From 0b43e0b4ddbd9fdac70658164f3dbf653067a4de Mon Sep 17 00:00:00 2001 From: marha Date: Mon, 11 Jul 2011 08:47:29 +0200 Subject: mesa update 11 july 2011 --- mesalib/src/glsl/ir_optimization.h | 2 - mesalib/src/glsl/ir_reader.cpp | 2012 ++++++++++++++++++------------------ mesalib/src/glsl/linker.cpp | 2 +- mesalib/src/glsl/lower_jumps.cpp | 1580 ++++++++++++++++++---------- 4 files changed, 2018 insertions(+), 1578 deletions(-) (limited to 'mesalib/src/glsl') diff --git a/mesalib/src/glsl/ir_optimization.h b/mesalib/src/glsl/ir_optimization.h index dd265673c..59a040751 100644 --- a/mesalib/src/glsl/ir_optimization.h +++ b/mesalib/src/glsl/ir_optimization.h @@ -56,10 +56,8 @@ bool do_if_simplification(exec_list *instructions); bool do_discard_simplification(exec_list *instructions); bool lower_if_to_cond_assign(exec_list *instructions, unsigned max_depth = 0); bool do_mat_op_to_vec(exec_list *instructions); -bool do_mod_to_fract(exec_list *instructions); bool do_noop_swizzle(exec_list *instructions); bool do_structure_splitting(exec_list *instructions); -bool do_sub_to_add_neg(exec_list *instructions); bool do_swizzle_swizzle(exec_list *instructions); bool do_tree_grafting(exec_list *instructions); bool do_vec_index_to_cond_assign(exec_list *instructions); diff --git a/mesalib/src/glsl/ir_reader.cpp b/mesalib/src/glsl/ir_reader.cpp index 201e436be..f3a621734 100644 --- a/mesalib/src/glsl/ir_reader.cpp +++ b/mesalib/src/glsl/ir_reader.cpp @@ -1,1005 +1,1007 @@ -/* - * Copyright © 2010 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. - */ - -#include "ir_reader.h" -#include "glsl_parser_extras.h" -#include "glsl_types.h" -#include "s_expression.h" - -const static bool debug = false; - -class ir_reader { -public: - ir_reader(_mesa_glsl_parse_state *); - - void read(exec_list *instructions, const char *src, bool scan_for_protos); - -private: - void *mem_ctx; - _mesa_glsl_parse_state *state; - - void ir_read_error(s_expression *, const char *fmt, ...); - - const glsl_type *read_type(s_expression *); - - void scan_for_prototypes(exec_list *, s_expression *); - ir_function *read_function(s_expression *, bool skip_body); - void read_function_sig(ir_function *, s_expression *, bool skip_body); - - void read_instructions(exec_list *, s_expression *, ir_loop *); - ir_instruction *read_instruction(s_expression *, ir_loop *); - ir_variable *read_declaration(s_expression *); - ir_if *read_if(s_expression *, ir_loop *); - ir_loop *read_loop(s_expression *); - ir_return *read_return(s_expression *); - ir_rvalue *read_rvalue(s_expression *); - ir_assignment *read_assignment(s_expression *); - ir_expression *read_expression(s_expression *); - ir_call *read_call(s_expression *); - ir_swizzle *read_swizzle(s_expression *); - ir_constant *read_constant(s_expression *); - ir_texture *read_texture(s_expression *); - - ir_dereference *read_dereference(s_expression *); -}; - -ir_reader::ir_reader(_mesa_glsl_parse_state *state) : state(state) -{ - this->mem_ctx = state; -} - -void -_mesa_glsl_read_ir(_mesa_glsl_parse_state *state, exec_list *instructions, - const char *src, bool scan_for_protos) -{ - ir_reader r(state); - r.read(instructions, src, scan_for_protos); -} - -void -ir_reader::read(exec_list *instructions, const char *src, bool scan_for_protos) -{ - s_expression *expr = s_expression::read_expression(mem_ctx, src); - if (expr == NULL) { - ir_read_error(NULL, "couldn't parse S-Expression."); - return; - } - - if (scan_for_protos) { - scan_for_prototypes(instructions, expr); - if (state->error) - return; - } - - read_instructions(instructions, expr, NULL); - ralloc_free(expr); - - if (debug) - validate_ir_tree(instructions); -} - -void -ir_reader::ir_read_error(s_expression *expr, const char *fmt, ...) -{ - va_list ap; - - state->error = true; - - if (state->current_function != NULL) - ralloc_asprintf_append(&state->info_log, "In function %s:\n", - state->current_function->function_name()); - ralloc_strcat(&state->info_log, "error: "); - - va_start(ap, fmt); - ralloc_vasprintf_append(&state->info_log, fmt, ap); - va_end(ap); - ralloc_strcat(&state->info_log, "\n"); - - if (expr != NULL) { - ralloc_strcat(&state->info_log, "...in this context:\n "); - expr->print(); - ralloc_strcat(&state->info_log, "\n\n"); - } -} - -const glsl_type * -ir_reader::read_type(s_expression *expr) -{ - s_expression *s_base_type; - s_int *s_size; - - s_pattern pat[] = { "array", s_base_type, s_size }; - if (MATCH(expr, pat)) { - const glsl_type *base_type = read_type(s_base_type); - if (base_type == NULL) { - ir_read_error(NULL, "when reading base type of array type"); - return NULL; - } - - return glsl_type::get_array_instance(base_type, s_size->value()); - } - - s_symbol *type_sym = SX_AS_SYMBOL(expr); - if (type_sym == NULL) { - ir_read_error(expr, "expected "); - return NULL; - } - - const glsl_type *type = state->symbols->get_type(type_sym->value()); - if (type == NULL) - ir_read_error(expr, "invalid type: %s", type_sym->value()); - - return type; -} - - -void -ir_reader::scan_for_prototypes(exec_list *instructions, s_expression *expr) -{ - s_list *list = SX_AS_LIST(expr); - if (list == NULL) { - ir_read_error(expr, "Expected ( ...); found an atom."); - return; - } - - foreach_iter(exec_list_iterator, it, list->subexpressions) { - s_list *sub = SX_AS_LIST(it.get()); - if (sub == NULL) - continue; // not a (function ...); ignore it. - - s_symbol *tag = SX_AS_SYMBOL(sub->subexpressions.get_head()); - if (tag == NULL || strcmp(tag->value(), "function") != 0) - continue; // not a (function ...); ignore it. - - ir_function *f = read_function(sub, true); - if (f == NULL) - return; - instructions->push_tail(f); - } -} - -ir_function * -ir_reader::read_function(s_expression *expr, bool skip_body) -{ - bool added = false; - s_symbol *name; - - s_pattern pat[] = { "function", name }; - if (!PARTIAL_MATCH(expr, pat)) { - ir_read_error(expr, "Expected (function (signature ...) ...)"); - return NULL; - } - - ir_function *f = state->symbols->get_function(name->value()); - if (f == NULL) { - f = new(mem_ctx) ir_function(name->value()); - added = state->symbols->add_function(f); - assert(added); - } - - exec_list_iterator it = ((s_list *) expr)->subexpressions.iterator(); - it.next(); // skip "function" tag - it.next(); // skip function name - for (/* nothing */; it.has_next(); it.next()) { - s_expression *s_sig = (s_expression *) it.get(); - read_function_sig(f, s_sig, skip_body); - } - return added ? f : NULL; -} - -void -ir_reader::read_function_sig(ir_function *f, s_expression *expr, bool skip_body) -{ - s_expression *type_expr; - s_list *paramlist; - s_list *body_list; - - s_pattern pat[] = { "signature", type_expr, paramlist, body_list }; - if (!MATCH(expr, pat)) { - ir_read_error(expr, "Expected (signature (parameters ...) " - "( ...))"); - return; - } - - const glsl_type *return_type = read_type(type_expr); - if (return_type == NULL) - return; - - s_symbol *paramtag = SX_AS_SYMBOL(paramlist->subexpressions.get_head()); - if (paramtag == NULL || strcmp(paramtag->value(), "parameters") != 0) { - ir_read_error(paramlist, "Expected (parameters ...)"); - return; - } - - // Read the parameters list into a temporary place. - exec_list hir_parameters; - state->symbols->push_scope(); - - exec_list_iterator it = paramlist->subexpressions.iterator(); - for (it.next() /* skip "parameters" */; it.has_next(); it.next()) { - ir_variable *var = read_declaration((s_expression *) it.get()); - if (var == NULL) - return; - - hir_parameters.push_tail(var); - } - - ir_function_signature *sig = f->exact_matching_signature(&hir_parameters); - if (sig == NULL && skip_body) { - /* If scanning for prototypes, generate a new signature. */ - sig = new(mem_ctx) ir_function_signature(return_type); - sig->is_builtin = true; - f->add_signature(sig); - } else if (sig != NULL) { - const char *badvar = sig->qualifiers_match(&hir_parameters); - if (badvar != NULL) { - ir_read_error(expr, "function `%s' parameter `%s' qualifiers " - "don't match prototype", f->name, badvar); - return; - } - - if (sig->return_type != return_type) { - ir_read_error(expr, "function `%s' return type doesn't " - "match prototype", f->name); - return; - } - } else { - /* No prototype for this body exists - skip it. */ - state->symbols->pop_scope(); - return; - } - assert(sig != NULL); - - sig->replace_parameters(&hir_parameters); - - if (!skip_body && !body_list->subexpressions.is_empty()) { - if (sig->is_defined) { - ir_read_error(expr, "function %s redefined", f->name); - return; - } - state->current_function = sig; - read_instructions(&sig->body, body_list, NULL); - state->current_function = NULL; - sig->is_defined = true; - } - - state->symbols->pop_scope(); -} - -void -ir_reader::read_instructions(exec_list *instructions, s_expression *expr, - ir_loop *loop_ctx) -{ - // Read in a list of instructions - s_list *list = SX_AS_LIST(expr); - if (list == NULL) { - ir_read_error(expr, "Expected ( ...); found an atom."); - return; - } - - foreach_iter(exec_list_iterator, it, list->subexpressions) { - s_expression *sub = (s_expression*) it.get(); - ir_instruction *ir = read_instruction(sub, loop_ctx); - if (ir != NULL) { - /* Global variable declarations should be moved to the top, before - * any functions that might use them. Functions are added to the - * instruction stream when scanning for prototypes, so without this - * hack, they always appear before variable declarations. - */ - if (state->current_function == NULL && ir->as_variable() != NULL) - instructions->push_head(ir); - else - instructions->push_tail(ir); - } - } -} - - -ir_instruction * -ir_reader::read_instruction(s_expression *expr, ir_loop *loop_ctx) -{ - s_symbol *symbol = SX_AS_SYMBOL(expr); - if (symbol != NULL) { - if (strcmp(symbol->value(), "break") == 0 && loop_ctx != NULL) - return new(mem_ctx) ir_loop_jump(ir_loop_jump::jump_break); - if (strcmp(symbol->value(), "continue") == 0 && loop_ctx != NULL) - return new(mem_ctx) ir_loop_jump(ir_loop_jump::jump_continue); - } - - s_list *list = SX_AS_LIST(expr); - if (list == NULL || list->subexpressions.is_empty()) { - ir_read_error(expr, "Invalid instruction.\n"); - return NULL; - } - - s_symbol *tag = SX_AS_SYMBOL(list->subexpressions.get_head()); - if (tag == NULL) { - ir_read_error(expr, "expected instruction tag"); - return NULL; - } - - ir_instruction *inst = NULL; - if (strcmp(tag->value(), "declare") == 0) { - inst = read_declaration(list); - } else if (strcmp(tag->value(), "assign") == 0) { - inst = read_assignment(list); - } else if (strcmp(tag->value(), "if") == 0) { - inst = read_if(list, loop_ctx); - } else if (strcmp(tag->value(), "loop") == 0) { - inst = read_loop(list); - } else if (strcmp(tag->value(), "return") == 0) { - inst = read_return(list); - } else if (strcmp(tag->value(), "function") == 0) { - inst = read_function(list, false); - } else { - inst = read_rvalue(list); - if (inst == NULL) - ir_read_error(NULL, "when reading instruction"); - } - return inst; -} - -ir_variable * -ir_reader::read_declaration(s_expression *expr) -{ - s_list *s_quals; - s_expression *s_type; - s_symbol *s_name; - - s_pattern pat[] = { "declare", s_quals, s_type, s_name }; - if (!MATCH(expr, pat)) { - ir_read_error(expr, "expected (declare () )"); - return NULL; - } - - const glsl_type *type = read_type(s_type); - if (type == NULL) - return NULL; - - ir_variable *var = new(mem_ctx) ir_variable(type, s_name->value(), - ir_var_auto); - - foreach_iter(exec_list_iterator, it, s_quals->subexpressions) { - s_symbol *qualifier = SX_AS_SYMBOL(it.get()); - if (qualifier == NULL) { - ir_read_error(expr, "qualifier list must contain only symbols"); - return NULL; - } - - // FINISHME: Check for duplicate/conflicting qualifiers. - if (strcmp(qualifier->value(), "centroid") == 0) { - var->centroid = 1; - } else if (strcmp(qualifier->value(), "invariant") == 0) { - var->invariant = 1; - } else if (strcmp(qualifier->value(), "uniform") == 0) { - var->mode = ir_var_uniform; - } else if (strcmp(qualifier->value(), "auto") == 0) { - var->mode = ir_var_auto; - } else if (strcmp(qualifier->value(), "in") == 0) { - var->mode = ir_var_in; - } else if (strcmp(qualifier->value(), "const_in") == 0) { - var->mode = ir_var_const_in; - } else if (strcmp(qualifier->value(), "out") == 0) { - var->mode = ir_var_out; - } else if (strcmp(qualifier->value(), "inout") == 0) { - var->mode = ir_var_inout; - } else if (strcmp(qualifier->value(), "smooth") == 0) { - var->interpolation = ir_var_smooth; - } else if (strcmp(qualifier->value(), "flat") == 0) { - var->interpolation = ir_var_flat; - } else if (strcmp(qualifier->value(), "noperspective") == 0) { - var->interpolation = ir_var_noperspective; - } else { - ir_read_error(expr, "unknown qualifier: %s", qualifier->value()); - return NULL; - } - } - - // Add the variable to the symbol table - state->symbols->add_variable(var); - - return var; -} - - -ir_if * -ir_reader::read_if(s_expression *expr, ir_loop *loop_ctx) -{ - s_expression *s_cond; - s_expression *s_then; - s_expression *s_else; - - s_pattern pat[] = { "if", s_cond, s_then, s_else }; - if (!MATCH(expr, pat)) { - ir_read_error(expr, "expected (if (...) (...))"); - return NULL; - } - - ir_rvalue *condition = read_rvalue(s_cond); - if (condition == NULL) { - ir_read_error(NULL, "when reading condition of (if ...)"); - return NULL; - } - - ir_if *iff = new(mem_ctx) ir_if(condition); - - read_instructions(&iff->then_instructions, s_then, loop_ctx); - read_instructions(&iff->else_instructions, s_else, loop_ctx); - if (state->error) { - delete iff; - iff = NULL; - } - return iff; -} - - -ir_loop * -ir_reader::read_loop(s_expression *expr) -{ - s_expression *s_counter, *s_from, *s_to, *s_inc, *s_body; - - s_pattern pat[] = { "loop", s_counter, s_from, s_to, s_inc, s_body }; - if (!MATCH(expr, pat)) { - ir_read_error(expr, "expected (loop " - " )"); - return NULL; - } - - // FINISHME: actually read the count/from/to fields. - - ir_loop *loop = new(mem_ctx) ir_loop; - read_instructions(&loop->body_instructions, s_body, loop); - if (state->error) { - delete loop; - loop = NULL; - } - return loop; -} - - -ir_return * -ir_reader::read_return(s_expression *expr) -{ - s_expression *s_retval; - - s_pattern pat[] = { "return", s_retval}; - if (!MATCH(expr, pat)) { - ir_read_error(expr, "expected (return )"); - return NULL; - } - - ir_rvalue *retval = read_rvalue(s_retval); - if (retval == NULL) { - ir_read_error(NULL, "when reading return value"); - return NULL; - } - - return new(mem_ctx) ir_return(retval); -} - - -ir_rvalue * -ir_reader::read_rvalue(s_expression *expr) -{ - s_list *list = SX_AS_LIST(expr); - if (list == NULL || list->subexpressions.is_empty()) - return NULL; - - s_symbol *tag = SX_AS_SYMBOL(list->subexpressions.get_head()); - if (tag == NULL) { - ir_read_error(expr, "expected rvalue tag"); - return NULL; - } - - ir_rvalue *rvalue = read_dereference(list); - if (rvalue != NULL || state->error) - return rvalue; - else if (strcmp(tag->value(), "swiz") == 0) { - rvalue = read_swizzle(list); - } else if (strcmp(tag->value(), "expression") == 0) { - rvalue = read_expression(list); - } else if (strcmp(tag->value(), "call") == 0) { - rvalue = read_call(list); - } else if (strcmp(tag->value(), "constant") == 0) { - rvalue = read_constant(list); - } else { - rvalue = read_texture(list); - if (rvalue == NULL && !state->error) - ir_read_error(expr, "unrecognized rvalue tag: %s", tag->value()); - } - - return rvalue; -} - -ir_assignment * -ir_reader::read_assignment(s_expression *expr) -{ - s_expression *cond_expr = NULL; - s_expression *lhs_expr, *rhs_expr; - s_list *mask_list; - - s_pattern pat4[] = { "assign", mask_list, lhs_expr, rhs_expr }; - s_pattern pat5[] = { "assign", cond_expr, mask_list, lhs_expr, rhs_expr }; - if (!MATCH(expr, pat4) && !MATCH(expr, pat5)) { - ir_read_error(expr, "expected (assign [] () " - " )"); - return NULL; - } - - ir_rvalue *condition = NULL; - if (cond_expr != NULL) { - condition = read_rvalue(cond_expr); - if (condition == NULL) { - ir_read_error(NULL, "when reading condition of assignment"); - return NULL; - } - } - - unsigned mask = 0; - - s_symbol *mask_symbol; - s_pattern mask_pat[] = { mask_symbol }; - if (MATCH(mask_list, mask_pat)) { - const char *mask_str = mask_symbol->value(); - unsigned mask_length = strlen(mask_str); - if (mask_length > 4) { - ir_read_error(expr, "invalid write mask: %s", mask_str); - return NULL; - } - - const unsigned idx_map[] = { 3, 0, 1, 2 }; /* w=bit 3, x=0, y=1, z=2 */ - - for (unsigned i = 0; i < mask_length; i++) { - if (mask_str[i] < 'w' || mask_str[i] > 'z') { - ir_read_error(expr, "write mask contains invalid character: %c", - mask_str[i]); - return NULL; - } - mask |= 1 << idx_map[mask_str[i] - 'w']; - } - } else if (!mask_list->subexpressions.is_empty()) { - ir_read_error(mask_list, "expected () or ()"); - return NULL; - } - - ir_dereference *lhs = read_dereference(lhs_expr); - if (lhs == NULL) { - ir_read_error(NULL, "when reading left-hand side of assignment"); - return NULL; - } - - ir_rvalue *rhs = read_rvalue(rhs_expr); - if (rhs == NULL) { - ir_read_error(NULL, "when reading right-hand side of assignment"); - return NULL; - } - - if (mask == 0 && (lhs->type->is_vector() || lhs->type->is_scalar())) { - ir_read_error(expr, "non-zero write mask required."); - return NULL; - } - - return new(mem_ctx) ir_assignment(lhs, rhs, condition, mask); -} - -ir_call * -ir_reader::read_call(s_expression *expr) -{ - s_symbol *name; - s_list *params; - - s_pattern pat[] = { "call", name, params }; - if (!MATCH(expr, pat)) { - ir_read_error(expr, "expected (call ( ...))"); - return NULL; - } - - exec_list parameters; - - foreach_iter(exec_list_iterator, it, params->subexpressions) { - s_expression *expr = (s_expression*) it.get(); - ir_rvalue *param = read_rvalue(expr); - if (param == NULL) { - ir_read_error(expr, "when reading parameter to function call"); - return NULL; - } - parameters.push_tail(param); - } - - ir_function *f = state->symbols->get_function(name->value()); - if (f == NULL) { - ir_read_error(expr, "found call to undefined function %s", - name->value()); - return NULL; - } - - ir_function_signature *callee = f->matching_signature(¶meters); - if (callee == NULL) { - ir_read_error(expr, "couldn't find matching signature for function " - "%s", name->value()); - return NULL; - } - - return new(mem_ctx) ir_call(callee, ¶meters); -} - -ir_expression * -ir_reader::read_expression(s_expression *expr) -{ - s_expression *s_type; - s_symbol *s_op; - s_expression *s_arg1; - - s_pattern pat[] = { "expression", s_type, s_op, s_arg1 }; - if (!PARTIAL_MATCH(expr, pat)) { - ir_read_error(expr, "expected (expression " - " [])"); - return NULL; - } - s_expression *s_arg2 = (s_expression *) s_arg1->next; // may be tail sentinel - - const glsl_type *type = read_type(s_type); - if (type == NULL) - return NULL; - - /* Read the operator */ - ir_expression_operation op = ir_expression::get_operator(s_op->value()); - if (op == (ir_expression_operation) -1) { - ir_read_error(expr, "invalid operator: %s", s_op->value()); - return NULL; - } - - unsigned num_operands = ir_expression::get_num_operands(op); - if (num_operands == 1 && !s_arg1->next->is_tail_sentinel()) { - ir_read_error(expr, "expected (expression %s )", - s_op->value()); - return NULL; - } - - ir_rvalue *arg1 = read_rvalue(s_arg1); - ir_rvalue *arg2 = NULL; - if (arg1 == NULL) { - ir_read_error(NULL, "when reading first operand of %s", s_op->value()); - return NULL; - } - - if (num_operands == 2) { - if (s_arg2->is_tail_sentinel() || !s_arg2->next->is_tail_sentinel()) { - ir_read_error(expr, "expected (expression %s " - ")", s_op->value()); - return NULL; - } - arg2 = read_rvalue(s_arg2); - if (arg2 == NULL) { - ir_read_error(NULL, "when reading second operand of %s", - s_op->value()); - return NULL; - } - } - - return new(mem_ctx) ir_expression(op, type, arg1, arg2); -} - -ir_swizzle * -ir_reader::read_swizzle(s_expression *expr) -{ - s_symbol *swiz; - s_expression *sub; - - s_pattern pat[] = { "swiz", swiz, sub }; - if (!MATCH(expr, pat)) { - ir_read_error(expr, "expected (swiz )"); - return NULL; - } - - if (strlen(swiz->value()) > 4) { - ir_read_error(expr, "expected a valid swizzle; found %s", swiz->value()); - return NULL; - } - - ir_rvalue *rvalue = read_rvalue(sub); - if (rvalue == NULL) - return NULL; - - ir_swizzle *ir = ir_swizzle::create(rvalue, swiz->value(), - rvalue->type->vector_elements); - if (ir == NULL) - ir_read_error(expr, "invalid swizzle"); - - return ir; -} - -ir_constant * -ir_reader::read_constant(s_expression *expr) -{ - s_expression *type_expr; - s_list *values; - - s_pattern pat[] = { "constant", type_expr, values }; - if (!MATCH(expr, pat)) { - ir_read_error(expr, "expected (constant (...))"); - return NULL; - } - - const glsl_type *type = read_type(type_expr); - if (type == NULL) - return NULL; - - if (values == NULL) { - ir_read_error(expr, "expected (constant (...))"); - return NULL; - } - - if (type->is_array()) { - unsigned elements_supplied = 0; - exec_list elements; - foreach_iter(exec_list_iterator, it, values->subexpressions) { - s_expression *elt = (s_expression *) it.get(); - ir_constant *ir_elt = read_constant(elt); - if (ir_elt == NULL) - return NULL; - elements.push_tail(ir_elt); - elements_supplied++; - } - - if (elements_supplied != type->length) { - ir_read_error(values, "expected exactly %u array elements, " - "given %u", type->length, elements_supplied); - return NULL; - } - return new(mem_ctx) ir_constant(type, &elements); - } - - const glsl_type *const base_type = type->get_base_type(); - - ir_constant_data data = { { 0 } }; - - // Read in list of values (at most 16). - int k = 0; - foreach_iter(exec_list_iterator, it, values->subexpressions) { - if (k >= 16) { - ir_read_error(values, "expected at most 16 numbers"); - return NULL; - } - - s_expression *expr = (s_expression*) it.get(); - - if (base_type->base_type == GLSL_TYPE_FLOAT) { - s_number *value = SX_AS_NUMBER(expr); - if (value == NULL) { - ir_read_error(values, "expected numbers"); - return NULL; - } - data.f[k] = value->fvalue(); - } else { - s_int *value = SX_AS_INT(expr); - if (value == NULL) { - ir_read_error(values, "expected integers"); - return NULL; - } - - switch (base_type->base_type) { - case GLSL_TYPE_UINT: { - data.u[k] = value->value(); - break; - } - case GLSL_TYPE_INT: { - data.i[k] = value->value(); - break; - } - case GLSL_TYPE_BOOL: { - data.b[k] = value->value(); - break; - } - default: - ir_read_error(values, "unsupported constant type"); - return NULL; - } - } - ++k; - } - - return new(mem_ctx) ir_constant(type, &data); -} - -ir_dereference * -ir_reader::read_dereference(s_expression *expr) -{ - s_symbol *s_var; - s_expression *s_subject; - s_expression *s_index; - s_symbol *s_field; - - s_pattern var_pat[] = { "var_ref", s_var }; - s_pattern array_pat[] = { "array_ref", s_subject, s_index }; - s_pattern record_pat[] = { "record_ref", s_subject, s_field }; - - if (MATCH(expr, var_pat)) { - ir_variable *var = state->symbols->get_variable(s_var->value()); - if (var == NULL) { - ir_read_error(expr, "undeclared variable: %s", s_var->value()); - return NULL; - } - return new(mem_ctx) ir_dereference_variable(var); - } else if (MATCH(expr, array_pat)) { - ir_rvalue *subject = read_rvalue(s_subject); - if (subject == NULL) { - ir_read_error(NULL, "when reading the subject of an array_ref"); - return NULL; - } - - ir_rvalue *idx = read_rvalue(s_index); - if (subject == NULL) { - ir_read_error(NULL, "when reading the index of an array_ref"); - return NULL; - } - return new(mem_ctx) ir_dereference_array(subject, idx); - } else if (MATCH(expr, record_pat)) { - ir_rvalue *subject = read_rvalue(s_subject); - if (subject == NULL) { - ir_read_error(NULL, "when reading the subject of a record_ref"); - return NULL; - } - return new(mem_ctx) ir_dereference_record(subject, s_field->value()); - } - return NULL; -} - -ir_texture * -ir_reader::read_texture(s_expression *expr) -{ - s_symbol *tag = NULL; - s_expression *s_type = NULL; - s_expression *s_sampler = NULL; - s_expression *s_coord = NULL; - s_expression *s_offset = NULL; - s_expression *s_proj = NULL; - s_list *s_shadow = NULL; - s_expression *s_lod = NULL; - - ir_texture_opcode op = ir_tex; /* silence warning */ - - s_pattern tex_pattern[] = - { "tex", s_type, s_sampler, s_coord, s_offset, s_proj, s_shadow }; - s_pattern txf_pattern[] = - { "txf", s_type, s_sampler, s_coord, s_offset, s_lod }; - s_pattern other_pattern[] = - { tag, s_type, s_sampler, s_coord, s_offset, s_proj, s_shadow, s_lod }; - - if (MATCH(expr, tex_pattern)) { - op = ir_tex; - } else if (MATCH(expr, txf_pattern)) { - op = ir_txf; - } else if (MATCH(expr, other_pattern)) { - op = ir_texture::get_opcode(tag->value()); - if (op == -1) - return NULL; - } else { - ir_read_error(NULL, "unexpected texture pattern"); - return NULL; - } - - ir_texture *tex = new(mem_ctx) ir_texture(op); - - // Read return type - const glsl_type *type = read_type(s_type); - if (type == NULL) { - ir_read_error(NULL, "when reading type in (%s ...)", - tex->opcode_string()); - return NULL; - } - - // Read sampler (must be a deref) - ir_dereference *sampler = read_dereference(s_sampler); - if (sampler == NULL) { - ir_read_error(NULL, "when reading sampler in (%s ...)", - tex->opcode_string()); - return NULL; - } - tex->set_sampler(sampler, type); - - // Read coordinate (any rvalue) - tex->coordinate = read_rvalue(s_coord); - if (tex->coordinate == NULL) { - ir_read_error(NULL, "when reading coordinate in (%s ...)", - tex->opcode_string()); - return NULL; - } - - // Read texel offset - either 0 or an rvalue. - s_int *si_offset = SX_AS_INT(s_offset); - if (si_offset == NULL || si_offset->value() != 0) { - tex->offset = read_rvalue(s_offset); - if (tex->offset == NULL) { - ir_read_error(s_offset, "expected 0 or an expression"); - return NULL; - } - } - - if (op != ir_txf) { - s_int *proj_as_int = SX_AS_INT(s_proj); - if (proj_as_int && proj_as_int->value() == 1) { - tex->projector = NULL; - } else { - tex->projector = read_rvalue(s_proj); - if (tex->projector == NULL) { - ir_read_error(NULL, "when reading projective divide in (%s ..)", - tex->opcode_string()); - return NULL; - } - } - - if (s_shadow->subexpressions.is_empty()) { - tex->shadow_comparitor = NULL; - } else { - tex->shadow_comparitor = read_rvalue(s_shadow); - if (tex->shadow_comparitor == NULL) { - ir_read_error(NULL, "when reading shadow comparitor in (%s ..)", - tex->opcode_string()); - return NULL; - } - } - } - - switch (op) { - case ir_txb: - tex->lod_info.bias = read_rvalue(s_lod); - if (tex->lod_info.bias == NULL) { - ir_read_error(NULL, "when reading LOD bias in (txb ...)"); - return NULL; - } - break; - case ir_txl: - case ir_txf: - tex->lod_info.lod = read_rvalue(s_lod); - if (tex->lod_info.lod == NULL) { - ir_read_error(NULL, "when reading LOD in (%s ...)", - tex->opcode_string()); - return NULL; - } - break; - case ir_txd: { - s_expression *s_dx, *s_dy; - s_pattern dxdy_pat[] = { s_dx, s_dy }; - if (!MATCH(s_lod, dxdy_pat)) { - ir_read_error(s_lod, "expected (dPdx dPdy) in (txd ...)"); - return NULL; - } - tex->lod_info.grad.dPdx = read_rvalue(s_dx); - if (tex->lod_info.grad.dPdx == NULL) { - ir_read_error(NULL, "when reading dPdx in (txd ...)"); - return NULL; - } - tex->lod_info.grad.dPdy = read_rvalue(s_dy); - if (tex->lod_info.grad.dPdy == NULL) { - ir_read_error(NULL, "when reading dPdy in (txd ...)"); - return NULL; - } - break; - } - default: - // tex doesn't have any extra parameters. - break; - }; - return tex; -} +/* + * Copyright © 2010 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. + */ + +#include "ir_reader.h" +#include "glsl_parser_extras.h" +#include "glsl_types.h" +#include "s_expression.h" + +const static bool debug = false; + +class ir_reader { +public: + ir_reader(_mesa_glsl_parse_state *); + + void read(exec_list *instructions, const char *src, bool scan_for_protos); + +private: + void *mem_ctx; + _mesa_glsl_parse_state *state; + + void ir_read_error(s_expression *, const char *fmt, ...); + + const glsl_type *read_type(s_expression *); + + void scan_for_prototypes(exec_list *, s_expression *); + ir_function *read_function(s_expression *, bool skip_body); + void read_function_sig(ir_function *, s_expression *, bool skip_body); + + void read_instructions(exec_list *, s_expression *, ir_loop *); + ir_instruction *read_instruction(s_expression *, ir_loop *); + ir_variable *read_declaration(s_expression *); + ir_if *read_if(s_expression *, ir_loop *); + ir_loop *read_loop(s_expression *); + ir_return *read_return(s_expression *); + ir_rvalue *read_rvalue(s_expression *); + ir_assignment *read_assignment(s_expression *); + ir_expression *read_expression(s_expression *); + ir_call *read_call(s_expression *); + ir_swizzle *read_swizzle(s_expression *); + ir_constant *read_constant(s_expression *); + ir_texture *read_texture(s_expression *); + + ir_dereference *read_dereference(s_expression *); +}; + +ir_reader::ir_reader(_mesa_glsl_parse_state *state) : state(state) +{ + this->mem_ctx = state; +} + +void +_mesa_glsl_read_ir(_mesa_glsl_parse_state *state, exec_list *instructions, + const char *src, bool scan_for_protos) +{ + ir_reader r(state); + r.read(instructions, src, scan_for_protos); +} + +void +ir_reader::read(exec_list *instructions, const char *src, bool scan_for_protos) +{ + s_expression *expr = s_expression::read_expression(mem_ctx, src); + if (expr == NULL) { + ir_read_error(NULL, "couldn't parse S-Expression."); + return; + } + + if (scan_for_protos) { + scan_for_prototypes(instructions, expr); + if (state->error) + return; + } + + read_instructions(instructions, expr, NULL); + ralloc_free(expr); + + if (debug) + validate_ir_tree(instructions); +} + +void +ir_reader::ir_read_error(s_expression *expr, const char *fmt, ...) +{ + va_list ap; + + state->error = true; + + if (state->current_function != NULL) + ralloc_asprintf_append(&state->info_log, "In function %s:\n", + state->current_function->function_name()); + ralloc_strcat(&state->info_log, "error: "); + + va_start(ap, fmt); + ralloc_vasprintf_append(&state->info_log, fmt, ap); + va_end(ap); + ralloc_strcat(&state->info_log, "\n"); + + if (expr != NULL) { + ralloc_strcat(&state->info_log, "...in this context:\n "); + expr->print(); + ralloc_strcat(&state->info_log, "\n\n"); + } +} + +const glsl_type * +ir_reader::read_type(s_expression *expr) +{ + s_expression *s_base_type; + s_int *s_size; + + s_pattern pat[] = { "array", s_base_type, s_size }; + if (MATCH(expr, pat)) { + const glsl_type *base_type = read_type(s_base_type); + if (base_type == NULL) { + ir_read_error(NULL, "when reading base type of array type"); + return NULL; + } + + return glsl_type::get_array_instance(base_type, s_size->value()); + } + + s_symbol *type_sym = SX_AS_SYMBOL(expr); + if (type_sym == NULL) { + ir_read_error(expr, "expected "); + return NULL; + } + + const glsl_type *type = state->symbols->get_type(type_sym->value()); + if (type == NULL) + ir_read_error(expr, "invalid type: %s", type_sym->value()); + + return type; +} + + +void +ir_reader::scan_for_prototypes(exec_list *instructions, s_expression *expr) +{ + s_list *list = SX_AS_LIST(expr); + if (list == NULL) { + ir_read_error(expr, "Expected ( ...); found an atom."); + return; + } + + foreach_iter(exec_list_iterator, it, list->subexpressions) { + s_list *sub = SX_AS_LIST(it.get()); + if (sub == NULL) + continue; // not a (function ...); ignore it. + + s_symbol *tag = SX_AS_SYMBOL(sub->subexpressions.get_head()); + if (tag == NULL || strcmp(tag->value(), "function") != 0) + continue; // not a (function ...); ignore it. + + ir_function *f = read_function(sub, true); + if (f == NULL) + return; + instructions->push_tail(f); + } +} + +ir_function * +ir_reader::read_function(s_expression *expr, bool skip_body) +{ + bool added = false; + s_symbol *name; + + s_pattern pat[] = { "function", name }; + if (!PARTIAL_MATCH(expr, pat)) { + ir_read_error(expr, "Expected (function (signature ...) ...)"); + return NULL; + } + + ir_function *f = state->symbols->get_function(name->value()); + if (f == NULL) { + f = new(mem_ctx) ir_function(name->value()); + added = state->symbols->add_function(f); + assert(added); + } + + exec_list_iterator it = ((s_list *) expr)->subexpressions.iterator(); + it.next(); // skip "function" tag + it.next(); // skip function name + for (/* nothing */; it.has_next(); it.next()) { + s_expression *s_sig = (s_expression *) it.get(); + read_function_sig(f, s_sig, skip_body); + } + return added ? f : NULL; +} + +void +ir_reader::read_function_sig(ir_function *f, s_expression *expr, bool skip_body) +{ + s_expression *type_expr; + s_list *paramlist; + s_list *body_list; + + s_pattern pat[] = { "signature", type_expr, paramlist, body_list }; + if (!MATCH(expr, pat)) { + ir_read_error(expr, "Expected (signature (parameters ...) " + "( ...))"); + return; + } + + const glsl_type *return_type = read_type(type_expr); + if (return_type == NULL) + return; + + s_symbol *paramtag = SX_AS_SYMBOL(paramlist->subexpressions.get_head()); + if (paramtag == NULL || strcmp(paramtag->value(), "parameters") != 0) { + ir_read_error(paramlist, "Expected (parameters ...)"); + return; + } + + // Read the parameters list into a temporary place. + exec_list hir_parameters; + state->symbols->push_scope(); + + exec_list_iterator it = paramlist->subexpressions.iterator(); + for (it.next() /* skip "parameters" */; it.has_next(); it.next()) { + ir_variable *var = read_declaration((s_expression *) it.get()); + if (var == NULL) + return; + + hir_parameters.push_tail(var); + } + + ir_function_signature *sig = f->exact_matching_signature(&hir_parameters); + if (sig == NULL && skip_body) { + /* If scanning for prototypes, generate a new signature. */ + sig = new(mem_ctx) ir_function_signature(return_type); + sig->is_builtin = true; + f->add_signature(sig); + } else if (sig != NULL) { + const char *badvar = sig->qualifiers_match(&hir_parameters); + if (badvar != NULL) { + ir_read_error(expr, "function `%s' parameter `%s' qualifiers " + "don't match prototype", f->name, badvar); + return; + } + + if (sig->return_type != return_type) { + ir_read_error(expr, "function `%s' return type doesn't " + "match prototype", f->name); + return; + } + } else { + /* No prototype for this body exists - skip it. */ + state->symbols->pop_scope(); + return; + } + assert(sig != NULL); + + sig->replace_parameters(&hir_parameters); + + if (!skip_body && !body_list->subexpressions.is_empty()) { + if (sig->is_defined) { + ir_read_error(expr, "function %s redefined", f->name); + return; + } + state->current_function = sig; + read_instructions(&sig->body, body_list, NULL); + state->current_function = NULL; + sig->is_defined = true; + } + + state->symbols->pop_scope(); +} + +void +ir_reader::read_instructions(exec_list *instructions, s_expression *expr, + ir_loop *loop_ctx) +{ + // Read in a list of instructions + s_list *list = SX_AS_LIST(expr); + if (list == NULL) { + ir_read_error(expr, "Expected ( ...); found an atom."); + return; + } + + foreach_iter(exec_list_iterator, it, list->subexpressions) { + s_expression *sub = (s_expression*) it.get(); + ir_instruction *ir = read_instruction(sub, loop_ctx); + if (ir != NULL) { + /* Global variable declarations should be moved to the top, before + * any functions that might use them. Functions are added to the + * instruction stream when scanning for prototypes, so without this + * hack, they always appear before variable declarations. + */ + if (state->current_function == NULL && ir->as_variable() != NULL) + instructions->push_head(ir); + else + instructions->push_tail(ir); + } + } +} + + +ir_instruction * +ir_reader::read_instruction(s_expression *expr, ir_loop *loop_ctx) +{ + s_symbol *symbol = SX_AS_SYMBOL(expr); + if (symbol != NULL) { + if (strcmp(symbol->value(), "break") == 0 && loop_ctx != NULL) + return new(mem_ctx) ir_loop_jump(ir_loop_jump::jump_break); + if (strcmp(symbol->value(), "continue") == 0 && loop_ctx != NULL) + return new(mem_ctx) ir_loop_jump(ir_loop_jump::jump_continue); + } + + s_list *list = SX_AS_LIST(expr); + if (list == NULL || list->subexpressions.is_empty()) { + ir_read_error(expr, "Invalid instruction.\n"); + return NULL; + } + + s_symbol *tag = SX_AS_SYMBOL(list->subexpressions.get_head()); + if (tag == NULL) { + ir_read_error(expr, "expected instruction tag"); + return NULL; + } + + ir_instruction *inst = NULL; + if (strcmp(tag->value(), "declare") == 0) { + inst = read_declaration(list); + } else if (strcmp(tag->value(), "assign") == 0) { + inst = read_assignment(list); + } else if (strcmp(tag->value(), "if") == 0) { + inst = read_if(list, loop_ctx); + } else if (strcmp(tag->value(), "loop") == 0) { + inst = read_loop(list); + } else if (strcmp(tag->value(), "return") == 0) { + inst = read_return(list); + } else if (strcmp(tag->value(), "function") == 0) { + inst = read_function(list, false); + } else { + inst = read_rvalue(list); + if (inst == NULL) + ir_read_error(NULL, "when reading instruction"); + } + return inst; +} + +ir_variable * +ir_reader::read_declaration(s_expression *expr) +{ + s_list *s_quals; + s_expression *s_type; + s_symbol *s_name; + + s_pattern pat[] = { "declare", s_quals, s_type, s_name }; + if (!MATCH(expr, pat)) { + ir_read_error(expr, "expected (declare () )"); + return NULL; + } + + const glsl_type *type = read_type(s_type); + if (type == NULL) + return NULL; + + ir_variable *var = new(mem_ctx) ir_variable(type, s_name->value(), + ir_var_auto); + + foreach_iter(exec_list_iterator, it, s_quals->subexpressions) { + s_symbol *qualifier = SX_AS_SYMBOL(it.get()); + if (qualifier == NULL) { + ir_read_error(expr, "qualifier list must contain only symbols"); + return NULL; + } + + // FINISHME: Check for duplicate/conflicting qualifiers. + if (strcmp(qualifier->value(), "centroid") == 0) { + var->centroid = 1; + } else if (strcmp(qualifier->value(), "invariant") == 0) { + var->invariant = 1; + } else if (strcmp(qualifier->value(), "uniform") == 0) { + var->mode = ir_var_uniform; + } else if (strcmp(qualifier->value(), "auto") == 0) { + var->mode = ir_var_auto; + } else if (strcmp(qualifier->value(), "in") == 0) { + var->mode = ir_var_in; + } else if (strcmp(qualifier->value(), "const_in") == 0) { + var->mode = ir_var_const_in; + } else if (strcmp(qualifier->value(), "out") == 0) { + var->mode = ir_var_out; + } else if (strcmp(qualifier->value(), "inout") == 0) { + var->mode = ir_var_inout; + } else if (strcmp(qualifier->value(), "smooth") == 0) { + var->interpolation = ir_var_smooth; + } else if (strcmp(qualifier->value(), "flat") == 0) { + var->interpolation = ir_var_flat; + } else if (strcmp(qualifier->value(), "noperspective") == 0) { + var->interpolation = ir_var_noperspective; + } else { + ir_read_error(expr, "unknown qualifier: %s", qualifier->value()); + return NULL; + } + } + + // Add the variable to the symbol table + state->symbols->add_variable(var); + + return var; +} + + +ir_if * +ir_reader::read_if(s_expression *expr, ir_loop *loop_ctx) +{ + s_expression *s_cond; + s_expression *s_then; + s_expression *s_else; + + s_pattern pat[] = { "if", s_cond, s_then, s_else }; + if (!MATCH(expr, pat)) { + ir_read_error(expr, "expected (if (...) (...))"); + return NULL; + } + + ir_rvalue *condition = read_rvalue(s_cond); + if (condition == NULL) { + ir_read_error(NULL, "when reading condition of (if ...)"); + return NULL; + } + + ir_if *iff = new(mem_ctx) ir_if(condition); + + read_instructions(&iff->then_instructions, s_then, loop_ctx); + read_instructions(&iff->else_instructions, s_else, loop_ctx); + if (state->error) { + delete iff; + iff = NULL; + } + return iff; +} + + +ir_loop * +ir_reader::read_loop(s_expression *expr) +{ + s_expression *s_counter, *s_from, *s_to, *s_inc, *s_body; + + s_pattern pat[] = { "loop", s_counter, s_from, s_to, s_inc, s_body }; + if (!MATCH(expr, pat)) { + ir_read_error(expr, "expected (loop " + " )"); + return NULL; + } + + // FINISHME: actually read the count/from/to fields. + + ir_loop *loop = new(mem_ctx) ir_loop; + read_instructions(&loop->body_instructions, s_body, loop); + if (state->error) { + delete loop; + loop = NULL; + } + return loop; +} + + +ir_return * +ir_reader::read_return(s_expression *expr) +{ + s_expression *s_retval; + + s_pattern return_value_pat[] = { "return", s_retval}; + s_pattern return_void_pat[] = { "return" }; + if (MATCH(expr, return_value_pat)) { + ir_rvalue *retval = read_rvalue(s_retval); + if (retval == NULL) { + ir_read_error(NULL, "when reading return value"); + return NULL; + } + return new(mem_ctx) ir_return(retval); + } else if (MATCH(expr, return_void_pat)) { + return new(mem_ctx) ir_return; + } else { + ir_read_error(expr, "expected (return ) or (return)"); + return NULL; + } +} + + +ir_rvalue * +ir_reader::read_rvalue(s_expression *expr) +{ + s_list *list = SX_AS_LIST(expr); + if (list == NULL || list->subexpressions.is_empty()) + return NULL; + + s_symbol *tag = SX_AS_SYMBOL(list->subexpressions.get_head()); + if (tag == NULL) { + ir_read_error(expr, "expected rvalue tag"); + return NULL; + } + + ir_rvalue *rvalue = read_dereference(list); + if (rvalue != NULL || state->error) + return rvalue; + else if (strcmp(tag->value(), "swiz") == 0) { + rvalue = read_swizzle(list); + } else if (strcmp(tag->value(), "expression") == 0) { + rvalue = read_expression(list); + } else if (strcmp(tag->value(), "call") == 0) { + rvalue = read_call(list); + } else if (strcmp(tag->value(), "constant") == 0) { + rvalue = read_constant(list); + } else { + rvalue = read_texture(list); + if (rvalue == NULL && !state->error) + ir_read_error(expr, "unrecognized rvalue tag: %s", tag->value()); + } + + return rvalue; +} + +ir_assignment * +ir_reader::read_assignment(s_expression *expr) +{ + s_expression *cond_expr = NULL; + s_expression *lhs_expr, *rhs_expr; + s_list *mask_list; + + s_pattern pat4[] = { "assign", mask_list, lhs_expr, rhs_expr }; + s_pattern pat5[] = { "assign", cond_expr, mask_list, lhs_expr, rhs_expr }; + if (!MATCH(expr, pat4) && !MATCH(expr, pat5)) { + ir_read_error(expr, "expected (assign [] () " + " )"); + return NULL; + } + + ir_rvalue *condition = NULL; + if (cond_expr != NULL) { + condition = read_rvalue(cond_expr); + if (condition == NULL) { + ir_read_error(NULL, "when reading condition of assignment"); + return NULL; + } + } + + unsigned mask = 0; + + s_symbol *mask_symbol; + s_pattern mask_pat[] = { mask_symbol }; + if (MATCH(mask_list, mask_pat)) { + const char *mask_str = mask_symbol->value(); + unsigned mask_length = strlen(mask_str); + if (mask_length > 4) { + ir_read_error(expr, "invalid write mask: %s", mask_str); + return NULL; + } + + const unsigned idx_map[] = { 3, 0, 1, 2 }; /* w=bit 3, x=0, y=1, z=2 */ + + for (unsigned i = 0; i < mask_length; i++) { + if (mask_str[i] < 'w' || mask_str[i] > 'z') { + ir_read_error(expr, "write mask contains invalid character: %c", + mask_str[i]); + return NULL; + } + mask |= 1 << idx_map[mask_str[i] - 'w']; + } + } else if (!mask_list->subexpressions.is_empty()) { + ir_read_error(mask_list, "expected () or ()"); + return NULL; + } + + ir_dereference *lhs = read_dereference(lhs_expr); + if (lhs == NULL) { + ir_read_error(NULL, "when reading left-hand side of assignment"); + return NULL; + } + + ir_rvalue *rhs = read_rvalue(rhs_expr); + if (rhs == NULL) { + ir_read_error(NULL, "when reading right-hand side of assignment"); + return NULL; + } + + if (mask == 0 && (lhs->type->is_vector() || lhs->type->is_scalar())) { + ir_read_error(expr, "non-zero write mask required."); + return NULL; + } + + return new(mem_ctx) ir_assignment(lhs, rhs, condition, mask); +} + +ir_call * +ir_reader::read_call(s_expression *expr) +{ + s_symbol *name; + s_list *params; + + s_pattern pat[] = { "call", name, params }; + if (!MATCH(expr, pat)) { + ir_read_error(expr, "expected (call ( ...))"); + return NULL; + } + + exec_list parameters; + + foreach_iter(exec_list_iterator, it, params->subexpressions) { + s_expression *expr = (s_expression*) it.get(); + ir_rvalue *param = read_rvalue(expr); + if (param == NULL) { + ir_read_error(expr, "when reading parameter to function call"); + return NULL; + } + parameters.push_tail(param); + } + + ir_function *f = state->symbols->get_function(name->value()); + if (f == NULL) { + ir_read_error(expr, "found call to undefined function %s", + name->value()); + return NULL; + } + + ir_function_signature *callee = f->matching_signature(¶meters); + if (callee == NULL) { + ir_read_error(expr, "couldn't find matching signature for function " + "%s", name->value()); + return NULL; + } + + return new(mem_ctx) ir_call(callee, ¶meters); +} + +ir_expression * +ir_reader::read_expression(s_expression *expr) +{ + s_expression *s_type; + s_symbol *s_op; + s_expression *s_arg1; + + s_pattern pat[] = { "expression", s_type, s_op, s_arg1 }; + if (!PARTIAL_MATCH(expr, pat)) { + ir_read_error(expr, "expected (expression " + " [])"); + return NULL; + } + s_expression *s_arg2 = (s_expression *) s_arg1->next; // may be tail sentinel + + const glsl_type *type = read_type(s_type); + if (type == NULL) + return NULL; + + /* Read the operator */ + ir_expression_operation op = ir_expression::get_operator(s_op->value()); + if (op == (ir_expression_operation) -1) { + ir_read_error(expr, "invalid operator: %s", s_op->value()); + return NULL; + } + + unsigned num_operands = ir_expression::get_num_operands(op); + if (num_operands == 1 && !s_arg1->next->is_tail_sentinel()) { + ir_read_error(expr, "expected (expression %s )", + s_op->value()); + return NULL; + } + + ir_rvalue *arg1 = read_rvalue(s_arg1); + ir_rvalue *arg2 = NULL; + if (arg1 == NULL) { + ir_read_error(NULL, "when reading first operand of %s", s_op->value()); + return NULL; + } + + if (num_operands == 2) { + if (s_arg2->is_tail_sentinel() || !s_arg2->next->is_tail_sentinel()) { + ir_read_error(expr, "expected (expression %s " + ")", s_op->value()); + return NULL; + } + arg2 = read_rvalue(s_arg2); + if (arg2 == NULL) { + ir_read_error(NULL, "when reading second operand of %s", + s_op->value()); + return NULL; + } + } + + return new(mem_ctx) ir_expression(op, type, arg1, arg2); +} + +ir_swizzle * +ir_reader::read_swizzle(s_expression *expr) +{ + s_symbol *swiz; + s_expression *sub; + + s_pattern pat[] = { "swiz", swiz, sub }; + if (!MATCH(expr, pat)) { + ir_read_error(expr, "expected (swiz )"); + return NULL; + } + + if (strlen(swiz->value()) > 4) { + ir_read_error(expr, "expected a valid swizzle; found %s", swiz->value()); + return NULL; + } + + ir_rvalue *rvalue = read_rvalue(sub); + if (rvalue == NULL) + return NULL; + + ir_swizzle *ir = ir_swizzle::create(rvalue, swiz->value(), + rvalue->type->vector_elements); + if (ir == NULL) + ir_read_error(expr, "invalid swizzle"); + + return ir; +} + +ir_constant * +ir_reader::read_constant(s_expression *expr) +{ + s_expression *type_expr; + s_list *values; + + s_pattern pat[] = { "constant", type_expr, values }; + if (!MATCH(expr, pat)) { + ir_read_error(expr, "expected (constant (...))"); + return NULL; + } + + const glsl_type *type = read_type(type_expr); + if (type == NULL) + return NULL; + + if (values == NULL) { + ir_read_error(expr, "expected (constant (...))"); + return NULL; + } + + if (type->is_array()) { + unsigned elements_supplied = 0; + exec_list elements; + foreach_iter(exec_list_iterator, it, values->subexpressions) { + s_expression *elt = (s_expression *) it.get(); + ir_constant *ir_elt = read_constant(elt); + if (ir_elt == NULL) + return NULL; + elements.push_tail(ir_elt); + elements_supplied++; + } + + if (elements_supplied != type->length) { + ir_read_error(values, "expected exactly %u array elements, " + "given %u", type->length, elements_supplied); + return NULL; + } + return new(mem_ctx) ir_constant(type, &elements); + } + + const glsl_type *const base_type = type->get_base_type(); + + ir_constant_data data = { { 0 } }; + + // Read in list of values (at most 16). + int k = 0; + foreach_iter(exec_list_iterator, it, values->subexpressions) { + if (k >= 16) { + ir_read_error(values, "expected at most 16 numbers"); + return NULL; + } + + s_expression *expr = (s_expression*) it.get(); + + if (base_type->base_type == GLSL_TYPE_FLOAT) { + s_number *value = SX_AS_NUMBER(expr); + if (value == NULL) { + ir_read_error(values, "expected numbers"); + return NULL; + } + data.f[k] = value->fvalue(); + } else { + s_int *value = SX_AS_INT(expr); + if (value == NULL) { + ir_read_error(values, "expected integers"); + return NULL; + } + + switch (base_type->base_type) { + case GLSL_TYPE_UINT: { + data.u[k] = value->value(); + break; + } + case GLSL_TYPE_INT: { + data.i[k] = value->value(); + break; + } + case GLSL_TYPE_BOOL: { + data.b[k] = value->value(); + break; + } + default: + ir_read_error(values, "unsupported constant type"); + return NULL; + } + } + ++k; + } + + return new(mem_ctx) ir_constant(type, &data); +} + +ir_dereference * +ir_reader::read_dereference(s_expression *expr) +{ + s_symbol *s_var; + s_expression *s_subject; + s_expression *s_index; + s_symbol *s_field; + + s_pattern var_pat[] = { "var_ref", s_var }; + s_pattern array_pat[] = { "array_ref", s_subject, s_index }; + s_pattern record_pat[] = { "record_ref", s_subject, s_field }; + + if (MATCH(expr, var_pat)) { + ir_variable *var = state->symbols->get_variable(s_var->value()); + if (var == NULL) { + ir_read_error(expr, "undeclared variable: %s", s_var->value()); + return NULL; + } + return new(mem_ctx) ir_dereference_variable(var); + } else if (MATCH(expr, array_pat)) { + ir_rvalue *subject = read_rvalue(s_subject); + if (subject == NULL) { + ir_read_error(NULL, "when reading the subject of an array_ref"); + return NULL; + } + + ir_rvalue *idx = read_rvalue(s_index); + if (subject == NULL) { + ir_read_error(NULL, "when reading the index of an array_ref"); + return NULL; + } + return new(mem_ctx) ir_dereference_array(subject, idx); + } else if (MATCH(expr, record_pat)) { + ir_rvalue *subject = read_rvalue(s_subject); + if (subject == NULL) { + ir_read_error(NULL, "when reading the subject of a record_ref"); + return NULL; + } + return new(mem_ctx) ir_dereference_record(subject, s_field->value()); + } + return NULL; +} + +ir_texture * +ir_reader::read_texture(s_expression *expr) +{ + s_symbol *tag = NULL; + s_expression *s_type = NULL; + s_expression *s_sampler = NULL; + s_expression *s_coord = NULL; + s_expression *s_offset = NULL; + s_expression *s_proj = NULL; + s_list *s_shadow = NULL; + s_expression *s_lod = NULL; + + ir_texture_opcode op = ir_tex; /* silence warning */ + + s_pattern tex_pattern[] = + { "tex", s_type, s_sampler, s_coord, s_offset, s_proj, s_shadow }; + s_pattern txf_pattern[] = + { "txf", s_type, s_sampler, s_coord, s_offset, s_lod }; + s_pattern other_pattern[] = + { tag, s_type, s_sampler, s_coord, s_offset, s_proj, s_shadow, s_lod }; + + if (MATCH(expr, tex_pattern)) { + op = ir_tex; + } else if (MATCH(expr, txf_pattern)) { + op = ir_txf; + } else if (MATCH(expr, other_pattern)) { + op = ir_texture::get_opcode(tag->value()); + if (op == -1) + return NULL; + } else { + ir_read_error(NULL, "unexpected texture pattern"); + return NULL; + } + + ir_texture *tex = new(mem_ctx) ir_texture(op); + + // Read return type + const glsl_type *type = read_type(s_type); + if (type == NULL) { + ir_read_error(NULL, "when reading type in (%s ...)", + tex->opcode_string()); + return NULL; + } + + // Read sampler (must be a deref) + ir_dereference *sampler = read_dereference(s_sampler); + if (sampler == NULL) { + ir_read_error(NULL, "when reading sampler in (%s ...)", + tex->opcode_string()); + return NULL; + } + tex->set_sampler(sampler, type); + + // Read coordinate (any rvalue) + tex->coordinate = read_rvalue(s_coord); + if (tex->coordinate == NULL) { + ir_read_error(NULL, "when reading coordinate in (%s ...)", + tex->opcode_string()); + return NULL; + } + + // Read texel offset - either 0 or an rvalue. + s_int *si_offset = SX_AS_INT(s_offset); + if (si_offset == NULL || si_offset->value() != 0) { + tex->offset = read_rvalue(s_offset); + if (tex->offset == NULL) { + ir_read_error(s_offset, "expected 0 or an expression"); + return NULL; + } + } + + if (op != ir_txf) { + s_int *proj_as_int = SX_AS_INT(s_proj); + if (proj_as_int && proj_as_int->value() == 1) { + tex->projector = NULL; + } else { + tex->projector = read_rvalue(s_proj); + if (tex->projector == NULL) { + ir_read_error(NULL, "when reading projective divide in (%s ..)", + tex->opcode_string()); + return NULL; + } + } + + if (s_shadow->subexpressions.is_empty()) { + tex->shadow_comparitor = NULL; + } else { + tex->shadow_comparitor = read_rvalue(s_shadow); + if (tex->shadow_comparitor == NULL) { + ir_read_error(NULL, "when reading shadow comparitor in (%s ..)", + tex->opcode_string()); + return NULL; + } + } + } + + switch (op) { + case ir_txb: + tex->lod_info.bias = read_rvalue(s_lod); + if (tex->lod_info.bias == NULL) { + ir_read_error(NULL, "when reading LOD bias in (txb ...)"); + return NULL; + } + break; + case ir_txl: + case ir_txf: + tex->lod_info.lod = read_rvalue(s_lod); + if (tex->lod_info.lod == NULL) { + ir_read_error(NULL, "when reading LOD in (%s ...)", + tex->opcode_string()); + return NULL; + } + break; + case ir_txd: { + s_expression *s_dx, *s_dy; + s_pattern dxdy_pat[] = { s_dx, s_dy }; + if (!MATCH(s_lod, dxdy_pat)) { + ir_read_error(s_lod, "expected (dPdx dPdy) in (txd ...)"); + return NULL; + } + tex->lod_info.grad.dPdx = read_rvalue(s_dx); + if (tex->lod_info.grad.dPdx == NULL) { + ir_read_error(NULL, "when reading dPdx in (txd ...)"); + return NULL; + } + tex->lod_info.grad.dPdy = read_rvalue(s_dy); + if (tex->lod_info.grad.dPdy == NULL) { + ir_read_error(NULL, "when reading dPdy in (txd ...)"); + return NULL; + } + break; + } + default: + // tex doesn't have any extra parameters. + break; + }; + return tex; +} diff --git a/mesalib/src/glsl/linker.cpp b/mesalib/src/glsl/linker.cpp index 265da84e5..34b64837a 100644 --- a/mesalib/src/glsl/linker.cpp +++ b/mesalib/src/glsl/linker.cpp @@ -1248,7 +1248,7 @@ assign_attribute_or_color_locations(gl_shader_program *prog, */ const int generic_base = (target_index == MESA_SHADER_VERTEX) - ? VERT_ATTRIB_GENERIC0 : FRAG_RESULT_DATA0; + ? (int) VERT_ATTRIB_GENERIC0 : (int) FRAG_RESULT_DATA0; const enum ir_variable_mode direction = (target_index == MESA_SHADER_VERTEX) ? ir_var_in : ir_var_out; diff --git a/mesalib/src/glsl/lower_jumps.cpp b/mesalib/src/glsl/lower_jumps.cpp index 17be39f5b..61874990a 100644 --- a/mesalib/src/glsl/lower_jumps.cpp +++ b/mesalib/src/glsl/lower_jumps.cpp @@ -1,570 +1,1010 @@ -/* - * Copyright © 2010 Luca Barbieri - * - * 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. - */ - -/** - * \file lower_jumps.cpp - * - * This pass lowers jumps (break, continue, and return) to if/else structures. - * - * It can be asked to: - * 1. Pull jumps out of ifs where possible - * 2. Remove all "continue"s, replacing them with an "execute flag" - * 3. Replace all "break" with a single conditional one at the end of the loop - * 4. Replace all "return"s with a single return at the end of the function, - * for the main function and/or other functions - * - * Applying this pass gives several benefits: - * 1. All functions can be inlined. - * 2. nv40 and other pre-DX10 chips without "continue" can be supported - * 3. nv30 and other pre-DX10 chips with no control flow at all are better - * supported - * - * Continues are lowered by adding a per-loop "execute flag", initialized to - * true, that when cleared inhibits all execution until the end of the loop. - * - * Breaks are lowered to continues, plus setting a "break flag" that is checked - * at the end of the loop, and trigger the unique "break". - * - * Returns are lowered to breaks/continues, plus adding a "return flag" that - * causes loops to break again out of their enclosing loops until all the - * loops are exited: then the "execute flag" logic will ignore everything - * until the end of the function. - * - * Note that "continue" and "return" can also be implemented by adding - * a dummy loop and using break. - * However, this is bad for hardware with limited nesting depth, and - * prevents further optimization, and thus is not currently performed. - */ - -#include "glsl_types.h" -#include -#include "ir.h" - -enum jump_strength -{ - strength_none, - strength_always_clears_execute_flag, - strength_continue, - strength_break, - strength_return -}; - -struct block_record -{ - /* minimum jump strength (of lowered IR, not pre-lowering IR) - * - * If the block ends with a jump, must be the strength of the jump. - * Otherwise, the jump would be dead and have been deleted before) - * - * If the block doesn't end with a jump, it can be different than strength_none if all paths before it lead to some jump - * (e.g. an if with a return in one branch, and a break in the other, while not lowering them) - * Note that identical jumps are usually unified though. - */ - jump_strength min_strength; - - /* can anything clear the execute flag? */ - bool may_clear_execute_flag; - - block_record() - { - this->min_strength = strength_none; - this->may_clear_execute_flag = false; - } -}; - -struct loop_record -{ - ir_function_signature* signature; - ir_loop* loop; - - /* used to avoid lowering the break used to represent lowered breaks */ - unsigned nesting_depth; - bool in_if_at_the_end_of_the_loop; - - bool may_set_return_flag; - - ir_variable* break_flag; - ir_variable* execute_flag; /* cleared to emulate continue */ - - loop_record(ir_function_signature* p_signature = 0, ir_loop* p_loop = 0) - { - this->signature = p_signature; - this->loop = p_loop; - this->nesting_depth = 0; - this->in_if_at_the_end_of_the_loop = false; - this->may_set_return_flag = false; - this->break_flag = 0; - this->execute_flag = 0; - } - - ir_variable* get_execute_flag() - { - /* also supported for the "function loop" */ - if(!this->execute_flag) { - exec_list& list = this->loop ? this->loop->body_instructions : signature->body; - this->execute_flag = new(this->signature) ir_variable(glsl_type::bool_type, "execute_flag", ir_var_temporary); - list.push_head(new(this->signature) ir_assignment(new(this->signature) ir_dereference_variable(execute_flag), new(this->signature) ir_constant(true), 0)); - list.push_head(this->execute_flag); - } - return this->execute_flag; - } - - ir_variable* get_break_flag() - { - assert(this->loop); - if(!this->break_flag) { - this->break_flag = new(this->signature) ir_variable(glsl_type::bool_type, "break_flag", ir_var_temporary); - this->loop->insert_before(this->break_flag); - this->loop->insert_before(new(this->signature) ir_assignment(new(this->signature) ir_dereference_variable(break_flag), new(this->signature) ir_constant(false), 0)); - } - return this->break_flag; - } -}; - -struct function_record -{ - ir_function_signature* signature; - ir_variable* return_flag; /* used to break out of all loops and then jump to the return instruction */ - ir_variable* return_value; - bool is_main; - unsigned nesting_depth; - - function_record(ir_function_signature* p_signature = 0) - { - this->signature = p_signature; - this->return_flag = 0; - this->return_value = 0; - this->nesting_depth = 0; - this->is_main = this->signature && (strcmp(this->signature->function_name(), "main") == 0); - } - - ir_variable* get_return_flag() - { - if(!this->return_flag) { - this->return_flag = new(this->signature) ir_variable(glsl_type::bool_type, "return_flag", ir_var_temporary); - this->signature->body.push_head(new(this->signature) ir_assignment(new(this->signature) ir_dereference_variable(return_flag), new(this->signature) ir_constant(false), 0)); - this->signature->body.push_head(this->return_flag); - } - return this->return_flag; - } - - ir_variable* get_return_value() - { - if(!this->return_value) { - assert(!this->signature->return_type->is_void()); - return_value = new(this->signature) ir_variable(this->signature->return_type, "return_value", ir_var_temporary); - this->signature->body.push_head(this->return_value); - } - return this->return_value; - } -}; - -struct ir_lower_jumps_visitor : public ir_control_flow_visitor { - bool progress; - - struct function_record function; - struct loop_record loop; - struct block_record block; - - bool pull_out_jumps; - bool lower_continue; - bool lower_break; - bool lower_sub_return; - bool lower_main_return; - - ir_lower_jumps_visitor() - { - this->progress = false; - } - - void truncate_after_instruction(exec_node *ir) - { - if (!ir) - return; - - while (!ir->get_next()->is_tail_sentinel()) { - ((ir_instruction *)ir->get_next())->remove(); - this->progress = true; - } - } - - void move_outer_block_inside(ir_instruction *ir, exec_list *inner_block) - { - while (!ir->get_next()->is_tail_sentinel()) { - ir_instruction *move_ir = (ir_instruction *)ir->get_next(); - - move_ir->remove(); - inner_block->push_tail(move_ir); - } - } - - virtual void visit(class ir_loop_jump * ir) - { - truncate_after_instruction(ir); - this->block.min_strength = ir->is_break() ? strength_break : strength_continue; - } - - virtual void visit(class ir_return * ir) - { - truncate_after_instruction(ir); - this->block.min_strength = strength_return; - } - - virtual void visit(class ir_discard * ir) - { - } - - enum jump_strength get_jump_strength(ir_instruction* ir) - { - if(!ir) - return strength_none; - else if(ir->ir_type == ir_type_loop_jump) { - if(((ir_loop_jump*)ir)->is_break()) - return strength_break; - else - return strength_continue; - } else if(ir->ir_type == ir_type_return) - return strength_return; - else - return strength_none; - } - - bool should_lower_jump(ir_jump* ir) - { - unsigned strength = get_jump_strength(ir); - bool lower; - switch(strength) - { - case strength_none: - lower = false; /* don't change this, code relies on it */ - break; - case strength_continue: - lower = lower_continue; - break; - case strength_break: - assert(this->loop.loop); - /* never lower "canonical break" */ - if(ir->get_next()->is_tail_sentinel() && (this->loop.nesting_depth == 0 - || (this->loop.nesting_depth == 1 && this->loop.in_if_at_the_end_of_the_loop))) - lower = false; - else - lower = lower_break; - break; - case strength_return: - /* never lower return at the end of a this->function */ - if(this->function.nesting_depth == 0 && ir->get_next()->is_tail_sentinel()) - lower = false; - else if (this->function.is_main) - lower = lower_main_return; - else - lower = lower_sub_return; - break; - } - return lower; - } - - block_record visit_block(exec_list* list) - { - block_record saved_block = this->block; - this->block = block_record(); - visit_exec_list(list, this); - block_record ret = this->block; - this->block = saved_block; - return ret; - } - - virtual void visit(ir_if *ir) - { - if(this->loop.nesting_depth == 0 && ir->get_next()->is_tail_sentinel()) - this->loop.in_if_at_the_end_of_the_loop = true; - - ++this->function.nesting_depth; - ++this->loop.nesting_depth; - - block_record block_records[2]; - ir_jump* jumps[2]; - - block_records[0] = visit_block(&ir->then_instructions); - block_records[1] = visit_block(&ir->else_instructions); - -retry: /* we get here if we put code after the if inside a branch */ - for(unsigned i = 0; i < 2; ++i) { - exec_list& list = i ? ir->else_instructions : ir->then_instructions; - jumps[i] = 0; - if(!list.is_empty() && get_jump_strength((ir_instruction*)list.get_tail())) - jumps[i] = (ir_jump*)list.get_tail(); - } - - for(;;) { - jump_strength jump_strengths[2]; - - for(unsigned i = 0; i < 2; ++i) { - if(jumps[i]) { - jump_strengths[i] = block_records[i].min_strength; - assert(jump_strengths[i] == get_jump_strength(jumps[i])); - } else - jump_strengths[i] = strength_none; - } - - /* move both jumps out if possible */ - if(pull_out_jumps && jump_strengths[0] == jump_strengths[1]) { - bool unify = true; - if(jump_strengths[0] == strength_continue) - ir->insert_after(new(ir) ir_loop_jump(ir_loop_jump::jump_continue)); - else if(jump_strengths[0] == strength_break) - ir->insert_after(new(ir) ir_loop_jump(ir_loop_jump::jump_break)); - /* FINISHME: unify returns with identical expressions */ - else if(jump_strengths[0] == strength_return && this->function.signature->return_type->is_void()) - ir->insert_after(new(ir) ir_return(NULL)); - else - unify = false; - - if(unify) { - jumps[0]->remove(); - jumps[1]->remove(); - this->progress = true; - - jumps[0] = 0; - jumps[1] = 0; - block_records[0].min_strength = strength_none; - block_records[1].min_strength = strength_none; - break; - } - } - - /* lower a jump: if both need to lowered, start with the strongest one, so that - * we might later unify the lowered version with the other one - */ - bool should_lower[2]; - for(unsigned i = 0; i < 2; ++i) - should_lower[i] = should_lower_jump(jumps[i]); - - int lower; - if(should_lower[1] && should_lower[0]) - lower = jump_strengths[1] > jump_strengths[0]; - else if(should_lower[0]) - lower = 0; - else if(should_lower[1]) - lower = 1; - else - break; - - if(jump_strengths[lower] == strength_return) { - ir_variable* return_flag = this->function.get_return_flag(); - if(!this->function.signature->return_type->is_void()) { - ir_variable* return_value = this->function.get_return_value(); - jumps[lower]->insert_before(new(ir) ir_assignment(new (ir) ir_dereference_variable(return_value), ((ir_return*)jumps[lower])->value, NULL)); - } - jumps[lower]->insert_before(new(ir) ir_assignment(new (ir) ir_dereference_variable(return_flag), new (ir) ir_constant(true), NULL)); - this->loop.may_set_return_flag = true; - if(this->loop.loop) { - ir_loop_jump* lowered = 0; - lowered = new(ir) ir_loop_jump(ir_loop_jump::jump_break); - block_records[lower].min_strength = strength_break; - jumps[lower]->replace_with(lowered); - jumps[lower] = lowered; - } else - goto lower_continue; - this->progress = true; - } else if(jump_strengths[lower] == strength_break) { - /* We can't lower to an actual continue because that would execute the increment. - * - * In the lowered code, we instead put the break check between the this->loop body and the increment, - * which is impossible with a real continue as defined by the GLSL IR currently. - * - * Smarter options (such as undoing the increment) are possible but it's not worth implementing them, - * because if break is lowered, continue is almost surely lowered too. - */ - jumps[lower]->insert_before(new(ir) ir_assignment(new (ir) ir_dereference_variable(this->loop.get_break_flag()), new (ir) ir_constant(true), 0)); - goto lower_continue; - } else if(jump_strengths[lower] == strength_continue) { -lower_continue: - ir_variable* execute_flag = this->loop.get_execute_flag(); - jumps[lower]->replace_with(new(ir) ir_assignment(new (ir) ir_dereference_variable(execute_flag), new (ir) ir_constant(false), 0)); - jumps[lower] = 0; - block_records[lower].min_strength = strength_always_clears_execute_flag; - block_records[lower].may_clear_execute_flag = true; - this->progress = true; - break; - } - } - - /* move out a jump out if possible */ - if(pull_out_jumps) { - int move_out = -1; - if(jumps[0] && block_records[1].min_strength >= strength_continue) - move_out = 0; - else if(jumps[1] && block_records[0].min_strength >= strength_continue) - move_out = 1; - - if(move_out >= 0) - { - jumps[move_out]->remove(); - ir->insert_after(jumps[move_out]); - jumps[move_out] = 0; - block_records[move_out].min_strength = strength_none; - this->progress = true; - } - } - - if(block_records[0].min_strength < block_records[1].min_strength) - this->block.min_strength = block_records[0].min_strength; - else - this->block.min_strength = block_records[1].min_strength; - this->block.may_clear_execute_flag = this->block.may_clear_execute_flag || block_records[0].may_clear_execute_flag || block_records[1].may_clear_execute_flag; - - if(this->block.min_strength) - truncate_after_instruction(ir); - else if(this->block.may_clear_execute_flag) - { - int move_into = -1; - if(block_records[0].min_strength && !block_records[1].may_clear_execute_flag) - move_into = 1; - else if(block_records[1].min_strength && !block_records[0].may_clear_execute_flag) - move_into = 0; - - if(move_into >= 0) { - assert(!block_records[move_into].min_strength && !block_records[move_into].may_clear_execute_flag); /* otherwise, we just truncated */ - - exec_list* list = move_into ? &ir->else_instructions : &ir->then_instructions; - exec_node* next = ir->get_next(); - if(!next->is_tail_sentinel()) { - move_outer_block_inside(ir, list); - - exec_list list; - list.head = next; - block_records[move_into] = visit_block(&list); - - this->progress = true; - goto retry; - } - } else { - ir_instruction* ir_after; - for(ir_after = (ir_instruction*)ir->get_next(); !ir_after->is_tail_sentinel();) - { - ir_if* ir_if = ir_after->as_if(); - if(ir_if && ir_if->else_instructions.is_empty()) { - ir_dereference_variable* ir_if_cond_deref = ir_if->condition->as_dereference_variable(); - if(ir_if_cond_deref && ir_if_cond_deref->var == this->loop.execute_flag) { - ir_instruction* ir_next = (ir_instruction*)ir_after->get_next(); - ir_after->insert_before(&ir_if->then_instructions); - ir_after->remove(); - ir_after = ir_next; - continue; - } - } - ir_after = (ir_instruction*)ir_after->get_next(); - - /* only set this if we find any unprotected instruction */ - this->progress = true; - } - - if(!ir->get_next()->is_tail_sentinel()) { - assert(this->loop.execute_flag); - ir_if* if_execute = new(ir) ir_if(new(ir) ir_dereference_variable(this->loop.execute_flag)); - move_outer_block_inside(ir, &if_execute->then_instructions); - ir->insert_after(if_execute); - } - } - } - --this->loop.nesting_depth; - --this->function.nesting_depth; - } - - virtual void visit(ir_loop *ir) - { - ++this->function.nesting_depth; - loop_record saved_loop = this->loop; - this->loop = loop_record(this->function.signature, ir); - - block_record body = visit_block(&ir->body_instructions); - - if(body.min_strength >= strength_break) { - /* FINISHME: turn the this->loop into an if, or replace it with its body */ - } - - if(this->loop.break_flag) { - ir_if* break_if = new(ir) ir_if(new(ir) ir_dereference_variable(this->loop.break_flag)); - break_if->then_instructions.push_tail(new(ir) ir_loop_jump(ir_loop_jump::jump_break)); - ir->body_instructions.push_tail(break_if); - } - - if(this->loop.may_set_return_flag) { - assert(this->function.return_flag); - ir_if* return_if = new(ir) ir_if(new(ir) ir_dereference_variable(this->function.return_flag)); - saved_loop.may_set_return_flag = true; - if(saved_loop.loop) - return_if->then_instructions.push_tail(new(ir) ir_loop_jump(ir_loop_jump::jump_break)); - else - move_outer_block_inside(ir, &return_if->else_instructions); - ir->insert_after(return_if); - } - - this->loop = saved_loop; - --this->function.nesting_depth; - } - - virtual void visit(ir_function_signature *ir) - { - /* these are not strictly necessary */ - assert(!this->function.signature); - assert(!this->loop.loop); - - function_record saved_function = this->function; - loop_record saved_loop = this->loop; - this->function = function_record(ir); - this->loop = loop_record(ir); - - assert(!this->loop.loop); - visit_block(&ir->body); - - if(this->function.return_value) - ir->body.push_tail(new(ir) ir_return(new (ir) ir_dereference_variable(this->function.return_value))); - - this->loop = saved_loop; - this->function = saved_function; - } - - virtual void visit(class ir_function * ir) - { - visit_block(&ir->signatures); - } -}; - -bool -do_lower_jumps(exec_list *instructions, bool pull_out_jumps, bool lower_sub_return, bool lower_main_return, bool lower_continue, bool lower_break) -{ - ir_lower_jumps_visitor v; - v.pull_out_jumps = pull_out_jumps; - v.lower_continue = lower_continue; - v.lower_break = lower_break; - v.lower_sub_return = lower_sub_return; - v.lower_main_return = lower_main_return; - - do { - v.progress = false; - visit_exec_list(instructions, &v); - } while (v.progress); - - return v.progress; -} +/* + * Copyright © 2010 Luca Barbieri + * + * 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. + */ + +/** + * \file lower_jumps.cpp + * + * This pass lowers jumps (break, continue, and return) to if/else structures. + * + * It can be asked to: + * 1. Pull jumps out of ifs where possible + * 2. Remove all "continue"s, replacing them with an "execute flag" + * 3. Replace all "break" with a single conditional one at the end of the loop + * 4. Replace all "return"s with a single return at the end of the function, + * for the main function and/or other functions + * + * Applying this pass gives several benefits: + * 1. All functions can be inlined. + * 2. nv40 and other pre-DX10 chips without "continue" can be supported + * 3. nv30 and other pre-DX10 chips with no control flow at all are better + * supported + * + * Continues are lowered by adding a per-loop "execute flag", initialized to + * true, that when cleared inhibits all execution until the end of the loop. + * + * Breaks are lowered to continues, plus setting a "break flag" that is checked + * at the end of the loop, and trigger the unique "break". + * + * Returns are lowered to breaks/continues, plus adding a "return flag" that + * causes loops to break again out of their enclosing loops until all the + * loops are exited: then the "execute flag" logic will ignore everything + * until the end of the function. + * + * Note that "continue" and "return" can also be implemented by adding + * a dummy loop and using break. + * However, this is bad for hardware with limited nesting depth, and + * prevents further optimization, and thus is not currently performed. + */ + +#include "glsl_types.h" +#include +#include "ir.h" + +/** + * Enum recording the result of analyzing how control flow might exit + * an IR node. + * + * Each possible value of jump_strength indicates a strictly stronger + * guarantee on control flow than the previous value. + * + * The ordering of strengths roughly reflects the way jumps are + * lowered: jumps with higher strength tend to be lowered to jumps of + * lower strength. Accordingly, strength is used as a heuristic to + * determine which lowering to perform first. + * + * This enum is also used by get_jump_strength() to categorize + * instructions as either break, continue, return, or other. When + * used in this fashion, strength_always_clears_execute_flag is not + * used. + * + * The control flow analysis made by this optimization pass makes two + * simplifying assumptions: + * + * - It ignores discard instructions, since they are lowered by a + * separate pass (lower_discard.cpp). + * + * - It assumes it is always possible for control to flow from a loop + * to the instruction immediately following it. Technically, this + * is not true (since all execution paths through the loop might + * jump back to the top, or return from the function). + * + * Both of these simplifying assumtions are safe, since they can never + * cause reachable code to be incorrectly classified as unreachable; + * they can only do the opposite. + */ +enum jump_strength +{ + /** + * Analysis has produced no guarantee on how control flow might + * exit this IR node. It might fall out the bottom (with or + * without clearing the execute flag, if present), or it might + * continue to the top of the innermost enclosing loop, break out + * of it, or return from the function. + */ + strength_none, + + /** + * The only way control can fall out the bottom of this node is + * through a code path that clears the execute flag. It might also + * continue to the top of the innermost enclosing loop, break out + * of it, or return from the function. + */ + strength_always_clears_execute_flag, + + /** + * Control cannot fall out the bottom of this node. It might + * continue to the top of the innermost enclosing loop, break out + * of it, or return from the function. + */ + strength_continue, + + /** + * Control cannot fall out the bottom of this node, or continue the + * top of the innermost enclosing loop. It can only break out of + * it or return from the function. + */ + strength_break, + + /** + * Control cannot fall out the bottom of this node, continue to the + * top of the innermost enclosing loop, or break out of it. It can + * only return from the function. + */ + strength_return +}; + +struct block_record +{ + /* minimum jump strength (of lowered IR, not pre-lowering IR) + * + * If the block ends with a jump, must be the strength of the jump. + * Otherwise, the jump would be dead and have been deleted before) + * + * If the block doesn't end with a jump, it can be different than strength_none if all paths before it lead to some jump + * (e.g. an if with a return in one branch, and a break in the other, while not lowering them) + * Note that identical jumps are usually unified though. + */ + jump_strength min_strength; + + /* can anything clear the execute flag? */ + bool may_clear_execute_flag; + + block_record() + { + this->min_strength = strength_none; + this->may_clear_execute_flag = false; + } +}; + +struct loop_record +{ + ir_function_signature* signature; + ir_loop* loop; + + /* used to avoid lowering the break used to represent lowered breaks */ + unsigned nesting_depth; + bool in_if_at_the_end_of_the_loop; + + bool may_set_return_flag; + + ir_variable* break_flag; + ir_variable* execute_flag; /* cleared to emulate continue */ + + loop_record(ir_function_signature* p_signature = 0, ir_loop* p_loop = 0) + { + this->signature = p_signature; + this->loop = p_loop; + this->nesting_depth = 0; + this->in_if_at_the_end_of_the_loop = false; + this->may_set_return_flag = false; + this->break_flag = 0; + this->execute_flag = 0; + } + + ir_variable* get_execute_flag() + { + /* also supported for the "function loop" */ + if(!this->execute_flag) { + exec_list& list = this->loop ? this->loop->body_instructions : signature->body; + this->execute_flag = new(this->signature) ir_variable(glsl_type::bool_type, "execute_flag", ir_var_temporary); + list.push_head(new(this->signature) ir_assignment(new(this->signature) ir_dereference_variable(execute_flag), new(this->signature) ir_constant(true), 0)); + list.push_head(this->execute_flag); + } + return this->execute_flag; + } + + ir_variable* get_break_flag() + { + assert(this->loop); + if(!this->break_flag) { + this->break_flag = new(this->signature) ir_variable(glsl_type::bool_type, "break_flag", ir_var_temporary); + this->loop->insert_before(this->break_flag); + this->loop->insert_before(new(this->signature) ir_assignment(new(this->signature) ir_dereference_variable(break_flag), new(this->signature) ir_constant(false), 0)); + } + return this->break_flag; + } +}; + +struct function_record +{ + ir_function_signature* signature; + ir_variable* return_flag; /* used to break out of all loops and then jump to the return instruction */ + ir_variable* return_value; + bool lower_return; + unsigned nesting_depth; + + function_record(ir_function_signature* p_signature = 0, + bool lower_return = false) + { + this->signature = p_signature; + this->return_flag = 0; + this->return_value = 0; + this->nesting_depth = 0; + this->lower_return = lower_return; + } + + ir_variable* get_return_flag() + { + if(!this->return_flag) { + this->return_flag = new(this->signature) ir_variable(glsl_type::bool_type, "return_flag", ir_var_temporary); + this->signature->body.push_head(new(this->signature) ir_assignment(new(this->signature) ir_dereference_variable(return_flag), new(this->signature) ir_constant(false), 0)); + this->signature->body.push_head(this->return_flag); + } + return this->return_flag; + } + + ir_variable* get_return_value() + { + if(!this->return_value) { + assert(!this->signature->return_type->is_void()); + return_value = new(this->signature) ir_variable(this->signature->return_type, "return_value", ir_var_temporary); + this->signature->body.push_head(this->return_value); + } + return this->return_value; + } +}; + +struct ir_lower_jumps_visitor : public ir_control_flow_visitor { + /* Postconditions: on exit of any visit() function: + * + * ANALYSIS: this->block.min_strength, + * this->block.may_clear_execute_flag, and + * this->loop.may_set_return_flag are updated to reflect the + * characteristics of the visited statement. + * + * DEAD_CODE_ELIMINATION: If this->block.min_strength is not + * strength_none, the visited node is at the end of its exec_list. + * In other words, any unreachable statements that follow the + * visited statement in its exec_list have been removed. + * + * CONTAINED_JUMPS_LOWERED: If the visited statement contains other + * statements, then should_lower_jump() is false for all of the + * return, break, or continue statements it contains. + * + * Note that visiting a jump does not lower it. That is the + * responsibility of the statement (or function signature) that + * contains the jump. + */ + + bool progress; + + struct function_record function; + struct loop_record loop; + struct block_record block; + + bool pull_out_jumps; + bool lower_continue; + bool lower_break; + bool lower_sub_return; + bool lower_main_return; + + ir_lower_jumps_visitor() + { + this->progress = false; + } + + void truncate_after_instruction(exec_node *ir) + { + if (!ir) + return; + + while (!ir->get_next()->is_tail_sentinel()) { + ((ir_instruction *)ir->get_next())->remove(); + this->progress = true; + } + } + + void move_outer_block_inside(ir_instruction *ir, exec_list *inner_block) + { + while (!ir->get_next()->is_tail_sentinel()) { + ir_instruction *move_ir = (ir_instruction *)ir->get_next(); + + move_ir->remove(); + inner_block->push_tail(move_ir); + } + } + + /** + * Insert the instructions necessary to lower a return statement, + * before the given return instruction. + */ + void insert_lowered_return(ir_return *ir) + { + ir_variable* return_flag = this->function.get_return_flag(); + if(!this->function.signature->return_type->is_void()) { + ir_variable* return_value = this->function.get_return_value(); + ir->insert_before( + new(ir) ir_assignment( + new (ir) ir_dereference_variable(return_value), + ir->value)); + } + ir->insert_before( + new(ir) ir_assignment( + new (ir) ir_dereference_variable(return_flag), + new (ir) ir_constant(true))); + this->loop.may_set_return_flag = true; + } + + /** + * If the given instruction is a return, lower it to instructions + * that store the return value (if there is one), set the return + * flag, and then break. + * + * It is safe to pass NULL to this function. + */ + void lower_return_unconditionally(ir_instruction *ir) + { + if (get_jump_strength(ir) != strength_return) { + return; + } + insert_lowered_return((ir_return*)ir); + ir->replace_with(new(ir) ir_loop_jump(ir_loop_jump::jump_break)); + } + + /** + * Create the necessary instruction to replace a break instruction. + */ + ir_instruction *create_lowered_break() + { + void *ctx = this->function.signature; + return new(ctx) ir_assignment( + new(ctx) ir_dereference_variable(this->loop.get_break_flag()), + new(ctx) ir_constant(true), + 0); + } + + /** + * If the given instruction is a break, lower it to an instruction + * that sets the break flag, without consulting + * should_lower_jump(). + * + * It is safe to pass NULL to this function. + */ + void lower_break_unconditionally(ir_instruction *ir) + { + if (get_jump_strength(ir) != strength_break) { + return; + } + ir->replace_with(create_lowered_break()); + } + + /** + * If the block ends in a conditional or unconditional break, lower + * it, even though should_lower_jump() says it needn't be lowered. + */ + void lower_final_breaks(exec_list *block) + { + ir_instruction *ir = (ir_instruction *) block->get_tail(); + lower_break_unconditionally(ir); + ir_if *ir_if = ir->as_if(); + if (ir_if) { + lower_break_unconditionally( + (ir_instruction *) ir_if->then_instructions.get_tail()); + lower_break_unconditionally( + (ir_instruction *) ir_if->else_instructions.get_tail()); + } + } + + virtual void visit(class ir_loop_jump * ir) + { + /* Eliminate all instructions after each one, since they are + * unreachable. This satisfies the DEAD_CODE_ELIMINATION + * postcondition. + */ + truncate_after_instruction(ir); + + /* Set this->block.min_strength based on this instruction. This + * satisfies the ANALYSIS postcondition. It is not necessary to + * update this->block.may_clear_execute_flag or + * this->loop.may_set_return_flag, because an unlowered jump + * instruction can't change any flags. + */ + this->block.min_strength = ir->is_break() ? strength_break : strength_continue; + + /* The CONTAINED_JUMPS_LOWERED postcondition is already + * satisfied, because jump statements can't contain other + * statements. + */ + } + + virtual void visit(class ir_return * ir) + { + /* Eliminate all instructions after each one, since they are + * unreachable. This satisfies the DEAD_CODE_ELIMINATION + * postcondition. + */ + truncate_after_instruction(ir); + + /* Set this->block.min_strength based on this instruction. This + * satisfies the ANALYSIS postcondition. It is not necessary to + * update this->block.may_clear_execute_flag or + * this->loop.may_set_return_flag, because an unlowered return + * instruction can't change any flags. + */ + this->block.min_strength = strength_return; + + /* The CONTAINED_JUMPS_LOWERED postcondition is already + * satisfied, because jump statements can't contain other + * statements. + */ + } + + virtual void visit(class ir_discard * ir) + { + /* Nothing needs to be done. The ANALYSIS and + * DEAD_CODE_ELIMINATION postconditions are already satisfied, + * because discard statements are ignored by this optimization + * pass. The CONTAINED_JUMPS_LOWERED postcondition is already + * satisfied, because discard statements can't contain other + * statements. + */ + } + + enum jump_strength get_jump_strength(ir_instruction* ir) + { + if(!ir) + return strength_none; + else if(ir->ir_type == ir_type_loop_jump) { + if(((ir_loop_jump*)ir)->is_break()) + return strength_break; + else + return strength_continue; + } else if(ir->ir_type == ir_type_return) + return strength_return; + else + return strength_none; + } + + bool should_lower_jump(ir_jump* ir) + { + unsigned strength = get_jump_strength(ir); + bool lower; + switch(strength) + { + case strength_none: + lower = false; /* don't change this, code relies on it */ + break; + case strength_continue: + lower = lower_continue; + break; + case strength_break: + assert(this->loop.loop); + /* never lower "canonical break" */ + if(ir->get_next()->is_tail_sentinel() && (this->loop.nesting_depth == 0 + || (this->loop.nesting_depth == 1 && this->loop.in_if_at_the_end_of_the_loop))) + lower = false; + else + lower = lower_break; + break; + case strength_return: + /* never lower return at the end of a this->function */ + if(this->function.nesting_depth == 0 && ir->get_next()->is_tail_sentinel()) + lower = false; + else + lower = this->function.lower_return; + break; + } + return lower; + } + + block_record visit_block(exec_list* list) + { + /* Note: since visiting a node may change that node's next + * pointer, we can't use visit_exec_list(), because + * visit_exec_list() caches the node's next pointer before + * visiting it. So we use foreach_list() instead. + * + * foreach_list() isn't safe if the node being visited gets + * removed, but fortunately this visitor doesn't do that. + */ + + block_record saved_block = this->block; + this->block = block_record(); + foreach_list(node, list) { + ((ir_instruction *) node)->accept(this); + } + block_record ret = this->block; + this->block = saved_block; + return ret; + } + + virtual void visit(ir_if *ir) + { + if(this->loop.nesting_depth == 0 && ir->get_next()->is_tail_sentinel()) + this->loop.in_if_at_the_end_of_the_loop = true; + + ++this->function.nesting_depth; + ++this->loop.nesting_depth; + + block_record block_records[2]; + ir_jump* jumps[2]; + + /* Recursively lower nested jumps. This satisfies the + * CONTAINED_JUMPS_LOWERED postcondition, except in the case of + * unconditional jumps at the end of ir->then_instructions and + * ir->else_instructions, which are handled below. + */ + block_records[0] = visit_block(&ir->then_instructions); + block_records[1] = visit_block(&ir->else_instructions); + +retry: /* we get here if we put code after the if inside a branch */ + + /* Determine which of ir->then_instructions and + * ir->else_instructions end with an unconditional jump. + */ + for(unsigned i = 0; i < 2; ++i) { + exec_list& list = i ? ir->else_instructions : ir->then_instructions; + jumps[i] = 0; + if(!list.is_empty() && get_jump_strength((ir_instruction*)list.get_tail())) + jumps[i] = (ir_jump*)list.get_tail(); + } + + /* Loop until we have satisfied the CONTAINED_JUMPS_LOWERED + * postcondition by lowering jumps in both then_instructions and + * else_instructions. + */ + for(;;) { + /* Determine the types of the jumps that terminate + * ir->then_instructions and ir->else_instructions. + */ + jump_strength jump_strengths[2]; + + for(unsigned i = 0; i < 2; ++i) { + if(jumps[i]) { + jump_strengths[i] = block_records[i].min_strength; + assert(jump_strengths[i] == get_jump_strength(jumps[i])); + } else + jump_strengths[i] = strength_none; + } + + /* If both code paths end in a jump, and the jumps are the + * same, and we are pulling out jumps, replace them with a + * single jump that comes after the if instruction. The new + * jump will be visited next, and it will be lowered if + * necessary by the loop or conditional that encloses it. + */ + if(pull_out_jumps && jump_strengths[0] == jump_strengths[1]) { + bool unify = true; + if(jump_strengths[0] == strength_continue) + ir->insert_after(new(ir) ir_loop_jump(ir_loop_jump::jump_continue)); + else if(jump_strengths[0] == strength_break) + ir->insert_after(new(ir) ir_loop_jump(ir_loop_jump::jump_break)); + /* FINISHME: unify returns with identical expressions */ + else if(jump_strengths[0] == strength_return && this->function.signature->return_type->is_void()) + ir->insert_after(new(ir) ir_return(NULL)); + else + unify = false; + + if(unify) { + jumps[0]->remove(); + jumps[1]->remove(); + this->progress = true; + + /* Update jumps[] to reflect the fact that the jumps + * are gone, and update block_records[] to reflect the + * fact that control can now flow to the next + * instruction. + */ + jumps[0] = 0; + jumps[1] = 0; + block_records[0].min_strength = strength_none; + block_records[1].min_strength = strength_none; + + /* The CONTAINED_JUMPS_LOWERED postcondition is now + * satisfied, so we can break out of the loop. + */ + break; + } + } + + /* lower a jump: if both need to lowered, start with the strongest one, so that + * we might later unify the lowered version with the other one + */ + bool should_lower[2]; + for(unsigned i = 0; i < 2; ++i) + should_lower[i] = should_lower_jump(jumps[i]); + + int lower; + if(should_lower[1] && should_lower[0]) + lower = jump_strengths[1] > jump_strengths[0]; + else if(should_lower[0]) + lower = 0; + else if(should_lower[1]) + lower = 1; + else + /* Neither code path ends in a jump that needs to be + * lowered, so the CONTAINED_JUMPS_LOWERED postcondition + * is satisfied and we can break out of the loop. + */ + break; + + if(jump_strengths[lower] == strength_return) { + /* To lower a return, we create a return flag (if the + * function doesn't have one already) and add instructions + * that: 1. store the return value (if this function has a + * non-void return) and 2. set the return flag + */ + insert_lowered_return((ir_return*)jumps[lower]); + if(this->loop.loop) { + /* If we are in a loop, replace the return instruction + * with a break instruction, and then loop so that the + * break instruction can be lowered if necessary. + */ + ir_loop_jump* lowered = 0; + lowered = new(ir) ir_loop_jump(ir_loop_jump::jump_break); + /* Note: we must update block_records and jumps to + * reflect the fact that the control path has been + * altered from a return to a break. + */ + block_records[lower].min_strength = strength_break; + jumps[lower]->replace_with(lowered); + jumps[lower] = lowered; + } else { + /* If we are not in a loop, we then proceed as we would + * for a continue statement (set the execute flag to + * false to prevent the rest of the function from + * executing). + */ + goto lower_continue; + } + this->progress = true; + } else if(jump_strengths[lower] == strength_break) { + /* To lower a break, we create a break flag (if the loop + * doesn't have one already) and add an instruction that + * sets it. + * + * Then we proceed as we would for a continue statement + * (set the execute flag to false to prevent the rest of + * the loop body from executing). + * + * The visit() function for the loop will ensure that the + * break flag is checked after executing the loop body. + */ + jumps[lower]->insert_before(create_lowered_break()); + goto lower_continue; + } else if(jump_strengths[lower] == strength_continue) { +lower_continue: + /* To lower a continue, we create an execute flag (if the + * loop doesn't have one already) and replace the continue + * with an instruction that clears it. + * + * Note that this code path gets exercised when lowering + * return statements that are not inside a loop, so + * this->loop must be initialized even outside of loops. + */ + ir_variable* execute_flag = this->loop.get_execute_flag(); + jumps[lower]->replace_with(new(ir) ir_assignment(new (ir) ir_dereference_variable(execute_flag), new (ir) ir_constant(false), 0)); + /* Note: we must update block_records and jumps to reflect + * the fact that the control path has been altered to an + * instruction that clears the execute flag. + */ + jumps[lower] = 0; + block_records[lower].min_strength = strength_always_clears_execute_flag; + block_records[lower].may_clear_execute_flag = true; + this->progress = true; + + /* Let the loop run again, in case the other branch of the + * if needs to be lowered too. + */ + } + } + + /* move out a jump out if possible */ + if(pull_out_jumps) { + /* If one of the branches ends in a jump, and control cannot + * fall out the bottom of the other branch, then we can move + * the jump after the if. + * + * Set move_out to the branch we are moving a jump out of. + */ + int move_out = -1; + if(jumps[0] && block_records[1].min_strength >= strength_continue) + move_out = 0; + else if(jumps[1] && block_records[0].min_strength >= strength_continue) + move_out = 1; + + if(move_out >= 0) + { + jumps[move_out]->remove(); + ir->insert_after(jumps[move_out]); + /* Note: we must update block_records and jumps to reflect + * the fact that the jump has been moved out of the if. + */ + jumps[move_out] = 0; + block_records[move_out].min_strength = strength_none; + this->progress = true; + } + } + + /* Now satisfy the ANALYSIS postcondition by setting + * this->block.min_strength and + * this->block.may_clear_execute_flag based on the + * characteristics of the two branches. + */ + if(block_records[0].min_strength < block_records[1].min_strength) + this->block.min_strength = block_records[0].min_strength; + else + this->block.min_strength = block_records[1].min_strength; + this->block.may_clear_execute_flag = this->block.may_clear_execute_flag || block_records[0].may_clear_execute_flag || block_records[1].may_clear_execute_flag; + + /* Now we need to clean up the instructions that follow the + * if. + * + * If those instructions are unreachable, then satisfy the + * DEAD_CODE_ELIMINATION postcondition by eliminating them. + * Otherwise that postcondition is already satisfied. + */ + if(this->block.min_strength) + truncate_after_instruction(ir); + else if(this->block.may_clear_execute_flag) + { + /* If the "if" instruction might clear the execute flag, then + * we need to guard any instructions that follow so that they + * are only executed if the execute flag is set. + * + * If one of the branches of the "if" always clears the + * execute flag, and the other branch never clears it, then + * this is easy: just move all the instructions following the + * "if" into the branch that never clears it. + */ + int move_into = -1; + if(block_records[0].min_strength && !block_records[1].may_clear_execute_flag) + move_into = 1; + else if(block_records[1].min_strength && !block_records[0].may_clear_execute_flag) + move_into = 0; + + if(move_into >= 0) { + assert(!block_records[move_into].min_strength && !block_records[move_into].may_clear_execute_flag); /* otherwise, we just truncated */ + + exec_list* list = move_into ? &ir->else_instructions : &ir->then_instructions; + exec_node* next = ir->get_next(); + if(!next->is_tail_sentinel()) { + move_outer_block_inside(ir, list); + + /* If any instructions moved, then we need to visit + * them (since they are now inside the "if"). Since + * block_records[move_into] is in its default state + * (see assertion above), we can safely replace + * block_records[move_into] with the result of this + * analysis. + */ + exec_list list; + list.head = next; + block_records[move_into] = visit_block(&list); + + /* + * Then we need to re-start our jump lowering, since one + * of the instructions we moved might be a jump that + * needs to be lowered. + */ + this->progress = true; + goto retry; + } + } else { + /* If we get here, then the simple case didn't apply; we + * need to actually guard the instructions that follow. + * + * To avoid creating unnecessarily-deep nesting, first + * look through the instructions that follow and unwrap + * any instructions that that are already wrapped in the + * appropriate guard. + */ + ir_instruction* ir_after; + for(ir_after = (ir_instruction*)ir->get_next(); !ir_after->is_tail_sentinel();) + { + ir_if* ir_if = ir_after->as_if(); + if(ir_if && ir_if->else_instructions.is_empty()) { + ir_dereference_variable* ir_if_cond_deref = ir_if->condition->as_dereference_variable(); + if(ir_if_cond_deref && ir_if_cond_deref->var == this->loop.execute_flag) { + ir_instruction* ir_next = (ir_instruction*)ir_after->get_next(); + ir_after->insert_before(&ir_if->then_instructions); + ir_after->remove(); + ir_after = ir_next; + continue; + } + } + ir_after = (ir_instruction*)ir_after->get_next(); + + /* only set this if we find any unprotected instruction */ + this->progress = true; + } + + /* Then, wrap all the instructions that follow in a single + * guard. + */ + if(!ir->get_next()->is_tail_sentinel()) { + assert(this->loop.execute_flag); + ir_if* if_execute = new(ir) ir_if(new(ir) ir_dereference_variable(this->loop.execute_flag)); + move_outer_block_inside(ir, &if_execute->then_instructions); + ir->insert_after(if_execute); + } + } + } + --this->loop.nesting_depth; + --this->function.nesting_depth; + } + + virtual void visit(ir_loop *ir) + { + /* Visit the body of the loop, with a fresh data structure in + * this->loop so that the analysis we do here won't bleed into + * enclosing loops. + * + * We assume that all code after a loop is reachable from the + * loop (see comments on enum jump_strength), so the + * DEAD_CODE_ELIMINATION postcondition is automatically + * satisfied, as is the block.min_strength portion of the + * ANALYSIS postcondition. + * + * The block.may_clear_execute_flag portion of the ANALYSIS + * postcondition is automatically satisfied because execute + * flags do not propagate outside of loops. + * + * The loop.may_set_return_flag portion of the ANALYSIS + * postcondition is handled below. + */ + ++this->function.nesting_depth; + loop_record saved_loop = this->loop; + this->loop = loop_record(this->function.signature, ir); + + /* Recursively lower nested jumps. This satisfies the + * CONTAINED_JUMPS_LOWERED postcondition, except in the case of + * an unconditional continue or return at the bottom of the + * loop, which are handled below. + */ + block_record body = visit_block(&ir->body_instructions); + + /* If the loop ends in an unconditional continue, eliminate it + * because it is redundant. + */ + ir_instruction *ir_last + = (ir_instruction *) ir->body_instructions.get_tail(); + if (get_jump_strength(ir_last) == strength_continue) { + ir_last->remove(); + } + + /* If the loop ends in an unconditional return, and we are + * lowering returns, lower it. + */ + if (this->function.lower_return) + lower_return_unconditionally(ir_last); + + if(body.min_strength >= strength_break) { + /* FINISHME: If the min_strength of the loop body is + * strength_break or strength_return, that means that it + * isn't a loop at all, since control flow always leaves the + * body of the loop via break or return. In principle the + * loop could be eliminated in this case. This optimization + * is not implemented yet. + */ + } + + if(this->loop.break_flag) { + /* We only get here if we are lowering breaks */ + assert (lower_break); + + /* If a break flag was generated while visiting the body of + * the loop, then at least one break was lowered, so we need + * to generate an if statement at the end of the loop that + * does a "break" if the break flag is set. The break we + * generate won't violate the CONTAINED_JUMPS_LOWERED + * postcondition, because should_lower_jump() always returns + * false for a break that happens at the end of a loop. + * + * However, if the loop already ends in a conditional or + * unconditional break, then we need to lower that break, + * because it won't be at the end of the loop anymore. + */ + lower_final_breaks(&ir->body_instructions); + + ir_if* break_if = new(ir) ir_if(new(ir) ir_dereference_variable(this->loop.break_flag)); + break_if->then_instructions.push_tail(new(ir) ir_loop_jump(ir_loop_jump::jump_break)); + ir->body_instructions.push_tail(break_if); + } + + /* If the body of the loop may set the return flag, then at + * least one return was lowered to a break, so we need to ensure + * that the return flag is checked after the body of the loop is + * executed. + */ + if(this->loop.may_set_return_flag) { + assert(this->function.return_flag); + /* Generate the if statement to check the return flag */ + ir_if* return_if = new(ir) ir_if(new(ir) ir_dereference_variable(this->function.return_flag)); + /* Note: we also need to propagate the knowledge that the + * return flag may get set to the outer context. This + * satisfies the loop.may_set_return_flag part of the + * ANALYSIS postcondition. + */ + saved_loop.may_set_return_flag = true; + if(saved_loop.loop) + /* If this loop is nested inside another one, then the if + * statement that we generated should break out of that + * loop if the return flag is set. Caller will lower that + * break statement if necessary. + */ + return_if->then_instructions.push_tail(new(ir) ir_loop_jump(ir_loop_jump::jump_break)); + else + /* Otherwise, all we need to do is ensure that the + * instructions that follow are only executed if the + * return flag is clear. We can do that by moving those + * instructions into the else clause of the generated if + * statement. + */ + move_outer_block_inside(ir, &return_if->else_instructions); + ir->insert_after(return_if); + } + + this->loop = saved_loop; + --this->function.nesting_depth; + } + + virtual void visit(ir_function_signature *ir) + { + /* these are not strictly necessary */ + assert(!this->function.signature); + assert(!this->loop.loop); + + bool lower_return; + if (strcmp(ir->function_name(), "main") == 0) + lower_return = lower_main_return; + else + lower_return = lower_sub_return; + + function_record saved_function = this->function; + loop_record saved_loop = this->loop; + this->function = function_record(ir, lower_return); + this->loop = loop_record(ir); + + assert(!this->loop.loop); + + /* Visit the body of the function to lower any jumps that occur + * in it, except possibly an unconditional return statement at + * the end of it. + */ + visit_block(&ir->body); + + /* If the body ended in an unconditional return of non-void, + * then we don't need to lower it because it's the one canonical + * return. + * + * If the body ended in a return of void, eliminate it because + * it is redundant. + */ + if (ir->return_type->is_void() && + get_jump_strength((ir_instruction *) ir->body.get_tail())) { + ir_jump *jump = (ir_jump *) ir->body.get_tail(); + assert (jump->ir_type == ir_type_return); + jump->remove(); + } + + if(this->function.return_value) + ir->body.push_tail(new(ir) ir_return(new (ir) ir_dereference_variable(this->function.return_value))); + + this->loop = saved_loop; + this->function = saved_function; + } + + virtual void visit(class ir_function * ir) + { + visit_block(&ir->signatures); + } +}; + +bool +do_lower_jumps(exec_list *instructions, bool pull_out_jumps, bool lower_sub_return, bool lower_main_return, bool lower_continue, bool lower_break) +{ + ir_lower_jumps_visitor v; + v.pull_out_jumps = pull_out_jumps; + v.lower_continue = lower_continue; + v.lower_break = lower_break; + v.lower_sub_return = lower_sub_return; + v.lower_main_return = lower_main_return; + + do { + v.progress = false; + visit_exec_list(instructions, &v); + } while (v.progress); + + return v.progress; +} -- cgit v1.2.3