lib/msun amd64: use AVX FMA instructions for fma(3) and fmaf(3) when available

Utilize the ARCHLEVEL framework from libc/amd64/string to provide the
way for runtime selection of the implementation, if wanted.

Reviewed by:	fuz, kfv
Discussed with:	kargl
Sponsored by:	The FreeBSD Foundation
MFC after:	1 week
Differential revision:	https://reviews.freebsd.org/D59462

lib/msun/amd64/Makefile.inc

@@ -1,8 +1,15 @@
-ARCH_SRCS = e_fmod.S e_fmodf.S e_fmodl.S \
+.PATH:	${LIBC_SRCTOP}/amd64/string
+
+ARCH_SRCS = amd64_archlevel.c \
+	    e_fmod.S e_fmodf.S e_fmodl.S \
 	    e_remainder.S e_remainderf.S e_remainderl.S \
 	    e_sqrt.S e_sqrtf.S e_sqrtl.S \
+	    s_fma_avx.S s_fmaf_avx.S \
 	    s_llrint.S s_llrintf.S s_llrintl.S \
 	    s_logbl.S s_lrint.S s_lrintf.S s_lrintl.S \
 	    s_remquo.S s_remquof.S s_remquol.S \
 	    s_rintl.S s_scalbn.S s_scalbnf.S s_scalbnl.S
 LDBL_PREC = 64
+
+CFLAGS+=-D_RENAME_FMA
+CFLAGS+=-D_RENAME_FMAF

lib/msun/amd64/s_fma_avx.S

@@ -0,0 +1,24 @@
+/*-
+ * SPDX-License-Identifier: BSD-2-Clause
+ *
+ * Copyright 2026 The FreeBSD Foundation
+ *
+ * This software were developed by
+ * Konstantin Belousov <kib@FreeBSD.org> under sponsorship from
+ * the FreeBSD Foundation.
+ */
+
+#include "amd64_archlevel.h"
+
+	.text
+ARCHENTRY(fma, avx)
+	vfmadd132sd	%xmm1, %xmm2, %xmm0
+	ret
+ARCHEND(fma, avx)
+
+ARCHFUNCS(fma)
+	ARCHFUNC(fma, sw)
+	NOARCHFUNC
+	NOARCHFUNC
+	ARCHFUNC(fma, avx)
+ENDARCHFUNCS(fma)

lib/msun/amd64/s_fmaf_avx.S

@@ -0,0 +1,24 @@
+/*-
+ * SPDX-License-Identifier: BSD-2-Clause
+ *
+ * Copyright 2026 The FreeBSD Foundation
+ *
+ * This software were developed by
+ * Konstantin Belousov <kib@FreeBSD.org> under sponsorship from
+ * the FreeBSD Foundation.
+ */
+
+#include "amd64_archlevel.h"
+
+	.text
+ARCHENTRY(fmaf, avx):
+	vfmadd132ss	%xmm1, %xmm2, %xmm0
+	ret
+ARCHEND(fmaf, avx):
+
+ARCHFUNCS(fmaf)
+	ARCHFUNC(fmaf, sw)
+	NOARCHFUNC
+	NOARCHFUNC
+	ARCHFUNC(fmaf, avx)
+ENDARCHFUNCS(fmaf)

lib/msun/src/s_fma.c

@@ -163,6 +163,10 @@ dd_mul(double a, double b)
 	return (ret);
 }
 
+#ifdef _RENAME_FMA
+double fma_sw(double, double, double);
+#endif
+
 /*
  * Fused multiply-add: Compute x * y + z with a single rounding error.
  *
@@ -181,7 +185,12 @@ dd_mul(double a, double b)
  * since this implementation will likely be several times slower.
  */
 double
-fma(double x, double y, double z)
+#ifdef _RENAME_FMA
+fma_sw
+#else
+fma
+#endif
+(double x, double y, double z)
 {
 	double xs, ys, zs, adj;
 	struct dd xy, r;

lib/msun/src/s_fmaf.c

@@ -38,6 +38,10 @@ fmaf(float x, float y, float z)
 	return (__builtin_fmaf(x, y, z));
 }
 #else
+#ifdef _RENAME_FMAF
+float fmaf_sw(float, float, float);
+#endif
+
 /*
  * Fused multiply-add: Compute x * y + z with a single rounding error.
  *
@@ -46,7 +50,12 @@ fmaf(float x, float y, float z)
  * rounding occurs.
  */
 float
-fmaf(float x, float y, float z)
+#ifdef _RENAME_FMAF
+fmaf_sw
+#else
+fmaf
+#endif
+(float x, float y, float z)
 {
 	double xy, result;
 	uint32_t hr, lr;