From e429f8cf05926852ed63949cad652bf1de5869a4 Mon Sep 17 00:00:00 2001
From: Walter <walter.kruger@hotmail.com>
Date: Tue, 16 Jun 2026 17:51:14 +1000
Subject: [PATCH] SPU LLVM + zero with saturation subtract

- Added a LLVM SPU recompiler implementation as LLVM can't take advantage of the rounding mode.
- Further optimized the method by replacing MSB check with saturation subtraction. Negative values are saturated to zero due to the float's sign bit.
---
 rpcs3/Emu/Cell/SPUASMJITRecompiler.cpp | 10 ++++------
 rpcs3/Emu/Cell/SPULLVMRecompiler.cpp   | 23 +++++++++++++++++++++++
 2 files changed, 27 insertions(+), 6 deletions(-)

diff --git a/rpcs3/Emu/Cell/SPUASMJITRecompiler.cpp b/rpcs3/Emu/Cell/SPUASMJITRecompiler.cpp
index 14cd7314dc..89694affd4 100644
--- a/rpcs3/Emu/Cell/SPUASMJITRecompiler.cpp
+++ b/rpcs3/Emu/Cell/SPUASMJITRecompiler.cpp
@@ -3354,8 +3354,8 @@ void spu_recompiler::CLZ(spu_opcode_t op)
 	}
 
 	// Use signed conversion to float, as exponent is ilog2
-	// Fixup "negative" cases by overwriting with zero
-	const u32 exp_bias = 127;
+	// "Negative" values are zeroed due to saturation subtract
+	constexpr u32 exp_bias = 127;
 
 	const XmmLink& vf = XmmAlloc();
 	const XmmLink& v1 = XmmAlloc();
@@ -3365,10 +3365,8 @@ void spu_recompiler::CLZ(spu_opcode_t op)
 	c->pcmpeqd(v1, va);
 	c->pand(v1, XmmConst(v128::from32p(32 ^ (31 + exp_bias))));
 	c->pxor(v1, XmmConst(v128::from32p(31 + exp_bias)));
-	c->psubd(v1, vf);	// (x==0)? 32 : 31 - (exponent - exp_bias)
-	c->psrad(va, 31);
-	c->pandn(va, v1);
-	c->movdqa(SPU_OFF_128(gpr, op.rt), va);
+	c->psubusw(v1, vf);	// (x==0)? 32 : 31 - (exponent - exp_bias)
+	c->movdqa(SPU_OFF_128(gpr, op.rt), v1);
 	return;
 }
 
diff --git a/rpcs3/Emu/Cell/SPULLVMRecompiler.cpp b/rpcs3/Emu/Cell/SPULLVMRecompiler.cpp
index 40e473d365..80d1100754 100644
--- a/rpcs3/Emu/Cell/SPULLVMRecompiler.cpp
+++ b/rpcs3/Emu/Cell/SPULLVMRecompiler.cpp
@@ -6329,7 +6329,30 @@ public:
 
 	void CLZ(spu_opcode_t op)
 	{
+#ifdef ARCH_ARM64
 		set_vr(op.rt, ctlz(get_vr(op.ra)));
+#else
+		if (m_use_avx512)
+		{
+			set_vr(op.rt, ctlz(get_vr(op.ra)));
+			return;
+		}
+
+		// Implement manually since LLVM can't take advantage of round-towards-zero.
+		// Helpful as when converting to a float the exponent is always floor(ilog2)
+
+		constexpr u32 exp_bias = 127;
+		value_t<f32[4]> flt;
+
+		const auto a = get_vr(op.ra);
+		flt.value = m_ir->CreateSIToFP(a.value, get_type<f32[4]>()); // only correct with round-towards-zero!
+		const auto exp = bitcast<u32[4]>(flt) >> 23;
+
+		// "Negative" values cause saturation due to float's sign bit
+		const auto offset = select(a == 0, splat<u32[4]>(32), splat<u32[4]>(exp_bias + 31));
+		const auto lzcnt = sub_sat(bitcast<u16[8]>(offset), bitcast<u16[8]>(exp));
+		set_vr(op.rt, bitcast<u32[4]>(lzcnt));
+#endif
 	}
 
 	void XSWD(spu_opcode_t op)