This is the mail archive of the mailing list for the GCC project.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: real.c on unicosmk

On Thu, May 16, 2002 at 10:33:15AM +0200, Roman Lechtchinsky wrote:
> On Wed, 15 May 2002, Zack Weinberg wrote:
> > Do you have a 32-bit and a 64-bit type?
> Yes.
> > If so, a patch to make real.c
> > work on your architecture is feasible -- it's a bit more involved than
> > I feel like doing right now, but it can be done.
> Does it require some fundamental changes or just a lot of tedious typing?
> I could do the latter.

Just a lot of tedious typing.  Lemme get you started.  Apply the
appended patch.  Go through the entire of real.c and change all the
uses of 2*NE to EBYTES.  Then inspect all the other uses of NE to make
sure they don't assume that EMUSHORT/UEMUSHORT are 16 bits.  You
should probably also take a close look at NI, NBITS, etc.  I'll help
as needed.

Bonus points if you can persuade EMUSHORT to be 16 bits on 32-bit
architectures and 32 bits on 64-bit architectures (this would be most
efficient on both).


Index: real.c
--- real.c	12 May 2002 17:09:23 -0000	1.71
+++ real.c	17 May 2002 04:06:01 -0000
@@ -171,21 +171,6 @@ unknown arithmetic type
-/* If no 16-bit type has been found and the compiler is GCC, try HImode.  */
-#if defined(__GNUC__) && EMUSHORT_SIZE != 16
-typedef int HItype __attribute__ ((mode (HI)));
-typedef unsigned int UHItype __attribute__ ((mode (HI)));
-#undef EMUSHORT
-#define EMUSHORT HItype
-#define UEMUSHORT UHItype
-#define EMUSHORT_SIZE 16
-#define EMULONG_SIZE 32
-#define UEMUSHORT unsigned EMUSHORT
 #define EMULONG short
@@ -204,36 +189,38 @@ typedef unsigned int UHItype __attribute
-#if EMUSHORT_SIZE != 16
-  #error "The host interface doesn't work if no 16-bit size exists."
 /* Calculate the size of the generic "e" type.  This always has
    identical in-memory size to REAL_VALUE_TYPE.  The sizes are supposed
    to be the same as well, but when REAL_VALUE_TYPE_SIZE is not evenly
    divisible by HOST_BITS_PER_WIDE_INT we have some padding in
-   There are only two supported sizes: ten and six 16-bit words (160
-   or 96 bits).  */
+   There are only two supported sizes: 160 and 96 bits.  */
 /* TFmode */
-# define NE 10
+# define EBITS 160 
 # define MAXDECEXP 4932
 # define MINDECEXP -4977
-# define NE 6
+# define EBITS 96
 # define MAXDECEXP 4932
 # define MINDECEXP -4956
-/* Fail compilation if 2*NE is not the appropriate size.
+/* There must not have been any rounding.  */
+ #error "EMUSHORT_SIZE does not divide EBITS"
+/* Fail compilation if EBYTES is not the appropriate size.
    If HOST_BITS_PER_WIDE_INT is 64, we're going to have padding
    at the end of the array, because neither 96 nor 160 is
    evenly divisible by 64.  */
 struct compile_test_dummy {
   char twice_NE_must_equal_sizeof_REAL_VALUE_TYPE
-  [(sizeof (REAL_VALUE_TYPE) >= 2*NE) ? 1 : -1];
+  [(sizeof (REAL_VALUE_TYPE) >= EBYTES) ? 1 : -1];
 /* Construct macros to translate between REAL_VALUE_TYPE and e type.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]