From dc8252bb9fd9c3ea3929533a4c50f1947d475ed9 Mon Sep 17 00:00:00 2001
From: Nekotekina <nekotekina@gmail.com>
Date: Tue, 20 Oct 2020 08:22:25 +0300
Subject: [PATCH] Remove XABORT in PPU/SPU transactions.

It's expensive for unknown reason. Simply XEND is usually much cheaper.
Add some minor improvements. Use g_sudo_addr.
---
 Utilities/JIT.cpp            |  37 ++---------
 Utilities/JIT.h              |   5 +-
 rpcs3/Emu/Cell/PPUThread.cpp |  94 +++++++++++++++++++--------
 rpcs3/Emu/Cell/SPUThread.cpp | 121 +++++++++++++++++++++++------------
 rpcs3/util/atomic2.cpp       |   8 ++-
 5 files changed, 161 insertions(+), 104 deletions(-)

diff --git a/Utilities/JIT.cpp b/Utilities/JIT.cpp
index 01e92a7241..0cfb1e16d4 100644
--- a/Utilities/JIT.cpp
+++ b/Utilities/JIT.cpp
@@ -272,33 +272,13 @@ asmjit::Label asmjit::build_transaction_enter(asmjit::X86Assembler& c, asmjit::L
 	Label begin = c.newLabel();
 	c.jmp(begin);
 	c.bind(fall);
+	c.add(ctr, 1);
 
-	if (less_than < 65)
-	{
-		c.add(ctr, 1);
-		c.test(x86::eax, _XABORT_RETRY);
-		c.jz(fallback);
-	}
-	else
-	{
-		// Don't repeat on explicit XABORT instruction (workaround)
-		c.test(x86::eax, _XABORT_EXPLICIT);
-		c.jnz(fallback);
-
-		// Don't repeat on weird zero status
-		c.test(x86::eax, x86::eax);
-		c.jz(fallback);
-
-		// Count an attempt without RETRY flag as 65 normal attempts and continue
-		c.push(x86::rax);
-		c.not_(x86::eax);
-		c.and_(x86::eax, _XABORT_RETRY);
-		c.shl(x86::eax, 5);
-		c.add(x86::eax, 1); // eax = RETRY ? 1 : 65
-		c.add(ctr, x86::rax);
-		c.pop(x86::rax);
-	}
+	// Don't repeat on zero status (may indicate syscall or interrupt)
+	c.test(x86::eax, x86::eax);
+	c.jz(fallback);
 
+	// Other bad statuses are ignored regardless of repeat flag (TODO)
 	c.cmp(ctr, less_than);
 	c.jae(fallback);
 	c.align(kAlignCode, 16);
@@ -309,13 +289,6 @@ asmjit::Label asmjit::build_transaction_enter(asmjit::X86Assembler& c, asmjit::L
 	//c.xbegin(fall);
 }
 
-void asmjit::build_transaction_abort(asmjit::X86Assembler& c, unsigned char code)
-{
-	c.db(0xc6);
-	c.db(0xf8);
-	c.db(code);
-}
-
 #ifdef LLVM_AVAILABLE
 
 #include <unordered_map>
diff --git a/Utilities/JIT.h b/Utilities/JIT.h
index fafcae8d1b..6085e2cf8e 100644
--- a/Utilities/JIT.h
+++ b/Utilities/JIT.h
@@ -55,11 +55,8 @@ namespace asmjit
 	// Should only be used to build global functions
 	asmjit::Runtime& get_global_runtime();
 
-	// Emit xbegin and adjacent loop, return label at xbegin
+	// Emit xbegin and adjacent loop, return label at xbegin (don't use xabort please)
 	[[nodiscard]] asmjit::Label build_transaction_enter(X86Assembler& c, Label fallback, const X86Gp& ctr, uint less_than);
-
-	// Emit xabort
-	void build_transaction_abort(X86Assembler& c, unsigned char code);
 }
 
 // Build runtime function with asmjit::X86Assembler
diff --git a/rpcs3/Emu/Cell/PPUThread.cpp b/rpcs3/Emu/Cell/PPUThread.cpp
index 7945681f51..c4f22b9969 100644
--- a/rpcs3/Emu/Cell/PPUThread.cpp
+++ b/rpcs3/Emu/Cell/PPUThread.cpp
@@ -1340,7 +1340,7 @@ const auto ppu_stcx_accurate_tx = build_function_asm<u32(*)(u32 raddr, u64 rtime
 
 	// Prepare registers
 	c.mov(x86::rbx, imm_ptr(+vm::g_reservations));
-	c.mov(x86::rax, imm_ptr(&vm::g_base_addr));
+	c.mov(x86::rax, imm_ptr(&vm::g_sudo_addr));
 	c.mov(x86::rbp, x86::qword_ptr(x86::rax));
 	c.lea(x86::rbp, x86::qword_ptr(x86::rbp, args[0]));
 	c.and_(x86::rbp, -128);
@@ -1350,8 +1350,9 @@ const auto ppu_stcx_accurate_tx = build_function_asm<u32(*)(u32 raddr, u64 rtime
 	c.shr(args[0].r32(), 1);
 	c.lea(x86::rbx, x86::qword_ptr(x86::rbx, args[0]));
 	c.and_(x86::rbx, -128 / 2);
+	c.prefetchw(x86::byte_ptr(x86::rbx));
 	c.and_(args[0].r32(), 63);
-	c.xor_(x86::r12d, x86::r12d);
+	c.mov(x86::r12d, 1);
 	c.mov(x86::r13, args[1]);
 	c.bswap(args[3]);
 
@@ -1376,7 +1377,7 @@ const auto ppu_stcx_accurate_tx = build_function_asm<u32(*)(u32 raddr, u64 rtime
 	}
 
 	// Begin transaction
-	Label tx0 = build_transaction_enter(c, fall, x86::r12, 4);
+	Label tx0 = build_transaction_enter(c, fall, x86::r12d, 4);
 	c.xbegin(tx0);
 	c.mov(x86::rax, x86::qword_ptr(x86::rbx));
 	c.test(x86::eax, 127);
@@ -1423,22 +1424,46 @@ const auto ppu_stcx_accurate_tx = build_function_asm<u32(*)(u32 raddr, u64 rtime
 	// Update reservation
 	c.sub(x86::qword_ptr(x86::rbx), -128);
 	c.xend();
-	c.mov(x86::eax, 1);
+	c.mov(x86::eax, x86::r12d);
+	c.jmp(_ret);
+
+	// XABORT is expensive so finish with xend instead
+	c.bind(fail);
+
+	// Load old data (unused)
+	if (s_tsx_avx)
+	{
+		c.vmovaps(x86::ymm0, x86::yword_ptr(x86::rbp, 0));
+		c.vmovaps(x86::ymm1, x86::yword_ptr(x86::rbp, 32));
+		c.vmovaps(x86::ymm2, x86::yword_ptr(x86::rbp, 64));
+		c.vmovaps(x86::ymm3, x86::yword_ptr(x86::rbp, 96));
+	}
+	else
+	{
+		c.movaps(x86::xmm0, x86::oword_ptr(x86::rbp, 0));
+		c.movaps(x86::xmm1, x86::oword_ptr(x86::rbp, 16));
+		c.movaps(x86::xmm2, x86::oword_ptr(x86::rbp, 32));
+		c.movaps(x86::xmm3, x86::oword_ptr(x86::rbp, 48));
+		c.movaps(x86::xmm4, x86::oword_ptr(x86::rbp, 64));
+		c.movaps(x86::xmm5, x86::oword_ptr(x86::rbp, 80));
+		c.movaps(x86::xmm6, x86::oword_ptr(x86::rbp, 96));
+		c.movaps(x86::xmm7, x86::oword_ptr(x86::rbp, 112));
+	}
+
+	c.xend();
+	c.xor_(x86::eax, x86::eax);
 	c.jmp(_ret);
 
 	c.bind(skip);
-	c.xor_(x86::eax, x86::eax);
-	c.xor_(x86::r12d, x86::r12d);
-	build_transaction_abort(c, 0);
+	c.xend();
+	c.mov(x86::eax, _XABORT_EXPLICIT);
 	//c.jmp(fall);
 
 	c.bind(fall);
-	c.sar(x86::eax, 24);
-	c.js(fail);
 
-	// Touch memory if transaction failed without RETRY flag on the first attempt
-	c.cmp(x86::r12, 1);
-	c.jne(next);
+	// Touch memory if transaction failed with status 0
+	c.test(x86::eax, x86::eax);
+	c.jnz(next);
 	c.xor_(x86::rbp, 0xf80);
 	c.lock().add(x86::dword_ptr(x86::rbp), 0);
 	c.xor_(x86::rbp, 0xf80);
@@ -1454,19 +1479,19 @@ const auto ppu_stcx_accurate_tx = build_function_asm<u32(*)(u32 raddr, u64 rtime
 	c.mov(x86::eax, 1);
 	c.lock().xadd(x86::qword_ptr(x86::rbx), x86::rax);
 	c.test(x86::eax, vm::rsrv_unique_lock);
-	c.jnz(fail3);
+	c.jnz(fall2);
 
 	// Allow only first shared lock to proceed
 	c.cmp(x86::rax, x86::r13);
 	c.jne(fail2);
 
-	Label tx1 = build_transaction_enter(c, fall2, x86::r12, 666);
+	Label tx1 = build_transaction_enter(c, fall2, x86::r12d, 666);
 	c.prefetchw(x86::byte_ptr(x86::rbp, 0));
 	c.prefetchw(x86::byte_ptr(x86::rbp, 64));
 
 	// Check pause flag
 	c.bt(x86::dword_ptr(args[2], ::offset32(&ppu_thread::state) - ::offset32(&ppu_thread::rdata)), static_cast<u32>(cpu_flag::pause));
-	c.jc(fail3);
+	c.jc(fall2);
 	c.mov(x86::rax, x86::qword_ptr(x86::rbx));
 	c.and_(x86::rax, -128);
 	c.cmp(x86::rax, x86::r13);
@@ -1504,30 +1529,47 @@ const auto ppu_stcx_accurate_tx = build_function_asm<u32(*)(u32 raddr, u64 rtime
 		c.ptest(x86::xmm0, x86::xmm0);
 	}
 
-	c.jnz(fail2);
+	c.jnz(fail3);
 
 	// Store 8 bytes
 	c.mov(x86::qword_ptr(x86::rbp, args[0], 1, 0), args[3]);
 
 	c.xend();
 	c.lock().add(x86::qword_ptr(x86::rbx), 127);
-	c.lea(x86::rax, x86::qword_ptr(x86::r12, 1));
+	c.mov(x86::eax, x86::r12d);
 	c.jmp(_ret);
 
-	c.bind(fall2);
-	c.sar(x86::eax, 24);
-	c.js(fail2);
+	// XABORT is expensive so try to finish with xend instead
 	c.bind(fail3);
+
+	// Load old data (unused)
+	if (s_tsx_avx)
+	{
+		c.vmovaps(x86::ymm0, x86::yword_ptr(x86::rbp, 0));
+		c.vmovaps(x86::ymm1, x86::yword_ptr(x86::rbp, 32));
+		c.vmovaps(x86::ymm2, x86::yword_ptr(x86::rbp, 64));
+		c.vmovaps(x86::ymm3, x86::yword_ptr(x86::rbp, 96));
+	}
+	else
+	{
+		c.movaps(x86::xmm0, x86::oword_ptr(x86::rbp, 0));
+		c.movaps(x86::xmm1, x86::oword_ptr(x86::rbp, 16));
+		c.movaps(x86::xmm2, x86::oword_ptr(x86::rbp, 32));
+		c.movaps(x86::xmm3, x86::oword_ptr(x86::rbp, 48));
+		c.movaps(x86::xmm4, x86::oword_ptr(x86::rbp, 64));
+		c.movaps(x86::xmm5, x86::oword_ptr(x86::rbp, 80));
+		c.movaps(x86::xmm6, x86::oword_ptr(x86::rbp, 96));
+		c.movaps(x86::xmm7, x86::oword_ptr(x86::rbp, 112));
+	}
+
+	c.xend();
+	c.jmp(fail2);
+
+	c.bind(fall2);
 	c.mov(x86::eax, -1);
 	c.jmp(_ret);
 
-	c.bind(fail);
-	build_transaction_abort(c, 0xff);
-	c.xor_(x86::eax, x86::eax);
-	c.jmp(_ret);
-
 	c.bind(fail2);
-	build_transaction_abort(c, 0xff);
 	c.lock().sub(x86::qword_ptr(x86::rbx), 1);
 	c.xor_(x86::eax, x86::eax);
 	//c.jmp(_ret);
diff --git a/rpcs3/Emu/Cell/SPUThread.cpp b/rpcs3/Emu/Cell/SPUThread.cpp
index 347b67d291..af53104b4b 100644
--- a/rpcs3/Emu/Cell/SPUThread.cpp
+++ b/rpcs3/Emu/Cell/SPUThread.cpp
@@ -308,7 +308,7 @@ namespace spu
 	}
 }
 
-const auto spu_putllc_tx = build_function_asm<u32(*)(u32 raddr, u64 rtime, const void* _old, const void* _new)>([](asmjit::X86Assembler& c, auto& args)
+const auto spu_putllc_tx = build_function_asm<u32(*)(u32 raddr, u64 rtime, void* _old, const void* _new)>([](asmjit::X86Assembler& c, auto& args)
 {
 	using namespace asmjit;
 
@@ -352,7 +352,7 @@ const auto spu_putllc_tx = build_function_asm<u32(*)(u32 raddr, u64 rtime, const
 
 	// Prepare registers
 	c.mov(x86::rbx, imm_ptr(+vm::g_reservations));
-	c.mov(x86::rax, imm_ptr(&vm::g_base_addr));
+	c.mov(x86::rax, imm_ptr(&vm::g_sudo_addr));
 	c.mov(x86::rbp, x86::qword_ptr(x86::rax));
 	c.lea(x86::rbp, x86::qword_ptr(x86::rbp, args[0]));
 	c.prefetchw(x86::byte_ptr(x86::rbp, 0));
@@ -360,7 +360,8 @@ const auto spu_putllc_tx = build_function_asm<u32(*)(u32 raddr, u64 rtime, const
 	c.and_(args[0].r32(), 0xff80);
 	c.shr(args[0].r32(), 1);
 	c.lea(x86::rbx, x86::qword_ptr(x86::rbx, args[0]));
-	c.xor_(x86::r12d, x86::r12d);
+	c.prefetchw(x86::byte_ptr(x86::rbx));
+	c.mov(x86::r12d, 1);
 	c.mov(x86::r13, args[1]);
 
 	// Prepare data
@@ -396,7 +397,7 @@ const auto spu_putllc_tx = build_function_asm<u32(*)(u32 raddr, u64 rtime, const
 	}
 
 	// Begin transaction
-	Label tx0 = build_transaction_enter(c, fall, x86::r12, 4);
+	Label tx0 = build_transaction_enter(c, fall, x86::r12d, 4);
 	c.xbegin(tx0);
 	c.mov(x86::rax, x86::qword_ptr(x86::rbx));
 	c.test(x86::eax, 127);
@@ -458,22 +459,46 @@ const auto spu_putllc_tx = build_function_asm<u32(*)(u32 raddr, u64 rtime, const
 
 	c.sub(x86::qword_ptr(x86::rbx), -128);
 	c.xend();
-	c.mov(x86::eax, 1);
+	c.mov(x86::eax, x86::r12d);
+	c.jmp(_ret);
+
+	// XABORT is expensive so finish with xend instead
+	c.bind(fail);
+
+	// Load old data (unused)
+	if (s_tsx_avx)
+	{
+		c.vmovaps(x86::ymm0, x86::yword_ptr(x86::rbp, 0));
+		c.vmovaps(x86::ymm1, x86::yword_ptr(x86::rbp, 32));
+		c.vmovaps(x86::ymm2, x86::yword_ptr(x86::rbp, 64));
+		c.vmovaps(x86::ymm3, x86::yword_ptr(x86::rbp, 96));
+	}
+	else
+	{
+		c.movaps(x86::xmm0, x86::oword_ptr(x86::rbp, 0));
+		c.movaps(x86::xmm1, x86::oword_ptr(x86::rbp, 16));
+		c.movaps(x86::xmm2, x86::oword_ptr(x86::rbp, 32));
+		c.movaps(x86::xmm3, x86::oword_ptr(x86::rbp, 48));
+		c.movaps(x86::xmm4, x86::oword_ptr(x86::rbp, 64));
+		c.movaps(x86::xmm5, x86::oword_ptr(x86::rbp, 80));
+		c.movaps(x86::xmm6, x86::oword_ptr(x86::rbp, 96));
+		c.movaps(x86::xmm7, x86::oword_ptr(x86::rbp, 112));
+	}
+
+	c.xend();
+	c.xor_(x86::eax, x86::eax);
 	c.jmp(_ret);
 
 	c.bind(skip);
-	c.xor_(x86::eax, x86::eax);
-	c.xor_(x86::r12d, x86::r12d);
-	build_transaction_abort(c, 0);
+	c.xend();
+	c.mov(x86::eax, _XABORT_EXPLICIT);
 	//c.jmp(fall);
 
 	c.bind(fall);
-	c.sar(x86::eax, 24);
-	c.js(fail);
 
-	// Touch memory if transaction failed without RETRY flag on the first attempt
-	c.cmp(x86::r12, 1);
-	c.jne(next);
+	// Touch memory if transaction failed with status 0
+	c.test(x86::eax, x86::eax);
+	c.jnz(next);
 	c.xor_(x86::rbp, 0xf80);
 	c.lock().add(x86::dword_ptr(x86::rbp), 0);
 	c.xor_(x86::rbp, 0xf80);
@@ -495,13 +520,13 @@ const auto spu_putllc_tx = build_function_asm<u32(*)(u32 raddr, u64 rtime, const
 	c.cmp(x86::rax, x86::r13);
 	c.jne(fail2);
 
-	Label tx1 = build_transaction_enter(c, fall2, x86::r12, 666);
+	Label tx1 = build_transaction_enter(c, fall2, x86::r12d, 666);
 	c.prefetchw(x86::byte_ptr(x86::rbp, 0));
 	c.prefetchw(x86::byte_ptr(x86::rbp, 64));
 
 	// Check pause flag
 	c.bt(x86::dword_ptr(args[2], ::offset32(&spu_thread::state) - ::offset32(&spu_thread::rdata)), static_cast<u32>(cpu_flag::pause));
-	c.jc(fail3);
+	c.jc(fall2);
 	c.mov(x86::rax, x86::qword_ptr(x86::rbx));
 	c.and_(x86::rax, -128);
 	c.cmp(x86::rax, x86::r13);
@@ -539,7 +564,7 @@ const auto spu_putllc_tx = build_function_asm<u32(*)(u32 raddr, u64 rtime, const
 		c.ptest(x86::xmm0, x86::xmm0);
 	}
 
-	c.jnz(fail2);
+	c.jnz(fail3);
 
 	if (s_tsx_avx)
 	{
@@ -562,23 +587,40 @@ const auto spu_putllc_tx = build_function_asm<u32(*)(u32 raddr, u64 rtime, const
 
 	c.xend();
 	c.lock().add(x86::qword_ptr(x86::rbx), 127);
-	c.lea(x86::rax, x86::qword_ptr(x86::r12, 1));
+	c.mov(x86::eax, x86::r12d);
 	c.jmp(_ret);
 
-	c.bind(fall2);
-	c.sar(x86::eax, 24);
-	c.js(fail2);
+	// XABORT is expensive so try to finish with xend instead
 	c.bind(fail3);
+
+	// Load old data (unused)
+	if (s_tsx_avx)
+	{
+		c.vmovaps(x86::ymm0, x86::yword_ptr(x86::rbp, 0));
+		c.vmovaps(x86::ymm1, x86::yword_ptr(x86::rbp, 32));
+		c.vmovaps(x86::ymm2, x86::yword_ptr(x86::rbp, 64));
+		c.vmovaps(x86::ymm3, x86::yword_ptr(x86::rbp, 96));
+	}
+	else
+	{
+		c.movaps(x86::xmm0, x86::oword_ptr(x86::rbp, 0));
+		c.movaps(x86::xmm1, x86::oword_ptr(x86::rbp, 16));
+		c.movaps(x86::xmm2, x86::oword_ptr(x86::rbp, 32));
+		c.movaps(x86::xmm3, x86::oword_ptr(x86::rbp, 48));
+		c.movaps(x86::xmm4, x86::oword_ptr(x86::rbp, 64));
+		c.movaps(x86::xmm5, x86::oword_ptr(x86::rbp, 80));
+		c.movaps(x86::xmm6, x86::oword_ptr(x86::rbp, 96));
+		c.movaps(x86::xmm7, x86::oword_ptr(x86::rbp, 112));
+	}
+
+	c.xend();
+	c.jmp(fail2);
+
+	c.bind(fall2);
 	c.mov(x86::eax, -1);
 	c.jmp(_ret);
 
-	c.bind(fail);
-	build_transaction_abort(c, 0xff);
-	c.xor_(x86::eax, x86::eax);
-	c.jmp(_ret);
-
 	c.bind(fail2);
-	build_transaction_abort(c, 0xff);
 	c.lock().sub(x86::qword_ptr(x86::rbx), 1);
 	c.xor_(x86::eax, x86::eax);
 	//c.jmp(_ret);
@@ -649,7 +691,7 @@ const auto spu_putlluc_tx = build_function_asm<u32(*)(u32 raddr, const void* rda
 
 	// Prepare registers
 	c.mov(x86::rbx, imm_ptr(+vm::g_reservations));
-	c.mov(x86::rax, imm_ptr(&vm::g_base_addr));
+	c.mov(x86::rax, imm_ptr(&vm::g_sudo_addr));
 	c.mov(x86::rbp, x86::qword_ptr(x86::rax));
 	c.lea(x86::rbp, x86::qword_ptr(x86::rbp, args[0]));
 	c.prefetchw(x86::byte_ptr(x86::rbp, 0));
@@ -657,7 +699,8 @@ const auto spu_putlluc_tx = build_function_asm<u32(*)(u32 raddr, const void* rda
 	c.and_(args[0].r32(), 0xff80);
 	c.shr(args[0].r32(), 1);
 	c.lea(x86::rbx, x86::qword_ptr(x86::rbx, args[0]));
-	c.xor_(x86::r12d, x86::r12d);
+	c.prefetchw(x86::byte_ptr(x86::rbx));
+	c.mov(x86::r12d, 1);
 	c.mov(x86::r13, args[1]);
 
 	// Prepare data
@@ -681,7 +724,7 @@ const auto spu_putlluc_tx = build_function_asm<u32(*)(u32 raddr, const void* rda
 	}
 
 	// Begin transaction
-	Label tx0 = build_transaction_enter(c, fall, x86::r12, 8);
+	Label tx0 = build_transaction_enter(c, fall, x86::r12d, 8);
 	c.xbegin(tx0);
 	c.test(x86::qword_ptr(x86::rbx), vm::rsrv_unique_lock);
 	c.jnz(skip);
@@ -711,16 +754,14 @@ const auto spu_putlluc_tx = build_function_asm<u32(*)(u32 raddr, const void* rda
 	c.jmp(_ret);
 
 	c.bind(skip);
-	c.xor_(x86::eax, x86::eax);
-	c.xor_(x86::r12d, x86::r12d);
-	build_transaction_abort(c, 0);
+	c.xend();
 	//c.jmp(fall);
 
 	c.bind(fall);
 
-	// Touch memory if transaction failed without RETRY flag on the first attempt
-	c.cmp(x86::r12, 1);
-	c.jne(next);
+	// Touch memory if transaction failed with status 0
+	c.test(x86::eax, x86::eax);
+	c.jnz(next);
 	c.xor_(x86::rbp, 0xf80);
 	c.lock().add(x86::dword_ptr(x86::rbp), 0);
 	c.xor_(x86::rbp, 0xf80);
@@ -736,7 +777,7 @@ const auto spu_putlluc_tx = build_function_asm<u32(*)(u32 raddr, const void* rda
 	c.test(x86::eax, vm::rsrv_unique_lock);
 	c.jnz(fall2);
 
-	Label tx1 = build_transaction_enter(c, fall2, x86::r12, 666);
+	Label tx1 = build_transaction_enter(c, fall2, x86::r12d, 666);
 	c.prefetchw(x86::byte_ptr(x86::rbp, 0));
 	c.prefetchw(x86::byte_ptr(x86::rbp, 64));
 
@@ -766,7 +807,7 @@ const auto spu_putlluc_tx = build_function_asm<u32(*)(u32 raddr, const void* rda
 
 	c.xend();
 	c.lock().add(x86::qword_ptr(x86::rbx), 127);
-	c.lea(x86::rax, x86::qword_ptr(x86::r12, 1));
+	c.mov(x86::eax, x86::r12d);
 	c.jmp(_ret);
 
 	c.bind(fall2);
@@ -824,17 +865,17 @@ const extern auto spu_getllar_tx = build_function_asm<u32(*)(u32 raddr, void* rd
 
 	// Prepare registers
 	c.mov(x86::rbx, imm_ptr(+vm::g_reservations));
-	c.mov(x86::rax, imm_ptr(&vm::g_base_addr));
+	c.mov(x86::rax, imm_ptr(&vm::g_sudo_addr));
 	c.mov(x86::rbp, x86::qword_ptr(x86::rax));
 	c.lea(x86::rbp, x86::qword_ptr(x86::rbp, args[0]));
 	c.and_(args[0].r32(), 0xff80);
 	c.shr(args[0].r32(), 1);
 	c.lea(x86::rbx, x86::qword_ptr(x86::rbx, args[0]));
-	c.xor_(x86::r12d, x86::r12d);
+	c.mov(x86::r12d, 1);
 	c.mov(x86::r13, args[1]);
 
 	// Begin transaction
-	Label tx0 = build_transaction_enter(c, fall, x86::r12, 8);
+	Label tx0 = build_transaction_enter(c, fall, x86::r12d, 8);
 
 	// Check pause flag
 	c.bt(x86::dword_ptr(args[2], ::offset32(&cpu_thread::state)), static_cast<u32>(cpu_flag::pause));
diff --git a/rpcs3/util/atomic2.cpp b/rpcs3/util/atomic2.cpp
index fdec8df1df..7702079a2e 100644
--- a/rpcs3/util/atomic2.cpp
+++ b/rpcs3/util/atomic2.cpp
@@ -211,11 +211,15 @@ static const auto commit_tx = build_function_asm<s32(*)(const stx::multi_cas_ite
 
 	// Transaction abort
 	c.bind(stop);
-	build_transaction_abort(c, 0xff);
+	c.xend();
+	c.xor_(x86::eax, x86::eax);
+	c.jmp(fall);
 
 	// Abort when there is still a chance of success
 	c.bind(wait);
-	build_transaction_abort(c, 0x00);
+	c.xend();
+	c.mov(x86::eax, 0xffu << 24);
+	c.jmp(fall);
 
 	// Transaction fallback: return zero
 	c.bind(fall);