From f9ab077908b001995a271761a7a4ce4cf2047d30 Mon Sep 17 00:00:00 2001
From: Malcolm Jestadt <MalcolmJestadt@gmail.com>
Date: Sun, 26 Sep 2021 04:40:47 -0400
Subject: [PATCH] SPU LLVM: Use VDBPSADBW in SUMB - This instruction can be
 used to sum bytes horrizontally if the second input vector is all zeroes.

---
 rpcs3/Emu/CPU/CPUTranslator.h    | 15 +++++++++++++++
 rpcs3/Emu/Cell/SPURecompiler.cpp | 17 +++++++++++++++++
 2 files changed, 32 insertions(+)

diff --git a/rpcs3/Emu/CPU/CPUTranslator.h b/rpcs3/Emu/CPU/CPUTranslator.h
index 52d3f690d0..d25783989a 100644
--- a/rpcs3/Emu/CPU/CPUTranslator.h
+++ b/rpcs3/Emu/CPU/CPUTranslator.h
@@ -3366,6 +3366,21 @@ public:
 		return result;
 	}
 
+	template <typename T1, typename T2>
+	value_t<u8[16]> vdbpsadbw(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<u32>{c});
+		const auto imm8 = immediate.eval(m_ir);
+
+		result.value = m_ir->CreateCall(get_intrinsic(llvm::Intrinsic::x86_avx512_dbpsadbw_128), {data0, data1, imm8});
+		return result;
+	}
+
 	template <typename T1, typename T2>
 	value_t<u8[16]> vpermb(T1 a, T2 b)
 	{
diff --git a/rpcs3/Emu/Cell/SPURecompiler.cpp b/rpcs3/Emu/Cell/SPURecompiler.cpp
index 000e61c0fd..cc34903816 100644
--- a/rpcs3/Emu/Cell/SPURecompiler.cpp
+++ b/rpcs3/Emu/Cell/SPURecompiler.cpp
@@ -7022,6 +7022,23 @@ public:
 
 	void SUMB(spu_opcode_t op)
 	{
+		if (m_use_avx512)
+		{
+			const auto [a, b] = get_vrs<u8[16]>(op.ra, op.rb);
+			const auto zeroes = splat<u8[16]>(0);
+
+			if (op.ra == op.rb && !m_interp_magn)
+			{
+				set_vr(op.rt, vdbpsadbw(a, zeroes, 0));
+				return;
+			}
+
+			const auto ax = vdbpsadbw(a, zeroes, 0);
+			const auto bx = vdbpsadbw(b, zeroes, 0);
+			set_vr(op.rt, shuffle2(ax, bx, 0, 8, 2, 10, 4, 12, 6, 14));
+			return;
+		}
+
 		if (m_use_vnni)
 		{
 			const auto [a, b] = get_vrs<u32[4]>(op.ra, op.rb);