https://gcc.gnu.org/bugzilla/show_bug.cgi?id=127693
--- Comment #18 from Thiago Macieira <thiago at kde dot org> ---
(In reply to Richard Biener from comment #16)
> Oh, and checking whether the original Qt issue is solved as well is of
> course appreciated!
No, it looks like it hasn't fixed the problem.
With the Qt reproducer from the original report, I still see ud2:
$ ~/dev/gcc/bin/g++ -I $QTOBJDIR/include -S -o - -O2 -g0 qstringview-repro.cpp
.file "qstringview-repro.cpp"
.text
.section
.text._ZN7Scanner12scanForSignsEv,"axG",@progbits,_ZN7Scanner12scanForSignsEv,comdat
.align 2
.p2align 4
.weak _ZN7Scanner12scanForSignsEv
.type _ZN7Scanner12scanForSignsEv, @function
_ZN7Scanner12scanForSignsEv:
.LFB7687:
.cfi_startproc
ud2
.cfi_endproc
You only need the Qt 6.11 headers for this, no need to link.
In any case, the AI created a new, C-only reproducer which still has the
problem:
$ cat r7.c
typedef struct { long size; const short *data; } SV;
typedef struct { SV given; } Scanner;
long findString(SV h, long from, SV n);
const short *strchr16(const short *b, const short *e, short c);
static inline long findChar(SV h, long from, short c)
{
if ((unsigned long)from < (unsigned long)h.size) {
const short *n = strchr16(h.data + from, h.data + h.size, c);
if (n != h.data + h.size)
return n - h.data;
}
return -1;
}
static inline long indexOf(const SV *h, SV n, long from)
{
if (__builtin_constant_p(n.size) && n.size == 1)
return findChar(*h, from, n.data[0]);
return findString(*h, from, n);
}
static int scanForToken(Scanner *s, SV sought)
{
long n = sought.size;
long idx = -n;
int count = 0;
while ((idx = indexOf(&s->given, sought, idx + n)) >= 0)
++count;
return count;
}
__attribute__((noinline)) int scanForSigns(Scanner *s)
{
static const short minus[] = { 0x2212 };
SV m = { 1, minus };
return scanForToken(s, m);
}
int f(SV h)
{
Scanner s = { h };
return scanForSigns(&s);
}
$ ~/dev/gcc/bin/gcc -S -o - -O2 r7.c
.file "r7.c"
.text
.p2align 4
.globl scanForSigns
.type scanForSigns, @function
scanForSigns:
.LFB3:
.cfi_startproc
ud2
.cfi_endproc
The insight from the AI (wrongly or not) was:
• The inlined char search has a bounds check, if ((unsigned long)from <
(unsigned long)h.size) . Removing it makes the test pass.
• The haystack is read through a pointer ( &s->given ).
