intel/fs: Emit better code for bfi(..., 0)

DG2, Tiger Lake, Ice Lake, and Skylake had similar results (Ice Lake shown)
total instructions in shared programs: 20570141 -> 20570063 (<.01%)
instructions in affected programs: 30679 -> 30601 (-0.25%)
helped: 77 / HURT: 0

total cycles in shared programs: 902113977 -> 902118723 (<.01%)
cycles in affected programs: 3255958 -> 3260704 (0.15%)
helped: 60 / HURT: 19

Broadwell
total instructions in shared programs: 18524633 -> 18524547 (<.01%)
instructions in affected programs: 34095 -> 34009 (-0.25%)
helped: 75 / HURT: 2

total cycles in shared programs: 949532394 -> 949543761 (<.01%)
cycles in affected programs: 3419107 -> 3430474 (0.33%)
helped: 57 / HURT: 24

total spills in shared programs: 22484 -> 22484 (0.00%)
spills in affected programs: 516 -> 516 (0.00%)
helped: 2 / HURT: 2

total fills in shared programs: 29346 -> 29338 (-0.03%)
fills in affected programs: 572 -> 564 (-1.40%)
helped: 4 / HURT: 0

Haswell
total instructions in shared programs: 17331356 -> 17331523 (<.01%)
instructions in affected programs: 27920 -> 28087 (0.60%)
helped: 41 / HURT: 4

total cycles in shared programs: 936603192 -> 936574664 (<.01%)
cycles in affected programs: 3417695 -> 3389167 (-0.83%)
helped: 28 / HURT: 21

total spills in shared programs: 19718 -> 19756 (0.19%)
spills in affected programs: 436 -> 474 (8.72%)
helped: 0 / HURT: 4

total fills in shared programs: 22547 -> 22607 (0.27%)
fills in affected programs: 444 -> 504 (13.51%)
helped: 0 / HURT: 4

Ivy Bridge
total cycles in shared programs: 463451277 -> 463451273 (<.01%)
cycles in affected programs: 95870 -> 95866 (<.01%)
helped: 3 / HURT: 2

DG2, Tiger Lake, Ice Lake, and Skylake had similar results (Ice Lake shown)
Totals:
Instrs: 152825278 -> 152819969 (-0.00%); split: -0.00%, +0.00%
Cycles: 15014075626 -> 15014628652 (+0.00%); split: -0.01%, +0.01%
Subgroup size: 8528536 -> 8528560 (+0.00%)
Send messages: 7711431 -> 7711464 (+0.00%)
Spill count: 99907 -> 99509 (-0.40%); split: -0.40%, +0.00%
Fill count: 202459 -> 201598 (-0.43%); split: -0.43%, +0.00%
Scratch Memory Size: 4376576 -> 4371456 (-0.12%)

Totals from 2915 (0.44% of 662497) affected shaders:
Instrs: 2288842 -> 2283533 (-0.23%); split: -0.24%, +0.01%
Cycles: 471633295 -> 472186321 (+0.12%); split: -0.27%, +0.39%
Subgroup size: 27488 -> 27512 (+0.09%)
Send messages: 151344 -> 151377 (+0.02%)
Spill count: 48091 -> 47693 (-0.83%); split: -0.83%, +0.00%
Fill count: 59053 -> 58192 (-1.46%); split: -1.46%, +0.00%
Scratch Memory Size: 1827840 -> 1822720 (-0.28%)

Reviewed-by: Matt Turner <mattst88@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19968>

diff --git a/src/intel/compiler/brw_fs_nir.cpp b/src/intel/compiler/brw_fs_nir.cpp
index 073a1ad..0c48d10 100644 (file)
--- a/src/intel/compiler/brw_fs_nir.cpp
+++ b/src/intel/compiler/brw_fs_nir.cpp
@@ -936,6 +936,12 @@ can_fuse_fmul_fsign(nir_alu_instr *instr, unsigned fsign_src)
           is_used_once(fsign_instr);
 }
 
+static bool
+is_const_zero(const nir_src &src)
+{
+   return nir_src_is_const(src) && nir_src_as_int(src) == 0;
+}
+
 void
 fs_visitor::nir_emit_alu(const fs_builder &bld, nir_alu_instr *instr,
                          bool need_dest)
@@ -1733,7 +1739,16 @@ fs_visitor::nir_emit_alu(const fs_builder &bld, nir_alu_instr *instr,
       break;
    case nir_op_bfi:
       assert(nir_dest_bit_size(instr->dest.dest) < 64);
-      bld.BFI2(result, op[0], op[1], op[2]);
+
+      /* bfi is ((...) | (~src0 & src2)). The second part is zero when src2 is
+       * either 0 or src0. Replacing the 0 with another value can eliminate a
+       * temporary register.
+       */
+      if (is_const_zero(instr->src[2].src))
+         bld.BFI2(result, op[0], op[1], op[0]);
+      else
+         bld.BFI2(result, op[0], op[1], op[2]);
+
       break;
 
    case nir_op_bitfield_insert: