work around regression building for armhf with clang (compiler bug)

commit e4355bd6 moved the math asm
from external source files to inline asm, but unfortunately, all
current releases of clang use the wrong inline asm constraint codes
for float and double ("w" and "P" instead of "t" and "w",
respectively). this patch adds detection for the bug in configure,
and, for now, just disables the affected asm on broken clang versions.

configure

@@ -588,6 +588,20 @@ fi
 if test "$ARCH" = "arm" ; then
 trycppif __ARMEB__ "$t" && SUBARCH=${SUBARCH}eb
 trycppif __ARM_PCS_VFP "$t" && SUBARCH=${SUBARCH}hf
+# Versions of clang up until at least 3.8 have the wrong constraint codes
+# for floating point operands to inline asm. Detect this so the affected
+# source files can just disable the asm.
+if test "$cc_family" = clang ; then
+printf "checking whether clang's vfp asm constraints work... "
+echo 'float f(float x) { __asm__("":"+t"(x)); return x; }' > "$tmpc"
+if $CC $CFLAGS_C99FSE $CPPFLAGS $CFLAGS -c -o /dev/null "$tmpc" >/dev/null 2>&1 ; then
+printf "yes\n"
+else
+printf "no\n"
+CFLAGS_AUTO="$CFLAGS_AUTO -DBROKEN_VFP_ASM"
+CFLAGS_AUTO="${CFLAGS_AUTO# }"
+fi
+fi
 fi
 
 if test "$ARCH" = "aarch64" ; then

src/math/arm/fabsf.c

 #include <math.h>
 
-#if __ARM_PCS_VFP
+#if __ARM_PCS_VFP && !BROKEN_VFP_ASM
 
 float fabsf(float x)
 {

src/math/arm/sqrtf.c

 #include <math.h>
 
-#if __ARM_PCS_VFP || (__VFP_FP__ && !__SOFTFP__)
+#if (__ARM_PCS_VFP || (__VFP_FP__ && !__SOFTFP__)) && !BROKEN_VFP_ASM
 
 float sqrtf(float x)
 {