shylie/llvm-project
1
0
Fork 0
Code Issues Pull Requests Actions 6 Packages Projects Releases Wiki Activity
llvm-project/llvm/test/Transforms/DeadStoreElimination/atomic.ll
Jin Huang ce2518f59d
[AA] Improve precision for monotonic atomic load/store operations (#158169) 
Refines `getModRefInfo` for `Monotonic` and `Unordered` atomic
operations to perform a full alias evaluation instead of conservatively
returning `ModRef`.

This allows AA to return more precise results (e.g., `NoModRef`),
unblocking optimizations in passes like DSE/PRE when the pointers are
proven not to alias.

LIT testing: 
1. Fixed the `XFAIL` test `@test9` in `DSE/atomic-todo.ll`, where the
improved AA now allows DSE to correctly remove a dead store.
2. Added test `@test9a` to the same file to ensure DSE is still blocked
when pointers may alias.
3. I am also reviewing tests in GVN/PRE to see where this change may be
beneficial. While all existing tests pass, I will update this PR with
any new GVN test cases in a follow-up commit.

---------

Co-authored-by: Jin Huang <jingold@google.com>
2025-12-05 15:59:49 -08:00

371 lines
11 KiB
LLVM
Raw Blame History

; 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.
@x = common global i32 0, align 4
@y = common global i32 0, align 4
declare void @randomop(ptr)
; DSE across unordered store (allowed)
define void @test1() {
; CHECK-LABEL: @test1(
; CHECK-NEXT: store atomic i32 0, ptr @y unordered, align 4
; CHECK-NEXT: store i32 1, ptr @x, align 4
; CHECK-NEXT: ret void
;
store i32 0, ptr @x
store atomic i32 0, ptr @y unordered, align 4
store i32 1, ptr @x
ret void
}
; DSE remove unordered store (allowed)
define void @test4() {
; CHECK-LABEL: @test4(
; CHECK-NEXT: store i32 1, ptr @x, align 4
; CHECK-NEXT: ret void
;
store atomic i32 0, ptr @x unordered, align 4
store i32 1, ptr @x
ret void
}
; 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
;
store i32 0, ptr @x
store atomic i32 1, ptr @x unordered, align 4
ret void
}
; DSE no-op unordered atomic store (allowed)
define void @test7() {
; CHECK-LABEL: @test7(
; CHECK-NEXT: ret void
;
%x = load atomic i32, ptr @x unordered, align 4
store atomic i32 %x, ptr @x unordered, align 4
ret void
}
; DSE seq_cst store (be conservative; DSE doesn't have infrastructure
; to reason about atomic operations).
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
;
%a = alloca i32
store atomic i32 0, ptr %a seq_cst, align 4
ret void
}
; DSE and seq_cst load (be conservative; DSE doesn't have infrastructure
; to reason about atomic operations).
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
; CHECK-NEXT: [[X:%.*]] = load atomic i32, ptr @x seq_cst, align 4
; CHECK-NEXT: ret i32 [[X]]
;
%a = alloca i32
call void @randomop(ptr %a)
store i32 0, ptr %a, align 4
%x = load atomic i32, ptr @x seq_cst, align 4
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 @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
ret void
}
; DSE across monotonic load (forbidden since the eliminated store is atomic)
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
; CHECK-NEXT: ret i32 [[X]]
;
store atomic i32 0, ptr @x monotonic, align 4
%x = load atomic i32, ptr @y monotonic, align 4
store atomic i32 1, ptr @x monotonic, align 4
ret i32 %x
}
; DSE across monotonic store (forbidden since the eliminated store is atomic)
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
; CHECK-NEXT: ret void
;
store atomic i32 0, ptr @x monotonic, align 4
store atomic i32 42, ptr @y monotonic, align 4
store atomic i32 1, ptr @x monotonic, align 4
ret void
}
; But DSE is not allowed across a release-acquire pair.
define i32 @test15() {
; CHECK-LABEL: @test15(
; CHECK-NEXT: store i32 0, ptr @x, align 4
; CHECK-NEXT: store atomic i32 0, ptr @y release, align 4
; CHECK-NEXT: [[X:%.*]] = load atomic i32, ptr @y acquire, align 4
; CHECK-NEXT: store i32 1, ptr @x, align 4
; CHECK-NEXT: ret i32 [[X]]
;
store i32 0, ptr @x
store atomic i32 0, ptr @y release, align 4
%x = load atomic i32, ptr @y acquire, align 4
store i32 1, ptr @x
ret i32 %x
}
@z = common global i64 0, align 4
@a = common global i64 0, align 4
; Be conservative, do not kill regular store.
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, align 8
; CHECK-NEXT: ret i64 [[RES]]
;
store i64 1, ptr @z
%res = atomicrmw add ptr @z, i64 -1 monotonic
ret i64 %res
}
; Be conservative, do not kill regular store.
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, align 8
; CHECK-NEXT: ret i64 [[RES]]
;
store i64 1, ptr @z
%res = atomicrmw add ptr @z, i64 -1 acq_rel
ret i64 %res
}
; 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, align 8
; CHECK-NEXT: store i64 2, ptr @z, align 8
; CHECK-NEXT: ret i64 [[RES]]
;
store i64 1, ptr @z
%res = atomicrmw add ptr @a, i64 -1 monotonic
store i64 2, ptr @z
ret i64 %res
}
; Be conservative, do not eliminate stores across atomic operations > monotonic.
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, align 8
; CHECK-NEXT: store i64 2, ptr @z, align 8
; CHECK-NEXT: ret i64 [[RES]]
;
store i64 1, ptr @z
%res = atomicrmw add ptr @a, i64 -1 release
store i64 2, ptr @z
ret i64 %res
}
; Be conservative, do not eliminate may-alias stores.
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, align 8
; CHECK-NEXT: store i64 2, ptr @z, align 8
; CHECK-NEXT: ret i64 [[RES]]
;
store i64 1, ptr @z
%res = atomicrmw add ptr %ptr, i64 -1 monotonic
store i64 2, ptr @z
ret i64 %res
}
; Be conservative, do not eliminate aliasing stores.
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, align 8
; CHECK-NEXT: store i64 2, ptr @z, align 8
; CHECK-NEXT: ret i64 [[RES]]
;
store i64 1, ptr @z
%res = atomicrmw add ptr @z, i64 -1 monotonic
store i64 2, ptr @z
ret i64 %res
}
; Be conservative, do not eliminate non-monotonic cmpxchg.
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, align 4
; CHECK-NEXT: store i32 2, ptr @x, align 4
; CHECK-NEXT: ret { i32, i1 } [[RET]]
;
store i32 1, ptr @x
%ret = cmpxchg volatile ptr @x, i32 10, i32 20 seq_cst monotonic
store i32 2, ptr @x
ret { i32, i1 } %ret
}
; 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, align 4
; CHECK-NEXT: store i32 2, ptr @x, align 4
; CHECK-NEXT: ret { i32, i1 } [[RET]]
;
store i32 1, ptr @x
%ret = cmpxchg volatile ptr @y, i32 10, i32 20 monotonic monotonic
store i32 2, ptr @x
ret { i32, i1 } %ret
}
; Be conservative, do not eliminate non-monotonic cmpxchg.
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, align 4
; CHECK-NEXT: store i32 2, ptr @x, align 4
; CHECK-NEXT: ret { i32, i1 } [[RET]]
;
store i32 1, ptr @x
%ret = cmpxchg volatile ptr @y, i32 10, i32 20 seq_cst seq_cst
store i32 2, ptr @x
ret { i32, i1 } %ret
}
; Be conservative, do not eliminate may-alias stores.
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, align 4
; CHECK-NEXT: store i32 2, ptr @x, align 4
; CHECK-NEXT: ret { i32, i1 } [[RET]]
;
store i32 1, ptr @x
%ret = cmpxchg volatile ptr %ptr, i32 10, i32 20 monotonic monotonic
store i32 2, ptr @x
ret { i32, i1 } %ret
}
; Be conservative, do not eliminate alias stores.
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, align 4
; CHECK-NEXT: store i32 2, ptr @x, align 4
; CHECK-NEXT: ret { i32, i1 } [[RET]]
;
store i32 1, ptr @x
%ret = cmpxchg volatile ptr @x, i32 10, i32 20 monotonic monotonic
store i32 2, ptr @x
ret { i32, i1 } %ret
}
; **** Noop load->store tests **************************************************
; We can optimize unordered atomic loads or stores.
define void @test_load_atomic(ptr %Q) {
; CHECK-LABEL: @test_load_atomic(
; CHECK-NEXT: ret void
;
%a = load atomic i32, ptr %Q unordered, align 4
store atomic i32 %a, ptr %Q unordered, align 4
ret void
}
; We can optimize unordered atomic loads or stores.
define void @test_store_atomic(ptr %Q) {
; CHECK-LABEL: @test_store_atomic(
; CHECK-NEXT: ret void
;
%a = load i32, ptr %Q
store atomic i32 %a, ptr %Q unordered, align 4
ret void
}
; We can NOT optimize release atomic loads or stores.
define void @test_store_atomic_release(ptr %Q) {
; CHECK-LABEL: @test_store_atomic_release(
; CHECK-NEXT: [[A:%.*]] = load i32, ptr [[Q:%.*]], align 4
; CHECK-NEXT: store atomic i32 [[A]], ptr [[Q]] release, align 4
; CHECK-NEXT: ret void
;
%a = load i32, ptr %Q
store atomic i32 %a, ptr %Q release, align 4
ret void
}
Reference in New Issue View Git Blame Copy Permalink