nir/opt_peephole_select: Don't peephole_select expensive math instructions

On some GPUs, especially older Intel GPUs, some math instructions are
very expensive.  On those architectures, don't reduce flow control to a
csel if one of the branches contains one of these expensive math
instructions.

This prevents a bunch of cycle count regressions on pre-Gen6 platforms
with a later patch (intel/compiler: More peephole select for pre-Gen6).

v2: Remove stray #if block.  Noticed by Thomas.

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Thomas Helland <thomashelland90@gmail.com>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>

9 files changed, 40 insertions, 17 deletions

diff --git a/src/amd/vulkan/radv_shader.c b/src/amd/vulkan/radv_shader.c
index f778e85b8d9..2eb5164ac97 100644
--- a/src/amd/vulkan/radv_shader.c
+++ b/src/amd/vulkan/radv_shader.c
@@ -159,7 +159,7 @@ radv_optimize_nir(struct nir_shader *shader, bool optimize_conservatively,
                 NIR_PASS(progress, shader, nir_opt_if);
                 NIR_PASS(progress, shader, nir_opt_dead_cf);
                 NIR_PASS(progress, shader, nir_opt_cse);
-                NIR_PASS(progress, shader, nir_opt_peephole_select, 8, true);
+                NIR_PASS(progress, shader, nir_opt_peephole_select, 8, true, true);
                 NIR_PASS(progress, shader, nir_opt_algebraic);
                 NIR_PASS(progress, shader, nir_opt_constant_folding);
                 NIR_PASS(progress, shader, nir_opt_undef);
diff --git a/src/broadcom/compiler/nir_to_vir.c b/src/broadcom/compiler/nir_to_vir.c
index 9b1066467cc..ada0dd69921 100644
--- a/src/broadcom/compiler/nir_to_vir.c
+++ b/src/broadcom/compiler/nir_to_vir.c
@@ -1241,7 +1241,7 @@ v3d_optimize_nir(struct nir_shader *s)
                 NIR_PASS(progress, s, nir_opt_dce);
                 NIR_PASS(progress, s, nir_opt_dead_cf);
                 NIR_PASS(progress, s, nir_opt_cse);
-                NIR_PASS(progress, s, nir_opt_peephole_select, 8, true);
+                NIR_PASS(progress, s, nir_opt_peephole_select, 8, true, true);
                 NIR_PASS(progress, s, nir_opt_algebraic);
                 NIR_PASS(progress, s, nir_opt_constant_folding);
                 NIR_PASS(progress, s, nir_opt_undef);
diff --git a/src/compiler/nir/nir.h b/src/compiler/nir/nir.h
index 2bbfb3c6b15..54f9c64a3a3 100644
--- a/src/compiler/nir/nir.h
+++ b/src/compiler/nir/nir.h
@@ -3198,7 +3198,7 @@ bool nir_opt_move_comparisons(nir_shader *shader);
 bool nir_opt_move_load_ubo(nir_shader *shader);
 
 bool nir_opt_peephole_select(nir_shader *shader, unsigned limit,
-                             bool indirect_load_ok);
+                             bool indirect_load_ok, bool expensive_alu_ok);
 
 bool nir_opt_remove_phis_impl(nir_function_impl *impl);
 bool nir_opt_remove_phis(nir_shader *shader);
diff --git a/src/compiler/nir/nir_opt_peephole_select.c b/src/compiler/nir/nir_opt_peephole_select.c
index 6808d3eda6c..b27cdd28a99 100644
--- a/src/compiler/nir/nir_opt_peephole_select.c
+++ b/src/compiler/nir/nir_opt_peephole_select.c
@@ -59,7 +59,8 @@
 
 static bool
 block_check_for_allowed_instrs(nir_block *block, unsigned *count,
-                               bool alu_ok, bool indirect_load_ok)
+                               bool alu_ok, bool indirect_load_ok,
+                               bool expensive_alu_ok)
 {
    nir_foreach_instr(instr, block) {
       switch (instr->type) {
@@ -117,6 +118,25 @@ block_check_for_allowed_instrs(nir_block *block, unsigned *count,
          case nir_op_vec3:
          case nir_op_vec4:
             break;
+
+         case nir_op_fcos:
+         case nir_op_fdiv:
+         case nir_op_fexp2:
+         case nir_op_flog2:
+         case nir_op_fmod:
+         case nir_op_fpow:
+         case nir_op_frcp:
+         case nir_op_frem:
+         case nir_op_frsq:
+         case nir_op_fsin:
+         case nir_op_idiv:
+         case nir_op_irem:
+         case nir_op_udiv:
+            if (!alu_ok || !expensive_alu_ok)
+               return false;
+
+            break;
+
          default:
             if (!alu_ok) {
                /* It must be a move-like operation. */
@@ -160,7 +180,8 @@ block_check_for_allowed_instrs(nir_block *block, unsigned *count,
 
 static bool
 nir_opt_peephole_select_block(nir_block *block, nir_shader *shader,
-                              unsigned limit, bool indirect_load_ok)
+                              unsigned limit, bool indirect_load_ok,
+                              bool expensive_alu_ok)
 {
    if (nir_cf_node_is_first(&block->cf_node))
       return false;
@@ -181,9 +202,9 @@ nir_opt_peephole_select_block(nir_block *block, nir_shader *shader,
    /* ... and those blocks must only contain "allowed" instructions. */
    unsigned count = 0;
    if (!block_check_for_allowed_instrs(then_block, &count, limit != 0,
-                                       indirect_load_ok) ||
+                                       indirect_load_ok, expensive_alu_ok) ||
        !block_check_for_allowed_instrs(else_block, &count, limit != 0,
-                                       indirect_load_ok))
+                                       indirect_load_ok, expensive_alu_ok))
       return false;
 
    if (count > limit)
@@ -250,14 +271,15 @@ nir_opt_peephole_select_block(nir_block *block, nir_shader *shader,
 
 static bool
 nir_opt_peephole_select_impl(nir_function_impl *impl, unsigned limit,
-                             bool indirect_load_ok)
+                             bool indirect_load_ok, bool expensive_alu_ok)
 {
    nir_shader *shader = impl->function->shader;
    bool progress = false;
 
    nir_foreach_block_safe(block, impl) {
       progress |= nir_opt_peephole_select_block(block, shader, limit,
-                                                indirect_load_ok);
+                                                indirect_load_ok,
+                                                expensive_alu_ok);
    }
 
    if (progress)
@@ -268,14 +290,15 @@ nir_opt_peephole_select_impl(nir_function_impl *impl, unsigned limit,
 
 bool
 nir_opt_peephole_select(nir_shader *shader, unsigned limit,
-                        bool indirect_load_ok)
+                        bool indirect_load_ok, bool expensive_alu_ok)
 {
    bool progress = false;
 
    nir_foreach_function(function, shader) {
       if (function->impl)
          progress |= nir_opt_peephole_select_impl(function->impl, limit,
-                                                  indirect_load_ok);
+                                                  indirect_load_ok,
+                                                  expensive_alu_ok);
    }
 
    return progress;
diff --git a/src/freedreno/ir3/ir3_nir.c b/src/freedreno/ir3/ir3_nir.c
index 112c0928529..f5db5d151ef 100644
--- a/src/freedreno/ir3/ir3_nir.c
+++ b/src/freedreno/ir3/ir3_nir.c
@@ -97,7 +97,7 @@ ir3_optimize_loop(nir_shader *s)
 			progress |= OPT(s, nir_opt_gcm, true);
 		else if (gcm == 2)
 			progress |= OPT(s, nir_opt_gcm, false);
-		progress |= OPT(s, nir_opt_peephole_select, 16, true);
+		progress |= OPT(s, nir_opt_peephole_select, 16, true, true);
 		progress |= OPT(s, nir_opt_intrinsics);
 		progress |= OPT(s, nir_opt_algebraic);
 		progress |= OPT(s, nir_opt_constant_folding);
diff --git a/src/gallium/drivers/radeonsi/si_shader_nir.c b/src/gallium/drivers/radeonsi/si_shader_nir.c
index 89acaab248a..0a692277f64 100644
--- a/src/gallium/drivers/radeonsi/si_shader_nir.c
+++ b/src/gallium/drivers/radeonsi/si_shader_nir.c
@@ -841,7 +841,7 @@ si_lower_nir(struct si_shader_selector* sel)
 		NIR_PASS(progress, sel->nir, nir_opt_if);
 		NIR_PASS(progress, sel->nir, nir_opt_dead_cf);
 		NIR_PASS(progress, sel->nir, nir_opt_cse);
-		NIR_PASS(progress, sel->nir, nir_opt_peephole_select, 8, true);
+		NIR_PASS(progress, sel->nir, nir_opt_peephole_select, 8, true, true);
 
 		/* Needed for algebraic lowering */
 		NIR_PASS(progress, sel->nir, nir_opt_algebraic);
diff --git a/src/gallium/drivers/vc4/vc4_program.c b/src/gallium/drivers/vc4/vc4_program.c
index 48d83061f9a..fb75c07e266 100644
--- a/src/gallium/drivers/vc4/vc4_program.c
+++ b/src/gallium/drivers/vc4/vc4_program.c
@@ -1591,7 +1591,7 @@ vc4_optimize_nir(struct nir_shader *s)
                 NIR_PASS(progress, s, nir_opt_dce);
                 NIR_PASS(progress, s, nir_opt_dead_cf);
                 NIR_PASS(progress, s, nir_opt_cse);
-                NIR_PASS(progress, s, nir_opt_peephole_select, 8, true);
+                NIR_PASS(progress, s, nir_opt_peephole_select, 8, true, true);
                 NIR_PASS(progress, s, nir_opt_algebraic);
                 NIR_PASS(progress, s, nir_opt_constant_folding);
                 NIR_PASS(progress, s, nir_opt_undef);
diff --git a/src/intel/compiler/brw_nir.c b/src/intel/compiler/brw_nir.c
index 3e547daf78a..8a5665076b9 100644
--- a/src/intel/compiler/brw_nir.c
+++ b/src/intel/compiler/brw_nir.c
@@ -589,9 +589,9 @@ brw_nir_optimize(nir_shader *nir, const struct brw_compiler *compiler,
       const bool is_vec4_tessellation = !is_scalar &&
          (nir->info.stage == MESA_SHADER_TESS_CTRL ||
           nir->info.stage == MESA_SHADER_TESS_EVAL);
-      OPT(nir_opt_peephole_select, 0, !is_vec4_tessellation);
+      OPT(nir_opt_peephole_select, 0, !is_vec4_tessellation, false);
       if (compiler->devinfo->gen >= 6)
-         OPT(nir_opt_peephole_select, 1, !is_vec4_tessellation);
+         OPT(nir_opt_peephole_select, 1, !is_vec4_tessellation, true);
 
       OPT(nir_opt_intrinsics);
       OPT(nir_opt_idiv_const, 32);
diff --git a/src/mesa/state_tracker/st_glsl_to_nir.cpp b/src/mesa/state_tracker/st_glsl_to_nir.cpp
index e6d5c86bfbc..ed9f643e890 100644
--- a/src/mesa/state_tracker/st_glsl_to_nir.cpp
+++ b/src/mesa/state_tracker/st_glsl_to_nir.cpp
@@ -328,7 +328,7 @@ st_nir_opts(nir_shader *nir, bool scalar)
       NIR_PASS(progress, nir, nir_opt_if);
       NIR_PASS(progress, nir, nir_opt_dead_cf);
       NIR_PASS(progress, nir, nir_opt_cse);
-      NIR_PASS(progress, nir, nir_opt_peephole_select, 8, true);
+      NIR_PASS(progress, nir, nir_opt_peephole_select, 8, true, true);
 
       NIR_PASS(progress, nir, nir_opt_algebraic);
       NIR_PASS(progress, nir, nir_opt_constant_folding);