X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fglsl%2Fnir%2Fnir_opt_peephole_select.c;h=90902b97ffcdf73f157e59212b2214b087c5e8ef;hb=aecbc93f2d1ff9de4e03a2b216e86dcb9a4ce414;hp=9f541fef19d85bc2749e164a71b0dfc31f16f1a7;hpb=534a4ec82f09fa37e2134c994082e71c79b7d5ec;p=mesa.git diff --git a/src/glsl/nir/nir_opt_peephole_select.c b/src/glsl/nir/nir_opt_peephole_select.c index 9f541fef19d..90902b97ffc 100644 --- a/src/glsl/nir/nir_opt_peephole_select.c +++ b/src/glsl/nir/nir_opt_peephole_select.c @@ -26,6 +26,7 @@ */ #include "nir.h" +#include "nir_control_flow.h" /* * Implements a small peephole optimization that looks for @@ -52,36 +53,76 @@ struct peephole_select_state { }; static bool -are_all_move_to_phi(nir_block *block) +block_check_for_allowed_instrs(nir_block *block) { nir_foreach_instr(block, instr) { - if (instr->type != nir_instr_type_alu) - return false; + switch (instr->type) { + case nir_instr_type_intrinsic: { + nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr); + + switch (intrin->intrinsic) { + case nir_intrinsic_load_var: + switch (intrin->variables[0]->var->data.mode) { + case nir_var_shader_in: + case nir_var_uniform: + break; + + default: + return false; + } + break; + + default: + return false; + } - /* It must be a move operation */ - nir_alu_instr *mov = nir_instr_as_alu(instr); - if (mov->op != nir_op_fmov && mov->op != nir_op_imov) - return false; + break; + } - /* Can't handle saturate */ - if (mov->dest.saturate) - return false; + case nir_instr_type_load_const: + break; - /* It must be SSA */ - if (!mov->dest.dest.is_ssa) - return false; + case nir_instr_type_alu: { + nir_alu_instr *mov = nir_instr_as_alu(instr); + switch (mov->op) { + case nir_op_fmov: + case nir_op_imov: + case nir_op_fneg: + case nir_op_ineg: + case nir_op_fabs: + case nir_op_iabs: + case nir_op_vec2: + case nir_op_vec3: + case nir_op_vec4: + /* It must be a move-like operation. */ + break; + default: + return false; + } - /* It cannot have any if-uses */ - if (mov->dest.dest.ssa.if_uses->entries != 0) - return false; + /* Can't handle saturate */ + if (mov->dest.saturate) + return false; + + /* It must be SSA */ + if (!mov->dest.dest.is_ssa) + return false; - /* The only uses of this definition must be phi's in the successor */ - struct set_entry *entry; - set_foreach(mov->dest.dest.ssa.uses, entry) { - const nir_instr *dest_instr = entry->key; - if (dest_instr->type != nir_instr_type_phi || - dest_instr->block != block->successors[0]) + /* It cannot have any if-uses */ + if (!list_empty(&mov->dest.dest.ssa.if_uses)) return false; + + /* The only uses of this definition must be phi's in the successor */ + nir_foreach_use(&mov->dest.dest.ssa, use) { + if (use->parent_instr->type != nir_instr_type_phi || + use->parent_instr->block != block->successors[0]) + return false; + } + break; + } + + default: + return false; } } @@ -119,8 +160,9 @@ nir_opt_peephole_select_block(nir_block *block, void *void_state) nir_block *then_block = nir_cf_node_as_block(then_node); nir_block *else_block = nir_cf_node_as_block(else_node); - /* ... and those blocks must only contain move-to-phi. */ - if (!are_all_move_to_phi(then_block) || !are_all_move_to_phi(else_block)) + /* ... and those blocks must only contain "allowed" instructions. */ + if (!block_check_for_allowed_instrs(then_block) || + !block_check_for_allowed_instrs(else_block)) return true; /* At this point, we know that the previous CFG node is an if-then @@ -129,13 +171,32 @@ nir_opt_peephole_select_block(nir_block *block, void *void_state) * 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. We have already guaranteed that this is safe by + * calling block_check_for_allowed_instrs() + */ + nir_foreach_instr_safe(then_block, instr) { + exec_node_remove(&instr->node); + instr->block = prev_block; + exec_list_push_tail(&prev_block->instr_list, &instr->node); + } + + nir_foreach_instr_safe(else_block, instr) { + exec_node_remove(&instr->node); + instr->block = prev_block; + exec_list_push_tail(&prev_block->instr_list, &instr->node); + } + nir_foreach_instr_safe(block, instr) { 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(state->mem_ctx, nir_op_bcsel); - sel->src[0].src = nir_src_copy(if_stmt->condition, state->mem_ctx); + nir_src_copy(&sel->src[0].src, &if_stmt->condition, sel); /* Splat the condition to all channels */ memset(sel->src[0].swizzle, 0, sizeof sel->src[0].swizzle); @@ -145,22 +206,7 @@ nir_opt_peephole_select_block(nir_block *block, void *void_state) assert(src->src.is_ssa); unsigned idx = src->pred == then_block ? 1 : 2; - - if (src->src.ssa->parent_instr->block == src->pred) { - /* We already know that this instruction must be a move with - * this phi's in this block as its only users. - */ - nir_alu_instr *mov = nir_instr_as_alu(src->src.ssa->parent_instr); - assert(mov->instr.type == nir_instr_type_alu); - assert(mov->op == nir_op_fmov || mov->op == nir_op_imov); - - sel->src[idx].src = nir_src_copy(mov->src[0].src, state->mem_ctx); - sel->src[idx].negate = mov->src[0].negate; - sel->src[idx].abs = mov->src[0].abs; - memcpy(sel->src[idx].swizzle, mov->src[0].swizzle, 4); - } else { - sel->src[idx].src = nir_src_copy(src->src, state->mem_ctx); - } + nir_src_copy(&sel->src[idx].src, &src->src, sel); } nir_ssa_dest_init(&sel->instr, &sel->dest.dest, @@ -168,8 +214,7 @@ nir_opt_peephole_select_block(nir_block *block, void *void_state) 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), - state->mem_ctx); + nir_src_for_ssa(&sel->dest.dest.ssa)); nir_instr_insert_before(&phi->instr, &sel->instr); nir_instr_remove(&phi->instr);