[committed] libffi patch for IA-64 denorms

[James E Wilson <wilson at specifix dot com>]

This addresses a problem reported in PR 26483, where denorm values
passed through libffi were accidentally rounded, when they should have
been passed through unchanged.

The problem here is that the code was written to use long double values
as a generic FP type.  Since the IA-64 architecture holds FP values in
registers in the long double format always, this is fine normally,
except for denorms, which get rounded inappropriately.  This was fixed
by rewriting the code to use macros, and appropriately typed local
variables, to avoid any FP type conversions.

While working on this, I noticed that the HFA (homogeneous
floating-point aggregate) support was also broken, and had to be fixed. 
This this patch is a bit bigger than my original patch.

This was tested with an ia64-linux bootstrap and make check, with and
without the patch.  There were 4 fewer libffi failures with the patch,
for the testcase I checked in last Friday.  There were no other
changes.  I also verified this by disassembling libffi/src/ia64/ffi.o,
and looking for fnorm (FP type conversion) instructions.  There are 8
such instructions in the unpatched toolchain.  There are 0 such
instructions in the patched toolchain.

I have added the patch to mainline.
--
Jim Wilson, GNU Tools Support, http://www.specifix.com

2006-04-12  James E Wilson  <wilson@specifix.com>

	PR libgcj/26483
	* src/ia64/ffi.c (stf_spill, ldf_fill): Rewrite as macros.
	(hfa_type_load): Call stf_spill.
	(hfa_type_store): Call ldf_fill.
	(ffi_call): Adjust calls to above routines.  Add local temps for
	macro result.
	
Index: ffi.c
===================================================================
--- ffi.c	(revision 112761)
+++ ffi.c	(working copy)
@@ -69,24 +69,19 @@ endian_adjust (void *addr, size_t len)
 #endif
 }
 
-/* Store VALUE to ADDR in the current cpu implementation's fp spill format.  */
+/* Store VALUE to ADDR in the current cpu implementation's fp spill format.
+   This is a macro instead of a function, so that it works for all 3 floating
+   point types without type conversions.  Type conversion to long double breaks
+   the denorm support.  */
 
-static inline void
-stf_spill(fpreg *addr, __float80 value)
-{
+#define stf_spill(addr, value)	\
   asm ("stf.spill %0 = %1%P0" : "=m" (*addr) : "f"(value));
-}
 
 /* Load a value from ADDR, which is in the current cpu implementation's
-   fp spill format.  */
+   fp spill format.  As above, this must also be a macro.  */
 
-static inline __float80
-ldf_fill(fpreg *addr)
-{
-  __float80 ret;
-  asm ("ldf.fill %0 = %1%P1" : "=f"(ret) : "m"(*addr));
-  return ret;
-}
+#define ldf_fill(result, addr)	\
+  asm ("ldf.fill %0 = %1%P1" : "=f"(result) : "m"(*addr));
 
 /* Return the size of the C type associated with with TYPE.  Which will
    be one of the FFI_IA64_TYPE_HFA_* values.  */
@@ -110,17 +105,20 @@ hfa_type_size (int type)
 /* Load from ADDR a value indicated by TYPE.  Which will be one of
    the FFI_IA64_TYPE_HFA_* values.  */
 
-static __float80
-hfa_type_load (int type, void *addr)
+static void
+hfa_type_load (fpreg *fpaddr, int type, void *addr)
 {
   switch (type)
     {
     case FFI_IA64_TYPE_HFA_FLOAT:
-      return *(float *) addr;
+      stf_spill (fpaddr, *(float *) addr);
+      return;
     case FFI_IA64_TYPE_HFA_DOUBLE:
-      return *(double *) addr;
+      stf_spill (fpaddr, *(double *) addr);
+      return;
     case FFI_IA64_TYPE_HFA_LDOUBLE:
-      return *(__float80 *) addr;
+      stf_spill (fpaddr, *(__float80 *) addr);
+      return;
     default:
       abort ();
     }
@@ -130,19 +128,31 @@ hfa_type_load (int type, void *addr)
    the FFI_IA64_TYPE_HFA_* values.  */
 
 static void
-hfa_type_store (int type, void *addr, __float80 value)
+hfa_type_store (int type, void *addr, fpreg *fpaddr)
 {
   switch (type)
     {
     case FFI_IA64_TYPE_HFA_FLOAT:
-      *(float *) addr = value;
-      break;
+      {
+	float result;
+	ldf_fill (result, fpaddr);
+	*(float *) addr = result;
+	break;
+      }
     case FFI_IA64_TYPE_HFA_DOUBLE:
-      *(double *) addr = value;
-      break;
+      {
+	double result;
+	ldf_fill (result, fpaddr);
+	*(double *) addr = result;
+	break;
+      }
     case FFI_IA64_TYPE_HFA_LDOUBLE:
-      *(__float80 *) addr = value;
-      break;
+      {
+	__float80 result;
+	ldf_fill (result, fpaddr);
+	*(__float80 *) addr = result;
+	break;
+      }
     default:
       abort ();
     }
@@ -351,8 +361,8 @@ ffi_call(ffi_cif *cif, void (*fn)(), voi
 		       && offset < size
 		       && gp_offset < 8 * 8)
 		  {
-		    stf_spill (&stack->fp_regs[fpcount],
-			       hfa_type_load (hfa_type, avalue[i] + offset));
+		    hfa_type_load (&stack->fp_regs[fpcount], hfa_type,
+				   avalue[i] + offset);
 		    offset += hfa_size;
 		    gp_offset += hfa_size;
 		    fpcount += 1;
@@ -475,9 +485,11 @@ ffi_closure_unix_inner (ffi_closure *clo
 	case FFI_TYPE_FLOAT:
 	  if (gpcount < 8 && fpcount < 8)
 	    {
-	      void *addr = &stack->fp_regs[fpcount++];
+	      fpreg *addr = &stack->fp_regs[fpcount++];
+	      float result;
 	      avalue[i] = addr;
-	      *(float *)addr = ldf_fill (addr);
+	      ldf_fill (result, addr);
+	      *(float *)addr = result;
 	    }
 	  else
 	    avalue[i] = endian_adjust(&stack->gp_regs[gpcount], 4);
@@ -487,9 +499,11 @@ ffi_closure_unix_inner (ffi_closure *clo
 	case FFI_TYPE_DOUBLE:
 	  if (gpcount < 8 && fpcount < 8)
 	    {
-	      void *addr = &stack->fp_regs[fpcount++];
+	      fpreg *addr = &stack->fp_regs[fpcount++];
+	      double result;
 	      avalue[i] = addr;
-	      *(double *)addr = ldf_fill (addr);
+	      ldf_fill (result, addr);
+	      *(double *)addr = result;
 	    }
 	  else
 	    avalue[i] = &stack->gp_regs[gpcount];
@@ -501,9 +515,11 @@ ffi_closure_unix_inner (ffi_closure *clo
 	    gpcount++;
 	  if (LDBL_MANT_DIG == 64 && gpcount < 8 && fpcount < 8)
 	    {
-	      void *addr = &stack->fp_regs[fpcount++];
+	      fpreg *addr = &stack->fp_regs[fpcount++];
+	      __float80 result;
 	      avalue[i] = addr;
-	      *(__float80 *)addr = ldf_fill (addr);
+	      ldf_fill (result, addr);
+	      *(__float80 *)addr = result;
 	    }
 	  else
 	    avalue[i] = &stack->gp_regs[gpcount];
@@ -533,8 +549,8 @@ ffi_closure_unix_inner (ffi_closure *clo
 		       && offset < size
 		       && gp_offset < 8 * 8)
 		  {
-		    hfa_type_store (hfa_type, addr + offset, 
-				    ldf_fill (&stack->fp_regs[fpcount]));
+		    hfa_type_store (hfa_type, addr + offset,
+				    &stack->fp_regs[fpcount]);
 		    offset += hfa_size;
 		    gp_offset += hfa_size;
 		    fpcount += 1;