namespace aco {
+#ifndef NDEBUG
+void perfwarn(Program *program, bool cond, const char *msg, Instruction *instr)
+{
+ if (cond) {
+ char *out;
+ size_t outsize;
+ FILE *memf = open_memstream(&out, &outsize);
+
+ fprintf(memf, "%s: ", msg);
+ aco_print_instr(instr, memf);
+ fclose(memf);
+
+ aco_perfwarn(program, out);
+ free(out);
+
+ if (debug_flags & DEBUG_PERFWARN)
+ exit(1);
+ }
+}
+#endif
+
/**
* The optimizer works in 4 phases:
* (1) The first pass collects information for each ssa-def,
return true;
}
-bool parse_base_offset(opt_ctx &ctx, Instruction* instr, unsigned op_index, Temp *base, uint32_t *offset)
+bool parse_base_offset(opt_ctx &ctx, Instruction* instr, unsigned op_index, Temp *base, uint32_t *offset, bool prevent_overflow)
{
Operand op = instr->operands[op_index];
default:
return false;
}
+ if (prevent_overflow && !add_instr->definitions[0].isNUW())
+ return false;
if (add_instr->usesModifiers())
return false;
continue;
uint32_t offset2 = 0;
- if (parse_base_offset(ctx, add_instr, !i, base, &offset2)) {
+ if (parse_base_offset(ctx, add_instr, !i, base, &offset2, prevent_overflow)) {
*offset += offset2;
} else {
*base = add_instr->operands[!i].getTemp();
ASSERTED bool all_const = false;
for (Operand& op : instr->operands)
all_const = all_const && (!op.isTemp() || ctx.info[op.tempId()].is_constant_or_literal(32));
- perfwarn(all_const, "All instruction operands are constant", instr.get());
+ perfwarn(ctx.program, all_const, "All instruction operands are constant", instr.get());
}
for (unsigned i = 0; i < instr->operands.size(); i++)
unsigned bits = get_operand_size(instr, i);
if (info.is_constant(bits) && alu_can_accept_constant(instr->opcode, i)) {
Operand op = get_constant_op(ctx, info, bits);
- perfwarn(instr->opcode == aco_opcode::v_cndmask_b32 && i == 2, "v_cndmask_b32 with a constant selector", instr.get());
+ perfwarn(ctx.program, instr->opcode == aco_opcode::v_cndmask_b32 && i == 2, "v_cndmask_b32 with a constant selector", instr.get());
if (i == 0 || instr->opcode == aco_opcode::v_readlane_b32 || instr->opcode == aco_opcode::v_writelane_b32) {
instr->operands[i] = op;
continue;
while (info.is_temp())
info = ctx.info[info.temp.id()];
+ /* According to AMDGPUDAGToDAGISel::SelectMUBUFScratchOffen(), vaddr
+ * overflow for scratch accesses works only on GFX9+ and saddr overflow
+ * never works. Since swizzling is the only thing that separates
+ * scratch accesses and other accesses and swizzling changing how
+ * addressing works significantly, this probably applies to swizzled
+ * MUBUF accesses. */
+ bool vaddr_prevent_overflow = mubuf->swizzled && ctx.program->chip_class < GFX9;
+ bool saddr_prevent_overflow = mubuf->swizzled;
+
if (mubuf->offen && i == 1 && info.is_constant_or_literal(32) && mubuf->offset + info.val < 4096) {
assert(!mubuf->idxen);
instr->operands[1] = Operand(v1);
instr->operands[2] = Operand((uint32_t) 0);
mubuf->offset += info.val;
continue;
- } else if (mubuf->offen && i == 1 && parse_base_offset(ctx, instr.get(), i, &base, &offset) && base.regClass() == v1 && mubuf->offset + offset < 4096) {
+ } else if (mubuf->offen && i == 1 && parse_base_offset(ctx, instr.get(), i, &base, &offset, vaddr_prevent_overflow) &&
+ base.regClass() == v1 && mubuf->offset + offset < 4096) {
assert(!mubuf->idxen);
instr->operands[1].setTemp(base);
mubuf->offset += offset;
continue;
- } else if (i == 2 && parse_base_offset(ctx, instr.get(), i, &base, &offset) && base.regClass() == s1 && mubuf->offset + offset < 4096) {
+ } else if (i == 2 && parse_base_offset(ctx, instr.get(), i, &base, &offset, saddr_prevent_overflow) &&
+ base.regClass() == s1 && mubuf->offset + offset < 4096) {
instr->operands[i].setTemp(base);
mubuf->offset += offset;
continue;
uint32_t offset;
bool has_usable_ds_offset = ctx.program->chip_class >= GFX7;
if (has_usable_ds_offset &&
- i == 0 && parse_base_offset(ctx, instr.get(), i, &base, &offset) &&
+ i == 0 && parse_base_offset(ctx, instr.get(), i, &base, &offset, false) &&
base.regClass() == instr->operands[i].regClass() &&
instr->opcode != aco_opcode::ds_swizzle_b32) {
if (instr->opcode == aco_opcode::ds_write2_b32 || instr->opcode == aco_opcode::ds_read2_b32 ||
SMEM_instruction *smem = static_cast<SMEM_instruction *>(instr.get());
Temp base;
uint32_t offset;
+ bool prevent_overflow = smem->operands[0].size() > 2 || smem->prevent_overflow;
if (i == 1 && info.is_constant_or_literal(32) &&
((ctx.program->chip_class == GFX6 && info.val <= 0x3FF) ||
(ctx.program->chip_class == GFX7 && info.val <= 0xFFFFFFFF) ||
(ctx.program->chip_class >= GFX8 && info.val <= 0xFFFFF))) {
instr->operands[i] = Operand(info.val);
continue;
- } else if (i == 1 && parse_base_offset(ctx, instr.get(), i, &base, &offset) && base.regClass() == s1 && offset <= 0xFFFFF && ctx.program->chip_class >= GFX9) {
+ } else if (i == 1 && parse_base_offset(ctx, instr.get(), i, &base, &offset, prevent_overflow) && base.regClass() == s1 && offset <= 0xFFFFF && ctx.program->chip_class >= GFX9) {
bool soe = smem->operands.size() >= (!smem->definitions.empty() ? 3 : 4);
if (soe &&
(!ctx.info[smem->operands.back().tempId()].is_constant_or_literal(32) ||
new_instr->operands.back() = Operand(base);
if (!smem->definitions.empty())
new_instr->definitions[0] = smem->definitions[0];
- new_instr->can_reorder = smem->can_reorder;
- new_instr->barrier = smem->barrier;
+ new_instr->sync = smem->sync;
new_instr->glc = smem->glc;
new_instr->dlc = smem->dlc;
new_instr->nv = smem->nv;
instr->opcode == aco_opcode::v_sub_f16 ||
instr->opcode == aco_opcode::v_subrev_f16;
if (mad16 || mad32) {
- bool need_fma = mad32 ? block.fp_mode.denorm32 != 0 :
+ bool need_fma = mad32 ? (block.fp_mode.denorm32 != 0 || ctx.program->chip_class >= GFX10_3) :
(block.fp_mode.denorm16_64 != 0 || ctx.program->chip_class >= GFX10);
if (need_fma && instr->definitions[0].isPrecise())
return;