| Summary: | incorrect -ffast-math division simplification | ||
|---|---|---|---|
| Product: | gcc | Reporter: | Laurent GUERBY <laurent> |
| Component: | target | Assignee: | Not yet assigned to anyone <unassigned> |
| Status: | RESOLVED DUPLICATE | ||
| Severity: | normal | CC: | gcc-bugs, laurent |
| Priority: | P2 | ||
| Version: | 4.0.0 | ||
| Target Milestone: | --- | ||
| Host: | i686-pc-linux-gnu | Target: | i686-pc-linux-gnu |
| Build: | i686-pc-linux-gnu | Known to work: | |
| Known to fail: | Last reconfirmed: | ||
Simple C testcase:
float x = 30.0;
int main()
{
if ( 90.0/x != 3.0)
abort();
return 0;
}
It works on PPC but not on i686 so this is a dup of bug 323 and invalid.
*** This bug has been marked as a duplicate of 323 ***
Note what -ffast-math does is converts: 90.0/x into 1/x *90.0, so there is a precision loss which causes 1./0f/30.0 * 90.0 != 3.0. So this is still invalid. Ok I see, but then it's not really a dup of 323 which has an incorrect program, here the program is correct and just broken when we allow unsafe math optimizations :) Laurent |
The following trivial program fails with "-O1 -ffast-math" whereas it works without -ffast-math. This comes from ACATS. $ gnatmake -f -O1 -ffast-math c34003a gcc -c -O1 -ffast-math c34003a.adb gnatbind -x c34003a.ali gnatlink c34003a.ali -ffast-math $ ./c34003a raised PROGRAM_ERROR : c34003a.adb:5 explicit raise $ cat c34003a.adb PROCEDURE C34003A IS procedure Do_Fail is begin raise Program_Error; end Do_Fail; X : Float := 30.0; BEGIN IF 90.0 / X /= 3.0 THEN Do_Fail; END IF; END C34003A;