shithub: dav1d

Download patch

ref: 3a6a55d8d5883cf094bc3a140e568420014bbf47
parent: fb348f64dc0bcd4c0aaf686afbdd810484a5536f
author: Henrik Gramner <gramner@twoorioles.com>
date: Wed Mar 4 18:05:45 EST 2020 

x86: Fix crash in AVX2 cdef_filter with <32-byte stack alignment


--- a/src/x86/cdef.asm
+++ b/src/x86/cdef.asm
@@ -446,7 +446,7 @@
     PUSH           r11
 %if %2 == 4
  %assign regs_used 12
- %if WIN64
+ %if STACK_ALIGNMENT < 32
     PUSH  r%+regs_used
   %assign regs_used regs_used+1
  %endif
@@ -458,24 +458,24 @@
     movu    [rsp+0x10], m0
     movu    [rsp+0x28], m1
     movu    [rsp+0x40], m2
-%else
+%elif %1 == 4
     PUSH           r12
- %if %1 == 4
-  %assign regs_used 13
-  %if WIN64
+ %assign regs_used 13
+ %if STACK_ALIGNMENT < 32
     PUSH  r%+regs_used
    %assign regs_used regs_used+1
-  %endif
+ %endif
     ALLOC_STACK 8*2+%1*%2*1, 16
     pmovzxwd        m0, [leftq]
     mova    [rsp+0x10], m0
- %else
+%else
+    PUSH           r12
     PUSH           r13
-  %assign regs_used 14
-  %if WIN64
+ %assign regs_used 14
+ %if STACK_ALIGNMENT < 32
     PUSH  r%+regs_used
-   %assign regs_used regs_used+1
-  %endif
+  %assign regs_used regs_used+1
+ %endif
     ALLOC_STACK 8*2+%1*%2*2+32, 16
     lea            r11, [strideq*3]
     movu           xm4, [dstq+strideq*2]
@@ -489,7 +489,6 @@
     mova    [rsp+0x50], m2
     mova    [rsp+0x70], m3
     mova    [rsp+0x90], m4
- %endif
 %endif
 
  DEFINE_ARGS dst, stride, left, top, pri, secdmp, zero, pridmp, damping
@@ -1255,7 +1254,7 @@
 %else
  %assign regs_used 9
 %endif
-%if WIN64
+%if STACK_ALIGNMENT < 32
     PUSH  r%+regs_used
  %assign regs_used regs_used+1
 %endif