From 4728478247b197b89d8ed2ee26730228ae0bf818 Mon Sep 17 00:00:00 2001
From: Nekotekina <nekotekina@gmail.com>
Date: Tue, 18 Jan 2022 00:08:41 +0300
Subject: [PATCH] PPU: improve interpreter codegen

Implement more sensible check for EVEX prefix and SSE
---
 rpcs3/Emu/Cell/PPUInterpreter.cpp | 21 +++++------------
 rpcs3/util/simd.hpp               | 38 +++++++++++++++++++++++++++----
 2 files changed, 40 insertions(+), 19 deletions(-)

diff --git a/rpcs3/Emu/Cell/PPUInterpreter.cpp b/rpcs3/Emu/Cell/PPUInterpreter.cpp
index 80d285f46f..26ad65d79f 100644
--- a/rpcs3/Emu/Cell/PPUInterpreter.cpp
+++ b/rpcs3/Emu/Cell/PPUInterpreter.cpp
@@ -223,14 +223,7 @@ namespace asmjit
 		template <auto MPtr, u32 Size = sizeof(std::declval<ppu_thread&>().*MPtr)>
 		x86::Mem ppu_mem()
 		{
-			if (ppu_base == 0)
-			{
-				return x86::ptr(arg_ppu, static_cast<s32>(::offset32(MPtr)), Size);
-			}
-			else
-			{
-				return x86::ptr(x86::r10, static_cast<s32>(::offset32(MPtr) - ppu_base), Size);
-			}
+			return x86::ptr(arg_ppu, static_cast<s32>(::offset32(MPtr)), Size);
 		}
 
 		template <u32 Size = 16, uint I, uint N>
@@ -316,22 +309,20 @@ struct ppu_abstract_t
 		}
 	} vr;
 
-	struct abstract_sat : asmjit::mem_lazy
+	struct abstract_sat : asmjit::mem_type
 	{
-		const asmjit::Operand& eval(bool)
-		{
+		abstract_sat()
 		#if defined(ARCH_X64)
-			Operand::operator=(static_cast<asmjit::ppu_builder*>(g_vc)->ppu_sat());
+			: asmjit::mem_type(static_cast<asmjit::ppu_builder*>(g_vc)->ppu_sat())
 		#endif
-
-			return *this;
+		{
 		}
 
 		template <typename T>
 		void operator=(T&& _val) const
 		{
 		#if defined(ARCH_X64)
-			FOR_X64(store_op, kIdMovaps, kIdVmovaps, static_cast<asmjit::ppu_builder*>(g_vc)->ppu_sat(), std::forward<T>(_val));
+			FOR_X64(store_op, kIdMovaps, kIdVmovaps, *this, std::forward<T>(_val));
 		#endif
 		}
 	} sat{};
diff --git a/rpcs3/util/simd.hpp b/rpcs3/util/simd.hpp
index a6095bc050..c3c1b52ff8 100644
--- a/rpcs3/util/simd.hpp
+++ b/rpcs3/util/simd.hpp
@@ -196,6 +196,11 @@ namespace asmjit
 		return mem.eval(std::is_reference_v<T>);
 	}
 
+	inline decltype(auto) arg_eval(const Operand& mem, u32)
+	{
+		return mem;
+	}
+
 	inline decltype(auto) arg_eval(Operand& mem, u32)
 	{
 		return mem;
@@ -206,6 +211,31 @@ namespace asmjit
 		return std::move(mem);
 	}
 
+	template <typename T>
+	inline bool arg_use_evex(const auto& op)
+	{
+		constexpr auto _class = arg_classify<T>;
+		if constexpr (_class == arg_class::imm_rv)
+			return true;
+		else if constexpr (_class == arg_class::imm_lv)
+			return false;
+		else if (op.isMem())
+		{
+			// Check if broadcast is set, or if the offset immediate can use disp8*N encoding
+			mem_type mem{};
+			mem.copyFrom(op);
+			if (mem.hasBaseLabel())
+				return false;
+			if (mem.hasBroadcast())
+				return true;
+			if (!mem.hasOffset() || mem.offset() % mem.size() || u64(mem.offset() + 128) < 256 || u64(mem.offset() / mem.size() + 128) >= 256)
+				return false;
+			return true;
+		}
+
+		return false;
+	}
+
 	template <typename A, typename... Args>
 	vec_type unary_op(x86::Inst::Id op, x86::Inst::Id op2, A&& a, Args&&... args)
 	{
@@ -259,7 +289,7 @@ namespace asmjit
 
 		if (utils::has_avx512() && evex_op)
 		{
-			if (!dst.hasBaseLabel() && dst.hasOffset() && dst.offset() % dst.size() == 0 && dst.offset() / dst.size() + 128 < 256)
+			if (!dst.hasBaseLabel() && dst.hasOffset() && dst.offset() % dst.size() == 0 && u64(dst.offset() + 128) >= 256 && u64(dst.offset() / dst.size() + 128) < 256)
 			{
 				ensure(!g_vc->evex().emit(evex_op, dst, arg_eval(std::forward<S>(s), 16)));
 				return;
@@ -279,7 +309,7 @@ namespace asmjit
 			// Use src1 as a destination
 			src1 = arg_eval(std::forward<A>(a), 16);
 
-			if (utils::has_avx512() && evex_op && (arg_classify<B> == arg_class::imm_rv || arg_classify<B> == arg_class::mem_rv || b.isMem()))
+			if (utils::has_avx512() && evex_op && arg_use_evex<B>(b))
 			{
 				ensure(!g_vc->evex().emit(evex_op, src1, src1, arg_eval(std::forward<B>(b), esize), std::forward<Args>(args)...));
 				return vec_type{src1.id()};
@@ -322,7 +352,7 @@ namespace asmjit
 				}
 			}
 
-			if (utils::has_avx512() && evex_op && (arg_classify<B> == arg_class::imm_rv || arg_classify<B> == arg_class::mem_rv || b.isMem()))
+			if (utils::has_avx512() && evex_op && arg_use_evex<B>(b))
 			{
 				ensure(!g_vc->evex().emit(evex_op, src1, vec_type{a.id()}, arg_eval(std::forward<B>(b), esize), std::forward<Args>(args)...));
 				return vec_type{src1.id()};
@@ -352,7 +382,7 @@ namespace asmjit
 		}
 		while (0);
 
-		if (utils::has_avx512() && evex_op && (arg_classify<B> == arg_class::imm_rv || arg_classify<B> == arg_class::mem_rv || b.isMem()))
+		if (utils::has_avx512() && evex_op && arg_use_evex<B>(b))
 		{
 			ensure(!g_vc->evex().emit(evex_op, src1, src1, arg_eval(std::forward<B>(b), esize), std::forward<Args>(args)...));
 		}