Fix macOS/Clang build.
This commit is contained in:
parent
dd27c42293
commit
f163aad38c
3 changed files with 30 additions and 24 deletions
|
@ -45,8 +45,8 @@
|
|||
movq xmm0, rcx
|
||||
punpcklqdq xmm4, xmm0
|
||||
|
||||
ALIGN 64
|
||||
$main_loop_ryzen:
|
||||
ALIGN 16
|
||||
main_loop_ryzen:
|
||||
movdqa xmm5, XMMWORD PTR [r10+rbx]
|
||||
movq xmm0, r11
|
||||
movq xmm6, r8
|
||||
|
@ -103,10 +103,10 @@ $main_loop_ryzen:
|
|||
sqrtsd xmm1, xmm0
|
||||
movq rdi, xmm1
|
||||
test rdi, 524287
|
||||
je $sqrt_fixup_ryzen
|
||||
je sqrt_fixup_ryzen
|
||||
shr rdi, 19
|
||||
|
||||
$sqrt_fixup_ryzen_ret:
|
||||
sqrt_fixup_ryzen_ret:
|
||||
mov rax, rsi
|
||||
mul r14
|
||||
movq xmm1, rax
|
||||
|
@ -141,7 +141,7 @@ $sqrt_fixup_ryzen_ret:
|
|||
and r10d, 2097136
|
||||
movdqa xmm3, xmm5
|
||||
dec ebp
|
||||
jne $main_loop_ryzen
|
||||
jne main_loop_ryzen
|
||||
|
||||
ldmxcsr DWORD PTR [rsp]
|
||||
movaps xmm6, XMMWORD PTR [rsp+48]
|
||||
|
@ -157,9 +157,9 @@ $sqrt_fixup_ryzen_ret:
|
|||
pop r13
|
||||
pop r12
|
||||
pop rdi
|
||||
jmp $cnv2_main_loop_ryzen_endp
|
||||
jmp cnv2_main_loop_ryzen_endp
|
||||
|
||||
$sqrt_fixup_ryzen:
|
||||
sqrt_fixup_ryzen:
|
||||
movq r9, xmm2
|
||||
dec rdi
|
||||
mov edx, -1022
|
||||
|
@ -174,6 +174,6 @@ $sqrt_fixup_ryzen:
|
|||
imul rcx, rax
|
||||
sub rcx, r9
|
||||
adc rdi, 0
|
||||
jmp $sqrt_fixup_ryzen_ret
|
||||
jmp sqrt_fixup_ryzen_ret
|
||||
|
||||
$cnv2_main_loop_ryzen_endp:
|
||||
cnv2_main_loop_ryzen_endp:
|
||||
|
|
Loading…
Add table
Add a link
Reference in a new issue