}
cat0->repeat = instr->repeat;
cat0->ss = !!(instr->flags & IR3_INSTR_SS);
- cat0->inv = instr->cat0.inv;
- cat0->comp = instr->cat0.comp;
+ cat0->inv0 = instr->cat0.inv;
+ cat0->comp0 = instr->cat0.comp;
cat0->opc = instr->opc;
+ cat0->opc_hi = instr->opc >= 16;
cat0->jmp_tgt = !!(instr->flags & IR3_INSTR_JP);
cat0->sync = !!(instr->flags & IR3_INSTR_SY);
cat0->opc_cat = 0;
- switch (instr->opc) {
- case OPC_IF:
- case OPC_ELSE:
- case OPC_ENDIF:
- cat0->dummy4 = 16;
- break;
- default:
- break;
- }
-
return 0;
}
cat5->src2 = reg(src2, info, instr->repeat, IR3_REG_HALF);
}
+ if (instr->flags & IR3_INSTR_B) {
+ cat5->s2en_bindless.base_hi = instr->cat5.tex_base >> 1;
+ cat5->base_lo = instr->cat5.tex_base & 1;
+ }
+
if (instr->flags & IR3_INSTR_S2EN) {
struct ir3_register *samp_tex = instr->regs[1];
iassert(samp_tex->flags & IR3_REG_HALF);
- cat5->s2en_bindless.src3 = reg(samp_tex, info, instr->repeat, IR3_REG_HALF);
- /* TODO: This should probably be CAT5_UNIFORM, at least on a6xx, as
- * this is what the blob does and it is presumably faster, but first
- * we should confirm it is actually nonuniform and figure out when the
- * whole descriptor mode mechanism was introduced.
- */
- cat5->s2en_bindless.desc_mode = CAT5_NONUNIFORM;
+ cat5->s2en_bindless.src3 = reg(samp_tex, info, instr->repeat,
+ (instr->flags & IR3_INSTR_B) ? 0 : IR3_REG_HALF);
+ if (instr->flags & IR3_INSTR_B) {
+ if (instr->flags & IR3_INSTR_A1EN) {
+ cat5->s2en_bindless.desc_mode = CAT5_BINDLESS_A1_UNIFORM;
+ } else {
+ cat5->s2en_bindless.desc_mode = CAT5_BINDLESS_UNIFORM;
+ }
+ } else {
+ /* TODO: This should probably be CAT5_UNIFORM, at least on a6xx,
+ * as this is what the blob does and it is presumably faster, but
+ * first we should confirm it is actually nonuniform and figure
+ * out when the whole descriptor mode mechanism was introduced.
+ */
+ cat5->s2en_bindless.desc_mode = CAT5_NONUNIFORM;
+ }
iassert(!(instr->cat5.samp | instr->cat5.tex));
+ } else if (instr->flags & IR3_INSTR_B) {
+ cat5->s2en_bindless.src3 = instr->cat5.samp;
+ if (instr->flags & IR3_INSTR_A1EN) {
+ cat5->s2en_bindless.desc_mode = CAT5_BINDLESS_A1_IMM;
+ } else {
+ cat5->s2en_bindless.desc_mode = CAT5_BINDLESS_IMM;
+ }
} else {
cat5->norm.samp = instr->cat5.samp;
cat5->norm.tex = instr->cat5.tex;
cat5->is_3d = !!(instr->flags & IR3_INSTR_3D);
cat5->is_a = !!(instr->flags & IR3_INSTR_A);
cat5->is_s = !!(instr->flags & IR3_INSTR_S);
- cat5->is_s2en_bindless = !!(instr->flags & IR3_INSTR_S2EN);
+ cat5->is_s2en_bindless = !!(instr->flags & (IR3_INSTR_S2EN | IR3_INSTR_B));
cat5->is_o = !!(instr->flags & IR3_INSTR_O);
cat5->is_p = !!(instr->flags & IR3_INSTR_P);
cat5->opc = instr->opc;
static int emit_cat6_a6xx(struct ir3_instruction *instr, void *ptr,
struct ir3_info *info)
{
- struct ir3_register *src1, *src2;
+ struct ir3_register *src1, *src2, *ssbo;
instr_cat6_a6xx_t *cat6 = ptr;
- bool has_dest = (instr->opc == OPC_LDIB);
-
- /* first reg should be SSBO binding point: */
- iassert(instr->regs[1]->flags & IR3_REG_IMMED);
+ bool has_dest = (instr->opc == OPC_LDIB || instr->opc == OPC_LDC);
+ ssbo = instr->regs[1];
src1 = instr->regs[2];
if (has_dest) {
}
cat6->type = instr->cat6.type;
- cat6->d = instr->cat6.d - 1;
+ cat6->d = instr->cat6.d - (instr->opc == OPC_LDC ? 0 : 1);
cat6->typed = instr->cat6.typed;
cat6->type_size = instr->cat6.iim_val - 1;
cat6->opc = instr->opc;
cat6->src1 = reg(src1, info, instr->repeat, 0);
cat6->src2 = reg(src2, info, instr->repeat, 0);
- cat6->ssbo = instr->regs[1]->iim_val;
+ cat6->ssbo = reg(ssbo, info, instr->repeat, IR3_REG_IMMED);
+
+ if (instr->flags & IR3_INSTR_B) {
+ if (ssbo->flags & IR3_REG_IMMED) {
+ cat6->desc_mode = CAT6_BINDLESS_IMM;
+ } else {
+ cat6->desc_mode = CAT6_BINDLESS_UNIFORM;
+ }
+ cat6->base = instr->cat6.base;
+ } else {
+ if (ssbo->flags & IR3_REG_IMMED)
+ cat6->desc_mode = CAT6_IMM;
+ else
+ cat6->desc_mode = CAT6_UNIFORM;
+ }
switch (instr->opc) {
case OPC_ATOMIC_ADD:
{
uint32_t *ptr, *dwords;
+ memset(info, 0, sizeof(*info));
info->gpu_id = gpu_id;
info->max_reg = -1;
info->max_half_reg = -1;
info->max_const = -1;
- info->instrs_count = 0;
- info->sizedwords = 0;
- info->ss = info->sy = 0;
foreach_block (block, &shader->block_list) {
foreach_instr (instr, &block->instr_list) {
info->nops_count += instr->nop;
if (instr->opc == OPC_NOP)
info->nops_count += 1 + instr->repeat;
+ if (instr->opc == OPC_MOV) {
+ if (instr->cat1.src_type == instr->cat1.dst_type) {
+ info->mov_count += 1 + instr->repeat;
+ } else {
+ info->cov_count += 1 + instr->repeat;
+ }
+ }
dwords += 2;
if (instr->flags & IR3_INSTR_SS) {
block->shader = shader;
list_inithead(&block->node);
list_inithead(&block->instr_list);
+ block->predecessors = _mesa_pointer_set_create(block);
return block;
}
debug_assert(instr->block == addr->block);
instr->address = addr;
- array_insert(ir, ir->indirects, instr);
+ debug_assert(reg_num(addr->regs[0]) == REG_A0);
+ unsigned comp = reg_comp(addr->regs[0]);
+ if (comp == 0) {
+ array_insert(ir, ir->a0_users, instr);
+ } else {
+ debug_assert(comp == 1);
+ array_insert(ir, ir->a1_users, instr);
+ }
}
}
}
}
-/* note: this will destroy instr->depth, don't do it until after sched! */
unsigned
ir3_count_instructions(struct ir3 *ir)
{
unsigned cnt = 1;
foreach_block (block, &ir->block_list) {
block->start_ip = cnt;
+ foreach_instr (instr, &block->instr_list) {
+ instr->ip = cnt++;
+ }
block->end_ip = cnt;
+ }
+ return cnt;
+}
+
+/* When counting instructions for RA, we insert extra fake instructions at the
+ * beginning of each block, where values become live, and at the end where
+ * values die. This prevents problems where values live-in at the beginning or
+ * live-out at the end of a block from being treated as if they were
+ * live-in/live-out at the first/last instruction, which would be incorrect.
+ * In ir3_legalize these ip's are assumed to be actual ip's of the final
+ * program, so it would be incorrect to use this everywhere.
+ */
+
+unsigned
+ir3_count_instructions_ra(struct ir3 *ir)
+{
+ unsigned cnt = 1;
+ foreach_block (block, &ir->block_list) {
+ block->start_ip = cnt++;
foreach_instr (instr, &block->instr_list) {
instr->ip = cnt++;
- block->end_ip = instr->ip;
}
+ block->end_ip = cnt++;
}
return cnt;
}
}
void
-ir3_find_ssa_uses(struct ir3 *ir, void *mem_ctx)
+ir3_find_ssa_uses(struct ir3 *ir, void *mem_ctx, bool falsedeps)
{
/* We could do this in a single pass if we can assume instructions
* are always sorted. Which currently might not always be true.
foreach_block (block, &ir->block_list) {
foreach_instr (instr, &block->instr_list) {
- struct ir3_instruction *src;
-
- foreach_ssa_src (src, instr) {
+ foreach_ssa_src_n (src, n, instr) {
+ if (__is_false_dep(instr, n) && !falsedeps)
+ continue;
if (!src->uses)
src->uses = _mesa_pointer_set_create(mem_ctx);
_mesa_set_add(src->uses, instr);
}
}
}
+
+/**
+ * Set the destination type of an instruction, for example if a
+ * conversion is folded in, handling the special cases where the
+ * instruction's dest type or opcode needs to be fixed up.
+ */
+void
+ir3_set_dst_type(struct ir3_instruction *instr, bool half)
+{
+ if (half) {
+ instr->regs[0]->flags |= IR3_REG_HALF;
+ } else {
+ instr->regs[0]->flags &= ~IR3_REG_HALF;
+ }
+
+ switch (opc_cat(instr->opc)) {
+ case 1: /* move instructions */
+ if (half) {
+ instr->cat1.dst_type = half_type(instr->cat1.dst_type);
+ } else {
+ instr->cat1.dst_type = full_type(instr->cat1.dst_type);
+ }
+ break;
+ case 4:
+ if (half) {
+ instr->opc = cat4_half_opc(instr->opc);
+ } else {
+ instr->opc = cat4_full_opc(instr->opc);
+ }
+ break;
+ case 5:
+ if (half) {
+ instr->cat5.type = half_type(instr->cat5.type);
+ } else {
+ instr->cat5.type = full_type(instr->cat5.type);
+ }
+ break;
+ }
+}
+
+/**
+ * One-time fixup for instruction src-types. Other than cov's that
+ * are folded, an instruction's src type does not change.
+ */
+void
+ir3_fixup_src_type(struct ir3_instruction *instr)
+{
+ bool half = !!(instr->regs[1]->flags & IR3_REG_HALF);
+
+ switch (opc_cat(instr->opc)) {
+ case 1: /* move instructions */
+ if (half) {
+ instr->cat1.src_type = half_type(instr->cat1.src_type);
+ } else {
+ instr->cat1.src_type = full_type(instr->cat1.src_type);
+ }
+ break;
+ case 3:
+ if (half) {
+ instr->opc = cat3_half_opc(instr->opc);
+ } else {
+ instr->opc = cat3_full_opc(instr->opc);
+ }
+ break;
+ }
+}
projects / mesa.git / blobdiff