Add save/restore xmm registers in x86 assembly code

Went through the code and fixed it. Verified on Windows.

Where possible, remove dependencies on xmm[67]

Current code relies on pushing rbp to the stack to get 16 byte
alignment. This broke when rbp wasn't pushed
(vp8/encoder/x86/sad_sse3.asm). Work around this by using unaligned
memory accesses. Revisit this and the offsets in
vp8/encoder/x86/sad_sse3.asm in another change to SAVE_XMM.

Change-Id: I5f940994d3ebfd977c3d68446cef20fd78b07877

diff --git a/vp8/common/x86/idctllm_sse2.asm b/vp8/common/x86/idctllm_sse2.asm
index edee157..c873869 100644 (file)
--- a/vp8/common/x86/idctllm_sse2.asm
+++ b/vp8/common/x86/idctllm_sse2.asm
@@ -32,9 +32,6 @@ sym(idct_dequant_0_2x_sse2):
         mov         rdx,            arg(1) ; dequant
         mov         rax,            arg(0) ; qcoeff
 
-    ; Zero out xmm7, for use unpacking
-        pxor        xmm7,           xmm7
-
         movd        xmm4,           [rax]
         movd        xmm5,           [rdx]
 
@@ -43,9 +40,12 @@ sym(idct_dequant_0_2x_sse2):
 
         pmullw      xmm4,           xmm5
 
+    ; Zero out xmm5, for use unpacking
+        pxor        xmm5,           xmm5
+
     ; clear coeffs
-        movd        [rax],          xmm7
-        movd        [rax+32],       xmm7
+        movd        [rax],          xmm5
+        movd        [rax+32],       xmm5
 ;pshufb
         pshuflw     xmm4,           xmm4,       00000000b
         pshufhw     xmm4,           xmm4,       00000000b
@@ -62,10 +62,10 @@ sym(idct_dequant_0_2x_sse2):
         lea         rcx,            [3*rcx]
         movq        xmm3,           [rax+rcx]
 
-        punpcklbw   xmm0,           xmm7
-        punpcklbw   xmm1,           xmm7
-        punpcklbw   xmm2,           xmm7
-        punpcklbw   xmm3,           xmm7
+        punpcklbw   xmm0,           xmm5
+        punpcklbw   xmm1,           xmm5
+        punpcklbw   xmm2,           xmm5
+        punpcklbw   xmm3,           xmm5
 
         mov         rax,            arg(3) ; dst
         movsxd      rdx,            dword ptr arg(4) ; dst_stride
@@ -77,10 +77,10 @@ sym(idct_dequant_0_2x_sse2):
         paddw       xmm3,           xmm4
 
     ; pack up before storing
-        packuswb    xmm0,           xmm7
-        packuswb    xmm1,           xmm7
-        packuswb    xmm2,           xmm7
-        packuswb    xmm3,           xmm7
+        packuswb    xmm0,           xmm5
+        packuswb    xmm1,           xmm5
+        packuswb    xmm2,           xmm5
+        packuswb    xmm3,           xmm5
 
     ; store blocks back out
         movq        [rax],          xmm0
@@ -102,6 +102,7 @@ sym(idct_dequant_full_2x_sse2):
     push        rbp
     mov         rbp, rsp
     SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM
     GET_GOT     rbx
     push        rsi
     push        rdi
@@ -347,6 +348,7 @@ sym(idct_dequant_full_2x_sse2):
     pop         rdi
     pop         rsi
     RESTORE_GOT
+    RESTORE_XMM
     UNSHADOW_ARGS
     pop         rbp
     ret
@@ -377,8 +379,8 @@ sym(idct_dequant_dc_0_2x_sse2):
         mov         rdi,            arg(3) ; dst
         mov         rdx,            arg(5) ; dc
 
-    ; Zero out xmm7, for use unpacking
-        pxor        xmm7,           xmm7
+    ; Zero out xmm5, for use unpacking
+        pxor        xmm5,           xmm5
 
     ; load up 2 dc words here == 2*16 = doubleword
         movd        xmm4,           [rdx]
@@ -398,10 +400,10 @@ sym(idct_dequant_dc_0_2x_sse2):
         psraw       xmm4,           3
 
     ; Predict buffer needs to be expanded from bytes to words
-        punpcklbw   xmm0,           xmm7
-        punpcklbw   xmm1,           xmm7
-        punpcklbw   xmm2,           xmm7
-        punpcklbw   xmm3,           xmm7
+        punpcklbw   xmm0,           xmm5
+        punpcklbw   xmm1,           xmm5
+        punpcklbw   xmm2,           xmm5
+        punpcklbw   xmm3,           xmm5
 
     ; Add to predict buffer
         paddw       xmm0,           xmm4
@@ -410,10 +412,10 @@ sym(idct_dequant_dc_0_2x_sse2):
         paddw       xmm3,           xmm4
 
     ; pack up before storing
-        packuswb    xmm0,           xmm7
-        packuswb    xmm1,           xmm7
-        packuswb    xmm2,           xmm7
-        packuswb    xmm3,           xmm7
+        packuswb    xmm0,           xmm5
+        packuswb    xmm1,           xmm5
+        packuswb    xmm2,           xmm5
+        packuswb    xmm3,           xmm5
 
     ; Load destination stride before writing out,
     ;   doesn't need to persist
@@ -441,6 +443,7 @@ sym(idct_dequant_dc_full_2x_sse2):
     push        rbp
     mov         rbp, rsp
     SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM
     GET_GOT     rbx
     push        rsi
     push        rdi
@@ -692,6 +695,7 @@ sym(idct_dequant_dc_full_2x_sse2):
     pop         rdi
     pop         rsi
     RESTORE_GOT
+    RESTORE_XMM
     UNSHADOW_ARGS
     pop         rbp
     ret

diff --git a/vp8/common/x86/subpixel_ssse3.asm b/vp8/common/x86/subpixel_ssse3.asm
index 0ec18de..1db3d62 100644 (file)
--- a/vp8/common/x86/subpixel_ssse3.asm
+++ b/vp8/common/x86/subpixel_ssse3.asm
@@ -39,6 +39,7 @@ sym(vp8_filter_block1d8_h6_ssse3):
     push        rbp
     mov         rbp, rsp
     SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM
     GET_GOT     rbx
     push        rsi
     push        rdi
@@ -107,6 +108,7 @@ filter_block1d8_h6_rowloop_ssse3:
     pop rdi
     pop rsi
     RESTORE_GOT
+    RESTORE_XMM
     UNSHADOW_ARGS
     pop         rbp
     ret
@@ -162,6 +164,7 @@ filter_block1d8_h4_rowloop_ssse3:
     pop rdi
     pop rsi
     RESTORE_GOT
+    RESTORE_XMM
     UNSHADOW_ARGS
     pop         rbp
     ret
@@ -286,6 +289,7 @@ sym(vp8_filter_block1d4_h6_ssse3):
     push        rbp
     mov         rbp, rsp
     SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM
     GET_GOT     rbx
     push        rsi
     push        rdi
@@ -393,6 +397,7 @@ filter_block1d4_h4_rowloop_ssse3:
     pop rdi
     pop rsi
     RESTORE_GOT
+    RESTORE_XMM
     UNSHADOW_ARGS
     pop         rbp
     ret
@@ -413,6 +418,7 @@ sym(vp8_filter_block1d16_v6_ssse3):
     push        rbp
     mov         rbp, rsp
     SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM
     GET_GOT     rbx
     push        rsi
     push        rdi
@@ -508,6 +514,7 @@ vp8_filter_block1d16_v6_ssse3_loop:
     pop rdi
     pop rsi
     RESTORE_GOT
+    RESTORE_XMM
     UNSHADOW_ARGS
     pop         rbp
     ret
@@ -580,6 +587,7 @@ vp8_filter_block1d16_v4_ssse3_loop:
     pop rdi
     pop rsi
     RESTORE_GOT
+    RESTORE_XMM
     UNSHADOW_ARGS
     pop         rbp
     ret
@@ -598,6 +606,7 @@ sym(vp8_filter_block1d8_v6_ssse3):
     push        rbp
     mov         rbp, rsp
     SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM
     GET_GOT     rbx
     push        rsi
     push        rdi
@@ -670,6 +679,7 @@ vp8_filter_block1d8_v6_ssse3_loop:
     pop rdi
     pop rsi
     RESTORE_GOT
+    RESTORE_XMM
     UNSHADOW_ARGS
     pop         rbp
     ret
@@ -718,6 +728,7 @@ vp8_filter_block1d8_v4_ssse3_loop:
     pop rdi
     pop rsi
     RESTORE_GOT
+    RESTORE_XMM
     UNSHADOW_ARGS
     pop         rbp
     ret
@@ -808,6 +819,7 @@ vp8_filter_block1d4_v6_ssse3_loop:
     pop rdi
     pop rsi
     RESTORE_GOT
+    RESTORE_XMM
     UNSHADOW_ARGS
     pop         rbp
     ret

diff --git a/vp8/encoder/x86/dct_sse2.asm b/vp8/encoder/x86/dct_sse2.asm
index 652dd98..287ad48 100644 (file)
--- a/vp8/encoder/x86/dct_sse2.asm
+++ b/vp8/encoder/x86/dct_sse2.asm
@@ -33,6 +33,7 @@
     %define     input       rcx
     %define     output      rdx
     %define     pitch       r8
+    SAVE_XMM
   %else
     %define     input       rdi
     %define     output      rsi
@@ -53,6 +54,7 @@
     pop         rbp
 %else
   %ifidn __OUTPUT_FORMAT__,x64
+    RESTORE_XMM
   %endif
 %endif
     ret

diff --git a/vp8/encoder/x86/encodeopt.asm b/vp8/encoder/x86/encodeopt.asm
index c0f06bb..e142a75 100644 (file)
--- a/vp8/encoder/x86/encodeopt.asm
+++ b/vp8/encoder/x86/encodeopt.asm
@@ -22,33 +22,33 @@ sym(vp8_block_error_xmm):
     ; end prologue
 
         mov         rsi,        arg(0) ;coeff_ptr
-
         mov         rdi,        arg(1) ;dcoef_ptr
-        movdqa      xmm3,       [rsi]
 
-        movdqa      xmm4,       [rdi]
-        movdqa      xmm5,       [rsi+16]
+        movdqa      xmm0,       [rsi]
+        movdqa      xmm1,       [rdi]
+
+        movdqa      xmm2,       [rsi+16]
+        movdqa      xmm3,       [rdi+16]
 
-        movdqa      xmm6,       [rdi+16]
-        psubw       xmm3,       xmm4
+        psubw       xmm0,       xmm1
+        psubw       xmm2,       xmm3
 
-        psubw       xmm5,       xmm6
-        pmaddwd     xmm3,       xmm3
-        pmaddwd     xmm5,       xmm5
+        pmaddwd     xmm0,       xmm0
+        pmaddwd     xmm2,       xmm2
 
-        paddd       xmm3,       xmm5
+        paddd       xmm0,       xmm2
 
-        pxor        xmm7,       xmm7
-        movdqa      xmm0,       xmm3
+        pxor        xmm5,       xmm5
+        movdqa      xmm1,       xmm0
 
-        punpckldq   xmm0,       xmm7
-        punpckhdq   xmm3,       xmm7
+        punpckldq   xmm0,       xmm5
+        punpckhdq   xmm1,       xmm5
 
-        paddd       xmm0,       xmm3
-        movdqa      xmm3,       xmm0
+        paddd       xmm0,       xmm1
+        movdqa      xmm1,       xmm0
 
         psrldq      xmm0,       8
-        paddd       xmm0,       xmm3
+        paddd       xmm0,       xmm1
 
         movq        rax,        xmm0
 
@@ -208,53 +208,54 @@ sym(vp8_mbblock_error_xmm_impl):
     push        rbp
     mov         rbp, rsp
     SHADOW_ARGS_TO_STACK 3
+    SAVE_XMM ; 6
     push rsi
     push rdi
     ; end prolog
 
 
         mov         rsi,        arg(0) ;coeff_ptr
-        pxor        xmm7,       xmm7
+        pxor        xmm6,       xmm6
 
         mov         rdi,        arg(1) ;dcoef_ptr
-        pxor        xmm2,       xmm2
+        pxor        xmm4,       xmm4
 
-        movd        xmm1,       dword ptr arg(2) ;dc
-        por         xmm1,       xmm2
+        movd        xmm5,       dword ptr arg(2) ;dc
+        por         xmm5,       xmm4
 
-        pcmpeqw     xmm1,       xmm7
+        pcmpeqw     xmm5,       xmm6
         mov         rcx,        16
 
 mberror_loop:
-        movdqa      xmm3,       [rsi]
-        movdqa      xmm4,       [rdi]
+        movdqa      xmm0,       [rsi]
+        movdqa      xmm1,       [rdi]
 
-        movdqa      xmm5,       [rsi+16]
-        movdqa      xmm6,       [rdi+16]
+        movdqa      xmm2,       [rsi+16]
+        movdqa      xmm3,       [rdi+16]
 
 
-        psubw       xmm5,       xmm6
-        pmaddwd     xmm5,       xmm5
+        psubw       xmm2,       xmm3
+        pmaddwd     xmm2,       xmm2
 
-        psubw       xmm3,       xmm4
-        pand        xmm3,       xmm1
+        psubw       xmm0,       xmm1
+        pand        xmm0,       xmm5
 
-        pmaddwd     xmm3,       xmm3
+        pmaddwd     xmm0,       xmm0
         add         rsi,        32
 
         add         rdi,        32
 
         sub         rcx,        1
-        paddd       xmm2,       xmm5
+        paddd       xmm4,       xmm2
 
-        paddd       xmm2,       xmm3
+        paddd       xmm4,       xmm0
         jnz         mberror_loop
 
-        movdqa      xmm0,       xmm2
-        punpckldq   xmm0,       xmm7
+        movdqa      xmm0,       xmm4
+        punpckldq   xmm0,       xmm6
 
-        punpckhdq   xmm2,       xmm7
-        paddd       xmm0,       xmm2
+        punpckhdq   xmm4,       xmm6
+        paddd       xmm0,       xmm4
 
         movdqa      xmm1,       xmm0
         psrldq      xmm0,       8
@@ -265,6 +266,7 @@ mberror_loop:
     pop rdi
     pop rsi
     ; begin epilog
+    RESTORE_XMM
     UNSHADOW_ARGS
     pop         rbp
     ret
@@ -342,7 +344,7 @@ sym(vp8_mbuverror_xmm_impl):
         mov             rdi,        arg(1) ;d_ptr
 
         mov             rcx,        16
-        pxor            xmm7,       xmm7
+        pxor            xmm3,       xmm3
 
 mbuverror_loop:
 
@@ -352,7 +354,7 @@ mbuverror_loop:
         psubw           xmm1,       xmm2
         pmaddwd         xmm1,       xmm1
 
-        paddd           xmm7,       xmm1
+        paddd           xmm3,       xmm1
 
         add             rsi,        16
         add             rdi,        16
@@ -361,7 +363,7 @@ mbuverror_loop:
         jnz             mbuverror_loop
 
         pxor        xmm0,           xmm0
-        movdqa      xmm1,           xmm7
+        movdqa      xmm1,           xmm3
 
         movdqa      xmm2,           xmm1
         punpckldq   xmm1,           xmm0

diff --git a/vp8/encoder/x86/sad_sse2.asm b/vp8/encoder/x86/sad_sse2.asm
index cc6bc3c..d9ac3ff 100644 (file)
--- a/vp8/encoder/x86/sad_sse2.asm
+++ b/vp8/encoder/x86/sad_sse2.asm
@@ -21,6 +21,7 @@ sym(vp8_sad16x16_wmt):
     push        rbp
     mov         rbp, rsp
     SHADOW_ARGS_TO_STACK 4
+    SAVE_XMM ; 6
     push        rsi
     push        rdi
     ; end prolog
@@ -34,7 +35,7 @@ sym(vp8_sad16x16_wmt):
         lea             rcx,        [rsi+rax*8]
 
         lea             rcx,        [rcx+rax*8]
-        pxor            xmm7,       xmm7
+        pxor            xmm6,       xmm6
 
 x16x16sad_wmt_loop:
 
@@ -52,32 +53,33 @@ x16x16sad_wmt_loop:
         punpcklbw       xmm1,       xmm3
 
         psadbw          xmm0,       xmm1
-        movq            xmm6,       QWORD PTR [rsi+rax+8]
+        movq            xmm2,       QWORD PTR [rsi+rax+8]
 
         movq            xmm3,       QWORD PTR [rdi+rdx+8]
         lea             rsi,        [rsi+rax*2]
 
         lea             rdi,        [rdi+rdx*2]
-        punpcklbw       xmm4,       xmm6
+        punpcklbw       xmm4,       xmm2
 
         punpcklbw       xmm5,       xmm3
         psadbw          xmm4,       xmm5
 
-        paddw           xmm7,       xmm0
-        paddw           xmm7,       xmm4
+        paddw           xmm6,       xmm0
+        paddw           xmm6,       xmm4
 
         cmp             rsi,        rcx
         jne             x16x16sad_wmt_loop
 
-        movq            xmm0,       xmm7
-        psrldq          xmm7,       8
+        movq            xmm0,       xmm6
+        psrldq          xmm6,       8
 
-        paddw           xmm0,       xmm7
+        paddw           xmm0,       xmm6
         movq            rax,        xmm0
 
     ; begin epilog
     pop rdi
     pop rsi
+    RESTORE_XMM
     UNSHADOW_ARGS
     pop         rbp
     ret

diff --git a/vp8/encoder/x86/sad_sse3.asm b/vp8/encoder/x86/sad_sse3.asm
index f0336ab..6668792 100644 (file)
--- a/vp8/encoder/x86/sad_sse3.asm
+++ b/vp8/encoder/x86/sad_sse3.asm
@@ -39,8 +39,9 @@
     %define     ref_stride  r9
     %define     end_ptr     r10
     %define     ret_var     r11
-    %define     result_ptr  [rsp+8+4*8]
-    %define     max_err     [rsp+8+4*8]
+    %define     result_ptr  [rsp+40+4*8]
+    %define     max_err     [rsp+40+4*8]
+    SAVE_XMM
   %else
     %define     src_ptr     rdi
     %define     src_stride  rsi
@@ -72,6 +73,7 @@
     pop         rbp
 %else
   %ifidn __OUTPUT_FORMAT__,x64
+    RESTORE_XMM
   %endif
 %endif
     ret
@@ -113,7 +115,8 @@
     %define     r2_ptr      r11
     %define     r3_ptr      r8
     %define     ref_stride  r9
-    %define     result_ptr  [rsp+16+4*8]
+    %define     result_ptr  [rsp+48+4*8]
+    SAVE_XMM
     push        rsi
 
     LOAD_X4_ADDRESSES r8, r0_ptr, r1_ptr, r2_ptr, r3_ptr
@@ -151,6 +154,7 @@
 %else
   %ifidn __OUTPUT_FORMAT__,x64
     pop         rsi
+    RESTORE_XMM
   %endif
 %endif
     ret

diff --git a/vp8/encoder/x86/sad_ssse3.asm b/vp8/encoder/x86/sad_ssse3.asm
index 69c5eae..7c7cd0a 100644 (file)
--- a/vp8/encoder/x86/sad_ssse3.asm
+++ b/vp8/encoder/x86/sad_ssse3.asm
@@ -157,6 +157,7 @@ sym(vp8_sad16x16x3_ssse3):
     push        rbp
     mov         rbp, rsp
     SHADOW_ARGS_TO_STACK 5
+    SAVE_XMM
     push        rsi
     push        rdi
     push        rcx
@@ -253,6 +254,7 @@ vp8_sad16x16x3_ssse3_store_off:
     pop         rcx
     pop         rdi
     pop         rsi
+    RESTORE_XMM
     UNSHADOW_ARGS
     pop         rbp
     ret
@@ -268,6 +270,7 @@ sym(vp8_sad16x8x3_ssse3):
     push        rbp
     mov         rbp, rsp
     SHADOW_ARGS_TO_STACK 5
+    SAVE_XMM
     push        rsi
     push        rdi
     push        rcx
@@ -361,6 +364,7 @@ vp8_sad16x8x3_ssse3_store_off:
     pop         rcx
     pop         rdi
     pop         rsi
+    RESTORE_XMM
     UNSHADOW_ARGS
     pop         rbp
     ret

diff --git a/vp8/encoder/x86/variance_impl_sse2.asm b/vp8/encoder/x86/variance_impl_sse2.asm
index c2c30de..2c0e170 100644 (file)
--- a/vp8/encoder/x86/variance_impl_sse2.asm
+++ b/vp8/encoder/x86/variance_impl_sse2.asm
@@ -85,6 +85,7 @@ sym(vp8_get16x16var_sse2):
     push        rbp
     mov         rbp, rsp
     SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM
     push rbx
     push rsi
     push rdi
@@ -206,6 +207,7 @@ var16loop:
     pop rdi
     pop rsi
     pop rbx
+    RESTORE_XMM
     UNSHADOW_ARGS
     pop         rbp
     ret
@@ -223,6 +225,7 @@ sym(vp8_get16x16pred_error_sse2):
     push        rbp
     mov         rbp, rsp
     SHADOW_ARGS_TO_STACK 4
+    SAVE_XMM
     GET_GOT     rbx
     push rsi
     push rdi
@@ -321,6 +324,7 @@ var16peloop:
     pop rdi
     pop rsi
     RESTORE_GOT
+    RESTORE_XMM
     UNSHADOW_ARGS
     pop         rbp
     ret
@@ -341,6 +345,7 @@ sym(vp8_get8x8var_sse2):
     push        rbp
     mov         rbp, rsp
     SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM
     GET_GOT     rbx
     push rsi
     push rdi
@@ -506,6 +511,7 @@ sym(vp8_get8x8var_sse2):
     pop rdi
     pop rsi
     RESTORE_GOT
+    RESTORE_XMM
     UNSHADOW_ARGS
     pop         rbp
     ret
@@ -805,6 +811,7 @@ sym(vp8_half_horiz_vert_variance8x_h_sse2):
     push        rbp
     mov         rbp, rsp
     SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM
     GET_GOT     rbx
     push rsi
     push rdi
@@ -906,6 +913,7 @@ vp8_half_horiz_vert_variance8x_h_1:
     pop rdi
     pop rsi
     RESTORE_GOT
+    RESTORE_XMM
     UNSHADOW_ARGS
     pop         rbp
     ret
@@ -1041,6 +1049,7 @@ sym(vp8_half_vert_variance8x_h_sse2):
     push        rbp
     mov         rbp, rsp
     SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM
     GET_GOT     rbx
     push rsi
     push rdi
@@ -1127,6 +1136,7 @@ vp8_half_vert_variance8x_h_1:
     pop rdi
     pop rsi
     RESTORE_GOT
+    RESTORE_XMM
     UNSHADOW_ARGS
     pop         rbp
     ret
@@ -1254,6 +1264,7 @@ sym(vp8_half_horiz_variance8x_h_sse2):
     push        rbp
     mov         rbp, rsp
     SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM
     GET_GOT     rbx
     push rsi
     push rdi
@@ -1338,6 +1349,7 @@ vp8_half_horiz_variance8x_h_1:
     pop rdi
     pop rsi
     RESTORE_GOT
+    RESTORE_XMM
     UNSHADOW_ARGS
     pop         rbp
     ret

diff --git a/vpx_ports/x86_abi_support.asm b/vpx_ports/x86_abi_support.asm
index be64cd7..60dff49 100644 (file)
--- a/vpx_ports/x86_abi_support.asm
+++ b/vpx_ports/x86_abi_support.asm
@@ -260,12 +260,12 @@
 %ifidn __OUTPUT_FORMAT__,x64
 %macro SAVE_XMM 0
   sub rsp, 32
-  movdqa XMMWORD PTR [rsp], xmm6
-  movdqa XMMWORD PTR [rsp+16], xmm7
+  movdqu XMMWORD PTR [rsp], xmm6
+  movdqu XMMWORD PTR [rsp+16], xmm7
 %endmacro
 %macro RESTORE_XMM 0
-  movdqa xmm6, XMMWORD PTR [rsp]
-  movdqa xmm7, XMMWORD PTR [rsp+16]
+  movdqu xmm6, XMMWORD PTR [rsp]
+  movdqu xmm7, XMMWORD PTR [rsp+16]
   add rsp, 32
 %endmacro
 %else