For PR 26483, IA-64 denorm failure due to unwanted rounding.

* testsuite/libffi.call/float4.c: New testcase.

From-SVN: r112768

libffi/ChangeLog

@@ -1,3 +1,7 @@
+2006-04-07  James E Wilson  <wilson@specifix.com>
+
+	* testsuite/libffi.call/float4.c: New testcase.
+
 2006-04-05  John David Anglin  <dave.anglin@nrc-cnrc.gc.ca>
 	    Andreas Tobler  <a.tobler@schweiz.ch>
 

libffi/testsuite/libffi.call/float4.c

@@ -0,0 +1,60 @@
+/* Area:	ffi_call
+   Purpose:	Check denorm double value.
+   Limitations:	none.
+   PR:		PR26483.
+   Originator:	From the original ffitest.c  */
+
+/* { dg-do run } */
+#include "ffitest.h"
+#include "float.h"
+
+typedef union
+{
+  double d;
+  unsigned char c[sizeof (double)];
+} value_type;
+
+#define CANARY 0xba
+
+static double dblit(double d)
+{
+  return d;
+}
+
+int main (void)
+{
+  ffi_cif cif;
+  ffi_type *args[MAX_ARGS];
+  void *values[MAX_ARGS];
+  double d;
+  value_type result[2];
+  unsigned int i;
+
+  args[0] = &ffi_type_double;
+  values[0] = &d;
+  
+  /* Initialize the cif */
+  CHECK(ffi_prep_cif(&cif, FFI_DEFAULT_ABI, 1,
+		     &ffi_type_double, args) == FFI_OK);
+  
+  d = DBL_MIN / 2;
+  
+  /* Put a canary in the return array.  This is a regression test for
+     a buffer overrun.  */
+  memset(result[1].c, CANARY, sizeof (double));
+
+  ffi_call(&cif, FFI_FN(dblit), &result[0].d, values);
+  
+  /* The standard delta check doesn't work for denorms.  Since we didn't do
+     any arithmetic, we should get the original result back, and hence an
+     exact check should be OK here.  */
+ 
+  CHECK(result[0].d == dblit(d));
+
+  /* Check the canary.  */
+  for (i = 0; i < sizeof (double); ++i)
+    CHECK(result[1].c[i] == CANARY);
+
+  exit(0);
+
+}