gallivm: replace the temp parameters of the JIT function with alloca'ed temps. This avoids useless writes of temporary results.

4 files changed, 41 insertions, 30 deletions

diff --git a/src/gallium/auxiliary/gallivm/gallivm_cpu.cpp b/src/gallium/auxiliary/gallivm/gallivm_cpu.cpp
index 3a2f2878a30..93a9748bdb3 100644
--- a/src/gallium/auxiliary/gallivm/gallivm_cpu.cpp
+++ b/src/gallium/auxiliary/gallivm/gallivm_cpu.cpp
@@ -179,8 +179,7 @@ struct gallivm_cpu_engine * gallivm_global_cpu_engine()
 
 typedef void (*vertex_shader_runner)(void *ainputs,
                                      void *dests,
-                                     float (*aconsts)[4],
-                                     void *temps);
+                                     float (*aconsts)[4]);
 
 #define MAX_TGSI_VERTICES 4
 /*!
@@ -223,8 +222,7 @@ int gallivm_cpu_vs_exec(struct gallivm_prog *prog,
       /* run shader */
       runner(machine->Inputs,
              machine->Outputs,
-             (float (*)[4]) constants,
-             machine->Temps);
+             (float (*)[4]) constants);
 
       /* Unswizzle all output results
        */
diff --git a/src/gallium/auxiliary/gallivm/storagesoa.cpp b/src/gallium/auxiliary/gallivm/storagesoa.cpp
index 4fc075cf6d4..e1e5cabcf55 100644
--- a/src/gallium/auxiliary/gallivm/storagesoa.cpp
+++ b/src/gallium/auxiliary/gallivm/storagesoa.cpp
@@ -48,13 +48,11 @@ using namespace llvm;
 StorageSoa::StorageSoa(llvm::BasicBlock *block,
                        llvm::Value *input,
                        llvm::Value *output,
-                       llvm::Value *consts,
-                       llvm::Value *temps)
+                       llvm::Value *consts)
    : m_block(block),
      m_input(input),
      m_output(output),
      m_consts(consts),
-     m_temps(temps),
      m_immediates(0),
      m_idx(0)
 {
@@ -169,7 +167,7 @@ std::vector<llvm::Value*> StorageSoa::constElement(llvm::IRBuilder<>* m_builder,
 {
    llvm::Value* res;
    std::vector<llvm::Value*> res2(4);
-   llvm::Value *xChannel, *yChannel, *zChannel, *wChannel;
+   llvm::Value *xChannel;
 
    xChannel = elementPointer(m_consts, idx, 0);
 
@@ -195,14 +193,15 @@ std::vector<llvm::Value*> StorageSoa::outputElement(llvm::Value *idx)
    return res;
 }
 
-std::vector<llvm::Value*> StorageSoa::tempElement(llvm::Value *idx)
+std::vector<llvm::Value*> StorageSoa::tempElement(llvm::IRBuilder<>* m_builder, int idx)
 {
    std::vector<llvm::Value*> res(4);
+   llvm::Value *temp = m_temps[idx];
 
-   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);
+   res[0] = element(temp, constantInt(0), 0);
+   res[1] = element(temp, constantInt(0), 1);
+   res[2] = element(temp, constantInt(0), 2);
+   res[3] = element(temp, constantInt(0), 3);
 
    return res;
 }
@@ -326,7 +325,7 @@ std::vector<llvm::Value*> StorageSoa::load(enum tgsi_file_type type, int idx, in
       val = outputElement(realIndex);
       break;
    case TGSI_FILE_TEMPORARY:
-      val = tempElement(realIndex);
+      val = tempElement(m_builder, idx);
       break;
    case TGSI_FILE_CONSTANT:
       val = constElement(m_builder, realIndex);
@@ -355,19 +354,39 @@ std::vector<llvm::Value*> StorageSoa::load(enum tgsi_file_type type, int idx, in
    return res;
 }
 
+llvm::Value * StorageSoa::allocaTemp(llvm::IRBuilder<>* m_builder)
+{
+   VectorType *vector   = VectorType::get(Type::FloatTy, 4);
+   ArrayType  *vecArray = ArrayType::get(vector, 4);
+   AllocaInst *alloca = new AllocaInst(vecArray, "temp",
+                                       m_builder->GetInsertBlock());
+
+   return alloca;
+}
+
+
 void StorageSoa::store(enum tgsi_file_type type, int idx, const std::vector<llvm::Value*> &val,
-                       int mask)
+                       int mask, llvm::IRBuilder<>* m_builder)
 {
    llvm::Value *out = 0;
+   llvm::Value *realIndex = 0;
    switch(type) {
    case TGSI_FILE_OUTPUT:
       out = m_output;
+      realIndex = constantInt(idx);
       break;
    case TGSI_FILE_TEMPORARY:
-      out = m_temps;
+      // if that temp doesn't already exist, alloca it
+      if (m_temps.find(idx) == m_temps.end())
+         m_temps[idx] = allocaTemp(m_builder);
+
+      out = m_temps[idx];
+
+      realIndex = constantInt(0);
       break;
    case TGSI_FILE_INPUT:
       out = m_input;
+      realIndex = constantInt(idx);
       break;
    case TGSI_FILE_ADDRESS: {
       llvm::Value *addr = m_addresses[idx];
@@ -385,7 +404,6 @@ void StorageSoa::store(enum tgsi_file_type type, int idx, const std::vector<llvm
       assert(0);
       break;
    }
-   llvm::Value *realIndex = constantInt(idx);
    if ((mask & TGSI_WRITEMASK_X)) {
       llvm::Value *xChannel = elementPointer(out, realIndex, 0);
       new StoreInst(val[0], xChannel, false, m_block);
diff --git a/src/gallium/auxiliary/gallivm/storagesoa.h b/src/gallium/auxiliary/gallivm/storagesoa.h
index f21ca6ec433..56886f85e7a 100644
--- a/src/gallium/auxiliary/gallivm/storagesoa.h
+++ b/src/gallium/auxiliary/gallivm/storagesoa.h
@@ -52,14 +52,13 @@ public:
    StorageSoa(llvm::BasicBlock *block,
               llvm::Value *input,
               llvm::Value *output,
-              llvm::Value *consts,
-              llvm::Value *temps);
+              llvm::Value *consts);
 
 
    std::vector<llvm::Value*> load(enum tgsi_file_type type, int idx, int swizzle, 
                                   llvm::IRBuilder<>* m_builder, llvm::Value *indIdx =0);
    void store(enum tgsi_file_type type, int idx, const std::vector<llvm::Value*> &val,
-              int mask);
+              int mask, llvm::IRBuilder<>* m_builder);
 
    void addImmediate(float *vec);
    void declareImmediates();
@@ -84,7 +83,7 @@ private:
    llvm::Value* unpackConstElement(llvm::IRBuilder<>* m_builder, llvm::Value *indIdx, int cc);
    std::vector<llvm::Value*> constElement(llvm::IRBuilder<>* m_builder, llvm::Value *indIdx);
    std::vector<llvm::Value*> outputElement(llvm::Value *indIdx);
-   std::vector<llvm::Value*> tempElement(llvm::Value *indIdx);
+   std::vector<llvm::Value*> tempElement(llvm::IRBuilder<>* m_builder, int idx);
    std::vector<llvm::Value*> immediateElement(llvm::Value *indIdx);
 private:
    llvm::BasicBlock *m_block;
@@ -92,12 +91,13 @@ private:
    llvm::Value *m_input;
    llvm::Value *m_output;
    llvm::Value *m_consts;
-   llvm::Value *m_temps;
+   std::map<int, llvm::Value*> m_temps;
    llvm::GlobalVariable *m_immediates;
 
    std::map<int, llvm::Value*> m_addresses;
 
    std::vector<std::vector<float> > m_immediatesToFlush;
+   llvm::Value * allocaTemp(llvm::IRBuilder<>* m_builder);
 
    mutable std::map<int, llvm::ConstantInt*> m_constInts;
    mutable char        m_name[32];
diff --git a/src/gallium/auxiliary/gallivm/tgsitollvm.cpp b/src/gallium/auxiliary/gallivm/tgsitollvm.cpp
index 1191a6cae97..c11b88af9ec 100644
--- a/src/gallium/auxiliary/gallivm/tgsitollvm.cpp
+++ b/src/gallium/auxiliary/gallivm/tgsitollvm.cpp
@@ -53,7 +53,6 @@ static inline FunctionType *vertexShaderFunctionType()
    // [4 x <4 x float>] inputs,
    // [4 x <4 x float>] output,
    // [4 x [1 x float]] consts,
-   // [4 x <4 x float>] temps
 
    std::vector<const Type*> funcArgs;
    VectorType *vectorType = VectorType::get(Type::FloatTy, 4);
@@ -67,7 +66,6 @@ static inline FunctionType *vertexShaderFunctionType()
    funcArgs.push_back(vectorArrayPtr);//inputs
    funcArgs.push_back(vectorArrayPtr);//output
    funcArgs.push_back(constsArrayPtr);//consts
-   funcArgs.push_back(vectorArrayPtr);//temps
 
    FunctionType *functionType = FunctionType::get(
       /*Result=*/Type::VoidTy,
@@ -246,7 +244,6 @@ translate_instruction(llvm::Module *module,
          val = storage->constElement(src->SrcRegister.Index, indIdx);
       } else if (src->SrcRegister.File == TGSI_FILE_INPUT) {
          val = storage->inputElement(src->SrcRegister.Index, indIdx);
-      // FIXME we should not be generating elements for temporaries, this creates useless memory writes
       } else if (src->SrcRegister.File == TGSI_FILE_TEMPORARY) {
          val = storage->tempElement(src->SrcRegister.Index);
       } else if (src->SrcRegister.File == TGSI_FILE_OUTPUT) {
@@ -677,7 +674,6 @@ translate_instruction(llvm::Module *module,
 
       if (dst->DstRegister.File == TGSI_FILE_OUTPUT) {
          storage->setOutputElement(dst->DstRegister.Index, out, dst->DstRegister.WriteMask);
-      // FIXME we should not be generating elements for temporaries, this creates useless memory writes
       } else if (dst->DstRegister.File == TGSI_FILE_TEMPORARY) {
          storage->setTempElement(dst->DstRegister.Index, out, dst->DstRegister.WriteMask);
       } else if (dst->DstRegister.File == TGSI_FILE_ADDRESS) {
@@ -1027,7 +1023,8 @@ translate_instructionir(llvm::Module *module,
    for (int i = 0; i < inst->Instruction.NumDstRegs; ++i) {
       struct tgsi_full_dst_register *dst = &inst->FullDstRegisters[i];
       storage->store((enum tgsi_file_type)dst->DstRegister.File,
-                     dst->DstRegister.Index, out, dst->DstRegister.WriteMask);
+                     dst->DstRegister.Index, out, dst->DstRegister.WriteMask,
+		     instr->getIRBuilder() );
    }
 }
 
@@ -1122,8 +1119,6 @@ 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 = BasicBlock::Create("entry", shader, 0);
 
@@ -1132,7 +1127,7 @@ 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, temps);
+   StorageSoa storage(label_entry, input, output, consts);
    InstructionsSoa instr(mod, shader, label_entry, &storage);
 
    while(!tgsi_parse_end_of_tokens(&parse)) {