[Mesa-dev] [PATCH v2 098/103] i965/vec4: support basic spilling of 64-bit registers

[Iago Toral Quiroga]

The current spilling code can't spill vgrf allocations larger than 1
but SIMD4x2 doubles require 2 vgrfs, so we need to permit this case (which
is handled properly for DF data types by emitting 2 scratch messages and
doing data shuffling). We accomplish this by not auto-disabling spilling
for vgrf allocations with a size of 2, and then disable spilling on any
register with an offset != 0B (which indicates array access).

Disable spilling of partial DF reads/writes because these don't read/write
data for both logical threads and our scratch messages for 64-bit data
need data for both threads to be present.
---
 .../drivers/dri/i965/brw_vec4_reg_allocate.cpp     | 34 ++++++++++++++++++----
 1 file changed, 28 insertions(+), 6 deletions(-)

diff --git a/src/mesa/drivers/dri/i965/brw_vec4_reg_allocate.cpp b/src/mesa/drivers/dri/i965/brw_vec4_reg_allocate.cpp
index 228e04c..7aff2d8 100644
--- a/src/mesa/drivers/dri/i965/brw_vec4_reg_allocate.cpp
+++ b/src/mesa/drivers/dri/i965/brw_vec4_reg_allocate.cpp
@@ -376,7 +376,7 @@ vec4_visitor::evaluate_spill_costs(float *spill_costs, bool *no_spill)
 
    for (unsigned i = 0; i < this->alloc.count; i++) {
       spill_costs[i] = 0.0;
-      no_spill[i] = alloc.sizes[i] != 1;
+      no_spill[i] = alloc.sizes[i] != 1 && alloc.sizes[i] != 2;
    }
 
    /* Calculate costs for spilling nodes.  Call it a cost of 1 per
@@ -393,7 +393,17 @@ vec4_visitor::evaluate_spill_costs(float *spill_costs, bool *no_spill)
             if (!can_use_scratch_for_source(inst, i, inst->src[i].nr)) {
                spill_costs[inst->src[i].nr] += loop_scale;
                if (inst->src[i].reladdr ||
-                   inst->src[i].offset % REG_SIZE != 0)
+                   inst->src[i].offset >= REG_SIZE)
+                  no_spill[inst->src[i].nr] = true;
+
+               /* We don't support unspills of partial DF reads.
+                *
+                * Our 64-bit unspills are implemented with two 32-bit scratch
+                * messages, each one reading that for both SIMD4x2 threads that
+                * we need to shuffle into correct 64-bit data. Ensure that we
+                * are reading data for both threads.
+                */
+               if (type_sz(inst->src[i].type) == 8 && inst->exec_size != 8)
                   no_spill[inst->src[i].nr] = true;
             }
          }
@@ -401,7 +411,16 @@ vec4_visitor::evaluate_spill_costs(float *spill_costs, bool *no_spill)
 
       if (inst->dst.file == VGRF) {
          spill_costs[inst->dst.nr] += loop_scale;
-         if (inst->dst.reladdr || inst->dst.offset % REG_SIZE != 0)
+         if (inst->dst.reladdr || inst->dst.offset >= REG_SIZE)
+            no_spill[inst->dst.nr] = true;
+
+         /* We don't support spills of partial DF writes.
+          *
+          * Our 64-bit spills are implemented with two 32-bit scratch messages,
+          * each one writing that for both SIMD4x2 threads. Ensure that we
+          * are writing data for both threads.
+          */
+         if (type_sz(inst->dst.type) == 8 && inst->exec_size != 8)
             no_spill[inst->dst.nr] = true;
       }
 
@@ -450,8 +469,9 @@ vec4_visitor::choose_spill_reg(struct ra_graph *g)
 void
 vec4_visitor::spill_reg(int spill_reg_nr)
 {
-   assert(alloc.sizes[spill_reg_nr] == 1);
-   unsigned int spill_offset = last_scratch++;
+   assert(alloc.sizes[spill_reg_nr] == 1 || alloc.sizes[spill_reg_nr] == 2);
+   unsigned int spill_offset = last_scratch;
+   last_scratch += alloc.sizes[spill_reg_nr];
 
    /* Generate spill/unspill instructions for the objects being spilled. */
    int scratch_reg = -1;
@@ -465,12 +485,14 @@ vec4_visitor::spill_reg(int spill_reg_nr)
                 * for consecutive instructions that read different channels of
                 * the same vec4.
                 */
-               scratch_reg = alloc.allocate(1);
+               scratch_reg = alloc.allocate(alloc.sizes[spill_reg_nr]);
                src_reg temp = inst->src[i];
                temp.nr = scratch_reg;
+               temp.offset = 0;
                temp.swizzle = BRW_SWIZZLE_XYZW;
                emit_scratch_read(block, inst,
                                  dst_reg(temp), inst->src[i], spill_offset);
+               temp.offset = inst->src[i].offset;
             }
             assert(scratch_reg != -1);
             inst->src[i].nr = scratch_reg;
-- 
2.7.4