MIPS: math-emu: Add support for the MIPS R6 MADDF FPU instruction

MIPS R6 introduced the following instruction:
Floating Point Fused Multiply Add:
MADDF.fmt To perform a fused multiply-add of FP values.

MADDF.fmt: FPR[fd] = FPR[fd] + (FPR[fs] x FPR[ft])

Signed-off-by: Markos Chandras <markos.chandras@imgtec.com>
Cc: linux-mips@linux-mips.org
Patchwork: https://patchwork.linux-mips.org/patch/10956/
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>

1 files changed, 255 insertions, 0 deletions

diff --git a/arch/mips/math-emu/sp_maddf.c b/arch/mips/math-emu/sp_maddf.c
new file mode 100644
index 000000000000..dd1dd83e34eb
--- /dev/null
+++ b/arch/mips/math-emu/sp_maddf.c
@@ -0,0 +1,255 @@
+/*
+ * IEEE754 floating point arithmetic
+ * single precision: MADDF.f (Fused Multiply Add)
+ * MADDF.fmt: FPR[fd] = FPR[fd] + (FPR[fs] x FPR[ft])
+ *
+ * MIPS floating point support
+ * Copyright (C) 2015 Imagination Technologies, Ltd.
+ * Author: Markos Chandras <markos.chandras@imgtec.com>
+ *
+ *  This program is free software; you can distribute it and/or modify it
+ *  under the terms of the GNU General Public License as published by the
+ *  Free Software Foundation; version 2 of the License.
+ */
+
+#include "ieee754sp.h"
+
+union ieee754sp ieee754sp_maddf(union ieee754sp z, union ieee754sp x,
+				union ieee754sp y)
+{
+	int re;
+	int rs;
+	unsigned rm;
+	unsigned short lxm;
+	unsigned short hxm;
+	unsigned short lym;
+	unsigned short hym;
+	unsigned lrm;
+	unsigned hrm;
+	unsigned t;
+	unsigned at;
+	int s;
+
+	COMPXSP;
+	COMPYSP;
+	u32 zm; int ze; int zs __maybe_unused; int zc;
+
+	EXPLODEXSP;
+	EXPLODEYSP;
+	EXPLODESP(z, zc, zs, ze, zm)
+
+	FLUSHXSP;
+	FLUSHYSP;
+	FLUSHSP(z, zc, zs, ze, zm);
+
+	ieee754_clearcx();
+
+	switch (zc) {
+	case IEEE754_CLASS_SNAN:
+		ieee754_setcx(IEEE754_INVALID_OPERATION);
+		return ieee754sp_nanxcpt(z);
+	case IEEE754_CLASS_DNORM:
+		SPDNORMx(zm, ze);
+	/* QNAN is handled separately below */
+	}
+
+	switch (CLPAIR(xc, yc)) {
+	case CLPAIR(IEEE754_CLASS_QNAN, IEEE754_CLASS_SNAN):
+	case CLPAIR(IEEE754_CLASS_ZERO, IEEE754_CLASS_SNAN):
+	case CLPAIR(IEEE754_CLASS_NORM, IEEE754_CLASS_SNAN):
+	case CLPAIR(IEEE754_CLASS_DNORM, IEEE754_CLASS_SNAN):
+	case CLPAIR(IEEE754_CLASS_INF, IEEE754_CLASS_SNAN):
+		return ieee754sp_nanxcpt(y);
+
+	case CLPAIR(IEEE754_CLASS_SNAN, IEEE754_CLASS_SNAN):
+	case CLPAIR(IEEE754_CLASS_SNAN, IEEE754_CLASS_QNAN):
+	case CLPAIR(IEEE754_CLASS_SNAN, IEEE754_CLASS_ZERO):
+	case CLPAIR(IEEE754_CLASS_SNAN, IEEE754_CLASS_NORM):
+	case CLPAIR(IEEE754_CLASS_SNAN, IEEE754_CLASS_DNORM):
+	case CLPAIR(IEEE754_CLASS_SNAN, IEEE754_CLASS_INF):
+		return ieee754sp_nanxcpt(x);
+
+	case CLPAIR(IEEE754_CLASS_ZERO, IEEE754_CLASS_QNAN):
+	case CLPAIR(IEEE754_CLASS_NORM, IEEE754_CLASS_QNAN):
+	case CLPAIR(IEEE754_CLASS_DNORM, IEEE754_CLASS_QNAN):
+	case CLPAIR(IEEE754_CLASS_INF, IEEE754_CLASS_QNAN):
+		return y;
+
+	case CLPAIR(IEEE754_CLASS_QNAN, IEEE754_CLASS_QNAN):
+	case CLPAIR(IEEE754_CLASS_QNAN, IEEE754_CLASS_ZERO):
+	case CLPAIR(IEEE754_CLASS_QNAN, IEEE754_CLASS_NORM):
+	case CLPAIR(IEEE754_CLASS_QNAN, IEEE754_CLASS_DNORM):
+	case CLPAIR(IEEE754_CLASS_QNAN, IEEE754_CLASS_INF):
+		return x;
+
+	/*
+	 * Infinity handling
+	 */
+	case CLPAIR(IEEE754_CLASS_INF, IEEE754_CLASS_ZERO):
+	case CLPAIR(IEEE754_CLASS_ZERO, IEEE754_CLASS_INF):
+		if (zc == IEEE754_CLASS_QNAN)
+			return z;
+		ieee754_setcx(IEEE754_INVALID_OPERATION);
+		return ieee754sp_indef();
+
+	case CLPAIR(IEEE754_CLASS_NORM, IEEE754_CLASS_INF):
+	case CLPAIR(IEEE754_CLASS_DNORM, IEEE754_CLASS_INF):
+	case CLPAIR(IEEE754_CLASS_INF, IEEE754_CLASS_NORM):
+	case CLPAIR(IEEE754_CLASS_INF, IEEE754_CLASS_DNORM):
+	case CLPAIR(IEEE754_CLASS_INF, IEEE754_CLASS_INF):
+		if (zc == IEEE754_CLASS_QNAN)
+			return z;
+		return ieee754sp_inf(xs ^ ys);
+
+	case CLPAIR(IEEE754_CLASS_ZERO, IEEE754_CLASS_ZERO):
+	case CLPAIR(IEEE754_CLASS_ZERO, IEEE754_CLASS_NORM):
+	case CLPAIR(IEEE754_CLASS_ZERO, IEEE754_CLASS_DNORM):
+	case CLPAIR(IEEE754_CLASS_NORM, IEEE754_CLASS_ZERO):
+	case CLPAIR(IEEE754_CLASS_DNORM, IEEE754_CLASS_ZERO):
+		if (zc == IEEE754_CLASS_INF)
+			return ieee754sp_inf(zs);
+		/* Multiplication is 0 so just return z */
+		return z;
+
+	case CLPAIR(IEEE754_CLASS_DNORM, IEEE754_CLASS_DNORM):
+		SPDNORMX;
+
+	case CLPAIR(IEEE754_CLASS_NORM, IEEE754_CLASS_DNORM):
+		if (zc == IEEE754_CLASS_QNAN)
+			return z;
+		else if (zc == IEEE754_CLASS_INF)
+			return ieee754sp_inf(zs);
+		SPDNORMY;
+		break;
+
+	case CLPAIR(IEEE754_CLASS_DNORM, IEEE754_CLASS_NORM):
+		if (zc == IEEE754_CLASS_QNAN)
+			return z;
+		else if (zc == IEEE754_CLASS_INF)
+			return ieee754sp_inf(zs);
+		SPDNORMX;
+		break;
+
+	case CLPAIR(IEEE754_CLASS_NORM, IEEE754_CLASS_NORM):
+		if (zc == IEEE754_CLASS_QNAN)
+			return z;
+		else if (zc == IEEE754_CLASS_INF)
+			return ieee754sp_inf(zs);
+		/* fall through to real computations */
+	}
+
+	/* Finally get to do some computation */
+
+	/*
+	 * Do the multiplication bit first
+	 *
+	 * rm = xm * ym, re = xe + ye basically
+	 *
+	 * At this point xm and ym should have been normalized.
+	 */
+
+	/* rm = xm * ym, re = xe+ye basically */
+	assert(xm & SP_HIDDEN_BIT);
+	assert(ym & SP_HIDDEN_BIT);
+
+	re = xe + ye;
+	rs = xs ^ ys;
+
+	/* shunt to top of word */
+	xm <<= 32 - (SP_FBITS + 1);
+	ym <<= 32 - (SP_FBITS + 1);
+
+	/*
+	 * Multiply 32 bits xm, ym to give high 32 bits rm with stickness.
+	 */
+	lxm = xm & 0xffff;
+	hxm = xm >> 16;
+	lym = ym & 0xffff;
+	hym = ym >> 16;
+
+	lrm = lxm * lym;	/* 16 * 16 => 32 */
+	hrm = hxm * hym;	/* 16 * 16 => 32 */
+
+	t = lxm * hym; /* 16 * 16 => 32 */
+	at = lrm + (t << 16);
+	hrm += at < lrm;
+	lrm = at;
+	hrm = hrm + (t >> 16);
+
+	t = hxm * lym; /* 16 * 16 => 32 */
+	at = lrm + (t << 16);
+	hrm += at < lrm;
+	lrm = at;
+	hrm = hrm + (t >> 16);
+
+	rm = hrm | (lrm != 0);
+
+	/*
+	 * Sticky shift down to normal rounding precision.
+	 */
+	if ((int) rm < 0) {
+		rm = (rm >> (32 - (SP_FBITS + 1 + 3))) |
+		    ((rm << (SP_FBITS + 1 + 3)) != 0);
+		re++;
+	} else {
+		rm = (rm >> (32 - (SP_FBITS + 1 + 3 + 1))) |
+		     ((rm << (SP_FBITS + 1 + 3 + 1)) != 0);
+	}
+	assert(rm & (SP_HIDDEN_BIT << 3));
+
+	/* And now the addition */
+
+	assert(zm & SP_HIDDEN_BIT);
+
+	/*
+	 * Provide guard,round and stick bit space.
+	 */
+	zm <<= 3;
+
+	if (ze > re) {
+		/*
+		 * Have to shift y fraction right to align.
+		 */
+		s = ze - re;
+		SPXSRSYn(s);
+	} else if (re > ze) {
+		/*
+		 * Have to shift x fraction right to align.
+		 */
+		s = re - ze;
+		SPXSRSYn(s);
+	}
+	assert(ze == re);
+	assert(ze <= SP_EMAX);
+
+	if (zs == rs) {
+		/*
+		 * Generate 28 bit result of adding two 27 bit numbers
+		 * leaving result in zm, zs and ze.
+		 */
+		zm = zm + rm;
+
+		if (zm >> (SP_FBITS + 1 + 3)) { /* carry out */
+			SPXSRSX1();
+		}
+	} else {
+		if (zm >= rm) {
+			zm = zm - rm;
+		} else {
+			zm = rm - zm;
+			zs = rs;
+		}
+		if (zm == 0)
+			return ieee754sp_zero(ieee754_csr.rm == FPU_CSR_RD);
+
+		/*
+		 * Normalize in extended single precision
+		 */
+		while ((zm >> (SP_MBITS + 3)) == 0) {
+			zm <<= 1;
+			ze--;
+		}
+
+	}
+	return ieee754sp_format(zs, ze, zm);
+}