http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54700
Bug #: 54700
Summary: Optimize away x<0 as mask argument of a blend.
Classification: Unclassified
Product: gcc
Version: 4.8.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
AssignedTo: unassig...@gcc.gnu.org
ReportedBy: gli...@gcc.gnu.org
Target: x86_64-linux-gnu
Hello,
blend instructions (vpblendvb, vblendvpd) only look at the high bit of the
elements of their mask. The mask may come from an operation, like v<0, which is
actually the identity on the high bits (for floats that may require
-fno-signed-zeros). In those cases, it would be nice to remove the useless
comparison.
Since the vectorizer is the only current producer of vec_cond_expr, here is an
example (compiled with -Ofast -mavx2):
#if 0
typedef double T;
#else
typedef signed char T;
#endif
void f(T* p){
for(int i=0;i<10000000;++i){
p[i]=(p[i]<0)?2:1;
}
}
(I haven't checked if there is a penalty for using a register last used for
floating point as mask)
Also, if vcond is passed constants to compare (say {-5,7}<{0,0}), it doesn't
get constant folded. I couldn't get the vectorizer to generate this (missed
optimization, using i%2 shouldn't prevent from vectorizing), so I don't have a
testcase for that one.
