i965: Remove duplicate MRF writes in the FS backend.

This is quite common for multitexture sampling, and not only cuts down
on the second and later set of MOVs, but typically also allows
compute-to-MRF on the first set.

No statistically siginficant performance difference in nexuiz (n=3),
but it reduces instruction count in one of its shaders and seems like
a good idea.

src/mesa/drivers/dri/i965/brw_fs.cpp

@@ -177,6 +177,46 @@ type_size(const struct glsl_type *type)
    }
 }
 
+/**
+ * Returns how many MRFs an FS opcode will write over.
+ *
+ * Note that this is not the 0 or 1 implied writes in an actual gen
+ * instruction -- the FS opcodes often generate MOVs in addition.
+ */
+int
+fs_visitor::implied_mrf_writes(fs_inst *inst)
+{
+   if (inst->mlen == 0)
+      return 0;
+
+   switch (inst->opcode) {
+   case FS_OPCODE_RCP:
+   case FS_OPCODE_RSQ:
+   case FS_OPCODE_SQRT:
+   case FS_OPCODE_EXP2:
+   case FS_OPCODE_LOG2:
+   case FS_OPCODE_SIN:
+   case FS_OPCODE_COS:
+      return 1;
+   case FS_OPCODE_POW:
+      return 2;
+   case FS_OPCODE_TEX:
+   case FS_OPCODE_TXB:
+   case FS_OPCODE_TXL:
+      return 1;
+   case FS_OPCODE_FB_WRITE:
+      return 2;
+   case FS_OPCODE_PULL_CONSTANT_LOAD:
+   case FS_OPCODE_UNSPILL:
+      return 1;
+   case FS_OPCODE_SPILL:
+      return 2;
+   default:
+      assert(!"not reached");
+      return inst->mlen;
+   }
+}
+
 int
 fs_visitor::virtual_grf_alloc(int size)
 {
@@ -3066,6 +3106,78 @@ fs_visitor::compute_to_mrf()
    return progress;
 }
 
+/**
+ * Walks through basic blocks, locking for repeated MRF writes and
+ * removing the later ones.
+ */
+bool
+fs_visitor::remove_duplicate_mrf_writes()
+{
+   fs_inst *last_mrf_move[16];
+   bool progress = false;
+
+   memset(last_mrf_move, 0, sizeof(last_mrf_move));
+
+   foreach_iter(exec_list_iterator, iter, this->instructions) {
+      fs_inst *inst = (fs_inst *)iter.get();
+
+      switch (inst->opcode) {
+      case BRW_OPCODE_DO:
+      case BRW_OPCODE_WHILE:
+      case BRW_OPCODE_IF:
+      case BRW_OPCODE_ELSE:
+      case BRW_OPCODE_ENDIF:
+	 memset(last_mrf_move, 0, sizeof(last_mrf_move));
+	 continue;
+      default:
+	 break;
+      }
+
+      if (inst->opcode == BRW_OPCODE_MOV &&
+	  inst->dst.file == MRF) {
+	 fs_inst *prev_inst = last_mrf_move[inst->dst.hw_reg];
+	 if (prev_inst && inst->equals(prev_inst)) {
+	    inst->remove();
+	    progress = true;
+	    continue;
+	 }
+      }
+
+      /* Clear out the last-write records for MRFs that were overwritten. */
+      if (inst->dst.file == MRF) {
+	 last_mrf_move[inst->dst.hw_reg] = NULL;
+      }
+
+      if (inst->mlen > 0) {
+	 /* Found a SEND instruction, which will include two of fewer
+	  * implied MRF writes.  We could do better here.
+	  */
+	 for (int i = 0; i < implied_mrf_writes(inst); i++) {
+	    last_mrf_move[inst->base_mrf + i] = NULL;
+	 }
+      }
+
+      /* Clear out any MRF move records whose sources got overwritten. */
+      if (inst->dst.file == GRF) {
+	 for (unsigned int i = 0; i < Elements(last_mrf_move); i++) {
+	    if (last_mrf_move[i] &&
+		last_mrf_move[i]->src[0].reg == inst->dst.reg) {
+	       last_mrf_move[i] = NULL;
+	    }
+	 }
+      }
+
+      if (inst->opcode == BRW_OPCODE_MOV &&
+	  inst->dst.file == MRF &&
+	  inst->src[0].file == GRF &&
+	  !inst->predicated) {
+	 last_mrf_move[inst->dst.hw_reg] = inst;
+      }
+   }
+
+   return progress;
+}
+
 bool
 fs_visitor::virtual_grf_interferes(int a, int b)
 {
@@ -3438,6 +3550,9 @@ brw_wm_fs_emit(struct brw_context *brw, struct brw_wm_compile *c)
       bool progress;
       do {
 	 progress = false;
+
+	 progress = v.remove_duplicate_mrf_writes() || progress;
+
 	 v.calculate_live_intervals();
 	 progress = v.propagate_constants() || progress;
 	 progress = v.register_coalesce() || progress;

src/mesa/drivers/dri/i965/brw_fs.h

@@ -148,6 +148,21 @@ public:
    fs_reg(enum register_file file, int hw_reg, uint32_t type);
    fs_reg(class fs_visitor *v, const struct glsl_type *type);
 
+   bool equals(fs_reg *r)
+   {
+      return (file == r->file &&
+	      reg == r->reg &&
+	      reg_offset == r->reg_offset &&
+	      hw_reg == r->hw_reg &&
+	      type == r->type &&
+	      negate == r->negate &&
+	      abs == r->abs &&
+	      memcmp(&fixed_hw_reg, &r->fixed_hw_reg,
+		     sizeof(fixed_hw_reg)) == 0 &&
+	      smear == r->smear &&
+	      imm.u == r->imm.u);
+   }
+
    /** Register file: ARF, GRF, MRF, IMM. */
    enum register_file file;
    /** virtual register number.  0 = fixed hw reg */
@@ -270,6 +285,26 @@ public:
 	 assert(src[2].reg_offset >= 0);
    }
 
+   bool equals(fs_inst *inst)
+   {
+      return (opcode == inst->opcode &&
+	      dst.equals(&inst->dst) &&
+	      src[0].equals(&inst->src[0]) &&
+	      src[1].equals(&inst->src[1]) &&
+	      src[2].equals(&inst->src[2]) &&
+	      saturate == inst->saturate &&
+	      predicated == inst->predicated &&
+	      conditional_mod == inst->conditional_mod &&
+	      mlen == inst->mlen &&
+	      base_mrf == inst->base_mrf &&
+	      sampler == inst->sampler &&
+	      target == inst->target &&
+	      eot == inst->eot &&
+	      header_present == inst->header_present &&
+	      shadow_compare == inst->shadow_compare &&
+	      offset == inst->offset);
+   }
+
    int opcode; /* BRW_OPCODE_* or FS_OPCODE_* */
    fs_reg dst;
    fs_reg src[3];
@@ -372,6 +407,7 @@ public:
    bool register_coalesce();
    bool compute_to_mrf();
    bool dead_code_eliminate();
+   bool remove_duplicate_mrf_writes();
    bool virtual_grf_interferes(int a, int b);
    void generate_code();
    void generate_fb_write(fs_inst *inst);
@@ -409,6 +445,7 @@ public:
    struct brw_reg interp_reg(int location, int channel);
    int setup_uniform_values(int loc, const glsl_type *type);
    void setup_builtin_uniform_values(ir_variable *ir);
+   int implied_mrf_writes(fs_inst *inst);
 
    struct brw_context *brw;
    const struct gl_fragment_program *fp;