From 9e4f43f4d1d92e4ef860b13ca7167a6a4a9f5a55 Mon Sep 17 00:00:00 2001
From: Whatcookie <MalcolmJestadt@gmail.com>
Date: Thu, 13 Aug 2020 10:00:56 -0400
Subject: [PATCH] SPU LLVM: Add icelake optimized paths for SHUFB (#8712)

---
 rpcs3/Emu/CPU/CPUTranslator.h    | 15 +++++++++++++++
 rpcs3/Emu/Cell/SPURecompiler.cpp | 24 ++++++++++++++++++++++--
 2 files changed, 37 insertions(+), 2 deletions(-)

diff --git a/rpcs3/Emu/CPU/CPUTranslator.h b/rpcs3/Emu/CPU/CPUTranslator.h
index f9f10610d4..60b273eedb 100644
--- a/rpcs3/Emu/CPU/CPUTranslator.h
+++ b/rpcs3/Emu/CPU/CPUTranslator.h
@@ -2785,6 +2785,21 @@ public:
 		return result;
 	}
 
+	template <typename T1, typename T2>
+	value_t<u8[16]> gf2p8affineqb(T1 a, T2 b, u8 c)
+	{
+		value_t<u8[16]> result;
+
+		const auto data0 = a.eval(m_ir);
+		const auto data1 = b.eval(m_ir);
+
+		const auto immediate = (llvm_const_int<u8>{c});
+		const auto imm8 = immediate.eval(m_ir);
+
+		result.value = m_ir->CreateCall(get_intrinsic(llvm::Intrinsic::x86_vgf2p8affineqb_128), {data0, data1, imm8});
+		return result;
+	}
+
 	template <typename T1, typename T2, typename T3>
 	value_t<u8[16]> vperm2b(T1 a, T2 b, T3 c)
 	{
diff --git a/rpcs3/Emu/Cell/SPURecompiler.cpp b/rpcs3/Emu/Cell/SPURecompiler.cpp
index aed86e32c6..baa7a3eb51 100644
--- a/rpcs3/Emu/Cell/SPURecompiler.cpp
+++ b/rpcs3/Emu/Cell/SPURecompiler.cpp
@@ -7277,10 +7277,20 @@ public:
 		{
 			if (auto [ok, v1] = match_expr(b, byteswap(match<u8[16]>())); ok)
 			{
-				// Undo endian swapping, and rely on pshufb to re-reverse endianness
-				const auto x = avg(noncast<u8[16]>(sext<s8[16]>((c & 0xc0) == 0xc0)), noncast<u8[16]>(sext<s8[16]>((c & 0xe0) == 0xc0)));
+				// Undo endian swapping, and rely on pshufb/vperm2b to re-reverse endianness
 				const auto as = byteswap(a);
 				const auto bs = byteswap(b);
+
+				if (m_use_avx512_icl && (op.ra != op.rb || m_interp_magn))
+				{
+					const auto m = gf2p8affineqb(build<u8[16]>(0x02, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x02, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04), c, 0x7f);
+					const auto mm = select(noncast<s8[16]>(m) >= 0, splat<u8[16]>(0), m);
+					const auto ab = vperm2b(as, bs, c);
+					set_vr(op.rt4, select(noncast<s8[16]>(c) >= 0, ab, mm));
+					return;
+				}
+
+				const auto x = avg(noncast<u8[16]>(sext<s8[16]>((c & 0xc0) == 0xc0)), noncast<u8[16]>(sext<s8[16]>((c & 0xe0) == 0xc0)));
 				const auto ax = pshufb(as, c);
 				const auto bx = pshufb(bs, c);
 				set_vr(op.rt4, select(noncast<s8[16]>(c << 3) >= 0, ax, bx) | x);
@@ -7319,6 +7329,16 @@ public:
 			}
 		}
 
+		if (m_use_avx512_icl && (op.ra != op.rb || m_interp_magn))
+		{
+			const auto m = gf2p8affineqb(build<u8[16]>(0x02, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x02, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04), c, 0x7f);
+			const auto mm = select(noncast<s8[16]>(m) >= 0, splat<u8[16]>(0), m);
+			const auto cr = eval(~c);
+			const auto ab = vperm2b(b, a, cr);
+			set_vr(op.rt4, select(noncast<s8[16]>(cr) >= 0, mm, ab));
+			return;
+		}
+
 		const auto x = avg(noncast<u8[16]>(sext<s8[16]>((c & 0xc0) == 0xc0)), noncast<u8[16]>(sext<s8[16]>((c & 0xe0) == 0xc0)));
 		const auto cr = eval(c ^ 0xf);
 		const auto ax = pshufb(a, cr);