Fix build problems for Mac.

strnlen not available on Mac OS X.
Fix worker thread synchronization.

Change-Id: I774ecb9f39753a9da35073a0121af0fd2cdecd01
Signed-off-by: David Li <davidxli@google.com>

7 files changed, 109 insertions, 72 deletions

diff --git a/Android.mk b/Android.mk
index 01044e3..6dd3223 100644
--- a/Android.mk
+++ b/Android.mk
@@ -149,9 +149,8 @@ endif
 
 LOCAL_C_INCLUDES :=	$(libMesa_C_INCLUDES)
 
-# libMesa only builds on Linux host for now
-#include $(LLVM_ROOT_PATH)/llvm-host-build.mk
-#include $(BUILD_HOST_STATIC_LIBRARY)
+include $(LLVM_ROOT_PATH)/llvm-host-build.mk
+include $(BUILD_HOST_STATIC_LIBRARY)
 
 
 # Static library for target
diff --git a/src/pixelflinger2/pixelflinger2.cpp b/src/pixelflinger2/pixelflinger2.cpp
index c5ee3a6..9c41ca9 100644
--- a/src/pixelflinger2/pixelflinger2.cpp
+++ b/src/pixelflinger2/pixelflinger2.cpp
@@ -185,6 +185,9 @@ static void EnableDisable(GGLInterface * iface, GLenum cap, GLboolean enable)
 
 void InitializeGGLState(GGLInterface * iface)
 {
+#if USE_DUAL_THREAD
+   reinterpret_cast<GGLContext *>(iface)->worker = GGLContext::Worker();
+#endif
    iface->DepthRangef = DepthRangef;
    iface->Viewport = Viewport;
    iface->CullFace = CullFace;
@@ -220,7 +223,7 @@ void InitializeGGLState(GGLInterface * iface)
 
    for (unsigned i = 0; i < GGL_MAXCOMBINEDTEXTUREIMAGEUNITS; i++)
       iface->SetSampler(iface, i, NULL);
-      
+
    iface->SetBuffer(iface, GL_COLOR_BUFFER_BIT, NULL);
    iface->SetBuffer(iface, GL_DEPTH_BUFFER_BIT, NULL);
    iface->SetBuffer(iface, GL_STENCIL_BUFFER_BIT, NULL);
@@ -230,12 +233,10 @@ void InitializeGGLState(GGLInterface * iface)
 
 GGLInterface * CreateGGLInterface()
 {
-   //GGLContext * ctx = (GGLContext *)calloc(1, sizeof(GGLContext) + sizeof(__GLcontextRec));
-   GGLContext * ctx = (GGLContext *)calloc(1, sizeof(GGLContext));
+   GGLContext * const ctx = (GGLContext *)calloc(1, sizeof(GGLContext));
    if (!ctx)
       return NULL;
    assert((void *)ctx == (void *)&ctx->interface);
-   //ctx->glCtx = (GLcontext *)((char *)ctx + sizeof(GGLContext));
 
    //_glapi_set_context(ctx->glCtx);
    //_mesa_init_constants(&Const);
@@ -247,19 +248,11 @@ GGLInterface * CreateGGLInterface()
 
 void UninitializeGGLState(GGLInterface * iface)
 {
-   GGLContext * ctx = (GGLContext *)iface;
-   assert((void *)ctx == (void *)iface);
-   
 #if USE_DUAL_THREAD
-   ctx->worker.hasWork = false;
-   ctx->worker.quit = true;
-   pthread_mutex_lock(&ctx->worker.lock);
-   pthread_cond_signal(&ctx->worker.cond);
-   pthread_mutex_unlock(&ctx->worker.lock);
+   reinterpret_cast<GGLContext *>(iface)->worker.~Worker();
 #endif
-
    DestroyShaderFunctions(iface);
-   
+
 #if USE_LLVM_TEXTURE_SAMPLER
    puts("USE_LLVM_TEXTURE_SAMPLER");
 #endif
@@ -277,8 +270,7 @@ void UninitializeGGLState(GGLInterface * iface)
 
 void DestroyGGLInterface(GGLInterface * iface)
 {
-   GGLContext * ctx = (GGLContext *)iface;
-   assert((void *)ctx == (void *)iface);
+   GGLContext * const ctx = reinterpret_cast<GGLContext *>(iface);
    UninitializeGGLState(iface);
    free(ctx);
 }
diff --git a/src/pixelflinger2/pixelflinger2.h b/src/pixelflinger2/pixelflinger2.h
index bd77a2e..f42bc78 100644
--- a/src/pixelflinger2/pixelflinger2.h
+++ b/src/pixelflinger2/pixelflinger2.h
@@ -108,14 +108,41 @@ struct GGLContext {
       unsigned startY, endY, varyingCount;
       VertexOutput bV, cV, bDx, cDx;
       int width, height;
-      volatile bool hasWork;
+      bool assignedWork; // only used by main; worker uses assignCond & quit
       bool quit;
 
-      pthread_cond_t cond;
-      pthread_mutex_t lock;
+      pthread_cond_t assignCond;
+      pthread_mutex_t assignLock; // held by worker execpt for during cond_wait assign
+      pthread_cond_t finishCond;
+      pthread_mutex_t finishLock; // held by main except for during cond_wait finish
       pthread_t thread;
 
-      Worker() : cond(PTHREAD_COND_INITIALIZER), lock(PTHREAD_MUTEX_INITIALIZER), thread(NULL) {}
+      Worker() : assignedWork(false), quit(false), thread(0)
+      {
+         pthread_cond_init(&assignCond, NULL);
+         pthread_cond_init(&finishCond, NULL);
+         pthread_mutex_init(&assignLock, NULL);
+         pthread_mutex_init(&finishLock, NULL);
+         pthread_mutex_lock(&finishLock);
+         // actual thread is created later in raster.cpp
+      }
+      ~Worker()
+      {
+         if (0 != thread)
+         {
+            pthread_mutex_lock(&assignLock);
+            quit = true;
+            pthread_cond_signal(&assignCond); // signal thread to quit
+            pthread_mutex_unlock(&assignLock);
+            pthread_join(thread, NULL);
+         }
+         pthread_mutex_unlock(&finishLock);
+
+         pthread_cond_destroy(&assignCond);
+         pthread_cond_destroy(&finishCond);
+         pthread_mutex_destroy(&assignLock);
+         pthread_mutex_destroy(&finishLock);
+      }
    } worker;
 #endif
 
diff --git a/src/pixelflinger2/raster.cpp b/src/pixelflinger2/raster.cpp
index bd69f59..07cdc0e 100644
--- a/src/pixelflinger2/raster.cpp
+++ b/src/pixelflinger2/raster.cpp
@@ -91,16 +91,21 @@ static void * RasterTrapezoidWorker(void * threadArgs)
 {
    GGLContext::Worker * args = (GGLContext::Worker *)threadArgs;
    VertexOutput clip0, clip1, * left, * right;
-   while (!args->quit) {
-      pthread_mutex_lock(&args->lock);
-      while (!args->hasWork && !args->quit)
-         pthread_cond_wait(&args->cond, &args->lock);
-      pthread_mutex_unlock(&args->lock);
-   
+
+   pthread_mutex_lock(&args->finishLock);
+   pthread_mutex_lock(&args->assignLock);
+   pthread_cond_signal(&args->finishCond);
+   pthread_mutex_unlock(&args->finishLock);
+
+   while (true) {
+      pthread_cond_wait(&args->assignCond, &args->assignLock);
       if (args->quit)
+      {
+         pthread_mutex_unlock(&args->assignLock);
          break;
-//         if (!args->hasWork)
-//            continue;
+      }
+      else
+          assert(args->assignedWork);
 
       for (unsigned y = args->startY; y <= args->endY; y += 2) {
          do {
@@ -133,7 +138,10 @@ static void * RasterTrapezoidWorker(void * threadArgs)
          args->bV.frontFacingPointCoord += args->bDx.frontFacingPointCoord;
          args->cV.frontFacingPointCoord += args->cDx.frontFacingPointCoord;
       }
-      args->hasWork = false;
+
+      pthread_mutex_lock(&args->finishLock);
+      pthread_cond_signal(&args->finishCond);
+      pthread_mutex_unlock(&args->finishLock);
    }
    pthread_exit(NULL);
    return NULL;
@@ -159,6 +167,7 @@ static void RasterTrapezoid(const GGLInterface * iface, const VertexOutput * tl,
    VertexOutput tmp;
 
    // vertically clip
+
    if ((int)tlv.position.y < 0) {
       InterpolateVertex(&tlv, &blv, (0 - tlv.position.y) / (blv.position.y - tlv.position.y),
                         &tmp, varyingCount);
@@ -240,38 +249,45 @@ static void RasterTrapezoid(const GGLInterface * iface, const VertexOutput * tl,
 #if USE_DUAL_THREAD
    GGLContext::Worker & args = ctx->worker;
    if (!ctx->worker.thread) {
-      int rc = pthread_create(&ctx->worker.thread, NULL, RasterTrapezoidWorker, &args);
+      pthread_attr_t attr;
+      pthread_attr_init(&attr);
+      pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_JOINABLE);
+      int rc = pthread_create(&ctx->worker.thread, &attr, RasterTrapezoidWorker, &args);
       assert(!rc);
+      // wait for worker to start
+      pthread_cond_wait(&args.finishCond, &args.finishLock);
    }
-   args.bV = bV;
-   args.cV = cV;
-   for (unsigned i = 0; i < varyingCount; i++) {
-      args.bV.varyings[i] += bDx.varyings[i];
-      bDx.varyings[i] += bDx.varyings[i];
-      args.cV.varyings[i] += cDx.varyings[i];
-      cDx.varyings[i] += cDx.varyings[i];
-   }
-   args.bV.position += bDx.position;
-   bDx.position += bDx.position;
-   args.cV.position += cDx.position;
-   cDx.position += cDx.position;
-   args.bV.frontFacingPointCoord += bDx.frontFacingPointCoord;
-   bDx.frontFacingPointCoord += bDx.frontFacingPointCoord;
-   args.cV.frontFacingPointCoord += cDx.frontFacingPointCoord;
-   cDx.frontFacingPointCoord += cDx.frontFacingPointCoord;
-   args.iface = iface;
-   args.bDx = bDx;
-   args.cDx = cDx;
-   args.varyingCount = varyingCount;
    args.startY = startY + 1;
    args.endY = endY;
-   args.width = width;
-   args.height = height;
    if (args.startY <= args.endY) {
-      pthread_mutex_lock(&args.lock);
-      args.hasWork = true;
-      pthread_cond_signal(&args.cond);
-      pthread_mutex_unlock(&args.lock);
+      pthread_mutex_lock(&args.assignLock);
+
+      args.bV = bV;
+      args.cV = cV;
+      for (unsigned i = 0; i < varyingCount; i++) {
+         args.bV.varyings[i] += bDx.varyings[i];
+         bDx.varyings[i] += bDx.varyings[i];
+         args.cV.varyings[i] += cDx.varyings[i];
+         cDx.varyings[i] += cDx.varyings[i];
+      }
+      args.bV.position += bDx.position;
+      bDx.position += bDx.position;
+      args.cV.position += cDx.position;
+      cDx.position += cDx.position;
+      args.bV.frontFacingPointCoord += bDx.frontFacingPointCoord;
+      bDx.frontFacingPointCoord += bDx.frontFacingPointCoord;
+      args.cV.frontFacingPointCoord += cDx.frontFacingPointCoord;
+      cDx.frontFacingPointCoord += cDx.frontFacingPointCoord;
+      args.iface = iface;
+      args.bDx = bDx;
+      args.cDx = cDx;
+      args.varyingCount = varyingCount;
+      args.width = width;
+      args.height = height;
+      args.assignedWork = true;
+
+      pthread_cond_signal(&args.assignCond);
+      pthread_mutex_unlock(&args.assignLock);
    }
 #endif
 
@@ -309,8 +325,11 @@ static void RasterTrapezoid(const GGLInterface * iface, const VertexOutput * tl,
    }
 
 #if USE_DUAL_THREAD
-   while (args.hasWork)
-      ; // wait
+   if (args.assignedWork)
+   {
+      pthread_cond_wait(&args.finishCond, &args.finishLock);
+      args.assignedWork = false;
+   }
 #endif
 }
 
@@ -373,8 +392,6 @@ static void DrawTriangle(const GGLInterface * iface, const VertexInput * vin1,
 
 //   LOGD("pf2: DrawTriangle");
 
-   const gl_shader_program * program = ctx->CurrentProgram;
-
 //   if (!strstr(program->Shaders[MESA_SHADER_FRAGMENT]->Source,
 //               "gl_FragColor = color * texture2D(sampler, outTexCoords).a;"))
 //      return;
diff --git a/src/pixelflinger2/shader.cpp b/src/pixelflinger2/shader.cpp
index 5cca627..ed1da57 100644
--- a/src/pixelflinger2/shader.cpp
+++ b/src/pixelflinger2/shader.cpp
@@ -336,7 +336,7 @@ static void GetShaderKeyString(const GLenum type, const ShaderKey * key,
    *str++ = '\0';
 }
 
-static const unsigned SCANLINE_KEY_STRING_LEN = 2 * sizeof(ShaderKey::scanLineKey) + 3 + SHADER_KEY_STRING_LEN;
+static const unsigned SCANLINE_KEY_STRING_LEN = 2 * sizeof(ShaderKey::ScanLineKey) + 3 + SHADER_KEY_STRING_LEN;
 
 static char * GetScanlineKeyString(const ShaderKey * key, char * buffer,
                                    const unsigned bufferSize)
diff --git a/src/talloc/hieralloc.c b/src/talloc/hieralloc.c
index 48b8179..ba6e6f7 100644
--- a/src/talloc/hieralloc.c
+++ b/src/talloc/hieralloc.c
@@ -339,8 +339,9 @@ char * hieralloc_strndup(const void * ctx, const char * str, unsigned len)
 	if (!str)
 		return NULL;
 	
-   len = strnlen(str, len);
-   char * ret = (char *)hieralloc_allocate(ctx, len + 1, str);
+    const char *p = (const char *)memchr(str, '\0', len);
+    len = (p ? p - str : len);
+    char * ret = (char *)hieralloc_allocate(ctx, len + 1, str);
 	if (!ret)
 		return NULL;
 	memcpy(ret, str, len);
@@ -389,7 +390,9 @@ char * hieralloc_strndup_append(char * str, const char * append, unsigned len)
 		return hieralloc_strdup(NULL, append);
 	if (!append)
 		return str;
-	return _hieralloc_strlendup_append(str, strlen(str), append, strnlen(append, len));
+    const char *p = (const char *)memchr(append, '\0', len);
+    len = (p ? p - append : len);
+	return _hieralloc_strlendup_append(str, strlen(str), append, len);
 }
 
 // allocate and vsprintf
diff --git a/test/Android.mk b/test/Android.mk
index 68df587..bf6728f 100644
--- a/test/Android.mk
+++ b/test/Android.mk
@@ -22,7 +22,7 @@ include $(CLEAR_VARS)
 LOCAL_MODULE_TAGS := optional
 
 ifeq ($(DEBUG_BUILD),true)
-LOCAL_CFLAGS += -DDEBUG -UNDEBUG -O0 -g 
+LOCAL_CFLAGS += -DDEBUG -UNDEBUG -O0 -g
 endif
 
 LOCAL_MODULE := mesa
@@ -31,9 +31,8 @@ LOCAL_STATIC_LIBRARIES := libMesa
 LOCAL_SHARED_LIBRARIES := libbcc
 LOCAL_C_INCLUDES := $(mesa_C_INCLUDES)
 
-# libMesa only builds on Linux host for now
-#include $(LLVM_ROOT_PATH)/llvm-host-build.mk
-#include $(BUILD_HOST_EXECUTABLE)
+include $(LLVM_ROOT_PATH)/llvm-host-build.mk
+include $(BUILD_HOST_EXECUTABLE)
 
 # Executable for target
 # ========================================================
@@ -42,7 +41,7 @@ include $(CLEAR_VARS)
 LOCAL_MODULE_TAGS := optional
 
 ifeq ($(DEBUG_BUILD),true)
-LOCAL_CFLAGS += -DDEBUG -UNDEBUG -O0 -g 
+LOCAL_CFLAGS += -DDEBUG -UNDEBUG -O0 -g
 endif
 
 LOCAL_MODULE := mesa