The following should be due to the same root cause, but it also fails at -Os:

[533] % clangtk -O1 small.c; ./a.out
[534] %
[534] % clangtk -Os small.c; ./a.out
Segmentation fault
[535] % clangtk -O2 small.c; ./a.out
Segmentation fault
[536] % clangtk -O3 small.c; ./a.out
Segmentation fault
[537] %
[537] % cat small.c
static int *a[2][7], b;
int main() {
for (b = 0; b < 2; b++)
a[1][b+1] = a[0][0];
return 0;
}

This reproduces for me with version 12.0.0. Not with 11.1.0.

The direct cause of the segfaults is MOVAPS to unaligned memory, as shown in the objdump:

0000000000401100

I've never looked at -globalopt before, but this seems wrong:

@​a = internal unnamed_addr global [2 x [7 x i32*]] zeroinitializer, align 16

define void @​#50253 () {
%x = load i32*, i32** getelementptr inbounds ([2 x [7 x i32*]], [2 x [7 x i32*]]* @​a, i64 0, i64 0, i64 0), align 16
store i32* %x, i32** getelementptr inbounds ([2 x [7 x i32*]], [2 x [7 x i32*]]* @​a, i64 0, i64 1, i64 1), align 16
store i32* %x, i32** getelementptr inbounds ([2 x [7 x i32*]], [2 x [7 x i32*]]* @​a, i64 0, i64 1, i64 2), align 8
ret void
}

$ opt -globalopt align.ll -S -debug
Args: ./opt -globalopt align.ll -S -debug
PERFORMING GLOBAL SRA ON: @​a = internal unnamed_addr global [2 x [7 x i32*]] zeroinitializer, align 16
MARKING CONSTANT: @​a.0 = internal unnamed_addr global [7 x i32*] zeroinitializer, align 16
PERFORMING GLOBAL SRA ON: @​a.0 = internal unnamed_addr constant [7 x i32*] zeroinitializer, align 16
GLOBAL NEVER LOADED: @​a.1 = internal unnamed_addr global [7 x i32*] zeroinitializer, align 16
MARKING CONSTANT: @​a.0.0 = internal unnamed_addr global i32* null, align 16
*** Marking constant allowed us to simplify all users and delete global!
GLOBAL NEVER LOADED: @​a.1 = internal unnamed_addr global [7 x i32*] zeroinitializer, align 16
; ModuleID = 'align.ll'
source_filename = "align.ll"

@​a.1 = internal unnamed_addr global [7 x i32*] zeroinitializer, align 16

define void @​#50253 () local_unnamed_addr {
store i32* null, i32** getelementptr inbounds ([7 x i32*], [7 x i32*]* @​a.1, i32 0, i64 1), align 16
store i32* null, i32** getelementptr inbounds ([7 x i32*], [7 x i32*]* @​a.1, i32 0, i64 2), align 8
ret void
}

In the original code, we accessed a[1][1], and that's correctly shown as "align 16" because it's the 8th element past a[0][0] which is "align 16" at the declaration.

But in the transformed code, we created a.1 with "align 16" and accessed a.1[0][1] which can't be "align 16" .

https://reviews.llvm.org/D102552

This was the same problem as bug 49661.

https://reviews.llvm.org/rGf34311c4024d

*** This bug has been marked as a duplicate of bug #49005 ***