i965/nir: Run the ffma peephole after the rest of the optimizations

The idea here is that fusing multiply-add combinations too early can reduce
our ability to perform CSE and value-numbering.  Instead, we split ffma
opcodes up-front, hope CSE cleans up, and then fuse after-the-fact.
Unless an algebraic pass does something silly where it inserts something
between the multiply and the add, splitting and re-fusing should never
cause a problem.  We run the late algebraic optimizations after this so
that things like compare-with-zero don't hurt our ability to fuse things.

shader-db results for fragment shaders on Haswell:
total instructions in shared programs: 4390538 -> 4379236 (-0.26%)
instructions in affected programs:     989359 -> 978057 (-1.14%)
helped:                                5308
HURT:                                  97
GAINED:                                78
LOST:                                  5

This does, unfortunately, cause some substantial hurt to a shader in Kerbal
Space Program.  However, the damage is caused by changing a single
instruction from a ffma to an add.  This, in turn, *decreases* register
pressure in one part of the program causing it to fail to register allocate
and spill.  Given the overwhelmingly positive results in other shaders and
the fact that the NIR for the Kerbal shaders is actually better, this
should be considered a positive.

Reviewed-by: Matt Turner <mattst88@gmail.com>

diff --git a/src/mesa/drivers/dri/i965/brw_context.c b/src/mesa/drivers/dri/i965/brw_context.c
index ed6fdffd265c8a53d87332d0853e6c98686c8a77..21c8bd331b55304941e0bd95cb34150869a3d9ec 100644 (file)
--- a/src/mesa/drivers/dri/i965/brw_context.c
+++ b/src/mesa/drivers/dri/i965/brw_context.c
@@ -558,6 +558,11 @@ brw_initialize_context_constants(struct brw_context *brw)
 
    static const nir_shader_compiler_options gen6_nir_options = {
       .native_integers = true,
+      /* In order to help allow for better CSE at the NIR level we tell NIR
+       * to split all ffma instructions during opt_algebraic and we then
+       * re-combine them as a later step.
+       */
+      .lower_ffma = true,
    };
 
    /* We want the GLSL compiler to emit code that uses condition codes */

diff --git a/src/mesa/drivers/dri/i965/brw_fs_nir.cpp b/src/mesa/drivers/dri/i965/brw_fs_nir.cpp
index 4f4b74620fed045b81da4462bf1c353a59f32482..94641cf2ec1e00ec21e02013ae7fbfdb2b244e86 100644 (file)
--- a/src/mesa/drivers/dri/i965/brw_fs_nir.cpp
+++ b/src/mesa/drivers/dri/i965/brw_fs_nir.cpp
@@ -52,8 +52,6 @@ nir_optimize(nir_shader *nir)
       nir_validate_shader(nir);
       progress |= nir_opt_algebraic(nir);
       nir_validate_shader(nir);
-      progress |= nir_opt_peephole_ffma(nir);
-      nir_validate_shader(nir);
       progress |= nir_opt_constant_folding(nir);
       nir_validate_shader(nir);
       progress |= nir_opt_remove_phis(nir);
@@ -149,6 +147,12 @@ fs_visitor::emit_nir_code()
 
    nir_optimize(nir);
 
+   if (brw->gen >= 6) {
+      /* Try and fuse multiply-adds */
+      nir_opt_peephole_ffma(nir);
+      nir_validate_shader(nir);
+   }
+
    nir_opt_algebraic_late(nir);
    nir_validate_shader(nir);