i965: Don't upload clip planes when gl_ClipDistance is in use.

[mesa.git] / src / mesa / drivers / dri / i965 / brw_vec4_emit.cpp

diff --git a/src/mesa/drivers/dri/i965/brw_vec4_emit.cpp b/src/mesa/drivers/dri/i965/brw_vec4_emit.cpp
index 61a1092bbff187db28586fd422ac9f9580acffde..b58ebc69b4c4a6d821aade994012210d067ade7f 100644 (file)
--- a/src/mesa/drivers/dri/i965/brw_vec4_emit.cpp
+++ b/src/mesa/drivers/dri/i965/brw_vec4_emit.cpp
@@ -93,7 +93,7 @@ vec4_visitor::setup_uniforms(int reg)
 {
    /* User clip planes from curbe:
     */
-   if (c->key.nr_userclip) {
+   if (c->key.nr_userclip && !c->key.uses_clip_distance) {
       if (intel->gen >= 6) {
         for (int i = 0; i < c->key.nr_userclip; i++) {
            c->userplane[i] = stride(brw_vec4_grf(reg + i / 2,
@@ -165,6 +165,12 @@ vec4_instruction::get_dst(void)
       brw_reg.dw1.bits.writemask = dst.writemask;
       break;
 
+   case MRF:
+      brw_reg = brw_message_reg(dst.reg + dst.reg_offset);
+      brw_reg = retype(brw_reg, dst.type);
+      brw_reg.dw1.bits.writemask = dst.writemask;
+      break;
+
    case HW_REG:
       brw_reg = dst.fixed_hw_reg;
       break;
@@ -391,14 +397,9 @@ vec4_visitor::generate_scratch_read(vec4_instruction *inst,
                                    struct brw_reg dst,
                                    struct brw_reg index)
 {
-   if (intel->gen >= 6) {
-      brw_push_insn_state(p);
-      brw_set_mask_control(p, BRW_MASK_DISABLE);
-      brw_MOV(p,
-             retype(brw_message_reg(inst->base_mrf), BRW_REGISTER_TYPE_D),
-             retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_D));
-      brw_pop_insn_state(p);
-   }
+   struct brw_reg header = brw_vec8_grf(0, 0);
+
+   gen6_resolve_implied_move(p, &header, inst->base_mrf);
 
    generate_oword_dual_block_offsets(brw_message_reg(inst->base_mrf + 1),
                                     index);
@@ -417,7 +418,9 @@ vec4_visitor::generate_scratch_read(vec4_instruction *inst,
     */
    struct brw_instruction *send = brw_next_insn(p, BRW_OPCODE_SEND);
    brw_set_dest(p, send, dst);
-   brw_set_src0(p, send, brw_message_reg(inst->base_mrf));
+   brw_set_src0(p, send, header);
+   if (intel->gen < 6)
+      send->header.destreg__conditionalmod = inst->base_mrf;
    brw_set_dp_read_message(p, send,
                           255, /* binding table index: stateless access */
                           BRW_DATAPORT_OWORD_DUAL_BLOCK_1OWORD,
@@ -433,19 +436,15 @@ vec4_visitor::generate_scratch_write(vec4_instruction *inst,
                                     struct brw_reg src,
                                     struct brw_reg index)
 {
+   struct brw_reg header = brw_vec8_grf(0, 0);
+   bool write_commit;
+
    /* If the instruction is predicated, we'll predicate the send, not
     * the header setup.
     */
    brw_set_predicate_control(p, false);
 
-   if (intel->gen >= 6) {
-      brw_push_insn_state(p);
-      brw_set_mask_control(p, BRW_MASK_DISABLE);
-      brw_MOV(p,
-             retype(brw_message_reg(inst->base_mrf), BRW_REGISTER_TYPE_D),
-             retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_D));
-      brw_pop_insn_state(p);
-   }
+   gen6_resolve_implied_move(p, &header, inst->base_mrf);
 
    generate_oword_dual_block_offsets(brw_message_reg(inst->base_mrf + 1),
                                     index);
@@ -463,12 +462,33 @@ vec4_visitor::generate_scratch_write(vec4_instruction *inst,
 
    brw_set_predicate_control(p, inst->predicate);
 
+   /* Pre-gen6, we have to specify write commits to ensure ordering
+    * between reads and writes within a thread.  Afterwards, that's
+    * guaranteed and write commits only matter for inter-thread
+    * synchronization.
+    */
+   if (intel->gen >= 6) {
+      write_commit = false;
+   } else {
+      /* The visitor set up our destination register to be g0.  This
+       * means that when the next read comes along, we will end up
+       * reading from g0 and causing a block on the write commit.  For
+       * write-after-read, we are relying on the value of the previous
+       * read being used (and thus blocking on completion) before our
+       * write is executed.  This means we have to be careful in
+       * instruction scheduling to not violate this assumption.
+       */
+      write_commit = true;
+   }
+
    /* Each of the 8 channel enables is considered for whether each
     * dword is written.
     */
    struct brw_instruction *send = brw_next_insn(p, BRW_OPCODE_SEND);
    brw_set_dest(p, send, dst);
-   brw_set_src0(p, send, brw_message_reg(inst->base_mrf));
+   brw_set_src0(p, send, header);
+   if (intel->gen < 6)
+      send->header.destreg__conditionalmod = inst->base_mrf;
    brw_set_dp_write_message(p, send,
                            255, /* binding table index: stateless access */
                            BRW_DATAPORT_OWORD_DUAL_BLOCK_1OWORD,
@@ -476,9 +496,9 @@ vec4_visitor::generate_scratch_write(vec4_instruction *inst,
                            3, /* mlen */
                            true, /* header present */
                            false, /* pixel scoreboard */
-                           0, /* rlen */
+                           write_commit, /* rlen */
                            false, /* eot */
-                           false /* commit */);
+                           write_commit);
 }
 
 void
@@ -508,6 +528,8 @@ vec4_visitor::generate_pull_constant_load(vec4_instruction *inst,
    struct brw_instruction *send = brw_next_insn(p, BRW_OPCODE_SEND);
    brw_set_dest(p, send, dst);
    brw_set_src0(p, send, header);
+   if (intel->gen < 6)
+      send->header.destreg__conditionalmod = inst->base_mrf;
    brw_set_dp_read_message(p, send,
                           SURF_INDEX_VERT_CONST_BUFFER,
                           BRW_DATAPORT_OWORD_DUAL_BLOCK_1OWORD,
@@ -591,14 +613,18 @@ vec4_visitor::run()
     */
    move_grf_array_access_to_scratch();
    move_uniform_array_access_to_pull_constants();
+   pack_uniform_registers();
+   move_push_constants_to_pull_constants();
 
    bool progress;
    do {
       progress = false;
       progress = dead_code_eliminate() || progress;
+      progress = opt_copy_propagation() || progress;
+      progress = opt_algebraic() || progress;
+      progress = opt_compute_to_mrf() || progress;
    } while (progress);
 
-   pack_uniform_registers();
 
    if (failed)
       return false;