https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114674
Bug ID: 114674
Summary: [aarch64] ldp_fusion fails to merge 2 strs due to
imprecise alignment info
Product: gcc
Version: 14.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: rtl-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: dizhao at os dot amperecomputing.com
Target Milestone: ---
For the case below:
typedef struct {
unsigned int f1;
unsigned int f2;
} test_struct;
static test_struct ts = {
123, 456
};
void foo(void)
{
ts.f2 = 36969 * (ts.f2 & 65535) + (ts.f1 >> 16);
ts.f1 = 18000 * (ts.f2 & 65535) + (ts.f2 >> 16);
}
When compiled with "-O3 --param=aarch64-stp-policy=aligned", gcc failed to fuse
the memory access instructions into stp/ldp, the dump file 312r.ldp_fusion1
says "ldp/stp policy says no". However, the accessing to ts.f1 is aligned to 64
bit:
(insn 26 23 33 2 (set (mem/c:SI (lo_sum:DI (reg/f:DI 113)
(symbol_ref:DI ("*.LANCHOR0") [flags 0x182])) [1 ts.f1+0 S4
A64])
...
So it looks like the 2 str instructions should be fused with
aarch64-stp-policy=aligned.
After debugged this a bit, it seems the problem is ldp_bb_info::fuse_pair
changed the alignment info when calling adjust_address_nv, to rewrite the base
of ts.f1.
