diff --git a/libs/vkd3d-shader/ir.c b/libs/vkd3d-shader/ir.c index dbce702e9..73dd157a9 100644 --- a/libs/vkd3d-shader/ir.c +++ b/libs/vkd3d-shader/ir.c @@ -8826,12 +8826,14 @@ struct temp_allocator uint8_t allocated_mask; uint32_t temp_id; } *ssa_regs, *temp_regs; + size_t ssa_count, temp_count; size_t allocated_ssa_count, allocated_temp_count; + unsigned int new_temp_count; enum vkd3d_result result; }; static uint8_t get_available_writemask(const struct temp_allocator *allocator, - struct liveness_tracker *tracker, unsigned int first_write, unsigned int last_access, uint32_t temp_id) + const struct liveness_tracker *tracker, unsigned int first_write, unsigned int last_access, uint32_t temp_id) { uint8_t writemask = VKD3DSP_WRITEMASK_ALL; @@ -8870,7 +8872,7 @@ static uint8_t get_available_writemask(const struct temp_allocator *allocator, return writemask; } -static bool temp_allocator_allocate(struct temp_allocator *allocator, struct liveness_tracker *tracker, +static bool temp_allocator_allocate(struct temp_allocator *allocator, const struct liveness_tracker *tracker, struct temp_allocator_reg *reg, const struct liveness_tracker_reg *liveness_reg) { if (!liveness_reg->written) @@ -9054,6 +9056,47 @@ static void temp_allocator_set_dst(struct temp_allocator *allocator, } } +static void temp_allocator_compute_allocation_map(struct temp_allocator *allocator, + const struct liveness_tracker *tracker) +{ + /* Reallocate temps first. We do this specifically to make sure that r0 is + * the first register to be allocated, and thus will be reallocated in + * place, and left alone. + * This is necessary because, in pixel shader model 1.x, r0 doubles as the + * output register, and needs to remain at r0. (Note that we need to already + * have the output in r0, rather than e.g. putting it in o0 and converting + * it to r0 after this pass, so that we know when r0 is live.) */ + for (unsigned int i = 0; i < allocator->temp_count; ++i) + { + const struct liveness_tracker_reg *liveness_reg = &tracker->temp_regs[i]; + struct temp_allocator_reg *reg = &allocator->temp_regs[i]; + + if (temp_allocator_allocate(allocator, tracker, reg, liveness_reg)) + { + TRACE("Reallocated r%u%s for r%u (liveness %u-%u).\n", + reg->temp_id, debug_vsir_writemask(reg->allocated_mask), i, + liveness_reg->first_write, liveness_reg->last_access); + allocator->new_temp_count = max(allocator->new_temp_count, reg->temp_id + 1); + } + ++allocator->allocated_temp_count; + } + + for (unsigned int i = 0; i < allocator->ssa_count; ++i) + { + const struct liveness_tracker_reg *liveness_reg = &tracker->ssa_regs[i]; + struct temp_allocator_reg *reg = &allocator->ssa_regs[i]; + + if (temp_allocator_allocate(allocator, tracker, reg, liveness_reg)) + { + TRACE("Allocated r%u%s for sr%u (liveness %u-%u).\n", + reg->temp_id, debug_vsir_writemask(reg->allocated_mask), i, + liveness_reg->first_write, liveness_reg->last_access); + allocator->new_temp_count = max(allocator->new_temp_count, reg->temp_id + 1); + } + ++allocator->allocated_ssa_count; + } +} + /* This pass does two things: * * - converts SSA registers (sr#) into temp registers (r#); @@ -9074,66 +9117,30 @@ enum vkd3d_result vsir_allocate_temp_registers(struct vsir_program *program, struct vkd3d_shader_message_context *message_context) { struct vsir_program_iterator it = vsir_program_iterator(&program->instructions); - const unsigned int prev_temp_count = program->temp_count; struct temp_allocator allocator = {0}; struct vkd3d_shader_instruction *ins; struct temp_allocator_reg *regs; struct liveness_tracker tracker; enum vkd3d_result ret; - if (!program->ssa_count && !prev_temp_count) + if (!program->ssa_count && !program->temp_count) return VKD3D_OK; if ((ret = track_liveness(program, &tracker))) return ret; - if (!(regs = vkd3d_calloc(program->ssa_count + prev_temp_count, sizeof(*regs)))) + if (!(regs = vkd3d_calloc(program->ssa_count + program->temp_count, sizeof(*regs)))) { liveness_tracker_cleanup(&tracker); return VKD3D_ERROR_OUT_OF_MEMORY; } allocator.message_context = message_context; + allocator.ssa_count = program->ssa_count; + allocator.temp_count = program->temp_count; allocator.ssa_regs = regs; allocator.temp_regs = regs + program->ssa_count; - program->temp_count = 0; - - /* Reallocate temps first. We do this specifically to make sure that r0 is - * the first register to be allocated, and thus will be reallocated in - * place, and left alone. - * This is necessary because, in pixel shader model 1.x, r0 doubles as the - * output register, and needs to remain at r0. (Note that we need to already - * have the output in r0, rather than e.g. putting it in o0 and converting - * it to r0 after this pass, so that we know when r0 is live.) */ - for (unsigned int i = 0; i < prev_temp_count; ++i) - { - const struct liveness_tracker_reg *liveness_reg = &tracker.temp_regs[i]; - struct temp_allocator_reg *reg = &allocator.temp_regs[i]; - - if (temp_allocator_allocate(&allocator, &tracker, reg, liveness_reg)) - { - TRACE("Reallocated r%u%s for r%u (liveness %u-%u).\n", - reg->temp_id, debug_vsir_writemask(reg->allocated_mask), i, - liveness_reg->first_write, liveness_reg->last_access); - program->temp_count = max(program->temp_count, reg->temp_id + 1); - } - ++allocator.allocated_temp_count; - } - - for (unsigned int i = 0; i < program->ssa_count; ++i) - { - const struct liveness_tracker_reg *liveness_reg = &tracker.ssa_regs[i]; - struct temp_allocator_reg *reg = &allocator.ssa_regs[i]; - - if (temp_allocator_allocate(&allocator, &tracker, reg, liveness_reg)) - { - TRACE("Allocated r%u%s for sr%u (liveness %u-%u).\n", - reg->temp_id, debug_vsir_writemask(reg->allocated_mask), i, - liveness_reg->first_write, liveness_reg->last_access); - program->temp_count = max(program->temp_count, reg->temp_id + 1); - } - ++allocator.allocated_ssa_count; - } + temp_allocator_compute_allocation_map(&allocator, &tracker); for (ins = vsir_program_iterator_head(&it); ins; ins = vsir_program_iterator_next(&it)) { @@ -9146,9 +9153,11 @@ enum vkd3d_result vsir_allocate_temp_registers(struct vsir_program *program, } program->ssa_count = 0; + program->temp_count = allocator.new_temp_count; vkd3d_free(regs); liveness_tracker_cleanup(&tracker); + return allocator.result; }