX]], 2
; CHECK-NEXT: ret i64 [[P2_IDX]]
;
%p2 = getelementptr inbounds [0 x i32], ptr %base, i64 0, i64 %idx
More information about the llvm-commits
mailing list