Determination of floating-point types parameters when cross-compiling.

diff --git a/m4/floatparam.m4 b/m4/floatparam.m4
new file mode 100644 (file)
index 0000000..56dd53e
--- /dev/null
+++ b/m4/floatparam.m4
@@ -0,0 +1,129 @@
+# floatparam.m4 serial 1
+dnl Copyright (C) 2005 Free Software Foundation, Inc.
+dnl This file is free software; the Free Software Foundation
+dnl gives unlimited permission to copy and/or distribute it,
+dnl with or without modifications, as long as this notice is preserved.
+
+dnl From Bruno Haible.
+
+AC_DEFUN([CL_FLOATPARAM_CROSS],
+[
+  AC_REQUIRE([CL_LONGDOUBLE])
+  cl_machine_file_h=$1
+  {
+    echo "/* Rounding modes, for use below */"
+    echo "#define rounds_to_nearest        0  /* 0.5 ulp */"
+    echo "#define rounds_to_zero           1  /* 1 ulp */"
+    echo "#define rounds_to_infinity       2  /* 1 ulp */"
+    echo "#define rounds_to_minus_infinity 3  /* 1 ulp */"
+    echo
+    for type in float double "`if test $cl_cv_c_longdouble = yes; then echo 'long double'; fi`"; do
+      if test -n "$type"; then
+        epsilon_bits=-1; y="($type)1.0"
+        while true; do
+          AC_TRY_COMPILE([],
+            [typedef int verify[2*(
+               (($type)(($type)1.0 + ($type)($y)) == ($type)1.0)
+               || ($type)(($type)(($type)1.0 + ($type)($y)) - ($type)1.0) != ($type)($y)
+             ) - 1];],
+            [break;])
+          epsilon_bits=`expr $epsilon_bits + 1`; y="$y * ($type)0.5"
+        done
+        negepsilon_bits=-1; y="($type)-1.0"
+        while true; do
+          AC_TRY_COMPILE([],
+            [typedef int verify[2*(
+               (($type)(($type)1.0 + ($type)($y)) == ($type)1.0)
+               || ($type)(($type)(($type)1.0 + ($type)($y)) - ($type)1.0) != ($type)($y)
+             ) - 1];],
+            [break;])
+          negepsilon_bits=`expr $negepsilon_bits + 1`; y="$y * ($type)0.5"
+        done
+        echo "/* Properties of type \`$type': */"
+        echo "/* Largest n for which 1+2^(-n) is exactly represented is $epsilon_bits. */"
+        echo "/* Largest n for which 1-2^(-n) is exactly represented is $negepsilon_bits. */"
+        if test `expr $negepsilon_bits '<=' $epsilon_bits` = 1; then
+          echo "#error \"No exponent jump at 1.0 for type $type!\""
+        else
+          if test `expr $negepsilon_bits '>' $epsilon_bits + 1` = 1; then
+            echo "/* Base for type '$type' is 2^"`expr $negepsilon_bits - $epsilon_bits`
+          fi
+          echo "#define "`echo $type | sed -e 's, ,_,g'`"_mant_bits "`expr $epsilon_bits + 1`
+        fi
+        x="($type)1.0"
+        i=$epsilon_bits
+        while test $i != 0; do
+          x="$x * ($type)0.5"
+          i=`expr $i - 1`
+        done
+        x="($type)($x)"
+        y1="($type)(($type)1.0 + ($type)5.0*$x)"
+        y2="($type)(($type)1.0 + ($type)6.0*$x)"
+        ys1="($type)(($type)1.0 + ($type)5.4*$x)"
+        ys2="($type)(($type)1.0 + ($type)5.6*$x)"
+        z1="($type)(($type)-1.0 + ($type)(-5.0)*$x)"
+        z2="($type)(($type)-1.0 + ($type)(-6.0)*$x)"
+        zs1="($type)(($type)-1.0 + ($type)(-5.4)*$x)"
+        zs2="($type)(($type)-1.0 + ($type)(-5.6)*$x)"
+        rounds=
+        if test -z "$rounds"; then
+          AC_TRY_COMPILE([],
+            [typedef int verify[2*(
+               $ys1 == $y1 && $ys2 == $y2 && $zs1 == $z1 && $zs2 == $z2
+             ) - 1];],
+            [rounds=rounds_to_nearest])
+        fi
+        if test -z "$rounds"; then
+          AC_TRY_COMPILE([],
+            [typedef int verify[2*(
+               $ys1 == $y1 && $ys2 == $y1 && $zs1 == $z1 && $zs2 == $z1
+             ) - 1];],
+            [rounds=rounds_to_zero])
+        fi
+        if test -z "$rounds"; then
+          AC_TRY_COMPILE([],
+            [typedef int verify[2*(
+               $ys1 == $y2 && $ys2 == $y2 && $zs1 == $z1 && $zs2 == $z1
+             ) - 1];],
+            [rounds=rounds_to_infinity])
+        fi
+        if test -z "$rounds"; then
+          AC_TRY_COMPILE([],
+            [typedef int verify[2*(
+               $ys1 == $y1 && $ys2 == $y1 && $zs1 == $z2 && $zs2 == $z2
+             ) - 1];],
+            [rounds=rounds_to_minus_infinity])
+        fi
+        if test -n "$rounds"; then
+          echo "#define "`echo $type | sed -e 's, ,_,g'`"_rounds $rounds"
+        else
+          echo "#error \"Unknown rounding mode for type $type!\""
+        fi
+        echo
+      fi
+    done
+    dnl Words-in-a-double endianness test. Note that, assuming IEEE 754 format,
+    dnl 2.5479915693083957     = { 0x40 0x04 0x62 0x49 0x67 0x65 0x4E 0x64 } ..bIgeNd
+    dnl 1.4396527506122064e164 = { 0x62 0x04 0x00 0x00 0x4E 0x65 0x67 0x49 } b...NegI
+    dnl 2.5495230282078065     = { 0x40 0x04 0x65 0x6C 0x54 0x54 0x69 0x4C } ..elTTiL
+    dnl 1.4139248369879473e214 = { 0x6C 0x65 0x00 0x00 0x4C 0x69 0x54 0x54 } le..LiTT
+    double_wordorder_bigendian_p=
+    AC_TRY_COMPILE([double a[9] = {
+       0, 2.5479915693083957, 0, 1.4396527506122064e164,
+       0, 2.5495230282078065, 0, 1.4139248369879473e214,
+       0 };], [], [
+      if grep LiTTle conftest.$ac_objext >/dev/null ; then
+        double_wordorder_bigendian_p=0
+      else
+        if grep bIgeN conftest.$ac_objext >/dev/null ; then
+          double_wordorder_bigendian_p=1
+        fi
+      fi])
+    if test -n "$double_wordorder_bigendian_p"; then
+      echo "#define double_wordorder_bigendian_p $double_wordorder_bigendian_p"
+    else
+      echo "/* Dazed and confused!  Better not define anything. */"
+    fi
+    echo
+  } > "$cl_machine_file_h"
+])