Skip to content

Commit 4e89d11

Browse files
authored
[InstCombine] Convert mem intrinsic with null into a noop (#100388)
When src/dest passed into memset/memcpy is null: ``` len == 0: this call is a noop. len != 0: the behavior is undefined. ``` See also https://llvm.org/docs/LangRef.html#llvm-memset-intrinsics Alive2: https://alive2.llvm.org/ce/z/tJeRNL This patch converts these mem intrinsic calls into an assumption `len == 0` to mitigate code-size bloat caused by JumpThreading.
1 parent 59ca618 commit 4e89d11

File tree

3 files changed

+92
-1
lines changed

3 files changed

+92
-1
lines changed

llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp

Lines changed: 14 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -1584,16 +1584,30 @@ Instruction *InstCombinerImpl::visitCallInst(CallInst &CI) {
15841584
return eraseInstFromFunction(CI);
15851585
}
15861586

1587+
auto IsPointerUndefined = [MI](Value *Ptr) {
1588+
return isa<ConstantPointerNull>(Ptr) &&
1589+
!NullPointerIsDefined(
1590+
MI->getFunction(),
1591+
cast<PointerType>(Ptr->getType())->getAddressSpace());
1592+
};
1593+
bool SrcIsUndefined = false;
15871594
// If we can determine a pointer alignment that is bigger than currently
15881595
// set, update the alignment.
15891596
if (auto *MTI = dyn_cast<AnyMemTransferInst>(MI)) {
15901597
if (Instruction *I = SimplifyAnyMemTransfer(MTI))
15911598
return I;
1599+
SrcIsUndefined = IsPointerUndefined(MTI->getRawSource());
15921600
} else if (auto *MSI = dyn_cast<AnyMemSetInst>(MI)) {
15931601
if (Instruction *I = SimplifyAnyMemSet(MSI))
15941602
return I;
15951603
}
15961604

1605+
// If src/dest is null, this memory intrinsic must be a noop.
1606+
if (SrcIsUndefined || IsPointerUndefined(MI->getRawDest())) {
1607+
Builder.CreateAssumption(Builder.CreateIsNull(MI->getLength()));
1608+
return eraseInstFromFunction(CI);
1609+
}
1610+
15971611
if (Changed) return II;
15981612
}
15991613

Lines changed: 77 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,77 @@
1+
; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
2+
; RUN: opt -S -passes=instcombine < %s | FileCheck %s
3+
4+
define void @memset_null(i64 %len) {
5+
; CHECK-LABEL: define void @memset_null(
6+
; CHECK-SAME: i64 [[LEN:%.*]]) {
7+
; CHECK-NEXT: [[TMP1:%.*]] = icmp eq i64 [[LEN]], 0
8+
; CHECK-NEXT: call void @llvm.assume(i1 [[TMP1]])
9+
; CHECK-NEXT: ret void
10+
;
11+
call void @llvm.memset.p0.i64(ptr null, i8 0, i64 %len, i1 false)
12+
ret void
13+
}
14+
15+
define void @memset_null_ub() {
16+
; CHECK-LABEL: define void @memset_null_ub() {
17+
; CHECK-NEXT: store i64 poison, ptr null, align 4294967296
18+
; CHECK-NEXT: ret void
19+
;
20+
call void @llvm.memset.p0.i64(ptr null, i8 0, i64 8, i1 false)
21+
ret void
22+
}
23+
24+
define void @memcpy_null_src(ptr %dst, i64 %len) {
25+
; CHECK-LABEL: define void @memcpy_null_src(
26+
; CHECK-SAME: ptr [[DST:%.*]], i64 [[LEN:%.*]]) {
27+
; CHECK-NEXT: [[TMP1:%.*]] = icmp eq i64 [[LEN]], 0
28+
; CHECK-NEXT: call void @llvm.assume(i1 [[TMP1]])
29+
; CHECK-NEXT: ret void
30+
;
31+
call void @llvm.memcpy.p0.i64(ptr %dst, ptr null, i64 %len, i1 false)
32+
ret void
33+
}
34+
35+
define void @memmove_null_src(ptr %dst, i64 %len) {
36+
; CHECK-LABEL: define void @memmove_null_src(
37+
; CHECK-SAME: ptr [[DST:%.*]], i64 [[LEN:%.*]]) {
38+
; CHECK-NEXT: [[TMP1:%.*]] = icmp eq i64 [[LEN]], 0
39+
; CHECK-NEXT: call void @llvm.assume(i1 [[TMP1]])
40+
; CHECK-NEXT: ret void
41+
;
42+
call void @llvm.memmove.p0.i64(ptr %dst, ptr null, i64 %len, i1 false)
43+
ret void
44+
}
45+
46+
define void @memset_element_atomic(i64 %len) {
47+
; CHECK-LABEL: define void @memset_element_atomic(
48+
; CHECK-SAME: i64 [[LEN:%.*]]) {
49+
; CHECK-NEXT: [[TMP1:%.*]] = icmp eq i64 [[LEN]], 0
50+
; CHECK-NEXT: call void @llvm.assume(i1 [[TMP1]])
51+
; CHECK-NEXT: ret void
52+
;
53+
call void @llvm.memset.element.unordered.atomic.p0.i64(ptr align 1 null, i8 0, i64 %len, i32 1)
54+
ret void
55+
}
56+
57+
; negative tests
58+
59+
define void @memset_null_volatile(i64 %len) {
60+
; CHECK-LABEL: define void @memset_null_volatile(
61+
; CHECK-SAME: i64 [[LEN:%.*]]) {
62+
; CHECK-NEXT: call void @llvm.memset.p0.i64(ptr null, i8 0, i64 [[LEN]], i1 true)
63+
; CHECK-NEXT: ret void
64+
;
65+
call void @llvm.memset.p0.i64(ptr null, i8 0, i64 %len, i1 true)
66+
ret void
67+
}
68+
69+
define void @memset_null_is_defined(i64 %len) null_pointer_is_valid {
70+
; CHECK-LABEL: define void @memset_null_is_defined(
71+
; CHECK-SAME: i64 [[LEN:%.*]]) #[[ATTR0:[0-9]+]] {
72+
; CHECK-NEXT: call void @llvm.memset.p0.i64(ptr align 4294967296 null, i8 0, i64 [[LEN]], i1 false)
73+
; CHECK-NEXT: ret void
74+
;
75+
call void @llvm.memset.p0.i64(ptr null, i8 0, i64 %len, i1 false)
76+
ret void
77+
}

llvm/test/Transforms/InstCombine/mempcpy.ll

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -55,7 +55,7 @@ define ptr @memcpy_big_const_n(ptr %d, ptr nocapture readonly %s) {
5555

5656
define i32 @PR48810() {
5757
; CHECK-LABEL: @PR48810(
58-
; CHECK-NEXT: call void @llvm.memcpy.p0.p0.i64(ptr align 1 undef, ptr align 4294967296 null, i64 undef, i1 false)
58+
; CHECK-NEXT: store i1 true, ptr poison, align 1
5959
; CHECK-NEXT: ret i32 undef
6060
;
6161
%r = call dereferenceable(1) ptr @mempcpy(ptr undef, ptr null, i64 undef)

0 commit comments

Comments
 (0)