From 52c178718ca888437eee900a5653ddf286a25f14 Mon Sep 17 00:00:00 2001
From: notaz <notasas@gmail.com>
Date: Fri, 22 Apr 2016 02:25:24 +0300
Subject: [PATCH] fix SoundMixer asm

SoundMixer386.asm forgot to save some callee-save registers too
---
 Sources/Engine/Sound/SoundMixer.cpp    | 24 ++++++++++++++++--------
 Sources/Engine/Sound/SoundMixer386.asm |  8 ++++++++
 2 files changed, 24 insertions(+), 8 deletions(-)

diff --git a/Sources/Engine/Sound/SoundMixer.cpp b/Sources/Engine/Sound/SoundMixer.cpp
index 6a9175a..e24ace2 100644
--- a/Sources/Engine/Sound/SoundMixer.cpp
+++ b/Sources/Engine/Sound/SoundMixer.cpp
@@ -96,11 +96,12 @@ void ResetMixer( const SLONG *pslBuffer, const SLONG slBufferSize)
   }
   #elif (defined __GNU_INLINE__)
   // !!! FIXME : rcg12172001 Is this REALLY any faster than memset()?
+  ULONG clob1, clob2;
   __asm__ __volatile__ (
     "cld                  \n\t"
     "rep                  \n\t"
     "stosl                \n\t"
-        : // no outputs.
+        : "=D" (clob1), "=c" (clob2)
         : "a" (0), "D" (pvMixerBuffer), "c" (slMixerBufferSize*2)
         : "cc", "memory"
   );
@@ -132,11 +133,12 @@ void CopyMixerBuffer_stereo( const SLONG slSrcOffset, void *pDstBuffer, const SL
   }
   #elif (defined __GNU_INLINE__)
   // !!! FIXME : rcg12172001 Is this REALLY any faster than memcpy()?
+  ULONG clob1, clob2, clob3;
   __asm__ __volatile__ (
     "cld                 \n\t"
     "rep                 \n\t"
     "movsl               \n\t"
-      : // no outputs.
+      : "=S" (clob1), "=D" (clob2), "=c" (clob3)
       : "S" (((char *)pvMixerBuffer) + slSrcOffset),
         "D" (pDstBuffer),
         "c" (slBytes >> 2)
@@ -184,6 +186,9 @@ copyLoop:
 
   #elif (defined __GNU_INLINE__)
   __asm__ __volatile__ (
+    "movl     %[pvMixerBuffer], %%esi         \n\t"
+    "movl     %[pDstBuffer], %%edi            \n\t"
+    "movl     %[slDW], %%ecx                  \n\t"
     "0:                                       \n\t" // copyLoop
     "movzwl   (%%esi), %%eax                  \n\t"
     "movw     %%ax, (%%edi)                   \n\t"
@@ -192,10 +197,10 @@ copyLoop:
     "decl     %%ecx                           \n\t"
     "jnz      0b                              \n\t" // copyLoop
       : // no outputs.
-      : "S" (((char *)pvMixerBuffer) + slSrcOffset),
-        "D" (pDstBuffer),
-        "c" (slBytes >> 2)
-      : "cc", "memory", "eax"
+      : [pvMixerBuffer] "g" (((char *)pvMixerBuffer) + slSrcOffset),
+        [pDstBuffer] "g" (pDstBuffer),
+        [slDW] "g" (slBytes >> 2)
+      : "eax", "ecx", "esi", "edi", "cc", "memory"
   );
 
   #else
@@ -247,6 +252,9 @@ copyLoop:
 
   #elif (defined __GNU_INLINE__)
   __asm__ __volatile__ (
+    "movl     %[pvMixerBuffer], %%esi      \n\t"
+    "movl     %[pvMixerBuffer], %%edi      \n\t"
+    "movl     %[slDW], %%ecx               \n\t"
     "cld                                   \n\t"
     "0:                                    \n\t" // copyLoop
     "movq     (%%esi), %%mm0               \n\t"
@@ -258,8 +266,8 @@ copyLoop:
     "jnz      0b                           \n\t" // copyLoop
     "emms                                  \n\t"
       : // no outputs.
-      : "S" (pvMixerBuffer), "D" (pvMixerBuffer), "c" (slBytes >> 2)
-      : "cc", "memory"
+      : [pvMixerBuffer] "g" (pvMixerBuffer), [slDW] "g" (slBytes >> 2)
+      : FPU_REGS, "mm0", "ecx", "esi", "edi", "cc", "memory"
   );
 
   #else
diff --git a/Sources/Engine/Sound/SoundMixer386.asm b/Sources/Engine/Sound/SoundMixer386.asm
index 9231c0c..bedc390 100644
--- a/Sources/Engine/Sound/SoundMixer386.asm
+++ b/Sources/Engine/Sound/SoundMixer386.asm
@@ -96,6 +96,8 @@ SEGMENT .text
 global MixMono_asm
 MixMono_asm:
     push    ebx   ; Save GCC register.
+    push    esi
+    push    edi
     ; convert from floats to fixints 32:16
     fld     D [fLeftOfs]
     fmul    D [f65536]
@@ -224,6 +226,8 @@ loopEnd_MixMono:
     shr     edx,16
     mov     D [slLastLeftSample],eax
     mov     D [slLastRightSample],edx
+    pop     edi
+    pop     esi
     pop     ebx   ; Restore GCC register.
     emms
     ret
@@ -232,6 +236,8 @@ loopEnd_MixMono:
 global MixStereo_asm
 MixStereo_asm:
     push    ebx   ; Save GCC register.
+    push    esi
+    push    edi
     ; convert from floats to fixints 32:16
     fld     D [fLeftOfs]
     fmul    D [f65536]
@@ -363,6 +369,8 @@ loopEnd_MixStereo:
     mov     D [slLastLeftSample],eax
     mov     D [slLastRightSample],edx
     emms
+    pop     edi
+    pop     esi
     pop     ebx   ; Restore GCC register.
     ret