handle temporaries in llvm code generated paths

6 files changed, 50 insertions, 25 deletions

diff --git a/src/mesa/pipe/draw/draw_vertex_shader.c b/src/mesa/pipe/draw/draw_vertex_shader.c
index 9c31df1e3e1..c824c1407e9 100644
--- a/src/mesa/pipe/draw/draw_vertex_shader.c
+++ b/src/mesa/pipe/draw/draw_vertex_shader.c
@@ -119,7 +119,8 @@ run_vertex_program(struct draw_context *draw,
    gallivm_cpu_vs_exec(prog,
                        machine->Inputs,
                        machine->Outputs,
-                       machine->Consts);
+                       machine->Consts,
+                       machine->Temps);
    } else
 #elif defined(__i386__) || defined(__386__)
    if (draw->use_sse) {
diff --git a/src/mesa/pipe/llvm/gallivm.h b/src/mesa/pipe/llvm/gallivm.h
index b104520cb7f..92da4bca7f4 100644
--- a/src/mesa/pipe/llvm/gallivm.h
+++ b/src/mesa/pipe/llvm/gallivm.h
@@ -82,7 +82,8 @@ struct gallivm_cpu_engine *gallivm_global_cpu_engine();
 int gallivm_cpu_vs_exec(struct gallivm_prog *prog,
                         struct tgsi_exec_vector       *inputs,
                         struct tgsi_exec_vector       *dests,
-                        float (*consts)[4]);
+                        float (*consts)[4],
+                        struct tgsi_exec_vector       *temps);
 int gallivm_cpu_fs_exec(struct gallivm_prog *prog,
                         float x, float y,
                         float (*dests)[PIPE_MAX_SHADER_INPUTS][4],
diff --git a/src/mesa/pipe/llvm/gallivm_cpu.cpp b/src/mesa/pipe/llvm/gallivm_cpu.cpp
index 011cba55c23..dc4d92a72a6 100644
--- a/src/mesa/pipe/llvm/gallivm_cpu.cpp
+++ b/src/mesa/pipe/llvm/gallivm_cpu.cpp
@@ -177,10 +177,7 @@ struct gallivm_cpu_engine * gallivm_global_cpu_engine()
 typedef void (*vertex_shader_runner)(void *ainputs,
                                      void *dests,
                                      float (*aconsts)[4],
-                                     int num_vertices,
-                                     int num_inputs,
-                                     int num_attribs,
-                                     int num_consts);
+                                     void *temps);
 
 
 /*!
@@ -191,12 +188,13 @@ typedef void (*vertex_shader_runner)(void *ainputs,
 int gallivm_cpu_vs_exec(struct gallivm_prog *prog,
                         struct tgsi_exec_vector       *inputs,
                         struct tgsi_exec_vector       *dests,
-                         float (*consts)[4])
+                        float (*consts)[4],
+                        struct tgsi_exec_vector       *temps)
 {
    vertex_shader_runner runner = reinterpret_cast<vertex_shader_runner>(prog->function);
    assert(runner);
    /*FIXME*/
-   runner(inputs, dests, consts, 4, 4, 4, prog->num_consts);
+   runner(inputs, dests, consts, temps);
 
    return 0;
 }
diff --git a/src/mesa/pipe/llvm/storagesoa.cpp b/src/mesa/pipe/llvm/storagesoa.cpp
index ef14e29bfed..e09e9e8fe7d 100644
--- a/src/mesa/pipe/llvm/storagesoa.cpp
+++ b/src/mesa/pipe/llvm/storagesoa.cpp
@@ -44,11 +44,13 @@ using namespace llvm;
 StorageSoa::StorageSoa(llvm::BasicBlock *block,
                        llvm::Value *input,
                        llvm::Value *output,
-                       llvm::Value *consts)
+                       llvm::Value *consts,
+                       llvm::Value *temps)
    : m_block(block),
      m_input(input),
      m_output(output),
      m_consts(consts),
+     m_temps(temps),
      m_idx(0)
 {
 }
@@ -101,6 +103,11 @@ std::vector<llvm::Value*> StorageSoa::tempElement(int idx, int swizzle,
 {
    std::vector<llvm::Value*> res(4);
 
+   res[0] = element(m_temps, idx, 0);
+   res[1] = element(m_temps, idx, 1);
+   res[2] = element(m_temps, idx, 2);
+   res[3] = element(m_temps, idx, 3);
+
    return res;
 }
 
@@ -138,6 +145,20 @@ void StorageSoa::storeOutput(int dstIdx, const std::vector<llvm::Value*> &val,
 void StorageSoa::storeTemp(int idx, const std::vector<llvm::Value*> &val,
                            int mask)
 {
+   if (mask != TGSI_WRITEMASK_XYZW) {
+      fprintf(stderr, "requires swizzle!!\n");
+      assert(0);
+   } else {
+      llvm::Value *xChannel = elementPointer(m_temps, idx, 0);
+      llvm::Value *yChannel = elementPointer(m_temps, idx, 1);
+      llvm::Value *zChannel = elementPointer(m_temps, idx, 2);
+      llvm::Value *wChannel = elementPointer(m_temps, idx, 3);
+
+      StoreInst *st = new StoreInst(val[0], xChannel, false, m_block);
+      st = new StoreInst(val[1], yChannel, false, m_block);
+      st = new StoreInst(val[2], zChannel, false, m_block);
+      st = new StoreInst(val[3], wChannel, false, m_block);
+   }
 }
 
 void StorageSoa::storeAddress(int idx, const std::vector<llvm::Value*> &val,
diff --git a/src/mesa/pipe/llvm/storagesoa.h b/src/mesa/pipe/llvm/storagesoa.h
index 9d5609f5398..84db7726a7c 100644
--- a/src/mesa/pipe/llvm/storagesoa.h
+++ b/src/mesa/pipe/llvm/storagesoa.h
@@ -46,7 +46,8 @@ public:
    StorageSoa(llvm::BasicBlock *block,
               llvm::Value *input,
               llvm::Value *output,
-              llvm::Value *consts);
+              llvm::Value *consts,
+              llvm::Value *temps);
 
    void addImmediate(float *vec);
 
@@ -79,6 +80,7 @@ private:
    llvm::Value *m_input;
    llvm::Value *m_output;
    llvm::Value *m_consts;
+   llvm::Value *m_temps;
 
    mutable std::map<int, llvm::ConstantInt*> m_constInts;
    mutable char        m_name[32];
diff --git a/src/mesa/pipe/llvm/tgsitollvm.cpp b/src/mesa/pipe/llvm/tgsitollvm.cpp
index 070c9a67f3c..3497eebcd97 100644
--- a/src/mesa/pipe/llvm/tgsitollvm.cpp
+++ b/src/mesa/pipe/llvm/tgsitollvm.cpp
@@ -53,23 +53,23 @@ static inline FunctionType *vertexShaderFunctionType()
    // pass are castable to the following:
    // [4 x <4 x float>] inputs,
    // [4 x <4 x float>] output,
-   // [4 x [4 x float]] consts
+   // [4 x [4 x float]] consts,
+   // [4 x <4 x float>] temps
+
    std::vector<const Type*> funcArgs;
-   {
-      VectorType *vectorType = VectorType::get(Type::FloatTy, 4);
-      ArrayType *vectorArray = ArrayType::get(vectorType, 4);
-      PointerType *vectorArrayPtr = PointerType::get(vectorArray, 0);
+   VectorType *vectorType = VectorType::get(Type::FloatTy, 4);
+   ArrayType *vectorArray = ArrayType::get(vectorType, 4);
+   PointerType *vectorArrayPtr = PointerType::get(vectorArray, 0);
 
-      funcArgs.push_back(vectorArrayPtr);//inputs
-      funcArgs.push_back(vectorArrayPtr);//output
-   }
-   {
-      ArrayType   *floatArray     = ArrayType::get(Type::FloatTy, 4);
-      ArrayType   *constsArray    = ArrayType::get(floatArray, 4);
-      PointerType *constsArrayPtr = PointerType::get(constsArray, 0);
+   ArrayType   *floatArray     = ArrayType::get(Type::FloatTy, 4);
+   ArrayType   *constsArray    = ArrayType::get(floatArray, 4);
+   PointerType *constsArrayPtr = PointerType::get(constsArray, 0);
+
+   funcArgs.push_back(vectorArrayPtr);//inputs
+   funcArgs.push_back(vectorArrayPtr);//output
+   funcArgs.push_back(constsArrayPtr);//consts
+   funcArgs.push_back(vectorArrayPtr);//temps
 
-      funcArgs.push_back(constsArrayPtr);//consts
-   }
    FunctionType *functionType = FunctionType::get(
       /*Result=*/Type::VoidTy,
       /*Params=*/funcArgs,
@@ -1162,6 +1162,8 @@ llvm::Module * tgsi_to_llvmir(struct gallivm_ir *ir,
    output->setName("outputs");
    Value *consts = args++;
    consts->setName("consts");
+   Value *temps = args++;
+   temps->setName("temps");
 
    BasicBlock *label_entry = new BasicBlock("entry", shader, 0);
 
@@ -1170,7 +1172,7 @@ llvm::Module * tgsi_to_llvmir(struct gallivm_ir *ir,
    fi = tgsi_default_full_instruction();
    fd = tgsi_default_full_declaration();
 
-   StorageSoa storage(label_entry, input, output, consts);
+   StorageSoa storage(label_entry, input, output, consts, temps);
    InstructionsSoa instr(mod, shader, label_entry, &storage);
 
    while(!tgsi_parse_end_of_tokens(&parse)) {