llvmbot
commented
4 days ago @llvm/issue-subscribers-clang-codegen
Author: Yunfei Li (mzyx-hnu)
In this case, I used `#pragma pack` to set the structure alignment to 1. Then, I loaded the first int member of the structure. When compiled with `-O2`, the align attribute in the IR was converted to 4.
~~~
// align.c
#pragma pack(1)
struct a {
volatile int b;
int c;
long d;
long e;
short f;
unsigned : 25;
volatile long g;
int h;
} i[];
static volatile struct a k = {1}, l = {2}, m = {3}, n = {4}, p = {5}, q = {6};
static int foo() {
int s = m.b; // load m with align 4
m.c = k.c = l.c = n.c = p.c = q.c = 1;
return s;
}
int main() {
foo();
return 0;
}
~~~
compile with main branch `-O2 -emit-llvm -S`
~~~
; ModuleID = './align.c'
source_filename = "./align.c"
target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128-Fn32"
target triple = "aarch64-unknown-linux-gnu"
%struct.a = type <{ i32, i32, i64, i64, i16, i32, i64, i32 }>
@i = dso_local local_unnamed_addr global [1 x %struct.a] zeroinitializer, align 1
@m = internal global <{ i32, i32, i64, i64, i16, [4 x i8], i64, i32 }> <{ i32 3, i32 0, i64 0, i64 0, i16 0, [4 x i8] zeroinitializer, i64 0, i32 0 }>, align 4
@q = internal global <{ i32, i32, i64, i64, i16, [4 x i8], i64, i32 }> <{ i32 6, i32 0, i64 0, i64 0, i16 0, [4 x i8] zeroinitializer, i64 0, i32 0 }>, align 1
@p = internal global <{ i32, i32, i64, i64, i16, [4 x i8], i64, i32 }> <{ i32 5, i32 0, i64 0, i64 0, i16 0, [4 x i8] zeroinitializer, i64 0, i32 0 }>, align 1
@n = internal global <{ i32, i32, i64, i64, i16, [4 x i8], i64, i32 }> <{ i32 4, i32 0, i64 0, i64 0, i16 0, [4 x i8] zeroinitializer, i64 0, i32 0 }>, align 1
@l = internal global <{ i32, i32, i64, i64, i16, [4 x i8], i64, i32 }> <{ i32 2, i32 0, i64 0, i64 0, i16 0, [4 x i8] zeroinitializer, i64 0, i32 0 }>, align 1
@k = internal global <{ i32, i32, i64, i64, i16, [4 x i8], i64, i32 }> <{ i32 1, i32 0, i64 0, i64 0, i16 0, [4 x i8] zeroinitializer, i64 0, i32 0 }>, align 1
; Function Attrs: nofree norecurse nounwind memory(readwrite, argmem: none) uwtable
define dso_local noundef i32 @main() local_unnamed_addr #0 {
entry:
%0 = load volatile i32, ptr @m, align 4, !tbaa !6
store volatile i32 1, ptr getelementptr inbounds nuw (i8, ptr @q, i64 4), align 1, !tbaa !13
store volatile i32 1, ptr getelementptr inbounds nuw (i8, ptr @p, i64 4), align 1, !tbaa !13
store volatile i32 1, ptr getelementptr inbounds nuw (i8, ptr @n, i64 4), align 1, !tbaa !13
store volatile i32 1, ptr getelementptr inbounds nuw (i8, ptr @l, i64 4), align 1, !tbaa !13
store volatile i32 1, ptr getelementptr inbounds nuw (i8, ptr @k, i64 4), align 1, !tbaa !13
store volatile i32 1, ptr getelementptr inbounds nuw (i8, ptr @m, i64 4), align 4, !tbaa !13
ret i32 0
}
~~~
The pass responsible for increasing the alignment attribute is `InferAlignmentPass`. Since the user code loaded the struct object m (Or m.b, the first int member of m) with a 4-byte alignment once, the alignment attribute for m was increased from 1 to 4. However, the struct's actual memory is still 1-byte alignment, and this inconsistency has caused issues in my development. Should llvm add a check in `InferAlignmentPass` for this case？
mzyx-hnu
In this case, I used
#pragma packto set the structure alignment to 1. Then, I loaded the first int member of the structure. When compiled with-O2, the align attribute in the IR was converted to 4.compile with main branch
-O2 -emit-llvm -SThe pass responsible for increasing the alignment attribute is
InferAlignmentPass. Since the user code loaded the struct object m (Or m.b, the first int member of m) with a 4-byte alignment once, the alignment attribute for m was increased from 1 to 4. However, the struct's actual memory is still 1-byte alignment, and this inconsistency has caused issues in my development. Should llvm add a check inInferAlignmentPassfor this case？