[llvm] [InstCombine] Set flags when canonicalizing GEP indices (PR #151516)
Nikita Popov via llvm-commits
llvm-commits at lists.llvm.org
Thu Jul 31 06:17:14 PDT 2025
https://github.com/nikic created https://github.com/llvm/llvm-project/pull/151516
When truncating set nsw/nuw based on nusw/nuw. When extending, use zext nneg if nusw+nuw.
Proof: https://alive2.llvm.org/ce/z/JA2Yzr
>From 9f280d0d0f642d2a2d2c3e75e6a70c1179784845 Mon Sep 17 00:00:00 2001
From: Nikita Popov <npopov at redhat.com>
Date: Thu, 31 Jul 2025 15:07:49 +0200
Subject: [PATCH] [InstCombine] Set flags when canonicalizing GEP indices
When truncating set nsw/nuw based on nusw/nuw. When extending,
use zext nneg if nusw+nuw.
Proof: https://alive2.llvm.org/ce/z/JA2Yzr
---
.../InstCombine/InstructionCombining.cpp | 11 ++-
.../Transforms/InstCombine/getelementptr.ll | 96 ++++++++++++++++++-
.../Transforms/InstCombine/icmp-custom-dl.ll | 10 +-
llvm/test/Transforms/InstCombine/icmp-gep.ll | 7 +-
.../InstCombine/indexed-gep-compares.ll | 2 +-
llvm/test/Transforms/InstCombine/load-cmp.ll | 5 +-
llvm/test/Transforms/InstCombine/pr39908.ll | 4 +-
llvm/test/Transforms/InstCombine/sub-gep.ll | 2 +-
8 files changed, 117 insertions(+), 20 deletions(-)
diff --git a/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp b/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp
index 9e333202fb2e7..9d0fcf0d59e17 100644
--- a/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp
@@ -3182,7 +3182,16 @@ Instruction *InstCombinerImpl::visitGetElementPtrInst(GetElementPtrInst &GEP) {
// If we are using a wider index than needed for this platform, shrink
// it to what we need. If narrower, sign-extend it to what we need.
// This explicit cast can make subsequent optimizations more obvious.
- *I = Builder.CreateIntCast(*I, NewIndexType, true);
+ if (IndexTy->getScalarSizeInBits() <
+ NewIndexType->getScalarSizeInBits()) {
+ if (GEP.hasNoUnsignedWrap() && GEP.hasNoUnsignedSignedWrap())
+ *I = Builder.CreateZExt(*I, NewIndexType, "", /*IsNonNeg=*/true);
+ else
+ *I = Builder.CreateSExt(*I, NewIndexType);
+ } else {
+ *I = Builder.CreateTrunc(*I, NewIndexType, "", GEP.hasNoUnsignedWrap(),
+ GEP.hasNoUnsignedSignedWrap());
+ }
MadeChange = true;
}
}
diff --git a/llvm/test/Transforms/InstCombine/getelementptr.ll b/llvm/test/Transforms/InstCombine/getelementptr.ll
index 7c1342a7db84b..55b5b5eb1fc0c 100644
--- a/llvm/test/Transforms/InstCombine/getelementptr.ll
+++ b/llvm/test/Transforms/InstCombine/getelementptr.ll
@@ -356,7 +356,7 @@ define i1 @test13_i16(i16 %X, ptr %P) {
define i1 @test13_i128(i128 %X, ptr %P) {
; CHECK-LABEL: @test13_i128(
-; CHECK-NEXT: [[TMP1:%.*]] = trunc i128 [[X:%.*]] to i64
+; CHECK-NEXT: [[TMP1:%.*]] = trunc nsw i128 [[X:%.*]] to i64
; CHECK-NEXT: [[C:%.*]] = icmp eq i64 [[TMP1]], -1
; CHECK-NEXT: ret i1 [[C]]
;
@@ -412,7 +412,7 @@ define ptr @test_index_canon_inbounds(ptr %X, i32 %Idx) {
define ptr @test_index_canon_nusw_nuw(ptr %X, i32 %Idx) {
; CHECK-LABEL: @test_index_canon_nusw_nuw(
-; CHECK-NEXT: [[TMP1:%.*]] = sext i32 [[IDX:%.*]] to i64
+; CHECK-NEXT: [[TMP1:%.*]] = zext nneg i32 [[IDX:%.*]] to i64
; CHECK-NEXT: [[R:%.*]] = getelementptr nusw nuw i32, ptr [[X:%.*]], i64 [[TMP1]]
; CHECK-NEXT: ret ptr [[R]]
;
@@ -568,7 +568,7 @@ define i32 @test20(ptr %P, i32 %A, i32 %B) {
define i32 @test20_as1(ptr addrspace(1) %P, i32 %A, i32 %B) {
; CHECK-LABEL: @test20_as1(
-; CHECK-NEXT: [[TMP1:%.*]] = trunc i32 [[A:%.*]] to i16
+; CHECK-NEXT: [[TMP1:%.*]] = trunc nsw i32 [[A:%.*]] to i16
; CHECK-NEXT: [[T6:%.*]] = icmp eq i16 [[TMP1]], 0
; CHECK-NEXT: [[T7:%.*]] = zext i1 [[T6]] to i32
; CHECK-NEXT: ret i32 [[T7]]
@@ -1978,4 +1978,94 @@ define ptr @gep_merge_nusw_const(ptr %p, i64 %idx, i64 %idx2) {
ret ptr %gep
}
+define ptr @gep_index_trunc_nothing(ptr %p, i128 %idx) {
+; CHECK-LABEL: @gep_index_trunc_nothing(
+; CHECK-NEXT: [[TMP1:%.*]] = trunc i128 [[IDX:%.*]] to i64
+; CHECK-NEXT: [[GEP:%.*]] = getelementptr i8, ptr [[P:%.*]], i64 [[TMP1]]
+; CHECK-NEXT: ret ptr [[GEP]]
+;
+ %gep = getelementptr i8, ptr %p, i128 %idx
+ ret ptr %gep
+}
+
+define ptr @gep_index_trunc_nuw(ptr %p, i128 %idx) {
+; CHECK-LABEL: @gep_index_trunc_nuw(
+; CHECK-NEXT: [[TMP1:%.*]] = trunc nuw i128 [[IDX:%.*]] to i64
+; CHECK-NEXT: [[GEP:%.*]] = getelementptr nuw i8, ptr [[P:%.*]], i64 [[TMP1]]
+; CHECK-NEXT: ret ptr [[GEP]]
+;
+ %gep = getelementptr nuw i8, ptr %p, i128 %idx
+ ret ptr %gep
+}
+
+define ptr @gep_index_trunc_nusw(ptr %p, i128 %idx) {
+; CHECK-LABEL: @gep_index_trunc_nusw(
+; CHECK-NEXT: [[TMP1:%.*]] = trunc nsw i128 [[IDX:%.*]] to i64
+; CHECK-NEXT: [[GEP:%.*]] = getelementptr nusw i8, ptr [[P:%.*]], i64 [[TMP1]]
+; CHECK-NEXT: ret ptr [[GEP]]
+;
+ %gep = getelementptr nusw i8, ptr %p, i128 %idx
+ ret ptr %gep
+}
+
+define ptr @gep_index_trunc_inbounds(ptr %p, i128 %idx) {
+; CHECK-LABEL: @gep_index_trunc_inbounds(
+; CHECK-NEXT: [[TMP1:%.*]] = trunc nsw i128 [[IDX:%.*]] to i64
+; CHECK-NEXT: [[GEP:%.*]] = getelementptr inbounds i8, ptr [[P:%.*]], i64 [[TMP1]]
+; CHECK-NEXT: ret ptr [[GEP]]
+;
+ %gep = getelementptr inbounds i8, ptr %p, i128 %idx
+ ret ptr %gep
+}
+
+define ptr @gep_index_trunc_nusw_nuw(ptr %p, i128 %idx) {
+; CHECK-LABEL: @gep_index_trunc_nusw_nuw(
+; CHECK-NEXT: [[TMP1:%.*]] = trunc nuw nsw i128 [[IDX:%.*]] to i64
+; CHECK-NEXT: [[GEP:%.*]] = getelementptr nusw nuw i8, ptr [[P:%.*]], i64 [[TMP1]]
+; CHECK-NEXT: ret ptr [[GEP]]
+;
+ %gep = getelementptr nusw nuw i8, ptr %p, i128 %idx
+ ret ptr %gep
+}
+
+define ptr @gep_index_ext_nothing(ptr %p, i32 %idx) {
+; CHECK-LABEL: @gep_index_ext_nothing(
+; CHECK-NEXT: [[TMP1:%.*]] = sext i32 [[IDX:%.*]] to i64
+; CHECK-NEXT: [[GEP:%.*]] = getelementptr i8, ptr [[P:%.*]], i64 [[TMP1]]
+; CHECK-NEXT: ret ptr [[GEP]]
+;
+ %gep = getelementptr i8, ptr %p, i32 %idx
+ ret ptr %gep
+}
+
+define ptr @gep_index_ext_nuw(ptr %p, i32 %idx) {
+; CHECK-LABEL: @gep_index_ext_nuw(
+; CHECK-NEXT: [[TMP1:%.*]] = sext i32 [[IDX:%.*]] to i64
+; CHECK-NEXT: [[GEP:%.*]] = getelementptr nuw i8, ptr [[P:%.*]], i64 [[TMP1]]
+; CHECK-NEXT: ret ptr [[GEP]]
+;
+ %gep = getelementptr nuw i8, ptr %p, i32 %idx
+ ret ptr %gep
+}
+
+define ptr @gep_index_ext_nusw(ptr %p, i32 %idx) {
+; CHECK-LABEL: @gep_index_ext_nusw(
+; CHECK-NEXT: [[TMP1:%.*]] = sext i32 [[IDX:%.*]] to i64
+; CHECK-NEXT: [[GEP:%.*]] = getelementptr nuw i8, ptr [[P:%.*]], i64 [[TMP1]]
+; CHECK-NEXT: ret ptr [[GEP]]
+;
+ %gep = getelementptr nuw i8, ptr %p, i32 %idx
+ ret ptr %gep
+}
+
+define ptr @gep_index_ext_nusw_nuw(ptr %p, i32 %idx) {
+; CHECK-LABEL: @gep_index_ext_nusw_nuw(
+; CHECK-NEXT: [[TMP1:%.*]] = zext nneg i32 [[IDX:%.*]] to i64
+; CHECK-NEXT: [[GEP:%.*]] = getelementptr nusw nuw i8, ptr [[P:%.*]], i64 [[TMP1]]
+; CHECK-NEXT: ret ptr [[GEP]]
+;
+ %gep = getelementptr nusw nuw i8, ptr %p, i32 %idx
+ ret ptr %gep
+}
+
!0 = !{!"branch_weights", i32 2, i32 10}
diff --git a/llvm/test/Transforms/InstCombine/icmp-custom-dl.ll b/llvm/test/Transforms/InstCombine/icmp-custom-dl.ll
index 1296dc6a33b5d..f873551512718 100644
--- a/llvm/test/Transforms/InstCombine/icmp-custom-dl.ll
+++ b/llvm/test/Transforms/InstCombine/icmp-custom-dl.ll
@@ -39,9 +39,9 @@ define i1 @test59_as1(ptr addrspace(1) %foo) {
define i1 @test60(ptr %foo, i64 %i, i64 %j) {
; CHECK-LABEL: @test60(
-; CHECK-NEXT: [[TMP1:%.*]] = trunc i64 [[I:%.*]] to i32
+; CHECK-NEXT: [[TMP1:%.*]] = trunc nsw i64 [[I:%.*]] to i32
; CHECK-NEXT: [[GEP1_IDX:%.*]] = shl nsw i32 [[TMP1]], 2
-; CHECK-NEXT: [[TMP2:%.*]] = trunc i64 [[J:%.*]] to i32
+; CHECK-NEXT: [[TMP2:%.*]] = trunc nsw i64 [[J:%.*]] to i32
; CHECK-NEXT: [[CMP:%.*]] = icmp slt i32 [[GEP1_IDX]], [[TMP2]]
; CHECK-NEXT: ret i1 [[CMP]]
;
@@ -53,9 +53,9 @@ define i1 @test60(ptr %foo, i64 %i, i64 %j) {
define i1 @test60_as1(ptr addrspace(1) %foo, i64 %i, i64 %j) {
; CHECK-LABEL: @test60_as1(
-; CHECK-NEXT: [[TMP1:%.*]] = trunc i64 [[I:%.*]] to i16
+; CHECK-NEXT: [[TMP1:%.*]] = trunc nsw i64 [[I:%.*]] to i16
; CHECK-NEXT: [[GEP1_IDX:%.*]] = shl nsw i16 [[TMP1]], 2
-; CHECK-NEXT: [[TMP2:%.*]] = trunc i64 [[J:%.*]] to i16
+; CHECK-NEXT: [[TMP2:%.*]] = trunc nsw i64 [[J:%.*]] to i16
; CHECK-NEXT: [[CMP:%.*]] = icmp slt i16 [[GEP1_IDX]], [[TMP2]]
; CHECK-NEXT: ret i1 [[CMP]]
;
@@ -69,7 +69,7 @@ define i1 @test60_as1(ptr addrspace(1) %foo, i64 %i, i64 %j) {
; bitcast. This uses the same sized addrspace.
define i1 @test60_addrspacecast(ptr %foo, i64 %i, i64 %j) {
; CHECK-LABEL: @test60_addrspacecast(
-; CHECK-NEXT: [[TMP1:%.*]] = trunc i64 [[J:%.*]] to i32
+; CHECK-NEXT: [[TMP1:%.*]] = trunc nsw i64 [[J:%.*]] to i32
; CHECK-NEXT: [[I_TR:%.*]] = trunc i64 [[I:%.*]] to i32
; CHECK-NEXT: [[TMP2:%.*]] = shl i32 [[I_TR]], 2
; CHECK-NEXT: [[CMP:%.*]] = icmp slt i32 [[TMP2]], [[TMP1]]
diff --git a/llvm/test/Transforms/InstCombine/icmp-gep.ll b/llvm/test/Transforms/InstCombine/icmp-gep.ll
index 5044850ec918f..1385dc3f625f1 100644
--- a/llvm/test/Transforms/InstCombine/icmp-gep.ll
+++ b/llvm/test/Transforms/InstCombine/icmp-gep.ll
@@ -286,8 +286,7 @@ define i1 @PR8882(i64 %i) {
define i1 @test24_as1(i64 %i) {
; CHECK-LABEL: @test24_as1(
-; CHECK-NEXT: [[TMP1:%.*]] = and i64 [[I:%.*]], 65535
-; CHECK-NEXT: [[CMP:%.*]] = icmp eq i64 [[TMP1]], 1000
+; CHECK-NEXT: [[CMP:%.*]] = icmp eq i64 [[TMP1:%.*]], 1000
; CHECK-NEXT: ret i1 [[CMP]]
;
%p1 = getelementptr inbounds i32, ptr addrspace(1) @X_as1, i64 %i
@@ -449,9 +448,9 @@ define i1 @test_gep_eq_no_inbounds(ptr %foo, i64 %i, i64 %j) {
define i1 @test60_as1(ptr addrspace(1) %foo, i64 %i, i64 %j) {
; CHECK-LABEL: @test60_as1(
-; CHECK-NEXT: [[TMP1:%.*]] = trunc i64 [[I:%.*]] to i16
+; CHECK-NEXT: [[TMP1:%.*]] = trunc nsw i64 [[I:%.*]] to i16
; CHECK-NEXT: [[GEP1_IDX:%.*]] = shl nsw i16 [[TMP1]], 2
-; CHECK-NEXT: [[TMP2:%.*]] = trunc i64 [[J:%.*]] to i16
+; CHECK-NEXT: [[TMP2:%.*]] = trunc nsw i64 [[J:%.*]] to i16
; CHECK-NEXT: [[CMP:%.*]] = icmp slt i16 [[GEP1_IDX]], [[TMP2]]
; CHECK-NEXT: ret i1 [[CMP]]
;
diff --git a/llvm/test/Transforms/InstCombine/indexed-gep-compares.ll b/llvm/test/Transforms/InstCombine/indexed-gep-compares.ll
index 53c9736da59c4..07486ff21d54d 100644
--- a/llvm/test/Transforms/InstCombine/indexed-gep-compares.ll
+++ b/llvm/test/Transforms/InstCombine/indexed-gep-compares.ll
@@ -376,7 +376,7 @@ define i1 @test8(ptr %in, i64 %offset) {
; CHECK-NEXT: [[LD:%.*]] = load i64, ptr [[IN:%.*]], align 8
; CHECK-NEXT: [[TMP0:%.*]] = trunc i64 [[LD]] to i32
; CHECK-NEXT: [[CASTI8:%.*]] = inttoptr i32 [[TMP0]] to ptr
-; CHECK-NEXT: [[TMP1:%.*]] = trunc i64 [[OFFSET:%.*]] to i32
+; CHECK-NEXT: [[TMP1:%.*]] = trunc nsw i64 [[OFFSET:%.*]] to i32
; CHECK-NEXT: [[GEPI8:%.*]] = getelementptr inbounds i8, ptr [[CASTI8]], i32 [[TMP1]]
; CHECK-NEXT: [[TMP2:%.*]] = trunc i64 [[LD]] to i32
; CHECK-NEXT: [[PTRCAST:%.*]] = inttoptr i32 [[TMP2]] to ptr
diff --git a/llvm/test/Transforms/InstCombine/load-cmp.ll b/llvm/test/Transforms/InstCombine/load-cmp.ll
index ccaf31f3084d6..df34e7d58bcba 100644
--- a/llvm/test/Transforms/InstCombine/load-cmp.ll
+++ b/llvm/test/Transforms/InstCombine/load-cmp.ll
@@ -293,8 +293,7 @@ define i1 @test10_struct_arr_i16(i16 %x) {
define i1 @test10_struct_arr_i64(i64 %x) {
; CHECK-LABEL: @test10_struct_arr_i64(
-; CHECK-NEXT: [[TMP1:%.*]] = and i64 [[X:%.*]], 4294967295
-; CHECK-NEXT: [[R:%.*]] = icmp ne i64 [[TMP1]], 1
+; CHECK-NEXT: [[R:%.*]] = icmp ne i64 [[TMP1:%.*]], 1
; CHECK-NEXT: ret i1 [[R]]
;
%p = getelementptr inbounds [4 x %Foo], ptr @GStructArr, i64 0, i64 %x, i32 2
@@ -331,7 +330,7 @@ define i1 @test10_struct_arr_noinbounds_i64(i64 %x) {
define i1 @pr93017(i64 %idx) {
; CHECK-LABEL: @pr93017(
-; CHECK-NEXT: [[TMP1:%.*]] = trunc i64 [[IDX:%.*]] to i32
+; CHECK-NEXT: [[TMP1:%.*]] = trunc nsw i64 [[IDX:%.*]] to i32
; CHECK-NEXT: [[GEP:%.*]] = getelementptr inbounds [2 x ptr], ptr @table, i32 0, i32 [[TMP1]]
; CHECK-NEXT: [[V:%.*]] = load ptr, ptr [[GEP]], align 4
; CHECK-NEXT: [[CMP:%.*]] = icmp ne ptr [[V]], null
diff --git a/llvm/test/Transforms/InstCombine/pr39908.ll b/llvm/test/Transforms/InstCombine/pr39908.ll
index 5d13a331c6d2e..c36495ddcf10f 100644
--- a/llvm/test/Transforms/InstCombine/pr39908.ll
+++ b/llvm/test/Transforms/InstCombine/pr39908.ll
@@ -19,7 +19,7 @@ define i1 @test(ptr %p, i32 %n) {
; Same test using 64-bit indices.
define i1 @test64(ptr %p, i64 %n) {
; CHECK-LABEL: @test64(
-; CHECK-NEXT: [[TMP1:%.*]] = trunc i64 [[N:%.*]] to i32
+; CHECK-NEXT: [[TMP1:%.*]] = trunc nsw i64 [[N:%.*]] to i32
; CHECK-NEXT: [[CMP:%.*]] = icmp eq i32 [[TMP1]], 1
; CHECK-NEXT: ret i1 [[CMP]]
;
@@ -32,7 +32,7 @@ define i1 @test64(ptr %p, i64 %n) {
; Here the offset overflows and is treated modulo 2^32. This is UB.
define i1 @test64_overflow(ptr %p, i64 %n) {
; CHECK-LABEL: @test64_overflow(
-; CHECK-NEXT: [[TMP1:%.*]] = trunc i64 [[N:%.*]] to i32
+; CHECK-NEXT: [[TMP1:%.*]] = trunc nsw i64 [[N:%.*]] to i32
; CHECK-NEXT: [[CMP:%.*]] = icmp eq i32 [[TMP1]], 1
; CHECK-NEXT: ret i1 [[CMP]]
;
diff --git a/llvm/test/Transforms/InstCombine/sub-gep.ll b/llvm/test/Transforms/InstCombine/sub-gep.ll
index 45e5686ad70e7..0db5cbee33009 100644
--- a/llvm/test/Transforms/InstCombine/sub-gep.ll
+++ b/llvm/test/Transforms/InstCombine/sub-gep.ll
@@ -394,7 +394,7 @@ define i64 @negative_ptrtoint_sub_zext_ptrtoint(ptr %p, i32 %offset) {
define i16 @test25_as1(ptr addrspace(1) %P, i64 %A) {
; CHECK-LABEL: @test25_as1(
-; CHECK-NEXT: [[TMP1:%.*]] = trunc i64 [[A:%.*]] to i16
+; CHECK-NEXT: [[TMP1:%.*]] = trunc nsw i64 [[A:%.*]] to i16
; CHECK-NEXT: [[B_IDX:%.*]] = shl nsw i16 [[TMP1]], 1
; CHECK-NEXT: [[GEPDIFF:%.*]] = add nsw i16 [[B_IDX]], -84
; CHECK-NEXT: ret i16 [[GEPDIFF]]
More information about the llvm-commits
mailing list