diff options
Diffstat (limited to 'sysdeps/x86_64/multiarch/rawmemchr-evex.S')
| -rw-r--r-- | sysdeps/x86_64/multiarch/rawmemchr-evex.S | 6 |
1 files changed, 3 insertions, 3 deletions
diff --git a/sysdeps/x86_64/multiarch/rawmemchr-evex.S b/sysdeps/x86_64/multiarch/rawmemchr-evex.S index 52e6b186c8..0175a5e98c 100644 --- a/sysdeps/x86_64/multiarch/rawmemchr-evex.S +++ b/sysdeps/x86_64/multiarch/rawmemchr-evex.S @@ -187,13 +187,13 @@ L(loop_4x_vec): ymm0-15 is used at all is because there is no EVEX encoding vpcmpeq and with vpcmpeq this loop can be performed more efficiently. The non-vzeroupper version is safe for RTM - while the vzeroupper version should be prefered if RTM are + while the vzeroupper version should be preferred if RTM are not supported. Which loop version we use is determined by USE_TERN_IN_LOOP. */ # if USE_TERN_IN_LOOP /* Since vptern can only take 3x vectors fastest to do 1 vec - seperately with EVEX vpcmp. */ + separately with EVEX vpcmp. */ VPCMPEQ (VEC_SIZE * 4)(%rdi), %VMATCH, %k1 /* Compare 3x with vpcmpeq and or them all together with vptern. */ @@ -256,7 +256,7 @@ L(loop_4x_vec): (only if used VEX encoded loop). */ COND_VZEROUPPER - /* Seperate logic for VEC_SIZE == 64 and VEC_SIZE == 32 for + /* Separate logic for VEC_SIZE == 64 and VEC_SIZE == 32 for returning last 2x VEC. For VEC_SIZE == 64 we test each VEC individually, for VEC_SIZE == 32 we combine them in a single 64-bit GPR. */ |