https://gcc.gnu.org/bugzilla/show_bug.cgi?id=77980
Victor Tong <vitong at microsoft dot com> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |vitong at microsoft dot com
--- Comment #4 from Victor Tong <vitong at microsoft dot com> ---
Would a fix for this be more appropriate in the WPA analysis pass? It seems
that if x2 can be proved to be a constant, then all instances of x2 should be
replaced with 1 and t1 would get optimized to 0.
I tried writing a pattern in match.pd for 1/n but I found that the expression
looked like this:
x2.0_1 = x2;
_2 = x2.0_1 * 2;
_3 = 1 / _2;
There wasn't any information on x2 so gimple-match wasn't able to reduce the
expression.
When I alter the definition of x2 to be "const unsigned int" instead of
"unsigned int", the code generated by gcc(trunk) and clang(trunk) are
equivalent.
