i965/fs: Make register spill/unspill only do the regs for that instruction.

Previously, if we were spilling the result of a texture call, we would store
all 4 regs, then for each use of one of those regs as the source of an
instruction, we would unspill all 4 regs even though only one was needed.

In both lightsmark and l4d2 with my current graphics config, the shaders that
produce spilling do so on split GRFs, so this doesn't help them out.  However,
in a capture of the l4d2 shaders with a different snapshot and playing the
game instead of using a demo, it reduced one shader from 2817 instructions to
2179, due to choosing a now-cheaper texture result to spill instead of piles
of texcoords.

v2: Fix comment noted by Ken, and fix the if condition associated with it for
    the current state of what constitutes a partial write of the destination.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org> (v1)

1 files changed, 33 insertions, 33 deletions

diff --git a/src/mesa/drivers/dri/i965/brw_fs_reg_allocate.cpp b/src/mesa/drivers/dri/i965/brw_fs_reg_allocate.cpp
index 3f10ca6fc93..7618047df58 100644
--- a/src/mesa/drivers/dri/i965/brw_fs_reg_allocate.cpp
+++ b/src/mesa/drivers/dri/i965/brw_fs_reg_allocate.cpp
@@ -283,20 +283,13 @@ fs_visitor::emit_unspill(fs_inst *inst, fs_reg dst, uint32_t spill_offset)
 {
-   int size = virtual_grf_sizes[dst.reg];
-   dst.reg_offset = 0;
-
-   for (int chan = 0; chan < size; chan++) {
-      fs_inst *unspill_inst = new(mem_ctx) fs_inst(FS_OPCODE_UNSPILL,
-						   dst);
-      dst.reg_offset++;
-      unspill_inst->offset = spill_offset + chan * REG_SIZE;
-      unspill_inst->ir = inst->ir;
-      unspill_inst->annotation = inst->annotation;
-
-      /* Choose a MRF that won't conflict with an MRF that's live across the
-       * spill.  Nothing else will make it up to MRF 14/15.
-       */
-      unspill_inst->base_mrf = 14;
-      unspill_inst->mlen = 1; /* header contains offset */
-      inst->insert_before(unspill_inst);
-   }
+   fs_inst *unspill_inst = new(mem_ctx) fs_inst(FS_OPCODE_UNSPILL, dst);
+   unspill_inst->offset = spill_offset;
+   unspill_inst->ir = inst->ir;
+   unspill_inst->annotation = inst->annotation;
+
+   /* Choose a MRF that won't conflict with an MRF that's live across the
+    * spill.  Nothing else will make it up to MRF 14/15.
+    */
+   unspill_inst->base_mrf = 14;
+   unspill_inst->mlen = 1; /* header contains offset */
+   inst->insert_before(unspill_inst);
 }
@@ -324,4 +317,3 @@ fs_visitor::choose_spill_reg(struct ra_graph *g)
 	 if (inst->src[i].file == GRF) {
-	    int size = virtual_grf_sizes[inst->src[i].reg];
-	    spill_costs[inst->src[i].reg] += size * loop_scale;
+	    spill_costs[inst->src[i].reg] += loop_scale;
 	 }
@@ -330,4 +322,3 @@ fs_visitor::choose_spill_reg(struct ra_graph *g)
       if (inst->dst.file == GRF) {
-	 int size = virtual_grf_sizes[inst->dst.reg];
-	 spill_costs[inst->dst.reg] += size * loop_scale;
+	 spill_costs[inst->dst.reg] += inst->regs_written() * loop_scale;
       }
@@ -386,4 +377,5 @@ fs_visitor::spill_reg(int spill_reg)
 	     inst->src[i].reg == spill_reg) {
-	    inst->src[i].reg = virtual_grf_alloc(size);
-	    emit_unspill(inst, inst->src[i], spill_offset);
+	    inst->src[i].reg = virtual_grf_alloc(1);
+	    emit_unspill(inst, inst->src[i],
+                         spill_offset + REG_SIZE * inst->src[i].reg_offset);
 	 }
@@ -393,10 +385,18 @@ fs_visitor::spill_reg(int spill_reg)
 	  inst->dst.reg == spill_reg) {
-	 inst->dst.reg = virtual_grf_alloc(size);
-
-	 /* Since we spill/unspill the whole thing even if we access
-	  * just a component, we may need to unspill before the
-	  * instruction we're spilling for.
+         int subset_spill_offset = (spill_offset +
+                                    REG_SIZE * inst->dst.reg_offset);
+         inst->dst.reg = virtual_grf_alloc(inst->regs_written());
+         inst->dst.reg_offset = 0;
+
+	 /* If our write is going to affect just part of the
+          * inst->regs_written(), then we need to unspill the destination
+          * since we write back out all of the regs_written().
 	  */
-	 if (size != 1 || inst->predicated) {
-	    emit_unspill(inst, inst->dst, spill_offset);
+	 if (inst->predicated || inst->force_uncompressed || inst->force_sechalf) {
+            fs_reg unspill_reg = inst->dst;
+            for (int chan = 0; chan < inst->regs_written(); chan++) {
+               emit_unspill(inst, unspill_reg,
+                            subset_spill_offset + REG_SIZE * chan);
+               unspill_reg.reg_offset++;
+            }
 	 }
@@ -409,3 +409,3 @@ fs_visitor::spill_reg(int spill_reg)
 
-	 for (int chan = 0; chan < size; chan++) {
+	 for (int chan = 0; chan < inst->regs_written(); chan++) {
 	    fs_inst *spill_inst = new(mem_ctx) fs_inst(FS_OPCODE_SPILL,
@@ -413,3 +413,3 @@ fs_visitor::spill_reg(int spill_reg)
 	    spill_src.reg_offset++;
-	    spill_inst->offset = spill_offset + chan * REG_SIZE;
+	    spill_inst->offset = subset_spill_offset + chan * REG_SIZE;
 	    spill_inst->ir = inst->ir;