[llvm] [GlobalOpt] Fix alignments of globals introduced for allocations (PR #216480)
Ömer Sinan Ağacan via llvm-commits
llvm-commits at lists.llvm.org
Mon Aug 17 01:43:04 PDT 2026
https://github.com/osa1 updated https://github.com/llvm/llvm-project/pull/216480
>From 95557d8496a8980af14956ec7246c736c6f23d83 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=C3=96mer=20Sinan=20A=C4=9Facan?= <omer at osa1.net>
Date: Mon, 17 Aug 2026 09:40:46 +0100
Subject: [PATCH 1/2] Add tests
---
.../Transforms/GlobalOpt/global-align-1.ll | 44 ++++++++++
.../Transforms/GlobalOpt/global-align-2.ll | 47 +++++++++++
.../Transforms/GlobalOpt/global-align-3.ll | 48 +++++++++++
.../GlobalOpt/global-align-memset.ll | 82 +++++++++++++++++++
4 files changed, 221 insertions(+)
create mode 100644 llvm/test/Transforms/GlobalOpt/global-align-1.ll
create mode 100644 llvm/test/Transforms/GlobalOpt/global-align-2.ll
create mode 100644 llvm/test/Transforms/GlobalOpt/global-align-3.ll
create mode 100644 llvm/test/Transforms/GlobalOpt/global-align-memset.ll
diff --git a/llvm/test/Transforms/GlobalOpt/global-align-1.ll b/llvm/test/Transforms/GlobalOpt/global-align-1.ll
new file mode 100644
index 0000000000000..0e4eaa4b54f0b
--- /dev/null
+++ b/llvm/test/Transforms/GlobalOpt/global-align-1.ll
@@ -0,0 +1,44 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --check-globals all --version 6
+; RUN: opt < %s -passes=globalopt -S | FileCheck %s
+
+; Test from issue #215533.
+
+target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-i128:128-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+ at g10 = internal global ptr null, align 8
+ at g26 = dso_local global <4 x i32> zeroinitializer, align 16
+ at g5 = dso_local global i8 0, align 1
+
+;.
+; CHECK: @g26 = dso_local local_unnamed_addr global <4 x i32> zeroinitializer, align 16
+; CHECK: @g5 = dso_local local_unnamed_addr global i8 0, align 1
+; CHECK: @g10.body = internal unnamed_addr global [16 x i8] undef{{$}}
+;.
+define dso_local i32 @main() {
+; CHECK-LABEL: define dso_local i32 @main() local_unnamed_addr {
+; CHECK-NEXT: [[ENTRY:.*:]]
+; CHECK-NEXT: [[TMP0:%.*]] = load <4 x i32>, ptr @g26, align 16
+; CHECK-NEXT: store <4 x i32> [[TMP0]], ptr @g10.body, align 16
+; CHECK-NEXT: [[BB3_0_COPYLOAD:%.*]] = load i32, ptr @g10.body, align 1
+; CHECK-NEXT: [[TOBOOL:%.*]] = icmp ne i32 [[BB3_0_COPYLOAD]], 0
+; CHECK-NEXT: [[STOREDV:%.*]] = zext i1 [[TOBOOL]] to i8
+; CHECK-NEXT: store i8 [[STOREDV]], ptr @g5, align 1
+; CHECK-NEXT: ret i32 0
+;
+entry:
+ %call = call noalias align 16 ptr @aligned_alloc(i64 noundef 16, i64 noundef 16)
+ store ptr %call, ptr @g10, align 8
+ %0 = load <4 x i32>, ptr @g26, align 16
+ store <4 x i32> %0, ptr %call, align 16
+ %1 = load ptr, ptr @g10, align 8
+ %bb3.0.copyload = load i32, ptr %1, align 1
+ %tobool = icmp ne i32 %bb3.0.copyload, 0
+ %storedv = zext i1 %tobool to i8
+ store i8 %storedv, ptr @g5, align 1
+ ret i32 0
+}
+
+declare noalias noundef ptr @aligned_alloc(i64 allocalign noundef, i64 noundef) #0
+
+attributes #0 = { allockind("alloc,uninitialized,aligned") allocsize(1) }
diff --git a/llvm/test/Transforms/GlobalOpt/global-align-2.ll b/llvm/test/Transforms/GlobalOpt/global-align-2.ll
new file mode 100644
index 0000000000000..5e729275426ec
--- /dev/null
+++ b/llvm/test/Transforms/GlobalOpt/global-align-2.ll
@@ -0,0 +1,47 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --check-globals all --version 6
+; RUN: opt < %s -passes=globalopt -S | FileCheck %s
+
+; Use the declared alignment of the allocation as the global's alignment.
+
+target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+ at g = internal global ptr null
+
+;.
+; CHECK: @g.body = internal unnamed_addr global [64 x i8] undef{{$}}
+;.
+define void @init() {
+; CHECK-LABEL: define void @init() local_unnamed_addr {
+; CHECK-NEXT: ret void
+;
+ %m = call noalias align 64 ptr @malloc(i64 64)
+ store ptr %m, ptr @g, align 8
+ ret void
+}
+
+define void @store(<8 x float> %v, i8 %c) {
+; CHECK-LABEL: define void @store(
+; CHECK-SAME: <8 x float> [[V:%.*]], i8 [[C:%.*]]) local_unnamed_addr {
+; CHECK-NEXT: store <8 x float> [[V]], ptr @g.body, align 32
+; CHECK-NEXT: store i8 [[C]], ptr getelementptr inbounds nuw (i8, ptr @g.body, i64 3), align 1
+; CHECK-NEXT: ret void
+;
+ %p = load ptr, ptr @g, align 8
+ store <8 x float> %v, ptr %p, align 32
+ %q = getelementptr i8, ptr %p, i64 3
+ store i8 %c, ptr %q, align 1
+ ret void
+}
+
+define <8 x float> @load() {
+; CHECK-LABEL: define <8 x float> @load() local_unnamed_addr {
+; CHECK-NEXT: [[V:%.*]] = load <8 x float>, ptr @g.body, align 32
+; CHECK-NEXT: ret <8 x float> [[V]]
+;
+ %p = load ptr, ptr @g, align 8
+ %v = load <8 x float>, ptr %p, align 32
+ ret <8 x float> %v
+}
+
+declare noalias ptr @malloc(i64) allockind("alloc,uninitialized") allocsize(0)
diff --git a/llvm/test/Transforms/GlobalOpt/global-align-3.ll b/llvm/test/Transforms/GlobalOpt/global-align-3.ll
new file mode 100644
index 0000000000000..a899ab5761583
--- /dev/null
+++ b/llvm/test/Transforms/GlobalOpt/global-align-3.ll
@@ -0,0 +1,48 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --check-globals all --version 6
+; RUN: opt < %s -passes=globalopt -S | FileCheck %s
+
+; Check the the alignment is not specified when it's smaller than the preferred
+; alignment of the type.
+
+target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+ at g = internal global ptr null
+
+;.
+; CHECK: @g.body = internal unnamed_addr global [32 x i8] undef{{$}}
+;.
+define void @init() {
+; CHECK-LABEL: define void @init() local_unnamed_addr {
+; CHECK-NEXT: ret void
+;
+ %m = call noalias ptr @aligned_alloc(i64 8, i64 32)
+ store ptr %m, ptr @g, align 8
+ ret void
+}
+
+define void @store(<4 x float> %v, i8 %c) {
+; CHECK-LABEL: define void @store(
+; CHECK-SAME: <4 x float> [[V:%.*]], i8 [[C:%.*]]) local_unnamed_addr {
+; CHECK-NEXT: store <4 x float> [[V]], ptr @g.body, align 16
+; CHECK-NEXT: store i8 [[C]], ptr getelementptr inbounds nuw (i8, ptr @g.body, i64 3), align 1
+; CHECK-NEXT: ret void
+;
+ %p = load ptr, ptr @g, align 8
+ store <4 x float> %v, ptr %p, align 16
+ %q = getelementptr i8, ptr %p, i64 3
+ store i8 %c, ptr %q, align 1
+ ret void
+}
+
+define <4 x float> @load() {
+; CHECK-LABEL: define <4 x float> @load() local_unnamed_addr {
+; CHECK-NEXT: [[V:%.*]] = load <4 x float>, ptr @g.body, align 16
+; CHECK-NEXT: ret <4 x float> [[V]]
+;
+ %p = load ptr, ptr @g, align 8
+ %v = load <4 x float>, ptr %p, align 16
+ ret <4 x float> %v
+}
+
+declare noalias ptr @aligned_alloc(i64 allocalign, i64) allockind("alloc,uninitialized,aligned") allocsize(1)
diff --git a/llvm/test/Transforms/GlobalOpt/global-align-memset.ll b/llvm/test/Transforms/GlobalOpt/global-align-memset.ll
new file mode 100644
index 0000000000000..344b6747135b0
--- /dev/null
+++ b/llvm/test/Transforms/GlobalOpt/global-align-memset.ll
@@ -0,0 +1,82 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --check-globals all --version 6
+; RUN: opt < %s -passes=globalopt -S | FileCheck %s
+
+; When initializing an aligned global pass the alignment to the memset call.
+
+target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+ at aligned = internal global ptr null
+ at unaligned = internal global ptr null
+
+;.
+; CHECK: @aligned.body = internal unnamed_addr global [64 x i8] undef{{$}}
+; CHECK: @unaligned.body = internal unnamed_addr global [64 x i8] undef{{$}}
+;.
+define void @init_aligned() {
+; CHECK-LABEL: define void @init_aligned() local_unnamed_addr {
+; CHECK-NEXT: call void @llvm.memset.p0.i64(ptr @aligned.body, i8 0, i64 64, i1 false)
+; CHECK-NEXT: ret void
+;
+ %m = call noalias align 32 ptr @aligned_calloc(i64 32, i64 64)
+ store ptr %m, ptr @aligned, align 8
+ ret void
+}
+
+define void @store_aligned(<8 x float> %v) {
+; CHECK-LABEL: define void @store_aligned(
+; CHECK-SAME: <8 x float> [[V:%.*]]) local_unnamed_addr {
+; CHECK-NEXT: store <8 x float> [[V]], ptr @aligned.body, align 32
+; CHECK-NEXT: ret void
+;
+ %p = load ptr, ptr @aligned, align 8
+ store <8 x float> %v, ptr %p, align 32
+ ret void
+}
+
+define <8 x float> @load_aligned() {
+; CHECK-LABEL: define <8 x float> @load_aligned() local_unnamed_addr {
+; CHECK-NEXT: [[V:%.*]] = load <8 x float>, ptr @aligned.body, align 32
+; CHECK-NEXT: ret <8 x float> [[V]]
+;
+ %p = load ptr, ptr @aligned, align 8
+ %v = load <8 x float>, ptr %p, align 32
+ ret <8 x float> %v
+}
+
+define void @init_unaligned() {
+; CHECK-LABEL: define void @init_unaligned() local_unnamed_addr {
+; CHECK-NEXT: call void @llvm.memset.p0.i64(ptr @unaligned.body, i8 0, i64 64, i1 false)
+; CHECK-NEXT: ret void
+;
+ %m = call noalias ptr @calloc(i64 1, i64 64)
+ store ptr %m, ptr @unaligned, align 8
+ ret void
+}
+
+define void @store_unaligned(i8 %c) {
+; CHECK-LABEL: define void @store_unaligned(
+; CHECK-SAME: i8 [[C:%.*]]) local_unnamed_addr {
+; CHECK-NEXT: store i8 [[C]], ptr @unaligned.body, align 1
+; CHECK-NEXT: ret void
+;
+ %p = load ptr, ptr @unaligned, align 8
+ store i8 %c, ptr %p, align 1
+ ret void
+}
+
+define i8 @load_unaligned() {
+; CHECK-LABEL: define i8 @load_unaligned() local_unnamed_addr {
+; CHECK-NEXT: [[V:%.*]] = load i8, ptr @unaligned.body, align 1
+; CHECK-NEXT: ret i8 [[V]]
+;
+ %p = load ptr, ptr @unaligned, align 8
+ %v = load i8, ptr %p, align 1
+ ret i8 %v
+}
+
+declare noalias ptr @aligned_calloc(i64, i64) allockind("alloc,zeroed,aligned") allocsize(1)
+declare noalias ptr @calloc(i64, i64) allockind("alloc,zeroed") allocsize(0,1)
+;.
+; CHECK: attributes #[[ATTR0:[0-9]+]] = { nocallback nofree nosync nounwind willreturn memory(argmem: write) }
+;.
>From 61c56ceb3860019bcfb205d56a8cedb72feb1eb0 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=C3=96mer=20Sinan=20A=C4=9Facan?= <omer at osa1.net>
Date: Mon, 17 Aug 2026 09:41:08 +0100
Subject: [PATCH 2/2] Fix and update tests
---
llvm/lib/Transforms/IPO/GlobalOpt.cpp | 12 ++++++++++--
llvm/test/Transforms/GlobalOpt/calloc-promote.ll | 2 +-
llvm/test/Transforms/GlobalOpt/global-align-1.ll | 2 +-
llvm/test/Transforms/GlobalOpt/global-align-2.ll | 2 +-
.../test/Transforms/GlobalOpt/global-align-memset.ll | 4 ++--
5 files changed, 15 insertions(+), 7 deletions(-)
diff --git a/llvm/lib/Transforms/IPO/GlobalOpt.cpp b/llvm/lib/Transforms/IPO/GlobalOpt.cpp
index f78b7169a6a26..ad611801195c8 100644
--- a/llvm/lib/Transforms/IPO/GlobalOpt.cpp
+++ b/llvm/lib/Transforms/IPO/GlobalOpt.cpp
@@ -944,6 +944,15 @@ OptimizeGlobalAddressOfAllocation(GlobalVariable *GV, CallInst *CI,
UndefValue::get(GlobalType), GV->getName() + ".body", nullptr,
GV->getThreadLocalMode());
+ // Alignment of the return value of the allocator call.
+ Align GVAlign = CI->getPointerAlignment(DL);
+
+ // Only specify the global alignment if it increases the preferred alignment.
+ // Otherwise leave it unset to allow other optimizations to increase it.
+ if (GVAlign > DL.getPreferredAlign(NewGV)) {
+ NewGV->setAlignment(GVAlign);
+ }
+
// Initialize the global at the point of the original call. Note that this
// is a different point from the initialization referred to below for the
// nullability handling. Sublety: We have not proven the original global was
@@ -951,8 +960,7 @@ OptimizeGlobalAddressOfAllocation(GlobalVariable *GV, CallInst *CI,
// of the new global as may need to re-init the storage multiple times.
if (!isa<UndefValue>(InitVal)) {
IRBuilder<> Builder(CI->getNextNode());
- // TODO: Use alignment above if align!=1
- Builder.CreateMemSet(NewGV, InitVal, AllocSize, std::nullopt);
+ Builder.CreateMemSet(NewGV, InitVal, AllocSize, NewGV->getAlign());
}
// Update users of the allocation to use the new global instead.
diff --git a/llvm/test/Transforms/GlobalOpt/calloc-promote.ll b/llvm/test/Transforms/GlobalOpt/calloc-promote.ll
index c369ed9b9fc11..d12b4a8f82cad 100644
--- a/llvm/test/Transforms/GlobalOpt/calloc-promote.ll
+++ b/llvm/test/Transforms/GlobalOpt/calloc-promote.ll
@@ -6,7 +6,7 @@
define signext i32 @f() local_unnamed_addr {
; CHECK-LABEL: @f(
; CHECK-NEXT: entry:
-; CHECK-NEXT: call void @llvm.memset.p0.i64(ptr @g.body, i8 0, i64 4, i1 false)
+; CHECK-NEXT: call void @llvm.memset.p0.i64(ptr align 16 @g.body, i8 0, i64 4, i1 false)
; CHECK-NEXT: store i16 -1, ptr @g.body, align 2
; CHECK-NEXT: ret i32 0
;
diff --git a/llvm/test/Transforms/GlobalOpt/global-align-1.ll b/llvm/test/Transforms/GlobalOpt/global-align-1.ll
index 0e4eaa4b54f0b..0ab7fae067c1d 100644
--- a/llvm/test/Transforms/GlobalOpt/global-align-1.ll
+++ b/llvm/test/Transforms/GlobalOpt/global-align-1.ll
@@ -13,7 +13,7 @@ target triple = "x86_64-unknown-linux-gnu"
;.
; CHECK: @g26 = dso_local local_unnamed_addr global <4 x i32> zeroinitializer, align 16
; CHECK: @g5 = dso_local local_unnamed_addr global i8 0, align 1
-; CHECK: @g10.body = internal unnamed_addr global [16 x i8] undef{{$}}
+; CHECK: @g10.body = internal unnamed_addr global [16 x i8] undef, align 16
;.
define dso_local i32 @main() {
; CHECK-LABEL: define dso_local i32 @main() local_unnamed_addr {
diff --git a/llvm/test/Transforms/GlobalOpt/global-align-2.ll b/llvm/test/Transforms/GlobalOpt/global-align-2.ll
index 5e729275426ec..9f8c6be68e4e4 100644
--- a/llvm/test/Transforms/GlobalOpt/global-align-2.ll
+++ b/llvm/test/Transforms/GlobalOpt/global-align-2.ll
@@ -9,7 +9,7 @@ target triple = "x86_64-unknown-linux-gnu"
@g = internal global ptr null
;.
-; CHECK: @g.body = internal unnamed_addr global [64 x i8] undef{{$}}
+; CHECK: @g.body = internal unnamed_addr global [64 x i8] undef, align 64
;.
define void @init() {
; CHECK-LABEL: define void @init() local_unnamed_addr {
diff --git a/llvm/test/Transforms/GlobalOpt/global-align-memset.ll b/llvm/test/Transforms/GlobalOpt/global-align-memset.ll
index 344b6747135b0..53716b733caad 100644
--- a/llvm/test/Transforms/GlobalOpt/global-align-memset.ll
+++ b/llvm/test/Transforms/GlobalOpt/global-align-memset.ll
@@ -10,12 +10,12 @@ target triple = "x86_64-unknown-linux-gnu"
@unaligned = internal global ptr null
;.
-; CHECK: @aligned.body = internal unnamed_addr global [64 x i8] undef{{$}}
+; CHECK: @aligned.body = internal unnamed_addr global [64 x i8] undef, align 32
; CHECK: @unaligned.body = internal unnamed_addr global [64 x i8] undef{{$}}
;.
define void @init_aligned() {
; CHECK-LABEL: define void @init_aligned() local_unnamed_addr {
-; CHECK-NEXT: call void @llvm.memset.p0.i64(ptr @aligned.body, i8 0, i64 64, i1 false)
+; CHECK-NEXT: call void @llvm.memset.p0.i64(ptr align 32 @aligned.body, i8 0, i64 64, i1 false)
; CHECK-NEXT: ret void
;
%m = call noalias align 32 ptr @aligned_calloc(i64 32, i64 64)
More information about the llvm-commits
mailing list