| /* | |
| * Copyright © 2014 Intel Corporation | |
| * | |
| * Permission is hereby granted, free of charge, to any person obtaining a | |
| * copy of this software and associated documentation files (the "Software"), | |
| * to deal in the Software without restriction, including without limitation | |
| * the rights to use, copy, modify, merge, publish, distribute, sublicense, | |
| * and/or sell copies of the Software, and to permit persons to whom the | |
| * Software is furnished to do so, subject to the following conditions: | |
| * | |
| * The above copyright notice and this permission notice (including the next | |
| * paragraph) shall be included in all copies or substantial portions of the | |
| * Software. | |
| * | |
| * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR | |
| * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, | |
| * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL | |
| * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER | |
| * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING | |
| * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS | |
| * IN THE SOFTWARE. | |
| * | |
| * Authors: | |
| * Connor Abbott (cwabbott0@gmail.com) | |
| * | |
| */ | |
| #include <assert.h> | |
| #include "c11/threads.h" | |
| #include "util/hash_table.h" | |
| #include "util/simple_mtx.h" | |
| #include "nir.h" | |
| #include "nir_xfb_info.h" | |
| /* | |
| * This file checks for invalid IR indicating a bug somewhere in the compiler. | |
| */ | |
| /* Since this file is just a pile of asserts, don't bother compiling it if | |
| * we're not building a debug build. | |
| */ | |
| #ifndef NDEBUG | |
| typedef struct { | |
| void *mem_ctx; | |
| /* the current shader being validated */ | |
| nir_shader *shader; | |
| /* the current instruction being validated */ | |
| nir_instr *instr; | |
| /* the current variable being validated */ | |
| nir_variable *var; | |
| /* the current basic block being validated */ | |
| nir_block *block; | |
| /* the current if statement being validated */ | |
| nir_if *if_stmt; | |
| /* the current loop being visited */ | |
| nir_loop *loop; | |
| /* weather the loop continue construct is being visited */ | |
| bool in_loop_continue_construct; | |
| /* the parent of the current cf node being visited */ | |
| nir_cf_node *parent_node; | |
| /* the current function implementation being validated */ | |
| nir_function_impl *impl; | |
| /* Set of all blocks in the list */ | |
| struct set *blocks; | |
| /* Number of tagged nir_src's. This is implicitly the cardinality of the set | |
| * of pending nir_src's. | |
| */ | |
| uint32_t nr_tagged_srcs; | |
| /* bitset of ssa definitions we have found; used to check uniqueness */ | |
| BITSET_WORD *ssa_defs_found; | |
| /* map of variable -> function implementation where it is defined or NULL | |
| * if it is a global variable | |
| */ | |
| struct hash_table *var_defs; | |
| /* map of instruction/var/etc to failed assert string */ | |
| struct hash_table *errors; | |
| } validate_state; | |
| static bool interactive; | |
| #if !DETECT_OS_WINDOWS | |
| static void | |
| init_interactive(void) | |
| { | |
| interactive = isatty(STDERR_FILENO); | |
| } | |
| #endif | |
| static void | |
| log_error(validate_state *state, const char *cond, const char *file, int line) | |
| { | |
| const void *obj; | |
| const char *color_red = interactive ? "\x1b[0;1;31m" : ""; | |
| const char *color_reset = interactive ? "\x1b[0m" : ""; | |
| if (state->instr) | |
| obj = state->instr; | |
| else if (state->var) | |
| obj = state->var; | |
| else | |
| obj = cond; | |
| char *msg = ralloc_asprintf(state->errors, "%serror: %s (%s:%d)%s", | |
| color_red, cond, file, line, color_reset); | |
| _mesa_hash_table_insert(state->errors, obj, msg); | |
| } | |
| static bool | |
| validate_assert_impl(validate_state *state, bool cond, const char *str, | |
| const char *file, unsigned line) | |
| { | |
| if (unlikely(!cond)) | |
| log_error(state, str, file, line); | |
| return cond; | |
| } | |
| #define validate_assert(state, cond) \ | |
| validate_assert_impl(state, (cond), #cond, __FILE__, __LINE__) | |
| static void | |
| validate_num_components(validate_state *state, unsigned num_components) | |
| { | |
| validate_assert(state, nir_num_components_valid(num_components)); | |
| } | |
| /* Tag used in nir_src::_parent to indicate that a source has been seen. */ | |
| #define SRC_TAG_SEEN (0x2) | |
| static_assert(SRC_TAG_SEEN == (~NIR_SRC_PARENT_MASK + 1), | |
| "Parent pointer tags chosen not to collide"); | |
| static void | |
| tag_src(nir_src *src, validate_state *state) | |
| { | |
| /* nir_src only appears once and only in one SSA def use list, since we | |
| * mark nir_src's as we go by tagging this pointer. | |
| */ | |
| if (validate_assert(state, (src->_parent & SRC_TAG_SEEN) == 0)) { | |
| src->_parent |= SRC_TAG_SEEN; | |
| state->nr_tagged_srcs++; | |
| } | |
| } | |
| /* Due to tagging, it's not safe to use nir_src_use_instr during the main | |
| * validate loop. This is a tagging-aware version. | |
| */ | |
| static nir_instr * | |
| src_parent_instr_safe(nir_src *src) | |
| { | |
| uintptr_t untagged = (src->_parent & ~SRC_TAG_SEEN); | |
| assert(!(untagged & NIR_SRC_PARENT_IS_IF) && "precondition"); | |
| return (nir_instr *)untagged; | |
| } | |
| /* | |
| * As we walk SSA defs, we mark every use as seen by tagging the parent pointer. | |
| * We need to make sure our use is seen in a use list. | |
| * | |
| * Then we unmark when we hit the source. This will let us prove that we've | |
| * seen all the sources. | |
| */ | |
| static void | |
| validate_src_tag(nir_src *src, validate_state *state) | |
| { | |
| if (validate_assert(state, src->_parent & SRC_TAG_SEEN)) { | |
| src->_parent &= ~SRC_TAG_SEEN; | |
| state->nr_tagged_srcs--; | |
| } | |
| } | |
| static void | |
| validate_if_src(nir_src *src, validate_state *state) | |
| { | |
| validate_src_tag(src, state); | |
| validate_assert(state, nir_src_use_if(src) == state->if_stmt); | |
| validate_assert(state, src->ssa != NULL); | |
| validate_assert(state, src->ssa->num_components == 1); | |
| } | |
| static void | |
| validate_src(nir_src *src, validate_state *state) | |
| { | |
| /* Validate the tag first, so that nir_src_use_instr is valid */ | |
| validate_src_tag(src, state); | |
| /* Source assumed to be instruction, use validate_if_src for if */ | |
| validate_assert(state, nir_src_use_instr(src) == state->instr); | |
| validate_assert(state, src->ssa != NULL); | |
| } | |
| static void | |
| validate_sized_src(nir_src *src, validate_state *state, | |
| unsigned bit_sizes, unsigned num_components) | |
| { | |
| validate_src(src, state); | |
| if (bit_sizes) | |
| validate_assert(state, src->ssa->bit_size & bit_sizes); | |
| if (num_components) | |
| validate_assert(state, src->ssa->num_components == num_components); | |
| } | |
| static void | |
| validate_alu_src(nir_alu_instr *instr, unsigned index, validate_state *state) | |
| { | |
| nir_alu_src *src = &instr->src[index]; | |
| unsigned num_instr_channels = nir_ssa_alu_instr_src_components(instr, index); | |
| unsigned num_components = nir_src_num_components(src->src); | |
| for (unsigned i = 0; i < num_instr_channels; i++) { | |
| validate_assert(state, src->swizzle[i] < num_components); | |
| } | |
| validate_src(&src->src, state); | |
| } | |
| static void | |
| validate_def(nir_def *def, validate_state *state) | |
| { | |
| validate_assert(state, def->index < state->impl->ssa_alloc); | |
| validate_assert(state, !BITSET_TEST(state->ssa_defs_found, def->index)); | |
| BITSET_SET(state->ssa_defs_found, def->index); | |
| validate_num_components(state, def->num_components); | |
| list_validate(&def->uses); | |
| nir_foreach_use_including_if(src, def) { | |
| /* Check that the def matches. */ | |
| validate_assert(state, src->ssa == def); | |
| /* Check that nir_src's are unique */ | |
| tag_src(src, state); | |
| } | |
| } | |
| static void | |
| validate_alu_instr(nir_alu_instr *instr, validate_state *state) | |
| { | |
| validate_assert(state, instr->op < nir_num_opcodes); | |
| unsigned instr_bit_size = 0; | |
| for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs; i++) { | |
| nir_alu_type src_type = nir_op_infos[instr->op].input_types[i]; | |
| unsigned src_bit_size = nir_src_bit_size(instr->src[i].src); | |
| if (nir_alu_type_get_type_size(src_type)) { | |
| validate_assert(state, src_bit_size == nir_alu_type_get_type_size(src_type)); | |
| } else if (instr_bit_size) { | |
| validate_assert(state, src_bit_size == instr_bit_size); | |
| } else { | |
| instr_bit_size = src_bit_size; | |
| } | |
| if (nir_alu_type_get_base_type(src_type) == nir_type_float) { | |
| /* 8-bit float isn't a thing */ | |
| validate_assert(state, src_bit_size == 16 || src_bit_size == 32 || | |
| src_bit_size == 64); | |
| } | |
| /* In nir_opcodes.py, these are defined to take general uint or int | |
| * sources. However, they're really only defined for 32-bit or 64-bit | |
| * sources. This seems to be the only place to enforce this | |
| * restriction. | |
| */ | |
| switch (instr->op) { | |
| case nir_op_ufind_msb: | |
| case nir_op_ufind_msb_rev: | |
| validate_assert(state, src_bit_size == 32 || src_bit_size == 64); | |
| break; | |
| /* In nir_opcodes.py, these are defined to take general int sources. | |
| * However, for downcasts they are redundant with u2uN opcodes and | |
| * in order make things consistent, we only want to use the latter. | |
| */ | |
| case nir_op_i2i8: | |
| case nir_op_i2i16: | |
| case nir_op_i2i32: | |
| validate_assert(state, src_bit_size <= instr->def.bit_size); | |
| break; | |
| default: | |
| break; | |
| } | |
| validate_alu_src(instr, i, state); | |
| } | |
| nir_alu_type dest_type = nir_op_infos[instr->op].output_type; | |
| unsigned dest_bit_size = instr->def.bit_size; | |
| if (nir_alu_type_get_type_size(dest_type)) { | |
| validate_assert(state, dest_bit_size == nir_alu_type_get_type_size(dest_type)); | |
| } else if (instr_bit_size) { | |
| validate_assert(state, dest_bit_size == instr_bit_size); | |
| } else { | |
| /* The only unsized thing is the destination so it's vacuously valid */ | |
| } | |
| if (nir_alu_type_get_base_type(dest_type) == nir_type_float) { | |
| /* 8-bit float isn't a thing */ | |
| validate_assert(state, dest_bit_size == 16 || dest_bit_size == 32 || | |
| dest_bit_size == 64); | |
| } | |
| validate_def(&instr->def, state); | |
| validate_assert(state, (instr->fp_math_ctrl & ~nir_op_infos[instr->op].valid_fp_math_ctrl) == 0); | |
| validate_assert(state, nir_alu_instr_no_transform(instr) || !nir_alu_instr_no_contract(instr)); | |
| validate_assert(state, nir_alu_instr_no_transform(instr) || !nir_alu_instr_no_reassoc(instr)); | |
| } | |
| static void | |
| validate_var_use(nir_variable *var, validate_state *state) | |
| { | |
| struct hash_entry *entry = _mesa_hash_table_search(state->var_defs, var); | |
| validate_assert(state, entry); | |
| if (entry && var->data.mode == nir_var_function_temp) | |
| validate_assert(state, (nir_function_impl *)entry->data == state->impl); | |
| } | |
| static void | |
| validate_deref_instr(nir_deref_instr *instr, validate_state *state) | |
| { | |
| if (instr->deref_type == nir_deref_type_var) { | |
| /* Variable dereferences are stupid simple. */ | |
| validate_assert(state, instr->modes == instr->var->data.mode); | |
| validate_assert(state, instr->type == instr->var->type); | |
| validate_var_use(instr->var, state); | |
| } else if (instr->deref_type == nir_deref_type_cast) { | |
| /* For cast, we simply have to trust the instruction. It's up to | |
| * lowering passes and front/back-ends to make them sane. | |
| */ | |
| validate_src(&instr->parent, state); | |
| /* Most variable modes in NIR can only exist by themselves. */ | |
| if (instr->modes & ~nir_var_mem_generic) | |
| validate_assert(state, util_bitcount(instr->modes) == 1); | |
| nir_deref_instr *parent = nir_src_as_deref(instr->parent); | |
| if (parent) { | |
| if (parent->modes & nir_var_resource_heap) { | |
| /* Some casts from the resource heap pointer are allowed. */ | |
| validate_assert(state, instr->modes & (nir_var_resource_heap | | |
| nir_var_image | | |
| nir_var_uniform | | |
| nir_var_mem_ubo | | |
| nir_var_mem_ssbo)); | |
| } else if (parent->modes & nir_var_sampler_heap) { | |
| /* Some casts from the sampler heap pointer are allowed. */ | |
| validate_assert(state, instr->modes & (nir_var_sampler_heap | | |
| nir_var_uniform)); | |
| } else { | |
| /* Casts can change the mode but it can't change completely. The | |
| * new mode must have some bits in common with the old. | |
| */ | |
| validate_assert(state, instr->modes & parent->modes); | |
| } | |
| } else { | |
| /* If our parent isn't a deref, just assert the mode is there */ | |
| validate_assert(state, instr->modes != 0); | |
| } | |
| /* We just validate that the type is there */ | |
| validate_assert(state, instr->type); | |
| if (instr->cast.align_mul > 0) { | |
| validate_assert(state, util_is_power_of_two_nonzero(instr->cast.align_mul)); | |
| validate_assert(state, instr->cast.align_offset < instr->cast.align_mul); | |
| } else { | |
| validate_assert(state, instr->cast.align_offset == 0); | |
| } | |
| } else { | |
| /* The parent pointer value must have the same number of components | |
| * as the destination. | |
| */ | |
| validate_sized_src(&instr->parent, state, instr->def.bit_size, | |
| instr->def.num_components); | |
| nir_instr *parent_instr = nir_def_instr(instr->parent.ssa); | |
| /* The parent must come from another deref instruction */ | |
| validate_assert(state, parent_instr->type == nir_instr_type_deref); | |
| nir_deref_instr *parent = nir_instr_as_deref(parent_instr); | |
| validate_assert(state, instr->modes == parent->modes); | |
| switch (instr->deref_type) { | |
| case nir_deref_type_struct: | |
| validate_assert(state, glsl_type_is_struct_or_ifc(parent->type)); | |
| validate_assert(state, | |
| instr->strct.index < glsl_get_length(parent->type)); | |
| validate_assert(state, instr->type == | |
| glsl_get_struct_field(parent->type, instr->strct.index)); | |
| break; | |
| case nir_deref_type_array: | |
| case nir_deref_type_array_wildcard: | |
| if (instr->modes & nir_var_vec_indexable_modes) { | |
| /* Shared variables and UBO/SSBOs have a bit more relaxed rules | |
| * because we need to be able to handle array derefs on vectors. | |
| * Fortunately, nir_lower_io handles these just fine. | |
| */ | |
| validate_assert(state, glsl_type_is_array(parent->type) || | |
| glsl_type_is_matrix(parent->type) || | |
| glsl_type_is_vector(parent->type)); | |
| } else { | |
| /* Most of NIR cannot handle array derefs on vectors */ | |
| validate_assert(state, glsl_type_is_array(parent->type) || | |
| glsl_type_is_matrix(parent->type)); | |
| } | |
| validate_assert(state, | |
| instr->type == glsl_get_array_element(parent->type)); | |
| if (instr->deref_type == nir_deref_type_array) { | |
| validate_sized_src(&instr->arr.index, state, | |
| instr->def.bit_size, 1); | |
| } | |
| break; | |
| case nir_deref_type_ptr_as_array: | |
| /* ptr_as_array derefs must have a parent that is either an array, | |
| * ptr_as_array, or cast. If the parent is a cast, we get the stride | |
| * information (if any) from the cast deref. | |
| */ | |
| validate_assert(state, | |
| parent->deref_type == nir_deref_type_array || | |
| parent->deref_type == nir_deref_type_ptr_as_array || | |
| parent->deref_type == nir_deref_type_cast); | |
| validate_sized_src(&instr->arr.index, state, | |
| instr->def.bit_size, 1); | |
| break; | |
| default: | |
| UNREACHABLE("Invalid deref instruction type"); | |
| } | |
| } | |
| /* We intentionally don't validate the size of the destination because we | |
| * want to let other compiler components such as SPIR-V decide how big | |
| * pointers should be. | |
| */ | |
| validate_def(&instr->def, state); | |
| /* Certain modes cannot be used as sources for phi instructions because | |
| * way too many passes assume that they can always chase deref chains. | |
| */ | |
| nir_foreach_use_including_if(use, &instr->def) { | |
| /* Deref instructions as if conditions don't make sense because if | |
| * conditions expect well-formed Booleans. If you want to compare with | |
| * NULL, an explicit comparison operation should be used. | |
| */ | |
| if (!validate_assert(state, !nir_src_is_if(use))) | |
| continue; | |
| if (src_parent_instr_safe(use)->type == nir_instr_type_phi) { | |
| validate_assert(state, !(instr->modes & (nir_var_shader_in | | |
| nir_var_shader_out | | |
| nir_var_shader_out | | |
| nir_var_uniform))); | |
| } | |
| } | |
| } | |
| static bool | |
| vectorized_intrinsic(nir_intrinsic_instr *intr) | |
| { | |
| const nir_intrinsic_info *info = &nir_intrinsic_infos[intr->intrinsic]; | |
| if (info->dest_components == 0) | |
| return true; | |
| for (unsigned i = 0; i < info->num_srcs; i++) | |
| if (info->src_components[i] == 0) | |
| return true; | |
| return false; | |
| } | |
| /** Returns the image format or PIPE_FORMAT_COUNT for incomplete derefs | |
| * | |
| * We use PIPE_FORMAT_COUNT for incomplete derefs because PIPE_FORMAT_NONE | |
| * indicates that we found the variable but it has no format specified. | |
| */ | |
| static enum pipe_format | |
| image_intrin_format(nir_intrinsic_instr *instr) | |
| { | |
| if (nir_intrinsic_format(instr) != PIPE_FORMAT_NONE) | |
| return nir_intrinsic_format(instr); | |
| /* If this not a deref intrinsic, PIPE_FORMAT_NONE is the best we can do */ | |
| if (nir_intrinsic_infos[instr->intrinsic].src_components[0] != -1) | |
| return PIPE_FORMAT_NONE; | |
| nir_variable *var = nir_intrinsic_get_var(instr, 0); | |
| if (var == NULL) | |
| return PIPE_FORMAT_COUNT; | |
| return var->data.image.format; | |
| } | |
| static void | |
| validate_register_handle(nir_src handle_src, | |
| unsigned num_components, | |
| unsigned bit_size, | |
| validate_state *state) | |
| { | |
| nir_def *handle = handle_src.ssa; | |
| nir_instr *parent = nir_def_instr(handle); | |
| if (!validate_assert(state, parent->type == nir_instr_type_intrinsic)) | |
| return; | |
| nir_intrinsic_instr *intr = nir_instr_as_intrinsic(parent); | |
| if (!validate_assert(state, intr->intrinsic == nir_intrinsic_decl_reg)) | |
| return; | |
| validate_assert(state, nir_intrinsic_num_components(intr) == num_components); | |
| validate_assert(state, nir_intrinsic_bit_size(intr) == bit_size); | |
| } | |
| static void | |
| validate_intrinsic_instr(nir_intrinsic_instr *instr, validate_state *state) | |
| { | |
| unsigned dest_bit_size = 0; | |
| unsigned src_bit_sizes[NIR_INTRINSIC_MAX_INPUTS] = { | |
| 0, | |
| }; | |
| switch (instr->intrinsic) { | |
| case nir_intrinsic_decl_reg: | |
| assert(state->block == nir_start_block(state->impl)); | |
| break; | |
| case nir_intrinsic_load_reg: | |
| case nir_intrinsic_load_reg_indirect: | |
| validate_register_handle(instr->src[0], | |
| instr->def.num_components, | |
| instr->def.bit_size, state); | |
| break; | |
| case nir_intrinsic_store_reg: | |
| case nir_intrinsic_store_reg_indirect: | |
| validate_register_handle(instr->src[1], | |
| nir_src_num_components(instr->src[0]), | |
| nir_src_bit_size(instr->src[0]), state); | |
| break; | |
| case nir_intrinsic_convert_alu_types: { | |
| nir_alu_type src_type = nir_intrinsic_src_type(instr); | |
| nir_alu_type dest_type = nir_intrinsic_dest_type(instr); | |
| dest_bit_size = nir_alu_type_get_type_size(dest_type); | |
| src_bit_sizes[0] = nir_alu_type_get_type_size(src_type); | |
| validate_assert(state, dest_bit_size != 0); | |
| validate_assert(state, src_bit_sizes[0] != 0); | |
| break; | |
| } | |
| case nir_intrinsic_load_param: { | |
| unsigned param_idx = nir_intrinsic_param_idx(instr); | |
| validate_assert(state, param_idx < state->impl->function->num_params); | |
| nir_parameter *param = &state->impl->function->params[param_idx]; | |
| validate_assert(state, instr->num_components == param->num_components); | |
| dest_bit_size = param->bit_size; | |
| break; | |
| } | |
| case nir_intrinsic_load_deref: { | |
| nir_deref_instr *src = nir_src_as_deref(instr->src[0]); | |
| assert(src); | |
| validate_assert(state, glsl_type_is_vector_or_scalar(src->type) || | |
| (src->modes == nir_var_uniform && | |
| glsl_get_base_type(src->type) == GLSL_TYPE_SUBROUTINE)); | |
| validate_assert(state, instr->num_components == | |
| glsl_get_vector_elements(src->type)); | |
| dest_bit_size = glsl_get_bit_size(src->type); | |
| /* Also allow 32-bit boolean load operations */ | |
| if (glsl_type_is_boolean(src->type)) | |
| dest_bit_size |= 32; | |
| break; | |
| } | |
| case nir_intrinsic_store_deref: { | |
| nir_deref_instr *dst = nir_src_as_deref(instr->src[0]); | |
| assert(dst); | |
| validate_assert(state, glsl_type_is_vector_or_scalar(dst->type)); | |
| validate_assert(state, instr->num_components == | |
| glsl_get_vector_elements(dst->type)); | |
| src_bit_sizes[1] = glsl_get_bit_size(dst->type); | |
| /* Also allow 32-bit boolean store operations */ | |
| if (glsl_type_is_boolean(dst->type)) | |
| src_bit_sizes[1] |= 32; | |
| validate_assert(state, !nir_deref_mode_may_be(dst, nir_var_read_only_modes)); | |
| break; | |
| } | |
| case nir_intrinsic_copy_deref: { | |
| nir_deref_instr *dst = nir_src_as_deref(instr->src[0]); | |
| nir_deref_instr *src = nir_src_as_deref(instr->src[1]); | |
| validate_assert(state, glsl_get_bare_type(dst->type) == | |
| glsl_get_bare_type(src->type)); | |
| validate_assert(state, !nir_deref_mode_may_be(dst, nir_var_read_only_modes)); | |
| /* FIXME: now that we track if the var copies were lowered, it would be | |
| * good to validate here that no new copy derefs were added. Right now | |
| * we can't as there are some specific cases where copies are added even | |
| * after the lowering. One example is the Intel compiler, that calls | |
| * nir_lower_io_vars_to_temporaries when linking some shader stages. | |
| */ | |
| break; | |
| } | |
| case nir_intrinsic_load_ubo_vec4: { | |
| int bit_size = instr->def.bit_size; | |
| validate_assert(state, bit_size >= 8); | |
| validate_assert(state, (nir_intrinsic_component(instr) + | |
| instr->num_components) * | |
| (bit_size / 8) <= | |
| 16); | |
| break; | |
| } | |
| case nir_intrinsic_load_ubo: | |
| /* Make sure that the creator didn't forget to set the range_base+range. */ | |
| validate_assert(state, nir_intrinsic_range(instr) != 0); | |
| FALLTHROUGH; | |
| case nir_intrinsic_load_ssbo: | |
| case nir_intrinsic_load_shared: | |
| case nir_intrinsic_load_global: | |
| case nir_intrinsic_load_global_constant: | |
| case nir_intrinsic_load_scratch: | |
| case nir_intrinsic_load_constant: | |
| /* These memory load operations must have alignments */ | |
| validate_assert(state, | |
| util_is_power_of_two_nonzero(nir_intrinsic_align_mul(instr))); | |
| validate_assert(state, nir_intrinsic_align_offset(instr) < | |
| nir_intrinsic_align_mul(instr)); | |
| FALLTHROUGH; | |
| case nir_intrinsic_load_uniform: | |
| case nir_intrinsic_load_input: | |
| case nir_intrinsic_load_per_primitive_input: | |
| case nir_intrinsic_load_per_vertex_input: | |
| case nir_intrinsic_load_interpolated_input: | |
| case nir_intrinsic_load_output: | |
| case nir_intrinsic_load_per_vertex_output: | |
| case nir_intrinsic_load_per_view_output: | |
| case nir_intrinsic_load_per_primitive_output: | |
| case nir_intrinsic_load_push_constant: | |
| case nir_intrinsic_load_attr_pan: | |
| /* All memory load operations must load at least a byte */ | |
| validate_assert(state, instr->def.bit_size >= 8); | |
| break; | |
| case nir_intrinsic_load_barycentric_pixel: | |
| case nir_intrinsic_load_barycentric_centroid: | |
| case nir_intrinsic_load_barycentric_sample: | |
| case nir_intrinsic_load_barycentric_at_offset: | |
| case nir_intrinsic_load_barycentric_at_sample: { | |
| enum glsl_interp_mode mode = nir_intrinsic_interp_mode(instr); | |
| validate_assert(state, | |
| mode == INTERP_MODE_NONE || | |
| mode == INTERP_MODE_SMOOTH || | |
| mode == INTERP_MODE_NOPERSPECTIVE); | |
| break; | |
| } | |
| case nir_intrinsic_store_ssbo: | |
| case nir_intrinsic_store_shared: | |
| case nir_intrinsic_store_global: | |
| case nir_intrinsic_store_scratch: | |
| /* These memory store operations must also have alignments */ | |
| validate_assert(state, | |
| util_is_power_of_two_nonzero(nir_intrinsic_align_mul(instr))); | |
| validate_assert(state, nir_intrinsic_align_offset(instr) < | |
| nir_intrinsic_align_mul(instr)); | |
| /* All memory store operations must store at least a byte */ | |
| validate_assert(state, nir_src_bit_size(instr->src[0]) >= 8); | |
| break; | |
| case nir_intrinsic_store_output: | |
| case nir_intrinsic_store_per_vertex_output: | |
| case nir_intrinsic_store_per_view_output: | |
| validate_assert(state, nir_src_bit_size(instr->src[0]) >= 8); | |
| validate_assert(state, | |
| nir_src_bit_size(instr->src[0]) == | |
| nir_alu_type_get_type_size(nir_intrinsic_src_type(instr))); | |
| break; | |
| case nir_intrinsic_deref_mode_is: | |
| case nir_intrinsic_addr_mode_is: | |
| validate_assert(state, | |
| util_bitcount(nir_intrinsic_memory_modes(instr)) == 1); | |
| break; | |
| case nir_intrinsic_image_deref_atomic: | |
| case nir_intrinsic_image_deref_atomic_swap: | |
| case nir_intrinsic_bindless_image_atomic: | |
| case nir_intrinsic_bindless_image_atomic_swap: | |
| case nir_intrinsic_image_atomic: | |
| case nir_intrinsic_image_atomic_swap: | |
| case nir_intrinsic_image_heap_atomic: | |
| case nir_intrinsic_image_heap_atomic_swap: { | |
| nir_atomic_op op = nir_intrinsic_atomic_op(instr); | |
| enum pipe_format format = image_intrin_format(instr); | |
| if (format != PIPE_FORMAT_COUNT) { | |
| bool allowed = false; | |
| bool is_float = (nir_atomic_op_type(op) == nir_type_float); | |
| switch (format) { | |
| case PIPE_FORMAT_R32_FLOAT: | |
| case PIPE_FORMAT_R16G16_FLOAT: | |
| case PIPE_FORMAT_R16G16B16A16_FLOAT: | |
| allowed = is_float || op == nir_atomic_op_xchg; | |
| break; | |
| case PIPE_FORMAT_R16_FLOAT: | |
| case PIPE_FORMAT_R64_FLOAT: | |
| allowed = op == nir_atomic_op_fmin || op == nir_atomic_op_fmax; | |
| break; | |
| case PIPE_FORMAT_R32_UINT: | |
| case PIPE_FORMAT_R32_SINT: | |
| case PIPE_FORMAT_R64_UINT: | |
| case PIPE_FORMAT_R64_SINT: | |
| allowed = !is_float; | |
| break; | |
| default: | |
| break; | |
| } | |
| validate_assert(state, allowed); | |
| const struct util_format_description *fmt_desc = | |
| util_format_description(format); | |
| validate_assert(state, instr->def.bit_size == | |
| fmt_desc->channel[0].size); | |
| validate_assert(state, | |
| instr->num_components >= 1 && | |
| instr->num_components <= 4); | |
| } | |
| break; | |
| } | |
| case nir_intrinsic_load_buffer_amd: | |
| if (nir_intrinsic_access(instr) & ACCESS_USES_FORMAT_AMD) { | |
| nir_alu_type dest_type = nir_intrinsic_dest_type(instr); | |
| validate_assert(state, nir_alu_type_get_type_size(dest_type) && | |
| nir_alu_type_get_base_type(dest_type)); | |
| validate_assert(state, nir_alu_type_get_type_size(dest_type) == | |
| instr->def.bit_size); | |
| } | |
| break; | |
| case nir_intrinsic_store_buffer_amd: | |
| if (nir_intrinsic_access(instr) & ACCESS_USES_FORMAT_AMD) { | |
| unsigned writemask = nir_intrinsic_write_mask(instr); | |
| /* Make sure the writemask is derived from the component count. */ | |
| validate_assert(state, | |
| writemask == | |
| BITFIELD_MASK(nir_src_num_components(instr->src[0]))); | |
| } | |
| break; | |
| case nir_intrinsic_load_deref_transpose_amd: { | |
| nir_deref_instr *src = nir_src_as_deref(instr->src[0]); | |
| assert(src); | |
| unsigned disallow_access = ACCESS_ATOMIC | ACCESS_SKIP_HELPERS | ACCESS_SMEM_AMD; | |
| validate_assert(state, !(nir_intrinsic_access(instr) & disallow_access)); | |
| validate_assert(state, glsl_type_is_scalar(src->type)); | |
| validate_assert(state, instr->num_components == 8 || instr->num_components == 4); | |
| dest_bit_size = glsl_get_bit_size(src->type); | |
| src_bit_sizes[0] = 64; | |
| break; | |
| } | |
| case nir_intrinsic_load_global_transpose_amd: | |
| case nir_intrinsic_load_global_tr_amd: { | |
| unsigned disallow_access = ACCESS_ATOMIC | ACCESS_SKIP_HELPERS | ACCESS_SMEM_AMD; | |
| validate_assert(state, !(nir_intrinsic_access(instr) & disallow_access)); | |
| validate_assert(state, instr->num_components == 8 || instr->num_components == 4); | |
| src_bit_sizes[0] = 64; | |
| src_bit_sizes[1] = 32; | |
| break; | |
| } | |
| case nir_intrinsic_global_atomic_nv: | |
| case nir_intrinsic_global_atomic_swap_nv: | |
| case nir_intrinsic_shared_atomic_nv: | |
| case nir_intrinsic_shared_atomic_swap_nv: | |
| case nir_intrinsic_ldc_nv: | |
| case nir_intrinsic_ldcx_nv: | |
| case nir_intrinsic_load_global_nv: | |
| case nir_intrinsic_load_scratch_nv: | |
| case nir_intrinsic_load_shared_nv: | |
| case nir_intrinsic_store_global_nv: | |
| case nir_intrinsic_store_scratch_nv: | |
| case nir_intrinsic_store_shared_nv: | |
| case nir_intrinsic_store_shared_unlock_nv: | |
| case nir_intrinsic_vild_nv: { | |
| int base = nir_intrinsic_base(instr); | |
| nir_src src = *nir_get_io_offset_src(instr); | |
| nir_src *uniform_src = nir_get_io_uniform_offset_src(instr); | |
| unsigned const_bits = nir_get_io_base_size_nv(instr); | |
| if (nir_src_is_const(src) && nir_src_as_int(src) == 0 && | |
| (!uniform_src || (nir_src_is_const(*uniform_src) && nir_src_as_int(*uniform_src) == 0))) { | |
| validate_assert(state, base >= 0 && base < BITFIELD_MASK(const_bits)); | |
| } else { | |
| int32_t max = BITFIELD_MASK(const_bits - 1); | |
| int32_t min = ~BITFIELD_MASK(const_bits - 1); | |
| validate_assert(state, base >= min && base < max); | |
| } | |
| if (uniform_src) | |
| validate_assert(state, uniform_src->ssa->bit_size >= src.ssa->bit_size); | |
| if (instr->intrinsic == nir_intrinsic_load_global_nv) { | |
| validate_assert(state, instr->src[2].ssa->bit_size == 1); | |
| } | |
| break; | |
| } | |
| case nir_intrinsic_barrier: { | |
| unsigned semantics = nir_intrinsic_memory_semantics(instr); | |
| bool is_arrive = semantics & NIR_MEMORY_CONTROL_ARRIVE; | |
| bool is_wait = semantics & NIR_MEMORY_CONTROL_WAIT; | |
| if (nir_intrinsic_execution_scope(instr) != SCOPE_NONE) { | |
| if (is_wait) | |
| validate_assert(state, is_arrive || !(semantics & NIR_MEMORY_RELEASE)); | |
| if (is_arrive) | |
| validate_assert(state, is_wait || !(semantics & NIR_MEMORY_ACQUIRE)); | |
| } else { | |
| validate_assert(state, !is_arrive && !is_wait); | |
| } | |
| break; | |
| } | |
| default: | |
| break; | |
| } | |
| if (instr->num_components > 0) | |
| validate_num_components(state, instr->num_components); | |
| const nir_intrinsic_info *info = &nir_intrinsic_infos[instr->intrinsic]; | |
| unsigned num_srcs = info->num_srcs; | |
| for (unsigned i = 0; i < num_srcs; i++) { | |
| unsigned components_read = nir_intrinsic_src_components(instr, i); | |
| validate_num_components(state, components_read); | |
| validate_sized_src(&instr->src[i], state, src_bit_sizes[i], components_read); | |
| } | |
| if (nir_intrinsic_infos[instr->intrinsic].has_dest) { | |
| unsigned components_written = nir_intrinsic_dest_components(instr); | |
| unsigned bit_sizes = info->dest_bit_sizes; | |
| if (!bit_sizes && info->bit_size_src >= 0) | |
| bit_sizes = nir_src_bit_size(instr->src[info->bit_size_src]); | |
| validate_num_components(state, components_written); | |
| if (dest_bit_size && bit_sizes) | |
| validate_assert(state, dest_bit_size & bit_sizes); | |
| else | |
| dest_bit_size = dest_bit_size ? dest_bit_size : bit_sizes; | |
| validate_def(&instr->def, state); | |
| validate_assert(state, instr->def.num_components == components_written); | |
| if (dest_bit_size) | |
| validate_assert(state, instr->def.bit_size & dest_bit_size); | |
| } | |
| if (!vectorized_intrinsic(instr)) | |
| validate_assert(state, instr->num_components == 0); | |
| if (nir_intrinsic_has_write_mask(instr)) { | |
| unsigned component_mask = BITFIELD_MASK(instr->num_components); | |
| validate_assert(state, (nir_intrinsic_write_mask(instr) & ~component_mask) == 0); | |
| } | |
| if (nir_intrinsic_has_io_xfb(instr)) { | |
| unsigned used_mask = 0; | |
| nir_io_xfb xfb = nir_intrinsic_io_xfb(instr); | |
| for (unsigned i = 0; i < 4; i++) { | |
| unsigned xfb_mask = BITFIELD_RANGE(i, xfb.out[i].num_components); | |
| /* Each component can be used only once by transform feedback info. */ | |
| validate_assert(state, (xfb_mask & used_mask) == 0); | |
| used_mask |= xfb_mask; | |
| } | |
| } | |
| if (nir_intrinsic_has_io_semantics(instr) && | |
| !nir_intrinsic_io_semantics(instr).no_validate) { | |
| bool is_input = instr->intrinsic == nir_intrinsic_load_input || | |
| instr->intrinsic == nir_intrinsic_load_per_vertex_input || | |
| instr->intrinsic == nir_intrinsic_load_per_primitive_input || | |
| instr->intrinsic == nir_intrinsic_load_interpolated_input || | |
| instr->intrinsic == nir_intrinsic_load_input_vertex; | |
| bool is_output = instr->intrinsic == nir_intrinsic_load_output || | |
| instr->intrinsic == nir_intrinsic_load_per_vertex_output || | |
| instr->intrinsic == nir_intrinsic_load_per_primitive_output || | |
| instr->intrinsic == nir_intrinsic_load_per_view_output || | |
| instr->intrinsic == nir_intrinsic_store_output || | |
| instr->intrinsic == nir_intrinsic_store_per_vertex_output || | |
| instr->intrinsic == nir_intrinsic_store_per_primitive_output || | |
| instr->intrinsic == nir_intrinsic_store_per_view_output; | |
| /* Driver-specific intrinsics with IO semantics are not validated. */ | |
| bool is_core_intrinsic = is_input || is_output; | |
| if (is_core_intrinsic) { | |
| nir_io_semantics sem = nir_intrinsic_io_semantics(instr); | |
| unsigned max_num_components; | |
| /* Validate that a single intrinsic doesn't access past the maximum | |
| * number of components per slot. | |
| */ | |
| /* TODO: remove nir_io_radv_intrinsic_component_workaround */ | |
| if (state->shader->options->io_options & | |
| nir_io_radv_intrinsic_component_workaround && sem.high_16bits) | |
| max_num_components = 8; | |
| else | |
| max_num_components = 4; | |
| if (nir_intrinsic_infos[instr->intrinsic].has_dest) { | |
| /* Input and output loads shouldn't load from multiple slots at once. */ | |
| validate_assert(state, nir_intrinsic_component(instr) + | |
| instr->def.num_components <= max_num_components); | |
| } else { | |
| /* Output stores shouldn't write to multiple slots at once. */ | |
| validate_assert(state, nir_intrinsic_component(instr) + | |
| instr->num_components <= max_num_components); | |
| } | |
| if (is_input) { | |
| validate_assert(state, !sem.no_sysval_output); | |
| validate_assert(state, !sem.no_varying); | |
| validate_assert(state, !sem.gs_streams); | |
| } else { | |
| /* An output that has no effect shouldn't be present in the IR. */ | |
| validate_assert(state, | |
| (nir_slot_is_sysval_output(sem.location, MESA_SHADER_NONE) && | |
| !sem.no_sysval_output) || | |
| (nir_slot_is_varying(sem.location, MESA_SHADER_NONE) && | |
| !sem.no_varying) || | |
| nir_instr_xfb_write_mask(instr) || | |
| /* TCS can set no_varying and no_sysval_output, meaning | |
| * that the output is only read by TCS and not TES. | |
| */ | |
| state->shader->info.stage == MESA_SHADER_TESS_CTRL); | |
| validate_assert(state, | |
| !sem.gs_streams || | |
| state->shader->info.stage == MESA_SHADER_GEOMETRY); | |
| } | |
| validate_assert(state, | |
| (!sem.dual_source_blend_index && | |
| !sem.fb_fetch_output && | |
| !sem.fb_fetch_output_coherent) || | |
| (state->shader->info.stage == MESA_SHADER_FRAGMENT && | |
| is_output)); | |
| validate_assert(state, | |
| !sem.high_dvec2 || | |
| (state->shader->info.stage == MESA_SHADER_VERTEX && | |
| is_input)); | |
| validate_assert(state, | |
| !sem.interp_explicit_strict || | |
| (state->shader->info.stage == MESA_SHADER_FRAGMENT && | |
| instr->intrinsic == nir_intrinsic_load_input_vertex)); | |
| /* Non-constant src offset with num_slots == 1 is disallowed. */ | |
| if (sem.num_slots == 1) { | |
| nir_src *offset_src = nir_get_io_offset_src(instr); | |
| /* TODO: nir_opt_loop produces phi 0, 0 for store_output as follows, | |
| * which isn't incorrect, but also isn't useful: | |
| * | |
| * Before: | |
| * %1 = load_const 0 | |
| * loop { | |
| * if { | |
| * } else { | |
| * } | |
| * } | |
| * store_output(, %1) | |
| * | |
| * After: | |
| * %1 = load_const 0 | |
| * if { | |
| * } else { | |
| * loop { | |
| * %38 = load_const 0 | |
| * } | |
| * } | |
| * %52 = phi %1, %38 | |
| * store_output(, %52) | |
| * | |
| * Test: KHR-GLES3.shaders.indexing.uniform_array.float_dynamic_loop_read_fragment | |
| * | |
| * So allow phis in offset_src with num_slots == 1 for now. | |
| */ | |
| validate_assert(state, | |
| nir_src_is_const(*offset_src) || | |
| nir_def_is_phi(offset_src->ssa)); | |
| } | |
| } | |
| } | |
| if (nir_intrinsic_has_offset_shift(instr) && | |
| nir_intrinsic_has_align(instr)) { | |
| unsigned min_align = 1 << nir_intrinsic_offset_shift(instr); | |
| validate_assert(state, nir_intrinsic_align(instr) >= min_align); | |
| } | |
| if (nir_intrinsic_has_offset_shift_nv(instr)) { | |
| unsigned shift = nir_intrinsic_offset_shift_nv(instr); | |
| validate_assert(state, shift == 0 || (shift >= 2 && shift <= 4)); | |
| } | |
| } | |
| static void | |
| validate_tex_src_texture_deref(nir_tex_instr *instr, validate_state *state, | |
| nir_deref_instr *deref) | |
| { | |
| validate_assert(state, glsl_type_is_image(deref->type) || | |
| glsl_type_is_texture(deref->type) || | |
| glsl_type_is_sampler(deref->type)); | |
| switch (instr->op) { | |
| case nir_texop_descriptor_amd: | |
| case nir_texop_sampler_descriptor_amd: | |
| case nir_texop_custom_border_color_agx: | |
| break; | |
| case nir_texop_lod: | |
| case nir_texop_lod_bias: | |
| validate_assert(state, nir_alu_type_get_base_type(instr->dest_type) == nir_type_float); | |
| break; | |
| case nir_texop_image_min_lod_agx: | |
| validate_assert(state, instr->dest_type == nir_type_float16 || | |
| instr->dest_type == nir_type_uint16); | |
| break; | |
| case nir_texop_samples_identical: | |
| case nir_texop_has_custom_border_color_agx: | |
| validate_assert(state, nir_alu_type_get_base_type(instr->dest_type) == nir_type_bool); | |
| break; | |
| case nir_texop_txs: | |
| case nir_texop_texture_samples: | |
| case nir_texop_query_levels: | |
| case nir_texop_fragment_mask_fetch_amd: | |
| case nir_texop_txf_ms_mcs_intel: | |
| case nir_texop_txs_lod0_intel: | |
| case nir_texop_resinfo_intel: | |
| validate_assert(state, nir_alu_type_get_base_type(instr->dest_type) == nir_type_int || | |
| nir_alu_type_get_base_type(instr->dest_type) == nir_type_uint); | |
| break; | |
| default: | |
| validate_assert(state, | |
| glsl_get_sampler_result_type(deref->type) == GLSL_TYPE_VOID || | |
| glsl_base_type_is_integer(glsl_get_sampler_result_type(deref->type)) == | |
| (nir_alu_type_get_base_type(instr->dest_type) == nir_type_int || | |
| nir_alu_type_get_base_type(instr->dest_type) == nir_type_uint)); | |
| } | |
| } | |
| static void | |
| validate_tex_instr(nir_tex_instr *instr, validate_state *state) | |
| { | |
| bool src_type_seen[nir_num_tex_src_types]; | |
| for (unsigned i = 0; i < nir_num_tex_src_types; i++) | |
| src_type_seen[i] = false; | |
| for (unsigned i = 0; i < instr->num_srcs; i++) { | |
| validate_assert(state, !src_type_seen[instr->src[i].src_type]); | |
| src_type_seen[instr->src[i].src_type] = true; | |
| validate_sized_src(&instr->src[i].src, state, | |
| 0, nir_tex_instr_src_size(instr, i)); | |
| switch (instr->src[i].src_type) { | |
| case nir_tex_src_comparator: | |
| validate_assert(state, instr->is_shadow); | |
| break; | |
| case nir_tex_src_bias: | |
| validate_assert(state, instr->op == nir_texop_txb || | |
| instr->op == nir_texop_tg4 || | |
| instr->op == nir_texop_lod || | |
| instr->op == nir_texop_sparse_residency_intel); | |
| break; | |
| case nir_tex_src_lod: | |
| validate_assert(state, instr->op != nir_texop_tex && | |
| instr->op != nir_texop_txb && | |
| instr->op != nir_texop_txd && | |
| instr->op != nir_texop_lod); | |
| break; | |
| case nir_tex_src_ddx: | |
| case nir_tex_src_ddy: | |
| validate_assert(state, | |
| instr->op == nir_texop_txd || | |
| instr->op == nir_texop_sparse_residency_intel); | |
| break; | |
| case nir_tex_src_texture_deref: | |
| case nir_tex_src_texture_2_deref: { | |
| nir_deref_instr *deref = nir_src_as_deref(instr->src[i].src); | |
| if (!validate_assert(state, deref)) | |
| break; | |
| validate_tex_src_texture_deref(instr, state, deref); | |
| break; | |
| } | |
| case nir_tex_src_sampler_deref: | |
| case nir_tex_src_sampler_2_deref: { | |
| nir_deref_instr *deref = nir_src_as_deref(instr->src[i].src); | |
| if (!validate_assert(state, deref)) | |
| break; | |
| validate_assert(state, glsl_type_is_sampler(deref->type)); | |
| break; | |
| } | |
| case nir_tex_src_sampler_deref_intrinsic: | |
| case nir_tex_src_texture_deref_intrinsic: { | |
| nir_intrinsic_instr *intrin = | |
| nir_def_as_intrinsic(instr->src[i].src.ssa); | |
| nir_deref_instr *deref = | |
| nir_def_as_deref(intrin->src[0].ssa); | |
| if (!validate_assert(state, deref)) | |
| break; | |
| if (instr->src[i].src_type == nir_tex_src_sampler_deref_intrinsic) | |
| validate_assert(state, glsl_type_is_sampler(deref->type)); | |
| else | |
| validate_tex_src_texture_deref(instr, state, deref); | |
| break; | |
| } | |
| case nir_tex_src_block_size: | |
| validate_assert(state, | |
| instr->op == nir_texop_block_match_sad_qcom || | |
| instr->op == nir_texop_block_match_ssd_qcom); | |
| break; | |
| case nir_tex_src_box_size: | |
| validate_assert(state, instr->op == nir_texop_box_filter_qcom); | |
| break; | |
| case nir_tex_src_coord: | |
| case nir_tex_src_ref_coord: | |
| case nir_tex_src_projector: | |
| case nir_tex_src_offset: | |
| case nir_tex_src_min_lod: | |
| case nir_tex_src_ms_index: | |
| case nir_tex_src_texture_offset: | |
| case nir_tex_src_sampler_offset: | |
| case nir_tex_src_plane: | |
| case nir_tex_src_texture_handle: | |
| case nir_tex_src_sampler_handle: | |
| case nir_tex_src_texture_2_handle: | |
| case nir_tex_src_sampler_2_handle: | |
| break; | |
| default: | |
| break; | |
| } | |
| } | |
| bool msaa = (instr->sampler_dim == GLSL_SAMPLER_DIM_MS || | |
| instr->sampler_dim == GLSL_SAMPLER_DIM_SUBPASS_MS); | |
| if (msaa) | |
| validate_assert(state, instr->op != nir_texop_txf); | |
| else | |
| validate_assert(state, instr->op != nir_texop_txf_ms); | |
| if (instr->op != nir_texop_tg4) | |
| validate_assert(state, instr->component == 0); | |
| if (nir_tex_instr_has_explicit_tg4_offsets(instr)) { | |
| validate_assert(state, instr->op == nir_texop_tg4); | |
| validate_assert(state, !src_type_seen[nir_tex_src_offset]); | |
| } | |
| if (instr->is_gather_implicit_lod) | |
| validate_assert(state, instr->op == nir_texop_tg4); | |
| validate_def(&instr->def, state); | |
| validate_assert(state, instr->def.num_components == | |
| nir_tex_instr_dest_size(instr)); | |
| unsigned bit_size = nir_alu_type_get_type_size(instr->dest_type); | |
| validate_assert(state, | |
| (bit_size ? bit_size : 32) == | |
| instr->def.bit_size); | |
| } | |
| static void | |
| validate_call_instr(nir_call_instr *instr, validate_state *state) | |
| { | |
| validate_assert(state, instr->num_params == instr->callee->num_params); | |
| if (instr->indirect_callee.ssa) { | |
| validate_assert(state, !instr->callee->impl); | |
| validate_src(&instr->indirect_callee, state); | |
| } | |
| for (unsigned i = 0; i < instr->num_params; i++) { | |
| validate_sized_src(&instr->params[i], state, | |
| instr->callee->params[i].bit_size, | |
| instr->callee->params[i].num_components); | |
| } | |
| } | |
| static void | |
| validate_cmat_call_instr(nir_cmat_call_instr *instr, validate_state *state) | |
| { | |
| validate_assert(state, instr->num_params == nir_cmat_call_op_params(instr->op, instr->callee)); | |
| for (unsigned i = 0; i < instr->num_params; i++) { | |
| validate_src(&instr->params[i], state); | |
| } | |
| } | |
| static void | |
| validate_const_value(nir_const_value *val, unsigned bit_size, | |
| bool is_null_constant, validate_state *state) | |
| { | |
| /* In order for block copies to work properly for things like instruction | |
| * comparisons and [de]serialization, we require the unused bits of the | |
| * nir_const_value to be zero. | |
| */ | |
| nir_const_value cmp_val; | |
| memset(&cmp_val, 0, sizeof(cmp_val)); | |
| if (!is_null_constant) { | |
| switch (bit_size) { | |
| case 1: | |
| cmp_val.b = val->b; | |
| break; | |
| case 8: | |
| cmp_val.u8 = val->u8; | |
| break; | |
| case 16: | |
| cmp_val.u16 = val->u16; | |
| break; | |
| case 32: | |
| cmp_val.u32 = val->u32; | |
| break; | |
| case 64: | |
| cmp_val.u64 = val->u64; | |
| break; | |
| default: | |
| validate_assert(state, !"Invalid load_const bit size"); | |
| } | |
| } | |
| validate_assert(state, memcmp(val, &cmp_val, sizeof(cmp_val)) == 0); | |
| } | |
| static void | |
| validate_load_const_instr(nir_load_const_instr *instr, validate_state *state) | |
| { | |
| validate_def(&instr->def, state); | |
| for (unsigned i = 0; i < instr->def.num_components; i++) | |
| validate_const_value(&instr->value[i], instr->def.bit_size, false, state); | |
| } | |
| static void | |
| validate_ssa_undef_instr(nir_undef_instr *instr, validate_state *state) | |
| { | |
| validate_def(&instr->def, state); | |
| } | |
| static void | |
| validate_phi_instr(nir_phi_instr *instr, validate_state *state) | |
| { | |
| /* | |
| * don't validate the sources until we get to them from their predecessor | |
| * basic blocks, to avoid validating an SSA use before its definition. | |
| */ | |
| validate_def(&instr->def, state); | |
| exec_list_validate(&instr->srcs); | |
| validate_assert(state, exec_list_length(&instr->srcs) == | |
| nir_block_num_preds(state->block)); | |
| } | |
| static void | |
| validate_jump_instr(nir_jump_instr *instr, validate_state *state) | |
| { | |
| nir_block *block = state->block; | |
| validate_assert(state, &instr->instr == nir_block_last_instr(block)); | |
| switch (instr->type) { | |
| case nir_jump_return: | |
| case nir_jump_halt: | |
| case nir_jump_abort: | |
| validate_assert(state, block->successors[0] == state->impl->end_block); | |
| validate_assert(state, block->successors[1] == NULL); | |
| validate_assert(state, instr->target == NULL); | |
| validate_assert(state, instr->else_target == NULL); | |
| validate_assert(state, !state->in_loop_continue_construct); | |
| break; | |
| case nir_jump_break: | |
| validate_assert(state, state->impl->structured); | |
| validate_assert(state, state->loop != NULL); | |
| if (state->loop) { | |
| nir_block *after = | |
| nir_cf_node_as_block(nir_cf_node_next(&state->loop->cf_node)); | |
| validate_assert(state, block->successors[0] == after); | |
| } | |
| validate_assert(state, block->successors[1] == NULL); | |
| validate_assert(state, instr->target == NULL); | |
| validate_assert(state, instr->else_target == NULL); | |
| break; | |
| case nir_jump_continue: | |
| validate_assert(state, state->impl->structured); | |
| validate_assert(state, state->loop != NULL); | |
| if (state->loop) { | |
| validate_assert(state, nir_loop_has_continue_construct(state->loop)); | |
| if (nir_loop_has_continue_construct(state->loop)) { | |
| nir_block *cont_block = nir_loop_first_continue_block(state->loop); | |
| validate_assert(state, block->successors[0] == cont_block); | |
| } | |
| } | |
| validate_assert(state, block->successors[1] == NULL); | |
| validate_assert(state, instr->target == NULL); | |
| validate_assert(state, instr->else_target == NULL); | |
| validate_assert(state, !state->in_loop_continue_construct); | |
| break; | |
| case nir_jump_goto: | |
| validate_assert(state, !state->impl->structured); | |
| validate_assert(state, instr->target == block->successors[0]); | |
| validate_assert(state, instr->target != NULL); | |
| validate_assert(state, instr->else_target == NULL); | |
| break; | |
| case nir_jump_goto_if: | |
| validate_assert(state, !state->impl->structured); | |
| validate_assert(state, instr->target == block->successors[1]); | |
| validate_assert(state, instr->else_target == block->successors[0]); | |
| validate_sized_src(&instr->condition, state, 0, 1); | |
| validate_assert(state, instr->target != NULL); | |
| validate_assert(state, instr->else_target != NULL); | |
| break; | |
| default: | |
| validate_assert(state, !"Invalid jump instruction type"); | |
| break; | |
| } | |
| } | |
| static void | |
| validate_instr(nir_instr *instr, validate_state *state) | |
| { | |
| validate_assert(state, instr->block == state->block); | |
| state->instr = instr; | |
| switch (instr->type) { | |
| case nir_instr_type_alu: | |
| validate_alu_instr(nir_instr_as_alu(instr), state); | |
| break; | |
| case nir_instr_type_deref: | |
| validate_deref_instr(nir_instr_as_deref(instr), state); | |
| break; | |
| case nir_instr_type_call: | |
| validate_call_instr(nir_instr_as_call(instr), state); | |
| break; | |
| case nir_instr_type_cmat_call: | |
| validate_cmat_call_instr(nir_instr_as_cmat_call(instr), state); | |
| break; | |
| case nir_instr_type_intrinsic: | |
| validate_intrinsic_instr(nir_instr_as_intrinsic(instr), state); | |
| break; | |
| case nir_instr_type_tex: | |
| validate_tex_instr(nir_instr_as_tex(instr), state); | |
| break; | |
| case nir_instr_type_load_const: | |
| validate_load_const_instr(nir_instr_as_load_const(instr), state); | |
| break; | |
| case nir_instr_type_phi: | |
| validate_phi_instr(nir_instr_as_phi(instr), state); | |
| break; | |
| case nir_instr_type_undef: | |
| validate_ssa_undef_instr(nir_instr_as_undef(instr), state); | |
| break; | |
| case nir_instr_type_jump: | |
| validate_jump_instr(nir_instr_as_jump(instr), state); | |
| break; | |
| default: | |
| validate_assert(state, !"Invalid ALU instruction type"); | |
| break; | |
| } | |
| state->instr = NULL; | |
| } | |
| static void | |
| validate_phi_src(nir_phi_instr *instr, nir_block *pred, validate_state *state) | |
| { | |
| state->instr = &instr->instr; | |
| exec_list_validate(&instr->srcs); | |
| nir_foreach_phi_src(src, instr) { | |
| if (src->pred == pred) { | |
| validate_sized_src(&src->src, state, instr->def.bit_size, | |
| instr->def.num_components); | |
| state->instr = NULL; | |
| return; | |
| } | |
| } | |
| validate_assert(state, !"Phi does not have a source corresponding to one " | |
| "of its predecessor blocks"); | |
| } | |
| static void | |
| validate_phi_srcs(nir_block *block, nir_block *succ, validate_state *state) | |
| { | |
| nir_foreach_phi(phi, succ) { | |
| validate_phi_src(phi, block, state); | |
| } | |
| } | |
| static void | |
| collect_blocks(struct exec_list *cf_list, validate_state *state) | |
| { | |
| /* We walk the blocks manually here rather than using nir_foreach_block for | |
| * a few reasons: | |
| * | |
| * 1. We want to call exec_list_validate() on every linked list in the IR | |
| * which means we need to touch every linked and just walking blocks | |
| * with nir_foreach_block() would make that difficult. In particular, | |
| * we want to validate each list before the first time we walk it so | |
| * that we catch broken lists in exec_list_validate() instead of | |
| * getting stuck in a hard-to-debug infinite loop in the validator. | |
| * | |
| * 2. nir_foreach_block() depends on several invariants of the CF node | |
| * hierarchy which nir_validate_shader() is responsible for verifying. | |
| * If we used nir_foreach_block() in nir_validate_shader(), we could | |
| * end up blowing up on a bad list walk instead of throwing the much | |
| * easier to debug validation error. | |
| */ | |
| exec_list_validate(cf_list); | |
| foreach_list_typed(nir_cf_node, node, node, cf_list) { | |
| switch (node->type) { | |
| case nir_cf_node_block: | |
| _mesa_set_add(state->blocks, nir_cf_node_as_block(node)); | |
| break; | |
| case nir_cf_node_if: | |
| collect_blocks(&nir_cf_node_as_if(node)->then_list, state); | |
| collect_blocks(&nir_cf_node_as_if(node)->else_list, state); | |
| break; | |
| case nir_cf_node_loop: | |
| collect_blocks(&nir_cf_node_as_loop(node)->body, state); | |
| collect_blocks(&nir_cf_node_as_loop(node)->continue_list, state); | |
| break; | |
| default: | |
| UNREACHABLE("Invalid CF node type"); | |
| } | |
| } | |
| } | |
| static void | |
| collect_blocks_pdfs(nir_function_impl *impl, nir_block *block, | |
| uint32_t *count, validate_state *state) | |
| { | |
| if (block == impl->end_block) | |
| return; | |
| if (_mesa_set_search(state->blocks, block)) | |
| return; | |
| _mesa_set_add(state->blocks, block); | |
| for (uint32_t i = 0; i < ARRAY_SIZE(block->successors); i++) { | |
| if (block->successors[i] != NULL) | |
| collect_blocks_pdfs(impl, block->successors[i], count, state); | |
| } | |
| /* Assert that the blocks are indexed in reverse PDFS order */ | |
| validate_assert(state, block->index == --(*count)); | |
| } | |
| static void | |
| collect_unstructured_blocks(nir_function_impl *impl, validate_state *state) | |
| { | |
| exec_list_validate(&impl->body); | |
| /* Assert that the blocks are properly indexed */ | |
| uint32_t count = 0; | |
| foreach_list_typed(nir_cf_node, node, node, &impl->body) { | |
| nir_block *block = nir_cf_node_as_block(node); | |
| validate_assert(state, block->index == count++); | |
| } | |
| validate_assert(state, impl->end_block->index == count); | |
| collect_blocks_pdfs(impl, nir_start_block(impl), &count, state); | |
| } | |
| static void validate_cf_node(nir_cf_node *node, validate_state *state); | |
| static void | |
| validate_block_predecessors(nir_block *block, validate_state *state) | |
| { | |
| for (unsigned i = 0; i < 2; i++) { | |
| if (block->successors[i] == NULL) | |
| continue; | |
| /* The block has to exist in the nir_function_impl */ | |
| validate_assert(state, _mesa_set_search(state->blocks, | |
| block->successors[i])); | |
| /* And we have to be in our successor's predecessors set */ | |
| bool has_pred = false; | |
| nir_foreach_pred(pred, block->successors[i]) | |
| has_pred |= pred == block; | |
| validate_assert(state, has_pred); | |
| validate_phi_srcs(block, block->successors[i], state); | |
| } | |
| /* The start block cannot have any predecessors */ | |
| if (block == nir_start_block(state->impl)) | |
| validate_assert(state, nir_block_num_preds(block) == 0); | |
| /* Check for duplicate predecessors. */ | |
| nir_foreach_pred(pred, block) { | |
| bool found = false; | |
| nir_foreach_pred(pred2, block) { | |
| if (pred == pred2) { | |
| validate_assert(state, !found); | |
| found = true; | |
| } | |
| } | |
| } | |
| nir_foreach_pred(pred, block) { | |
| validate_assert(state, _mesa_set_search(state->blocks, pred)); | |
| validate_assert(state, pred->successors[0] == block || | |
| pred->successors[1] == block); | |
| } | |
| } | |
| static void | |
| validate_block(nir_block *block, validate_state *state) | |
| { | |
| validate_assert(state, block->cf_node.parent == state->parent_node); | |
| validate_assert(state, block->impl == state->impl); | |
| state->block = block; | |
| exec_list_validate(&block->instr_list); | |
| nir_foreach_instr(instr, block) { | |
| if (instr->type == nir_instr_type_phi) { | |
| validate_assert(state, instr == nir_block_first_instr(block) || | |
| nir_instr_prev(instr)->type == nir_instr_type_phi); | |
| } | |
| validate_instr(instr, state); | |
| } | |
| validate_assert(state, block->successors[0] != NULL); | |
| validate_assert(state, block->successors[0] != block->successors[1]); | |
| validate_block_predecessors(block, state); | |
| if (!state->impl->structured) { | |
| validate_assert(state, nir_block_ends_in_jump(block)); | |
| } else if (!nir_block_ends_in_jump(block)) { | |
| nir_cf_node *next = nir_cf_node_next(&block->cf_node); | |
| if (next == NULL) { | |
| switch (state->parent_node->type) { | |
| case nir_cf_node_loop: { | |
| if (!nir_loop_has_continue_construct(state->loop) || | |
| block == nir_loop_last_continue_block(state->loop)) { | |
| nir_block *header = nir_loop_first_block(state->loop); | |
| validate_assert(state, block->successors[0] == header); | |
| } else { | |
| nir_block *cont = nir_loop_first_continue_block(state->loop); | |
| validate_assert(state, block->successors[0] == cont); | |
| } | |
| /* due to the hack for infinite loops, block->successors[1] may | |
| * point to the block after the loop. | |
| */ | |
| break; | |
| } | |
| case nir_cf_node_if: { | |
| nir_block *after = | |
| nir_cf_node_as_block(nir_cf_node_next(state->parent_node)); | |
| validate_assert(state, block->successors[0] == after); | |
| validate_assert(state, block->successors[1] == NULL); | |
| break; | |
| } | |
| case nir_cf_node_function: | |
| validate_assert(state, block->successors[0] == state->impl->end_block); | |
| validate_assert(state, block->successors[1] == NULL); | |
| break; | |
| default: | |
| UNREACHABLE("unknown control flow node type"); | |
| } | |
| } else { | |
| if (next->type == nir_cf_node_if) { | |
| nir_if *if_stmt = nir_cf_node_as_if(next); | |
| validate_assert(state, block->successors[0] == | |
| nir_if_first_then_block(if_stmt)); | |
| validate_assert(state, block->successors[1] == | |
| nir_if_first_else_block(if_stmt)); | |
| } else if (next->type == nir_cf_node_loop) { | |
| nir_loop *loop = nir_cf_node_as_loop(next); | |
| validate_assert(state, block->successors[0] == | |
| nir_loop_first_block(loop)); | |
| validate_assert(state, block->successors[1] == NULL); | |
| } else { | |
| validate_assert(state, | |
| !"Structured NIR cannot have consecutive blocks"); | |
| } | |
| } | |
| } | |
| } | |
| static void | |
| validate_end_block(nir_block *block, validate_state *state) | |
| { | |
| validate_assert(state, block->cf_node.parent == &state->impl->cf_node); | |
| exec_list_validate(&block->instr_list); | |
| validate_assert(state, exec_list_is_empty(&block->instr_list)); | |
| validate_assert(state, block->successors[0] == NULL); | |
| validate_assert(state, block->successors[1] == NULL); | |
| validate_block_predecessors(block, state); | |
| } | |
| static void | |
| validate_if(nir_if *if_stmt, validate_state *state) | |
| { | |
| validate_assert(state, state->impl->structured); | |
| state->if_stmt = if_stmt; | |
| validate_assert(state, !exec_node_is_head_sentinel(if_stmt->cf_node.node.prev)); | |
| nir_cf_node *prev_node = nir_cf_node_prev(&if_stmt->cf_node); | |
| validate_assert(state, prev_node->type == nir_cf_node_block); | |
| validate_assert(state, !exec_node_is_tail_sentinel(if_stmt->cf_node.node.next)); | |
| nir_cf_node *next_node = nir_cf_node_next(&if_stmt->cf_node); | |
| validate_assert(state, next_node->type == nir_cf_node_block); | |
| validate_assert(state, nir_src_is_if(&if_stmt->condition)); | |
| validate_if_src(&if_stmt->condition, state); | |
| validate_assert(state, !exec_list_is_empty(&if_stmt->then_list)); | |
| validate_assert(state, !exec_list_is_empty(&if_stmt->else_list)); | |
| nir_cf_node *old_parent = state->parent_node; | |
| state->parent_node = &if_stmt->cf_node; | |
| foreach_list_typed(nir_cf_node, cf_node, node, &if_stmt->then_list) { | |
| validate_cf_node(cf_node, state); | |
| } | |
| foreach_list_typed(nir_cf_node, cf_node, node, &if_stmt->else_list) { | |
| validate_cf_node(cf_node, state); | |
| } | |
| state->parent_node = old_parent; | |
| state->if_stmt = NULL; | |
| } | |
| static void | |
| validate_loop(nir_loop *loop, validate_state *state) | |
| { | |
| validate_assert(state, state->impl->structured); | |
| validate_assert(state, !exec_node_is_head_sentinel(loop->cf_node.node.prev)); | |
| nir_cf_node *prev_node = nir_cf_node_prev(&loop->cf_node); | |
| validate_assert(state, prev_node->type == nir_cf_node_block); | |
| validate_assert(state, !exec_node_is_tail_sentinel(loop->cf_node.node.next)); | |
| nir_cf_node *next_node = nir_cf_node_next(&loop->cf_node); | |
| validate_assert(state, next_node->type == nir_cf_node_block); | |
| validate_assert(state, !exec_list_is_empty(&loop->body)); | |
| validate_assert(state, nir_block_num_preds(nir_loop_first_block(loop)) <= 2); | |
| nir_cf_node *old_parent = state->parent_node; | |
| state->parent_node = &loop->cf_node; | |
| nir_loop *old_loop = state->loop; | |
| bool old_continue_construct = state->in_loop_continue_construct; | |
| state->loop = loop; | |
| state->in_loop_continue_construct = false; | |
| foreach_list_typed(nir_cf_node, cf_node, node, &loop->body) { | |
| validate_cf_node(cf_node, state); | |
| } | |
| state->in_loop_continue_construct = true; | |
| foreach_list_typed(nir_cf_node, cf_node, node, &loop->continue_list) { | |
| validate_cf_node(cf_node, state); | |
| } | |
| state->in_loop_continue_construct = false; | |
| state->parent_node = old_parent; | |
| state->loop = old_loop; | |
| state->in_loop_continue_construct = old_continue_construct; | |
| } | |
| static void | |
| validate_cf_node(nir_cf_node *node, validate_state *state) | |
| { | |
| validate_assert(state, node->parent == state->parent_node); | |
| switch (node->type) { | |
| case nir_cf_node_block: | |
| validate_block(nir_cf_node_as_block(node), state); | |
| break; | |
| case nir_cf_node_if: | |
| validate_if(nir_cf_node_as_if(node), state); | |
| break; | |
| case nir_cf_node_loop: | |
| validate_loop(nir_cf_node_as_loop(node), state); | |
| break; | |
| default: | |
| UNREACHABLE("Invalid CF node type"); | |
| } | |
| } | |
| static void | |
| validate_constant(nir_constant *c, const struct glsl_type *type, | |
| validate_state *state) | |
| { | |
| if (glsl_type_is_vector_or_scalar(type)) { | |
| unsigned num_components = glsl_get_vector_elements(type); | |
| unsigned bit_size = glsl_get_bit_size(type); | |
| for (unsigned i = 0; i < num_components; i++) | |
| validate_const_value(&c->values[i], bit_size, c->is_null_constant, state); | |
| for (unsigned i = num_components; i < NIR_MAX_VEC_COMPONENTS; i++) | |
| validate_assert(state, c->values[i].u64 == 0); | |
| } else { | |
| validate_assert(state, c->num_elements == glsl_get_length(type)); | |
| if (glsl_type_is_struct_or_ifc(type)) { | |
| for (unsigned i = 0; i < c->num_elements; i++) { | |
| const struct glsl_type *elem_type = glsl_get_struct_field(type, i); | |
| validate_constant(c->elements[i], elem_type, state); | |
| validate_assert(state, !c->is_null_constant || c->elements[i]->is_null_constant); | |
| } | |
| } else if (glsl_type_is_array_or_matrix(type)) { | |
| const struct glsl_type *elem_type = glsl_get_array_element(type); | |
| for (unsigned i = 0; i < c->num_elements; i++) { | |
| validate_constant(c->elements[i], elem_type, state); | |
| validate_assert(state, !c->is_null_constant || c->elements[i]->is_null_constant); | |
| } | |
| } else { | |
| validate_assert(state, !"Invalid type for nir_constant"); | |
| } | |
| } | |
| } | |
| static void | |
| validate_var_decl(nir_variable *var, nir_variable_mode valid_modes, | |
| validate_state *state) | |
| { | |
| state->var = var; | |
| /* Must have exactly one mode set */ | |
| validate_assert(state, util_is_power_of_two_nonzero(var->data.mode)); | |
| validate_assert(state, var->data.mode & valid_modes); | |
| if (var->data.compact) { | |
| /* The "compact" flag is only valid on arrays of scalars. */ | |
| assert(glsl_type_is_array(var->type)); | |
| const struct glsl_type *type = glsl_get_array_element(var->type); | |
| if (nir_is_arrayed_io(var, state->shader->info.stage)) { | |
| assert(glsl_type_is_array(type)); | |
| assert(glsl_type_is_scalar(glsl_get_array_element(type))); | |
| } else { | |
| assert(glsl_type_is_scalar(type)); | |
| } | |
| } | |
| if (var->num_members > 0) { | |
| const struct glsl_type *without_array = glsl_without_array(var->type); | |
| validate_assert(state, glsl_type_is_struct_or_ifc(without_array)); | |
| validate_assert(state, var->num_members == glsl_get_length(without_array)); | |
| validate_assert(state, var->members != NULL); | |
| } | |
| if (var->data.per_view) | |
| validate_assert(state, glsl_type_is_array(var->type)); | |
| if (var->constant_initializer) | |
| validate_constant(var->constant_initializer, var->type, state); | |
| if (var->data.mode == nir_var_image) { | |
| validate_assert(state, !var->data.bindless); | |
| validate_assert(state, glsl_type_is_image(glsl_without_array(var->type))); | |
| } | |
| if (var->data.per_vertex) | |
| validate_assert(state, state->shader->info.stage == MESA_SHADER_FRAGMENT); | |
| /* | |
| * TODO validate some things ir_validate.cpp does (requires more GLSL type | |
| * support) | |
| */ | |
| _mesa_hash_table_insert(state->var_defs, var, | |
| valid_modes == nir_var_function_temp ? state->impl : NULL); | |
| state->var = NULL; | |
| } | |
| static bool | |
| validate_ssa_def_dominance(nir_def *def, void *_state) | |
| { | |
| validate_state *state = _state; | |
| validate_assert(state, def->index < state->impl->ssa_alloc); | |
| validate_assert(state, !BITSET_TEST(state->ssa_defs_found, def->index)); | |
| BITSET_SET(state->ssa_defs_found, def->index); | |
| return true; | |
| } | |
| static bool | |
| validate_src_dominance(nir_src *src, void *_state) | |
| { | |
| validate_state *state = _state; | |
| if (nir_def_block(src->ssa) == nir_src_use_instr(src)->block) { | |
| validate_assert(state, src->ssa->index < state->impl->ssa_alloc); | |
| validate_assert(state, BITSET_TEST(state->ssa_defs_found, | |
| src->ssa->index)); | |
| } else { | |
| validate_assert(state, nir_block_dominates(nir_def_block(src->ssa), | |
| nir_src_use_instr(src)->block)); | |
| } | |
| return true; | |
| } | |
| static void | |
| validate_ssa_dominance(nir_function_impl *impl, validate_state *state) | |
| { | |
| nir_metadata_require(impl, nir_metadata_dominance); | |
| nir_foreach_block_unstructured(block, impl) { | |
| state->block = block; | |
| nir_foreach_instr(instr, block) { | |
| state->instr = instr; | |
| if (instr->type == nir_instr_type_phi) { | |
| nir_phi_instr *phi = nir_instr_as_phi(instr); | |
| nir_foreach_phi_src(src, phi) { | |
| validate_assert(state, | |
| nir_block_dominates(nir_def_block(src->src.ssa), | |
| src->pred)); | |
| } | |
| } else { | |
| nir_foreach_src(instr, validate_src_dominance, state); | |
| } | |
| nir_foreach_def(instr, validate_ssa_def_dominance, state); | |
| } | |
| nir_if *nif = nir_block_get_following_if(block); | |
| if (nif) { | |
| validate_assert(state, nir_block_dominates(nir_def_block(nif->condition.ssa), | |
| block)); | |
| } | |
| } | |
| } | |
| static void | |
| validate_block_index(nir_function_impl *impl, validate_state *state) | |
| { | |
| unsigned index = 0; | |
| nir_foreach_block_unstructured(block, impl) { | |
| state->block = block; | |
| validate_assert(state, block->index == index); | |
| index++; | |
| } | |
| state->block = NULL; | |
| validate_assert(state, impl->num_blocks == index); | |
| validate_assert(state, impl->end_block->index == impl->num_blocks); | |
| } | |
| static void | |
| validate_instr_index(nir_function_impl *impl, validate_state *state) | |
| { | |
| int index = -1; | |
| nir_foreach_block(block, impl) { | |
| state->block = block; | |
| validate_assert(state, (int)block->start_ip > index); | |
| index = block->start_ip; | |
| nir_foreach_instr(instr, block) { | |
| state->instr = instr; | |
| validate_assert(state, instr->index > index); | |
| index = instr->index; | |
| } | |
| state->instr = NULL; | |
| validate_assert(state, block->end_ip > index); | |
| index = block->end_ip; | |
| } | |
| state->block = NULL; | |
| } | |
| typedef struct { | |
| uint32_t index; | |
| unsigned num_dom_children; | |
| struct nir_block **dom_children; | |
| struct set dom_frontier; | |
| uint32_t dom_pre_index, dom_post_index; | |
| } block_dom_metadata; | |
| static void | |
| validate_dominance(nir_function_impl *impl, validate_state *state) | |
| { | |
| nir_metadata valid_metadata = impl->valid_metadata; | |
| /* Preserve dominance */ | |
| block_dom_metadata *blocks = ralloc_array(state->mem_ctx, block_dom_metadata, | |
| state->blocks->size); | |
| set_foreach(state->blocks, entry) { | |
| nir_block *block = (nir_block *)entry->key; | |
| block_dom_metadata *md = &blocks[entry - state->blocks->table]; | |
| md->index = block->index; | |
| md->num_dom_children = block->num_dom_children; | |
| md->dom_pre_index = block->dom_pre_index; | |
| md->dom_post_index = block->dom_post_index; | |
| md->dom_children = block->dom_children; | |
| md->dom_frontier = block->dom_frontier; | |
| memset(&block->dom_frontier, 0, sizeof(struct set)); | |
| block->dom_children = NULL; | |
| _mesa_pointer_set_init(&block->dom_frontier, block); | |
| } | |
| /* Call metadata passes and compare it against the preserved metadata and call SSA dominance | |
| * validation. Dominance requires block indices, but we should ignore the current ones, since | |
| * we don't trust them. | |
| */ | |
| impl->valid_metadata &= ~(nir_metadata_block_index | nir_metadata_dominance); | |
| nir_metadata_require(impl, nir_metadata_block_index | nir_metadata_dominance); | |
| assert(impl->valid_metadata == (valid_metadata | nir_metadata_block_index | nir_metadata_dominance)); | |
| if (valid_metadata & nir_metadata_dominance) { | |
| set_foreach(state->blocks, entry) { | |
| nir_block *block = (nir_block *)entry->key; | |
| block_dom_metadata *md = &blocks[entry - state->blocks->table]; | |
| state->block = (nir_block *)block; | |
| validate_assert(state, block->num_dom_children == md->num_dom_children); | |
| validate_assert(state, block->dom_pre_index == md->dom_pre_index); | |
| validate_assert(state, block->dom_post_index == md->dom_post_index); | |
| if (block->num_dom_children == md->num_dom_children && block->num_dom_children) { | |
| validate_assert(state, !memcmp(block->dom_children, md->dom_children, | |
| block->num_dom_children * sizeof(md->dom_children[0]))); | |
| } | |
| validate_assert(state, block->dom_frontier.entries == md->dom_frontier.entries); | |
| set_foreach(&block->dom_frontier, entry) { | |
| validate_assert(state, _mesa_set_search_pre_hashed(&md->dom_frontier, | |
| entry->hash, entry->key)); | |
| } | |
| } | |
| state->block = NULL; | |
| } | |
| memset(state->ssa_defs_found, 0, BITSET_BYTES(impl->ssa_alloc)); | |
| validate_ssa_dominance(impl, state); | |
| /* Restore the old dominance metadata */ | |
| set_foreach(state->blocks, entry) { | |
| nir_block *block = (nir_block *)entry->key; | |
| block_dom_metadata *md = &blocks[entry - state->blocks->table]; | |
| if (block->dom_children != block->_dom_children_storage) | |
| ralloc_free(block->dom_children); | |
| _mesa_set_fini(&block->dom_frontier, NULL); | |
| block->index = md->index; | |
| block->num_dom_children = md->num_dom_children; | |
| block->dom_pre_index = md->dom_pre_index; | |
| block->dom_post_index = md->dom_post_index; | |
| block->dom_children = md->dom_children; | |
| block->dom_frontier = md->dom_frontier; | |
| } | |
| ralloc_free(blocks); | |
| impl->valid_metadata = valid_metadata; | |
| } | |
| typedef struct { | |
| struct u_sparse_bitset live_in; | |
| struct u_sparse_bitset live_out; | |
| } block_liveness_metadata; | |
| static void | |
| validate_live_defs(nir_function_impl *impl, validate_state *state) | |
| { | |
| nir_metadata valid_metadata = impl->valid_metadata; | |
| /* Preserve live defs */ | |
| block_liveness_metadata *blocks = ralloc_array(state->mem_ctx, block_liveness_metadata, | |
| state->blocks->size); | |
| set_foreach(state->blocks, entry) { | |
| nir_block *block = (nir_block *)entry->key; | |
| block_liveness_metadata *md = &blocks[entry - state->blocks->table]; | |
| md->live_in = block->live_in; | |
| md->live_out = block->live_out; | |
| u_sparse_bitset_init(&block->live_in, impl->ssa_alloc, NULL); | |
| u_sparse_bitset_init(&block->live_out, impl->ssa_alloc, NULL); | |
| } | |
| /* Call metadata passes and compare it against the preserved metadata */ | |
| impl->valid_metadata &= ~nir_metadata_live_defs; | |
| nir_metadata_require(impl, nir_metadata_live_defs); | |
| assert(impl->valid_metadata == valid_metadata); | |
| set_foreach(state->blocks, entry) { | |
| nir_block *block = (nir_block *)entry->key; | |
| block_liveness_metadata *md = &blocks[entry - state->blocks->table]; | |
| state->block = (nir_block *)block; | |
| if (block == impl->end_block) | |
| continue; | |
| size_t bitset_words = BITSET_WORDS(impl->ssa_alloc); | |
| if (bitset_words) { | |
| validate_assert(state, !u_sparse_bitset_cmp(&md->live_in, &block->live_in)); | |
| validate_assert(state, !u_sparse_bitset_cmp(&md->live_out, &block->live_out)); | |
| } | |
| } | |
| state->block = NULL; | |
| /* Restore the old live defs metadata */ | |
| set_foreach(state->blocks, entry) { | |
| nir_block *block = (nir_block *)entry->key; | |
| block_liveness_metadata *md = &blocks[entry - state->blocks->table]; | |
| u_sparse_bitset_free(&block->live_in); | |
| u_sparse_bitset_free(&block->live_out); | |
| block->live_in = md->live_in; | |
| block->live_out = md->live_out; | |
| } | |
| ralloc_free(blocks); | |
| } | |
| typedef struct { | |
| uint32_t index; | |
| bool divergent; | |
| bool divergent_break; | |
| bool divergent_continue; | |
| bool is_loop_header; | |
| } block_divergence_metadata; | |
| static void | |
| validate_divergence(nir_function_impl *impl, validate_state *state) | |
| { | |
| nir_metadata valid_metadata = impl->valid_metadata; | |
| /* Preserve divergence information. */ | |
| unsigned num_blocks = impl->num_blocks; | |
| block_divergence_metadata *blocks = ralloc_array(state->mem_ctx, | |
| block_divergence_metadata, | |
| state->blocks->size); | |
| BITSET_WORD *ssa_divergence = BITSET_RZALLOC(state->mem_ctx, | |
| impl->ssa_alloc); | |
| BITSET_WORD *loop_invariance = BITSET_RZALLOC(state->mem_ctx, | |
| impl->ssa_alloc); | |
| set_foreach(state->blocks, entry) { | |
| nir_block *block = (nir_block *)entry->key; | |
| block_divergence_metadata *md = &blocks[entry - state->blocks->table]; | |
| md->index = block->index; | |
| md->divergent = block->divergent; | |
| md->is_loop_header = false; | |
| if (block->cf_node.parent->type == nir_cf_node_loop && | |
| nir_cf_node_is_first(&block->cf_node)) { | |
| md->is_loop_header = true; | |
| nir_loop *loop = nir_cf_node_as_loop(block->cf_node.parent); | |
| md->divergent_break = loop->divergent_break; | |
| md->divergent_continue = loop->divergent_continue; | |
| } | |
| nir_foreach_instr(instr, block) { | |
| nir_def *def = nir_instr_def(instr); | |
| if (def && def->divergent) | |
| BITSET_SET(ssa_divergence, def->index); | |
| if (def && def->loop_invariant) | |
| BITSET_SET(loop_invariance, def->index); | |
| } | |
| } | |
| /* Call metadata passes and compare it against the preserved metadata */ | |
| impl->valid_metadata &= ~nir_metadata_divergence; | |
| nir_metadata_require(impl, nir_metadata_divergence); | |
| assert(impl->valid_metadata == (valid_metadata | nir_metadata_block_index)); | |
| set_foreach(state->blocks, entry) { | |
| nir_block *block = (nir_block *)entry->key; | |
| block_divergence_metadata *md = &blocks[entry - state->blocks->table]; | |
| state->block = (nir_block *)block; | |
| validate_assert(state, block->divergent == md->divergent); | |
| if (md->is_loop_header) { | |
| nir_loop *loop = nir_cf_node_as_loop(block->cf_node.parent); | |
| validate_assert(state, loop->divergent_break == md->divergent_break); | |
| validate_assert(state, loop->divergent_continue == md->divergent_continue); | |
| } | |
| nir_foreach_instr(instr, block) { | |
| nir_def *def = nir_instr_def(instr); | |
| if (def) { | |
| state->instr = instr; | |
| validate_assert(state, def->divergent == BITSET_TEST(ssa_divergence, def->index)); | |
| validate_assert(state, def->loop_invariant == BITSET_TEST(loop_invariance, def->index)); | |
| } | |
| } | |
| state->instr = NULL; | |
| } | |
| state->block = NULL; | |
| /* Restore the old divergence metadata */ | |
| set_foreach(state->blocks, entry) { | |
| nir_block *block = (nir_block *)entry->key; | |
| block_divergence_metadata *md = &blocks[entry - state->blocks->table]; | |
| block->index = md->index; | |
| block->divergent = md->divergent; | |
| if (md->is_loop_header) { | |
| nir_loop *loop = nir_cf_node_as_loop(block->cf_node.parent); | |
| loop->divergent_break = md->divergent_break; | |
| loop->divergent_continue = md->divergent_continue; | |
| } | |
| nir_foreach_instr(instr, block) { | |
| nir_def *def = nir_instr_def(instr); | |
| if (def) { | |
| def->divergent = BITSET_TEST(ssa_divergence, def->index); | |
| def->loop_invariant = BITSET_TEST(loop_invariance, def->index); | |
| } | |
| } | |
| } | |
| impl->num_blocks = num_blocks; | |
| ralloc_free(blocks); | |
| ralloc_free(ssa_divergence); | |
| ralloc_free(loop_invariance); | |
| impl->valid_metadata = valid_metadata; | |
| } | |
| static bool | |
| are_loop_terminators_equal(const nir_loop_terminator *a, const nir_loop_terminator *b) | |
| { | |
| if (!a || !b) | |
| return !a && !b; | |
| return a->nif == b->nif && | |
| a->conditional_instr == b->conditional_instr && | |
| a->break_block == b->break_block && | |
| a->continue_from_block == b->continue_from_block && | |
| a->continue_from_then == b->continue_from_then && | |
| a->induction_rhs == b->induction_rhs && | |
| a->exact_trip_count_unknown == b->exact_trip_count_unknown; | |
| } | |
| static void | |
| validate_loop_info(nir_function_impl *impl, validate_state *state) | |
| { | |
| nir_metadata valid_metadata = impl->valid_metadata; | |
| /* Preserve loop info */ | |
| struct hash_table *loops = _mesa_pointer_hash_table_create(state->mem_ctx); | |
| set_foreach(state->blocks, entry) { | |
| nir_block *block = (nir_block *)entry->key; | |
| if (block->cf_node.parent->type == nir_cf_node_loop && | |
| nir_cf_node_is_first(&block->cf_node)) { | |
| nir_loop *loop = nir_cf_node_as_loop(block->cf_node.parent); | |
| _mesa_hash_table_insert(loops, loop, loop->info); | |
| loop->info = NULL; | |
| } | |
| } | |
| /* Call metadata passes and compare it against the preserved metadata */ | |
| impl->valid_metadata &= ~nir_metadata_loop_analysis; | |
| nir_metadata_require(impl, nir_metadata_loop_analysis, impl->loop_analysis_indirect_mask, | |
| impl->loop_analysis_force_unroll_sampler_indirect); | |
| assert(impl->valid_metadata == valid_metadata); | |
| hash_table_foreach(loops, entry) { | |
| const nir_loop *loop = entry->key; | |
| const nir_loop_info *md = entry->data; | |
| validate_assert(state, loop->info->instr_cost == md->instr_cost); | |
| validate_assert(state, loop->info->has_soft_fp64 == md->has_soft_fp64); | |
| validate_assert(state, loop->info->guessed_trip_count == md->guessed_trip_count); | |
| validate_assert(state, loop->info->max_trip_count == md->max_trip_count); | |
| validate_assert(state, loop->info->exact_trip_count_known == md->exact_trip_count_known); | |
| validate_assert(state, loop->info->force_unroll == md->force_unroll); | |
| validate_assert(state, loop->info->complex_loop == md->complex_loop); | |
| validate_assert(state, are_loop_terminators_equal(loop->info->limiting_terminator, | |
| md->limiting_terminator)); | |
| validate_assert(state, list_length(&loop->info->loop_terminator_list) == | |
| list_length(&md->loop_terminator_list)); | |
| list_pair_for_each_entry(nir_loop_terminator, a, b, &loop->info->loop_terminator_list, | |
| &md->loop_terminator_list, loop_terminator_link) { | |
| validate_assert(state, are_loop_terminators_equal(a, b)); | |
| } | |
| validate_assert(state, _mesa_hash_table_num_entries(loop->info->induction_vars) == | |
| _mesa_hash_table_num_entries(md->induction_vars)); | |
| hash_table_foreach(loop->info->induction_vars, var) { | |
| struct hash_entry *prev_var = _mesa_hash_table_search(md->induction_vars, var->key); | |
| validate_assert(state, prev_var != NULL); | |
| if (prev_var) { | |
| nir_loop_induction_variable *a = var->data; | |
| nir_loop_induction_variable *b = prev_var->data; | |
| validate_assert(state, a->basis == b->basis); | |
| validate_assert(state, a->def == b->def); | |
| validate_assert(state, a->init_src == b->init_src); | |
| validate_assert(state, a->update_src == b->update_src); | |
| } | |
| } | |
| } | |
| /* Restore the old loop info */ | |
| hash_table_foreach(loops, entry) { | |
| nir_loop *loop = (nir_loop *)entry->key; | |
| nir_loop_info *md = entry->data; | |
| ralloc_free(loop->info); | |
| loop->info = md; | |
| } | |
| } | |
| static void | |
| validate_metadata_and_ssa_dominance(nir_function_impl *impl, validate_state *state) | |
| { | |
| /* We should preserve and restore metadata when necessary, so that passes do not accidentally | |
| * depend on nir_validate_shader(). | |
| */ | |
| if (impl->valid_metadata & nir_metadata_block_index) | |
| validate_block_index(impl, state); | |
| if (impl->valid_metadata & nir_metadata_instr_index) | |
| validate_instr_index(impl, state); | |
| /* This validates both metadata and SSA dominance. */ | |
| validate_dominance(impl, state); | |
| if (impl->valid_metadata & nir_metadata_live_defs) | |
| validate_live_defs(impl, state); | |
| if (!impl->structured) | |
| return; | |
| if (impl->valid_metadata & nir_metadata_divergence) | |
| validate_divergence(impl, state); | |
| if (impl->valid_metadata & nir_metadata_loop_analysis) | |
| validate_loop_info(impl, state); | |
| } | |
| static void | |
| validate_function_impl(nir_function_impl *impl, validate_state *state) | |
| { | |
| validate_assert(state, impl->function->impl == impl); | |
| validate_assert(state, impl->cf_node.parent == NULL); | |
| if (impl->preamble) { | |
| validate_assert(state, impl->function->is_entrypoint); | |
| validate_assert(state, impl->preamble->is_preamble); | |
| } | |
| validate_assert(state, exec_list_is_empty(&impl->end_block->instr_list)); | |
| validate_assert(state, impl->end_block->successors[0] == NULL); | |
| validate_assert(state, impl->end_block->successors[1] == NULL); | |
| state->impl = impl; | |
| state->parent_node = &impl->cf_node; | |
| exec_list_validate(&impl->locals); | |
| nir_foreach_function_temp_variable(var, impl) { | |
| validate_var_decl(var, nir_var_function_temp, state); | |
| } | |
| state->ssa_defs_found = reralloc(state->mem_ctx, state->ssa_defs_found, | |
| BITSET_WORD, BITSET_WORDS(impl->ssa_alloc)); | |
| memset(state->ssa_defs_found, 0, BITSET_BYTES(impl->ssa_alloc)); | |
| _mesa_set_clear(state->blocks, NULL); | |
| _mesa_set_resize(state->blocks, impl->num_blocks); | |
| if (impl->structured) | |
| collect_blocks(&impl->body, state); | |
| else | |
| collect_unstructured_blocks(impl, state); | |
| _mesa_set_add(state->blocks, impl->end_block); | |
| validate_assert(state, !exec_list_is_empty(&impl->body)); | |
| foreach_list_typed(nir_cf_node, node, node, &impl->body) { | |
| validate_cf_node(node, state); | |
| } | |
| validate_end_block(impl->end_block, state); | |
| /* We must have seen every source by now. This also means that we've untagged | |
| * every source, so we have valid (unaugmented) NIR once again. | |
| */ | |
| validate_assert(state, state->nr_tagged_srcs == 0); | |
| /* Metadata validation assumes a valid NIR shader. */ | |
| if (_mesa_hash_table_num_entries(state->errors) == 0) | |
| validate_metadata_and_ssa_dominance(impl, state); | |
| } | |
| static void | |
| validate_function(nir_function *func, validate_state *state) | |
| { | |
| if (func->impl != NULL) { | |
| validate_assert(state, func->impl->function == func); | |
| validate_function_impl(func->impl, state); | |
| } | |
| } | |
| static void | |
| validate_float_mul_add(nir_float_muladd_support muladd_support, validate_state *state) | |
| { | |
| if (muladd_support & nir_float_muladd_support_fuse) { | |
| validate_assert(state, muladd_support & | |
| (nir_float_muladd_support_has_ffma | nir_float_muladd_support_has_fmad)); | |
| } | |
| if (muladd_support & nir_float_muladd_support_prefers_split) | |
| validate_assert(state, muladd_support & nir_float_muladd_support_has_ffma); | |
| } | |
| static void | |
| validate_options(const nir_shader_compiler_options *options, validate_state *state) | |
| { | |
| if (!options) | |
| return; | |
| validate_float_mul_add(options->float_mul_add16, state); | |
| validate_float_mul_add(options->float_mul_add32, state); | |
| validate_float_mul_add(options->float_mul_add64, state); | |
| } | |
| static void | |
| init_validate_state(validate_state *state) | |
| { | |
| static struct util_once_flag once = UTIL_ONCE_FLAG_INIT; | |
| #if !DETECT_OS_WINDOWS | |
| util_call_once(&once, init_interactive); | |
| #endif | |
| state->mem_ctx = ralloc_context(NULL); | |
| state->ssa_defs_found = NULL; | |
| state->blocks = _mesa_pointer_set_create(state->mem_ctx); | |
| state->var_defs = _mesa_pointer_hash_table_create(state->mem_ctx); | |
| state->errors = _mesa_pointer_hash_table_create(state->mem_ctx); | |
| state->nr_tagged_srcs = 0; | |
| state->loop = NULL; | |
| state->in_loop_continue_construct = false; | |
| state->instr = NULL; | |
| state->var = NULL; | |
| } | |
| static void | |
| destroy_validate_state(validate_state *state) | |
| { | |
| ralloc_free(state->mem_ctx); | |
| } | |
| simple_mtx_t fail_dump_mutex = SIMPLE_MTX_INITIALIZER; | |
| static void | |
| dump_errors(validate_state *state, const char *when) | |
| { | |
| struct hash_table *errors = state->errors; | |
| /* Lock around dumping so that we get clean dumps in a multi-threaded | |
| * scenario | |
| */ | |
| simple_mtx_lock(&fail_dump_mutex); | |
| if (when) { | |
| fprintf(stderr, "NIR validation failed %s\n", when); | |
| fprintf(stderr, "%d errors:\n", _mesa_hash_table_num_entries(errors)); | |
| } else { | |
| fprintf(stderr, "NIR validation failed with %d errors:\n", | |
| _mesa_hash_table_num_entries(errors)); | |
| } | |
| nir_print_shader_annotated(state->shader, stderr, errors); | |
| if (_mesa_hash_table_num_entries(errors) > 0) { | |
| fprintf(stderr, "%d additional errors:\n", | |
| _mesa_hash_table_num_entries(errors)); | |
| hash_table_foreach(errors, entry) { | |
| fprintf(stderr, "%s\n", (char *)entry->data); | |
| } | |
| } | |
| fflush(stderr); | |
| simple_mtx_unlock(&fail_dump_mutex); | |
| abort(); | |
| } | |
| void | |
| nir_validate_shader(nir_shader *shader, const char *when) | |
| { | |
| if (NIR_DEBUG(NOVALIDATE)) | |
| return; | |
| validate_state state; | |
| init_validate_state(&state); | |
| state.shader = shader; | |
| nir_variable_mode valid_modes = | |
| nir_var_shader_in | | |
| nir_var_shader_out | | |
| nir_var_shader_temp | | |
| nir_var_uniform | | |
| nir_var_mem_ubo | | |
| nir_var_system_value | | |
| nir_var_mem_ssbo | | |
| nir_var_mem_shared | | |
| nir_var_mem_global | | |
| nir_var_mem_push_const | | |
| nir_var_mem_constant | | |
| nir_var_mem_pixel_local_in | | |
| nir_var_mem_pixel_local_out | | |
| nir_var_mem_pixel_local_inout | | |
| nir_var_image | | |
| nir_var_resource_heap | | |
| nir_var_sampler_heap; | |
| if (mesa_shader_stage_is_callable(shader->info.stage)) | |
| valid_modes |= nir_var_shader_call_data; | |
| if (shader->info.stage == MESA_SHADER_ANY_HIT || | |
| shader->info.stage == MESA_SHADER_CLOSEST_HIT || | |
| shader->info.stage == MESA_SHADER_INTERSECTION) | |
| valid_modes |= nir_var_ray_hit_attrib; | |
| if (shader->info.stage == MESA_SHADER_TASK || | |
| shader->info.stage == MESA_SHADER_MESH) | |
| valid_modes |= nir_var_mem_task_payload; | |
| if (shader->info.stage == MESA_SHADER_COMPUTE) | |
| valid_modes |= nir_var_mem_node_payload | | |
| nir_var_mem_node_payload_in; | |
| validate_options(shader->options, &state); | |
| exec_list_validate(&shader->variables); | |
| nir_foreach_variable_in_shader(var, shader) | |
| validate_var_decl(var, valid_modes, &state); | |
| exec_list_validate(&shader->functions); | |
| foreach_list_typed(nir_function, func, node, &shader->functions) { | |
| validate_function(func, &state); | |
| } | |
| if (shader->xfb_info != NULL) { | |
| /* At least validate that, if nir_shader::xfb_info exists, the shader | |
| * has real transform feedback going on. | |
| */ | |
| validate_assert(&state, shader->info.stage == MESA_SHADER_VERTEX || | |
| shader->info.stage == MESA_SHADER_TESS_EVAL || | |
| shader->info.stage == MESA_SHADER_GEOMETRY); | |
| validate_assert(&state, shader->xfb_info->buffers_written != 0); | |
| validate_assert(&state, shader->xfb_info->streams_written != 0); | |
| validate_assert(&state, shader->xfb_info->output_count > 0); | |
| } | |
| validate_assert(&state, | |
| util_is_power_of_two_or_zero(shader->info.api_subgroup_size)); | |
| validate_assert(&state, | |
| util_is_power_of_two_nonzero(shader->info.max_subgroup_size)); | |
| validate_assert(&state, | |
| util_is_power_of_two_nonzero(shader->info.min_subgroup_size)); | |
| validate_assert(&state, shader->info.max_subgroup_size <= 128); | |
| validate_assert(&state, shader->info.min_subgroup_size <= 128); | |
| validate_assert(&state, shader->info.api_subgroup_size == 0 || | |
| shader->info.api_subgroup_size >= shader->info.max_subgroup_size); | |
| validate_assert(&state, | |
| shader->info.min_subgroup_size <= shader->info.max_subgroup_size); | |
| if (_mesa_hash_table_num_entries(state.errors) > 0) | |
| dump_errors(&state, when); | |
| destroy_validate_state(&state); | |
| } | |
| void | |
| nir_validate_ssa_dominance(nir_shader *shader, const char *when) | |
| { | |
| if (NIR_DEBUG(NOVALIDATE)) | |
| return; | |
| validate_state state; | |
| init_validate_state(&state); | |
| state.shader = shader; | |
| nir_foreach_function_impl(impl, shader) { | |
| state.ssa_defs_found = reralloc(state.mem_ctx, state.ssa_defs_found, | |
| BITSET_WORD, | |
| BITSET_WORDS(impl->ssa_alloc)); | |
| memset(state.ssa_defs_found, 0, BITSET_BYTES(impl->ssa_alloc)); | |
| state.impl = impl; | |
| validate_ssa_dominance(impl, &state); | |
| } | |
| if (_mesa_hash_table_num_entries(state.errors) > 0) | |
| dump_errors(&state, when); | |
| destroy_validate_state(&state); | |
| } | |
| void | |
| _nir_assert_no_progress(bool progress, const char *when) | |
| { | |
| if (!progress) | |
| return; | |
| /* Lock around dumping so that we get clean dumps in a multi-threaded | |
| * scenario. | |
| */ | |
| simple_mtx_lock(&fail_dump_mutex); | |
| fprintf(stderr, "NIR assertion failed: Expected no progress from %s.\n", | |
| when); | |
| fflush(stderr); | |
| simple_mtx_unlock(&fail_dump_mutex); | |
| abort(); | |
| } | |
| #endif /* NDEBUG */ |