[dynarmic] backport WAITPKG based spinlocks (#2614)

Alder Lake/Tremont or superior feature only, based on archived UMWAIT by @Wunkolo. With some fixes due to register corruption/clobber (regalloc was blind to this)

Signed-off-by: lizzie <lizzie@eden-emu.dev>

Reviewed-on: https://git.eden-emu.dev/eden-emu/eden/pulls/2614
Reviewed-by: MaranBr <maranbr@eden-emu.dev>
Reviewed-by: DraVee <dravee@eden-emu.dev>
Reviewed-by: CamilleLaVey <camillelavey99@gmail.com>
Co-authored-by: lizzie <lizzie@eden-emu.dev>
Co-committed-by: lizzie <lizzie@eden-emu.dev>

src/dynarmic/src/dynarmic/backend/x64/block_of_code.cpp

@@ -194,6 +194,8 @@ HostFeature GetHostFeatures() {
         features |= HostFeature::LZCNT;
     if (cpu_info.has(Cpu::tGFNI))
         features |= HostFeature::GFNI;
+    if (cpu_info.has(Cpu::tWAITPKG))
+        features |= HostFeature::WAITPKG;
 
     if (cpu_info.has(Cpu::tBMI2)) {
         // BMI2 instructions such as pdep and pext have been very slow up until Zen 3.

src/dynarmic/src/dynarmic/backend/x64/emit_x64_memory.cpp.inc

@@ -420,10 +420,11 @@ void AxxEmitX64::EmitExclusiveWriteMemoryInline(AxxEmitContext& ctx, IR::Inst* i
     const Xbyak::Reg64 vaddr = ctx.reg_alloc.UseGpr(code, args[1]);
     const Xbyak::Reg32 status = ctx.reg_alloc.ScratchGpr(code).cvt32();
     const Xbyak::Reg64 tmp = ctx.reg_alloc.ScratchGpr(code);
+    const Xbyak::Reg64 tmp2 = ctx.reg_alloc.ScratchGpr(code);
 
     const auto wrapped_fn = exclusive_write_fallbacks[std::make_tuple(ordered, bitsize, vaddr.getIdx(), value.getIdx())];
 
-    EmitExclusiveLock(code, conf, tmp, eax);
+    EmitExclusiveLock(code, conf, tmp, tmp2.cvt32());
 
     SharedLabel end = GenSharedLabel();
 

src/dynarmic/src/dynarmic/backend/x64/emit_x64_memory.h

@@ -348,7 +348,7 @@ void EmitExclusiveLock(BlockOfCode& code, const UserConfig& conf, Xbyak::Reg64 p
     }
 
     code.mov(pointer, std::bit_cast<u64>(GetExclusiveMonitorLockPointer(conf.global_monitor)));
-    EmitSpinLockLock(code, pointer, tmp);
+    EmitSpinLockLock(code, pointer, tmp, code.HasHostFeature(HostFeature::WAITPKG));
 }
 
 template<typename UserConfig>

src/dynarmic/src/dynarmic/backend/x64/host_feature.h

@@ -35,9 +35,10 @@ enum class HostFeature : u64 {
     BMI2 = 1ULL << 19,
     LZCNT = 1ULL << 20,
     GFNI = 1ULL << 21,
+    WAITPKG = 1ULL << 22,
 
     // Zen-based BMI2
-    FastBMI2 = 1ULL << 22,
+    FastBMI2 = 1ULL << 23,
 
     // Orthographic AVX512 features on 128 and 256 vectors
     AVX512_Ortho = AVX512F | AVX512VL,

src/dynarmic/src/dynarmic/common/spin_lock_x64.cpp

@@ -22,17 +22,46 @@ static const auto default_cg_mode = nullptr; //Allow RWE
 
 namespace Dynarmic {
 
-void EmitSpinLockLock(Xbyak::CodeGenerator& code, Xbyak::Reg64 ptr, Xbyak::Reg32 tmp) {
+void EmitSpinLockLock(Xbyak::CodeGenerator& code, Xbyak::Reg64 ptr, Xbyak::Reg32 tmp, bool waitpkg) {
+    // TODO: this is because we lack regalloc - so better to be safe :(
+    if (waitpkg) {
+        code.push(Xbyak::util::eax);
+        code.push(Xbyak::util::ebx);
+        code.push(Xbyak::util::edx);
+    }
     Xbyak::Label start, loop;
-
     code.jmp(start, code.T_NEAR);
     code.L(loop);
-    code.pause();
+    if (waitpkg) {
+        // TODO: This clobbers EAX and EDX did we tell the regalloc?
+        // ARM ptr for address-monitoring
+        code.umonitor(ptr);
+        // tmp.bit[0] = 0: C0.1 | Slow Wakup | Better Savings
+        // tmp.bit[0] = 1: C0.2 | Fast Wakup | Lesser Savings
+        // edx:eax is implicitly used as a 64-bit deadline timestamp
+        // Use the maximum so that we use the operating system's maximum
+        // allowed wait time within the IA32_UMWAIT_CONTROL register
+        // Enter power state designated by tmp and wait for a write to lock_ptr
+        code.mov(Xbyak::util::eax, 0xFFFFFFFF);
+        code.mov(Xbyak::util::edx, Xbyak::util::eax);
+        // TODO: We can only be here because tmp is 1 already - however we repeatedly overwrite it...
+        code.mov(Xbyak::util::ebx, 1);
+        code.umwait(Xbyak::util::ebx);
+        // CF == 1 if we hit the OS-timeout in IA32_UMWAIT_CONTROL without a write
+        // CF == 0 if we exited the wait for any other reason
+    } else {
+        code.pause();
+    }
     code.L(start);
     code.mov(tmp, 1);
     /*code.lock();*/ code.xchg(code.dword[ptr], tmp);
     code.test(tmp, tmp);
     code.jnz(loop, code.T_NEAR);
+    if (waitpkg) {
+        code.pop(Xbyak::util::edx);
+        code.pop(Xbyak::util::ebx);
+        code.pop(Xbyak::util::eax);
+    }
 }
 
 void EmitSpinLockUnlock(Xbyak::CodeGenerator& code, Xbyak::Reg64 ptr, Xbyak::Reg32 tmp) {
@@ -60,7 +89,7 @@ void SpinLockImpl::Initialize() noexcept {
     Xbyak::Reg64 const ABI_PARAM1 = Backend::X64::HostLocToReg64(Backend::X64::ABI_PARAM1);
     code.align();
     lock = code.getCurr<void (*)(volatile int*)>();
-    EmitSpinLockLock(code, ABI_PARAM1, code.eax);
+    EmitSpinLockLock(code, ABI_PARAM1, code.eax, false);
     code.ret();
     code.align();
     unlock = code.getCurr<void (*)(volatile int*)>();

src/dynarmic/src/dynarmic/common/spin_lock_x64.h

@@ -1,3 +1,6 @@
+// SPDX-FileCopyrightText: Copyright 2025 Eden Emulator Project
+// SPDX-License-Identifier: GPL-3.0-or-later
+
 /* This file is part of the dynarmic project.
  * Copyright (c) 2022 MerryMage
  * SPDX-License-Identifier: 0BSD
@@ -9,7 +12,7 @@
 
 namespace Dynarmic {
 
-void EmitSpinLockLock(Xbyak::CodeGenerator& code, Xbyak::Reg64 ptr, Xbyak::Reg32 tmp);
+void EmitSpinLockLock(Xbyak::CodeGenerator& code, Xbyak::Reg64 ptr, Xbyak::Reg32 tmp, bool waitpkg);
 void EmitSpinLockUnlock(Xbyak::CodeGenerator& code, Xbyak::Reg64 ptr, Xbyak::Reg32 tmp);
 
 }  // namespace Dynarmic