Files
sharpemu/src/SharpEmu.Core/Cpu/Native/DirectExecutionBackend.Amd64Compat.cs
T
999sian ada67a1924 cpu: recover SSE4a EXTRQ/INSERTQ faults on Linux (#482)
The fault-time SSE4a fallback was Windows-only because the POSIX signal
bridge never carried XMM state: the CONTEXT scratch buffer only held the
17 general-purpose registers, so emulating EXTRQ/INSERTQ there would
have computed results from zeroed bytes and discarded the write. Bridge
the XMM registers on Linux by copying them between the mcontext's
FXSAVE image (kernel sigcontext ABI, libc-independent) and the CONTEXT
FltSave slots on capture and write-back, and gate the recovery on that
bridge instead of on Windows. Darwin still declines: its XMM area
remains unbridged.

With this, guest EXTRQ/INSERTQ on Linux hosts without SSE4a (any Intel
CPU) resumes with correct register state instead of dying on an
unrecovered SIGILL (#328).
2026-07-21 14:18:21 +03:00

200 lines
9.4 KiB
C#

// 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, and on Linux the bridge copies the mcontext's FXSAVE image into the
// Xmm0.. slots and writes them back through sigreturn (_posixXmmContextBridged). On
// Darwin the XMM area is still a zeroed scratch buffer - running this there would
// silently compute a result from stale bytes and then discard whatever it "wrote", so
// the recovery declines until that bridge exists.
return (OperatingSystem.IsWindows() || _posixXmmContextBridged) &&
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() && !_posixXmmContextBridged ||
!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;
}
}
}