From a7b569d60c72726f0f630acab0a1ce160321de83 Mon Sep 17 00:00:00 2001 From: Rob Clark Date: Mon, 29 Jan 2018 16:09:44 -0500 Subject: [PATCH] freedreno/ir3: remove lower_if_else pass Now that it is unused. Signed-off-by: Rob Clark --- .../drivers/freedreno/Makefile.sources | 1 - src/gallium/drivers/freedreno/ir3/ir3_nir.h | 1 - .../freedreno/ir3/ir3_nir_lower_if_else.c | 325 ------------------ src/gallium/drivers/freedreno/meson.build | 1 - 4 files changed, 328 deletions(-) delete mode 100644 src/gallium/drivers/freedreno/ir3/ir3_nir_lower_if_else.c diff --git a/src/gallium/drivers/freedreno/Makefile.sources b/src/gallium/drivers/freedreno/Makefile.sources index 18e65fbbfac..e0f0586653c 100644 --- a/src/gallium/drivers/freedreno/Makefile.sources +++ b/src/gallium/drivers/freedreno/Makefile.sources @@ -173,7 +173,6 @@ ir3_SOURCES := \ ir3/ir3_legalize.c \ ir3/ir3_nir.c \ ir3/ir3_nir.h \ - ir3/ir3_nir_lower_if_else.c \ ir3/ir3_nir_lower_tg4_to_tex.c \ ir3/ir3_print.c \ ir3/ir3_ra.c \ diff --git a/src/gallium/drivers/freedreno/ir3/ir3_nir.h b/src/gallium/drivers/freedreno/ir3/ir3_nir.h index ca873407fae..c4892c34713 100644 --- a/src/gallium/drivers/freedreno/ir3/ir3_nir.h +++ b/src/gallium/drivers/freedreno/ir3/ir3_nir.h @@ -36,7 +36,6 @@ void ir3_nir_scan_driver_consts(nir_shader *shader, struct ir3_driver_const_layout *layout); -bool ir3_nir_lower_if_else(nir_shader *shader); bool ir3_nir_apply_trig_workarounds(nir_shader *shader); bool ir3_nir_lower_tg4_to_tex(nir_shader *shader); diff --git a/src/gallium/drivers/freedreno/ir3/ir3_nir_lower_if_else.c b/src/gallium/drivers/freedreno/ir3/ir3_nir_lower_if_else.c deleted file mode 100644 index b781fa7663f..00000000000 --- a/src/gallium/drivers/freedreno/ir3/ir3_nir_lower_if_else.c +++ /dev/null @@ -1,325 +0,0 @@ -/* - * Copyright © 2014 Intel Corporation - * Copyright © 2015 Red Hat - * - * 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: - * Jason Ekstrand (jason@jlekstrand.net) - * Rob Clark (robclark@freedesktop.org) - * - */ - -#include "ir3_nir.h" -#include "compiler/nir/nir_builder.h" -#include "compiler/nir/nir_control_flow.h" - -/* Based on nir_opt_peephole_select, and hacked up to more aggressively - * flatten anything that can be flattened - * - * This *might* be something that other drivers could use. On the other - * hand, I think most other hw has predicated instructions or similar - * to select which side of if/else writes back result (and therefore - * not having to assign unique registers to both sides of the if/else. - * (And hopefully those drivers don't also have crazy scheduling reqs - * and can more easily do this in their backend.) - * - * TODO eventually when we have proper flow control in the backend: - * - * + Probably weight differently normal ALUs vs SFUs (cos/rcp/exp) - * since executing extra SFUs for the branch-not-taken path will - * generally be much more expensive. - * - * Possibly what constitutes an ALU vs SFU differs between hw - * backends.. but that seems doubtful. - * - * + Account for texture fetch and memory accesses (incl UBOs) - * since these will be more expensive.. - * - * + When if-condition is const (or uniform) or we have some way - * to know that all threads in the warp take the same branch - * then we should prefer to not flatten the if/else.. - */ - -static bool -valid_dest(nir_block *block, nir_dest *dest) -{ - /* It must be SSA */ - if (!dest->is_ssa) - return false; - - /* We only lower blocks that do not contain other blocks - * (so this is run iteratively in a loop). Therefore if - * we get this far, it should not have any if_uses: - */ - assert(list_empty(&dest->ssa.if_uses)); - - /* The only uses of this definition must be phi's in the - * successor or in the current block - */ - nir_foreach_use(use, &dest->ssa) { - nir_instr *dest_instr = use->parent_instr; - if (dest_instr->block == block) - continue; - if ((dest_instr->type == nir_instr_type_phi) && - (dest_instr->block == block->successors[0])) - continue; - return false; - } - - return true; -} - -static bool -block_check_for_allowed_instrs(nir_block *block) -{ - nir_foreach_instr(instr, block) { - switch (instr->type) { - case nir_instr_type_intrinsic: { - nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr); - const nir_intrinsic_info *info = - &nir_intrinsic_infos[intr->intrinsic]; - - switch (intr->intrinsic) { - case nir_intrinsic_discard_if: - /* to simplify things, we want discard_if src in ssa: */ - if (!intr->src[0].is_ssa) - return false; - /* fallthrough */ - case nir_intrinsic_discard: - /* discard/discard_if can be reordered, but only - * with some special care - */ - break; - case nir_intrinsic_store_output: - /* TODO technically, if both if and else store - * the same output, we can hoist that out to - * the end of the block w/ a phi.. - * In practice, the tgsi shaders we already get - * do this for us, so I think we don't need to - */ - default: - if (!(info->flags & NIR_INTRINSIC_CAN_REORDER)) - return false; - } - - break; - } - - case nir_instr_type_tex: { - nir_tex_instr *tex = nir_instr_as_tex(instr); - if (!valid_dest(block, &tex->dest)) - return false; - break; - } - case nir_instr_type_phi: { - nir_phi_instr *phi = nir_instr_as_phi(instr); - if (!valid_dest(block, &phi->dest)) - return false; - break; - } - case nir_instr_type_alu: { - nir_alu_instr *alu = nir_instr_as_alu(instr); - if (!valid_dest(block, &alu->dest.dest)) - return false; - break; - } - - case nir_instr_type_load_const: - case nir_instr_type_ssa_undef: - break; /* always ssa dest */ - - default: - return false; - } - } - - return true; -} - -/* flatten an then or else block: */ -static void -flatten_block(nir_builder *bld, nir_block *if_block, nir_block *prev_block, - nir_ssa_def *condition, bool invert) -{ - nir_foreach_instr_safe(instr, if_block) { - if (instr->type == nir_instr_type_intrinsic) { - nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr); - if ((intr->intrinsic == nir_intrinsic_discard) || - (intr->intrinsic == nir_intrinsic_discard_if)) { - nir_ssa_def *discard_cond; - - bld->cursor = nir_after_instr( - nir_block_last_instr(prev_block)); - - if (invert) { - condition = nir_inot(bld, condition); - invert = false; - } - - if (intr->intrinsic == nir_intrinsic_discard) { - discard_cond = condition; - } else { - assert(intr->src[0].is_ssa); - /* discard_if gets re-written w/ src and'd: */ - discard_cond = nir_iand(bld, condition, intr->src[0].ssa); - } - - nir_intrinsic_instr *discard_if = - nir_intrinsic_instr_create(bld->shader, - nir_intrinsic_discard_if); - discard_if->src[0] = nir_src_for_ssa(discard_cond); - - nir_instr_insert_after(nir_block_last_instr(prev_block), - &discard_if->instr); - nir_instr_remove(instr); - instr = NULL; - } - } - /* if not an handled specially, just move to prev block: */ - if (instr) { - /* NOTE: exec_node_remove() is safe here (vs nir_instr_remove() - * since we are re-adding the instructin back in to the prev - * block (so no dangling SSA uses) - */ - exec_node_remove(&instr->node); - instr->block = prev_block; - exec_list_push_tail(&prev_block->instr_list, &instr->node); - } - } -} - -static bool -lower_if_else_block(nir_block *block, nir_builder *b, void *mem_ctx) -{ - /* If the block is empty, then it certainly doesn't have any phi nodes, - * so we can skip it. This also ensures that we do an early skip on the - * end block of the function which isn't actually attached to the CFG. - */ - if (exec_list_is_empty(&block->instr_list)) - return false; - - if (nir_cf_node_is_first(&block->cf_node)) - return false; - - nir_cf_node *prev_node = nir_cf_node_prev(&block->cf_node); - if (prev_node->type != nir_cf_node_if) - return false; - - nir_if *if_stmt = nir_cf_node_as_if(prev_node); - nir_block *then_block = nir_if_first_then_block(if_stmt); - nir_block *else_block = nir_if_first_else_block(if_stmt); - - /* We can only have one block in each side ... */ - if (nir_if_last_then_block(if_stmt) != then_block || - nir_if_last_else_block(if_stmt) != else_block) - return false; - - /* ... and those blocks must only contain "allowed" instructions. */ - if (!block_check_for_allowed_instrs(then_block) || - !block_check_for_allowed_instrs(else_block)) - return false; - - /* condition should be ssa too, which simplifies flatten_block: */ - if (!if_stmt->condition.is_ssa) - return false; - - /* At this point, we know that the previous CFG node is an if-then - * statement containing only moves to phi nodes in this block. We can - * just remove that entire CF node and replace all of the phi nodes with - * selects. - */ - - nir_block *prev_block = nir_cf_node_as_block(nir_cf_node_prev(prev_node)); - assert(prev_block->cf_node.type == nir_cf_node_block); - - /* First, we move the remaining instructions from the blocks to the - * block before. There are a few things that need handling specially - * like discard/discard_if. - */ - flatten_block(b, then_block, prev_block, - if_stmt->condition.ssa, false); - flatten_block(b, else_block, prev_block, - if_stmt->condition.ssa, true); - - nir_foreach_instr_safe(instr, block) { - if (instr->type != nir_instr_type_phi) - break; - - nir_phi_instr *phi = nir_instr_as_phi(instr); - nir_alu_instr *sel = nir_alu_instr_create(mem_ctx, nir_op_bcsel); - nir_src_copy(&sel->src[0].src, &if_stmt->condition, mem_ctx); - /* Splat the condition to all channels */ - memset(sel->src[0].swizzle, 0, sizeof sel->src[0].swizzle); - - assert(exec_list_length(&phi->srcs) == 2); - nir_foreach_phi_src(src, phi) { - assert(src->pred == then_block || src->pred == else_block); - assert(src->src.is_ssa); - - unsigned idx = src->pred == then_block ? 1 : 2; - nir_src_copy(&sel->src[idx].src, &src->src, mem_ctx); - } - - nir_ssa_dest_init(&sel->instr, &sel->dest.dest, - phi->dest.ssa.num_components, 32, phi->dest.ssa.name); - sel->dest.write_mask = (1 << phi->dest.ssa.num_components) - 1; - - nir_ssa_def_rewrite_uses(&phi->dest.ssa, - nir_src_for_ssa(&sel->dest.dest.ssa)); - - nir_instr_insert_before(&phi->instr, &sel->instr); - nir_instr_remove(&phi->instr); - } - - nir_cf_node_remove(&if_stmt->cf_node); - return true; -} - -static bool -lower_if_else_impl(nir_function_impl *impl) -{ - void *mem_ctx = ralloc_parent(impl); - nir_builder b; - nir_builder_init(&b, impl); - - bool progress = false; - nir_foreach_block_safe(block, impl) { - progress |= lower_if_else_block(block, &b, mem_ctx); - } - - if (progress) - nir_metadata_preserve(impl, nir_metadata_none); - - return progress; -} - -bool -ir3_nir_lower_if_else(nir_shader *shader) -{ - bool progress = false; - - nir_foreach_function(function, shader) { - if (function->impl) - progress |= lower_if_else_impl(function->impl); - } - - return progress; -} diff --git a/src/gallium/drivers/freedreno/meson.build b/src/gallium/drivers/freedreno/meson.build index f334d2d8ef8..8094e141b4b 100644 --- a/src/gallium/drivers/freedreno/meson.build +++ b/src/gallium/drivers/freedreno/meson.build @@ -195,7 +195,6 @@ files_libfreedreno = files( 'ir3/ir3_legalize.c', 'ir3/ir3_nir.c', 'ir3/ir3_nir.h', - 'ir3/ir3_nir_lower_if_else.c', 'ir3/ir3_nir_lower_tg4_to_tex.c', 'ir3/ir3_print.c', 'ir3/ir3_ra.c', -- 2.30.2