renderer: reduce synchronization and allocation overhead

Performance changes from #1495:

- Sync changed BDA ranges instead of scanning all mappings.
- Let VMA suballocate BDA buffers.
- Remove the renderer lock from presentation.
- Keep compatible pipeline caches across builds.
- Fold masked shifts that always produce zero.
This commit is contained in:
shadowbeat070
2026-10-16 16:58:07 +02:00
committed by nmzik
parent 4c6be6cd1d
commit d58aefde49
12 changed files with 277 additions and 38 deletions
@@ -608,13 +608,23 @@ void FoldInstruction(Block& block, Block::iterator instruction,
ReplaceBinaryIdentity(inst, Type::U32, UINT32_MAX)) return;
for (uint32_t arm = 0; arm < 2u; ++arm) {
const auto mask = Arg(inst, arm);
if (!IsImmediate(mask, Type::U32)) continue;
if (mask.U32() == 0u) {
Replace(inst, Value(0u));
return;
}
const auto* shifted = Arg(inst, arm ^ 1u).TryInstruction();
if (!IsImmediate(mask, Type::U32) || shifted == nullptr ||
shifted->GetOpcode() != ValueOpcode::ShiftLeftLogical32 || !HasOneUse(*shifted)) continue;
if (shifted == nullptr || shifted->GetOpcode() != ValueOpcode::ShiftLeftLogical32) continue;
const auto shift = Arg(*shifted, 1);
if (!IsImmediate(shift, Type::U32) || shift.U32() >= 32u) continue;
if (!IsImmediate(shift, Type::U32)) continue;
const auto shifted_mask = mask.U32() >> (shift.U32() & 31u);
if (shifted_mask == 0u) {
Replace(inst, Value(0u));
return;
}
if (!HasOneUse(*shifted) || shift.U32() >= 32u) continue;
const auto masked = worklist.Prepend(block, instruction, ValueOpcode::BitwiseAnd32,
{Arg(*shifted, 0), Value(mask.U32() >> shift.U32())});
{Arg(*shifted, 0), Value(shifted_mask)});
Replace(inst, Value(&*worklist.Prepend(block, instruction, ValueOpcode::ShiftLeftLogical32,
{Value(&*masked), shift})));
return;