Check for SSE2 and set the denormals-are-zero bit for mixing if available

Alc/ALc.c

@@ -845,7 +845,7 @@ static void alc_initconfig(void)
 
     capfilter = 0;
 #ifdef HAVE_SSE
-    capfilter |= CPU_CAP_SSE;
+    capfilter |= CPU_CAP_SSE | CPU_CAP_SSE2;
 #endif
 #ifdef HAVE_NEON
     capfilter |= CPU_CAP_NEON;
@@ -872,6 +872,8 @@ static void alc_initconfig(void)
                 len = (next ? ((size_t)(next-str)) : strlen(str));
                 if(strncasecmp(str, "sse", len) == 0)
                     capfilter &= ~CPU_CAP_SSE;
+                else if(strncasecmp(str, "sse2", len) == 0)
+                    capfilter &= ~CPU_CAP_SSE2;
                 else if(strncasecmp(str, "neon", len) == 0)
                     capfilter &= ~CPU_CAP_NEON;
                 else

Alc/helpers.c

@@ -106,7 +106,11 @@ void FillCPUCaps(ALuint capfilter)
         {
 #ifdef bit_SSE
             if((cpuinf[0].regs[3]&bit_SSE))
+            {
                 caps |= CPU_CAP_SSE;
+                if((cpuinf[0].regs[3]&bit_SSE2))
+                    caps |= CPU_CAP_SSE2;
+            }
 #endif
         }
     }
@@ -119,7 +123,11 @@ void FillCPUCaps(ALuint capfilter)
     else
     {
         if(IsProcessorFeaturePresent(PF_XMMI_INSTRUCTIONS_AVAILABLE))
+        {
             caps |= CPU_CAP_SSE;
+            if(IsProcessorFeaturePresent(PF_XMMI64_INSTRUCTIONS_AVAILABLE))
+                caps |= CPU_CAP_SSE2;
+        }
     }
 #endif
 #ifdef HAVE_NEON
@@ -127,9 +135,10 @@ void FillCPUCaps(ALuint capfilter)
     caps |= CPU_CAP_NEON;
 #endif
 
-    TRACE("Got caps:%s%s%s\n", ((caps&CPU_CAP_SSE)?((capfilter&CPU_CAP_SSE)?" SSE":" (SSE)"):""),
-                               ((caps&CPU_CAP_NEON)?((capfilter&CPU_CAP_NEON)?" Neon":" (Neon)"):""),
-                               ((!caps)?" -none-":""));
+    TRACE("Got caps:%s%s%s%s\n", ((caps&CPU_CAP_SSE)?((capfilter&CPU_CAP_SSE)?" SSE":" (SSE)"):""),
+                                 ((caps&CPU_CAP_SSE)?((capfilter&CPU_CAP_SSE2)?" SSE2":" (SSE2)"):""),
+                                 ((caps&CPU_CAP_NEON)?((capfilter&CPU_CAP_NEON)?" Neon":" (Neon)"):""),
+                                 ((!caps)?" -none-":""));
     CPUCapFlags = caps & capfilter;
 }
 
@@ -207,6 +216,8 @@ void SetMixerFPUMode(FPUCtl *ctl)
         ctl->sse_state = sseState;
         sseState |= 0x6000; /* set round-to-zero */
         sseState |= 0x8000; /* set flush-to-zero */
+        if((CPUCapFlags&CPU_CAP_SSE2))
+            sseState |= 0x0040; /* set denormals-are-zero */
         __asm__ __volatile__("ldmxcsr %0" : : "m" (*&sseState));
     }
 #endif

CMakeLists.txt

@@ -26,10 +26,10 @@ PROJECT(OpenAL C)
 SET(CMAKE_ALLOW_LOOSE_LOOP_CONSTRUCTS TRUE)
 
 
-OPTION(ALSOFT_CPUEXT_SSE  "Check for SSE CPU extensions"      ON)
+OPTION(ALSOFT_CPUEXT_SSE  "Check for SSE/SSE2 CPU extensions" ON)
 OPTION(ALSOFT_CPUEXT_NEON "Check for ARM Neon CPU extensions" ON)
 
-OPTION(ALSOFT_REQUIRE_SSE  "Require SSE CPU extensions"      OFF)
+OPTION(ALSOFT_REQUIRE_SSE  "Require SSE/SSE2 CPU extensions" OFF)
 OPTION(ALSOFT_REQUIRE_NEON "Require ARM Neon CPU extensions" OFF)
 
 

OpenAL32/Include/alMain.h

@@ -837,7 +837,8 @@ extern ALint RTPrioLevel;
 extern ALuint CPUCapFlags;
 enum {
     CPU_CAP_SSE    = 1<<0,
-    CPU_CAP_NEON   = 1<<1,
+    CPU_CAP_SSE2   = 1<<1,
+    CPU_CAP_NEON   = 1<<2,
 };
 
 void FillCPUCaps(ALuint capfilter);

alsoftrc.sample

@@ -14,7 +14,7 @@
 ## disable-cpu-exts:
 #  Disables use of the listed CPU extensions. Certain methods may utilize CPU
 #  extensions when detected, and this option is useful for preventing those
-#  extensions from being used. The available extensions are: sse, neon.
+#  extensions from being used. The available extensions are: sse, sse2, neon.
 #  Specifying 'all' disables use of all extensions.
 #disable-cpu-exts =