From: Connor Abbott Date: Tue, 22 Jul 2014 21:05:06 +0000 (-0700) Subject: nir: add a pass to convert to SSA X-Git-Url: https://git.libre-soc.org/?a=commitdiff_plain;h=4553887d4a4501a75c91033e9e8e5fe4cd6bc3f2;p=mesa.git nir: add a pass to convert to SSA v2: Jason Ekstrand : whitespace fixes --- diff --git a/src/glsl/Makefile.sources b/src/glsl/Makefile.sources index c736d6a56ec..0aaea580816 100644 --- a/src/glsl/Makefile.sources +++ b/src/glsl/Makefile.sources @@ -28,6 +28,7 @@ NIR_FILES = \ $(GLSL_SRCDIR)/nir/nir_opt_global_to_local.c \ $(GLSL_SRCDIR)/nir/nir_print.c \ $(GLSL_SRCDIR)/nir/nir_remove_dead_variables.c \ + $(GLSL_SRCDIR)/nir/nir_to_ssa.c \ $(GLSL_SRCDIR)/nir/nir_validate.c \ $(GLSL_SRCDIR)/nir/nir_types.cpp \ $(GLSL_SRCDIR)/nir/glsl_to_nir.cpp diff --git a/src/glsl/nir/nir.h b/src/glsl/nir/nir.h index 97687cda8f0..feaf21dd656 100644 --- a/src/glsl/nir/nir.h +++ b/src/glsl/nir/nir.h @@ -1288,6 +1288,9 @@ void nir_lower_system_values(nir_shader *shader); void nir_lower_atomics(nir_shader *shader); +void nir_convert_to_ssa_impl(nir_function_impl *impl); +void nir_convert_to_ssa(nir_shader *shader); + bool nir_opt_global_to_local(nir_shader *shader); #ifdef __cplusplus diff --git a/src/glsl/nir/nir_to_ssa.c b/src/glsl/nir/nir_to_ssa.c new file mode 100644 index 00000000000..9f70225f211 --- /dev/null +++ b/src/glsl/nir/nir_to_ssa.c @@ -0,0 +1,682 @@ +/* + * Copyright © 2014 Intel Corporation + * + * 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 "malloc.h" +#include + +/* + * Implements the classic to-SSA algorithm described by Cytron et. al. in + * "Efficiently Computing Static Single Assignment Form and the Control + * Dependence Graph." + */ + +/* inserts a phi node of the form reg = phi(reg, reg, reg, ...) */ + +static void +insert_trivial_phi(nir_register *reg, nir_block *block, void *mem_ctx) +{ + nir_phi_instr *instr = nir_phi_instr_create(mem_ctx); + + instr->dest.reg.reg = reg; + struct set_entry *entry; + set_foreach(block->predecessors, entry) { + nir_block *pred = (nir_block *) entry->key; + + nir_phi_src *src = ralloc(mem_ctx, nir_phi_src); + src->pred = pred; + src->src.is_ssa = false; + src->src.reg.base_offset = 0; + src->src.reg.indirect = NULL; + src->src.reg.reg = reg; + exec_list_push_tail(&instr->srcs, &src->node); + } + + nir_instr_insert_before_block(block, &instr->instr); +} + +static void +insert_phi_nodes(nir_function_impl *impl) +{ + void *mem_ctx = ralloc_parent(impl); + + unsigned *work = calloc(impl->num_blocks, sizeof(unsigned)); + unsigned *has_already = calloc(impl->num_blocks, sizeof(unsigned)); + + /* + * Since the work flags already prevent us from inserting a node that has + * ever been inserted into W, we don't need to use a set to represent W. + * Also, since no block can ever be inserted into W more than once, we know + * that the maximum size of W is the number of basic blocks in the + * function. So all we need to handle W is an array and a pointer to the + * next element to be inserted and the next element to be removed. + */ + nir_block **W = malloc(impl->num_blocks * sizeof(nir_block *)); + unsigned w_start, w_end; + + unsigned iter_count = 0; + + nir_index_blocks(impl); + + foreach_list_typed(nir_register, reg, node, &impl->registers) { + if (reg->num_array_elems != 0) + continue; + + w_start = w_end = 0; + iter_count++; + + struct set_entry *entry; + set_foreach(reg->defs, entry) { + nir_instr *def = (nir_instr *) entry->key; + if (work[def->block->index] < iter_count) + W[w_end++] = def->block; + work[def->block->index] = iter_count; + } + + while (w_start != w_end) { + nir_block *cur = W[w_start++]; + set_foreach(cur->dom_frontier, entry) { + nir_block *next = (nir_block *) entry->key; + + /* + * If there's more than one return statement, then the end block + * can be a join point for some definitions. However, there are + * no instructions in the end block, so nothing would use those + * phi nodes. Of course, we couldn't place those phi nodes + * anyways due to the restriction of having no instructions in the + * end block... + */ + if (next == impl->end_block) + continue; + + if (has_already[next->index] < iter_count) { + insert_trivial_phi(reg, next, mem_ctx); + has_already[next->index] = iter_count; + if (work[next->index] < iter_count) { + work[next->index] = iter_count; + W[w_end++] = next; + } + } + } + } + } + + free(work); + free(has_already); + free(W); +} + +typedef struct { + nir_ssa_def **stack; + int index; + unsigned num_defs; /** < used to add indices to debug names */ +#ifdef DEBUG + unsigned stack_size; +#endif +} reg_state; + +typedef struct { + reg_state *states; + void *mem_ctx; + nir_instr *parent_instr; + nir_if *parent_if; + nir_function_impl *impl; + + /* map from SSA value -> original register */ + struct hash_table *ssa_map; + + /* predicate for this instruction */ + nir_src *predicate; +} rewrite_state; + +static nir_ssa_def *get_ssa_src(nir_register *reg, rewrite_state *state) +{ + unsigned index = reg->index; + + if (state->states[index].index == -1) { + /* + * We're using an undefined register, create a new undefined SSA value + * to preserve the information that this source is undefined + */ + nir_ssa_undef_instr *instr = nir_ssa_undef_instr_create(state->mem_ctx); + instr->def.index = state->impl->ssa_alloc++; + instr->def.num_components = reg->num_components; + instr->def.uses = _mesa_set_create(state->mem_ctx, + _mesa_key_pointer_equal); + instr->def.if_uses = _mesa_set_create(state->mem_ctx, + _mesa_key_pointer_equal); + /* + * We could just insert the undefined instruction before the instruction + * we're rewriting, but we could be rewriting a phi source in which case + * we can't do that, so do the next easiest thing - insert it at the + * beginning of the program. In the end, it doesn't really matter where + * the undefined instructions are because they're going to be ignored + * in the backend. + */ + nir_instr_insert_before_cf_list(&state->impl->body, &instr->instr); + return &instr->def; + } + + return state->states[index].stack[state->states[index].index]; +} + +static bool +rewrite_use(nir_src *src, void *_state) +{ + rewrite_state *state = (rewrite_state *) _state; + + if (src->is_ssa) + return true; + + unsigned index = src->reg.reg->index; + + if (state->states[index].stack == NULL) + return true; + + src->is_ssa = true; + src->ssa = get_ssa_src(src->reg.reg, state); + + if (state->parent_instr) + _mesa_set_add(src->ssa->uses, _mesa_hash_pointer(state->parent_instr), + state->parent_instr); + else + _mesa_set_add(src->ssa->if_uses, _mesa_hash_pointer(state->parent_if), + state->parent_if); + return true; +} + +static bool +rewrite_def_forwards(nir_dest *dest, void *_state) +{ + rewrite_state *state = (rewrite_state *) _state; + + if (dest->is_ssa) + return true; + + nir_register *reg = dest->reg.reg; + unsigned index = reg->index; + + if (state->states[index].stack == NULL) + return true; + + nir_alu_instr *csel = NULL; + if (state->predicate) { + /* + * To capture the information that we may or may not overwrite this + * register due to the predicate, we need to emit a conditional select + * that takes the old version of the register and the new version. + * This is basically a watered-down version of the Psi-SSA + * representation, without any of the optimizations. + * + * TODO: do we actually need full-blown Psi-SSA? + */ + csel = nir_alu_instr_create(state->mem_ctx, nir_op_bcsel); + csel->dest.dest.reg.reg = dest->reg.reg; + csel->dest.write_mask = (1 << dest->reg.reg->num_components) - 1; + csel->src[0].src = nir_src_copy(*state->predicate, state->mem_ctx); + if (csel->src[0].src.is_ssa) + _mesa_set_add(csel->src[0].src.ssa->uses, + _mesa_hash_pointer(&csel->instr), &csel->instr); + csel->src[2].src.is_ssa = true; + csel->src[2].src.ssa = get_ssa_src(dest->reg.reg, state); + _mesa_set_add(csel->src[2].src.ssa->uses, + _mesa_hash_pointer(&csel->instr), &csel->instr); + } + + dest->is_ssa = true; + + char *name = NULL; + if (dest->reg.reg->name) + name = ralloc_asprintf(state->mem_ctx, "%s_%u", dest->reg.reg->name, + state->states[index].num_defs); + + dest->ssa.index = state->impl->ssa_alloc++; + dest->ssa.num_components = reg->num_components; + dest->ssa.parent_instr = state->parent_instr; + dest->ssa.uses = _mesa_set_create(state->mem_ctx, _mesa_key_pointer_equal); + dest->ssa.if_uses = _mesa_set_create(state->mem_ctx, + _mesa_key_pointer_equal); + dest->ssa.name = name; + + /* push our SSA destination on the stack */ + state->states[index].index++; + assert(state->states[index].index < state->states[index].stack_size); + state->states[index].stack[state->states[index].index] = &dest->ssa; + state->states[index].num_defs++; + + _mesa_hash_table_insert(state->ssa_map, &dest->ssa, reg); + + if (state->predicate) { + csel->src[1].src.is_ssa = true; + csel->src[1].src.ssa = &dest->ssa; + _mesa_set_add(dest->ssa.uses, _mesa_hash_pointer(&csel->instr), + &csel->instr); + + nir_instr *old_parent_instr = state->parent_instr; + nir_src *old_predicate = state->predicate; + state->parent_instr = &csel->instr; + state->predicate = NULL; + rewrite_def_forwards(&csel->dest.dest, state); + state->parent_instr = old_parent_instr; + state->predicate = old_predicate; + + nir_instr_insert_after(state->parent_instr, &csel->instr); + } + + return true; +} + +static void +rewrite_alu_instr_forward(nir_alu_instr *instr, rewrite_state *state) +{ + state->parent_instr = &instr->instr; + state->predicate = instr->has_predicate ? &instr->predicate : NULL; + + nir_foreach_src(&instr->instr, rewrite_use, state); + + nir_register *reg = instr->dest.dest.reg.reg; + unsigned index = reg->index; + + if (state->states[index].stack == NULL) + return; + + unsigned write_mask = instr->dest.write_mask; + if (write_mask != (1 << instr->dest.dest.reg.reg->num_components) - 1) { + /* + * Calculate the number of components the final instruction, which for + * per-component things is the number of output components of the + * instruction and non-per-component things is the number of enabled + * channels in the write mask. + */ + unsigned num_components; + if (nir_op_infos[instr->op].output_size == 0) { + unsigned temp = (write_mask & 0x5) + ((write_mask >> 1) & 0x5); + num_components = (temp & 0x3) + ((temp >> 2) & 0x3); + } else { + num_components = nir_op_infos[instr->op].output_size; + } + + char *name = NULL; + if (instr->dest.dest.reg.reg->name) + name = ralloc_asprintf(state->mem_ctx, "%s_%u", + reg->name, state->states[index].num_defs); + + instr->dest.write_mask = (1 << num_components) - 1; + instr->dest.dest.is_ssa = true; + instr->dest.dest.ssa.index = state->impl->ssa_alloc++; + instr->dest.dest.ssa.num_components = num_components; + instr->dest.dest.ssa.name = name; + instr->dest.dest.ssa.parent_instr = &instr->instr; + instr->dest.dest.ssa.uses = _mesa_set_create(state->mem_ctx, + _mesa_key_pointer_equal); + instr->dest.dest.ssa.if_uses = _mesa_set_create(state->mem_ctx, + _mesa_key_pointer_equal); + + if (nir_op_infos[instr->op].output_size == 0) { + /* + * When we change the output writemask, we need to change the + * swizzles for per-component inputs too + */ + for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs; i++) { + if (nir_op_infos[instr->op].input_sizes[i] != 0) + continue; + + unsigned new_swizzle[4] = {0, 0, 0, 0}; + + /* + * We keep two indices: + * 1. The index of the original (non-SSA) component + * 2. The index of the post-SSA, compacted, component + * + * We need to map the swizzle component at index 1 to the swizzle + * component at index 2. + */ + + unsigned ssa_index = 0; + for (unsigned index = 0; index < 4; index++) { + if (!((write_mask >> index) & 1)) + continue; + + new_swizzle[ssa_index] = instr->src[i].swizzle[index]; + ssa_index++; + } + + for (unsigned j = 0; j < 4; j++) + instr->src[i].swizzle[j] = new_swizzle[j]; + } + } + + nir_op op; + switch (reg->num_components) { + case 2: op = nir_op_vec2; break; + case 3: op = nir_op_vec3; break; + case 4: op = nir_op_vec4; break; + default: assert(0); break; + } + + nir_alu_instr *vec = nir_alu_instr_create(state->mem_ctx, op); + + vec->dest.dest.reg.reg = reg; + vec->dest.write_mask = (1 << reg->num_components) - 1; + + nir_ssa_def *old_src = get_ssa_src(reg, state); + nir_ssa_def *new_src = &instr->dest.dest.ssa; + + unsigned ssa_index = 0; + for (unsigned i = 0; i < reg->num_components; i++) { + vec->src[i].src.is_ssa = true; + if ((write_mask >> i) & 1) { + vec->src[i].src.ssa = new_src; + if (nir_op_infos[instr->op].output_size == 0) + vec->src[i].swizzle[0] = ssa_index; + else + vec->src[i].swizzle[0] = i; + ssa_index++; + } else { + vec->src[i].src.ssa = old_src; + vec->src[i].swizzle[0] = i; + } + + _mesa_set_add(vec->src[i].src.ssa->uses, + _mesa_hash_pointer(&vec->instr), &vec->instr); + } + + vec->has_predicate = instr->has_predicate; + if (instr->has_predicate) { + vec->predicate = nir_src_copy(instr->predicate, state->mem_ctx); + if (vec->predicate.is_ssa) + _mesa_set_add(vec->predicate.ssa->uses, + _mesa_hash_pointer(&vec->instr), &vec->instr); + } + + nir_instr_insert_after(&instr->instr, &vec->instr); + + state->parent_instr = &vec->instr; + state->predicate = vec->has_predicate ? &vec->predicate : NULL; + rewrite_def_forwards(&vec->dest.dest, state); + } else { + rewrite_def_forwards(&instr->dest.dest, state); + } +} + +static void +rewrite_phi_instr(nir_phi_instr *instr, rewrite_state *state) +{ + state->parent_instr = &instr->instr; + state->predicate = NULL; + rewrite_def_forwards(&instr->dest, state); +} + +static nir_src * +get_instr_predicate(nir_instr *instr) +{ + nir_alu_instr *alu_instr; + nir_load_const_instr *load_const_instr; + nir_intrinsic_instr *intrinsic_instr; + nir_tex_instr *tex_instr; + + switch (instr->type) { + case nir_instr_type_alu: + alu_instr = nir_instr_as_alu(instr); + if (alu_instr->has_predicate) + return &alu_instr->predicate; + else + return NULL; + + case nir_instr_type_load_const: + load_const_instr = nir_instr_as_load_const(instr); + if (load_const_instr->has_predicate) + return &load_const_instr->predicate; + else + return NULL; + + case nir_instr_type_intrinsic: + intrinsic_instr = nir_instr_as_intrinsic(instr); + if (intrinsic_instr->has_predicate) + return &intrinsic_instr->predicate; + else + return NULL; + + case nir_instr_type_texture: + tex_instr = nir_instr_as_texture(instr); + if (tex_instr->has_predicate) + return &tex_instr->predicate; + else + return NULL; + + default: + break; + } + + return NULL; +} + +static void +rewrite_instr_forward(nir_instr *instr, rewrite_state *state) +{ + if (instr->type == nir_instr_type_alu) { + rewrite_alu_instr_forward(nir_instr_as_alu(instr), state); + return; + } + + if (instr->type == nir_instr_type_phi) { + rewrite_phi_instr(nir_instr_as_phi(instr), state); + return; + } + + state->parent_instr = instr; + state->predicate = get_instr_predicate(instr); + + nir_foreach_src(instr, rewrite_use, state); + nir_foreach_dest(instr, rewrite_def_forwards, state); +} + +static void +rewrite_phi_sources(nir_block *block, nir_block *pred, rewrite_state *state) +{ + nir_foreach_instr(block, instr) { + if (instr->type != nir_instr_type_phi) + break; + + nir_phi_instr *phi_instr = nir_instr_as_phi(instr); + + state->parent_instr = instr; + + foreach_list_typed(nir_phi_src, src, node, &phi_instr->srcs) { + if (src->pred == pred) { + rewrite_use(&src->src, state); + break; + } + } + } +} + +static bool +rewrite_def_backwards(nir_dest *dest, void *_state) +{ + rewrite_state *state = (rewrite_state *) _state; + + if (!dest->is_ssa) + return true; + + struct hash_entry *entry = + _mesa_hash_table_search(state->ssa_map, &dest->ssa); + + if (!entry) + return true; + + nir_register *reg = (nir_register *) entry->data; + unsigned index = reg->index; + + state->states[index].index--; + assert(state->states[index].index >= -1); + + return true; +} + +static void +rewrite_instr_backwards(nir_instr *instr, rewrite_state *state) +{ + nir_foreach_dest(instr, rewrite_def_backwards, state); +} + +static void +rewrite_block(nir_block *block, rewrite_state *state) +{ + /* This will skip over any instructions after the current one, which is + * what we want because those instructions (vector gather, conditional + * select) will already be in SSA form. + */ + nir_foreach_instr_safe(block, instr) { + rewrite_instr_forward(instr, state); + } + + if (block != state->impl->end_block && + !nir_cf_node_is_last(&block->cf_node) && + nir_cf_node_next(&block->cf_node)->type == nir_cf_node_if) { + nir_if *if_stmt = nir_cf_node_as_if(nir_cf_node_next(&block->cf_node)); + state->parent_instr = NULL; + state->parent_if = if_stmt; + rewrite_use(&if_stmt->condition, state); + } + + if (block->successors[0]) + rewrite_phi_sources(block->successors[0], block, state); + if (block->successors[1]) + rewrite_phi_sources(block->successors[1], block, state); + + for (unsigned i = 0; i < block->num_dom_children; i++) + rewrite_block(block->dom_children[i], state); + + nir_foreach_instr_reverse(block, instr) { + rewrite_instr_backwards(instr, state); + } +} + +static void +remove_unused_regs(nir_function_impl *impl, rewrite_state *state) +{ + foreach_list_typed_safe(nir_register, reg, node, &impl->registers) { + if (state->states[reg->index].stack != NULL) + exec_node_remove(®->node); + } +} + +static void +init_rewrite_state(nir_function_impl *impl, rewrite_state *state) +{ + state->impl = impl; + state->mem_ctx = ralloc_parent(impl); + state->ssa_map = _mesa_hash_table_create(NULL, _mesa_hash_pointer, + _mesa_key_pointer_equal); + state->states = ralloc_array(NULL, reg_state, impl->reg_alloc); + + foreach_list_typed(nir_register, reg, node, &impl->registers) { + assert(reg->index < impl->reg_alloc); + if (reg->num_array_elems > 0) { + state->states[reg->index].stack = NULL; + } else { + /* + * Calculate a conservative estimate of the stack size based on the + * number of definitions there are. Note that this function *must* be + * called after phi nodes are inserted so we can count phi node + * definitions too. + */ + unsigned stack_size = 0; + struct set_entry *entry; + set_foreach(reg->defs, entry) { + nir_instr *def = (nir_instr *) entry->key; + + stack_size++; + + /* + * predicates generate an additional predicate destination that + * gets pushed on the stack + * + * Note: ALU instructions generate an additional instruction too, + * but as of now only the additional instruction is pushed onto + * the stack, and not the original instruction because it doesn't + * need to be (actually, we could do the same with predicates, + * but it was easier to just use the existing codepath). + */ + + if (def->type == nir_instr_type_intrinsic) { + nir_intrinsic_instr *intrinsic_instr = + nir_instr_as_intrinsic(def); + if (nir_intrinsic_infos[intrinsic_instr->intrinsic].has_dest && + intrinsic_instr->has_predicate) + stack_size++; + } else { + if (get_instr_predicate(def) != NULL) + stack_size++; + } + } + + state->states[reg->index].stack = ralloc_array(state->states, + nir_ssa_def *, + stack_size); +#ifdef DEBUG + state->states[reg->index].stack_size = stack_size; +#endif + state->states[reg->index].index = -1; + state->states[reg->index].num_defs = 0; + } + } +} + +static void +destroy_rewrite_state(rewrite_state *state) +{ + _mesa_hash_table_destroy(state->ssa_map, NULL); + ralloc_free(state->states); +} + +void +nir_convert_to_ssa_impl(nir_function_impl *impl) +{ + nir_calc_dominance_impl(impl); + + insert_phi_nodes(impl); + + rewrite_state state; + init_rewrite_state(impl, &state); + + rewrite_block(impl->start_block, &state); + + remove_unused_regs(impl, &state); + + destroy_rewrite_state(&state); +} + +void +nir_convert_to_ssa(nir_shader *shader) +{ + nir_foreach_overload(shader, overload) { + if (overload->impl) + nir_convert_to_ssa_impl(overload->impl); + } +}