| 1 |
/* { dg-do compile } */ |
| 2 |
/* { dg-options "-Ofast -march=skylake-avx512 -mfpmath=sse -fno-tree-vectorize -fdump-tree-cddce3-details -fdump-tree-lsplit-optimized -fno-ssa-phiopt" } */ |
| 3 |
/* Load of d2/d3 is hoisted out, the loop is split, store of d1 and sqrt |
| 4 |
are sunk out of the loop and the loop is elided. One vsqrtsd with |
| 5 |
memory operand needs a xor to avoid partial dependence. */ |
| 6 |
/* Phi-OPT needs to be disabled otherwise, sqrt calls are merged which is better |
| 7 |
but we are testing to make sure the partial register stall for SSE is still avoided |
| 8 |
for sqrts. */ |
| 9 |
|
| 10 |
#include<math.h> |
| 11 |
|
| 12 |
extern double d1, d2, d3; |
| 13 |
void |
| 14 |
foo (int n, int k) |
| 15 |
{ |
| 16 |
for (int i = 0; i != n; i++) |
| 17 |
if(i < k) |
| 18 |
d1 = sqrt (d2); |
| 19 |
else |
| 20 |
d1 = sqrt (d3); |
| 21 |
} |
| 22 |
|
| 23 |
/* { dg-final { scan-tree-dump "optimized: loop split" "lsplit" } } */ |
| 24 |
/* { dg-final { scan-tree-dump-times "removing loop" 2 "cddce3" } } */ |
| 25 |
/* { dg-final { scan-assembler-times "vxorps\[^\n\r\]*xmm\[0-9\]" 1 } } */ |