some peepholes that should match horizontal add/sub operations.
llvm-svn: 163103
This commit is contained in:
parent
30c0cb2f30
commit
ba3ba8fa1f
|
@ -941,3 +941,15 @@ and inversion with an rsqrtss instruction, which computes 1/sqrt faster at the
|
|||
cost of reduced accuracy.
|
||||
|
||||
//===---------------------------------------------------------------------===//
|
||||
|
||||
This function should be matched to haddpd when the appropriate CPU is enabled:
|
||||
|
||||
#include <x86intrin.h>
|
||||
double f (__m128d p) {
|
||||
return p[0] + p[1];
|
||||
}
|
||||
|
||||
similarly, v[0]-v[1] should match to hsubpd, and {v[0]-v[1], w[0]-w[1]} should
|
||||
turn into hsubpd also.
|
||||
|
||||
//===---------------------------------------------------------------------===//
|
||||
|
|
Loading…
Reference in New Issue