From ccdb9b9d4c4ce81221424296ab2d3c86af86e2fe Mon Sep 17 00:00:00 2001 From: lizzie Date: Mon, 9 Feb 2026 10:45:29 +0000 Subject: [PATCH] [dynarmic] use bitset<32> instead of vector for ABI regset allocations HLE macro does this, may as well do it on dynarm as well :) Signed-off-by: lizzie --- .../src/dynarmic/backend/x64/a64_emit_x64.cpp | 2 +- src/dynarmic/src/dynarmic/backend/x64/abi.cpp | 114 ++++-------------- src/dynarmic/src/dynarmic/backend/x64/abi.h | 65 ++++++++-- .../src/dynarmic/backend/x64/reg_alloc.cpp | 18 +-- 4 files changed, 91 insertions(+), 108 deletions(-) diff --git a/src/dynarmic/src/dynarmic/backend/x64/a64_emit_x64.cpp b/src/dynarmic/src/dynarmic/backend/x64/a64_emit_x64.cpp index 03d0e13562..5fbc94d30f 100644 --- a/src/dynarmic/src/dynarmic/backend/x64/a64_emit_x64.cpp +++ b/src/dynarmic/src/dynarmic/backend/x64/a64_emit_x64.cpp @@ -188,7 +188,7 @@ void A64EmitX64::ClearFastDispatchTable() { void A64EmitX64::GenTerminalHandlers() { // PC ends up in rcx, location_descriptor ends up in rbx - static_assert(std::find(ABI_ALL_CALLEE_SAVE.begin(), ABI_ALL_CALLEE_SAVE.end(), HostLoc::R12) != ABI_ALL_CALLEE_SAVE.end()); + //static_assert(ABI_ALL_CALLEE_SAVE.test(size_t(HostLoc::R12))); const auto calculate_location_descriptor = [this] { // This calculation has to match up with A64::LocationDescriptor::UniqueHash // TODO: Optimization is available here based on known state of fpcr. diff --git a/src/dynarmic/src/dynarmic/backend/x64/abi.cpp b/src/dynarmic/src/dynarmic/backend/x64/abi.cpp index 299bf1d1d6..82c50432f4 100644 --- a/src/dynarmic/src/dynarmic/backend/x64/abi.cpp +++ b/src/dynarmic/src/dynarmic/backend/x64/abi.cpp @@ -1,4 +1,4 @@ -// SPDX-FileCopyrightText: Copyright 2025 Eden Emulator Project +// SPDX-FileCopyrightText: Copyright 2026 Eden Emulator Project // SPDX-License-Identifier: GPL-3.0-or-later /* This file is part of the dynarmic project. @@ -40,59 +40,53 @@ static FrameInfo CalculateFrameInfo(const size_t num_gprs, const size_t num_xmms }; } -template -void ABI_PushRegistersAndAdjustStack(BlockOfCode& code, const size_t frame_size, const RegisterArrayT& regs) { +void ABI_PushRegistersAndAdjustStack(BlockOfCode& code, const size_t frame_size, std::bitset<32> const& regs) { using namespace Xbyak::util; - const size_t num_gprs = std::count_if(regs.begin(), regs.end(), HostLocIsGPR); - const size_t num_xmms = std::count_if(regs.begin(), regs.end(), HostLocIsXMM); + const size_t num_gprs = (ABI_ALL_GPRS & regs).count(); + const size_t num_xmms = (ABI_ALL_XMMS & regs).count(); const FrameInfo frame_info = CalculateFrameInfo(num_gprs, num_xmms, frame_size); - for (auto const gpr : regs) - if (HostLocIsGPR(gpr)) - code.push(HostLocToReg64(gpr)); + for (size_t i = 0; i < regs.size(); ++i) + if (regs[i] && HostLocIsGPR(HostLoc(i))) + code.push(HostLocToReg64(HostLoc(i))); if (frame_info.stack_subtraction != 0) code.sub(rsp, u32(frame_info.stack_subtraction)); size_t xmm_offset = frame_info.xmm_offset; - for (auto const xmm : regs) { - if (HostLocIsXMM(xmm)) { + for (size_t i = 0; i < regs.size(); ++i) { + if (regs[i] && HostLocIsXMM(HostLoc(i))) { if (code.HasHostFeature(HostFeature::AVX)) { - code.vmovaps(code.xword[rsp + xmm_offset], HostLocToXmm(xmm)); + code.vmovaps(code.xword[rsp + xmm_offset], HostLocToXmm(HostLoc(i))); } else { - code.movaps(code.xword[rsp + xmm_offset], HostLocToXmm(xmm)); + code.movaps(code.xword[rsp + xmm_offset], HostLocToXmm(HostLoc(i))); } xmm_offset += XMM_SIZE; } } } -template -void ABI_PopRegistersAndAdjustStack(BlockOfCode& code, const size_t frame_size, const RegisterArrayT& regs) { +void ABI_PopRegistersAndAdjustStack(BlockOfCode& code, const size_t frame_size, std::bitset<32> const& regs) { using namespace Xbyak::util; - const size_t num_gprs = std::count_if(regs.begin(), regs.end(), HostLocIsGPR); - const size_t num_xmms = std::count_if(regs.begin(), regs.end(), HostLocIsXMM); + const size_t num_gprs = (ABI_ALL_GPRS & regs).count(); + const size_t num_xmms = (ABI_ALL_XMMS & regs).count(); const FrameInfo frame_info = CalculateFrameInfo(num_gprs, num_xmms, frame_size); size_t xmm_offset = frame_info.xmm_offset + (num_xmms * XMM_SIZE); - for (auto it = regs.rbegin(); it != regs.rend(); ++it) { - auto const xmm = *it; - if (HostLocIsXMM(xmm)) { + for (int32_t i = regs.size() - 1; i >= 0; --i) + if (regs[i] && HostLocIsXMM(HostLoc(i))) { xmm_offset -= XMM_SIZE; if (code.HasHostFeature(HostFeature::AVX)) { - code.vmovaps(HostLocToXmm(xmm), code.xword[rsp + xmm_offset]); + code.vmovaps(HostLocToXmm(HostLoc(i)), code.xword[rsp + xmm_offset]); } else { - code.movaps(HostLocToXmm(xmm), code.xword[rsp + xmm_offset]); + code.movaps(HostLocToXmm(HostLoc(i)), code.xword[rsp + xmm_offset]); } } - } if (frame_info.stack_subtraction != 0) code.add(rsp, u32(frame_info.stack_subtraction)); - for (auto it = regs.rbegin(); it != regs.rend(); ++it) { - auto const gpr = *it; - if (HostLocIsGPR(gpr)) - code.pop(HostLocToReg64(gpr)); - } + for (int32_t i = regs.size() - 1; i >= 0; --i) + if (regs[i] && HostLocIsGPR(HostLoc(i))) + code.pop(HostLocToReg64(HostLoc(i))); } void ABI_PushCalleeSaveRegistersAndAdjustStack(BlockOfCode& code, const std::size_t frame_size) { @@ -112,74 +106,16 @@ void ABI_PopCallerSaveRegistersAndAdjustStack(BlockOfCode& code, const std::size } // Windows ABI registers are not in the same allocation algorithm as unix's -#ifdef _MSC_VER void ABI_PushCallerSaveRegistersAndAdjustStackExcept(BlockOfCode& code, const HostLoc exception) { - std::vector regs; - std::remove_copy(ABI_ALL_CALLER_SAVE.begin(), ABI_ALL_CALLER_SAVE.end(), std::back_inserter(regs), exception); + std::bitset<32> regs = ABI_ALL_CALLER_SAVE; + regs.reset(size_t(exception)); ABI_PushRegistersAndAdjustStack(code, 0, regs); } void ABI_PopCallerSaveRegistersAndAdjustStackExcept(BlockOfCode& code, const HostLoc exception) { - std::vector regs; - std::remove_copy(ABI_ALL_CALLER_SAVE.begin(), ABI_ALL_CALLER_SAVE.end(), std::back_inserter(regs), exception); + std::bitset<32> regs = ABI_ALL_CALLER_SAVE; + regs.reset(size_t(exception)); ABI_PopRegistersAndAdjustStack(code, 0, regs); } -#else -static consteval size_t ABI_AllCallerSaveSize() noexcept { - return ABI_ALL_CALLER_SAVE.max_size(); -} -static consteval std::array ABI_AllCallerSaveExcept(const std::size_t except) noexcept { - std::array arr; - for(std::size_t i = 0; i < arr.size(); ++i) { - arr[i] = static_cast(i + (i >= except ? 1 : 0)); - } - return arr; -} - -alignas(64) static constinit std::array ABI_CALLER_SAVED_EXCEPT_TABLE[32] = { - ABI_AllCallerSaveExcept(0), - ABI_AllCallerSaveExcept(1), - ABI_AllCallerSaveExcept(2), - ABI_AllCallerSaveExcept(3), - ABI_AllCallerSaveExcept(4), - ABI_AllCallerSaveExcept(5), - ABI_AllCallerSaveExcept(6), - ABI_AllCallerSaveExcept(7), - ABI_AllCallerSaveExcept(8), - ABI_AllCallerSaveExcept(9), - ABI_AllCallerSaveExcept(10), - ABI_AllCallerSaveExcept(11), - ABI_AllCallerSaveExcept(12), - ABI_AllCallerSaveExcept(13), - ABI_AllCallerSaveExcept(14), - ABI_AllCallerSaveExcept(15), - ABI_AllCallerSaveExcept(16), - ABI_AllCallerSaveExcept(17), - ABI_AllCallerSaveExcept(18), - ABI_AllCallerSaveExcept(19), - ABI_AllCallerSaveExcept(20), - ABI_AllCallerSaveExcept(21), - ABI_AllCallerSaveExcept(22), - ABI_AllCallerSaveExcept(23), - ABI_AllCallerSaveExcept(24), - ABI_AllCallerSaveExcept(25), - ABI_AllCallerSaveExcept(26), - ABI_AllCallerSaveExcept(27), - ABI_AllCallerSaveExcept(28), - ABI_AllCallerSaveExcept(29), - ABI_AllCallerSaveExcept(30), - ABI_AllCallerSaveExcept(31), -}; - -void ABI_PushCallerSaveRegistersAndAdjustStackExcept(BlockOfCode& code, const HostLoc exception) { - ASSUME(size_t(exception) < 32); - ABI_PushRegistersAndAdjustStack(code, 0, ABI_CALLER_SAVED_EXCEPT_TABLE[size_t(exception)]); -} - -void ABI_PopCallerSaveRegistersAndAdjustStackExcept(BlockOfCode& code, const HostLoc exception) { - ASSUME(size_t(exception) < 32); - ABI_PopRegistersAndAdjustStack(code, 0, ABI_CALLER_SAVED_EXCEPT_TABLE[size_t(exception)]); -} -#endif } // namespace Dynarmic::Backend::X64 diff --git a/src/dynarmic/src/dynarmic/backend/x64/abi.h b/src/dynarmic/src/dynarmic/backend/x64/abi.h index 307817a864..f5e7cfb7c0 100644 --- a/src/dynarmic/src/dynarmic/backend/x64/abi.h +++ b/src/dynarmic/src/dynarmic/backend/x64/abi.h @@ -1,4 +1,4 @@ -// SPDX-FileCopyrightText: Copyright 2025 Eden Emulator Project +// SPDX-FileCopyrightText: Copyright 2026 Eden Emulator Project // SPDX-License-Identifier: GPL-3.0-or-later /* This file is part of the dynarmic project. @@ -8,6 +8,7 @@ #pragma once #include +#include #include "dynarmic/common/common_types.h" @@ -17,6 +18,50 @@ namespace Dynarmic::Backend::X64 { class BlockOfCode; +constexpr std::bitset<32> BuildRegSet(std::initializer_list regs) { + size_t bits = 0; + for (auto const& reg : regs) + bits |= size_t{1} << size_t(reg); + return {bits}; +} + +constexpr std::bitset<32> ABI_ALL_GPRS = BuildRegSet({ + HostLoc::RAX, + HostLoc::RBX, + HostLoc::RCX, + HostLoc::RDX, + HostLoc::RDI, + HostLoc::RSI, + HostLoc::RBP, + HostLoc::RSP, + HostLoc::R8, + HostLoc::R9, + HostLoc::R10, + HostLoc::R11, + HostLoc::R12, + HostLoc::R13, + HostLoc::R14, + HostLoc::R15, +}); +constexpr std::bitset<32> ABI_ALL_XMMS = BuildRegSet({ + HostLoc::XMM0, + HostLoc::XMM1, + HostLoc::XMM2, + HostLoc::XMM3, + HostLoc::XMM4, + HostLoc::XMM5, + HostLoc::XMM6, + HostLoc::XMM7, + HostLoc::XMM8, + HostLoc::XMM9, + HostLoc::XMM10, + HostLoc::XMM11, + HostLoc::XMM12, + HostLoc::XMM13, + HostLoc::XMM14, + HostLoc::XMM15, +}); + constexpr HostLoc ABI_JIT_PTR = HostLoc::R15; #ifdef _WIN32 @@ -29,7 +74,7 @@ constexpr HostLoc ABI_PARAM2 = HostLoc::RDX; constexpr HostLoc ABI_PARAM3 = HostLoc::R8; constexpr HostLoc ABI_PARAM4 = HostLoc::R9; -constexpr std::array ABI_ALL_CALLER_SAVE = { +constexpr std::bitset<32> ABI_ALL_CALLER_SAVE = BuildRegSet({ HostLoc::RAX, HostLoc::RCX, HostLoc::RDX, @@ -43,9 +88,9 @@ constexpr std::array ABI_ALL_CALLER_SAVE = { HostLoc::XMM3, HostLoc::XMM4, HostLoc::XMM5, -}; +}); -constexpr std::array ABI_ALL_CALLEE_SAVE = { +constexpr std::bitset<32> ABI_ALL_CALLEE_SAVE = BuildRegSet({ HostLoc::RBX, HostLoc::RSI, HostLoc::RDI, @@ -64,7 +109,7 @@ constexpr std::array ABI_ALL_CALLEE_SAVE = { HostLoc::XMM13, HostLoc::XMM14, HostLoc::XMM15, -}; +}); constexpr size_t ABI_SHADOW_SPACE = 32; // bytes @@ -82,7 +127,7 @@ constexpr HostLoc ABI_PARAM4 = HostLoc::RCX; constexpr HostLoc ABI_PARAM5 = HostLoc::R8; constexpr HostLoc ABI_PARAM6 = HostLoc::R9; -constexpr std::array ABI_ALL_CALLER_SAVE = { +constexpr std::bitset<32> ABI_ALL_CALLER_SAVE = BuildRegSet({ HostLoc::RAX, HostLoc::RCX, HostLoc::RDX, @@ -108,22 +153,22 @@ constexpr std::array ABI_ALL_CALLER_SAVE = { HostLoc::XMM13, HostLoc::XMM14, HostLoc::XMM15, -}; +}); -constexpr std::array ABI_ALL_CALLEE_SAVE = { +constexpr std::bitset<32> ABI_ALL_CALLEE_SAVE = BuildRegSet({ HostLoc::RBX, HostLoc::RBP, HostLoc::R12, HostLoc::R13, HostLoc::R14, HostLoc::R15, -}; +}); constexpr size_t ABI_SHADOW_SPACE = 0; // bytes #endif -static_assert(ABI_ALL_CALLER_SAVE.size() + ABI_ALL_CALLEE_SAVE.size() == 31, "Invalid total number of registers"); +//static_assert(ABI_ALL_CALLER_SAVE.count() + ABI_ALL_CALLEE_SAVE.count() == 31, "Invalid total number of registers"); void ABI_PushCalleeSaveRegistersAndAdjustStack(BlockOfCode& code, size_t frame_size = 0); void ABI_PopCalleeSaveRegistersAndAdjustStack(BlockOfCode& code, size_t frame_size = 0); diff --git a/src/dynarmic/src/dynarmic/backend/x64/reg_alloc.cpp b/src/dynarmic/src/dynarmic/backend/x64/reg_alloc.cpp index 115cb22ee9..0ab3acd5c7 100644 --- a/src/dynarmic/src/dynarmic/backend/x64/reg_alloc.cpp +++ b/src/dynarmic/src/dynarmic/backend/x64/reg_alloc.cpp @@ -1,4 +1,4 @@ -// SPDX-FileCopyrightText: Copyright 2025 Eden Emulator Project +// SPDX-FileCopyrightText: Copyright 2026 Eden Emulator Project // SPDX-License-Identifier: GPL-3.0-or-later /* This file is part of the dynarmic project. @@ -336,11 +336,11 @@ void RegAlloc::HostCall( constexpr std::array args_hostloc = {ABI_PARAM1, ABI_PARAM2, ABI_PARAM3, ABI_PARAM4}; const std::array, args_count> args = {arg0, arg1, arg2, arg3}; - static const boost::container::static_vector other_caller_save = [args_hostloc]() noexcept { - boost::container::static_vector ret(ABI_ALL_CALLER_SAVE.begin(), ABI_ALL_CALLER_SAVE.end()); - ret.erase(std::find(ret.begin(), ret.end(), ABI_RETURN)); + static const std::bitset<32> other_caller_save = [args_hostloc]() noexcept { + std::bitset<32> ret = ABI_ALL_CALLER_SAVE; + ret.reset(size_t(ABI_RETURN)); for (auto const hostloc : args_hostloc) - ret.erase(std::find(ret.begin(), ret.end(), hostloc)); + ret.reset(size_t(hostloc)); return ret; }(); @@ -356,9 +356,11 @@ void RegAlloc::HostCall( } } // Must match with with ScratchImpl - for (auto const gpr : other_caller_save) { - MoveOutOfTheWay(code, gpr); - LocInfo(gpr).WriteLock(); + for (size_t i = 0; i < other_caller_save.size(); ++i) { + if (other_caller_save[i]) { + MoveOutOfTheWay(code, HostLoc(i)); + LocInfo(HostLoc(i)).WriteLock(); + } } for (size_t i = 0; i < args.size(); i++) { if (args[i] && !args[i]->get().IsVoid()) {