Fix exponentl.m4 test

[Bruno Haible]

Tim Rühsen wrote:
> Attached is the config.log and config.cache for
> - gcc-9 with unset CFLAGS (*.gcc-9)
> - clang-9 with unset CFLAGS (*.clang-9)
> - gcc-10 with unset CFLAGS (*.gcc-10)
> - clang-10 with unset CFLAGS (*.clang-10)
> 
> And the same set with -fsanitize... (*.sanitize).

Thanks, that's much more useful.

The first interesting finding is this difference between config.cache.gcc-9
and config.cache.gcc-10:

-gl_cv_cc_long_double_expbit0=${gl_cv_cc_long_double_expbit0='word 2 bit 0'}
+gl_cv_cc_long_double_expbit0=${gl_cv_cc_long_double_expbit0=unknown}

Indeed, with GCC 10 I see:

  checking where to find the exponent in a 'long double'... unknown

I see it also with older versions of GCC, with "-O2".

What happens is that on x86_64, 'long double' is the 80-bit "extended double"
format, which needs 3 'unsigned int' values. But sizeof (long double) = 16
[whereas on x86, sizeof (long double) = 12]. So, in the test, NWORDS = 4.
But when optimizing, the compiler apparently copies only 3 words, not 4 words,
when reading or writing a parameter of type 'long double'.

I would like to avoid hard coding the particular representation (since
possible some other compilers will use 'float128' as 'long double').

This patch brings back

  checking where to find the exponent in a 'long double'... word 2 bit 0


2020-05-23  Bruno Haible  <address@hidden>

        isnanl, isnanl-nolibm: Make a test work better with "gcc -O2" on x86_64.
        * m4/exponentl.m4 (gl_LONG_DOUBLE_EXPONENT_LOCATION): Pass the
        'long double' values by reference, with values taken from a statically
        allocated array.

diff --git a/m4/exponentl.m4 b/m4/exponentl.m4
index b33b3bf..0a35c11 100644
--- a/m4/exponentl.m4
+++ b/m4/exponentl.m4
@@ -1,4 +1,4 @@
-# exponentl.m4 serial 4
+# exponentl.m4 serial 5
 dnl Copyright (C) 2007-2020 Free Software Foundation, Inc.
 dnl This file is free software; the Free Software Foundation
 dnl gives unlimited permission to copy and/or distribute it,
@@ -22,14 +22,14 @@ typedef union { long double value; unsigned int 
word[NWORDS]; }
         memory_long_double;
 static unsigned int ored_words[NWORDS];
 static unsigned int anded_words[NWORDS];
-static void add_to_ored_words (long double x)
+static void add_to_ored_words (long double *x)
 {
   memory_long_double m;
   size_t i;
   /* Clear it first, in case
      sizeof (long double) < sizeof (memory_long_double).  */
   memset (&m, 0, sizeof (memory_long_double));
-  m.value = x;
+  m.value = *x;
   for (i = 0; i < NWORDS; i++)
     {
       ored_words[i] |= m.word[i];
@@ -38,17 +38,15 @@ static void add_to_ored_words (long double x)
 }
 int main ()
 {
+  static long double samples[5] = { 0.25L, 0.5L, 1.0L, 2.0L, 4.0L };
   size_t j;
   FILE *fp = fopen ("conftest.out", "w");
   if (fp == NULL)
     return 1;
   for (j = 0; j < NWORDS; j++)
     anded_words[j] = ~ (unsigned int) 0;
-  add_to_ored_words (0.25L);
-  add_to_ored_words (0.5L);
-  add_to_ored_words (1.0L);
-  add_to_ored_words (2.0L);
-  add_to_ored_words (4.0L);
+  for (j = 0; j < 5; j++)
+    add_to_ored_words (&samples[j]);
   /* Remove bits that are common (e.g. if representation of the first mantissa
      bit is explicit).  */
   for (j = 0; j < NWORDS; j++)