From 57876cbe5ab13ff2f5939ab4ba096f9f946f91cd Mon Sep 17 00:00:00 2001
From: TellowKrinkle <tellowkrinkle@gmail.com>
Date: Mon, 17 Feb 2025 23:26:39 -0600
Subject: [PATCH] microVU: Accurate CLIP

---
 pcsx2/x86/microVU_Misc.h    |  1 +
 pcsx2/x86/microVU_Upper.inl | 36 ++++++++++++++++++------------------
 2 files changed, 19 insertions(+), 18 deletions(-)

diff --git a/pcsx2/x86/microVU_Misc.h b/pcsx2/x86/microVU_Misc.h
index c51cb22ba8..101feffd0b 100644
--- a/pcsx2/x86/microVU_Misc.h
+++ b/pcsx2/x86/microVU_Misc.h
@@ -21,6 +21,7 @@ struct mVU_Globals
 	u32   signbit [4] = __four(0x80000000);
 	u32   minvals [4] = __four(0xff7fffff);
 	u32   maxvals [4] = __four(0x7f7fffff);
+	u32   exponent[4] = __four(0x7f800000);
 	u32   one     [4] = __four(0x3f800000);
 	u32   Pi4     [4] = __four(0x3f490fdb);
 	u32   T1      [4] = __four(0x3f7ffff5);
diff --git a/pcsx2/x86/microVU_Upper.inl b/pcsx2/x86/microVU_Upper.inl
index f2d1573d09..3c1f426313 100644
--- a/pcsx2/x86/microVU_Upper.inl
+++ b/pcsx2/x86/microVU_Upper.inl
@@ -544,36 +544,36 @@ mVUop(mVU_CLIP)
 		const xmm& Fs = mVU.regAlloc->allocReg(_Fs_, 0, 0xf);
 		const xmm& Ft = mVU.regAlloc->allocReg(_Ft_, 0, 0x1);
 		const xmm& t1 = mVU.regAlloc->allocReg();
+		const xmm& t2 = mVU.regAlloc->allocReg();
 
 		mVUunpack_xyzw(Ft, Ft, 0);
 		mVUallocCFLAGa(mVU, gprT1, cFLAG.lastWrite);
 		xSHL(gprT1, 6);
 
-		xAND.PS(Ft, ptr128[mVUglob.absclip]);
-		xMOVAPS(t1, Ft);
-		xPOR(t1, ptr128[mVUglob.signbit]);
+		xMOVAPS  (t1, ptr128[mVUglob.exponent]);
+		xPAND    (t1, Fs);
+		xPXOR    (t2, t2);
+		xPCMP.EQD(t1, t2); // Denormal check
+		xPANDN   (t1, Fs); // If denormal, set to zero, which can't be greater than any nonnegative denormal in Ft
+		xPAND    (Ft, ptr128[mVUglob.absclip]);
 
-		xCMPNLE.PS(t1, Fs); // -w, -z, -y, -x
-		xCMPLT.PS(Ft, Fs);  // +w, +z, +y, +x
+		xMOVAPS  (Fs, ptr128[mVUglob.signbit]);
+		xPXOR    (Fs, t1); // Negate
+		xPCMP.GTD(t1, Ft); // +w, +z, +y, +x
+		xPCMP.GTD(Fs, Ft); // -w, -z, -y, -x
 
-		xMOVAPS(Fs, Ft);    // Fs = +w, +z, +y, +x
-		xUNPCK.LPS(Ft, t1); // Ft = -y,+y,-x,+x
-		xUNPCK.HPS(Fs, t1); // Fs = -w,+w,-z,+z
-
-		xMOVMSKPS(gprT2, Fs); // -w,+w,-z,+z
-		xAND(gprT2, 0x3);
-		xSHL(gprT2, 4);
-		xOR(gprT1, gprT2);
-
-		xMOVMSKPS(gprT2, Ft); // -y,+y,-x,+x
-		xAND(gprT2, 0xf);
-		xOR(gprT1, gprT2);
-		xAND(gprT1, 0xffffff);
+		xPBLEND.W (Fs, t1, 0x55); // Squish together
+		xPACK.SSWB(Fs, Fs);       // Convert u16 to u8
+		xPMOVMSKB (gprT2, Fs);    // Get bitmask
+		xAND      (gprT2, 0x3f);  // Mask unused stuff
+		xAND      (gprT1, 0xffffff);
+		xOR       (gprT1, gprT2);
 
 		mVUallocCFLAGb(mVU, gprT1, cFLAG.write);
 		mVU.regAlloc->clearNeeded(Fs);
 		mVU.regAlloc->clearNeeded(Ft);
 		mVU.regAlloc->clearNeeded(t1);
+		mVU.regAlloc->clearNeeded(t2);
 		mVU.profiler.EmitOp(opCLIP);
 	}
 	pass3