| Summary: | [5 Regression] curve25519-donna-c64 miscompilation | ||
|---|---|---|---|
| Product: | gcc | Reporter: | Sami Farin <hvtaifwkbgefbaei> |
| Component: | tree-optimization | Assignee: | Jakub Jelinek <jakub> |
| Status: | RESOLVED FIXED | ||
| Severity: | major | CC: | jakub |
| Priority: | P1 | ||
| Version: | 5.0 | ||
| Target Milestone: | 5.0 | ||
| Host: | Target: | ||
| Build: | Known to work: | ||
| Known to fail: | Last reconfirmed: | 2015-02-23 00:00:00 | |
| Attachments: |
curve25519-donna-c64.c
gcc5-pr65170.patch |
||
Started with r220164. So, either the VR info before vrp2 pass is already wrong, or something goes wrong during the vrp2 pass. Note the testcase uses TImode heavily. I believe this boils down to
__attribute__((noinline, noclone)) void
foo (__uint128_t b, __uint128_t c)
{
__uint128_t a;
b &= -1ULL;
c &= -1ULL;
a = b * c;
if (a != 1)
__builtin_abort ();
}
int
main ()
{
foo (1, 1);
return 0;
}
which at -O2 regressed with r210113 aka the wide-int merge. Will look at it.
Created attachment 34844 [details] gcc5-pr65170.patch Untested fix. Author: jakub Date: Tue Feb 24 08:07:10 2015 New Revision: 220931 URL: https://gcc.gnu.org/viewcvs?rev=220931&root=gcc&view=rev Log: PR tree-optimization/65170 * wide-int.cc (wi::mul_internal): For the umul_ppmm optimization, if val[1] < 0, clear also val[2] and return 3. * gcc.c-torture/execute/pr65170.c: New test. * gcc.dg/tree-ssa/vrp96.c: New test. Added: trunk/gcc/testsuite/gcc.c-torture/execute/pr65170.c trunk/gcc/testsuite/gcc.dg/tree-ssa/vrp96.c Modified: trunk/gcc/ChangeLog trunk/gcc/testsuite/ChangeLog trunk/gcc/wide-int.cc trunk/libsanitizer/ChangeLog Fixed now. |
Created attachment 34841 [details] curve25519-donna-c64.c curve25519-donna-c64.c function curve25519_donna with gcc-5.0.0-0.15.fc22, flag -O3 , sets mypublic to all-zeros. If I remove force_inline from fmul(), it compiles ok. Could not produce a smaller test case.