1 files changed, 116 insertions, 35 deletions

diff --git a/src/mesa/shader/slang/slang_emit.c b/src/mesa/shader/slang/slang_emit.c
index 3a837c9c10..6a47585c7a 100644
--- a/src/mesa/shader/slang/slang_emit.c
+++ b/src/mesa/shader/slang/slang_emit.c
@@ -164,7 +164,7 @@ _slang_var_swizzle(GLint size, GLint comp)
 {
    switch (size) {
    case 1:
-      return MAKE_SWIZZLE4(comp, comp, comp, comp);
+      return MAKE_SWIZZLE4(comp, SWIZZLE_NIL, SWIZZLE_NIL, SWIZZLE_NIL);
    case 2:
       return MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_Y, SWIZZLE_NIL, SWIZZLE_NIL);
    case 3:
@@ -451,7 +451,7 @@ emit_arl_load(slang_emit_info *emitInfo,
    struct prog_instruction *inst = new_instruction(emitInfo, OPCODE_ARL);
    inst->SrcReg[0].File = file;
    inst->SrcReg[0].Index = index;
-   inst->SrcReg[0].Swizzle = swizzle;
+   inst->SrcReg[0].Swizzle = fix_swizzle(swizzle);
    inst->DstReg.File = PROGRAM_ADDRESS;
    inst->DstReg.Index = 0;
    inst->DstReg.WriteMask = WRITEMASK_X;
@@ -717,6 +717,9 @@ instruction_annotation(gl_inst_opcode opcode, char *dstAnnot,
    case OPCODE_MUL:
       operator = "*";
       break;
+   case OPCODE_DP2:
+      operator = "DP2";
+      break;
    case OPCODE_DP3:
       operator = "DP3";
       break;
@@ -870,6 +873,7 @@ emit_compare(slang_emit_info *emitInfo, slang_ir_node *n)
 
    if (n->Children[0]->Store->Size != n->Children[1]->Store->Size) {
       slang_info_log_error(emitInfo->log, "invalid operands to == or !=");
+      n->Store = NULL;
       return NULL;
    }
 
@@ -899,6 +903,7 @@ emit_compare(slang_emit_info *emitInfo, slang_ir_node *n)
       slang_ir_storage tempStore;
 
       if (!alloc_local_temp(emitInfo, &tempStore, 4)) {
+         n->Store = NULL;
          return NULL;
          /* out of temps */
       }
@@ -913,7 +918,7 @@ emit_compare(slang_emit_info *emitInfo, slang_ir_node *n)
       }
       else {
          assert(size == 2);
-         dotOp = OPCODE_DP3;
+         dotOp = OPCODE_DP3; /* XXX use OPCODE_DP2 eventually */
          swizzle = MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_Y, SWIZZLE_Y, SWIZZLE_Y);
       }
 
@@ -1256,16 +1261,47 @@ emit_tex(slang_emit_info *emitInfo, slang_ir_node *n)
 {
    struct prog_instruction *inst;
    gl_inst_opcode opcode;
+   GLboolean shadow = GL_FALSE;
 
-   if (n->Opcode == IR_TEX) {
+   switch (n->Opcode) {
+   case IR_TEX:
       opcode = OPCODE_TEX;
-   }
-   else if (n->Opcode == IR_TEXB) {
+      break;
+   case IR_TEX_SH:
+      opcode = OPCODE_TEX;
+      shadow = GL_TRUE;
+      break;
+   case IR_TEXB:
       opcode = OPCODE_TXB;
-   }
-   else {
-      assert(n->Opcode == IR_TEXP);
+      break;
+   case IR_TEXB_SH:
+      opcode = OPCODE_TXB;
+      shadow = GL_TRUE;
+      break;
+   case IR_TEXP:
+      opcode = OPCODE_TXP;
+      break;
+   case IR_TEXP_SH:
       opcode = OPCODE_TXP;
+      shadow = GL_TRUE;
+      break;
+   default:
+      _mesa_problem(NULL, "Bad IR TEX code");
+      return NULL;
+   }
+
+   if (n->Children[0]->Opcode == IR_ELEMENT) {
+      /* array is the sampler (a uniform which'll indicate the texture unit) */
+      assert(n->Children[0]->Children[0]->Store);
+      assert(n->Children[0]->Children[0]->Store->File == PROGRAM_SAMPLER);
+
+      emit(emitInfo, n->Children[0]);
+
+      n->Children[0]->Var = n->Children[0]->Children[0]->Var;
+   } else {
+      /* this is the sampler (a uniform which'll indicate the texture unit) */
+      assert(n->Children[0]->Store);
+      assert(n->Children[0]->Store->File == PROGRAM_SAMPLER);
    }
 
    /* emit code for the texcoord operand */
@@ -1283,17 +1319,12 @@ emit_tex(slang_emit_info *emitInfo, slang_ir_node *n)
                            NULL,
                            NULL);
 
-   /* Child[0] is the sampler (a uniform which'll indicate the texture unit) */
-   assert(n->Children[0]->Store);
-   assert(n->Children[0]->Store->File == PROGRAM_SAMPLER);
-   /* Store->Index is the sampler index */
-   assert(n->Children[0]->Store->Index >= 0);
-   /* Store->Size is the texture target */
-   assert(n->Children[0]->Store->Size >= TEXTURE_1D_INDEX);
-   assert(n->Children[0]->Store->Size <= TEXTURE_RECT_INDEX);
+   inst->TexShadow = shadow;
 
-   inst->TexSrcTarget = n->Children[0]->Store->Size;
-   inst->TexSrcUnit = n->Children[0]->Store->Index; /* i.e. uniform's index */
+   /* Store->Index is the uniform/sampler index */
+   assert(n->Children[0]->Store->Index >= 0);
+   inst->TexSrcUnit = n->Children[0]->Store->Index;
+   inst->TexSrcTarget = n->Children[0]->Store->TexTarget;
 
    /* mark the sampler as being used */
    _mesa_use_uniform(emitInfo->prog->Parameters,
@@ -1339,14 +1370,16 @@ emit_copy(slang_emit_info *emitInfo, slang_ir_node *n)
 
    if (n->Store->File == PROGRAM_SAMPLER) {
       /* no code generated for sampler assignments,
-       * just copy the sampler index at compile time.
+       * just copy the sampler index/target at compile time.
        */
       n->Store->Index = n->Children[1]->Store->Index;
+      n->Store->TexTarget = n->Children[1]->Store->TexTarget;
       return NULL;
    }
 
 #if PEEPHOLE_OPTIMIZATIONS
    if (inst &&
+       (n->Children[1]->Opcode != IR_SWIZZLE) &&
        _slang_is_temp(emitInfo->vt, n->Children[1]->Store) &&
        (inst->DstReg.File == n->Children[1]->Store->File) &&
        (inst->DstReg.Index == n->Children[1]->Store->Index) &&
@@ -1363,13 +1396,9 @@ emit_copy(slang_emit_info *emitInfo, slang_ir_node *n)
        * becomes:
        *   MUL a, x, y;
        */
-      if (n->Children[1]->Opcode != IR_SWIZZLE)
-         _slang_free_temp(emitInfo->vt, n->Children[1]->Store);
-      *n->Children[1]->Store = *n->Children[0]->Store;
 
       /* fixup the previous instruction (which stored the RHS result) */
       assert(n->Children[0]->Store->Index >= 0);
-
       storage_to_dst_reg(&inst->DstReg, n->Children[0]->Store);
       return inst;
    }
@@ -1802,6 +1831,25 @@ emit_cont_break_if_true(slang_emit_info *emitInfo, slang_ir_node *n)
 }
 
 
+/**
+ * Return the size of a swizzle mask given that some swizzle components
+ * may be NIL/undefined.  For example:
+ *  swizzle_size(".zzxx") = 4
+ *  swizzle_size(".xy??") = 2
+ *  swizzle_size(".w???") = 1
+ */
+static GLuint
+swizzle_size(GLuint swizzle)
+{
+   GLuint i;
+   for (i = 0; i < 4; i++) {
+      if (GET_SWZ(swizzle, i) == SWIZZLE_NIL)
+         return i;
+   }
+   return 4;
+}
+
+
 static struct prog_instruction *
 emit_swizzle(slang_emit_info *emitInfo, slang_ir_node *n)
 {
@@ -1809,14 +1857,25 @@ emit_swizzle(slang_emit_info *emitInfo, slang_ir_node *n)
 
    inst = emit(emitInfo, n->Children[0]);
 
-#if 0
-   assert(n->Store->Parent);
-   /* Apply this node's swizzle to parent's storage */
-   GLuint swizzle = n->Store->Swizzle;
-   _slang_copy_ir_storage(n->Store, n->Store->Parent);
-   n->Store->Swizzle = _slang_swizzle_swizzle(n->Store->Swizzle, swizzle);
+   if (!n->Store->Parent) {
+      /* this covers a case such as "(b ? p : q).x" */
+      n->Store->Parent = n->Children[0]->Store;
+      assert(n->Store->Parent);
+   }
+
+   {
+      const GLuint swizzle = n->Store->Swizzle;
+      /* new storage is parent storage with updated Swizzle + Size fields */
+      _slang_copy_ir_storage(n->Store, n->Store->Parent);
+      /* Apply this node's swizzle to parent's storage */
+      n->Store->Swizzle = _slang_swizzle_swizzle(n->Store->Swizzle, swizzle);
+      /* Update size */
+      n->Store->Size = swizzle_size(n->Store->Swizzle);
+   }
+
    assert(!n->Store->Parent);
-#endif
+   assert(n->Store->Index >= 0);
+
    return inst;
 }
 
@@ -2109,6 +2168,10 @@ emit_var_ref(slang_emit_info *emitInfo, slang_ir_node *n)
       /* mark var as used */
       _mesa_use_uniform(emitInfo->prog->Parameters, (char *) n->Var->a_name);
    }
+   else if (n->Store->File == PROGRAM_INPUT) {
+      assert(n->Store->Index >= 0);
+      emitInfo->prog->InputsRead |= (1 << n->Store->Index);
+   }
 
    if (n->Store->Index < 0) {
       /* probably ran out of registers */
@@ -2192,12 +2255,15 @@ emit(slang_emit_info *emitInfo, slang_ir_node *n)
    case IR_NOISE2:
    case IR_NOISE3:
    case IR_NOISE4:
+   case IR_NRM4:
+   case IR_NRM3:
    /* binary */
    case IR_ADD:
    case IR_SUB:
    case IR_MUL:
    case IR_DOT4:
    case IR_DOT3:
+   case IR_DOT2:
    case IR_CROSS:
    case IR_MIN:
    case IR_MAX:
@@ -2221,6 +2287,9 @@ emit(slang_emit_info *emitInfo, slang_ir_node *n)
    case IR_TEX:
    case IR_TEXB:
    case IR_TEXP:
+   case IR_TEX_SH:
+   case IR_TEXB_SH:
+   case IR_TEXP_SH:
       return emit_tex(emitInfo, n);
    case IR_NEG:
       return emit_negation(emitInfo, n);
@@ -2364,10 +2433,20 @@ _slang_resolve_subroutines(slang_emit_info *emitInfo)
 
 
 
-
+/**
+ * Convert the IR tree into GPU instructions.
+ * \param n  root of IR tree
+ * \param vt  variable table
+ * \param prog  program to put GPU instructions into
+ * \param pragmas  controls codegen options
+ * \param withEnd  if true, emit END opcode at end
+ * \param log  log for emitting errors/warnings/info
+ */
 GLboolean
 _slang_emit_code(slang_ir_node *n, slang_var_table *vt,
-                 struct gl_program *prog, GLboolean withEnd,
+                 struct gl_program *prog,
+                 const struct gl_sl_pragmas *pragmas,
+                 GLboolean withEnd,
                  slang_info_log *log)
 {
    GET_CURRENT_CONTEXT(ctx);
@@ -2384,7 +2463,7 @@ _slang_emit_code(slang_ir_node *n, slang_var_table *vt,
 
    emitInfo.EmitHighLevelInstructions = ctx->Shader.EmitHighLevelInstructions;
    emitInfo.EmitCondCodes = ctx->Shader.EmitCondCodes;
-   emitInfo.EmitComments = ctx->Shader.EmitComments;
+   emitInfo.EmitComments = ctx->Shader.EmitComments || pragmas->Debug;
    emitInfo.EmitBeginEndSub = GL_TRUE;
 
    if (!emitInfo.EmitCondCodes) {
@@ -2400,7 +2479,9 @@ _slang_emit_code(slang_ir_node *n, slang_var_table *vt,
       maxUniforms = ctx->Const.VertexProgram.MaxUniformComponents / 4;
    }
    if (prog->Parameters->NumParameters > maxUniforms) {
-      slang_info_log_error(log, "Constant/uniform register limit exceeded");
+      slang_info_log_error(log, "Constant/uniform register limit exceeded "
+                           "(max=%u vec4)", maxUniforms);
+
       return GL_FALSE;
    }