#include "st_format.h"
#include "st_nir.h"
#include "st_shader_cache.h"
+#include "st_glsl_to_tgsi_temprename.h"
#include "util/hash_table.h"
#include <algorithm>
#define MAX_GLSL_TEXTURE_OFFSET 4
-class st_src_reg;
-class st_dst_reg;
-
-static int swizzle_for_size(int size);
-
-static int swizzle_for_type(const glsl_type *type, int component = 0)
-{
- unsigned num_elements = 4;
-
- if (type) {
- type = type->without_array();
- if (type->is_scalar() || type->is_vector() || type->is_matrix())
- num_elements = type->vector_elements;
- }
-
- int swizzle = swizzle_for_size(num_elements);
- assert(num_elements + component <= 4);
-
- swizzle += component * MAKE_SWIZZLE4(1, 1, 1, 1);
- return swizzle;
-}
-
static unsigned is_precise(const ir_variable *ir)
{
if (!ir)
return ir->data.precise || ir->data.invariant;
}
-/**
- * This struct is a corresponding struct to TGSI ureg_src.
- */
-class st_src_reg {
-public:
- st_src_reg(gl_register_file file, int index, const glsl_type *type,
- int component = 0, unsigned array_id = 0)
- {
- assert(file != PROGRAM_ARRAY || array_id != 0);
- this->file = file;
- this->index = index;
- this->swizzle = swizzle_for_type(type, component);
- this->negate = 0;
- this->abs = 0;
- this->index2D = 0;
- this->type = type ? type->base_type : GLSL_TYPE_ERROR;
- this->reladdr = NULL;
- this->reladdr2 = NULL;
- this->has_index2 = false;
- this->double_reg2 = false;
- this->array_id = array_id;
- this->is_double_vertex_input = false;
- }
-
- st_src_reg(gl_register_file file, int index, enum glsl_base_type type)
- {
- assert(file != PROGRAM_ARRAY); /* need array_id > 0 */
- this->type = type;
- this->file = file;
- this->index = index;
- this->index2D = 0;
- this->swizzle = SWIZZLE_XYZW;
- this->negate = 0;
- this->abs = 0;
- this->reladdr = NULL;
- this->reladdr2 = NULL;
- this->has_index2 = false;
- this->double_reg2 = false;
- this->array_id = 0;
- this->is_double_vertex_input = false;
- }
-
- st_src_reg(gl_register_file file, int index, enum glsl_base_type type, int index2D)
- {
- assert(file != PROGRAM_ARRAY); /* need array_id > 0 */
- this->type = type;
- this->file = file;
- this->index = index;
- this->index2D = index2D;
- this->swizzle = SWIZZLE_XYZW;
- this->negate = 0;
- this->abs = 0;
- this->reladdr = NULL;
- this->reladdr2 = NULL;
- this->has_index2 = false;
- this->double_reg2 = false;
- this->array_id = 0;
- this->is_double_vertex_input = false;
- }
-
- st_src_reg()
- {
- this->type = GLSL_TYPE_ERROR;
- this->file = PROGRAM_UNDEFINED;
- this->index = 0;
- this->index2D = 0;
- this->swizzle = 0;
- this->negate = 0;
- this->abs = 0;
- this->reladdr = NULL;
- this->reladdr2 = NULL;
- this->has_index2 = false;
- this->double_reg2 = false;
- this->array_id = 0;
- this->is_double_vertex_input = false;
- }
-
- explicit st_src_reg(st_dst_reg reg);
-
- int32_t index; /**< temporary index, VERT_ATTRIB_*, VARYING_SLOT_*, etc. */
- int16_t index2D;
- uint16_t swizzle; /**< SWIZZLE_XYZWONEZERO swizzles from Mesa. */
- int negate:4; /**< NEGATE_XYZW mask from mesa */
- unsigned abs:1;
- enum glsl_base_type type:5; /** GLSL_TYPE_* from GLSL IR (enum glsl_base_type) */
- unsigned has_index2:1;
- gl_register_file file:5; /**< PROGRAM_* from Mesa */
- /*
- * Is this the second half of a double register pair?
- * currently used for input mapping only.
- */
- unsigned double_reg2:1;
- unsigned is_double_vertex_input:1;
- unsigned array_id:10;
-
- /** Register index should be offset by the integer in this reg. */
- st_src_reg *reladdr;
- st_src_reg *reladdr2;
-
- st_src_reg get_abs()
- {
- st_src_reg reg = *this;
- reg.negate = 0;
- reg.abs = 1;
- return reg;
- }
-};
-
-class st_dst_reg {
-public:
- st_dst_reg(gl_register_file file, int writemask, enum glsl_base_type type, int index)
- {
- assert(file != PROGRAM_ARRAY); /* need array_id > 0 */
- this->file = file;
- this->index = index;
- this->index2D = 0;
- this->writemask = writemask;
- this->reladdr = NULL;
- this->reladdr2 = NULL;
- this->has_index2 = false;
- this->type = type;
- this->array_id = 0;
- }
-
- st_dst_reg(gl_register_file file, int writemask, enum glsl_base_type type)
- {
- assert(file != PROGRAM_ARRAY); /* need array_id > 0 */
- this->file = file;
- this->index = 0;
- this->index2D = 0;
- this->writemask = writemask;
- this->reladdr = NULL;
- this->reladdr2 = NULL;
- this->has_index2 = false;
- this->type = type;
- this->array_id = 0;
- }
-
- st_dst_reg()
- {
- this->type = GLSL_TYPE_ERROR;
- this->file = PROGRAM_UNDEFINED;
- this->index = 0;
- this->index2D = 0;
- this->writemask = 0;
- this->reladdr = NULL;
- this->reladdr2 = NULL;
- this->has_index2 = false;
- this->array_id = 0;
- }
-
- explicit st_dst_reg(st_src_reg reg);
-
- int32_t index; /**< temporary index, VERT_ATTRIB_*, VARYING_SLOT_*, etc. */
- int16_t index2D;
- gl_register_file file:5; /**< PROGRAM_* from Mesa */
- unsigned writemask:4; /**< Bitfield of WRITEMASK_[XYZW] */
- enum glsl_base_type type:5; /** GLSL_TYPE_* from GLSL IR (enum glsl_base_type) */
- unsigned has_index2:1;
- unsigned array_id:10;
-
- /** Register index should be offset by the integer in this reg. */
- st_src_reg *reladdr;
- st_src_reg *reladdr2;
-};
-
-st_src_reg::st_src_reg(st_dst_reg reg)
-{
- this->type = reg.type;
- this->file = reg.file;
- this->index = reg.index;
- this->swizzle = SWIZZLE_XYZW;
- this->negate = 0;
- this->abs = 0;
- this->reladdr = reg.reladdr;
- this->index2D = reg.index2D;
- this->reladdr2 = reg.reladdr2;
- this->has_index2 = reg.has_index2;
- this->double_reg2 = false;
- this->array_id = reg.array_id;
- this->is_double_vertex_input = false;
-}
-
-st_dst_reg::st_dst_reg(st_src_reg reg)
-{
- this->type = reg.type;
- this->file = reg.file;
- this->index = reg.index;
- this->writemask = WRITEMASK_XYZW;
- this->reladdr = reg.reladdr;
- this->index2D = reg.index2D;
- this->reladdr2 = reg.reladdr2;
- this->has_index2 = reg.has_index2;
- this->array_id = reg.array_id;
-}
-
-class glsl_to_tgsi_instruction : public exec_node {
-public:
- DECLARE_RALLOC_CXX_OPERATORS(glsl_to_tgsi_instruction)
-
- st_dst_reg dst[2];
- st_src_reg src[4];
- st_src_reg resource; /**< sampler, image or buffer register */
- st_src_reg *tex_offsets;
-
- /** Pointer to the ir source this tree came from for debugging */
- ir_instruction *ir;
-
- unsigned op:8; /**< TGSI opcode */
- unsigned precise:1;
- unsigned saturate:1;
- unsigned is_64bit_expanded:1;
- unsigned sampler_base:5;
- unsigned sampler_array_size:6; /**< 1-based size of sampler array, 1 if not array */
- unsigned tex_target:4; /**< One of TEXTURE_*_INDEX */
- glsl_base_type tex_type:5;
- unsigned tex_shadow:1;
- unsigned image_format:9;
- unsigned tex_offset_num_offset:3;
- unsigned dead_mask:4; /**< Used in dead code elimination */
- unsigned buffer_access:3; /**< buffer access type */
-
- const struct tgsi_opcode_info *info;
-};
-
class variable_storage {
DECLARE_RZALLOC_CXX_OPERATORS(variable_storage)
return GLSL_TYPE_ERROR;
}
-struct rename_reg_pair {
- int old_reg;
- int new_reg;
-};
-
struct glsl_to_tgsi_visitor : public ir_visitor {
public:
glsl_to_tgsi_visitor();
st_src_reg st_src_reg_for_double(double val);
st_src_reg st_src_reg_for_float(float val);
st_src_reg st_src_reg_for_int(int val);
+ st_src_reg st_src_reg_for_int64(int64_t val);
st_src_reg st_src_reg_for_type(enum glsl_base_type type, int val);
/**
void simplify_cmp(void);
- void rename_temp_registers(int num_renames, struct rename_reg_pair *renames);
+ void rename_temp_registers(struct rename_reg_pair *renames);
void get_first_temp_read(int *first_reads);
void get_first_temp_write(int *first_writes);
void get_last_temp_read_first_temp_write(int *last_reads, int *first_writes);
prog->data->LinkStatus = linking_failure;
}
-static int
+int
swizzle_for_size(int size)
{
static const int size_swizzles[4] = {
return size_swizzles[size - 1];
}
-static bool
-is_resource_instruction(unsigned opcode)
-{
- switch (opcode) {
- case TGSI_OPCODE_RESQ:
- case TGSI_OPCODE_LOAD:
- case TGSI_OPCODE_ATOMUADD:
- case TGSI_OPCODE_ATOMXCHG:
- case TGSI_OPCODE_ATOMCAS:
- case TGSI_OPCODE_ATOMAND:
- case TGSI_OPCODE_ATOMOR:
- case TGSI_OPCODE_ATOMXOR:
- case TGSI_OPCODE_ATOMUMIN:
- case TGSI_OPCODE_ATOMUMAX:
- case TGSI_OPCODE_ATOMIMIN:
- case TGSI_OPCODE_ATOMIMAX:
- return true;
- default:
- return false;
- }
-}
-
-static unsigned
-num_inst_dst_regs(const glsl_to_tgsi_instruction *op)
-{
- return op->info->num_dst;
-}
-
-static unsigned
-num_inst_src_regs(const glsl_to_tgsi_instruction *op)
-{
- return op->info->is_tex || is_resource_instruction(op->op) ?
- op->info->num_src - 1 : op->info->num_src;
-}
glsl_to_tgsi_instruction *
glsl_to_tgsi_visitor::emit_asm(ir_instruction *ir, unsigned op,
if (swz > 1) {
dinst->src[j].double_reg2 = true;
dinst->src[j].index++;
- }
+ }
if (swz & 1)
dinst->src[j].swizzle = MAKE_SWIZZLE4(SWIZZLE_Z, SWIZZLE_W, SWIZZLE_Z, SWIZZLE_W);
return src;
}
+st_src_reg
+glsl_to_tgsi_visitor::st_src_reg_for_int64(int64_t val)
+{
+ st_src_reg src(PROGRAM_IMMEDIATE, -1, GLSL_TYPE_INT64);
+ union gl_constant_value uval[2];
+
+ memcpy(uval, &val, sizeof(uval));
+ src.index = add_constant(src.file, uval, 1, GL_DOUBLE, &src.swizzle);
+ src.swizzle = MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_Y, SWIZZLE_X, SWIZZLE_Y);
+
+ return src;
+}
+
st_src_reg
glsl_to_tgsi_visitor::st_src_reg_for_type(enum glsl_base_type type, int val)
{
return type->count_attribute_slots(false);
}
+static void
+add_buffer_to_load_and_stores(glsl_to_tgsi_instruction *inst, st_src_reg *buf,
+ exec_list *instructions, ir_constant *access)
+{
+ /**
+ * emit_asm() might have actually split the op into pieces, e.g. for
+ * double stores. We have to go back and fix up all the generated ops.
+ */
+ unsigned op = inst->op;
+ do {
+ inst->resource = *buf;
+ if (access)
+ inst->buffer_access = access->value.u[0];
+
+ if (inst == instructions->get_head_raw())
+ break;
+ inst = (glsl_to_tgsi_instruction *)inst->get_prev();
+
+ if (inst->op == TGSI_OPCODE_UADD) {
+ if (inst == instructions->get_head_raw())
+ break;
+ inst = (glsl_to_tgsi_instruction *)inst->get_prev();
+ }
+ } while (inst->op == op && inst->resource.file == PROGRAM_UNDEFINED);
+}
+
/**
* If the given GLSL type is an array or matrix or a structure containing
* an array/matrix member, return true. Else return false.
break;
case ir_binop_ubo_load: {
- ir_constant *const_uniform_block = ir->operands[0]->as_constant();
- ir_constant *const_offset_ir = ir->operands[1]->as_constant();
- unsigned const_offset = const_offset_ir ? const_offset_ir->value.u[0] : 0;
- unsigned const_block = const_uniform_block ? const_uniform_block->value.u[0] + 1 : 1;
- st_src_reg index_reg = get_temp(glsl_type::uint_type);
- st_src_reg cbuf;
-
- cbuf.type = ir->type->base_type;
- cbuf.file = PROGRAM_CONSTANT;
- cbuf.index = 0;
- cbuf.reladdr = NULL;
- cbuf.negate = 0;
- cbuf.abs = 0;
- cbuf.index2D = const_block;
-
- assert(ir->type->is_vector() || ir->type->is_scalar());
-
- if (const_offset_ir) {
- /* Constant index into constant buffer */
- cbuf.reladdr = NULL;
- cbuf.index = const_offset / 16;
- }
- else {
- ir_expression *offset_expr = ir->operands[1]->as_expression();
- st_src_reg offset = op[1];
-
- /* The OpenGL spec is written in such a way that accesses with
- * non-constant offset are almost always vec4-aligned. The only
- * exception to this are members of structs in arrays of structs:
- * each struct in an array of structs is at least vec4-aligned,
- * but single-element and [ui]vec2 members of the struct may be at
- * an offset that is not a multiple of 16 bytes.
- *
- * Here, we extract that offset, relying on previous passes to always
- * generate offset expressions of the form (+ expr constant_offset).
- *
- * Note that the std430 layout, which allows more cases of alignment
- * less than vec4 in arrays, is not supported for uniform blocks, so
- * we do not have to deal with it here.
- */
- if (offset_expr && offset_expr->operation == ir_binop_add) {
- const_offset_ir = offset_expr->operands[1]->as_constant();
- if (const_offset_ir) {
- const_offset = const_offset_ir->value.u[0];
- cbuf.index = const_offset / 16;
- offset_expr->operands[0]->accept(this);
- offset = this->result;
- }
+ if (ctx->Const.UseSTD430AsDefaultPacking) {
+ ir_rvalue *block = ir->operands[0];
+ ir_rvalue *offset = ir->operands[1];
+ ir_constant *const_block = block->as_constant();
+
+ st_src_reg cbuf(PROGRAM_CONSTANT,
+ (const_block ? const_block->value.u[0] + 1 : 1),
+ ir->type->base_type);
+
+ cbuf.has_index2 = true;
+
+ if (!const_block) {
+ block->accept(this);
+ cbuf.reladdr = ralloc(mem_ctx, st_src_reg);
+ *cbuf.reladdr = this->result;
+ emit_arl(ir, sampler_reladdr, this->result);
}
- /* Relative/variable index into constant buffer */
- emit_asm(ir, TGSI_OPCODE_USHR, st_dst_reg(index_reg), offset,
- st_src_reg_for_int(4));
- cbuf.reladdr = ralloc(mem_ctx, st_src_reg);
- memcpy(cbuf.reladdr, &index_reg, sizeof(index_reg));
- }
+ /* Calculate the surface offset */
+ offset->accept(this);
+ st_src_reg off = this->result;
- if (const_uniform_block) {
- /* Constant constant buffer */
- cbuf.reladdr2 = NULL;
- }
- else {
- /* Relative/variable constant buffer */
- cbuf.reladdr2 = ralloc(mem_ctx, st_src_reg);
- memcpy(cbuf.reladdr2, &op[0], sizeof(st_src_reg));
- }
- cbuf.has_index2 = true;
-
- cbuf.swizzle = swizzle_for_size(ir->type->vector_elements);
- if (glsl_base_type_is_64bit(cbuf.type))
- cbuf.swizzle += MAKE_SWIZZLE4(const_offset % 16 / 8,
- const_offset % 16 / 8,
- const_offset % 16 / 8,
- const_offset % 16 / 8);
- else
- cbuf.swizzle += MAKE_SWIZZLE4(const_offset % 16 / 4,
- const_offset % 16 / 4,
- const_offset % 16 / 4,
- const_offset % 16 / 4);
+ glsl_to_tgsi_instruction *inst =
+ emit_asm(ir, TGSI_OPCODE_LOAD, result_dst, off);
- if (ir->type->is_boolean()) {
- emit_asm(ir, TGSI_OPCODE_USNE, result_dst, cbuf, st_src_reg_for_int(0));
+ if (result_dst.type == GLSL_TYPE_BOOL)
+ emit_asm(ir, TGSI_OPCODE_USNE, result_dst, st_src_reg(result_dst),
+ st_src_reg_for_int(0));
+
+ add_buffer_to_load_and_stores(inst, &cbuf, &this->instructions,
+ NULL);
} else {
- emit_asm(ir, TGSI_OPCODE_MOV, result_dst, cbuf);
+ ir_constant *const_uniform_block = ir->operands[0]->as_constant();
+ ir_constant *const_offset_ir = ir->operands[1]->as_constant();
+ unsigned const_offset = const_offset_ir ?
+ const_offset_ir->value.u[0] : 0;
+ unsigned const_block = const_uniform_block ?
+ const_uniform_block->value.u[0] + 1 : 1;
+ st_src_reg index_reg = get_temp(glsl_type::uint_type);
+ st_src_reg cbuf;
+
+ cbuf.type = ir->type->base_type;
+ cbuf.file = PROGRAM_CONSTANT;
+ cbuf.index = 0;
+ cbuf.reladdr = NULL;
+ cbuf.negate = 0;
+ cbuf.abs = 0;
+ cbuf.index2D = const_block;
+
+ assert(ir->type->is_vector() || ir->type->is_scalar());
+
+ if (const_offset_ir) {
+ /* Constant index into constant buffer */
+ cbuf.reladdr = NULL;
+ cbuf.index = const_offset / 16;
+ } else {
+ ir_expression *offset_expr = ir->operands[1]->as_expression();
+ st_src_reg offset = op[1];
+
+ /* The OpenGL spec is written in such a way that accesses with
+ * non-constant offset are almost always vec4-aligned. The only
+ * exception to this are members of structs in arrays of structs:
+ * each struct in an array of structs is at least vec4-aligned,
+ * but single-element and [ui]vec2 members of the struct may be at
+ * an offset that is not a multiple of 16 bytes.
+ *
+ * Here, we extract that offset, relying on previous passes to
+ * always generate offset expressions of the form
+ * (+ expr constant_offset).
+ *
+ * Note that the std430 layout, which allows more cases of
+ * alignment less than vec4 in arrays, is not supported for
+ * uniform blocks, so we do not have to deal with it here.
+ */
+ if (offset_expr && offset_expr->operation == ir_binop_add) {
+ const_offset_ir = offset_expr->operands[1]->as_constant();
+ if (const_offset_ir) {
+ const_offset = const_offset_ir->value.u[0];
+ cbuf.index = const_offset / 16;
+ offset_expr->operands[0]->accept(this);
+ offset = this->result;
+ }
+ }
+
+ /* Relative/variable index into constant buffer */
+ emit_asm(ir, TGSI_OPCODE_USHR, st_dst_reg(index_reg), offset,
+ st_src_reg_for_int(4));
+ cbuf.reladdr = ralloc(mem_ctx, st_src_reg);
+ memcpy(cbuf.reladdr, &index_reg, sizeof(index_reg));
+ }
+
+ if (const_uniform_block) {
+ /* Constant constant buffer */
+ cbuf.reladdr2 = NULL;
+ } else {
+ /* Relative/variable constant buffer */
+ cbuf.reladdr2 = ralloc(mem_ctx, st_src_reg);
+ memcpy(cbuf.reladdr2, &op[0], sizeof(st_src_reg));
+ }
+ cbuf.has_index2 = true;
+
+ cbuf.swizzle = swizzle_for_size(ir->type->vector_elements);
+ if (glsl_base_type_is_64bit(cbuf.type))
+ cbuf.swizzle += MAKE_SWIZZLE4(const_offset % 16 / 8,
+ const_offset % 16 / 8,
+ const_offset % 16 / 8,
+ const_offset % 16 / 8);
+ else
+ cbuf.swizzle += MAKE_SWIZZLE4(const_offset % 16 / 4,
+ const_offset % 16 / 4,
+ const_offset % 16 / 4,
+ const_offset % 16 / 4);
+
+ if (ir->type->is_boolean()) {
+ emit_asm(ir, TGSI_OPCODE_USNE, result_dst, cbuf,
+ st_src_reg_for_int(0));
+ } else {
+ emit_asm(ir, TGSI_OPCODE_MOV, result_dst, cbuf);
+ }
}
break;
}
st_src_reg temp = get_temp(glsl_type::uvec4_type);
st_dst_reg temp_dst = st_dst_reg(temp);
unsigned orig_swz = op[0].swizzle;
- /*
+ /*
* To convert unsigned to 64-bit:
* zero Y channel, copy X channel.
*/
break;
}
case ir_unop_i642b:
- emit_asm(ir, TGSI_OPCODE_U64SNE, result_dst, op[0], st_src_reg_for_int(0));
+ emit_asm(ir, TGSI_OPCODE_U64SNE, result_dst, op[0], st_src_reg_for_int64(0));
break;
case ir_unop_i642f:
emit_asm(ir, TGSI_OPCODE_I642F, result_dst, op[0]);
ir->array->accept(this);
src = this->result;
- if (ir->array->ir_type != ir_type_dereference_array) {
+ if (!src.has_index2) {
switch (this->prog->Target) {
case GL_TESS_CONTROL_PROGRAM_NV:
is_2D = (src.file == PROGRAM_INPUT || src.file == PROGRAM_OUTPUT) &&
if (index) {
if (this->prog->Target == GL_VERTEX_PROGRAM_ARB &&
- src.file == PROGRAM_INPUT)
- element_size = attrib_type_size(ir->type, true);
+ src.file == PROGRAM_INPUT)
+ element_size = attrib_type_size(ir->type, true);
if (is_2D) {
src.index2D = index->value.i[0];
src.has_index2 = true;
if (type->is_dual_slot()) {
l->index++;
if (r->is_double_vertex_input == false)
- r->index++;
+ r->index++;
}
}
inst = (glsl_to_tgsi_instruction *)this->instructions.get_tail();
new_inst = emit_asm(ir, inst->op, l, inst->src[0], inst->src[1], inst->src[2], inst->src[3]);
new_inst->saturate = inst->saturate;
+ new_inst->resource = inst->resource;
inst->dead_mask = inst->dst[0].writemask;
} else {
emit_block_mov(ir, ir->rhs->type, &l, &r, NULL, false);
st_src_reg temp_base = get_temp(ir->type);
st_dst_reg temp = st_dst_reg(temp_base);
- foreach_in_list(ir_constant, field_value, &ir->components) {
+ for (i = 0; i < ir->type->length; i++) {
+ ir_constant *const field_value = ir->get_record_field(i);
int size = type_size(field_value->type);
assert(size > 0);
in_array++;
for (i = 0; i < ir->type->length; i++) {
- ir->array_elements[i]->accept(this);
+ ir->const_elements[i]->accept(this);
src = this->result;
for (int j = 0; j < size; j++) {
emit_asm(ir, TGSI_OPCODE_MOV, temp, src);
assert(access);
}
- /* The emit_asm() might have actually split the op into pieces, e.g. for
- * double stores. We have to go back and fix up all the generated ops.
- */
- unsigned op = inst->op;
- do {
- inst->resource = buffer;
- if (access)
- inst->buffer_access = access->value.u[0];
-
- if (inst == this->instructions.get_head_raw())
- break;
- inst = (glsl_to_tgsi_instruction *)inst->get_prev();
-
- if (inst->op == TGSI_OPCODE_UADD) {
- if (inst == this->instructions.get_head_raw())
- break;
- inst = (glsl_to_tgsi_instruction *)inst->get_prev();
- }
- } while (inst->op == op && inst->resource.file == PROGRAM_UNDEFINED);
+ add_buffer_to_load_and_stores(inst, &buffer, &this->instructions, access);
}
void
/* Replaces all references to a temporary register index with another index. */
void
-glsl_to_tgsi_visitor::rename_temp_registers(int num_renames, struct rename_reg_pair *renames)
+glsl_to_tgsi_visitor::rename_temp_registers(struct rename_reg_pair *renames)
{
foreach_in_list(glsl_to_tgsi_instruction, inst, &this->instructions) {
unsigned j;
- int k;
for (j = 0; j < num_inst_src_regs(inst); j++) {
- if (inst->src[j].file == PROGRAM_TEMPORARY)
- for (k = 0; k < num_renames; k++)
- if (inst->src[j].index == renames[k].old_reg)
- inst->src[j].index = renames[k].new_reg;
+ if (inst->src[j].file == PROGRAM_TEMPORARY) {
+ int old_idx = inst->src[j].index;
+ if (renames[old_idx].valid)
+ inst->src[j].index = renames[old_idx].new_reg;
+ }
}
for (j = 0; j < inst->tex_offset_num_offset; j++) {
- if (inst->tex_offsets[j].file == PROGRAM_TEMPORARY)
- for (k = 0; k < num_renames; k++)
- if (inst->tex_offsets[j].index == renames[k].old_reg)
- inst->tex_offsets[j].index = renames[k].new_reg;
+ if (inst->tex_offsets[j].file == PROGRAM_TEMPORARY) {
+ int old_idx = inst->tex_offsets[j].index;
+ if (renames[old_idx].valid)
+ inst->tex_offsets[j].index = renames[old_idx].new_reg;
+ }
}
if (inst->resource.file == PROGRAM_TEMPORARY) {
- for (k = 0; k < num_renames; k++)
- if (inst->resource.index == renames[k].old_reg)
- inst->resource.index = renames[k].new_reg;
+ int old_idx = inst->resource.index;
+ if (renames[old_idx].valid)
+ inst->resource.index = renames[old_idx].new_reg;
}
for (j = 0; j < num_inst_dst_regs(inst); j++) {
- if (inst->dst[j].file == PROGRAM_TEMPORARY)
- for (k = 0; k < num_renames; k++)
- if (inst->dst[j].index == renames[k].old_reg)
- inst->dst[j].index = renames[k].new_reg;
+ if (inst->dst[j].file == PROGRAM_TEMPORARY) {
+ int old_idx = inst->dst[j].index;
+ if (renames[old_idx].valid)
+ inst->dst[j].index = renames[old_idx].new_reg;}
}
}
}
void
glsl_to_tgsi_visitor::merge_registers(void)
{
- int *last_reads = ralloc_array(mem_ctx, int, this->next_temp);
- int *first_writes = ralloc_array(mem_ctx, int, this->next_temp);
- struct rename_reg_pair *renames = rzalloc_array(mem_ctx, struct rename_reg_pair, this->next_temp);
- int i, j;
- int num_renames = 0;
- /* Read the indices of the last read and first write to each temp register
- * into an array so that we don't have to traverse the instruction list as
- * much. */
- for (i = 0; i < this->next_temp; i++) {
- last_reads[i] = -1;
- first_writes[i] = -1;
- }
- get_last_temp_read_first_temp_write(last_reads, first_writes);
+ struct lifetime *lifetimes =
+ rzalloc_array(mem_ctx, struct lifetime, this->next_temp);
- /* Start looking for registers with non-overlapping usages that can be
- * merged together. */
- for (i = 0; i < this->next_temp; i++) {
- /* Don't touch unused registers. */
- if (last_reads[i] < 0 || first_writes[i] < 0) continue;
-
- for (j = 0; j < this->next_temp; j++) {
- /* Don't touch unused registers. */
- if (last_reads[j] < 0 || first_writes[j] < 0) continue;
-
- /* We can merge the two registers if the first write to j is after or
- * in the same instruction as the last read from i. Note that the
- * register at index i will always be used earlier or at the same time
- * as the register at index j. */
- if (first_writes[i] <= first_writes[j] &&
- last_reads[i] <= first_writes[j]) {
- renames[num_renames].old_reg = j;
- renames[num_renames].new_reg = i;
- num_renames++;
-
- /* Update the first_writes and last_reads arrays with the new
- * values for the merged register index, and mark the newly unused
- * register index as such. */
- assert(last_reads[j] >= last_reads[i]);
- last_reads[i] = last_reads[j];
- first_writes[j] = -1;
- last_reads[j] = -1;
- }
- }
+ if (get_temp_registers_required_lifetimes(mem_ctx, &this->instructions,
+ this->next_temp, lifetimes)) {
+ struct rename_reg_pair *renames =
+ rzalloc_array(mem_ctx, struct rename_reg_pair, this->next_temp);
+ get_temp_registers_remapping(mem_ctx, this->next_temp, lifetimes, renames);
+ rename_temp_registers(renames);
+ ralloc_free(renames);
}
- rename_temp_registers(num_renames, renames);
- ralloc_free(renames);
- ralloc_free(last_reads);
- ralloc_free(first_writes);
+ ralloc_free(lifetimes);
}
/* Reassign indices to temporary registers by reusing unused indices created
int new_index = 0;
int *first_writes = ralloc_array(mem_ctx, int, this->next_temp);
struct rename_reg_pair *renames = rzalloc_array(mem_ctx, struct rename_reg_pair, this->next_temp);
- int num_renames = 0;
for (i = 0; i < this->next_temp; i++) {
first_writes[i] = -1;
for (i = 0; i < this->next_temp; i++) {
if (first_writes[i] < 0) continue;
if (i != new_index) {
- renames[num_renames].old_reg = i;
- renames[num_renames].new_reg = new_index;
- num_renames++;
+ renames[i].new_reg = new_index;
+ renames[i].valid = true;
}
new_index++;
}
- rename_temp_registers(num_renames, renames);
+ rename_temp_registers(renames);
this->next_temp = new_index;
ralloc_free(renames);
ralloc_free(first_writes);
}
/**
- * Map a glsl_to_tgsi src register to a TGSI ureg_src register.
+ * Create a TGSI ureg_dst register from an st_dst_reg.
+ */
+static struct ureg_dst
+translate_dst(struct st_translate *t,
+ const st_dst_reg *dst_reg,
+ bool saturate)
+{
+ struct ureg_dst dst = dst_register(t, dst_reg->file, dst_reg->index,
+ dst_reg->array_id);
+
+ if (dst.File == TGSI_FILE_NULL)
+ return dst;
+
+ dst = ureg_writemask(dst, dst_reg->writemask);
+
+ if (saturate)
+ dst = ureg_saturate(dst);
+
+ if (dst_reg->reladdr != NULL) {
+ assert(dst_reg->file != PROGRAM_TEMPORARY);
+ dst = ureg_dst_indirect(dst, ureg_src(t->address[0]));
+ }
+
+ if (dst_reg->has_index2) {
+ if (dst_reg->reladdr2)
+ dst = ureg_dst_dimension_indirect(dst, ureg_src(t->address[1]),
+ dst_reg->index2D);
+ else
+ dst = ureg_dst_dimension(dst, dst_reg->index2D);
+ }
+
+ return dst;
+}
+
+/**
+ * Create a TGSI ureg_src register from an st_src_reg.
*/
static struct ureg_src
-src_register(struct st_translate *t, const st_src_reg *reg)
+translate_src(struct st_translate *t, const st_src_reg *src_reg)
{
- int index = reg->index;
- int double_reg2 = reg->double_reg2 ? 1 : 0;
+ struct ureg_src src;
+ int index = src_reg->index;
+ int double_reg2 = src_reg->double_reg2 ? 1 : 0;
- switch(reg->file) {
+ switch(src_reg->file) {
case PROGRAM_UNDEFINED:
- return ureg_imm4f(t->ureg, 0, 0, 0, 0);
+ src = ureg_imm4f(t->ureg, 0, 0, 0, 0);
+ break;
case PROGRAM_TEMPORARY:
case PROGRAM_ARRAY:
- return ureg_src(dst_register(t, reg->file, reg->index, reg->array_id));
+ src = ureg_src(dst_register(t, src_reg->file, src_reg->index, src_reg->array_id));
+ break;
case PROGRAM_OUTPUT: {
- struct ureg_dst dst = dst_register(t, reg->file, reg->index, reg->array_id);
+ struct ureg_dst dst = dst_register(t, src_reg->file, src_reg->index, src_reg->array_id);
assert(dst.WriteMask != 0);
unsigned shift = ffs(dst.WriteMask) - 1;
- return ureg_swizzle(ureg_src(dst),
- shift,
- MIN2(shift + 1, 3),
- MIN2(shift + 2, 3),
- MIN2(shift + 3, 3));
+ src = ureg_swizzle(ureg_src(dst),
+ shift,
+ MIN2(shift + 1, 3),
+ MIN2(shift + 2, 3),
+ MIN2(shift + 3, 3));
+ break;
}
case PROGRAM_UNIFORM:
- assert(reg->index >= 0);
- return reg->index < t->num_constants ?
- t->constants[reg->index] : ureg_imm4f(t->ureg, 0, 0, 0, 0);
+ assert(src_reg->index >= 0);
+ src = src_reg->index < t->num_constants ?
+ t->constants[src_reg->index] : ureg_imm4f(t->ureg, 0, 0, 0, 0);
+ break;
case PROGRAM_STATE_VAR:
case PROGRAM_CONSTANT: /* ie, immediate */
- if (reg->has_index2)
- return ureg_src_register(TGSI_FILE_CONSTANT, reg->index);
+ if (src_reg->has_index2)
+ src = ureg_src_register(TGSI_FILE_CONSTANT, src_reg->index);
else
- return reg->index >= 0 && reg->index < t->num_constants ?
- t->constants[reg->index] : ureg_imm4f(t->ureg, 0, 0, 0, 0);
+ src = src_reg->index >= 0 && src_reg->index < t->num_constants ?
+ t->constants[src_reg->index] : ureg_imm4f(t->ureg, 0, 0, 0, 0);
+ break;
case PROGRAM_IMMEDIATE:
- assert(reg->index >= 0 && reg->index < t->num_immediates);
- return t->immediates[reg->index];
+ assert(src_reg->index >= 0 && src_reg->index < t->num_immediates);
+ src = t->immediates[src_reg->index];
+ break;
case PROGRAM_INPUT:
/* GLSL inputs are 64-bit containers, so we have to
* map back to the original index and add the offset after
* mapping. */
index -= double_reg2;
- if (!reg->array_id) {
+ if (!src_reg->array_id) {
assert(t->inputMapping[index] < ARRAY_SIZE(t->inputs));
assert(t->inputs[t->inputMapping[index]].File != TGSI_FILE_NULL);
- return t->inputs[t->inputMapping[index] + double_reg2];
+ src = t->inputs[t->inputMapping[index] + double_reg2];
}
else {
- struct inout_decl *decl = find_inout_array(t->input_decls, t->num_input_decls, reg->array_id);
+ struct inout_decl *decl = find_inout_array(t->input_decls, t->num_input_decls,
+ src_reg->array_id);
unsigned mesa_index = decl->mesa_index;
int slot = t->inputMapping[mesa_index];
assert(slot != -1 && t->inputs[slot].File == TGSI_FILE_INPUT);
- struct ureg_src src = t->inputs[slot];
- src.ArrayID = reg->array_id;
- return ureg_src_array_offset(src, index + double_reg2 - mesa_index);
+ src = t->inputs[slot];
+ src.ArrayID = src_reg->array_id;
+ src = ureg_src_array_offset(src, index + double_reg2 - mesa_index);
}
+ break;
case PROGRAM_ADDRESS:
- return ureg_src(t->address[reg->index]);
+ src = ureg_src(t->address[src_reg->index]);
+ break;
case PROGRAM_SYSTEM_VALUE:
- assert(reg->index < (int) ARRAY_SIZE(t->systemValues));
- return t->systemValues[reg->index];
+ assert(src_reg->index < (int) ARRAY_SIZE(t->systemValues));
+ src = t->systemValues[src_reg->index];
+ break;
default:
assert(!"unknown src register file");
return ureg_src_undef();
}
-}
-
-/**
- * Create a TGSI ureg_dst register from an st_dst_reg.
- */
-static struct ureg_dst
-translate_dst(struct st_translate *t,
- const st_dst_reg *dst_reg,
- bool saturate)
-{
- struct ureg_dst dst = dst_register(t, dst_reg->file, dst_reg->index,
- dst_reg->array_id);
-
- if (dst.File == TGSI_FILE_NULL)
- return dst;
-
- dst = ureg_writemask(dst, dst_reg->writemask);
-
- if (saturate)
- dst = ureg_saturate(dst);
-
- if (dst_reg->reladdr != NULL) {
- assert(dst_reg->file != PROGRAM_TEMPORARY);
- dst = ureg_dst_indirect(dst, ureg_src(t->address[0]));
- }
-
- if (dst_reg->has_index2) {
- if (dst_reg->reladdr2)
- dst = ureg_dst_dimension_indirect(dst, ureg_src(t->address[1]),
- dst_reg->index2D);
- else
- dst = ureg_dst_dimension(dst, dst_reg->index2D);
- }
-
- return dst;
-}
-
-/**
- * Create a TGSI ureg_src register from an st_src_reg.
- */
-static struct ureg_src
-translate_src(struct st_translate *t, const st_src_reg *src_reg)
-{
- struct ureg_src src = src_register(t, src_reg);
if (src_reg->has_index2) {
/* 2D indexes occur with geometry shader inputs (attrib, vertex)
src[0] = t->shared_memory;
} else if (inst->resource.file == PROGRAM_BUFFER) {
src[0] = t->buffers[inst->resource.index];
+ } else if (inst->resource.file == PROGRAM_CONSTANT) {
+ assert(inst->resource.has_index2);
+ src[0] = ureg_src_register(TGSI_FILE_CONSTBUF, inst->resource.index);
} else {
+ assert(inst->resource.file != PROGRAM_UNDEFINED);
if (inst->resource.file == PROGRAM_IMAGE) {
src[0] = t->images[inst->resource.index];
} else {
projects / mesa.git / blobdiff