vkd3d-shader/dxil: Emit phi instructions during parsing.

This commit is contained in:
Conor McCarthy
2024-09-30 15:49:23 +10:00
committed by Henri Verbeet
parent 28e58e0c85
commit 6643c848fd
Notes: Henri Verbeet 2025-11-20 18:36:32 +01:00
Approved-by: Giovanni Mascellani (@giomasce)
Approved-by: Henri Verbeet (@hverbeet)
Merge-Request: https://gitlab.winehq.org/wine/vkd3d/-/merge_requests/1133

View File

@@ -748,16 +748,8 @@ struct sm6_symbol
struct incoming_value
{
const struct sm6_block *block;
struct vkd3d_shader_register reg;
};
struct sm6_phi
{
struct sm6_value value;
struct incoming_value *incoming;
size_t incoming_capacity;
size_t incoming_count;
unsigned int block_idx;
const struct sm6_value *src;
};
struct sm6_block
@@ -768,10 +760,6 @@ struct sm6_block
/* A nonzero id. */
unsigned int id;
struct sm6_phi *phi;
size_t phi_capacity;
size_t phi_count;
};
struct sm6_function
@@ -3117,6 +3105,37 @@ static size_t sm6_parser_get_value_idx_by_ref(struct sm6_parser *sm6, const stru
return operand;
}
static uint64_t decode_rotated_signed_value(uint64_t value)
{
if (value != 1)
{
bool neg = value & 1;
value >>= 1;
return neg ? -value : value;
}
return value << 63;
}
static const struct sm6_value *sm6_parser_get_value_by_rotated_signed_idx(struct sm6_parser *dxil,
uint64_t idx, const struct sm6_type *fwd_type)
{
int64_t rotated_idx;
size_t operand;
rotated_idx = decode_rotated_signed_value(idx);
if (rotated_idx > INT32_MAX || rotated_idx < INT32_MIN)
vkd3d_shader_parser_warning(&dxil->p, VKD3D_SHADER_WARNING_DXIL_IGNORING_OPERANDS,
"Ignoring upper 32 bits of DXIL SSA value signed relative index %"PRIx64".", rotated_idx);
if ((operand = sm6_parser_get_value_index(dxil, rotated_idx)) == SIZE_MAX)
return NULL;
sm6_parser_pre_init_or_validate_referenced_value(dxil, operand, fwd_type);
return &dxil->values[operand];
}
static const struct sm6_value *sm6_parser_get_value_by_ref(struct sm6_parser *sm6,
const struct dxil_record *record, const struct sm6_type *type, unsigned int *rec_idx)
{
@@ -3188,17 +3207,6 @@ static bool sm6_parser_declare_function(struct sm6_parser *sm6, const struct dxi
return true;
}
static inline uint64_t decode_rotated_signed_value(uint64_t value)
{
if (value != 1)
{
bool neg = value & 1;
value >>= 1;
return neg ? -value : value;
}
return value << 63;
}
static struct sm6_index *sm6_get_value_index(struct sm6_parser *sm6, struct sm6_value *value)
{
switch (value->value_type)
@@ -4283,26 +4291,6 @@ static struct sm6_block *sm6_block_create()
return block;
}
static struct sm6_phi *sm6_block_phi_require_space(struct sm6_block *block, struct sm6_parser *sm6)
{
struct sm6_phi *phi;
if (!vkd3d_array_reserve((void **)&block->phi, &block->phi_capacity, block->phi_count + 1, sizeof(*block->phi)))
{
ERR("Failed to allocate phi array.\n");
vkd3d_shader_parser_error(&sm6->p, VKD3D_SHADER_ERROR_DXIL_OUT_OF_MEMORY,
"Out of memory allocating a phi instruction.");
return NULL;
}
phi = &block->phi[block->phi_count++];
phi->incoming = NULL;
phi->incoming_capacity = 0;
phi->incoming_count = 0;
return phi;
}
struct function_emission_state
{
struct sm6_block *code_block;
@@ -4736,15 +4724,6 @@ static bool sm6_function_validate_block_index(const struct sm6_function *functio
return true;
}
static const struct sm6_block *sm6_function_get_block(const struct sm6_function *function,
uint64_t index, struct sm6_parser *dxil)
{
if (!sm6_function_validate_block_index(function, index, dxil))
return NULL;
return function->blocks[index];
}
static void sm6_parser_emit_br(struct sm6_parser *dxil, const struct dxil_record *record,
struct sm6_function *function, struct vkd3d_shader_instruction *ins)
{
@@ -7653,95 +7632,105 @@ static int phi_incoming_compare(const void *a, const void *b)
{
const struct incoming_value *incoming_a = a, *incoming_b = b;
return (incoming_a->block > incoming_b->block) - (incoming_a->block < incoming_b->block);
return vkd3d_u32_compare(incoming_a->block_idx, incoming_b->block_idx);
}
static void sm6_parser_emit_phi(struct sm6_parser *sm6, const struct dxil_record *record,
struct sm6_function *function, struct sm6_block *code_block, struct vkd3d_shader_instruction *ins,
struct sm6_value *dst)
static void sm6_parser_emit_phi(struct sm6_parser *dxil, const struct dxil_record *record,
struct sm6_function *function, struct vkd3d_shader_instruction *ins, struct sm6_value *dst)
{
struct vkd3d_shader_src_param *src_params;
unsigned int i, j, incoming_count;
struct incoming_value *incoming;
const struct sm6_value *src;
const struct sm6_type *type;
struct sm6_phi *phi;
unsigned int i, j;
uint64_t src_idx;
if (!(record->operand_count & 1))
{
WARN("Invalid operand count %u.\n", record->operand_count);
vkd3d_shader_parser_error(&sm6->p, VKD3D_SHADER_ERROR_DXIL_INVALID_OPERAND_COUNT,
vkd3d_shader_parser_error(&dxil->p, VKD3D_SHADER_ERROR_DXIL_INVALID_OPERAND_COUNT,
"Invalid operand count %u for phi instruction.", record->operand_count);
return;
}
if (!(type = sm6_parser_get_type(sm6, record->operands[0])))
if (!(type = sm6_parser_get_type(dxil, record->operands[0])))
return;
if (!sm6_type_is_numeric(type))
{
/* dxc doesn't seem to use buffer/resource read return types here. */
FIXME("Only scalar numeric types are supported.\n");
vkd3d_shader_parser_error(&sm6->p, VKD3D_SHADER_ERROR_DXIL_INVALID_OPERAND_COUNT,
vkd3d_shader_parser_error(&dxil->p, VKD3D_SHADER_ERROR_DXIL_INVALID_OPERAND_COUNT,
"Result type class %u of a phi instruction is not scalar numeric.", type->class);
return;
}
dst->type = type;
sm6_parser_init_ssa_value(sm6, dst);
if (!(phi = sm6_block_phi_require_space(code_block, sm6)))
return;
phi->incoming_count = record->operand_count / 2u;
if (!vkd3d_array_reserve((void **)&phi->incoming, &phi->incoming_capacity, phi->incoming_count,
sizeof(*phi->incoming)))
incoming_count = record->operand_count / 2u;
if (!(incoming = vkd3d_calloc(incoming_count, sizeof(*incoming))))
{
ERR("Failed to allocate phi incoming array.\n");
vkd3d_shader_parser_error(&sm6->p, VKD3D_SHADER_ERROR_DXIL_OUT_OF_MEMORY,
vkd3d_shader_parser_error(&dxil->p, VKD3D_SHADER_ERROR_DXIL_OUT_OF_MEMORY,
"Out of memory allocating a phi incoming array.");
return;
}
incoming = phi->incoming;
for (i = 1; i < record->operand_count; i += 2)
{
src_idx = sm6->value_count - decode_rotated_signed_value(record->operands[i]);
/* May be a forward reference. */
if (src_idx >= sm6->cur_max_value)
/* Phi forward references are handled by the same mechanism as all
* others. Constant and undefined values are never forward references,
* and the only other valid incoming is an SSA value, which will be
* initialised if necessary. */
if (!(src = sm6_parser_get_value_by_rotated_signed_idx(dxil, record->operands[i], type)))
goto done;
if (!sm6_value_is_constant(src) && !sm6_value_is_undef(src) && !sm6_value_is_ssa(src))
{
WARN("Invalid value index %"PRIu64".\n", src_idx);
vkd3d_shader_parser_error(&sm6->p, VKD3D_SHADER_ERROR_DXIL_INVALID_OPERAND,
"Invalid value index %"PRIu64" for a phi incoming value.", src_idx);
return;
vkd3d_shader_parser_error(&dxil->p, VKD3D_SHADER_ERROR_DXIL_INVALID_OPERAND,
"A PHI incoming value is not a constant or SSA register.");
goto done;
}
if (src->type != type)
vkd3d_shader_parser_warning(&dxil->p, VKD3D_SHADER_WARNING_DXIL_TYPE_MISMATCH,
"The type of a phi incoming value does not match the result type.");
if (!sm6_function_validate_block_index(function, record->operands[i + 1], dxil))
goto done;
j = i / 2u;
/* Store the value index in the register for later resolution. */
incoming[j].reg.idx[0].offset = src_idx;
incoming[j].block = sm6_function_get_block(function, record->operands[i + 1], sm6);
incoming[j].src = src;
incoming[j].block_idx = record->operands[i + 1];
}
ins->opcode = VSIR_OP_NOP;
qsort(incoming, incoming_count, sizeof(*incoming), phi_incoming_compare);
qsort(incoming, phi->incoming_count, sizeof(*incoming), phi_incoming_compare);
for (i = 1, j = 1; i < phi->incoming_count; ++i)
/* Deduplicate incomings. DXIL phi instructions can contain duplicates. */
for (i = 1, j = 1; i < incoming_count; ++i)
{
if (incoming[i].block != incoming[i - 1].block)
if (incoming[i].block_idx != incoming[i - 1].block_idx)
{
incoming[j++] = incoming[i];
continue;
}
if (incoming[i].reg.idx[0].offset != incoming[i - 1].reg.idx[0].offset)
{
WARN("PHI conflict.\n");
vkd3d_shader_parser_error(&sm6->p, VKD3D_SHADER_ERROR_DXIL_INVALID_OPERAND,
if (incoming[i].src != incoming[i - 1].src)
vkd3d_shader_parser_error(&dxil->p, VKD3D_SHADER_ERROR_DXIL_INVALID_OPERAND,
"Two phi incomings have the same block but different values.");
}
}
/* if (j == 1) we should be able to set dst->u.reg to incoming[0].reg, but structurisation
* may potentially add new incomings. */
phi->incoming_count = j;
incoming_count = j;
vsir_instruction_init(ins, &dxil->p.location, VSIR_OP_PHI);
if (!(src_params = instruction_src_params_alloc(ins, incoming_count * 2u, dxil)))
goto done;
for (i = 0; i < incoming_count; ++i)
{
j = i * 2u;
src_param_init_from_value(&src_params[j], incoming[i].src, 0, dxil);
vsir_src_param_init_label(&src_params[j + 1], incoming[i].block_idx + 1);
}
instruction_dst_param_init_ssa_scalar(ins, 0, dxil);
done:
vkd3d_free(incoming);
}
static void sm6_parser_emit_ret(struct sm6_parser *dxil,
@@ -8274,45 +8263,6 @@ static struct sm6_block *sm6_function_create_block(struct sm6_function *function
return block;
}
static enum vkd3d_result sm6_function_resolve_phi_incomings(const struct sm6_function *function,
struct sm6_parser *sm6)
{
const struct sm6_block *block;
size_t i, j, block_idx;
for (block_idx = 0; block_idx < function->block_count; ++block_idx)
{
block = function->blocks[block_idx];
for (i = 0; i < block->phi_count; ++i)
{
struct sm6_phi *phi = &block->phi[i];
const struct sm6_value *src;
for (j = 0; j < phi->incoming_count; ++j)
{
src = &sm6->values[phi->incoming[j].reg.idx[0].offset];
if (!sm6_value_is_constant(src) && !sm6_value_is_undef(src) && !sm6_value_is_ssa(src))
{
FIXME("PHI incoming value is not a constant or SSA register.\n");
vkd3d_shader_parser_error(&sm6->p, VKD3D_SHADER_ERROR_DXIL_INVALID_OPERAND,
"A PHI incoming value is not a constant or SSA register.");
return VKD3D_ERROR_INVALID_SHADER;
}
if (src->type != phi->value.type)
{
WARN("Type mismatch.\n");
vkd3d_shader_parser_warning(&sm6->p, VKD3D_SHADER_WARNING_DXIL_TYPE_MISMATCH,
"The type of a phi incoming value does not match the result type.");
}
vsir_register_from_dxil_value(&phi->incoming[j].reg, src, 0, sm6);
}
}
}
return VKD3D_OK;
}
static enum vkd3d_result sm6_parser_function_init(struct sm6_parser *sm6, const struct dxil_block *block,
struct sm6_function *function)
{
@@ -8443,7 +8393,7 @@ static enum vkd3d_result sm6_parser_function_init(struct sm6_parser *sm6, const
sm6_parser_emit_load(sm6, record, ins, dst);
break;
case FUNC_CODE_INST_PHI:
sm6_parser_emit_phi(sm6, record, function, code_block, ins, dst);
sm6_parser_emit_phi(sm6, record, function, ins, dst);
break;
case FUNC_CODE_INST_RET:
sm6_parser_emit_ret(sm6, record, ins);
@@ -8471,10 +8421,6 @@ static enum vkd3d_result sm6_parser_function_init(struct sm6_parser *sm6, const
if (record->attachment)
metadata_attachment_record_apply(record->attachment, record->code, ins, dst, sm6);
/* This is specific for PHI nodes, but must happen after attachments have been applied. */
if (record->code == FUNC_CODE_INST_PHI)
code_block->phi[code_block->phi_count - 1].value = *dst;
/* A vsir instruction can be emitted for a block terminator, so handle
* a possible instruction count increment before moving to the next
* code block. */
@@ -8503,52 +8449,7 @@ static enum vkd3d_result sm6_parser_function_init(struct sm6_parser *sm6, const
return VKD3D_ERROR_INVALID_SHADER;
}
return sm6_function_resolve_phi_incomings(function, sm6);
}
static void sm6_block_emit_phi(const struct sm6_block *block, struct sm6_parser *sm6)
{
struct vkd3d_shader_instruction *ins;
unsigned int i, j, incoming_count;
const struct sm6_phi *src_phi;
for (i = 0; i < block->phi_count; ++i)
{
struct vkd3d_shader_src_param *src_params;
struct vkd3d_shader_dst_param *dst_param;
src_phi = &block->phi[i];
incoming_count = src_phi->incoming_count;
if (!(ins = sm6_parser_add_instruction(sm6, VSIR_OP_PHI)))
return;
if (!(src_params = instruction_src_params_alloc(ins, incoming_count * 2u, sm6)))
{
vkd3d_shader_instruction_make_nop(ins);
return;
}
if (!(dst_param = instruction_dst_params_alloc(ins, 1, sm6)))
{
vkd3d_shader_instruction_make_nop(ins);
return;
}
for (j = 0; j < incoming_count; ++j)
{
const struct sm6_block *incoming_block = src_phi->incoming[j].block;
unsigned int index = j * 2;
src_param_init(&src_params[index]);
src_params[index].reg = src_phi->incoming[j].reg;
if (incoming_block)
vsir_src_param_init_label(&src_params[index + 1], incoming_block->id);
else
VKD3D_ASSERT(sm6->p.status < 0);
}
dst_param_init(dst_param);
vsir_register_from_dxil_value(&dst_param->reg, &src_phi->value, 0, sm6);
}
return VKD3D_OK;
}
static enum vkd3d_result sm6_parser_module_init(struct sm6_parser *sm6, const struct dxil_block *block,
@@ -8636,7 +8537,6 @@ static enum vkd3d_result sm6_function_emit_blocks(const struct sm6_function *fun
const struct sm6_block *block = function->blocks[i];
sm6_parser_emit_label(sm6, block->id);
sm6_block_emit_phi(block, sm6);
for (j = 0; j < block->instruction_count; ++j)
{
@@ -10629,19 +10529,9 @@ static void sm6_symtab_cleanup(struct sm6_symbol *symbols, size_t count)
vkd3d_free(symbols);
}
static void sm6_phi_destroy(struct sm6_phi *phi)
{
vkd3d_free(phi->incoming);
}
static void sm6_block_destroy(struct sm6_block *block)
{
unsigned int i;
vkd3d_free(block->instructions);
for (i = 0; i < block->phi_count; ++i)
sm6_phi_destroy(&block->phi[i]);
vkd3d_free(block->phi);
vkd3d_free(block);
}