const struct glsl_type *last_interface_type;
struct nir_variable_data last_var_data;
+ /* For skipping equal ALU headers (typical after scalarization). */
+ nir_instr_type last_instr_type;
+ uintptr_t last_alu_header_offset;
+
/* Don't write optional data such as variable names. */
bool strip;
} write_ctx;
unsigned writemask:4;
unsigned op:9;
unsigned packed_src_ssa_16bit:1;
- unsigned _pad:2;
+ /* Scalarized ALUs always have the same header. */
+ unsigned num_followup_alu_sharing_header:2;
unsigned dest:8;
} alu;
struct {
unsigned deref_type:3;
unsigned cast_type_same_as_last:1;
unsigned mode:10; /* deref_var redefines this */
- unsigned _pad:6; /* deref_var redefines this */
+ unsigned packed_src_ssa_16bit:1; /* deref_var redefines this */
+ unsigned _pad:5; /* deref_var redefines this */
unsigned dest:8;
} deref;
struct {
/* Write "lo24" as low 24 bits in the first uint32. */
static void
-write_dest(write_ctx *ctx, const nir_dest *dst, union packed_instr header)
+write_dest(write_ctx *ctx, const nir_dest *dst, union packed_instr header,
+ nir_instr_type instr_type)
{
STATIC_ASSERT(sizeof(union packed_dest) == 1);
union packed_dest dest;
} else {
dest.reg.is_indirect = !!(dst->reg.indirect);
}
-
header.any.dest = dest.u8;
- blob_write_uint32(ctx->blob, header.u32);
+
+ /* Check if the current ALU instruction has the same header as the previous
+ * instruction that is also ALU. If it is, we don't have to write
+ * the current header. This is a typical occurence after scalarization.
+ */
+ if (instr_type == nir_instr_type_alu) {
+ bool equal_header = false;
+
+ if (ctx->last_instr_type == nir_instr_type_alu) {
+ assert(ctx->last_alu_header_offset);
+ union packed_instr *last_header =
+ (union packed_instr *)(ctx->blob->data +
+ ctx->last_alu_header_offset);
+
+ /* Clear the field that counts ALUs with equal headers. */
+ union packed_instr clean_header;
+ clean_header.u32 = last_header->u32;
+ clean_header.alu.num_followup_alu_sharing_header = 0;
+
+ /* There can be at most 4 consecutive ALU instructions
+ * sharing the same header.
+ */
+ if (last_header->alu.num_followup_alu_sharing_header < 3 &&
+ header.u32 == clean_header.u32) {
+ last_header->alu.num_followup_alu_sharing_header++;
+ equal_header = true;
+ }
+ }
+
+ if (!equal_header) {
+ ctx->last_alu_header_offset = ctx->blob->size;
+ blob_write_uint32(ctx->blob, header.u32);
+ }
+ } else {
+ blob_write_uint32(ctx->blob, header.u32);
+ }
if (dst->is_ssa) {
write_add_object(ctx, &dst->ssa);
header.alu.op = alu->op;
header.alu.packed_src_ssa_16bit = is_alu_src_ssa_16bit(ctx, alu);
- write_dest(ctx, &alu->dest.dest, header);
+ write_dest(ctx, &alu->dest.dest, header, alu->instr.type);
if (header.alu.packed_src_ssa_16bit) {
for (unsigned i = 0; i < num_srcs; i++) {
header.deref_var.object_idx = var_idx;
}
- write_dest(ctx, &deref->dest, header);
+ if (deref->deref_type == nir_deref_type_array ||
+ deref->deref_type == nir_deref_type_ptr_as_array) {
+ header.deref.packed_src_ssa_16bit =
+ deref->parent.is_ssa && deref->arr.index.is_ssa &&
+ are_object_ids_16bit(ctx);
+ }
+
+ write_dest(ctx, &deref->dest, header, deref->instr.type);
switch (deref->deref_type) {
case nir_deref_type_var:
case nir_deref_type_array:
case nir_deref_type_ptr_as_array:
- write_src(ctx, &deref->parent);
- write_src(ctx, &deref->arr.index);
+ if (header.deref.packed_src_ssa_16bit) {
+ blob_write_uint16(ctx->blob,
+ write_lookup_object(ctx, deref->parent.ssa));
+ blob_write_uint16(ctx->blob,
+ write_lookup_object(ctx, deref->arr.index.ssa));
+ } else {
+ write_src(ctx, &deref->parent);
+ write_src(ctx, &deref->arr.index);
+ }
break;
case nir_deref_type_cast:
case nir_deref_type_array:
case nir_deref_type_ptr_as_array:
- read_src(ctx, &deref->parent, &deref->instr);
+ if (header.deref.packed_src_ssa_16bit) {
+ deref->parent.is_ssa = true;
+ deref->parent.ssa = read_lookup_object(ctx, blob_read_uint16(ctx->blob));
+ deref->arr.index.is_ssa = true;
+ deref->arr.index.ssa = read_lookup_object(ctx, blob_read_uint16(ctx->blob));
+ } else {
+ read_src(ctx, &deref->parent, &deref->instr);
+ read_src(ctx, &deref->arr.index, &deref->instr);
+ }
+
parent = nir_src_as_deref(deref->parent);
if (deref->deref_type == nir_deref_type_array)
deref->type = glsl_get_array_element(parent->type);
else
deref->type = parent->type;
- read_src(ctx, &deref->arr.index, &deref->instr);
break;
case nir_deref_type_cast:
}
if (nir_intrinsic_infos[intrin->intrinsic].has_dest)
- write_dest(ctx, &intrin->dest, header);
+ write_dest(ctx, &intrin->dest, header, intrin->instr.type);
else
blob_write_uint32(ctx->blob, header.u32);
header.tex.op = tex->op;
header.tex.texture_array_size = tex->texture_array_size;
- write_dest(ctx, &tex->dest, header);
+ write_dest(ctx, &tex->dest, header, tex->instr.type);
blob_write_uint32(ctx->blob, tex->texture_index);
blob_write_uint32(ctx->blob, tex->sampler_index);
* and then store enough information so that a later fixup pass can fill
* them in correctly.
*/
- write_dest(ctx, &phi->dest, header);
+ write_dest(ctx, &phi->dest, header, phi->instr.type);
nir_foreach_phi_src(src, phi) {
assert(src->src.is_ssa);
}
}
-static void
+/* Return the number of instructions read. */
+static unsigned
read_instr(read_ctx *ctx, nir_block *block)
{
STATIC_ASSERT(sizeof(union packed_instr) == 4);
switch (header.any.instr_type) {
case nir_instr_type_alu:
- instr = &read_alu(ctx, header)->instr;
- break;
+ for (unsigned i = 0; i <= header.alu.num_followup_alu_sharing_header; i++)
+ nir_instr_insert_after_block(block, &read_alu(ctx, header)->instr);
+ return header.alu.num_followup_alu_sharing_header + 1;
case nir_instr_type_deref:
instr = &read_deref(ctx, header)->instr;
break;
* are read so that we can set their sources up.
*/
read_phi(ctx, block, header);
- return;
+ return 1;
case nir_instr_type_jump:
instr = &read_jump(ctx, header)->instr;
break;
}
nir_instr_insert_after_block(block, instr);
+ return 1;
}
static void
{
write_add_object(ctx, block);
blob_write_uint32(ctx->blob, exec_list_length(&block->instr_list));
- nir_foreach_instr(instr, block)
+
+ ctx->last_instr_type = ~0;
+ ctx->last_alu_header_offset = 0;
+
+ nir_foreach_instr(instr, block) {
write_instr(ctx, instr);
+ ctx->last_instr_type = instr->type;
+ }
}
static void
read_add_object(ctx, block);
unsigned num_instrs = blob_read_uint32(ctx->blob);
- for (unsigned i = 0; i < num_instrs; i++) {
- read_instr(ctx, block);
+ for (unsigned i = 0; i < num_instrs;) {
+ i += read_instr(ctx, block);
}
}
projects / mesa.git / blobdiff