https://gcc.gnu.org/bugzilla/show_bug.cgi?id=125296
Bug ID: 125296
Summary: forwprop increases alignment requirements
Product: gcc
Version: 17.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: kristerw at gcc dot gnu.org
Blocks: 118443
Target Milestone: ---
I believe the forwprop pass is incorrectly increasing alignment requirements
when the function below is compiled on x86_64 with -O2 -fno-strict-aliasing:
struct S {
struct S *next;
int vals[5];
};
int foo(struct S *s, int idx) {
return *(volatile int *)&(s->vals[idx]);
}
The original GIMPLE for the function is:
int foo (struct S * s, int idx)
{
int * _1;
int _5;
<bb 2> :
_1 = &s_2(D)->vals[idx_3(D)];
_5 ={v} MEM[(volatile int *)_1];
return _5;
}
GIMPLE only requires correct alignment for the memory access, so I believe this
only requires s to be 32-bit aligned.
The forwprop pass modifies it in a way that requires 64-bit alignment:
int foo (struct S * s, int idx)
{
int _5;
<bb 2> :
_5 ={v} MEM <struct S> [(volatile int *)s_2(D)].vals[idx_3(D)];
return _5;
}
Referenced Bugs:
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=118443
[Bug 118443] [Meta bug] Bugs triggered by and blocking more smtgcc testing