* subsequent partial writes to r0.xy. So the 'add r0.z, ...' is the
* defining instruction, as it is the first to partially write r0.xyz.
*
- * Note i965 has a similar scenario, which they solve with a virtual
- * LOAD_PAYLOAD instruction which gets turned into multiple MOV's after
- * register assignment. But for us that is horrible from a scheduling
- * standpoint. Instead what we do is use idea of 'definer' instruction.
- * Ie. the first instruction (lowest ip) to write to the variable is the
- * one we consider from use/def perspective when building interference
- * graph. (Other instructions which write other variable components
- * just define the variable some more.)
+ * To address the fragmentation that this can potentially cause, a
+ * two pass register allocation is used. After the first pass the
+ * assignment of scalars is discarded, but the assignment of vecN (for
+ * N > 1) is used to pre-color in the second pass, which considers
+ * only scalars.
*
* Arrays of arbitrary size are handled via pre-coloring a consecutive
* sequence of registers. Additional scalar (single component) reg
for (unsigned i = 0; i < half_class_count; i++) {
/* NOTE there are fewer half class sizes, but they match the
* first N full class sizes.. but assert in case that ever
- * accidentially changes:
+ * accidentally changes:
*/
debug_assert(class_sizes[i] == half_class_sizes[i]);
for (unsigned j = 0; j < CLASS_REGS(i) / 2; j++) {
unsigned hreg0 = set->gpr_to_ra_reg[i + HALF_OFFSET][(j * 2) + 0];
unsigned hreg1 = set->gpr_to_ra_reg[i + HALF_OFFSET][(j * 2) + 1];
- ra_add_transitive_reg_conflict(set->regs, freg, hreg0);
- ra_add_transitive_reg_conflict(set->regs, freg, hreg1);
+ ra_add_transitive_reg_pair_conflict(set->regs, freg, hreg0, hreg1);
}
}
struct ir3_ra_reg_set *set;
struct ra_graph *g;
+
+ /* Are we in the scalar assignment pass? In this pass, all larger-
+ * than-vec1 vales have already been assigned and pre-colored, so
+ * we only consider scalar values.
+ */
+ bool scalar_pass;
+
unsigned alloc_count;
/* one per class, plus one slot for arrays: */
unsigned class_alloc_count[total_class_count + 1];
static bool
writes_gpr(struct ir3_instruction *instr)
{
- if (is_store(instr))
- return false;
- if (instr->regs_count == 0)
+ if (dest_regs(instr) == 0)
return false;
/* is dest a normal temp register: */
struct ir3_register *reg = instr->regs[0];
- if (reg->flags & (IR3_REG_CONST | IR3_REG_IMMED))
- return false;
+ debug_assert(!(reg->flags & (IR3_REG_CONST | IR3_REG_IMMED)));
if ((reg->num == regid(REG_A0, 0)) ||
(reg->num == regid(REG_P0, 0)))
return false;
struct ir3_ra_instr_data *id = &ctx->instrd[instr->ip];
struct ir3_instruction *d = NULL;
+ if (ctx->scalar_pass) {
+ id->defn = instr;
+ id->off = 0;
+ id->sz = 1; /* considering things as N scalar regs now */
+ }
+
if (id->defn) {
*sz = id->sz;
*off = id->off;
return id->defn;
}
- if (instr->opc == OPC_META_FI) {
+ if (instr->opc == OPC_META_COLLECT) {
/* What about the case where collect is subset of array, we
* need to find the distance between where actual array starts
- * and fanin.. that probably doesn't happen currently.
+ * and collect.. that probably doesn't happen currently.
*/
struct ir3_register *src;
int dsz, doff;
/* note: don't use foreach_ssa_src as this gets called once
* while assigning regs (which clears SSA flag)
*/
- foreach_src_n(src, n, instr) {
+ foreach_src_n (src, n, instr) {
struct ir3_instruction *dd;
if (!src->instr)
continue;
/* by definition, the entire sequence forms one linked list
* of single scalar register nodes (even if some of them may
- * be fanouts from a texture sample (for example) instr. We
+ * be splits from a texture sample (for example) instr. We
* just need to walk the list finding the first element of
* the group defined (lowest ip)
*/
} else {
/* second case is looking directly at the instruction which
* produces multiple values (eg, texture sample), rather
- * than the fanout nodes that point back to that instruction.
+ * than the split nodes that point back to that instruction.
* This isn't quite right, because it may be part of a larger
* group, such as:
*
d = instr;
}
- if (d->opc == OPC_META_FO) {
+ if (d->opc == OPC_META_SPLIT) {
struct ir3_instruction *dd;
int dsz, doff;
*sz = MAX2(*sz, dsz);
- if (instr->opc == OPC_META_FO)
- *off = MAX2(*off, instr->fo.off);
+ if (instr->opc == OPC_META_SPLIT)
+ *off = MAX2(*off, instr->split.off);
d = dd;
}
- debug_assert(d->opc != OPC_META_FO);
+ debug_assert(d->opc != OPC_META_SPLIT);
id->defn = d;
id->sz = *sz;
static void
ra_block_find_definers(struct ir3_ra_ctx *ctx, struct ir3_block *block)
{
- list_for_each_entry (struct ir3_instruction, instr, &block->instr_list, node) {
+ foreach_instr (instr, &block->instr_list) {
struct ir3_ra_instr_data *id = &ctx->instrd[instr->ip];
if (instr->regs_count == 0)
continue;
static void
ra_block_name_instructions(struct ir3_ra_ctx *ctx, struct ir3_block *block)
{
- list_for_each_entry (struct ir3_instruction, instr, &block->instr_list, node) {
+ foreach_instr (instr, &block->instr_list) {
struct ir3_ra_instr_data *id = &ctx->instrd[instr->ip];
#ifdef DEBUG
if (id->defn != instr)
continue;
+ /* In scalar pass, collect/split don't get their own names,
+ * but instead inherit them from their src(s):
+ *
+ * Possibly we don't need this because of scalar_name(), but
+ * it does make the ir3_print() dumps easier to read.
+ */
+ if (ctx->scalar_pass) {
+ if (instr->opc == OPC_META_SPLIT) {
+ instr->name = instr->regs[1]->instr->name + instr->split.off;
+ continue;
+ }
+
+ if (instr->opc == OPC_META_COLLECT) {
+ instr->name = instr->regs[1]->instr->name;
+ continue;
+ }
+ }
+
/* arrays which don't fit in one of the pre-defined class
* sizes are pre-colored:
*/
if ((id->cls >= 0) && (id->cls < total_class_count)) {
- instr->name = ctx->class_alloc_count[id->cls]++;
- ctx->alloc_count++;
+ /* in the scalar pass, we generate a name for each
+ * scalar component, instr->name is the name of the
+ * first component.
+ */
+ unsigned n = ctx->scalar_pass ? dest_regs(instr) : 1;
+ instr->name = ctx->class_alloc_count[id->cls];
+ ctx->class_alloc_count[id->cls] += n;
+ ctx->alloc_count += n;
}
}
}
ctx->instrd = rzalloc_array(NULL, struct ir3_ra_instr_data, n);
- list_for_each_entry (struct ir3_block, block, &ctx->ir->block_list, node) {
+ foreach_block (block, &ctx->ir->block_list) {
ra_block_find_definers(ctx, block);
}
- list_for_each_entry (struct ir3_block, block, &ctx->ir->block_list, node) {
+ foreach_block (block, &ctx->ir->block_list) {
ra_block_name_instructions(ctx, block);
}
/* and vreg names for array elements: */
base = ctx->class_base[total_class_count];
- list_for_each_entry (struct ir3_array, arr, &ctx->ir->array_list, node) {
+ foreach_array (arr, &ctx->ir->array_list) {
arr->base = base;
ctx->class_alloc_count[total_class_count] += arr->length;
base += arr->length;
return __ra_name(ctx, id->cls, id->defn);
}
+/* Get the scalar name of the n'th component of an instruction dst: */
+static int
+scalar_name(struct ir3_ra_ctx *ctx, struct ir3_instruction *instr, unsigned n)
+{
+ if (ctx->scalar_pass) {
+ if (instr->opc == OPC_META_SPLIT) {
+ debug_assert(n == 0); /* split results in a scalar */
+ struct ir3_instruction *src = instr->regs[1]->instr;
+ return scalar_name(ctx, src, instr->split.off);
+ } else if (instr->opc == OPC_META_COLLECT) {
+ debug_assert(n < (instr->regs_count + 1));
+ struct ir3_instruction *src = instr->regs[n + 1]->instr;
+ return scalar_name(ctx, src, 0);
+ }
+ } else {
+ debug_assert(n == 0);
+ }
+
+ return ra_name(ctx, &ctx->instrd[instr->ip]) + n;
+}
+
static void
ra_destroy(struct ir3_ra_ctx *ctx)
{
ralloc_free(ctx->g);
}
+static void
+__def(struct ir3_ra_ctx *ctx, struct ir3_ra_block_data *bd, unsigned name,
+ struct ir3_instruction *instr)
+{
+ debug_assert(name < ctx->alloc_count);
+ /* defined on first write: */
+ if (!ctx->def[name])
+ ctx->def[name] = instr->ip;
+ ctx->use[name] = MAX2(ctx->use[name], instr->ip);
+ BITSET_SET(bd->def, name);
+}
+
+static void
+__use(struct ir3_ra_ctx *ctx, struct ir3_ra_block_data *bd, unsigned name,
+ struct ir3_instruction *instr)
+{
+ debug_assert(name < ctx->alloc_count);
+ ctx->use[name] = MAX2(ctx->use[name], instr->ip);
+ if (!BITSET_TEST(bd->def, name))
+ BITSET_SET(bd->use, name);
+}
+
static void
ra_block_compute_live_ranges(struct ir3_ra_ctx *ctx, struct ir3_block *block)
{
struct ir3_ra_block_data *bd;
unsigned bitset_words = BITSET_WORDS(ctx->alloc_count);
-#define def(name, instr) \
- do { \
- /* defined on first write: */ \
- if (!ctx->def[name]) \
- ctx->def[name] = instr->ip; \
- ctx->use[name] = instr->ip; \
- BITSET_SET(bd->def, name); \
- } while(0);
-
-#define use(name, instr) \
- do { \
- ctx->use[name] = MAX2(ctx->use[name], instr->ip); \
- if (!BITSET_TEST(bd->def, name)) \
- BITSET_SET(bd->use, name); \
- } while(0);
+#define def(name, instr) __def(ctx, bd, name, instr)
+#define use(name, instr) __use(ctx, bd, name, instr)
bd = rzalloc(ctx->g, struct ir3_ra_block_data);
block->data = bd;
- list_for_each_entry (struct ir3_instruction, instr, &block->instr_list, node) {
+ struct ir3_instruction *first_non_input = NULL;
+ foreach_instr (instr, &block->instr_list) {
+ if (instr->opc != OPC_META_INPUT) {
+ first_non_input = instr;
+ break;
+ }
+ }
+
+ foreach_instr (instr, &block->instr_list) {
struct ir3_instruction *src;
struct ir3_register *reg;
- /* There are a couple special cases to deal with here:
- *
- * fanout: used to split values from a higher class to a lower
- * class, for example split the results of a texture fetch
- * into individual scalar values; We skip over these from
- * a 'def' perspective, and for a 'use' we walk the chain
- * up to the defining instruction.
- *
- * fanin: used to collect values from lower class and assemble
- * them together into a higher class, for example arguments
- * to texture sample instructions; We consider these to be
- * defined at the earliest fanin source.
- *
- * Most of this is handled in the get_definer() helper.
- *
- * In either case, we trace the instruction back to the original
- * definer and consider that as the def/use ip.
- */
-
if (writes_gpr(instr)) {
struct ir3_ra_instr_data *id = &ctx->instrd[instr->ip];
struct ir3_register *dst = instr->regs[0];
unsigned name = arr->base + dst->array.offset;
def(name, instr);
}
-
} else if (id->defn == instr) {
- unsigned name = ra_name(ctx, id);
-
- /* since we are in SSA at this point: */
- debug_assert(!BITSET_TEST(bd->use, name));
+ /* in scalar pass, we aren't considering virtual register
+ * classes, ie. if an instruction writes a vec2, then it
+ * defines two different scalar register names.
+ */
+ unsigned n = ctx->scalar_pass ? dest_regs(instr) : 1;
+ for (unsigned i = 0; i < n; i++) {
+ unsigned name = scalar_name(ctx, instr, i);
+
+ /* tex instructions actually have a wrmask, and
+ * don't touch masked out components. We can't do
+ * anything useful about that in the first pass,
+ * but in the scalar pass we can realize these
+ * registers are available:
+ */
+ if (ctx->scalar_pass && is_tex_or_prefetch(instr) &&
+ !(instr->regs[0]->wrmask & (1 << i)))
+ continue;
- def(name, id->defn);
+ def(name, instr);
- if (is_high(id->defn)) {
- ra_set_node_class(ctx->g, name,
- ctx->set->high_classes[id->cls - HIGH_OFFSET]);
- } else if (is_half(id->defn)) {
- ra_set_node_class(ctx->g, name,
- ctx->set->half_classes[id->cls - HALF_OFFSET]);
- } else {
- ra_set_node_class(ctx->g, name,
- ctx->set->classes[id->cls]);
+ if ((instr->opc == OPC_META_INPUT) && first_non_input)
+ use(name, first_non_input);
+
+ if (is_high(instr)) {
+ ra_set_node_class(ctx->g, name,
+ ctx->set->high_classes[id->cls - HIGH_OFFSET]);
+ } else if (is_half(instr)) {
+ ra_set_node_class(ctx->g, name,
+ ctx->set->half_classes[id->cls - HALF_OFFSET]);
+ } else {
+ ra_set_node_class(ctx->g, name,
+ ctx->set->classes[id->cls]);
+ }
}
}
}
arr->start_ip = MIN2(arr->start_ip, instr->ip);
arr->end_ip = MAX2(arr->end_ip, instr->ip);
- /* indirect read is treated like a read fromall array
+ /* indirect read is treated like a read from all array
* elements, since we don't know which one is actually
* read:
*/
for (i = 0; i < arr->length; i++) {
unsigned name = arr->base + i;
use(name, instr);
+ BITSET_SET(bd->use, name);
}
} else {
unsigned name = arr->base + reg->array.offset;
BITSET_SET(bd->use, name);
debug_assert(reg->array.offset < arr->length);
}
+ } else if (ctx->scalar_pass) {
+ struct ir3_instruction *src = reg->instr;
+ /* skip things that aren't SSA: */
+ unsigned n = src ? dest_regs(src) : 0;
+
+ /* in scalar pass, we aren't considering virtual register
+ * classes, ie. if an instruction writes a vec2, then it
+ * defines two different scalar register names.
+ *
+ * We need to traverse up thru collect/split to find the
+ * actual non-meta instruction names for each of the
+ * components:
+ */
+ for (unsigned i = 0; i < n; i++) {
+ /* Need to filter out a couple special cases, ie.
+ * writes to a0.x or p0.x:
+ */
+ if (!writes_gpr(src))
+ continue;
+
+ /* split takes a src w/ wrmask potentially greater
+ * than 0x1, but it really only cares about a single
+ * component. This shows up in splits coming out of
+ * a tex instruction w/ wrmask=.z, for example.
+ */
+ if (ctx->scalar_pass && (instr->opc == OPC_META_SPLIT) &&
+ !(i == instr->split.off))
+ continue;
+
+ use(scalar_name(ctx, src, i), instr);
+ }
} else if ((src = ssa(reg)) && writes_gpr(src)) {
unsigned name = ra_name(ctx, &ctx->instrd[src->ip]);
use(name, instr);
unsigned bitset_words = BITSET_WORDS(ctx->alloc_count);
bool progress = false;
- list_for_each_entry (struct ir3_block, block, &ctx->ir->block_list, node) {
+ foreach_block (block, &ctx->ir->block_list) {
struct ir3_ra_block_data *bd = block->data;
/* update livein: */
struct ir3 *ir = ctx->ir;
/* initialize array live ranges: */
- list_for_each_entry (struct ir3_array, arr, &ir->array_list, node) {
+ foreach_array (arr, &ir->array_list) {
arr->start_ip = ~0;
arr->end_ip = 0;
}
* block's def/use bitmasks (used below to calculate per-block
* livein/liveout):
*/
- list_for_each_entry (struct ir3_block, block, &ir->block_list, node) {
+ foreach_block (block, &ir->block_list) {
ra_block_compute_live_ranges(ctx, block);
}
if (ir3_shader_debug & IR3_DBG_OPTMSGS) {
debug_printf("AFTER LIVEIN/OUT:\n");
- ir3_print(ir);
- list_for_each_entry (struct ir3_block, block, &ir->block_list, node) {
+ foreach_block (block, &ir->block_list) {
struct ir3_ra_block_data *bd = block->data;
debug_printf("block%u:\n", block_id(block));
print_bitset(" def", bd->def, ctx->alloc_count);
print_bitset(" l/i", bd->livein, ctx->alloc_count);
print_bitset(" l/o", bd->liveout, ctx->alloc_count);
}
- list_for_each_entry (struct ir3_array, arr, &ir->array_list, node) {
+ foreach_array (arr, &ir->array_list) {
debug_printf("array%u:\n", arr->id);
debug_printf(" length: %u\n", arr->length);
debug_printf(" start_ip: %u\n", arr->start_ip);
debug_printf(" end_ip: %u\n", arr->end_ip);
}
+ debug_printf("INSTRUCTION VREG NAMES:\n");
+ foreach_block (block, &ctx->ir->block_list) {
+ foreach_instr (instr, &block->instr_list) {
+ if (!ctx->instrd[instr->ip].defn)
+ continue;
+ debug_printf("%04u: ", scalar_name(ctx, instr, 0));
+ ir3_print_instr(instr);
+ }
+ }
+ debug_printf("ARRAY VREG NAMES:\n");
+ foreach_array (arr, &ctx->ir->array_list) {
+ debug_printf("%04u: arr%u\n", arr->base, arr->id);
+ }
}
/* extend start/end ranges based on livein/liveout info from cfg: */
- list_for_each_entry (struct ir3_block, block, &ir->block_list, node) {
+ foreach_block (block, &ir->block_list) {
struct ir3_ra_block_data *bd = block->data;
for (unsigned i = 0; i < ctx->alloc_count; i++) {
}
}
- list_for_each_entry (struct ir3_array, arr, &ctx->ir->array_list, node) {
+ foreach_array (arr, &ctx->ir->array_list) {
for (unsigned i = 0; i < arr->length; i++) {
if (BITSET_TEST(bd->livein, i + arr->base)) {
arr->start_ip = MIN2(arr->start_ip, block->start_ip);
}
/* need to fix things up to keep outputs live: */
- for (unsigned i = 0; i < ir->noutputs; i++) {
- struct ir3_instruction *instr = ir->outputs[i];
- if (!instr)
- continue;
- unsigned name = ra_name(ctx, &ctx->instrd[instr->ip]);
+ struct ir3_instruction *out;
+ foreach_output(out, ir) {
+ unsigned name = ra_name(ctx, &ctx->instrd[out->ip]);
ctx->use[name] = ctx->instr_cnt;
}
case 3:
switch (instr->opc) {
case OPC_MAD_F32:
- instr->opc = OPC_MAD_F16;
+ /* Available for that dest is half and srcs are full.
+ * eg. mad.f32 hr0, r0.x, r0.y, r0.z
+ */
+ if (instr->regs[1]->flags & IR3_REG_HALF)
+ instr->opc = OPC_MAD_F16;
break;
case OPC_SEL_B32:
instr->opc = OPC_SEL_B16;
break;
}
break;
+ case 4:
+ switch (instr->opc) {
+ case OPC_RSQ:
+ instr->opc = OPC_HRSQ;
+ break;
+ case OPC_LOG2:
+ instr->opc = OPC_HLOG2;
+ break;
+ case OPC_EXP2:
+ instr->opc = OPC_HEXP2;
+ break;
+ default:
+ break;
+ }
+ break;
case 5:
instr->cat5.type = half_type(instr->cat5.type);
break;
reg->flags &= ~IR3_REG_ARRAY;
} else if ((id = &ctx->instrd[instr->ip]) && id->defn) {
- unsigned name = ra_name(ctx, id);
+ unsigned first_component = 0;
+
+ /* Special case for tex instructions, which may use the wrmask
+ * to mask off the first component(s). In the scalar pass,
+ * this means the masked off component(s) are not def'd/use'd,
+ * so we get a bogus value when we ask the register_allocate
+ * algo to get the assigned reg for the unused/untouched
+ * component. So we need to consider the first used component:
+ */
+ if (ctx->scalar_pass && is_tex_or_prefetch(id->defn)) {
+ unsigned n = ffs(id->defn->regs[0]->wrmask);
+ debug_assert(n > 0);
+ first_component = n - 1;
+ }
+
+ unsigned name = scalar_name(ctx, id->defn, first_component);
unsigned r = ra_get_node_reg(ctx->g, name);
unsigned num = ctx->set->ra_reg_to_gpr[r] + id->off;
debug_assert(!(reg->flags & IR3_REG_RELATIV));
+ debug_assert(num >= first_component);
+
if (is_high(id->defn))
num += FIRST_HIGH_REG;
- reg->num = num;
+ reg->num = num - first_component;
+
reg->flags &= ~IR3_REG_SSA;
if (is_half(id->defn))
}
}
+/* helper to determine which regs to assign in which pass: */
+static bool
+should_assign(struct ir3_ra_ctx *ctx, struct ir3_instruction *instr)
+{
+ if ((instr->opc == OPC_META_SPLIT) ||
+ (instr->opc == OPC_META_COLLECT))
+ return !ctx->scalar_pass;
+ return ctx->scalar_pass;
+}
+
static void
ra_block_alloc(struct ir3_ra_ctx *ctx, struct ir3_block *block)
{
- list_for_each_entry (struct ir3_instruction, instr, &block->instr_list, node) {
+ foreach_instr (instr, &block->instr_list) {
struct ir3_register *reg;
if (writes_gpr(instr)) {
- reg_assign(ctx, instr->regs[0], instr);
- if (instr->regs[0]->flags & IR3_REG_HALF)
- fixup_half_instr_dst(instr);
+ if (should_assign(ctx, instr)) {
+ reg_assign(ctx, instr->regs[0], instr);
+ if (instr->regs[0]->flags & IR3_REG_HALF)
+ fixup_half_instr_dst(instr);
+ }
}
foreach_src_n(reg, n, instr) {
struct ir3_instruction *src = reg->instr;
+
+ if (src && !should_assign(ctx, src) && !should_assign(ctx, instr))
+ continue;
+
+ if (src && should_assign(ctx, instr))
+ reg_assign(ctx, src->regs[0], src);
+
/* Note: reg->instr could be null for IR3_REG_ARRAY */
if (src || (reg->flags & IR3_REG_ARRAY))
reg_assign(ctx, instr->regs[n+1], src);
+
if (instr->regs[n+1]->flags & IR3_REG_HALF)
fixup_half_instr_src(instr);
}
}
+
+ /* We need to pre-color outputs for the scalar pass in
+ * ra_precolor_assigned(), so we need to actually assign
+ * them in the first pass:
+ */
+ if (!ctx->scalar_pass) {
+ struct ir3_instruction *in, *out;
+
+ foreach_input (in, ctx->ir) {
+ reg_assign(ctx, in->regs[0], in);
+ }
+ foreach_output (out, ctx->ir) {
+ reg_assign(ctx, out->regs[0], out);
+ }
+ }
}
/* handle pre-colored registers. This includes "arrays" (which could be of
for (unsigned i = 0; i < nprecolor; i++) {
if (precolor[i] && !(precolor[i]->flags & IR3_INSTR_UNUSED)) {
struct ir3_instruction *instr = precolor[i];
+
struct ir3_ra_instr_data *id = &ctx->instrd[instr->ip];
debug_assert(!(instr->regs[0]->flags & (IR3_REG_HALF | IR3_REG_HIGH)));
if (id->off > 0)
continue;
+ if (ctx->scalar_pass && !should_assign(ctx, instr))
+ continue;
+
/* 'base' is in scalar (class 0) but we need to map that
* the conflicting register of the appropriate class (ie.
* input could be vec2/vec3/etc)
}
/* pre-assign array elements:
+ *
+ * TODO this is going to need some work for half-precision.. possibly
+ * this is easier on a6xx, where we can just divide array size by two?
+ * But on a5xx and earlier it will need to track two bases.
*/
- list_for_each_entry (struct ir3_array, arr, &ctx->ir->array_list, node) {
+ foreach_array (arr, &ctx->ir->array_list) {
unsigned base = 0;
if (arr->end_ip == 0)
* been assigned:
*/
retry:
- list_for_each_entry (struct ir3_array, arr2, &ctx->ir->array_list, node) {
+ foreach_array (arr2, &ctx->ir->array_list) {
if (arr2 == arr)
break;
if (arr2->end_ip == 0)
for (unsigned i = 0; i < nprecolor; i++) {
struct ir3_instruction *instr = precolor[i];
- if (!instr)
+ if (!instr || (instr->flags & IR3_INSTR_UNUSED))
continue;
struct ir3_ra_instr_data *id = &ctx->instrd[instr->ip];
ra_set_node_reg(ctx->g, name, reg);
}
}
+
+ if (ir3_shader_debug & IR3_DBG_OPTMSGS) {
+ foreach_array (arr, &ctx->ir->array_list) {
+ unsigned first = arr->reg;
+ unsigned last = arr->reg + arr->length - 1;
+ debug_printf("arr[%d] at r%d.%c->r%d.%c\n", arr->id,
+ (first >> 2), "xyzw"[first & 0x3],
+ (last >> 2), "xyzw"[last & 0x3]);
+ }
+ }
+}
+
+static void
+precolor(struct ir3_ra_ctx *ctx, struct ir3_instruction *instr)
+{
+ struct ir3_ra_instr_data *id = &ctx->instrd[instr->ip];
+ unsigned n = dest_regs(instr);
+ for (unsigned i = 0; i < n; i++) {
+ /* tex instructions actually have a wrmask, and
+ * don't touch masked out components. So we
+ * shouldn't precolor them::
+ */
+ if (is_tex_or_prefetch(instr) &&
+ !(instr->regs[0]->wrmask & (1 << i)))
+ continue;
+
+ unsigned name = scalar_name(ctx, instr, i);
+ unsigned regid = instr->regs[0]->num + i;
+
+ if (instr->regs[0]->flags & IR3_REG_HIGH)
+ regid -= FIRST_HIGH_REG;
+
+ unsigned vreg = ctx->set->gpr_to_ra_reg[id->cls][regid];
+ ra_set_node_reg(ctx->g, name, vreg);
+ }
+}
+
+/* pre-color non-scalar registers based on the registers assigned in previous
+ * pass. Do this by looking actually at the fanout instructions.
+ */
+static void
+ra_precolor_assigned(struct ir3_ra_ctx *ctx)
+{
+ debug_assert(ctx->scalar_pass);
+
+ foreach_block (block, &ctx->ir->block_list) {
+ foreach_instr (instr, &block->instr_list) {
+
+ if ((instr->opc != OPC_META_SPLIT) &&
+ (instr->opc != OPC_META_COLLECT))
+ continue;
+
+ precolor(ctx, instr);
+
+ struct ir3_register *src;
+ foreach_src (src, instr) {
+ if (!src->instr)
+ continue;
+ precolor(ctx, src->instr);
+ }
+ }
+ }
}
static int
if (!ra_allocate(ctx->g))
return -1;
- list_for_each_entry (struct ir3_block, block, &ctx->ir->block_list, node) {
+ foreach_block (block, &ctx->ir->block_list) {
ra_block_alloc(ctx, block);
}
return 0;
}
-int ir3_ra(struct ir3_shader_variant *v, struct ir3_instruction **precolor, unsigned nprecolor)
+/* if we end up with split/collect instructions with non-matching src
+ * and dest regs, that means something has gone wrong. Which makes it
+ * a pretty good sanity check.
+ */
+static void
+ra_sanity_check(struct ir3 *ir)
+{
+ foreach_block (block, &ir->block_list) {
+ foreach_instr (instr, &block->instr_list) {
+ if (instr->opc == OPC_META_SPLIT) {
+ struct ir3_register *dst = instr->regs[0];
+ struct ir3_register *src = instr->regs[1];
+ debug_assert(dst->num == (src->num + instr->split.off));
+ } else if (instr->opc == OPC_META_COLLECT) {
+ struct ir3_register *dst = instr->regs[0];
+ struct ir3_register *src;
+
+ foreach_src_n (src, n, instr) {
+ debug_assert(dst->num == (src->num - n));
+ }
+ }
+ }
+ }
+}
+
+static int
+ir3_ra_pass(struct ir3_shader_variant *v, struct ir3_instruction **precolor,
+ unsigned nprecolor, bool scalar_pass)
{
struct ir3_ra_ctx ctx = {
.v = v,
.ir = v->ir,
.set = v->ir->compiler->set,
+ .scalar_pass = scalar_pass,
};
int ret;
ra_init(&ctx);
ra_add_interference(&ctx);
ra_precolor(&ctx, precolor, nprecolor);
+ if (scalar_pass)
+ ra_precolor_assigned(&ctx);
ret = ra_alloc(&ctx);
ra_destroy(&ctx);
return ret;
}
+
+int
+ir3_ra(struct ir3_shader_variant *v, struct ir3_instruction **precolor,
+ unsigned nprecolor)
+{
+ int ret;
+
+ /* First pass, assign the vecN (non-scalar) registers: */
+ ret = ir3_ra_pass(v, precolor, nprecolor, false);
+ if (ret)
+ return ret;
+
+ if (ir3_shader_debug & IR3_DBG_OPTMSGS) {
+ printf("AFTER RA (1st pass):\n");
+ ir3_print(v->ir);
+ }
+
+ /* Second pass, assign the scalar registers: */
+ ret = ir3_ra_pass(v, precolor, nprecolor, true);
+ if (ret)
+ return ret;
+
+ if (ir3_shader_debug & IR3_DBG_OPTMSGS) {
+ printf("AFTER RA (2nd pass):\n");
+ ir3_print(v->ir);
+ }
+
+#ifdef DEBUG
+# define SANITY_CHECK DEBUG
+#else
+# define SANITY_CHECK 0
+#endif
+ if (SANITY_CHECK)
+ ra_sanity_check(v->ir);
+
+ return ret;
+}