https://gcc.gnu.org/bugzilla/show_bug.cgi?id=120356
Alexey Merzlyakov <alexey.merzlyakov at samsung dot com> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |alexey.merzlyakov at samsung
dot c
| |om
--- Comment #4 from Alexey Merzlyakov <alexey.merzlyakov at samsung dot com> ---
Confirm the problem appears since
https://gcc.gnu.org/cgit/gcc/commit/?id=7b815107f40 and un-hides by
https://gcc.gnu.org/cgit/gcc/commit/?id=28b2ad5341f which fixes ICE then. These
commits are seem to uncover the issue but are not the cause of the problem.
Reduced test-case which does not require -flto to ease debug process:
unsigned char a = 5;
long long c[18];
static void d () {
for (short i = 0; i < 60; i += 65413)
for (char j = 0; j < 18; j++) {
for (char k = 0; k < 18; k++) {
a *= 143;
}
for (char k = 0; k < 6; k++)
for (char l = 0; l < 18; l++)
c[l] = 0;
}
}
int main() {
d();
__builtin_printf("%llu\n", a + c[0]);
}
It works on -O2 and above.
>From the first sight, it generates the vector calculation chain in the
beginning of the function, like this:
vid.v v1
li a3,-113
vsll.vi v2,v1,1
vand.vi v1,v1,1
vsll.vi v3,v2,5
vmseq.vi v0,v1,1
vadd.vx v3,v3,a3
li a3,-32
vmul.vx v2,v2,a3
li a3,-31
vadd.vx v2,v2,a3
vmerge.vvm v2,v3,v2,v0
However, compiler inserts it again in the loop body when d() is being inlined,
which leads to incorrect "5" result. Further investigations are in progress on
it.
