x86inc: support stack mem allocation and re-alignment in PROLOGUE

Use this in VP8/H264-8bit loopfilter functions so they can be used if
there is no aligned stack (e.g. MSVC 32bit or ICC 10.x).

Signed-off-by: Luca Barbato <lu_zero@gentoo.org>

5 files changed, 191 insertions, 101 deletions

diff --git a/libavcodec/x86/h264_deblock.asm b/libavcodec/x86/h264_deblock.asm
index c124c4daa0..1f1dbc6a2f 100644
--- a/libavcodec/x86/h264_deblock.asm
+++ b/libavcodec/x86/h264_deblock.asm
@@ -398,14 +398,12 @@ DEBLOCK_LUMA
 ;-----------------------------------------------------------------------------
 ; void deblock_v8_luma( uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0 )
 ;-----------------------------------------------------------------------------
-cglobal deblock_%1_luma_8, 5,5
+cglobal deblock_%1_luma_8, 5,5,8,2*%2
     lea     r4, [r1*3]
     dec     r2     ; alpha-1
     neg     r4
     dec     r3     ; beta-1
     add     r4, r0 ; pix-3*stride
-    %assign pad 2*%2+12-(stack_offset&15)
-    SUB     esp, pad
 
     mova    m0, [r4+r1]   ; p1
     mova    m1, [r4+2*r1] ; p0
@@ -443,22 +441,19 @@ cglobal deblock_%1_luma_8, 5,5
     DEBLOCK_P0_Q0
     mova    [r4+2*r1], m1
     mova    [r0], m2
-    ADD     esp, pad
     RET
 
 ;-----------------------------------------------------------------------------
 ; void deblock_h_luma( uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0 )
 ;-----------------------------------------------------------------------------
 INIT_MMX cpuname
-cglobal deblock_h_luma_8, 0,5
+cglobal deblock_h_luma_8, 0,5,8,0x60+HAVE_ALIGNED_STACK*12
     mov    r0, r0mp
     mov    r3, r1m
     lea    r4, [r3*3]
     sub    r0, 4
     lea    r1, [r0+r4]
-    %assign pad 0x78-(stack_offset&15)
-    SUB    esp, pad
-%define pix_tmp esp+12
+%define pix_tmp esp+12*HAVE_ALIGNED_STACK
 
     ; transpose 6x16 -> tmp space
     TRANSPOSE6x8_MEM  PASS8ROWS(r0, r1, r3, r4), pix_tmp
@@ -500,7 +495,6 @@ cglobal deblock_h_luma_8, 0,5
     movq   m3, [pix_tmp+0x48]
     TRANSPOSE8x4B_STORE  PASS8ROWS(r0, r1, r3, r4)
 
-    ADD    esp, pad
     RET
 %endmacro ; DEBLOCK_LUMA
 
@@ -631,7 +625,7 @@ DEBLOCK_LUMA v, 16
     %define mpb_0 m14
     %define mpb_1 m15
 %else
-    %define spill(x) [esp+16*x+((stack_offset+4)&15)]
+    %define spill(x) [esp+16*x]
     %define p2 [r4+r1]
     %define q2 [r0+2*r1]
     %define t4 spill(0)
@@ -646,10 +640,7 @@ DEBLOCK_LUMA v, 16
 ;-----------------------------------------------------------------------------
 ; void deblock_v_luma_intra( uint8_t *pix, int stride, int alpha, int beta )
 ;-----------------------------------------------------------------------------
-cglobal deblock_%1_luma_intra_8, 4,6,16
-%if ARCH_X86_64 == 0
-    sub     esp, 0x60
-%endif
+cglobal deblock_%1_luma_intra_8, 4,6,16,ARCH_X86_64*0x50-0x50
     lea     r4, [r1*4]
     lea     r5, [r1*3] ; 3*stride
     dec     r2d        ; alpha-1
@@ -698,9 +689,6 @@ cglobal deblock_%1_luma_intra_8, 4,6,16
     LUMA_INTRA_SWAP_PQ
     LUMA_INTRA_P012 [r0], [r0+r1], [r0+2*r1], [r0+r5]
 .end:
-%if ARCH_X86_64 == 0
-    add     esp, 0x60
-%endif
     RET
 
 INIT_MMX cpuname
@@ -737,12 +725,10 @@ cglobal deblock_h_luma_intra_8, 4,9
     add    rsp, 0x88
     RET
 %else
-cglobal deblock_h_luma_intra_8, 2,4
+cglobal deblock_h_luma_intra_8, 2,4,8,0x80
     lea    r3,  [r1*3]
     sub    r0,  4
     lea    r2,  [r0+r3]
-%assign pad 0x8c-(stack_offset&15)
-    SUB    rsp, pad
     %define pix_tmp rsp
 
     ; transpose 8x16 -> tmp space
@@ -773,7 +759,6 @@ cglobal deblock_h_luma_intra_8, 2,4
     lea    r0,  [r0+r1*8]
     lea    r2,  [r2+r1*8]
     TRANSPOSE8x8_MEM  PASS8ROWS(pix_tmp+8, pix_tmp+0x38, 0x10, 0x30), PASS8ROWS(r0, r2, r1, r3)
-    ADD    rsp, pad
     RET
 %endif ; ARCH_X86_64
 %endmacro ; DEBLOCK_LUMA_INTRA
diff --git a/libavcodec/x86/h264dsp_init.c b/libavcodec/x86/h264dsp_init.c
index ac231cbb08..73d4990899 100644
--- a/libavcodec/x86/h264dsp_init.c
+++ b/libavcodec/x86/h264dsp_init.c
@@ -275,18 +275,16 @@ void ff_h264dsp_init_x86(H264DSPContext *c, const int bit_depth,
                     c->biweight_h264_pixels_tab[0] = ff_h264_biweight_16_sse2;
                     c->biweight_h264_pixels_tab[1] = ff_h264_biweight_8_sse2;
 
-#if HAVE_ALIGNED_STACK
                     c->h264_v_loop_filter_luma       = ff_deblock_v_luma_8_sse2;
                     c->h264_h_loop_filter_luma       = ff_deblock_h_luma_8_sse2;
                     c->h264_v_loop_filter_luma_intra = ff_deblock_v_luma_intra_8_sse2;
                     c->h264_h_loop_filter_luma_intra = ff_deblock_h_luma_intra_8_sse2;
-#endif /* HAVE_ALIGNED_STACK */
                 }
                 if (EXTERNAL_SSSE3(mm_flags)) {
                     c->biweight_h264_pixels_tab[0] = ff_h264_biweight_16_ssse3;
                     c->biweight_h264_pixels_tab[1] = ff_h264_biweight_8_ssse3;
                 }
-                if (EXTERNAL_AVX(mm_flags) && HAVE_ALIGNED_STACK) {
+                if (EXTERNAL_AVX(mm_flags)) {
                     c->h264_v_loop_filter_luma       = ff_deblock_v_luma_8_avx;
                     c->h264_h_loop_filter_luma       = ff_deblock_h_luma_8_avx;
                     c->h264_v_loop_filter_luma_intra = ff_deblock_v_luma_intra_8_avx;
diff --git a/libavcodec/x86/vp8dsp.asm b/libavcodec/x86/vp8dsp.asm
index d732bf45eb..af8403e991 100644
--- a/libavcodec/x86/vp8dsp.asm
+++ b/libavcodec/x86/vp8dsp.asm
@@ -1631,28 +1631,31 @@ SIMPLE_LOOPFILTER h, 5
 ;-----------------------------------------------------------------------------
 
 %macro INNER_LOOPFILTER 2
+%define stack_size 0
+%ifndef m8   ; stack layout: [0]=E, [1]=I, [2]=hev_thr
+%ifidn %1, v ;               [3]=hev() result
+%define stack_size mmsize * -4
+%else ; h    ; extra storage space for transposes
+%define stack_size mmsize * -5
+%endif
+%endif
+
 %if %2 == 8 ; chroma
-cglobal vp8_%1_loop_filter8uv_inner, 6, 6, 13, dst, dst8, stride, flimE, flimI, hevthr
+cglobal vp8_%1_loop_filter8uv_inner, 6, 6, 13, stack_size, dst, dst8, stride, flimE, flimI, hevthr
 %else ; luma
-cglobal vp8_%1_loop_filter16y_inner, 5, 5, 13, dst, stride, flimE, flimI, hevthr
+cglobal vp8_%1_loop_filter16y_inner, 5, 5, 13, stack_size, dst, stride, flimE, flimI, hevthr
 %endif
 
 %if cpuflag(ssse3)
     pxor             m7, m7
 %endif
-%ifndef m8   ; stack layout: [0]=E, [1]=I, [2]=hev_thr
-%ifidn %1, v ;               [3]=hev() result
-%assign pad 16 + mmsize * 4 - gprsize - (stack_offset & 15)
-%else ; h    ; extra storage space for transposes
-%assign pad 16 + mmsize * 5 - gprsize - (stack_offset & 15)
-%endif
+
+%ifndef m8
     ; splat function arguments
     SPLATB_REG       m0, flimEq, m7   ; E
     SPLATB_REG       m1, flimIq, m7   ; I
     SPLATB_REG       m2, hevthrq, m7  ; hev_thresh
 
-    SUB             rsp, pad
-
 %define m_flimE    [rsp]
 %define m_flimI    [rsp+mmsize]
 %define m_hevthr   [rsp+mmsize*2]
@@ -2082,12 +2085,10 @@ cglobal vp8_%1_loop_filter16y_inner, 5, 5, 13, dst, stride, flimE, flimI, hevthr
     dec           cntrq
     jg .next8px
 %endif
-%endif
-
-%ifndef m8 ; sse2 on x86-32 or mmx/mmxext
-    ADD             rsp, pad
-%endif
+    REP_RET
+%else ; mmsize == 16
     RET
+%endif
 %endmacro
 
 %if ARCH_X86_32
@@ -2122,31 +2123,34 @@ INNER_LOOPFILTER h,  8
 ;-----------------------------------------------------------------------------
 
 %macro MBEDGE_LOOPFILTER 2
-%if %2 == 8 ; chroma
-cglobal vp8_%1_loop_filter8uv_mbedge, 6, 6, 15, dst1, dst8, stride, flimE, flimI, hevthr
-%else ; luma
-cglobal vp8_%1_loop_filter16y_mbedge, 5, 5, 15, dst1, stride, flimE, flimI, hevthr
-%endif
-
-%if cpuflag(ssse3)
-    pxor             m7, m7
-%endif
+%define stack_size 0
 %ifndef m8       ; stack layout: [0]=E, [1]=I, [2]=hev_thr
 %if mmsize == 16 ;               [3]=hev() result
                  ;               [4]=filter tmp result
                  ;               [5]/[6] = p2/q2 backup
                  ;               [7]=lim_res sign result
-%assign pad 16 + mmsize * 7 - gprsize - (stack_offset & 15)
+%define stack_size mmsize * -7
 %else ; 8        ; extra storage space for transposes
-%assign pad 16 + mmsize * 8 - gprsize - (stack_offset & 15)
+%define stack_size mmsize * -8
+%endif
 %endif
+
+%if %2 == 8 ; chroma
+cglobal vp8_%1_loop_filter8uv_mbedge, 6, 6, 15, stack_size, dst1, dst8, stride, flimE, flimI, hevthr
+%else ; luma
+cglobal vp8_%1_loop_filter16y_mbedge, 5, 5, 15, stack_size, dst1, stride, flimE, flimI, hevthr
+%endif
+
+%if cpuflag(ssse3)
+    pxor             m7, m7
+%endif
+
+%ifndef m8
     ; splat function arguments
     SPLATB_REG       m0, flimEq, m7   ; E
     SPLATB_REG       m1, flimIq, m7   ; I
     SPLATB_REG       m2, hevthrq, m7  ; hev_thresh
 
-    SUB             rsp, pad
-
 %define m_flimE    [rsp]
 %define m_flimI    [rsp+mmsize]
 %define m_hevthr   [rsp+mmsize*2]
@@ -2740,12 +2744,10 @@ cglobal vp8_%1_loop_filter16y_mbedge, 5, 5, 15, dst1, stride, flimE, flimI, hevt
     dec          cntrq
     jg .next8px
 %endif
-%endif
-
-%ifndef m8 ; sse2 on x86-32 or mmx/mmxext
-    ADD            rsp, pad
-%endif
+    REP_RET
+%else ; mmsize == 16
     RET
+%endif
 %endmacro
 
 %if ARCH_X86_32
diff --git a/libavcodec/x86/vp8dsp_init.c b/libavcodec/x86/vp8dsp_init.c
index bf5463f080..a9d3541807 100644
--- a/libavcodec/x86/vp8dsp_init.c
+++ b/libavcodec/x86/vp8dsp_init.c
@@ -390,13 +390,11 @@ av_cold void ff_vp8dsp_init_x86(VP8DSPContext* c)
 
         c->vp8_v_loop_filter_simple = ff_vp8_v_loop_filter_simple_sse2;
 
-#if ARCH_X86_64 || HAVE_ALIGNED_STACK
         c->vp8_v_loop_filter16y_inner = ff_vp8_v_loop_filter16y_inner_sse2;
         c->vp8_v_loop_filter8uv_inner = ff_vp8_v_loop_filter8uv_inner_sse2;
 
         c->vp8_v_loop_filter16y       = ff_vp8_v_loop_filter16y_mbedge_sse2;
         c->vp8_v_loop_filter8uv       = ff_vp8_v_loop_filter8uv_mbedge_sse2;
-#endif
     }
 
     if (mm_flags & AV_CPU_FLAG_SSE2) {
@@ -404,13 +402,11 @@ av_cold void ff_vp8dsp_init_x86(VP8DSPContext* c)
 
         c->vp8_h_loop_filter_simple = ff_vp8_h_loop_filter_simple_sse2;
 
-#if ARCH_X86_64 || HAVE_ALIGNED_STACK
         c->vp8_h_loop_filter16y_inner = ff_vp8_h_loop_filter16y_inner_sse2;
         c->vp8_h_loop_filter8uv_inner = ff_vp8_h_loop_filter8uv_inner_sse2;
 
         c->vp8_h_loop_filter16y       = ff_vp8_h_loop_filter16y_mbedge_sse2;
         c->vp8_h_loop_filter8uv       = ff_vp8_h_loop_filter8uv_mbedge_sse2;
-#endif
     }
 
     if (mm_flags & AV_CPU_FLAG_SSSE3) {
@@ -424,7 +420,6 @@ av_cold void ff_vp8dsp_init_x86(VP8DSPContext* c)
         c->vp8_v_loop_filter_simple = ff_vp8_v_loop_filter_simple_ssse3;
         c->vp8_h_loop_filter_simple = ff_vp8_h_loop_filter_simple_ssse3;
 
-#if ARCH_X86_64 || HAVE_ALIGNED_STACK
         c->vp8_v_loop_filter16y_inner = ff_vp8_v_loop_filter16y_inner_ssse3;
         c->vp8_h_loop_filter16y_inner = ff_vp8_h_loop_filter16y_inner_ssse3;
         c->vp8_v_loop_filter8uv_inner = ff_vp8_v_loop_filter8uv_inner_ssse3;
@@ -434,17 +429,14 @@ av_cold void ff_vp8dsp_init_x86(VP8DSPContext* c)
         c->vp8_h_loop_filter16y       = ff_vp8_h_loop_filter16y_mbedge_ssse3;
         c->vp8_v_loop_filter8uv       = ff_vp8_v_loop_filter8uv_mbedge_ssse3;
         c->vp8_h_loop_filter8uv       = ff_vp8_h_loop_filter8uv_mbedge_ssse3;
-#endif
     }
 
     if (mm_flags & AV_CPU_FLAG_SSE4) {
         c->vp8_idct_dc_add                  = ff_vp8_idct_dc_add_sse4;
 
         c->vp8_h_loop_filter_simple   = ff_vp8_h_loop_filter_simple_sse4;
-#if ARCH_X86_64 || HAVE_ALIGNED_STACK
         c->vp8_h_loop_filter16y       = ff_vp8_h_loop_filter16y_mbedge_sse4;
         c->vp8_h_loop_filter8uv       = ff_vp8_h_loop_filter8uv_mbedge_sse4;
-#endif
     }
 #endif /* HAVE_YASM */
 }
diff --git a/libavutil/x86/x86inc.asm b/libavutil/x86/x86inc.asm
index 8a30689f00..483ae7b048 100644
--- a/libavutil/x86/x86inc.asm
+++ b/libavutil/x86/x86inc.asm
@@ -111,7 +111,12 @@ CPUNOP amdnop
 ; %1 = number of arguments. loads them from stack if needed.
 ; %2 = number of registers used. pushes callee-saved regs if needed.
 ; %3 = number of xmm registers used. pushes callee-saved xmm regs if needed.
-; %4 = list of names to define to registers
+; %4 = (optional) stack size to be allocated. If not aligned (x86-32 ICC 10.x,
+;      MSVC or YMM), the stack will be manually aligned (to 16 or 32 bytes),
+;      and an extra register will be allocated to hold the original stack
+;      pointer (to not invalidate r0m etc.). To prevent the use of an extra
+;      register as stack pointer, request a negative stack size.
+; %4+/%5+ = list of names to define to registers
 ; PROLOGUE can also be invoked by adding the same options to cglobal
 
 ; e.g.
@@ -147,11 +152,11 @@ CPUNOP amdnop
         %define r%1m  %2d
         %define r%1mp %2
     %elif ARCH_X86_64 ; memory
-        %define r%1m [rsp + stack_offset + %3]
-        %define r%1mp qword r %+ %1 %+ m
+        %define r%1m [rstk + stack_offset + %3]
+        %define r%1mp qword r %+ %1m
     %else
-        %define r%1m [esp + stack_offset + %3]
-        %define r%1mp dword r %+ %1 %+ m
+        %define r%1m [rstk + stack_offset + %3]
+        %define r%1mp dword r %+ %1m
     %endif
     %define r%1  %2
 %endmacro
@@ -212,12 +217,16 @@ DECLARE_REG_TMP_SIZE 0,1,2,3,4,5,6,7,8,9,10,11,12,13,14
 
 %macro PUSH 1
     push %1
-    %assign stack_offset stack_offset+gprsize
+    %ifidn rstk, rsp
+        %assign stack_offset stack_offset+gprsize
+    %endif
 %endmacro
 
 %macro POP 1
     pop %1
-    %assign stack_offset stack_offset-gprsize
+    %ifidn rstk, rsp
+        %assign stack_offset stack_offset-gprsize
+    %endif
 %endmacro
 
 %macro PUSH_IF_USED 1-*
@@ -249,14 +258,14 @@ DECLARE_REG_TMP_SIZE 0,1,2,3,4,5,6,7,8,9,10,11,12,13,14
 
 %macro SUB 2
     sub %1, %2
-    %ifidn %1, rsp
+    %ifidn %1, rstk
         %assign stack_offset stack_offset+(%2)
     %endif
 %endmacro
 
 %macro ADD 2
     add %1, %2
-    %ifidn %1, rsp
+    %ifidn %1, rstk
         %assign stack_offset stack_offset-(%2)
     %endif
 %endmacro
@@ -314,6 +323,73 @@ DECLARE_REG_TMP_SIZE 0,1,2,3,4,5,6,7,8,9,10,11,12,13,14
     %assign n_arg_names %0
 %endmacro
 
+%macro ALLOC_STACK 1-2 0 ; stack_size, n_xmm_regs (for win64 only)
+    %ifnum %1
+        %if %1 != 0
+            %assign %%stack_alignment ((mmsize + 15) & ~15)
+            %assign stack_size %1
+            %if stack_size < 0
+                %assign stack_size -stack_size
+            %endif
+            %assign xmm_regs_used %2
+            %if mmsize <= 16 && HAVE_ALIGNED_STACK
+                %assign stack_size_padded stack_size + %%stack_alignment - gprsize - (stack_offset & (%%stack_alignment - 1))
+                %if xmm_regs_used > 6
+                    %assign stack_size_padded stack_size_padded + (xmm_regs_used - 6) * 16
+                %endif
+                SUB rsp, stack_size_padded
+            %else
+                %assign reg_num (regs_used - 1)
+                %xdefine rstk r %+ reg_num
+                ; align stack, and save original stack location directly above
+                ; it, i.e. in [rsp+stack_size_padded], so we can restore the
+                ; stack in a single instruction (i.e. mov rsp, rstk or mov
+                ; rsp, [rsp+stack_size_padded])
+                mov  rstk, rsp
+                %assign stack_size_padded stack_size
+                %if xmm_regs_used > 6
+                    %assign stack_size_padded stack_size_padded + (xmm_regs_used - 6) * 16
+                %endif
+                %if %1 < 0 ; need to store rsp on stack
+                    sub  rsp, gprsize+stack_size_padded
+                    and  rsp, ~(%%stack_alignment-1)
+                    %xdefine rstkm [rsp+stack_size_padded]
+                    mov rstkm, rstk
+                %else ; can keep rsp in rstk during whole function
+                    sub  rsp, stack_size_padded
+                    and  rsp, ~(%%stack_alignment-1)
+                    %xdefine rstkm rstk
+                %endif
+            %endif
+            %if xmm_regs_used > 6
+                WIN64_PUSH_XMM
+            %endif
+        %endif
+    %endif
+%endmacro
+
+%macro SETUP_STACK_POINTER 1
+    %ifnum %1
+        %if %1 != 0 && (HAVE_ALIGNED_STACK == 0 || mmsize == 32)
+            %if %1 > 0
+                %assign regs_used (regs_used + 1)
+            %elif ARCH_X86_64 && regs_used == num_args && num_args <= 4 + UNIX64 * 2
+                %warning "Stack pointer will overwrite register argument"
+            %endif
+        %endif
+    %endif
+%endmacro
+
+%macro DEFINE_ARGS_INTERNAL 3+
+    %ifnum %2
+        DEFINE_ARGS %3
+    %elif %1 == 4
+        DEFINE_ARGS %2
+    %elif %1 > 4
+        DEFINE_ARGS %2, %3
+    %endif
+%endmacro
+
 %if WIN64 ; Windows x64 ;=================================================
 
 DECLARE_REG 0,  rcx
@@ -332,31 +408,37 @@ DECLARE_REG 12, R13, 104
 DECLARE_REG 13, R14, 112
 DECLARE_REG 14, R15, 120
 
-%macro PROLOGUE 2-4+ 0 ; #args, #regs, #xmm_regs, arg_names...
+%macro PROLOGUE 2-5+ 0 ; #args, #regs, #xmm_regs, [stack_size,] arg_names...
     %assign num_args %1
     %assign regs_used %2
+    SETUP_STACK_POINTER %4
     ASSERT regs_used >= num_args
     ASSERT regs_used <= 15
     PUSH_IF_USED 7, 8, 9, 10, 11, 12, 13, 14
-    %if mmsize == 8
-        %assign xmm_regs_used 0
-    %else
+    %assign xmm_regs_used 0
+    ALLOC_STACK %4, %3
+    %if mmsize != 8 && stack_size == 0
         WIN64_SPILL_XMM %3
     %endif
     LOAD_IF_USED 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14
-    DEFINE_ARGS %4
+    DEFINE_ARGS_INTERNAL %0, %4, %5
+%endmacro
+
+%macro WIN64_PUSH_XMM 0
+    %assign %%i xmm_regs_used
+    %rep (xmm_regs_used-6)
+        %assign %%i %%i-1
+        movdqa [rsp + (%%i-6)*16 + stack_size], xmm %+ %%i
+    %endrep
 %endmacro
 
 %macro WIN64_SPILL_XMM 1
     %assign xmm_regs_used %1
     ASSERT xmm_regs_used <= 16
     %if xmm_regs_used > 6
-        SUB rsp, (xmm_regs_used-6)*16+16
-        %assign %%i xmm_regs_used
-        %rep (xmm_regs_used-6)
-            %assign %%i %%i-1
-            movdqa [rsp + (%%i-6)*16+(~stack_offset&8)], xmm %+ %%i
-        %endrep
+        %assign stack_size_padded (xmm_regs_used-6)*16+16-gprsize-(stack_offset&15)
+        SUB rsp, stack_size_padded
+        WIN64_PUSH_XMM
     %endif
 %endmacro
 
@@ -365,19 +447,25 @@ DECLARE_REG 14, R15, 120
         %assign %%i xmm_regs_used
         %rep (xmm_regs_used-6)
             %assign %%i %%i-1
-            movdqa xmm %+ %%i, [%1 + (%%i-6)*16+(~stack_offset&8)]
+            movdqa xmm %+ %%i, [%1 + (%%i-6)*16+stack_size]
         %endrep
-        add %1, (xmm_regs_used-6)*16+16
+    %endif
+    %if stack_size_padded > 0
+        %if stack_size > 0 && (mmsize == 32 || HAVE_ALIGNED_STACK == 0)
+            mov rsp, rstkm
+        %else
+            add %1, stack_size_padded
+        %endif
     %endif
 %endmacro
 
 %macro WIN64_RESTORE_XMM 1
     WIN64_RESTORE_XMM_INTERNAL %1
-    %assign stack_offset stack_offset-(xmm_regs_used-6)*16+16
+    %assign stack_offset (stack_offset-stack_size_padded)
     %assign xmm_regs_used 0
 %endmacro
 
-%define has_epilogue regs_used > 7 || xmm_regs_used > 6 || mmsize == 32
+%define has_epilogue regs_used > 7 || xmm_regs_used > 6 || mmsize == 32 || stack_size > 0
 
 %macro RET 0
     WIN64_RESTORE_XMM_INTERNAL rsp
@@ -406,19 +494,28 @@ DECLARE_REG 12, R13, 56
 DECLARE_REG 13, R14, 64
 DECLARE_REG 14, R15, 72
 
-%macro PROLOGUE 2-4+ ; #args, #regs, #xmm_regs, arg_names...
+%macro PROLOGUE 2-5+ ; #args, #regs, #xmm_regs, [stack_size,] arg_names...
     %assign num_args %1
     %assign regs_used %2
+    SETUP_STACK_POINTER %4
     ASSERT regs_used >= num_args
     ASSERT regs_used <= 15
     PUSH_IF_USED 9, 10, 11, 12, 13, 14
+    ALLOC_STACK %4
     LOAD_IF_USED 6, 7, 8, 9, 10, 11, 12, 13, 14
-    DEFINE_ARGS %4
+    DEFINE_ARGS_INTERNAL %0, %4, %5
 %endmacro
 
-%define has_epilogue regs_used > 9 || mmsize == 32
+%define has_epilogue regs_used > 9 || mmsize == 32 || stack_size > 0
 
 %macro RET 0
+%if stack_size_padded > 0
+%if mmsize == 32 || HAVE_ALIGNED_STACK == 0
+    mov rsp, rstkm
+%else
+    add rsp, stack_size_padded
+%endif
+%endif
     POP_IF_USED 14, 13, 12, 11, 10, 9
 %if mmsize == 32
     vzeroupper
@@ -439,7 +536,7 @@ DECLARE_REG 6, ebp, 28
 
 %macro DECLARE_ARG 1-*
     %rep %0
-        %define r%1m [esp + stack_offset + 4*%1 + 4]
+        %define r%1m [rstk + stack_offset + 4*%1 + 4]
         %define r%1mp dword r%1m
         %rotate 1
     %endrep
@@ -447,24 +544,31 @@ DECLARE_REG 6, ebp, 28
 
 DECLARE_ARG 7, 8, 9, 10, 11, 12, 13, 14
 
-%macro PROLOGUE 2-4+ ; #args, #regs, #xmm_regs, arg_names...
+%macro PROLOGUE 2-5+ ; #args, #regs, #xmm_regs, [stack_size,] arg_names...
     %assign num_args %1
     %assign regs_used %2
-    %if num_args > 7
-        %assign num_args 7
-    %endif
     %if regs_used > 7
         %assign regs_used 7
     %endif
+    SETUP_STACK_POINTER %4
+    ASSERT regs_used <= 7
     ASSERT regs_used >= num_args
     PUSH_IF_USED 3, 4, 5, 6
+    ALLOC_STACK %4
     LOAD_IF_USED 0, 1, 2, 3, 4, 5, 6
-    DEFINE_ARGS %4
+    DEFINE_ARGS_INTERNAL %0, %4, %5
 %endmacro
 
-%define has_epilogue regs_used > 3 || mmsize == 32
+%define has_epilogue regs_used > 3 || mmsize == 32 || stack_size > 0
 
 %macro RET 0
+%if stack_size_padded > 0
+%if mmsize == 32 || HAVE_ALIGNED_STACK == 0
+    mov rsp, rstkm
+%else
+    add rsp, stack_size_padded
+%endif
+%endif
     POP_IF_USED 6, 5, 4, 3
 %if mmsize == 32
     vzeroupper
@@ -479,6 +583,8 @@ DECLARE_ARG 7, 8, 9, 10, 11, 12, 13, 14
 %endmacro
 %macro WIN64_RESTORE_XMM 1
 %endmacro
+%macro WIN64_PUSH_XMM 0
+%endmacro
 %endif
 
 %macro REP_RET 0
@@ -508,8 +614,12 @@ DECLARE_ARG 7, 8, 9, 10, 11, 12, 13, 14
 ; Applies any symbol mangling needed for C linkage, and sets up a define such that
 ; subsequent uses of the function name automatically refer to the mangled version.
 ; Appends cpuflags to the function name if cpuflags has been specified.
-%macro cglobal 1-2+ "" ; name, [PROLOGUE args]
+%macro cglobal 1-2+ ; name, [PROLOGUE args]
+%if %0 == 1
+    cglobal_internal %1 %+ SUFFIX
+%else
     cglobal_internal %1 %+ SUFFIX, %2
+%endif
 %endmacro
 %macro cglobal_internal 1-2+
     %ifndef cglobaled_%1
@@ -526,8 +636,11 @@ DECLARE_ARG 7, 8, 9, 10, 11, 12, 13, 14
     align function_align
     %1:
     RESET_MM_PERMUTATION ; not really needed, but makes disassembly somewhat nicer
+    %xdefine rstk rsp
     %assign stack_offset 0
-    %ifnidn %2, ""
+    %assign stack_size 0
+    %assign stack_size_padded 0
+    %if %0 > 1
         PROLOGUE %2
     %endif
 %endmacro