// Copyright 2021 yuzu Emulator Project
// Licensed under GPLv2 or any later version
// Refer to the license.txt file included.
#include <algorithm>
#include <compare>
#include <optional>
#include <ranges>
#include <boost/container/flat_set.hpp>
#include <boost/container/small_vector.hpp>
#include "shader_recompiler/frontend/ir/basic_block.h"
#include "shader_recompiler/frontend/ir/ir_emitter.h"
#include "shader_recompiler/frontend/ir/microinstruction.h"
#include "shader_recompiler/ir_opt/passes.h"
namespace Shader::Optimization {
namespace {
/// Address in constant buffers to the storage buffer descriptor
struct StorageBufferAddr {
auto operator<=>(const StorageBufferAddr&) const noexcept = default;
u32 index;
u32 offset;
};
/// Block iterator to a global memory instruction and the storage buffer it uses
struct StorageInst {
StorageBufferAddr storage_buffer;
IR::Block::iterator inst;
};
/// Bias towards a certain range of constant buffers when looking for storage buffers
struct Bias {
u32 index;
u32 offset_begin;
u32 offset_end;
};
using StorageBufferSet =
boost::container::flat_set<StorageBufferAddr, std::less<StorageBufferAddr>,
boost::container::small_vector<StorageBufferAddr, 16>>;
using StorageInstVector = boost::container::small_vector<StorageInst, 32>;
/// Returns true when the instruction is a global memory instruction
bool IsGlobalMemory(const IR::Inst& inst) {
switch (inst.Opcode()) {
case IR::Opcode::LoadGlobalS8:
case IR::Opcode::LoadGlobalU8:
case IR::Opcode::LoadGlobalS16:
case IR::Opcode::LoadGlobalU16:
case IR::Opcode::LoadGlobal32:
case IR::Opcode::LoadGlobal64:
case IR::Opcode::LoadGlobal128:
case IR::Opcode::WriteGlobalS8:
case IR::Opcode::WriteGlobalU8:
case IR::Opcode::WriteGlobalS16:
case IR::Opcode::WriteGlobalU16:
case IR::Opcode::WriteGlobal32:
case IR::Opcode::WriteGlobal64:
case IR::Opcode::WriteGlobal128:
return true;
default:
return false;
}
}
/// Converts a global memory opcode to its storage buffer equivalent
IR::Opcode GlobalToStorage(IR::Opcode opcode) {
switch (opcode) {
case IR::Opcode::LoadGlobalS8:
return IR::Opcode::LoadStorageS8;
case IR::Opcode::LoadGlobalU8:
return IR::Opcode::LoadStorageU8;
case IR::Opcode::LoadGlobalS16:
return IR::Opcode::LoadStorageS16;
case IR::Opcode::LoadGlobalU16:
return IR::Opcode::LoadStorageU16;
case IR::Opcode::LoadGlobal32:
return IR::Opcode::LoadStorage32;
case IR::Opcode::LoadGlobal64:
return IR::Opcode::LoadStorage64;
case IR::Opcode::LoadGlobal128:
return IR::Opcode::LoadStorage128;
case IR::Opcode::WriteGlobalS8:
return IR::Opcode::WriteStorageS8;
case IR::Opcode::WriteGlobalU8:
return IR::Opcode::WriteStorageU8;
case IR::Opcode::WriteGlobalS16:
return IR::Opcode::WriteStorageS16;
case IR::Opcode::WriteGlobalU16:
return IR::Opcode::WriteStorageU16;
case IR::Opcode::WriteGlobal32:
return IR::Opcode::WriteStorage32;
case IR::Opcode::WriteGlobal64:
return IR::Opcode::WriteStorage64;
case IR::Opcode::WriteGlobal128:
return IR::Opcode::WriteStorage128;
default:
throw InvalidArgument("Invalid global memory opcode {}", opcode);
}
}
/// Returns true when a storage buffer address satisfies a bias
bool MeetsBias(const StorageBufferAddr& storage_buffer, const Bias& bias) noexcept {
return storage_buffer.index == bias.index && storage_buffer.offset >= bias.offset_begin &&
storage_buffer.offset < bias.offset_end;
}
/// Ignores a global memory operation, reads return zero and writes are ignored
void IgnoreGlobalMemory(IR::Block& block, IR::Block::iterator inst) {
const IR::Value zero{u32{0}};
switch (inst->Opcode()) {
case IR::Opcode::LoadGlobalS8:
case IR::Opcode::LoadGlobalU8:
case IR::Opcode::LoadGlobalS16:
case IR::Opcode::LoadGlobalU16:
case IR::Opcode::LoadGlobal32:
inst->ReplaceUsesWith(zero);
break;
case IR::Opcode::LoadGlobal64:
inst->ReplaceUsesWith(
IR::Value{&*block.PrependNewInst(inst, IR::Opcode::CompositeConstruct2, {zero, zero})});
break;
case IR::Opcode::LoadGlobal128:
inst->ReplaceUsesWith(IR::Value{&*block.PrependNewInst(
inst, IR::Opcode::CompositeConstruct4, {zero, zero, zero, zero})});
break;
case IR::Opcode::WriteGlobalS8:
case IR::Opcode::WriteGlobalU8:
case IR::Opcode::WriteGlobalS16:
case IR::Opcode::WriteGlobalU16:
case IR::Opcode::WriteGlobal32:
case IR::Opcode::WriteGlobal64:
case IR::Opcode::WriteGlobal128:
inst->Invalidate();
break;
default:
throw LogicError("Invalid opcode to ignore its global memory operation {}", inst->Opcode());
}
}
/// Recursively tries to track the storage buffer address used by a global memory instruction
std::optional<StorageBufferAddr> Track(const IR::Value& value, const Bias* bias) {
if (value.IsImmediate()) {
// Immediates can't be a storage buffer
return std::nullopt;
}
const IR::Inst* const inst{value.InstRecursive()};
if (inst->Opcode() == IR::Opcode::GetCbuf) {
const IR::Value index{inst->Arg(0)};
const IR::Value offset{inst->Arg(1)};
if (!index.IsImmediate()) {
// Definitely not a storage buffer if it's read from a non-immediate index
return std::nullopt;
}
if (!offset.IsImmediate()) {
// TODO: Support SSBO arrays
return std::nullopt;
}
const StorageBufferAddr storage_buffer{
.index = index.U32(),
.offset = offset.U32(),
};
if (bias && !MeetsBias(storage_buffer, *bias)) {
// We have to blacklist some addresses in case we wrongly point to them
return std::nullopt;
}
return storage_buffer;
}
// Reversed loops are more likely to find the right result
for (size_t arg = inst->NumArgs(); arg--;) {
if (const std::optional storage_buffer{Track(inst->Arg(arg), bias)}) {
return *storage_buffer;
}
}
return std::nullopt;
}
/// Collects the storage buffer used by a global memory instruction and the instruction itself
void CollectStorageBuffers(IR::Block& block, IR::Block::iterator inst,
StorageBufferSet& storage_buffer_set, StorageInstVector& to_replace) {
// NVN puts storage buffers in a specific range, we have to bias towards these addresses to
// avoid getting false positives
static constexpr Bias nvn_bias{
.index{0},
.offset_begin{0x110},
.offset_end{0x610},
};
// First try to find storage buffers in the NVN address
const IR::U64 addr{inst->Arg(0)};
std::optional<StorageBufferAddr> storage_buffer{Track(addr, &nvn_bias)};
if (!storage_buffer) {
// If it fails, track without a bias
storage_buffer = Track(addr, nullptr);
if (!storage_buffer) {
// If that also failed, drop the global memory usage
IgnoreGlobalMemory(block, inst);
}
}
// Collect storage buffer and the instruction
storage_buffer_set.insert(*storage_buffer);
to_replace.push_back(StorageInst{
.storage_buffer{*storage_buffer},
.inst{inst},
});
}
/// Tries to track the first 32-bits of a global memory instruction
std::optional<IR::U32> TrackLowAddress(IR::IREmitter& ir, IR::Inst* inst) {
// The first argument is the low level GPU pointer to the global memory instruction
const IR::U64 addr{inst->Arg(0)};
if (addr.IsImmediate()) {
// Not much we can do if it's an immediate
return std::nullopt;
}
// This address is expected to either be a PackUint2x32 or a IAdd64
IR::Inst* addr_inst{addr.InstRecursive()};
s32 imm_offset{0};
if (addr_inst->Opcode() == IR::Opcode::IAdd64) {
// If it's an IAdd64, get the immediate offset it is applying and grab the address
// instruction. This expects for the instruction to be canonicalized having the address on
// the first argument and the immediate offset on the second one.
const IR::U64 imm_offset_value{addr_inst->Arg(1)};
if (!imm_offset_value.IsImmediate()) {
return std::nullopt;
}
imm_offset = static_cast<s32>(static_cast<s64>(imm_offset_value.U64()));
const IR::U64 iadd_addr{addr_inst->Arg(0)};
if (iadd_addr.IsImmediate()) {
return std::nullopt;
}
addr_inst = iadd_addr.Inst();
}
// With IAdd64 handled, now PackUint2x32 is expected without exceptions
if (addr_inst->Opcode() != IR::Opcode::PackUint2x32) {
return std::nullopt;
}
// PackUint2x32 is expected to be generated from a vector
const IR::Value vector{addr_inst->Arg(0)};
if (vector.IsImmediate()) {
return std::nullopt;
}
// This vector is expected to be a CompositeConstruct2
IR::Inst* const vector_inst{vector.InstRecursive()};
if (vector_inst->Opcode() != IR::Opcode::CompositeConstruct2) {
return std::nullopt;
}
// Grab the first argument from the CompositeConstruct2, this is the low address.
// Re-apply the offset in case we found one.
const IR::U32 low_addr{vector_inst->Arg(0)};
return imm_offset != 0 ? IR::U32{ir.IAdd(low_addr, ir.Imm32(imm_offset))} : low_addr;
}
/// Returns the offset in indices (not bytes) for an equivalent storage instruction
IR::U32 StorageOffset(IR::Block& block, IR::Block::iterator inst, StorageBufferAddr buffer) {
IR::IREmitter ir{block, inst};
IR::U32 offset;
if (const std::optional<IR::U32> low_addr{TrackLowAddress(ir, &*inst)}) {
offset = *low_addr;
} else {
offset = ir.ConvertU(32, IR::U64{inst->Arg(0)});
}
// Subtract the least significant 32 bits from the guest offset. The result is the storage
// buffer offset in bytes.
const IR::U32 low_cbuf{ir.GetCbuf(ir.Imm32(buffer.index), ir.Imm32(buffer.offset))};
return ir.ISub(offset, low_cbuf);
}
/// Replace a global memory load instruction with its storage buffer equivalent
void ReplaceLoad(IR::Block& block, IR::Block::iterator inst, const IR::U32& storage_index,
const IR::U32& offset) {
const IR::Opcode new_opcode{GlobalToStorage(inst->Opcode())};
const IR::Value value{&*block.PrependNewInst(inst, new_opcode, {storage_index, offset})};
inst->ReplaceUsesWith(value);
}
/// Replace a global memory write instruction with its storage buffer equivalent
void ReplaceWrite(IR::Block& block, IR::Block::iterator inst, const IR::U32& storage_index,
const IR::U32& offset) {
const IR::Opcode new_opcode{GlobalToStorage(inst->Opcode())};
block.PrependNewInst(inst, new_opcode, {storage_index, offset, inst->Arg(1)});
inst->Invalidate();
}
/// Replace a global memory instruction with its storage buffer equivalent
void Replace(IR::Block& block, IR::Block::iterator inst, const IR::U32& storage_index,
const IR::U32& offset) {
switch (inst->Opcode()) {
case IR::Opcode::LoadGlobalS8:
case IR::Opcode::LoadGlobalU8:
case IR::Opcode::LoadGlobalS16:
case IR::Opcode::LoadGlobalU16:
case IR::Opcode::LoadGlobal32:
case IR::Opcode::LoadGlobal64:
case IR::Opcode::LoadGlobal128:
return ReplaceLoad(block, inst, storage_index, offset);
case IR::Opcode::WriteGlobalS8:
case IR::Opcode::WriteGlobalU8:
case IR::Opcode::WriteGlobalS16:
case IR::Opcode::WriteGlobalU16:
case IR::Opcode::WriteGlobal32:
case IR::Opcode::WriteGlobal64:
case IR::Opcode::WriteGlobal128:
return ReplaceWrite(block, inst, storage_index, offset);
default:
throw InvalidArgument("Invalid global memory opcode {}", inst->Opcode());
}
}
} // Anonymous namespace
void GlobalMemoryToStorageBufferPass(IR::Block& block) {
StorageBufferSet storage_buffers;
StorageInstVector to_replace;
for (IR::Block::iterator inst{block.begin()}; inst != block.end(); ++inst) {
if (!IsGlobalMemory(*inst)) {
continue;
}
CollectStorageBuffers(block, inst, storage_buffers, to_replace);
}
for (const auto [storage_buffer, inst] : to_replace) {
const auto it{storage_buffers.find(storage_buffer)};
const IR::U32 storage_index{IR::Value{static_cast<u32>(storage_buffers.index_of(it))}};
const IR::U32 offset{StorageOffset(block, inst, storage_buffer)};
Replace(block, inst, storage_index, offset);
}
}
} // namespace Shader::Optimization