diff --git a/libs/vkd3d-shader/ir.c b/libs/vkd3d-shader/ir.c index 17215af58..1dc691152 100644 --- a/libs/vkd3d-shader/ir.c +++ b/libs/vkd3d-shader/ir.c @@ -1454,8 +1454,8 @@ static enum vkd3d_result vsir_program_lower_imul(struct vsir_program *program, static enum vkd3d_result vsir_program_lower_udiv(struct vsir_program *program, struct vsir_program_iterator *it, struct vsir_transformation_context *ctx) { + unsigned int count = 3, src0_id, src1_id, divisor_id; struct vkd3d_shader_instruction *udiv, *ins, *mov; - unsigned int count = 2; udiv = vsir_program_iterator_current(it); @@ -1469,9 +1469,9 @@ static enum vkd3d_result vsir_program_lower_udiv(struct vsir_program *program, } if (udiv->dst[0].reg.type != VKD3DSPR_NULL) - ++count; + count += 2; if (udiv->dst[1].reg.type != VKD3DSPR_NULL) - ++count; + count += 2; if (!vsir_program_iterator_insert_after(it, count)) return VKD3D_ERROR_OUT_OF_MEMORY; @@ -1483,14 +1483,33 @@ static enum vkd3d_result vsir_program_lower_udiv(struct vsir_program *program, return VKD3D_ERROR_OUT_OF_MEMORY; mov->src[0] = udiv->src[0]; - dst_param_init_ssa(&mov->dst[0], program->ssa_count, udiv->src[0].reg.data_type, udiv->src[0].reg.dimension); + src0_id = program->ssa_count++; + dst_param_init_ssa(&mov->dst[0], src0_id, udiv->src[0].reg.data_type, udiv->src[0].reg.dimension); mov = vsir_program_iterator_next(it); if (!(vsir_instruction_init_with_params(program, mov, &udiv->location, VSIR_OP_MOV, 1, 1))) return VKD3D_ERROR_OUT_OF_MEMORY; mov->src[0] = udiv->src[1]; - dst_param_init_ssa(&mov->dst[0], program->ssa_count + 1, udiv->src[1].reg.data_type, udiv->src[1].reg.dimension); + src1_id = program->ssa_count++; + dst_param_init_ssa(&mov->dst[0], src1_id, udiv->src[1].reg.data_type, udiv->src[1].reg.dimension); + + mov = vsir_program_iterator_next(it); + if (!(vsir_instruction_init_with_params(program, mov, &udiv->location, VSIR_OP_MOVC, 1, 3))) + return VKD3D_ERROR_OUT_OF_MEMORY; + + src_param_init_ssa(&mov->src[0], src1_id, udiv->src[1].reg.data_type, udiv->src[1].reg.dimension); + src_param_init_ssa(&mov->src[1], src1_id, udiv->src[1].reg.data_type, udiv->src[1].reg.dimension); + vsir_register_init(&mov->src[2].reg, VKD3DSPR_IMMCONST, VSIR_DATA_U32, 0); + mov->src[2].reg.dimension = udiv->src[1].reg.dimension; + mov->src[2].reg.u.immconst_u32[0] = 1; + mov->src[2].reg.u.immconst_u32[1] = 1; + mov->src[2].reg.u.immconst_u32[2] = 1; + mov->src[2].reg.u.immconst_u32[3] = 1; + if (mov->src[2].reg.dimension == VSIR_DIMENSION_VEC4) + mov->src[2].swizzle = VKD3D_SHADER_NO_SWIZZLE; + divisor_id = program->ssa_count++; + dst_param_init_ssa(&mov->dst[0], divisor_id, mov->src[1].reg.data_type, mov->src[1].reg.dimension); if (udiv->dst[0].reg.type != VKD3DSPR_NULL) { @@ -1501,11 +1520,30 @@ static enum vkd3d_result vsir_program_lower_udiv(struct vsir_program *program, ins->flags = udiv->flags; - src_param_init_ssa(&ins->src[0], program->ssa_count, - udiv->src[0].reg.data_type, udiv->src[0].reg.dimension); - src_param_init_ssa(&ins->src[1], program->ssa_count + 1, - udiv->src[1].reg.data_type, udiv->src[1].reg.dimension); + src_param_init_ssa(&ins->src[0], src0_id, udiv->src[0].reg.data_type, udiv->src[0].reg.dimension); + src_param_init_ssa(&ins->src[1], divisor_id, udiv->src[1].reg.data_type, udiv->src[1].reg.dimension); + dst_param_init_ssa(&ins->dst[0], program->ssa_count, udiv->src[1].reg.data_type, udiv->src[1].reg.dimension); + + /* Like its TPF equivalent, division by zero is well-defined for + * VSIR_OP_UDIV, and returns UINT_MAX. Division by zero is undefined + * for VSIR_OP_UDIV_SIMPLE and VSIR_OP_UREM, so handle it here. */ + ins = vsir_program_iterator_next(it); + if (!(vsir_instruction_init_with_params(program, ins, &udiv->location, VSIR_OP_MOVC, 1, 3))) + return VKD3D_ERROR_OUT_OF_MEMORY; + + src_param_init_ssa(&ins->src[0], src1_id, udiv->src[1].reg.data_type, udiv->src[1].reg.dimension); + src_param_init_ssa(&ins->src[1], program->ssa_count, udiv->src[1].reg.data_type, udiv->src[1].reg.dimension); + vsir_register_init(&ins->src[2].reg, VKD3DSPR_IMMCONST, VSIR_DATA_U32, 0); + ins->src[2].reg.dimension = udiv->src[1].reg.dimension; + ins->src[2].reg.u.immconst_u32[0] = UINT_MAX; + ins->src[2].reg.u.immconst_u32[1] = UINT_MAX; + ins->src[2].reg.u.immconst_u32[2] = UINT_MAX; + ins->src[2].reg.u.immconst_u32[3] = UINT_MAX; + if (ins->src[2].reg.dimension == VSIR_DIMENSION_VEC4) + ins->src[2].swizzle = VKD3D_SHADER_NO_SWIZZLE; ins->dst[0] = udiv->dst[0]; + + ++program->ssa_count; } if (udiv->dst[1].reg.type != VKD3DSPR_NULL) @@ -1517,15 +1555,30 @@ static enum vkd3d_result vsir_program_lower_udiv(struct vsir_program *program, ins->flags = udiv->flags; - src_param_init_ssa(&ins->src[0], program->ssa_count, - udiv->src[0].reg.data_type, udiv->src[0].reg.dimension); - src_param_init_ssa(&ins->src[1], program->ssa_count + 1, - udiv->src[1].reg.data_type, udiv->src[1].reg.dimension); + src_param_init_ssa(&ins->src[0], src0_id, udiv->src[0].reg.data_type, udiv->src[0].reg.dimension); + src_param_init_ssa(&ins->src[1], divisor_id, udiv->src[1].reg.data_type, udiv->src[1].reg.dimension); + dst_param_init_ssa(&ins->dst[0], program->ssa_count, udiv->src[1].reg.data_type, udiv->src[1].reg.dimension); + + ins = vsir_program_iterator_next(it); + if (!(vsir_instruction_init_with_params(program, ins, &udiv->location, VSIR_OP_MOVC, 1, 3))) + return VKD3D_ERROR_OUT_OF_MEMORY; + + src_param_init_ssa(&ins->src[0], src1_id, udiv->src[1].reg.data_type, udiv->src[1].reg.dimension); + src_param_init_ssa(&ins->src[1], program->ssa_count, udiv->src[1].reg.data_type, udiv->src[1].reg.dimension); + vsir_register_init(&ins->src[2].reg, VKD3DSPR_IMMCONST, VSIR_DATA_U32, 0); + ins->src[2].reg.dimension = udiv->src[1].reg.dimension; + ins->src[2].reg.u.immconst_u32[0] = UINT_MAX; + ins->src[2].reg.u.immconst_u32[1] = UINT_MAX; + ins->src[2].reg.u.immconst_u32[2] = UINT_MAX; + ins->src[2].reg.u.immconst_u32[3] = UINT_MAX; + if (ins->src[2].reg.dimension == VSIR_DIMENSION_VEC4) + ins->src[2].swizzle = VKD3D_SHADER_NO_SWIZZLE; ins->dst[0] = udiv->dst[1]; + + ++program->ssa_count; } vkd3d_shader_instruction_make_nop(udiv); - program->ssa_count += 2; return VKD3D_OK; } diff --git a/libs/vkd3d-shader/spirv.c b/libs/vkd3d-shader/spirv.c index b563459f4..a67299ecb 100644 --- a/libs/vkd3d-shader/spirv.c +++ b/libs/vkd3d-shader/spirv.c @@ -7625,14 +7625,13 @@ static void spirv_compiler_emit_bool_cast(struct spirv_compiler *compiler, static enum vkd3d_result spirv_compiler_emit_alu_instruction(struct spirv_compiler *compiler, const struct vkd3d_shader_instruction *instruction) { - uint32_t src_ids[SPIRV_MAX_SRC_COUNT], condition_id = 0, uint_max_id = 0; struct vkd3d_spirv_builder *builder = &compiler->spirv_builder; const struct vkd3d_shader_dst_param *dst = instruction->dst; const struct vkd3d_shader_src_param *src = instruction->src; - unsigned int i, component_count; + uint32_t src_ids[SPIRV_MAX_SRC_COUNT]; uint32_t type_id, val_id; SpvOp op = SpvOpMax; - bool check_zero; + unsigned int i; if (src->reg.data_type == VSIR_DATA_U64 && instruction->opcode == VSIR_OP_COUNTBITS) { @@ -7672,42 +7671,14 @@ static enum vkd3d_result spirv_compiler_emit_alu_instruction(struct spirv_compil return VKD3D_ERROR_INVALID_SHADER; } - /* SPIR-V doesn't mandate a behaviour when a denominator is zero, - * so we have an explicit check. */ - switch (instruction->opcode) - { - case VSIR_OP_UDIV_SIMPLE: - case VSIR_OP_UREM: - check_zero = true; - break; - - default: - check_zero = false; - break; - } - VKD3D_ASSERT(instruction->dst_count == 1); VKD3D_ASSERT(instruction->src_count <= SPIRV_MAX_SRC_COUNT); - if (check_zero) - VKD3D_ASSERT(instruction->src_count == 2); - component_count = vsir_write_mask_component_count(dst[0].write_mask); type_id = spirv_compiler_get_type_id_for_dst(compiler, dst); for (i = 0; i < instruction->src_count; ++i) src_ids[i] = spirv_compiler_emit_load_src(compiler, &src[i], dst->write_mask); - if (check_zero) - { - condition_id = spirv_compiler_emit_int_to_bool(compiler, - VKD3D_SHADER_CONDITIONAL_OP_NZ, src[1].reg.data_type, component_count, src_ids[1]); - - if (data_type_is_64_bit(dst[0].reg.data_type)) - uint_max_id = spirv_compiler_get_constant_uint64_vector(compiler, UINT64_MAX, component_count); - else - uint_max_id = spirv_compiler_get_constant_uint_vector(compiler, UINT_MAX, component_count); - } - /* The SPIR-V specification states, "The resulting value is undefined if * Shift is greater than or equal to the bit width of the components of * Base." Direct3D applies only the lowest 5 bits of the shift. @@ -7728,9 +7699,6 @@ static enum vkd3d_result spirv_compiler_emit_alu_instruction(struct spirv_compil if (instruction->flags & VKD3DSI_PRECISE_XYZW) vkd3d_spirv_build_op_decorate(builder, val_id, SpvDecorationNoContraction, NULL, 0); - if (check_zero) - val_id = vkd3d_spirv_build_op_select(builder, type_id, condition_id, val_id, uint_max_id); - spirv_compiler_emit_store_dst(compiler, dst, val_id); return VKD3D_OK; } diff --git a/tests/hlsl/arithmetic-uint.shader_test b/tests/hlsl/arithmetic-uint.shader_test index 93e7dfdc4..0ea94b194 100644 --- a/tests/hlsl/arithmetic-uint.shader_test +++ b/tests/hlsl/arithmetic-uint.shader_test @@ -69,7 +69,7 @@ void main(out uint4 dst : sv_target) uniform 0 uint4 0 0 1 0 todo(msl) draw quad % Integer division by zero is undefined for shader model 6 targets. -if(sm<6) todo(glsl & !llvmpipe) probe (0, 0) u32(0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff) +if(sm<6) probe (0, 0) u32(0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff) uniform 0 uint4 1 1 7 1 todo(msl) draw quad probe (0, 0) u32(1, 7, 0, 0)