Merge pull request #6382 from unknownbrackets/armjit-minor

Fix vmin/vmax to follow the PSP's rules

Core/MIPS/ARM/ArmCompVFPU.cpp

@@ -742,25 +742,51 @@ namespace MIPSComp
 			case 27: //VFPU3
 				switch ((op >> 23) & 7)	{
 				case 2:  // vmin
+				{
 					VCMP(fpr.V(sregs[i]), fpr.V(tregs[i]));
 					VMRS_APSR();
-					// TODO: Technically should use NaN sign bit.
+					FixupBranch skipNAN = B_CC(CC_VC);
+					VMOV(SCRATCHREG1, fpr.V(sregs[i]));
+					VMOV(SCRATCHREG2, fpr.V(tregs[i]));
+					// If both are negative, we reverse the comparison.  We want the highest mantissa then.
+					// Also, between -NAN and -5.0, we want -NAN to be less.
+					TST(SCRATCHREG1, SCRATCHREG2);
+					SetCC(CC_MI);
+					CMP(SCRATCHREG2, SCRATCHREG1);
+					SetCC(CC_PL);
+					CMP(SCRATCHREG1, SCRATCHREG2);
+					SetCC(CC_AL);
+					SetJumpTarget(skipNAN);
 					SetCC(CC_LT);
 					VMOV(fpr.V(tempregs[i]), fpr.V(sregs[i]));
 					SetCC(CC_GE);
 					VMOV(fpr.V(tempregs[i]), fpr.V(tregs[i]));
 					SetCC(CC_AL);
 					break;
+				}
 				case 3:  // vmax
+				{
 					VCMP(fpr.V(tregs[i]), fpr.V(sregs[i]));
 					VMRS_APSR();
-					// TODO: Technically should use NaN sign bit.
+					FixupBranch skipNAN = B_CC(CC_VC);
+					VMOV(SCRATCHREG1, fpr.V(sregs[i]));
+					VMOV(SCRATCHREG2, fpr.V(tregs[i]));
+					// If both are negative, we reverse the comparison.  We want the lowest mantissa then.
+					// Also, between -NAN and -5.0, we want -5.0 to be greater.
+					TST(SCRATCHREG2, SCRATCHREG1);
+					SetCC(CC_MI);
+					CMP(SCRATCHREG1, SCRATCHREG2);
+					SetCC(CC_PL);
+					CMP(SCRATCHREG2, SCRATCHREG1);
+					SetCC(CC_AL);
+					SetJumpTarget(skipNAN);
 					SetCC(CC_LT);
 					VMOV(fpr.V(tempregs[i]), fpr.V(sregs[i]));
 					SetCC(CC_GE);
 					VMOV(fpr.V(tempregs[i]), fpr.V(tregs[i]));
 					SetCC(CC_AL);
 					break;
+				}
 				case 6:  // vsge
 					DISABLE;  // pending testing
 					VCMP(fpr.V(tregs[i]), fpr.V(sregs[i]));

Core/MIPS/MIPSInt.cpp

@@ -893,7 +893,18 @@ namespace MIPSInt
 			switch (op & 0x3f)
 			{
 			case 12: FsI(fd) = (int)floorf(F(fs)+0.5f); break; //round.w.s
-			case 13: FsI(fd) = F(fs)>=0 ? (int)floorf(F(fs)) : (int)ceilf(F(fs)); break;//trunc.w.s
+			case 13: //trunc.w.s
+				if (F(fs) >= 0.0f) {
+					FsI(fd) = (int)floorf(F(fs));
+					// Overflow, but it was positive.
+					if (FsI(fd) == -2147483648) {
+						FsI(fd) = 2147483647;
+					}
+				} else {
+					// Overflow happens to be the right value anyway.
+					FsI(fd) = (int)ceilf(F(fs));
+				}
+				break;
 			case 14: FsI(fd) = (int)ceilf (F(fs)); break; //ceil.w.s
 			case 15: FsI(fd) = (int)floorf(F(fs)); break; //floor.w.s
 			}

Core/MIPS/MIPSIntVFPU.cpp

@@ -1530,26 +1530,54 @@ namespace MIPSInt
 		int cond = op&15;
 		VectorSize sz = GetVecSize(op);
 		int numElements = GetNumVectorElements(sz);
-		float s[4];
-		float t[4];
-		float d[4];
-		ReadVector(s, sz, vs);
-		ApplySwizzleS(s, sz);
-		ReadVector(t, sz, vt);
-		ApplySwizzleT(t, sz);
+
+		union FloatBits {
+			float f[4];
+			u32 u[4];
+			int i[4];
+		};
+
+		FloatBits s;
+		FloatBits t;
+		FloatBits d;
+		ReadVector(s.f, sz, vs);
+		ApplySwizzleS(s.f, sz);
+		ReadVector(t.f, sz, vt);
+		ApplySwizzleT(t.f, sz);
 
 		// If both are zero, take t's sign.
-		// TODO: Otherwise: -NAN < -INF < real < INF < NAN
+		// Otherwise: -NAN < -INF < real < INF < NAN (higher mantissa is farther from 0.)
 
 		switch ((op >> 23) & 3) {
 		case 2: // vmin
 			for (int i = 0; i < numElements; i++) {
-				d[i] = my_isnan(t[i]) ? s[i] : (my_isnan(s[i]) ? t[i] : std::min(t[i], s[i]));
+				if (my_isnanorinf(s.f[i]) || my_isnanorinf(t.f[i])) {
+					// If both are negative, we flip the comparison (not two's compliment.)
+					if (s.i[i] < 0 && t.i[i] < 0) {
+						// If at least one side is NAN, we take the highest mantissa bits.
+						d.i[i] = std::max(t.i[i], s.i[i]);
+					} else {
+						// Otherwise, we take the lowest value (negative or lowest mantissa.)
+						d.i[i] = std::min(t.i[i], s.i[i]);
+					}
+				} else {
+					d.f[i] = std::min(t.f[i], s.f[i]);
+				}
 			}
 			break;
 		case 3: // vmax
-			for (int i = 0; i < numElements; i++)
-				d[i] = my_isnan(t[i]) ? t[i] : (my_isnan(s[i]) ? s[i] : std::max(t[i], s[i]));
+			for (int i = 0; i < numElements; i++) {
+				// This is the same logic as vmin, just reversed.
+				if (my_isnanorinf(s.f[i]) || my_isnanorinf(t.f[i])) {
+					if (s.i[i] < 0 && t.i[i] < 0) {
+						d.i[i] = std::min(t.i[i], s.i[i]);
+					} else {
+						d.i[i] = std::max(t.i[i], s.i[i]);
+					}
+				} else {
+					d.f[i] = std::max(t.f[i], s.f[i]);
+				}
+			}
 			break;
 		default:
 			_dbg_assert_msg_(CPU,0,"unknown min/max op %d", cond);
@@ -1557,8 +1585,8 @@ namespace MIPSInt
 			EatPrefixes();
 			return;
 		}
-		ApplyPrefixD(d, sz);
-		WriteVector(d, sz, vd);
+		ApplyPrefixD(d.f, sz);
+		WriteVector(d.f, sz, vd);
 		PC += 4;
 		EatPrefixes();
 	}