intel/fs: Handle UNDEF in split_virtual_grfs

When the UNDEF instruction was added, we didn't do anything special in
split_virtual_grfs.  This mean that anything with an UNDEF wasn't
getting split which causes problems for the compiler.  Among other
things, it makes RA harder because things are in bigger chunks.  It also
meant that dvec4s weren't getting split which means that they are larger
than the maximum register size.

Shader-db results on Kaby Lake:

    total instructions in shared programs: 14959202 -> 14960035 (<.01%)
    instructions in affected programs: 96197 -> 97030 (0.87%)
    helped: 140
    HURT: 128
    helped stats (abs) min: 1 max: 17 x̄: 1.62 x̃: 1
    helped stats (rel) min: 0.09% max: 6.15% x̄: 0.65% x̃: 0.45%
    HURT stats (abs)   min: 1 max: 825 x̄: 8.28 x̃: 1
    HURT stats (rel)   min: 0.13% max: 139.83% x̄: 1.70% x̃: 0.50%
    95% mean confidence interval for instructions value: -2.96 9.18
    95% mean confidence interval for instructions %-change: -0.56% 1.51%
    Inconclusive result (value mean confidence interval includes 0).

    total loops in shared programs: 4372 -> 4372 (0.00%)
    loops in affected programs: 0 -> 0
    helped: 0
    HURT: 0

    total cycles in shared programs: 352646771 -> 352840997 (0.06%)
    cycles in affected programs: 218600800 -> 218795026 (0.09%)
    helped: 21167
    HURT: 21411
    helped stats (abs) min: 1 max: 2924 x̄: 36.89 x̃: 10
    helped stats (rel) min: <.01% max: 41.90% x̄: 2.97% x̃: 0.98%
    HURT stats (abs)   min: 1 max: 26027 x̄: 45.54 x̃: 10
    HURT stats (rel)   min: <.01% max: 324.46% x̄: 3.88% x̃: 1.06%
    95% mean confidence interval for cycles value: 2.87 6.26
    95% mean confidence interval for cycles %-change: 0.40% 0.55%
    Cycles are HURT.

    total spills in shared programs: 8840 -> 8953 (1.28%)
    spills in affected programs: 126 -> 239 (89.68%)
    helped: 1
    HURT: 2

    total fills in shared programs: 21782 -> 21914 (0.61%)
    fills in affected programs: 431 -> 563 (30.63%)
    helped: 1
    HURT: 3

    LOST:   0
    GAINED: 5

Shader-db results on Haswell:

    total instructions in shared programs: 13320918 -> 13320769 (<.01%)
    instructions in affected programs: 40998 -> 40849 (-0.36%)
    helped: 146
    HURT: 56
    helped stats (abs) min: 1 max: 8 x̄: 2.73 x̃: 2
    helped stats (rel) min: 0.16% max: 8.60% x̄: 2.52% x̃: 2.22%
    HURT stats (abs)   min: 2 max: 23 x̄: 4.45 x̃: 4
    HURT stats (rel)   min: 0.21% max: 10.26% x̄: 6.83% x̃: 10.26%
    95% mean confidence interval for instructions value: -1.26 -0.21
    95% mean confidence interval for instructions %-change: -0.62% 0.77%
    Inconclusive result (%-change mean confidence interval includes 0).

    total loops in shared programs: 4373 -> 4373 (0.00%)
    loops in affected programs: 0 -> 0
    helped: 0
    HURT: 0

    total cycles in shared programs: 374518258 -> 374384193 (-0.04%)
    cycles in affected programs: 231101954 -> 230967889 (-0.06%)
    helped: 21427
    HURT: 19438
    helped stats (abs) min: 1 max: 2035 x̄: 31.09 x̃: 8
    helped stats (rel) min: <.01% max: 40.95% x̄: 2.42% x̃: 0.86%
    HURT stats (abs)   min: 1 max: 20875 x̄: 27.38 x̃: 8
    HURT stats (rel)   min: <.01% max: 59.09% x̄: 2.49% x̃: 0.80%
    95% mean confidence interval for cycles value: -4.49 -2.07
    95% mean confidence interval for cycles %-change: -0.14% -0.04%
    Cycles are helped.

    total spills in shared programs: 23406 -> 23411 (0.02%)
    spills in affected programs: 3 -> 8 (166.67%)
    helped: 0
    HURT: 2

    total fills in shared programs: 34845 -> 34850 (0.01%)
    fills in affected programs: 3 -> 8 (166.67%)
    helped: 0
    HURT: 2

    LOST:   0
    GAINED: 0

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=111566
Fixes: f4ef34f207d1 "intel/fs: Add an UNDEF instruction to avoid..."
Reviewed-by: Francisco Jerez <currojerez@riseup.net>

diff --git a/src/intel/compiler/brw_fs.cpp b/src/intel/compiler/brw_fs.cpp
index f614fa858841b9de1b0b829d9da075729007ad6f..4d54dd0e4888cfc1b0353c58f8ed2e8451f30392 100644 (file)
--- a/src/intel/compiler/brw_fs.cpp
+++ b/src/intel/compiler/brw_fs.cpp
@@ -1910,6 +1910,17 @@ fs_visitor::split_virtual_grfs()
    }
 
    foreach_block_and_inst(block, fs_inst, inst, cfg) {
+      /* We fix up undef instructions later */
+      if (inst->opcode == SHADER_OPCODE_UNDEF) {
+         /* UNDEF instructions are currently only used to undef entire
+          * registers.  We need this invariant later when we split them.
+          */
+         assert(inst->dst.file == VGRF);
+         assert(inst->dst.offset == 0);
+         assert(inst->size_written == alloc.sizes[inst->dst.nr] * REG_SIZE);
+         continue;
+      }
+
       if (inst->dst.file == VGRF) {
          int reg = vgrf_to_reg[inst->dst.nr] + inst->dst.offset / REG_SIZE;
          for (unsigned j = 1; j < regs_written(inst); j++)
@@ -1962,7 +1973,20 @@ fs_visitor::split_virtual_grfs()
    }
    assert(reg == reg_count);
 
-   foreach_block_and_inst(block, fs_inst, inst, cfg) {
+   foreach_block_and_inst_safe(block, fs_inst, inst, cfg) {
+      if (inst->opcode == SHADER_OPCODE_UNDEF) {
+         const fs_builder ibld(this, block, inst);
+         assert(inst->size_written % REG_SIZE == 0);
+         unsigned reg_offset = 0;
+         while (reg_offset < inst->size_written / REG_SIZE) {
+            reg = vgrf_to_reg[inst->dst.nr] + reg_offset;
+            ibld.UNDEF(fs_reg(VGRF, new_virtual_grf[reg], inst->dst.type));
+            reg_offset += alloc.sizes[new_virtual_grf[reg]];
+         }
+         inst->remove(block);
+         continue;
+      }
+
       if (inst->dst.file == VGRF) {
          reg = vgrf_to_reg[inst->dst.nr] + inst->dst.offset / REG_SIZE;
          inst->dst.nr = new_virtual_grf[reg];