https://gcc.gnu.org/bugzilla/show_bug.cgi?id=120916
Bug ID: 120916
Summary: debug info for IV increment is lost
Product: gcc
Version: 16.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: driver
Assignee: unassigned at gcc dot gnu.org
Reporter: hubicka at gcc dot gnu.org
Target Milestone: ---
jan@padlo:/tmp> cat t.c
int s = 1023;
int a[1024];
__attribute__ ((noipa))
int test()
{
for (
int i = 0; /* Line 7 */
i < s; /* Line 8 */
i++) /* Line 9 */
a[i]++;
}
int
main()
{
for (int i = 0; i < 1000000; i++)
test ();
}
jan@padlo:/tmp> gcc -O3 t.c -g
jan@padlo:/tmp> perf record ./a.out
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.020 MB perf.data (497 samples) ]
perf report then shows:
│ a[i]++; ▒
│ movd %esi,%xmm1 ▒
│ shl $0x4,%rdx ▒
│ pshufd $0x0,%xmm1,%xmm1 ▒
0.37 │ data16 cs nopw 0x0(%rax,%rax,1) ▒
│ xchg %ax,%ax ▒
│40:┌─→movdqa 0x404040(%rax),%xmm0 ▒
0.44 │ │ add $0x10,%rax ▒
0.22 │ │ paddd %xmm1,%xmm0 ▒
1.74 │ │ movaps %xmm0,0x404030(%rax) ▒
│ │i < s; /* Line 8 */ ▒
91.10 │ ├──cmp %rdx,%rax ▒
0.22 │ └──jne 40 ◆
Notice that a[I]++ is misplaced and should be just before
│40:┌─→movdqa 0x404040(%rax),%xmm0 ▒
Moreover we lose location of "add" that should correspond to i++; /* Line 9 */
LLVM does:
│ a[i]++; ▒
0.50 │60:┌─→movdqa (%r8,%rdi,1),%xmm1 ▒
│ │ movdqa 0x10(%r8,%rdi,1),%xmm2 ▒
│ │ psubd %xmm0,%xmm1 ▒
0.36 │ │ psubd %xmm0,%xmm2 ▒
0.48 │ │ movdqa %xmm1,(%r8,%rdi,1) ▒
43.46 │ │ movdqa %xmm2,0x10(%r8,%rdi,1) ▒
│ │i++) /* Line 9 */ ▒
46.83 │ │ add $0x20,%r8 ▒
│ ├──cmp %r8,%rdx ▒
│ └──jne 60 ▒
It loses location of the "cmp" instruction, but that is less harmful since it
is not an header of the loop.
In optimized dup we have:
<bb 2> [local count: 118111600]:
[t.c:6:2] # DEBUG BEGIN_STMT
[t.c:7:7] # DEBUG BEGIN_STMT
# DEBUG i => 0
[t.c:8:9] # DEBUG BEGIN_STMT
[t.c:8:9] s.0_11 = s;
[t.c:8:9] if (s.0_11 > 0)
goto <bb 3>; [89.00%]
else
goto <bb 11>; [11.00%]
<bb 3> [local count: 105119324]:
niters.6_5 = (unsigned int) s.0_11;
_21 = niters.6_5 + 4294967295;
if (_21 <= 2)
goto <bb 7>; [10.00%]
else
goto <bb 4>; [90.00%]
<bb 4> [local count: 94607391]:
bnd.7_24 = niters.6_5 >> 2;
_20 = (sizetype) bnd.7_24;
_7 = _20 * 16;
<bb 5> [local count: 860067200]:
# ivtmp.21_17 = PHI <ivtmp.21_1(5), 0(4)>
# DEBUG i => NULL
[t.c:10:4] # DEBUG BEGIN_STMT
[t.c:10:5] vect__1.12_29 = MEM <vector(4) int> [(int *)&a + ivtmp.21_17 * 1];
[t.c:10:8] vect__2.13_31 = vect__1.12_29 + { 1, 1, 1, 1 };
[t.c:10:8] MEM <vector(4) int> [(int *)&a + ivtmp.21_17 * 1] = vect__2.13_31;
[t.c:9:8] # DEBUG BEGIN_STMT
# DEBUG i => NULL
[t.c:8:9] # DEBUG BEGIN_STMT
^^^^ maybe we want to just move this after the plus to "if"
ivtmp.21_1 = ivtmp.21_17 + 16;
if (ivtmp.21_1 != _7)
goto <bb 5>; [89.00%]
else
goto <bb 6>; [11.00%]
