mesa/src/compiler/nir/nir_opt_remove_phis.c
Georg Lehmann 0e6d32777f nir/opt_remove_phis: rematerialize equal alu
Foz-DB Navi31:
Totals from 943 (1.19% of 79395) affected shaders:
MaxWaves: 24672 -> 24722 (+0.20%)
Instrs: 1541665 -> 1544956 (+0.21%); split: -0.23%, +0.44%
CodeSize: 8085180 -> 8109212 (+0.30%); split: -0.16%, +0.46%
VGPRs: 57768 -> 57624 (-0.25%)
Latency: 18043743 -> 17948245 (-0.53%); split: -1.28%, +0.75%
InvThroughput: 2692605 -> 2677049 (-0.58%); split: -2.07%, +1.49%
VClause: 25321 -> 25343 (+0.09%); split: -0.48%, +0.57%
SClause: 38473 -> 38614 (+0.37%); split: -0.00%, +0.37%
Copies: 86089 -> 86236 (+0.17%); split: -0.46%, +0.63%
Branches: 36719 -> 36777 (+0.16%); split: -0.60%, +0.76%
PreSGPRs: 44138 -> 44303 (+0.37%); split: -0.05%, +0.42%
PreVGPRs: 43319 -> 43009 (-0.72%)
VALU: 893684 -> 894272 (+0.07%); split: -0.42%, +0.48%
SALU: 189561 -> 191358 (+0.95%); split: -0.05%, +1.00%
VMEM: 42294 -> 42313 (+0.04%); split: -0.44%, +0.49%
SMEM: 72916 -> 73144 (+0.31%)

Instruction count regressions are largly caused by additional
loop unrolling.

Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/31028>
2024-12-16 20:38:38 +00:00

211 lines
6.4 KiB
C

/*
* Copyright © 2015 Connor Abbott
*
* Permission is hereby granted, free of charge, to any person obtaining a
* copy of this software and associated documentation files (the "Software"),
* to deal in the Software without restriction, including without limitation
* the rights to use, copy, modify, merge, publish, distribute, sublicense,
* and/or sell copies of the Software, and to permit persons to whom the
* Software is furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice (including the next
* paragraph) shall be included in all copies or substantial portions of the
* Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
* THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
* FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
* IN THE SOFTWARE.
*
* Authors:
* Connor Abbott (cwabbott0@gmail.com)
*
*/
#include "nir.h"
#include "nir_builder.h"
static bool
phi_srcs_equal(nir_def *a, nir_def *b)
{
if (a == b)
return true;
if (a->parent_instr->type != b->parent_instr->type)
return false;
if (a->parent_instr->type != nir_instr_type_alu &&
a->parent_instr->type != nir_instr_type_load_const)
return false;
if (!nir_instrs_equal(a->parent_instr, b->parent_instr))
return false;
/* nir_instrs_equal ignores exact/fast_math */
if (a->parent_instr->type == nir_instr_type_alu) {
nir_alu_instr *a_alu = nir_instr_as_alu(a->parent_instr);
nir_alu_instr *b_alu = nir_instr_as_alu(b->parent_instr);
if (a_alu->exact != b_alu->exact || a_alu->fp_fast_math != b_alu->fp_fast_math)
return false;
}
return true;
}
static bool
src_dominates_block(nir_src *src, void *state)
{
nir_block *block = state;
return nir_block_dominates(src->ssa->parent_instr->block, block);
}
static bool
can_rematerialize_phi_src(nir_block *imm_dom, nir_def *def)
{
if (def->parent_instr->type == nir_instr_type_alu) {
return nir_foreach_src(def->parent_instr, src_dominates_block, imm_dom);
} else if (def->parent_instr->type == nir_instr_type_load_const) {
return true;
}
return false;
}
/*
* This is a pass for removing phi nodes that look like:
* a = phi(b, b, b, ...)
*
* Note that we can't always ignore undef sources here, or else we may create a
* situation where the definition of b isn't dominated by its uses. We're
* allowed to do this since the definition of b must dominate all of the
* phi node's predecessors, which means it must dominate the phi node as well
* as all of the phi node's uses. In essence, the phi node acts as a copy
* instruction. b can't be another phi node in the same block, since the only
* time when phi nodes can source other phi nodes defined in the same block is
* at the loop header, and in that case one of the sources of the phi has to
* be from before the loop and that source can't be b.
*/
static bool
remove_phis_block(nir_block *block, nir_builder *b)
{
bool progress = false;
nir_foreach_phi_safe(phi, block) {
nir_def *def = NULL;
bool srcs_same = true;
bool needs_remat = false;
nir_foreach_phi_src(src, phi) {
/* For phi nodes at the beginning of loops, we may encounter some
* sources from backedges that point back to the destination of the
* same phi, i.e. something like:
*
* a = phi(a, b, ...)
*
* We can safely ignore these sources, since if all of the normal
* sources point to the same definition, then that definition must
* still dominate the phi node, and the phi will still always take
* the value of that definition.
*/
if (src->src.ssa == &phi->def)
continue;
/* Ignore undef sources. */
if (nir_src_is_undef(src->src))
continue;
if (def == NULL) {
def = src->src.ssa;
if (!nir_block_dominates(def->parent_instr->block, block->imm_dom)) {
if (!can_rematerialize_phi_src(block->imm_dom, def)) {
srcs_same = false;
break;
}
needs_remat = true;
}
} else if (!phi_srcs_equal(src->src.ssa, def)) {
srcs_same = false;
break;
}
}
if (!srcs_same)
continue;
if (!def) {
/* In this case, the phi had no non undef sources. So turn it into an undef. */
b->cursor = nir_after_phis(block);
def = nir_undef(b, phi->def.num_components, phi->def.bit_size);
} else if (needs_remat) {
b->cursor = nir_after_block_before_jump(block->imm_dom);
nir_instr *remat = nir_instr_clone(b->shader, def->parent_instr);
nir_builder_instr_insert(b, remat);
def = nir_instr_def(remat);
}
nir_def_replace(&phi->def, def);
progress = true;
}
return progress;
}
bool
nir_remove_single_src_phis_block(nir_block *block)
{
assert(block->predecessors->entries <= 1);
bool progress = false;
nir_foreach_phi_safe(phi, block) {
nir_def *def = NULL;
nir_foreach_phi_src(src, phi) {
def = src->src.ssa;
break;
}
if (!def) {
nir_builder b = nir_builder_create(nir_cf_node_get_function(&block->cf_node));
b.cursor = nir_after_phis(block);
def = nir_undef(&b, phi->def.num_components, phi->def.bit_size);
}
nir_def_replace(&phi->def, def);
progress = true;
}
return progress;
}
static bool
nir_opt_remove_phis_impl(nir_function_impl *impl)
{
bool progress = false;
nir_builder bld = nir_builder_create(impl);
nir_metadata_require(impl, nir_metadata_dominance);
nir_foreach_block(block, impl) {
progress |= remove_phis_block(block, &bld);
}
if (progress) {
nir_metadata_preserve(impl, nir_metadata_control_flow);
} else {
nir_metadata_preserve(impl, nir_metadata_all);
}
return progress;
}
bool
nir_opt_remove_phis(nir_shader *shader)
{
bool progress = false;
nir_foreach_function_impl(impl, shader)
progress = nir_opt_remove_phis_impl(impl) || progress;
return progress;
}