Re: Java floating-point mess

Re: Java floating-point mess

[Robert Dewar]

<<The front end is required to do these computations in some contexts,
namely compile-time constants.  This is a concern to us, but it isn't
immediately pressing.  We already violate the spec with respect to
compile-time constants in several ways.
>>

Sure, of course I understand that, but these computations should be done
using simulated IEEE arithmetic rather than host arithmetic.

Re: Java floating-point mess

[Tom Tromey]

>>>>> "Stephen" == Stephen L Moshier <steve@moshier.net> writes:

Stephen> From the IEEE fanatics though, I got the impression that Java
Stephen> was supposed to work the same on all computers so you would
Stephen> have to provide a run-time emulator if your computer was not
Stephen> IEEE.  Is that the specification or was that just political
Stephen> propaganda?

If your class or method is strictfp then the results of your
computation are tightly specified.  Otherwise, in most contexts, there
is a bit of leeway.  I think the rule is 1 bit of difference is
acceptable, but you'd have to check the spec to make sure.  Earlier
versions of the spec didn't have strictfp; that was added to allow
code to be faster (but less accurate) on x86 (and presumably others).

Anyway, yes, it is the specification and not just propaganda.

Tom

Re: Java floating-point mess

[Tom Tromey]

>>>>> "Robert" == Robert Dewar <dewar@gnat.com> writes:

Andrew> The only issue here is that we won't be able to cross-compile
Andrew> Java on a VAX or somesuch, isn't it?  Or am I missing
Andrew> something here?

Robert> That sounds right provided that the front end is not trying to
Robert> use host floating-point arithmetic to precompute expressions
Robert> (IEEE is not nearly enough of an absolute standard for this to
Robert> give the right results in all cases).

The front end is required to do these computations in some contexts,
namely compile-time constants.  This is a concern to us, but it isn't
immediately pressing.  We already violate the spec with respect to
compile-time constants in several ways.

Tom

Re: Java floating-point mess

[Stephen L Moshier]

Here are the external functions that real.o seems to reference:

  mode_class
  mode_bitsize

  warning
  extra_warnings
  fancy_abort
  _hex_value
  _sch_istable

The reference to mode_class could just be deleted because it occurs
only in a sanity test that is probably not vitally important.

mode_bitsize is used to determine the width of types such as SFmode
etc.  I think that this code originally just tested for the
machine_mode enum, and it could revert to that.  (If you want
DFmode to be 32 bits wide, or some such, you are already in
deep trouble.)  Alternatively the two functions that use mode_bitsize,
REAL_VALUE_FROM_INT and REAL_VALUE_FROM_UNSIGNED_INT, might be removed to
a separate source file if you don't need them in java.

The other external references, fancy_abort, etc., do not look like they
are specific to the back end though I didn't check to see if they
are present in java.   _hex_value and _sch_istable were not used and
could be eliminated by reverting to an earlier version of the code.

Java floating-point mess

[Stephen L Moshier]

> The Java front end has a number of places that assume not only that
> the host and target floating point formats are identical, but that
> both are IEEE.  I've tried to get rid of them, but run afoul of its
> not being possible to link real.o into gcjh -- it wants to drag in
> most of the back end.

Originally, real.o did not drag in anything, as it had to fit into
an independent test harness.  That will have to happen again, in order
to test C99 features.  Maybe this would be an excuse to do some of that
work, or at least find out why you are having trouble.

From the IEEE fanatics though, I got the impression that Java was
supposed to work the same on all computers so you would have to
provide a run-time emulator if your computer was not IEEE.  Is that
the specification or was that just political propaganda?

Re: Java floating-point mess

[Robert Dewar]

<<The only issue here is that we won't be able to cross-compile Java on
a VAX or somesuch, isn't it?  Or am I missing something here?
>>

That sounds right provided that the front end is not trying to use host
floating-point arithmetic to precompute expressions (IEEE is not nearly
enough of an absolute standard for this to give the right results in
all cases).

Re: Java floating-point mess

[Andrew Haley]

Zack Weinberg writes:
 > On Thu, Apr 04, 2002 at 06:56:15AM -0500, Robert Dewar wrote:
 > > <<The Java front end has a number of places that assume not only that
 > > the host and target floating point formats are identical, but that
 > > both are IEEE.  I've tried to get rid of them, but run afoul of its
 > > not being possible to link real.o into gcjh -- it wants to drag in
 > > most of the back end.
 > > >>
 > > 
 > > Presumably the complaint is about assuming that the host fpt
 > > format is IEEE, the target fpt format presumably MUST be IEEE,
 > > since this is required by Java, no?
 > 
 > Well, the Java front end doesn't appear to take any special care to
 > ensure that SFmode is IEEE single and DFmode double.  (We'd have to
 > have a floating point emulator library built into libjava that
 > implemented IEEE fp, to get it right for non-IEEE targets.)  But yes,
 > the code I'm looking at right now is only for the host side.
 > 
 > Thinking out loud... The major problem is with gjavah, which wants to
 > read in a .class file and print out things like
 > 
 > const jdouble pi = 3.1415926... ;
 > 
 > which are then going to be interpreted by the C++ compiler -- come to
 > think of it, there is no way to win here on a non-IEEE target.

Right -- Java requires that the target FP format be IEEE.  We don't
support Java on non-IEEE targets.  We can solve any build problems by
not building Java on hosts that are non-IEEE.

The only issue here is that we won't be able to cross-compile Java on
a VAX or somesuch, isn't it?  Or am I missing something here?

Andrew.

Re: Java floating-point mess

[Zack Weinberg]

On Thu, Apr 04, 2002 at 06:56:15AM -0500, Robert Dewar wrote:
> <<The Java front end has a number of places that assume not only that
> the host and target floating point formats are identical, but that
> both are IEEE.  I've tried to get rid of them, but run afoul of its
> not being possible to link real.o into gcjh -- it wants to drag in
> most of the back end.
> >>
> 
> Presumably the complaint is about assuming that the host fpt format is
> IEEE, the target fpt format presumably MUST be IEEE, since this is required
> by Java, no?

Well, the Java front end doesn't appear to take any special care to
ensure that SFmode is IEEE single and DFmode double.  (We'd have to
have a floating point emulator library built into libjava that
implemented IEEE fp, to get it right for non-IEEE targets.)  But yes,
the code I'm looking at right now is only for the host side.

Thinking out loud... The major problem is with gjavah, which wants to
read in a .class file and print out things like

const jdouble pi = 3.1415926... ;

which are then going to be interpreted by the C++ compiler -- come to
think of it, there is no way to win here on a non-IEEE target.
Ignoring the language spec and using the target's floating point,
whatever it happens to be, might be the sanest thing to do from an
interoperability standpoint.

I think I can get around the immediate problem by using C99
hexadecimal floating-point constants.

[I should have cc:ed java@ from the beginning -- the original message
is at http://gcc.gnu.org/ml/gcc/2002-04/msg00135.html if anyone's
interested.]

zw

Re: Java floating-point mess

[Robert Dewar]

<<The Java front end has a number of places that assume not only that
the host and target floating point formats are identical, but that
both are IEEE.  I've tried to get rid of them, but run afoul of its
not being possible to link real.o into gcjh -- it wants to drag in
most of the back end.
>>

Presumably the complaint is about assuming that the host fpt format is
IEEE, the target fpt format presumably MUST be IEEE, since this is required
by Java, no?

Java floating-point mess

[Zack Weinberg]

The Java front end has a number of places that assume not only that
the host and target floating point formats are identical, but that
both are IEEE.  I've tried to get rid of them, but run afoul of its
not being possible to link real.o into gcjh -- it wants to drag in
most of the back end.

I've appended what I've got to this message.  Suggestions would be
appreciated.

zw

===================================================================
Index: java/Make-lang.in
--- java/Make-lang.in	2002/03/29 21:46:26	1.82
+++ java/Make-lang.in	2002/04/04 10:02:52
@@ -107,7 +107,7 @@ JAVA_OBJS = java/parse.o java/class.o ja
   java/jcf-path.o java/xref.o java/boehm.o mkdeps.o
 
 GCJH_OBJS = java/gjavah.o java/jcf-io.o java/jcf-depend.o java/jcf-path.o \
-  java/zextract.o version.o mkdeps.o errors.o
+  java/zextract.o version.o mkdeps.o errors.o real.o
 
 JVSCAN_OBJS = java/parse-scan.o java/jv-scan.o version.o
 
@@ -278,7 +278,8 @@ java/expr.o: java/expr.c $(CONFIG_H) $(J
   $(RTL_H) $(EXPR_H) java/javaop.h java/java-opcodes.h except.h \
   java/java-except.h java/java-except.h java/parse.h toplev.h \
   $(SYSTEM_H) $(GGC_H)
-java/jcf-depend.o: java/jcf-depend.c $(CONFIG_H) $(SYSTEM_H) java/jcf.h
+java/jcf-depend.o: java/jcf-depend.c $(CONFIG_H) $(SYSTEM_H) java/jcf.h \
+  real.h $(MACHMODE_H)
 java/jcf-parse.o: java/jcf-parse.c $(CONFIG_H) $(JAVA_TREE_H) flags.h \
   input.h java/java-except.h $(SYSTEM_H) toplev.h java/parse.h $(GGC_H) \
   debug.h
@@ -310,7 +311,8 @@ java/jcf-io.o: java/jcf-io.c $(CONFIG_H)
 		$(srcdir)/java/jcf-io.c $(OUTPUT_OPTION)
 
 # jcf-path.o needs a -D.
-java/jcf-path.o: java/jcf-path.c $(CONFIG_H) $(SYSTEM_H) java/jcf.h
+java/jcf-path.o: java/jcf-path.c $(CONFIG_H) $(SYSTEM_H) java/jcf.h \
+  real.h $(MACHMODE_H)
 	$(CC) -c $(ALL_CFLAGS) $(ALL_CPPFLAGS) $(INCLUDES) \
 	  -DLIBGCJ_ZIP_FILE='"$(prefix)/share/libgcj.jar"' \
 	  $(srcdir)/java/jcf-path.c $(OUTPUT_OPTION)
===================================================================
Index: java/constants.c
--- java/constants.c	2001/04/26 18:32:22	1.19
+++ java/constants.c	2002/04/04 10:02:52
@@ -23,9 +23,9 @@ The Free Software Foundation is independ
 
 #include "config.h"
 #include "system.h"
-#include "jcf.h"
 #include "tree.h"
 #include "java-tree.h"
+#include "jcf.h"
 #include "toplev.h"
 #include "ggc.h"
 
===================================================================
Index: java/gjavah.c
--- java/gjavah.c	2002/03/03 08:35:11	1.85
+++ java/gjavah.c	2002/04/04 10:02:52
@@ -28,10 +28,10 @@ The Free Software Foundation is independ
 #include "system.h"
 #include <math.h>
 
-#include "jcf.h"
 #include "tree.h"
-#include "javaop.h"
 #include "java-tree.h"
+#include "javaop.h"
+#include "jcf.h"
 #include "java-opcodes.h"
 
 #include <getopt.h>
@@ -130,8 +130,6 @@ static void print_full_cxx_name PARAMS (
 static void decompile_method PARAMS ((FILE*, JCF*, int));
 static void add_class_decl PARAMS ((FILE*, JCF*, JCF_u2));
 
-static int java_float_finite PARAMS ((jfloat));
-static int java_double_finite PARAMS ((jdouble));
 static void print_name PARAMS ((FILE *, JCF *, int));
 static void print_base_classname PARAMS ((FILE *, JCF *, int));
 static int utf8_cmp PARAMS ((const unsigned char *, int, const char *));
@@ -237,40 +235,6 @@ static int decompiled = 0;
 
 #include "jcf-reader.c"
 
-/* Some useful constants.  */
-#define F_NAN_MASK 0x7f800000
-#if (1 == HOST_FLOAT_WORDS_BIG_ENDIAN) && ! defined (HOST_WORDS_BIG_ENDIAN)
-#define D_NAN_MASK 0x000000007ff00000LL
-#else
-#define D_NAN_MASK 0x7ff0000000000000LL
-#endif
-
-/* Return 1 if F is not Inf or NaN.  */
-static int
-java_float_finite (f)
-     jfloat f;
-{
-  union Word u;
-  u.f = f;
-
-  /* We happen to know that F_NAN_MASK will match all NaN values, and
-     also positive and negative infinity.  That's why we only need one
-     test here.  See The Java Language Specification, section 20.9.  */
-  return (u.i & F_NAN_MASK) != F_NAN_MASK;
-}
-
-/* Return 1 if D is not Inf or NaN.  */
-static int
-java_double_finite (d)
-     jdouble d;
-{
-  union DWord u;
-  u.d = d;
-
-  /* Now check for all NaNs.  */
-  return (u.l & D_NAN_MASK) != D_NAN_MASK;
-}
-
 /* Print a character, appropriately mangled for JNI.  */
 
 static void
@@ -747,24 +711,32 @@ DEFUN(print_field_info, (stream, jcf, na
 	      break;
 	    case CONSTANT_Float:
 	      {
-		jfloat fnum = JPOOL_FLOAT (jcf, current_field_value);
+		REAL_VALUE_TYPE fnum = JPOOL_FLOAT (jcf, current_field_value);
+		char s[30];
 		fputs ("const jfloat ", out);
 		print_field_name (out, jcf, name_index, 0);
-		if (! java_float_finite (fnum))
+		if (REAL_VALUE_ISINF (fnum) || REAL_VALUE_ISNAN (fnum))
 		  fputs (";\n", out);
 		else
-		  fprintf (out, " = %.10g;\n",  fnum);
+		  {
+		    REAL_VALUE_TO_DECIMAL (fnum, "%.10g", s);
+		    fprintf (out, " = %s;\n", s);
+		  }
 	      }
 	      break;
 	    case CONSTANT_Double:
 	      {
-		jdouble dnum = JPOOL_DOUBLE (jcf, current_field_value);
+		REAL_VALUE_TYPE fnum = JPOOL_DOUBLE (jcf, current_field_value);
+		char s[30];
 		fputs ("const jdouble ", out);
 		print_field_name (out, jcf, name_index, 0);
-		if (! java_double_finite (dnum))
+		if (REAL_VALUE_ISINF (fnum) || REAL_VALUE_ISNAN (fnum))
 		  fputs (";\n", out);
 		else
-		  fprintf (out, " = %.17g;\n",  dnum);
+		  {
+		    REAL_VALUE_TO_DECIMAL (fnum, "%.17g", s);
+		    fprintf (out, " = %s;\n", s);
+		  }
 	      }
 	      break;
 	    default:
===================================================================
Index: java/javaop.h
--- java/javaop.h	2001/05/26 01:31:47	1.10
+++ java/javaop.h	2002/04/04 10:02:52
@@ -57,27 +57,15 @@ typedef int32                   jint;
 typedef int64                   jlong;
 typedef void*                   jref;
 
-/* A 32-bit IEEE single-precision float. */
-#ifndef jfloat 
-#define jfloat float
-#endif
-
-/* A 32-bit IEEE double-precision float. */
-#ifndef jdouble
-#define jdouble double
-#endif
-
 union Word {
   jint i;
-  jfloat f;
   void *p;
 };
 
 /* A jword is an unsigned integral type big enough for a 32-bit jint
-   or jfloat *or* a pointer.  It is the type appropriate for stack
-   locations and local variables in a Java interpreter. */
+   or a pointer.  It is the type appropriate for stack locations and
+   local variables in a Java interpreter. */
 
-
 #ifndef jword
 #define jword uint32
 #endif
@@ -102,11 +90,10 @@ union Word {
          | (BCODE[PC-2] << 8) | (BCODE[PC-1]))))
 #endif
 
-static inline jfloat
+static inline REAL_VALUE_TYPE
 WORD_TO_FLOAT(jword w)
-{ union Word wu;
-  wu.i = w;
-  return wu.f;
+{
+  return REAL_VALUE_FROM_TARGET_SINGLE (w);
 } 
 
 /* Sign extend w.  If the host on which this cross-compiler runs uses
@@ -129,20 +116,30 @@ WORDS_TO_LONG(jword hi, jword lo)
 }
 
 union DWord {
-  jdouble d;
   jlong l;
   jword w[2];
 };
 
-static inline jdouble
+static inline REAL_VALUE_TYPE
 WORDS_TO_DOUBLE(jword hi, jword lo)
-{ union DWord wu;
-#if (1 == HOST_FLOAT_WORDS_BIG_ENDIAN)
-  wu.l = WORDS_TO_LONG(lo, hi);
-#else
-  wu.l = WORDS_TO_LONG(hi, lo);
-#endif
-  return wu.d;
+{
+  HOST_WIDE_INT h[2];
+  if (HOST_BITS_PER_WIDE_INT == 64)
+    {
+      h[0] = WORDS_TO_LONG (hi, lo);
+      h[1] = 0;
+    }
+  else if (FLOAT_WORDS_BIG_ENDIAN)
+    {
+      h[0] = hi;
+      h[1] = lo;
+    }
+  else
+    {
+      h[1] = hi;
+      h[0] = lo;
+    }
+  return REAL_VALUE_FROM_TARGET_DOUBLE (h);
 } 
 
 /* If PREFIX_CHAR is the first character of the Utf8 encoding of a character,
===================================================================
Index: java/jcf-depend.c
--- java/jcf-depend.c	2001/01/05 07:50:24	1.13
+++ java/jcf-depend.c	2002/04/04 10:02:52
@@ -29,6 +29,8 @@ The Free Software Foundation is independ
 
 #include <assert.h>
 
+#include "machmode.h"
+#include "real.h"
 #include "jcf.h"
 
 \f
===================================================================
Index: java/jcf-dump.c
--- java/jcf-dump.c	2002/03/03 08:35:11	1.43
+++ java/jcf-dump.c	2002/04/04 10:02:53
@@ -504,16 +504,22 @@ DEFUN(print_constant, (out, jcf, index, 
       break;
     case CONSTANT_Float:
       {
-	jfloat fnum = JPOOL_FLOAT (jcf, index);
-	fprintf (out, "%s%.10g", verbosity > 0 ? "Float " : "", (double) fnum);
+	char s[30];
+	REAL_VALUE_TYPE r = JPOOL_FLOAT (jcf, index);
+	REAL_VALUE_TO_DECIMAL (r, "%.10g", s);
+	
+	fprintf (out, "%s%s", verbosity > 0 ? "Float " : "", s);
 	if (verbosity > 1)
-	  fprintf (out, ", bits = 0x%08lx", (long) (* (int32 *) &fnum));
+	  fprintf (out, ", bits = 0x%08lx", (long) JPOOL_UINT (jcf, index));
 	break;
       }
     case CONSTANT_Double:
       {
-	jdouble dnum = JPOOL_DOUBLE (jcf, index);
-	fprintf (out, "%s%.20g", verbosity > 0 ? "Double " : "", dnum);
+	char s[30];
+	REAL_VALUE_TYPE r = JPOOL_DOUBLE (jcf, index);
+	REAL_VALUE_TO_DECIMAL (r, "%.20g", s);
+
+	fprintf (out, "%s%s", verbosity > 0 ? "Double " : "", s);
 	if (verbosity > 1)
 	  {
 	    int32 hi, lo;
===================================================================
Index: java/jcf-io.c
--- java/jcf-io.c	2001/12/03 19:13:40	1.33
+++ java/jcf-io.c	2002/04/04 10:02:53
@@ -25,10 +25,10 @@ The Free Software Foundation is independ
 #include "config.h"
 #include "system.h"
 
-#include "jcf.h"
 #include "tree.h"
 #include "toplev.h"
 #include "java-tree.h"
+#include "jcf.h"
 
 #include "zlib.h"
 
===================================================================
Index: java/jcf-path.c
--- java/jcf-path.c	2002/03/03 08:35:11	1.19
+++ java/jcf-path.c	2002/04/04 10:02:53
@@ -26,6 +26,8 @@ The Free Software Foundation is independ
 #include "config.h"
 #include "system.h"
 
+#include "machmode.h"
+#include "real.h"
 #include "jcf.h"
 
 /* Some boilerplate that really belongs in a header.  */
===================================================================
Index: java/jcf-write.c
--- java/jcf-write.c	2002/03/12 10:32:38	1.100
+++ java/jcf-write.c	2002/04/04 10:02:53
@@ -23,9 +23,9 @@ The Free Software Foundation is independ
 
 #include "config.h"
 #include "system.h"
-#include "jcf.h"
 #include "tree.h"
 #include "java-tree.h"
+#include "jcf.h"
 #include "obstack.h"
 #undef AND
 #include "rtl.h"
===================================================================
Index: java/jvgenmain.c
--- java/jvgenmain.c	2001/11/27 17:31:38	1.28
+++ java/jvgenmain.c	2002/04/04 10:02:54
@@ -27,9 +27,9 @@ The Free Software Foundation is independ
 #include "config.h"
 #include "system.h"
 #include "obstack.h"
-#include "jcf.h"
 #include "tree.h"
 #include "java-tree.h"
+#include "jcf.h"
 
 static char * do_mangle_classname PARAMS ((const char *string));
 
===================================================================
Index: java/mangle.c
--- java/mangle.c	2001/12/20 22:48:51	1.21
+++ java/mangle.c	2002/04/04 10:02:54
@@ -27,9 +27,9 @@ The Free Software Foundation is independ
 
 #include "config.h"
 #include "system.h"
-#include "jcf.h"
 #include "tree.h"
 #include "java-tree.h"
+#include "jcf.h"
 #include "obstack.h"
 #include "toplev.h"
 #include "obstack.h"
===================================================================
Index: java/mangle_name.c
--- java/mangle_name.c	2001/12/20 22:48:51	1.4
+++ java/mangle_name.c	2002/04/04 10:02:54
@@ -27,9 +27,9 @@ The Free Software Foundation is independ
 
 #include "config.h"
 #include "system.h"
-#include "jcf.h"
 #include "tree.h"
 #include "java-tree.h"
+#include "jcf.h"
 #include "obstack.h"
 #include "toplev.h"
 #include "obstack.h"