cpu: emulate AMD-only Zen 2 instructions in software (#449)

Handle immediate EXTRQ and INSERTQ as well as MONITORX and MWAITX when the host raises illegal-instruction faults. Add unit coverage for SSE4a bit-field semantics and preserve existing load-time patching.

Co-authored-by: zocomputer <help@zocomputer.com>
This commit is contained in:
Nekono
2026-07-19 21:57:42 +03:00
committed by GitHub
parent 0c467e8c57
commit 8ef5a54ee4
4 changed files with 420 additions and 0 deletions
@@ -0,0 +1,71 @@
// Copyright (C) 2026 SharpEmu Emulator Project
// SPDX-License-Identifier: GPL-2.0-or-later
namespace SharpEmu.Core.Cpu.Emulation;
/// <summary>
/// Pure software implementation of the bit-field math behind AMD's SSE4a EXTRQ/INSERTQ
/// (immediate-form) instructions.
///
/// The direct-execution backend runs guest PS5 code natively on the host CPU. The PS5's Zen 2
/// cores implement AMD-only SSE4a (EXTRQ/INSERTQ), but Intel hosts - and Rosetta 2 on Apple
/// Silicon - do not, so they raise #UD (STATUS_ILLEGAL_INSTRUCTION) instead of executing the
/// opcode. SharpEmu already rewrites one specific compiled EXTRQ+VPBLENDD idiom at load time
/// (see <see cref="Native.Sse4aExtrqBlendPatch"/>), but any other occurrence of EXTRQ/INSERTQ -
/// a different register allocation, a title built with a different compiler version, and so on
/// - still aborts the title. This class ported from Kyty's
/// <c>Loader::X64InstructionEmulator::TryEmulateSse4a</c> provides the general bit-field
/// extract/insert so the illegal-instruction handler can finish *any* immediate-form
/// EXTRQ/INSERTQ in software and resume, instead of relying on a single hard-coded byte pattern.
///
/// The methods operate on plain 64-bit integers rather than the OS CONTEXT record so the bit
/// math can be unit-tested in isolation; the unsafe CONTEXT/XMM plumbing lives in the backend
/// adapter (<see cref="Native.DirectExecutionBackend"/>).
/// </summary>
public static class Sse4aBitFieldEmulator
{
public static bool IsValidBitField(int length, int index)
{
var len = length & 0x3F;
var idx = index & 0x3F;
return (len != 0 || idx == 0) && (len == 0 ? idx == 0 : idx + len <= 64);
}
public static ulong ExtractBitField(ulong value, int length, int index)
{
var len = length & 0x3F;
var idx = index & 0x3F;
if (!IsValidBitField(length, index))
{
return 0;
}
if (len == 0)
{
return value;
}
var mask = len == 64 ? ulong.MaxValue : (1UL << len) - 1;
return (value >> idx) & mask;
}
public static ulong InsertBitField(ulong destination, ulong source, int length, int index)
{
var len = length & 0x3F;
var idx = index & 0x3F;
if (!IsValidBitField(length, index))
{
return destination;
}
if (len == 0)
{
return source;
}
var fieldMask = len == 64 ? ulong.MaxValue : (1UL << len) - 1;
var destinationClearMask = fieldMask << idx;
var sourceField = (source & fieldMask) << idx;
return (destination & ~destinationClearMask) | sourceField;
}
}
@@ -0,0 +1,197 @@
// Copyright (C) 2026 SharpEmu Emulator Project
// SPDX-License-Identifier: GPL-2.0-or-later
using System.Threading;
using Iced.Intel;
using SharpEmu.Core.Cpu.Emulation;
namespace SharpEmu.Core.Cpu.Native;
// General software fallback for the AMD-only instructions PS5 titles occasionally emit that a
// Zen 2-only host implements but Intel hosts (and Rosetta 2 on Apple Silicon) do not:
// - SSE4a EXTRQ/INSERTQ, immediate form
// - MONITORX/MWAITX
//
// This is a direct port of Kyty's Loader::X64InstructionEmulator (TryEmulateSse4a /
// TryEmulateMonitorxMwaitx). SharpEmu already special-cases exactly one compiled EXTRQ+VPBLENDD
// byte sequence at load time (Sse4aExtrqBlendPatch), which only helps the one idiom it was
// reverse-engineered from. This file is a general, fault-time fallback that engages for any
// immediate-form EXTRQ/INSERTQ or MONITORX/MWAITX the narrower patch (or a title using a
// different compiler/register allocation) does not cover, complementing rather than replacing
// it: the load-time patch still avoids paying the fault-and-recover cost on the hot path it was
// built for, while this method is the safety net for everything else.
//
// This is deliberately additive: DirectExecutionBackend.IllegalInstruction.cs (the BMI1/BMI2/ABM
// fallback) is untouched, and this method is only reached from VectoredHandler after that one
// has already declined to handle the fault.
public sealed partial class DirectExecutionBackend
{
// Byte offset of Xmm0 within the Win64 CONTEXT record: FltSave (the XMM_SAVE_AREA32/FXSAVE
// image) starts right after Rip at offset 256, and XmmRegisters[0] sits 160 bytes into that
// area (32-byte header + 8 legacy x87/MMX slots x 16 bytes). 256 + 160 = 416 (0x1A0). Cross-
// checked against this file's own Win64ContextSize (0x4D0): rebuilding the whole CONTEXT
// layout field-by-field from offset 0 lands on the same 0x4D0 total, which would not happen
// if this offset (or anything before it) were wrong.
private const int Win64ContextXmm0Offset = 0x1A0;
private static int _sse4aSoftwareFallbackAnnounced;
private static long _sse4aInstructionsEmulated;
private static int _monitorxSoftwareFallbackAnnounced;
private static long _monitorxInstructionsEmulated;
private unsafe bool TryRecoverAmdCompatInstruction(void* contextRecord, ulong rip)
{
if (TryRecoverMonitorxMwaitx(contextRecord, rip))
{
return true;
}
// MONITORX/MWAITX above only ever reads guest code memory and rewrites RIP, both of
// which the POSIX signal bridge (DirectExecutionBackend.PosixSignals.cs) faithfully
// round-trips through the real ucontext, so it works on every supported OS. EXTRQ/
// INSERTQ additionally read and write an XMM register: on Windows contextRecord is the
// live CONTEXT the OS resumes the thread from, so touching the Xmm0.. slots is visible
// to the guest, but on POSIX contextRecord is a CONTEXT-shaped scratch buffer that the
// bridge only populates with the 17 general-purpose registers - the XMM region is never
// read from or written back to the real mcontext/ucontext. Running this on POSIX would
// silently compute a result from stale/zeroed XMM bytes and then discard whatever it
// "wrote", so keep it Windows-only, matching Kyty's own scope for the identical fix.
return OperatingSystem.IsWindows() && TryRecoverSse4aExtractInsert(contextRecord, rip);
}
private unsafe bool TryRecoverMonitorxMwaitx(void* contextRecord, ulong rip)
{
// MONITORX (0F 01 FA) and MWAITX (0F 01 FB) are fixed 3-byte encodings with no
// ModRM/SIB/displacement/immediate, so a raw byte compare is sufficient and unambiguous.
var opcode = new byte[3];
if (!TryReadHostBytes(rip, opcode) ||
opcode[0] != 0x0F || opcode[1] != 0x01 || (opcode[2] != 0xFA && opcode[2] != 0xFB))
{
return false;
}
// PS5 titles use this pair in idle/wait loops: MONITORX arms a monitor on a cache line
// and MWAITX blocks until that line is written (or a timeout elapses). Hosts without
// the extension raise #UD on either one. We do not model the monitor itself, only its
// observable effect on guest forward progress: MONITORX becomes a no-op (arming a
// watch we never honour has no side effect of its own) and MWAITX becomes a plain
// thread yield, i.e. treat the awaited condition as already satisfied so the guest
// loop keeps making progress instead of executing an illegal opcode forever.
if (opcode[2] == 0xFB)
{
Thread.Yield();
}
WriteCtxU64(contextRecord, CTX_RIP, rip + 3);
Interlocked.Increment(ref _monitorxInstructionsEmulated);
if (Interlocked.Exchange(ref _monitorxSoftwareFallbackAnnounced, 1) == 0)
{
Console.Error.WriteLine(
"[LOADER][INFO] Host lacks AMD MONITORX/MWAITX used by the guest; " +
"emulating those instructions in software.");
}
return true;
}
private unsafe bool TryRecoverSse4aExtractInsert(void* contextRecord, ulong rip)
{
if (!OperatingSystem.IsWindows() || !TryReadFaultingInstruction(rip, out var instruction))
{
return false;
}
var isExtrq = instruction.Mnemonic == Mnemonic.Extrq;
var isInsertq = instruction.Mnemonic == Mnemonic.Insertq;
if (!isExtrq && !isInsertq)
{
return false;
}
if (isExtrq && instruction.OpCount != 3 || isInsertq && instruction.OpCount != 4)
{
return false;
}
if (instruction.GetOpKind(0) != OpKind.Register ||
!TryGetXmmOffset(instruction.GetOpRegister(0), out var destOffset))
{
return false;
}
var destLow = ReadCtxU64(contextRecord, destOffset);
if (isExtrq)
{
var length = (int)instruction.GetImmediate(1);
var index = (int)instruction.GetImmediate(2);
if (!Sse4aBitFieldEmulator.IsValidBitField(length, index))
{
return false;
}
WriteCtxU64(contextRecord, destOffset, Sse4aBitFieldEmulator.ExtractBitField(destLow, length, index));
WriteCtxU64(contextRecord, destOffset + 8, 0);
}
else
{
if (instruction.GetOpKind(1) != OpKind.Register ||
!TryGetXmmOffset(instruction.GetOpRegister(1), out var srcOffset))
{
return false;
}
var length = (int)instruction.GetImmediate(2);
var index = (int)instruction.GetImmediate(3);
if (!Sse4aBitFieldEmulator.IsValidBitField(length, index))
{
return false;
}
WriteCtxU64(contextRecord, destOffset, Sse4aBitFieldEmulator.InsertBitField(
destLow, ReadCtxU64(contextRecord, srcOffset), length, index));
WriteCtxU64(contextRecord, destOffset + 8, 0);
}
WriteCtxU64(contextRecord, CTX_RIP, rip + (ulong)instruction.Length);
Interlocked.Increment(ref _sse4aInstructionsEmulated);
if (Interlocked.Exchange(ref _sse4aSoftwareFallbackAnnounced, 1) == 0)
{
Console.Error.WriteLine(
"[LOADER][INFO] Host lacks SSE4a EXTRQ/INSERTQ used by the guest; " +
"emulating those instructions in software.");
}
return true;
}
// Maps an Iced XMM register to its byte offset in the Win64 CONTEXT record. Written as an
// explicit switch (rather than arithmetic on the Register enum) to match the style already
// used by TryGetGprSlot/TryGetGpr64Offset in DirectExecutionBackend.IllegalInstruction.cs.
private static bool TryGetXmmOffset(Register register, out int offset)
{
switch (register)
{
case Register.XMM0: offset = Win64ContextXmm0Offset + 16 * 0; return true;
case Register.XMM1: offset = Win64ContextXmm0Offset + 16 * 1; return true;
case Register.XMM2: offset = Win64ContextXmm0Offset + 16 * 2; return true;
case Register.XMM3: offset = Win64ContextXmm0Offset + 16 * 3; return true;
case Register.XMM4: offset = Win64ContextXmm0Offset + 16 * 4; return true;
case Register.XMM5: offset = Win64ContextXmm0Offset + 16 * 5; return true;
case Register.XMM6: offset = Win64ContextXmm0Offset + 16 * 6; return true;
case Register.XMM7: offset = Win64ContextXmm0Offset + 16 * 7; return true;
case Register.XMM8: offset = Win64ContextXmm0Offset + 16 * 8; return true;
case Register.XMM9: offset = Win64ContextXmm0Offset + 16 * 9; return true;
case Register.XMM10: offset = Win64ContextXmm0Offset + 16 * 10; return true;
case Register.XMM11: offset = Win64ContextXmm0Offset + 16 * 11; return true;
case Register.XMM12: offset = Win64ContextXmm0Offset + 16 * 12; return true;
case Register.XMM13: offset = Win64ContextXmm0Offset + 16 * 13; return true;
case Register.XMM14: offset = Win64ContextXmm0Offset + 16 * 14; return true;
case Register.XMM15: offset = Win64ContextXmm0Offset + 16 * 15; return true;
default:
offset = 0;
return false;
}
}
}
@@ -133,6 +133,11 @@ public sealed partial class DirectExecutionBackend
{
return -1;
}
if (exceptionCode == StatusIllegalInstruction &&
TryRecoverAmdCompatInstruction(contextRecord, rip))
{
return -1;
}
if (IsBenignHostDebugException(exceptionCode))
{
return -1;