#define WORK_STRIDE 10
+/* We have overlapping register classes for special registers, handled via
+ * shadows */
+
+#define SHADOW_R28 18
+#define SHADOW_R29 19
+
/* Prepacked masks/swizzles for virtual register types */
static unsigned reg_type_to_mask[WORK_STRIDE] = {
0xF, /* xyzw */
return shifted;
}
-/* When we're 'squeezing down' the values in the IR, we maintain a hash
- * as such */
-
-static unsigned
-find_or_allocate_temp(compiler_context *ctx, unsigned hash)
-{
- if ((hash < 0) || (hash >= SSA_FIXED_MINIMUM))
- return hash;
-
- unsigned temp = (uintptr_t) _mesa_hash_table_u64_search(
- ctx->hash_to_temp, hash + 1);
-
- if (temp)
- return temp - 1;
-
- /* If no temp is find, allocate one */
- temp = ctx->temp_count++;
- ctx->max_hash = MAX2(ctx->max_hash, hash);
-
- _mesa_hash_table_u64_insert(ctx->hash_to_temp,
- hash + 1, (void *) ((uintptr_t) temp + 1));
-
- return temp;
-}
-
/* Helper to return the default phys_reg for a given register */
static struct phys_reg
int phys = virt / WORK_STRIDE;
int type = virt % WORK_STRIDE;
+ /* Apply shadow registers */
+
+ if (phys >= SHADOW_R28 && phys <= SHADOW_R29)
+ phys += 28 - SHADOW_R28;
+
struct phys_reg r = {
.reg = phys,
.mask = reg_type_to_mask[type],
};
/* Report that we actually use this register, and return it */
- ctx->work_registers = MAX2(ctx->work_registers, phys);
+
+ if (phys < 16)
+ ctx->work_registers = MAX2(ctx->work_registers, phys);
+
return r;
}
-/* This routine performs the actual register allocation. It should be succeeded
- * by install_registers */
+/* This routine creates a register set. Should be called infrequently since
+ * it's slow and can be cached. For legibility, variables are named in terms of
+ * work registers, although it is also used to create the register set for
+ * special register allocation */
-struct ra_graph *
-allocate_registers(compiler_context *ctx, bool *spilled)
+static void
+add_shadow_conflicts (struct ra_regs *regs, unsigned base, unsigned shadow)
{
- /* The number of vec4 work registers available depends on when the
- * uniforms start, so compute that first */
+ for (unsigned a = 0; a < WORK_STRIDE; ++a) {
+ unsigned reg_a = (WORK_STRIDE * base) + a;
- int work_count = 16 - MAX2((ctx->uniform_cutoff - 8), 0);
+ for (unsigned b = 0; b < WORK_STRIDE; ++b) {
+ unsigned reg_b = (WORK_STRIDE * shadow) + b;
- int virtual_count = work_count * WORK_STRIDE;
+ ra_add_reg_conflict(regs, reg_a, reg_b);
+ ra_add_reg_conflict(regs, reg_b, reg_a);
+ }
+ }
+}
+
+static struct ra_regs *
+create_register_set(unsigned work_count, unsigned *classes)
+{
+ int virtual_count = 32 * WORK_STRIDE;
/* First, initialize the RA */
struct ra_regs *regs = ra_alloc_reg_set(NULL, virtual_count, true);
- int work_vec4 = ra_alloc_reg_class(regs);
- int work_vec3 = ra_alloc_reg_class(regs);
- int work_vec2 = ra_alloc_reg_class(regs);
- int work_vec1 = ra_alloc_reg_class(regs);
-
- unsigned classes[4] = {
- work_vec1,
- work_vec2,
- work_vec3,
- work_vec4
- };
-
- /* Add the full set of work registers */
- for (unsigned i = 0; i < work_count; ++i) {
- int base = WORK_STRIDE * i;
-
- /* Build a full set of subdivisions */
- ra_class_add_reg(regs, work_vec4, base);
- ra_class_add_reg(regs, work_vec3, base + 1);
- ra_class_add_reg(regs, work_vec3, base + 2);
- ra_class_add_reg(regs, work_vec2, base + 3);
- ra_class_add_reg(regs, work_vec2, base + 4);
- ra_class_add_reg(regs, work_vec2, base + 5);
- ra_class_add_reg(regs, work_vec1, base + 6);
- ra_class_add_reg(regs, work_vec1, base + 7);
- ra_class_add_reg(regs, work_vec1, base + 8);
- ra_class_add_reg(regs, work_vec1, base + 9);
-
- for (unsigned a = 0; a < 10; ++a) {
- unsigned mask1 = reg_type_to_mask[a];
-
- for (unsigned b = 0; b < 10; ++b) {
- unsigned mask2 = reg_type_to_mask[b];
-
- if (mask1 & mask2)
- ra_add_reg_conflict(regs,
- base + a, base + b);
+ for (unsigned c = 0; c < NR_REG_CLASSES; ++c) {
+ int work_vec4 = ra_alloc_reg_class(regs);
+ int work_vec3 = ra_alloc_reg_class(regs);
+ int work_vec2 = ra_alloc_reg_class(regs);
+ int work_vec1 = ra_alloc_reg_class(regs);
+
+ classes[4*c + 0] = work_vec1;
+ classes[4*c + 1] = work_vec2;
+ classes[4*c + 2] = work_vec3;
+ classes[4*c + 3] = work_vec4;
+
+ /* Special register classes have other register counts */
+ unsigned count =
+ (c == REG_CLASS_WORK) ? work_count : 2;
+
+ unsigned first_reg =
+ (c == REG_CLASS_LDST) ? 26 :
+ (c == REG_CLASS_TEXR) ? 28 :
+ (c == REG_CLASS_TEXW) ? SHADOW_R28 :
+ 0;
+
+ /* Add the full set of work registers */
+ for (unsigned i = first_reg; i < (first_reg + count); ++i) {
+ int base = WORK_STRIDE * i;
+
+ /* Build a full set of subdivisions */
+ ra_class_add_reg(regs, work_vec4, base);
+ ra_class_add_reg(regs, work_vec3, base + 1);
+ ra_class_add_reg(regs, work_vec3, base + 2);
+ ra_class_add_reg(regs, work_vec2, base + 3);
+ ra_class_add_reg(regs, work_vec2, base + 4);
+ ra_class_add_reg(regs, work_vec2, base + 5);
+ ra_class_add_reg(regs, work_vec1, base + 6);
+ ra_class_add_reg(regs, work_vec1, base + 7);
+ ra_class_add_reg(regs, work_vec1, base + 8);
+ ra_class_add_reg(regs, work_vec1, base + 9);
+
+ for (unsigned a = 0; a < 10; ++a) {
+ unsigned mask1 = reg_type_to_mask[a];
+
+ for (unsigned b = 0; b < 10; ++b) {
+ unsigned mask2 = reg_type_to_mask[b];
+
+ if (mask1 & mask2)
+ ra_add_reg_conflict(regs,
+ base + a, base + b);
+ }
}
}
}
+
+ /* We have duplicate classes */
+ add_shadow_conflicts(regs, 28, SHADOW_R28);
+ add_shadow_conflicts(regs, 29, SHADOW_R29);
+
/* We're done setting up */
ra_set_finalize(regs, NULL);
- /* Transform the MIR into squeezed index form */
- mir_foreach_block(ctx, block) {
- mir_foreach_instr_in_block(block, ins) {
- if (ins->compact_branch) continue;
+ return regs;
+}
+
+/* This routine gets a precomputed register set off the screen if it's able, or
+ * otherwise it computes one on the fly */
+
+static struct ra_regs *
+get_register_set(struct midgard_screen *screen, unsigned work_count, unsigned **classes)
+{
+ /* Bounds check */
+ assert(work_count >= 8);
+ assert(work_count <= 16);
+
+ /* Compute index */
+ unsigned index = work_count - 8;
+
+ /* Find the reg set */
+ struct ra_regs *cached = screen->regs[index];
+
+ if (cached) {
+ assert(screen->reg_classes[index]);
+ *classes = screen->reg_classes[index];
+ return cached;
+ }
+
+ /* Otherwise, create one */
+ struct ra_regs *created = create_register_set(work_count, screen->reg_classes[index]);
+
+ /* Cache it and use it */
+ screen->regs[index] = created;
+
+ *classes = screen->reg_classes[index];
+ return created;
+}
+
+/* Assign a (special) class, ensuring that it is compatible with whatever class
+ * was already set */
+
+static void
+set_class(unsigned *classes, unsigned node, unsigned class)
+{
+ /* Check that we're even a node */
+ if ((node < 0) || (node >= SSA_FIXED_MINIMUM))
+ return;
+
+ /* First 4 are work, next 4 are load/store.. */
+ unsigned current_class = classes[node] >> 2;
+
+ /* Nothing to do */
+ if (class == current_class)
+ return;
+
+ /* If we're changing, we haven't assigned a special class */
+ assert(current_class == REG_CLASS_WORK);
+
+ classes[node] &= 0x3;
+ classes[node] |= (class << 2);
+}
+
+static void
+force_vec4(unsigned *classes, unsigned node)
+{
+ if ((node < 0) || (node >= SSA_FIXED_MINIMUM))
+ return;
+
+ /* Force vec4 = 3 */
+ classes[node] |= 0x3;
+}
+
+/* Special register classes impose special constraints on who can read their
+ * values, so check that */
+
+static bool
+check_read_class(unsigned *classes, unsigned tag, unsigned node)
+{
+ /* Non-nodes are implicitly ok */
+ if ((node < 0) || (node >= SSA_FIXED_MINIMUM))
+ return true;
+
+ unsigned current_class = classes[node] >> 2;
+
+ switch (current_class) {
+ case REG_CLASS_LDST:
+ return (tag == TAG_LOAD_STORE_4);
+ case REG_CLASS_TEXR:
+ return (tag == TAG_TEXTURE_4);
+ case REG_CLASS_TEXW:
+ return (tag != TAG_LOAD_STORE_4);
+ case REG_CLASS_WORK:
+ return (tag == TAG_ALU_4);
+ default:
+ unreachable("Invalid class");
+ }
+}
+
+static bool
+check_write_class(unsigned *classes, unsigned tag, unsigned node)
+{
+ /* Non-nodes are implicitly ok */
+ if ((node < 0) || (node >= SSA_FIXED_MINIMUM))
+ return true;
+
+ unsigned current_class = classes[node] >> 2;
+
+ switch (current_class) {
+ case REG_CLASS_TEXR:
+ return true;
+ case REG_CLASS_TEXW:
+ return (tag == TAG_TEXTURE_4);
+ case REG_CLASS_LDST:
+ case REG_CLASS_WORK:
+ return (tag == TAG_ALU_4) || (tag == TAG_LOAD_STORE_4);
+ default:
+ unreachable("Invalid class");
+ }
+}
+
+/* Prepass before RA to ensure special class restrictions are met. The idea is
+ * to create a bit field of types of instructions that read a particular index.
+ * Later, we'll add moves as appropriate and rewrite to specialize by type. */
- ins->ssa_args.dest = find_or_allocate_temp(ctx, ins->ssa_args.dest);
- ins->ssa_args.src0 = find_or_allocate_temp(ctx, ins->ssa_args.src0);
+static void
+mark_node_class (unsigned *bitfield, unsigned node)
+{
+ if ((node >= 0) && (node < SSA_FIXED_MINIMUM))
+ BITSET_SET(bitfield, node);
+}
+
+void
+mir_lower_special_reads(compiler_context *ctx)
+{
+ size_t sz = BITSET_WORDS(ctx->temp_count) * sizeof(BITSET_WORD);
+
+ /* Bitfields for the various types of registers we could have */
+
+ unsigned *alur = calloc(sz, 1);
+ unsigned *aluw = calloc(sz, 1);
+ unsigned *ldst = calloc(sz, 1);
+ unsigned *texr = calloc(sz, 1);
+ unsigned *texw = calloc(sz, 1);
+
+ /* Pass #1 is analysis, a linear scan to fill out the bitfields */
+
+ mir_foreach_instr_global(ctx, ins) {
+ switch (ins->type) {
+ case TAG_ALU_4:
+ mark_node_class(aluw, ins->ssa_args.dest);
+ mark_node_class(alur, ins->ssa_args.src[0]);
+ mark_node_class(alur, ins->ssa_args.src[1]);
+ break;
+
+ case TAG_LOAD_STORE_4:
+ mark_node_class(ldst, ins->ssa_args.src[0]);
+ mark_node_class(ldst, ins->ssa_args.src[1]);
+ mark_node_class(ldst, ins->ssa_args.src[2]);
+ break;
+
+ case TAG_TEXTURE_4:
+ mark_node_class(texr, ins->ssa_args.src[0]);
+ mark_node_class(texr, ins->ssa_args.src[1]);
+ mark_node_class(texr, ins->ssa_args.src[2]);
+ mark_node_class(texw, ins->ssa_args.dest);
+ break;
+ }
+ }
- if (!ins->ssa_args.inline_constant)
- ins->ssa_args.src1 = find_or_allocate_temp(ctx, ins->ssa_args.src1);
+ /* Pass #2 is lowering now that we've analyzed all the classes.
+ * Conceptually, if an index is only marked for a single type of use,
+ * there is nothing to lower. If it is marked for different uses, we
+ * split up based on the number of types of uses. To do so, we divide
+ * into N distinct classes of use (where N>1 by definition), emit N-1
+ * moves from the index to copies of the index, and finally rewrite N-1
+ * of the types of uses to use the corresponding move */
+ unsigned spill_idx = ctx->temp_count;
+
+ for (unsigned i = 0; i < ctx->temp_count; ++i) {
+ bool is_alur = BITSET_TEST(alur, i);
+ bool is_aluw = BITSET_TEST(aluw, i);
+ bool is_ldst = BITSET_TEST(ldst, i);
+ bool is_texr = BITSET_TEST(texr, i);
+ bool is_texw = BITSET_TEST(texw, i);
+
+ /* Analyse to check how many distinct uses there are. ALU ops
+ * (alur) can read the results of the texture pipeline (texw)
+ * but not ldst or texr. Load/store ops (ldst) cannot read
+ * anything but load/store inputs. Texture pipeline cannot read
+ * anything but texture inputs. TODO: Simplify. */
+
+ bool collision =
+ (is_alur && (is_ldst || is_texr)) ||
+ (is_ldst && (is_alur || is_texr || is_texw)) ||
+ (is_texr && (is_alur || is_ldst || is_texw)) ||
+ (is_texw && (is_aluw || is_ldst || is_texr));
+
+ if (!collision)
+ continue;
+
+ /* Use the index as-is as the work copy. Emit copies for
+ * special uses */
+
+ unsigned classes[] = { TAG_LOAD_STORE_4, TAG_TEXTURE_4, TAG_TEXTURE_4 };
+ bool collisions[] = { is_ldst, is_texr, is_texw && is_aluw };
+
+ for (unsigned j = 0; j < ARRAY_SIZE(collisions); ++j) {
+ if (!collisions[j]) continue;
+
+ /* When the hazard is from reading, we move and rewrite
+ * sources (typical case). When it's from writing, we
+ * flip the move and rewrite destinations (obscure,
+ * only from control flow -- impossible in SSA) */
+
+ bool hazard_write = (j == 2);
+
+ unsigned idx = spill_idx++;
+
+ midgard_instruction m = hazard_write ?
+ v_mov(idx, blank_alu_src, i) :
+ v_mov(i, blank_alu_src, idx);
+
+ /* Insert move before each read/write, depending on the
+ * hazard we're trying to account for */
+
+ mir_foreach_instr_global_safe(ctx, pre_use) {
+ if (pre_use->type != classes[j])
+ continue;
+
+ if (hazard_write) {
+ if (pre_use->ssa_args.dest != i)
+ continue;
+ } else {
+ if (!mir_has_arg(pre_use, i))
+ continue;
+ }
+
+ if (hazard_write) {
+ midgard_instruction *use = mir_next_op(pre_use);
+ assert(use);
+ mir_insert_instruction_before(use, m);
+ } else {
+ mir_insert_instruction_before(pre_use, m);
+ }
+ }
+
+ /* Rewrite to use */
+ if (hazard_write)
+ mir_rewrite_index_dst_tag(ctx, i, idx, classes[j]);
+ else
+ mir_rewrite_index_src_tag(ctx, i, idx, classes[j]);
}
}
- /* No register allocation to do with no SSA */
+ free(alur);
+ free(aluw);
+ free(ldst);
+ free(texr);
+ free(texw);
+}
+
+/* This routine performs the actual register allocation. It should be succeeded
+ * by install_registers */
+
+struct ra_graph *
+allocate_registers(compiler_context *ctx, bool *spilled)
+{
+ /* The number of vec4 work registers available depends on when the
+ * uniforms start, so compute that first */
+ int work_count = 16 - MAX2((ctx->uniform_cutoff - 8), 0);
+ unsigned *classes = NULL;
+ struct ra_regs *regs = get_register_set(ctx->screen, work_count, &classes);
+
+ assert(regs != NULL);
+ assert(classes != NULL);
+
+ /* No register allocation to do with no SSA */
if (!ctx->temp_count)
return NULL;
/* Let's actually do register allocation */
int nodes = ctx->temp_count;
struct ra_graph *g = ra_alloc_interference_graph(regs, nodes);
-
- /* Determine minimum size needed to hold values, to indirectly
- * determine class */
+
+ /* Register class (as known to the Mesa register allocator) is actually
+ * the product of both semantic class (work, load/store, texture..) and
+ * size (vec2/vec3..). First, we'll go through and determine the
+ * minimum size needed to hold values */
unsigned *found_class = calloc(sizeof(unsigned), ctx->temp_count);
- mir_foreach_block(ctx, block) {
- mir_foreach_instr_in_block(block, ins) {
- if (ins->compact_branch) continue;
- if (ins->ssa_args.dest < 0) continue;
- if (ins->ssa_args.dest >= SSA_FIXED_MINIMUM) continue;
+ mir_foreach_instr_global(ctx, ins) {
+ if (ins->ssa_args.dest < 0) continue;
+ if (ins->ssa_args.dest >= SSA_FIXED_MINIMUM) continue;
+
+ /* 0 for x, 1 for xy, 2 for xyz, 3 for xyzw */
+ int class = util_logbase2(ins->mask);
+
+ /* Use the largest class if there's ambiguity, this
+ * handles partial writes */
+
+ int dest = ins->ssa_args.dest;
+ found_class[dest] = MAX2(found_class[dest], class);
+ }
+
+ /* Next, we'll determine semantic class. We default to zero (work).
+ * But, if we're used with a special operation, that will force us to a
+ * particular class. Each node must be assigned to exactly one class; a
+ * prepass before RA should have lowered what-would-have-been
+ * multiclass nodes into a series of moves to break it up into multiple
+ * nodes (TODO) */
- int class = util_logbase2(ins->mask) + 1;
+ mir_foreach_instr_global(ctx, ins) {
+ /* Check if this operation imposes any classes */
- /* Use the largest class if there's ambiguity, this
- * handles partial writes */
+ if (ins->type == TAG_LOAD_STORE_4) {
+ bool force_vec4_only = OP_IS_VEC4_ONLY(ins->load_store.op);
- int dest = ins->ssa_args.dest;
- found_class[dest] = MAX2(found_class[dest], class);
+ set_class(found_class, ins->ssa_args.src[0], REG_CLASS_LDST);
+ set_class(found_class, ins->ssa_args.src[1], REG_CLASS_LDST);
+ set_class(found_class, ins->ssa_args.src[2], REG_CLASS_LDST);
+
+ if (force_vec4_only) {
+ force_vec4(found_class, ins->ssa_args.dest);
+ force_vec4(found_class, ins->ssa_args.src[0]);
+ force_vec4(found_class, ins->ssa_args.src[1]);
+ force_vec4(found_class, ins->ssa_args.src[2]);
+ }
+ } else if (ins->type == TAG_TEXTURE_4) {
+ set_class(found_class, ins->ssa_args.dest, REG_CLASS_TEXW);
+ set_class(found_class, ins->ssa_args.src[0], REG_CLASS_TEXR);
+ set_class(found_class, ins->ssa_args.src[1], REG_CLASS_TEXR);
+ set_class(found_class, ins->ssa_args.src[2], REG_CLASS_TEXR);
}
}
+ /* Check that the semantics of the class are respected */
+ mir_foreach_instr_global(ctx, ins) {
+ assert(check_write_class(found_class, ins->type, ins->ssa_args.dest));
+ assert(check_read_class(found_class, ins->type, ins->ssa_args.src[0]));
+ assert(check_read_class(found_class, ins->type, ins->ssa_args.src[1]));
+ assert(check_read_class(found_class, ins->type, ins->ssa_args.src[2]));
+ }
+
for (unsigned i = 0; i < ctx->temp_count; ++i) {
unsigned class = found_class[i];
- if (!class) continue;
- ra_set_node_class(g, i, classes[class - 1]);
+ ra_set_node_class(g, i, classes[class]);
}
/* Determine liveness */
mir_foreach_block(ctx, block) {
mir_foreach_instr_in_block(block, ins) {
- if (ins->compact_branch) continue;
-
- /* Dest is < 0 for st_vary instructions, which break
- * the usual SSA conventions. Liveness analysis doesn't
- * make sense on these instructions, so skip them to
- * avoid memory corruption */
-
- if (ins->ssa_args.dest < 0) continue;
-
if (ins->ssa_args.dest < SSA_FIXED_MINIMUM) {
/* If this destination is not yet live, it is
* now since we just wrote it */
int dest = ins->ssa_args.dest;
- if (live_start[dest] == -1)
+ if (dest >= 0 && live_start[dest] == -1)
live_start[dest] = d;
}
* invocations, and if there are none, the source dies
* */
- int sources[2] = {
- ins->ssa_args.src0, ins->ssa_args.src1
- };
-
- for (int src = 0; src < 2; ++src) {
- int s = sources[src];
-
- if (ins->ssa_args.inline_constant && src == 1)
- continue;
+ for (int src = 0; src < ARRAY_SIZE(ins->ssa_args.src); ++src) {
+ int s = ins->ssa_args.src[src];
if (s < 0) continue;
if (!ra_allocate(g)) {
*spilled = true;
- return NULL;
+ } else {
+ *spilled = false;
}
+ /* Whether we were successful or not, report the graph so we can
+ * compute spill nodes */
+
return g;
}
switch (ins->type) {
case TAG_ALU_4: {
- int adjusted_src = args.inline_constant ? -1 : args.src1;
- struct phys_reg src1 = index_to_reg(ctx, g, args.src0);
- struct phys_reg src2 = index_to_reg(ctx, g, adjusted_src);
+ struct phys_reg src1 = index_to_reg(ctx, g, args.src[0]);
+ struct phys_reg src2 = index_to_reg(ctx, g, args.src[1]);
struct phys_reg dest = index_to_reg(ctx, g, args.dest);
unsigned uncomposed_mask = ins->mask;
}
case TAG_LOAD_STORE_4: {
- if (OP_IS_STORE_R26(ins->load_store.op)) {
- /* TODO: use ssa_args for st_vary */
- ins->load_store.reg = 0;
- } else {
- /* Which physical register we read off depends on
- * whether we are loading or storing -- think about the
- * logical dataflow */
+ /* Which physical register we read off depends on
+ * whether we are loading or storing -- think about the
+ * logical dataflow */
+
+ bool encodes_src =
+ OP_IS_STORE(ins->load_store.op) &&
+ ins->load_store.op != midgard_op_st_cubemap_coords;
+
+ if (encodes_src) {
+ struct phys_reg src = index_to_reg(ctx, g, args.src[0]);
+ assert(src.reg == 26 || src.reg == 27);
+
+ ins->load_store.reg = src.reg - 26;
+
+ unsigned shift = __builtin_ctz(src.mask);
+ unsigned adjusted_mask = src.mask >> shift;
+ assert(((adjusted_mask + 1) & adjusted_mask) == 0);
+
+ unsigned new_swizzle = 0;
+ for (unsigned q = 0; q < 4; ++q) {
+ unsigned c = (ins->load_store.swizzle >> (2*q)) & 3;
+ new_swizzle |= (c + shift) << (2*q);
+ }
+
+ ins->load_store.swizzle = compose_swizzle(
+ new_swizzle, src.mask,
+ default_phys_reg(0), src);
+ } else {
+ unsigned r = encodes_src ?
+ args.src[0] : args.dest;
- unsigned r = OP_IS_STORE(ins->load_store.op) ?
- args.src0 : args.dest;
struct phys_reg src = index_to_reg(ctx, g, r);
ins->load_store.reg = src.reg;
ins->mask, src);
}
+ /* We also follow up by actual arguments */
+
+ int src2 =
+ encodes_src ? args.src[1] : args.src[0];
+
+ int src3 =
+ encodes_src ? args.src[2] : args.src[1];
+
+ if (src2 >= 0) {
+ struct phys_reg src = index_to_reg(ctx, g, src2);
+ unsigned component = __builtin_ctz(src.mask);
+ ins->load_store.arg_1 |= midgard_ldst_reg(src.reg, component);
+ }
+
+ if (src3 >= 0) {
+ struct phys_reg src = index_to_reg(ctx, g, src3);
+ unsigned component = __builtin_ctz(src.mask);
+ ins->load_store.arg_2 |= midgard_ldst_reg(src.reg, component);
+ }
+
+ break;
+ }
+
+ case TAG_TEXTURE_4: {
+ /* Grab RA results */
+ struct phys_reg dest = index_to_reg(ctx, g, args.dest);
+ struct phys_reg coord = index_to_reg(ctx, g, args.src[0]);
+ struct phys_reg lod = index_to_reg(ctx, g, args.src[1]);
+
+ assert(dest.reg == 28 || dest.reg == 29);
+ assert(coord.reg == 28 || coord.reg == 29);
+
+ /* First, install the texture coordinate */
+ ins->texture.in_reg_full = 1;
+ ins->texture.in_reg_upper = 0;
+ ins->texture.in_reg_select = coord.reg - 28;
+ ins->texture.in_reg_swizzle =
+ compose_swizzle(ins->texture.in_reg_swizzle, 0xF, coord, dest);
+
+ /* Next, install the destination */
+ ins->texture.out_full = 1;
+ ins->texture.out_upper = 0;
+ ins->texture.out_reg_select = dest.reg - 28;
+ ins->texture.swizzle =
+ compose_swizzle(ins->texture.swizzle, dest.mask, dest, dest);
+ ins->mask =
+ compose_writemask(ins->mask, dest);
+
+ /* If there is a register LOD/bias, use it */
+ if (args.src[1] > -1) {
+ midgard_tex_register_select sel = {
+ .select = lod.reg,
+ .full = 1,
+ .component = lod.swizzle & 3,
+ };
+
+ uint8_t packed;
+ memcpy(&packed, &sel, sizeof(packed));
+ ins->texture.bias = packed;
+ }
+
break;
}
{
mir_foreach_block(ctx, block) {
mir_foreach_instr_in_block(block, ins) {
- if (ins->compact_branch) continue;
install_registers_instr(ctx, g, ins);
}
}