[llvm] Revert "Revert "[AA] Improve precision for monotonic atomic load/stor… (PR #195015)
Jin Huang via llvm-commits
llvm-commits at lists.llvm.org
Sun May 3 21:37:22 PDT 2026
https://github.com/jinhuang1102 updated https://github.com/llvm/llvm-project/pull/195015
>From 2f0302b10cbc1b8aaed9f42b829241663927d607 Mon Sep 17 00:00:00 2001
From: Jin Huang <jinhuang1102 at gmail.com>
Date: Wed, 29 Apr 2026 22:42:58 -0700
Subject: [PATCH] =?UTF-8?q?Revert=20"Revert=20"[AA]=20Improve=20precision?=
=?UTF-8?q?=20for=20monotonic=20atomic=20load/store=20operat=E2=80=A6"?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
This reverts commit c8941df6a1e4ed5e1ba3287985a60e1d7512c250.
---
llvm/lib/Analysis/AliasAnalysis.cpp | 19 ++-
.../DeadStoreElimination/atomic-todo.ll | 23 ---
.../Transforms/DeadStoreElimination/atomic.ll | 133 ++++++++++++++----
.../DeadStoreElimination/atomic_tmp.ll | 54 +++++++
4 files changed, 177 insertions(+), 52 deletions(-)
delete mode 100644 llvm/test/Transforms/DeadStoreElimination/atomic-todo.ll
create mode 100644 llvm/test/Transforms/DeadStoreElimination/atomic_tmp.ll
diff --git a/llvm/lib/Analysis/AliasAnalysis.cpp b/llvm/lib/Analysis/AliasAnalysis.cpp
index 102f0c070d517..4f1bf9058ed32 100644
--- a/llvm/lib/Analysis/AliasAnalysis.cpp
+++ b/llvm/lib/Analysis/AliasAnalysis.cpp
@@ -462,7 +462,7 @@ ModRefInfo AAResults::getModRefInfo(const LoadInst *L,
const MemoryLocation &Loc,
AAQueryInfo &AAQI) {
// Be conservative in the face of atomic.
- if (isStrongerThan(L->getOrdering(), AtomicOrdering::Unordered))
+ if (isStrongerThanMonotonic(L->getOrdering()))
return ModRefInfo::ModRef;
// If the load address doesn't alias the given address, it doesn't read
@@ -472,6 +472,13 @@ ModRefInfo AAResults::getModRefInfo(const LoadInst *L,
if (AR == AliasResult::NoAlias)
return ModRefInfo::NoModRef;
}
+
+ assert(!isStrongerThanMonotonic(L->getOrdering()) &&
+ "Stronger atomic orderings should have been handled above!");
+
+ if (isStrongerThanUnordered(L->getOrdering()))
+ return ModRefInfo::ModRef;
+
// Otherwise, a load just reads.
return ModRefInfo::Ref;
}
@@ -480,7 +487,7 @@ ModRefInfo AAResults::getModRefInfo(const StoreInst *S,
const MemoryLocation &Loc,
AAQueryInfo &AAQI) {
// Be conservative in the face of atomic.
- if (isStrongerThan(S->getOrdering(), AtomicOrdering::Unordered))
+ if (isStrongerThanMonotonic(S->getOrdering()))
return ModRefInfo::ModRef;
if (Loc.Ptr) {
@@ -498,7 +505,13 @@ ModRefInfo AAResults::getModRefInfo(const StoreInst *S,
return ModRefInfo::NoModRef;
}
- // Otherwise, a store just writes.
+ assert(!isStrongerThanMonotonic(S->getOrdering()) &&
+ "Stronger atomic orderings should have been handled above!");
+
+ if (isStrongerThanUnordered(S->getOrdering()))
+ return ModRefInfo::ModRef;
+
+ // A store just writes.
return ModRefInfo::Mod;
}
diff --git a/llvm/test/Transforms/DeadStoreElimination/atomic-todo.ll b/llvm/test/Transforms/DeadStoreElimination/atomic-todo.ll
deleted file mode 100644
index 1c160442f8579..0000000000000
--- a/llvm/test/Transforms/DeadStoreElimination/atomic-todo.ll
+++ /dev/null
@@ -1,23 +0,0 @@
-; XFAIL: *
-; RUN: opt -passes=dse -S < %s | FileCheck %s
-
-target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64"
-target triple = "x86_64-apple-macosx10.7.0"
-
-; Basic correctness tests for atomic stores.
-; Note that it turns out essentially every transformation DSE does is legal on
-; atomic ops, just some transformations are not allowed across release-acquire pairs.
-
- at x = common global i32 0, align 4
- at y = common global i32 0, align 4
-
-; DSE across monotonic load (allowed as long as the eliminated store isUnordered)
-define i32 @test9() {
-; CHECK-LABEL: test9
-; CHECK-NOT: store i32 0
-; CHECK: store i32 1
- store i32 0, ptr @x
- %x = load atomic i32, ptr @y monotonic, align 4
- store i32 1, ptr @x
- ret i32 %x
-}
diff --git a/llvm/test/Transforms/DeadStoreElimination/atomic.ll b/llvm/test/Transforms/DeadStoreElimination/atomic.ll
index 55b9384e88d93..5d5bf95291da0 100644
--- a/llvm/test/Transforms/DeadStoreElimination/atomic.ll
+++ b/llvm/test/Transforms/DeadStoreElimination/atomic.ll
@@ -10,8 +10,10 @@ target triple = "x86_64-apple-macosx10.7.0"
@x = common global i32 0, align 4
@y = common global i32 0, align 4
+ at global_ptr = global ptr null, align 8
declare void @randomop(ptr)
+declare void @external_sync()
; DSE across unordered store (allowed)
define void @test1() {
@@ -37,9 +39,21 @@ define void @test4() {
ret void
}
-; DSE unordered store overwriting non-atomic store (allowed)
+; DSE doesn't remove monotonic store.
define void @test5() {
; CHECK-LABEL: @test5(
+; CHECK-NEXT: store atomic i32 2, ptr @x monotonic, align 4
+; CHECK-NEXT: store i32 1, ptr @x, align 4
+; CHECK-NEXT: ret void
+;
+ store atomic i32 2, ptr @x monotonic, align 4
+ store i32 1, ptr @x
+ ret void
+}
+
+; DSE unordered store overwriting non-atomic store (allowed)
+define void @test6() {
+; CHECK-LABEL: @test6(
; CHECK-NEXT: store atomic i32 1, ptr @x unordered, align 4
; CHECK-NEXT: ret void
;
@@ -49,8 +63,8 @@ define void @test5() {
}
; DSE no-op unordered atomic store (allowed)
-define void @test6() {
-; CHECK-LABEL: @test6(
+define void @test7() {
+; CHECK-LABEL: @test7(
; CHECK-NEXT: ret void
;
%x = load atomic i32, ptr @x unordered, align 4
@@ -60,8 +74,8 @@ define void @test6() {
; DSE seq_cst store (be conservative; DSE doesn't have infrastructure
; to reason about atomic operations).
-define void @test7() {
-; CHECK-LABEL: @test7(
+define void @test8() {
+; CHECK-LABEL: @test8(
; CHECK-NEXT: [[A:%.*]] = alloca i32, align 4
; CHECK-NEXT: store atomic i32 0, ptr [[A]] seq_cst, align 4
; CHECK-NEXT: ret void
@@ -73,8 +87,8 @@ define void @test7() {
; DSE and seq_cst load (be conservative; DSE doesn't have infrastructure
; to reason about atomic operations).
-define i32 @test8() {
-; CHECK-LABEL: @test8(
+define i32 @test9() {
+; CHECK-LABEL: @test9(
; CHECK-NEXT: [[A:%.*]] = alloca i32, align 4
; CHECK-NEXT: call void @randomop(ptr [[A]])
; CHECK-NEXT: store i32 0, ptr [[A]], align 4
@@ -88,11 +102,40 @@ define i32 @test8() {
ret i32 %x
}
+; DSE across monotonic load (allowed if the monotonic load's address is NoAlias)
+define i32 @test10() {
+; CHECK-LABEL: @test10(
+; CHECK-NEXT: [[X:%.*]] = load atomic i32, ptr @y monotonic, align 4
+; CHECK-NEXT: store i32 1, ptr @x, align 4
+; CHECK-NEXT: ret i32 [[X]]
+;
+ store i32 0, ptr @x
+ %x = load atomic i32, ptr @y monotonic, align 4
+ store i32 1, ptr @x
+ ret i32 %x
+}
+
+; DSE across monotonic load (blocked if the atomic load's address isn't NoAlias)
+define i32 @test11(ptr %ptr) {
+; CHECK-LABEL: @test11(
+; CHECK-NEXT: store i32 0, ptr @x, align 4
+; CHECK-NEXT: [[X:%.*]] = load atomic i32, ptr [[PTR:%.*]] monotonic, align 4
+; CHECK-NEXT: store i32 1, ptr @x, align 4
+; CHECK-NEXT: ret i32 [[X]]
+;
+ store i32 0, ptr @x
+ %x = load atomic i32, ptr %ptr monotonic, align 4
+ store i32 1, ptr @x
+ ret i32 %x
+}
+
; DSE across monotonic store (allowed as long as the eliminated store isUnordered)
-define void @test10() {
-; CHECK-LABEL: test10
-; CHECK-NOT: store i32 0
-; CHECK: store i32 1
+define void @test12() {
+; CHECK-LABEL: @test12(
+; CHECK-NEXT: store atomic i32 42, ptr @y monotonic, align 4
+; CHECK-NEXT: store i32 1, ptr @x, align 4
+; CHECK-NEXT: ret void
+;
store i32 0, ptr @x
store atomic i32 42, ptr @y monotonic, align 4
store i32 1, ptr @x
@@ -100,8 +143,8 @@ define void @test10() {
}
; DSE across monotonic load (forbidden since the eliminated store is atomic)
-define i32 @test11() {
-; CHECK-LABEL: @test11(
+define i32 @test13() {
+; CHECK-LABEL: @test13(
; CHECK-NEXT: store atomic i32 0, ptr @x monotonic, align 4
; CHECK-NEXT: [[X:%.*]] = load atomic i32, ptr @y monotonic, align 4
; CHECK-NEXT: store atomic i32 1, ptr @x monotonic, align 4
@@ -114,8 +157,8 @@ define i32 @test11() {
}
; DSE across monotonic store (forbidden since the eliminated store is atomic)
-define void @test12() {
-; CHECK-LABEL: @test12(
+define void @test14() {
+; CHECK-LABEL: @test14(
; CHECK-NEXT: store atomic i32 0, ptr @x monotonic, align 4
; CHECK-NEXT: store atomic i32 42, ptr @y monotonic, align 4
; CHECK-NEXT: store atomic i32 1, ptr @x monotonic, align 4
@@ -150,7 +193,7 @@ define i32 @test15() {
define i64 @test_atomicrmw_0() {
; CHECK-LABEL: @test_atomicrmw_0(
; CHECK-NEXT: store i64 1, ptr @z, align 8
-; CHECK-NEXT: [[RES:%.*]] = atomicrmw add ptr @z, i64 -1 monotonic
+; CHECK-NEXT: [[RES:%.*]] = atomicrmw add ptr @z, i64 -1 monotonic, align 8
; CHECK-NEXT: ret i64 [[RES]]
;
store i64 1, ptr @z
@@ -162,7 +205,7 @@ define i64 @test_atomicrmw_0() {
define i64 @test_atomicrmw_1() {
; CHECK-LABEL: @test_atomicrmw_1(
; CHECK-NEXT: store i64 1, ptr @z, align 8
-; CHECK-NEXT: [[RES:%.*]] = atomicrmw add ptr @z, i64 -1 acq_rel
+; CHECK-NEXT: [[RES:%.*]] = atomicrmw add ptr @z, i64 -1 acq_rel, align 8
; CHECK-NEXT: ret i64 [[RES]]
;
store i64 1, ptr @z
@@ -173,7 +216,7 @@ define i64 @test_atomicrmw_1() {
; Monotonic atomicrmw should not block eliminating no-aliasing stores.
define i64 @test_atomicrmw_2() {
; CHECK-LABEL: @test_atomicrmw_2(
-; CHECK-NEXT: [[RES:%.*]] = atomicrmw add ptr @a, i64 -1 monotonic
+; CHECK-NEXT: [[RES:%.*]] = atomicrmw add ptr @a, i64 -1 monotonic, align 8
; CHECK-NEXT: store i64 2, ptr @z, align 8
; CHECK-NEXT: ret i64 [[RES]]
;
@@ -187,7 +230,7 @@ define i64 @test_atomicrmw_2() {
define i64 @test_atomicrmw_3() {
; CHECK-LABEL: @test_atomicrmw_3(
; CHECK-NEXT: store i64 1, ptr @z, align 8
-; CHECK-NEXT: [[RES:%.*]] = atomicrmw add ptr @a, i64 -1 release
+; CHECK-NEXT: [[RES:%.*]] = atomicrmw add ptr @a, i64 -1 release, align 8
; CHECK-NEXT: store i64 2, ptr @z, align 8
; CHECK-NEXT: ret i64 [[RES]]
;
@@ -201,7 +244,7 @@ define i64 @test_atomicrmw_3() {
define i64 @test_atomicrmw_4(ptr %ptr) {
; CHECK-LABEL: @test_atomicrmw_4(
; CHECK-NEXT: store i64 1, ptr @z, align 8
-; CHECK-NEXT: [[RES:%.*]] = atomicrmw add ptr [[PTR:%.*]], i64 -1 monotonic
+; CHECK-NEXT: [[RES:%.*]] = atomicrmw add ptr [[PTR:%.*]], i64 -1 monotonic, align 8
; CHECK-NEXT: store i64 2, ptr @z, align 8
; CHECK-NEXT: ret i64 [[RES]]
;
@@ -215,7 +258,7 @@ define i64 @test_atomicrmw_4(ptr %ptr) {
define i64 @test_atomicrmw_5() {
; CHECK-LABEL: @test_atomicrmw_5(
; CHECK-NEXT: store i64 1, ptr @z, align 8
-; CHECK-NEXT: [[RES:%.*]] = atomicrmw add ptr @z, i64 -1 monotonic
+; CHECK-NEXT: [[RES:%.*]] = atomicrmw add ptr @z, i64 -1 monotonic, align 8
; CHECK-NEXT: store i64 2, ptr @z, align 8
; CHECK-NEXT: ret i64 [[RES]]
;
@@ -229,7 +272,7 @@ define i64 @test_atomicrmw_5() {
define { i32, i1} @test_cmpxchg_1() {
; CHECK-LABEL: @test_cmpxchg_1(
; CHECK-NEXT: store i32 1, ptr @x, align 4
-; CHECK-NEXT: [[RET:%.*]] = cmpxchg volatile ptr @x, i32 10, i32 20 seq_cst monotonic
+; CHECK-NEXT: [[RET:%.*]] = cmpxchg volatile ptr @x, i32 10, i32 20 seq_cst monotonic, align 4
; CHECK-NEXT: store i32 2, ptr @x, align 4
; CHECK-NEXT: ret { i32, i1 } [[RET]]
;
@@ -242,7 +285,7 @@ define { i32, i1} @test_cmpxchg_1() {
; Monotonic cmpxchg should not block DSE for non-aliasing stores.
define { i32, i1} @test_cmpxchg_2() {
; CHECK-LABEL: @test_cmpxchg_2(
-; CHECK-NEXT: [[RET:%.*]] = cmpxchg volatile ptr @y, i32 10, i32 20 monotonic monotonic
+; CHECK-NEXT: [[RET:%.*]] = cmpxchg volatile ptr @y, i32 10, i32 20 monotonic monotonic, align 4
; CHECK-NEXT: store i32 2, ptr @x, align 4
; CHECK-NEXT: ret { i32, i1 } [[RET]]
;
@@ -256,7 +299,7 @@ define { i32, i1} @test_cmpxchg_2() {
define { i32, i1} @test_cmpxchg_3() {
; CHECK-LABEL: @test_cmpxchg_3(
; CHECK-NEXT: store i32 1, ptr @x, align 4
-; CHECK-NEXT: [[RET:%.*]] = cmpxchg volatile ptr @y, i32 10, i32 20 seq_cst seq_cst
+; CHECK-NEXT: [[RET:%.*]] = cmpxchg volatile ptr @y, i32 10, i32 20 seq_cst seq_cst, align 4
; CHECK-NEXT: store i32 2, ptr @x, align 4
; CHECK-NEXT: ret { i32, i1 } [[RET]]
;
@@ -270,7 +313,7 @@ define { i32, i1} @test_cmpxchg_3() {
define { i32, i1} @test_cmpxchg_4(ptr %ptr) {
; CHECK-LABEL: @test_cmpxchg_4(
; CHECK-NEXT: store i32 1, ptr @x, align 4
-; CHECK-NEXT: [[RET:%.*]] = cmpxchg volatile ptr [[PTR:%.*]], i32 10, i32 20 monotonic monotonic
+; CHECK-NEXT: [[RET:%.*]] = cmpxchg volatile ptr [[PTR:%.*]], i32 10, i32 20 monotonic monotonic, align 4
; CHECK-NEXT: store i32 2, ptr @x, align 4
; CHECK-NEXT: ret { i32, i1 } [[RET]]
;
@@ -284,7 +327,7 @@ define { i32, i1} @test_cmpxchg_4(ptr %ptr) {
define { i32, i1} @test_cmpxchg_5(ptr %ptr) {
; CHECK-LABEL: @test_cmpxchg_5(
; CHECK-NEXT: store i32 1, ptr @x, align 4
-; CHECK-NEXT: [[RET:%.*]] = cmpxchg volatile ptr @x, i32 10, i32 20 monotonic monotonic
+; CHECK-NEXT: [[RET:%.*]] = cmpxchg volatile ptr @x, i32 10, i32 20 monotonic monotonic, align 4
; CHECK-NEXT: store i32 2, ptr @x, align 4
; CHECK-NEXT: ret { i32, i1 } [[RET]]
;
@@ -327,3 +370,41 @@ define void @test_store_atomic_release(ptr %Q) {
store atomic i32 %a, ptr %Q release, align 4
ret void
}
+
+define void @test_atomic_release_escape() {
+entry:
+ %cs = alloca { i64, i64, i64 }, align 8
+ %version_ptr = getelementptr inbounds { i64, i64, i64 }, ptr %cs, i32 0, i32 0
+ store i64 0, ptr %version_ptr, align 8
+ %start_ip_ptr = getelementptr inbounds { i64, i64, i64 }, ptr %cs, i32 0, i32 1
+ store i64 4096, ptr %start_ip_ptr, align 8
+ %abort_ip_ptr = getelementptr inbounds { i64, i64, i64 }, ptr %cs, i32 0, i32 2
+ store i64 8192, ptr %abort_ip_ptr, align 8
+ store atomic ptr %cs, ptr @global_ptr release, align 8
+
+ ret void
+}
+; CHECK-LABEL: @test_atomic_release_escape(
+; CHECK: store i64 0, ptr %version_ptr, align 8
+; CHECK: store i64 4096, ptr %start_ip_ptr, align 8
+; CHECK: store i64 8192, ptr %abort_ip_ptr, align 8
+; CHECK: store atomic ptr %cs, ptr @global_ptr release, align 8
+
+define void @test_atomic_monotonic_escape() {
+entry:
+ %cs = alloca { i64, i64, i64 }, align 8
+ %version_ptr = getelementptr inbounds { i64, i64, i64 }, ptr %cs, i32 0, i32 0
+ store i64 0, ptr %version_ptr, align 8
+ %start_ip_ptr = getelementptr inbounds { i64, i64, i64 }, ptr %cs, i32 0, i32 1
+ store i64 4096, ptr %start_ip_ptr, align 8
+ %abort_ip_ptr = getelementptr inbounds { i64, i64, i64 }, ptr %cs, i32 0, i32 2
+ store i64 8192, ptr %abort_ip_ptr, align 8
+ store atomic ptr %cs, ptr @global_ptr monotonic, align 8
+
+ ret void
+}
+; CHECK-LABEL: @test_atomic_monotonic_escape(
+; CHECK-NOT: store i64 0, ptr %version_ptr
+; CHECK-NOT: store i64 4096, ptr %start_ip_ptr
+; CHECK-NOT: store i64 8192, ptr %abort_ip_ptr
+; CHECK: store atomic ptr %cs, ptr @global_ptr monotonic, align 8
diff --git a/llvm/test/Transforms/DeadStoreElimination/atomic_tmp.ll b/llvm/test/Transforms/DeadStoreElimination/atomic_tmp.ll
new file mode 100644
index 0000000000000..20a9b03dbcc66
--- /dev/null
+++ b/llvm/test/Transforms/DeadStoreElimination/atomic_tmp.ll
@@ -0,0 +1,54 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt -passes=dse -S < %s | FileCheck %s
+
+target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64"
+target triple = "x86_64-apple-macosx10.7.0"
+
+; Basic correctness tests for atomic stores.
+; Note that it turns out essentially every transformation DSE does is legal on
+; atomic ops, just some transformations are not allowed across release-acquire pairs.
+
+ at x = common global i32 0, align 4
+ at y = common global i32 0, align 4
+ at global_ptr = global ptr null, align 8
+
+declare void @randomop(ptr)
+declare void @external_sync()
+
+define void @test_atomic_release_escape() {
+entry:
+ %cs = alloca { i64, i64, i64 }, align 8
+ %version_ptr = getelementptr inbounds { i64, i64, i64 }, ptr %cs, i32 0, i32 0
+ store i64 0, ptr %version_ptr, align 8
+ %start_ip_ptr = getelementptr inbounds { i64, i64, i64 }, ptr %cs, i32 0, i32 1
+ store i64 4096, ptr %start_ip_ptr, align 8
+ %abort_ip_ptr = getelementptr inbounds { i64, i64, i64 }, ptr %cs, i32 0, i32 2
+ store i64 8192, ptr %abort_ip_ptr, align 8
+ store atomic ptr %cs, ptr @global_ptr release, align 8
+
+ ret void
+}
+; CHECK-LABEL: @test_atomic_release_escape(
+; CHECK: store i64 0, ptr %version_ptr, align 8
+; CHECK: store i64 4096, ptr %start_ip_ptr, align 8
+; CHECK: store i64 8192, ptr %abort_ip_ptr, align 8
+; CHECK: store atomic ptr %cs, ptr @global_ptr release, align 8
+
+define void @test_atomic_monotonic_escape() {
+entry:
+ %cs = alloca { i64, i64, i64 }, align 8
+ %version_ptr = getelementptr inbounds { i64, i64, i64 }, ptr %cs, i32 0, i32 0
+ store i64 0, ptr %version_ptr, align 8
+ %start_ip_ptr = getelementptr inbounds { i64, i64, i64 }, ptr %cs, i32 0, i32 1
+ store i64 4096, ptr %start_ip_ptr, align 8
+ %abort_ip_ptr = getelementptr inbounds { i64, i64, i64 }, ptr %cs, i32 0, i32 2
+ store i64 8192, ptr %abort_ip_ptr, align 8
+ store atomic ptr %cs, ptr @global_ptr monotonic, align 8
+
+ ret void
+}
+; CHECK-LABEL: @test_atomic_monotonic_escape(
+; CHECK-NOT: store i64 0, ptr %version_ptr
+; CHECK-NOT: store i64 4096, ptr %start_ip_ptr
+; CHECK-NOT: store i64 8192, ptr %abort_ip_ptr
+; CHECK: store atomic ptr %cs, ptr @global_ptr monotonic, align 8
More information about the llvm-commits
mailing list