Created attachment 15097 [details]
preprocessed source of file producing ICE

Confirmed.

The testcase:

float test(unsigned int x)
{
  return (float)x;
}

gcc -O2 -mno-80387

t.c: In function âtestâ:
t.c:4: error: unrecognizable insn:
(insn 20 19 21 5 t.c:2 (set (reg:SF 60)
        (plus:SF (reg:SF 61)
            (reg:SF 61))) -1 (nil))
t.c:4: internal compiler error: in extract_insn, at recog.c:1990
Please submit a full bug report,

Mine.

Following patch fixes ICE (and avoids nasty runtime/code-size regression for x87 math where we don't use fildll):

--cut here--
Index: config/i386/i386.md
===================================================================
--- config/i386/i386.md (revision 132088)
+++ config/i386/i386.md (working copy)
@@ -5306,9 +5306,9 @@
 (define_expand "floatunssisf2"
   [(use (match_operand:SF 0 "register_operand" ""))
    (use (match_operand:SI 1 "nonimmediate_operand" ""))]
-  "!TARGET_64BIT"
+  "!TARGET_64BIT && TARGET_SSE_MATH"
 {
-  if (TARGET_SSE_MATH && TARGET_SSE2)
+  if (TARGET_SSE2)
     ix86_expand_convert_uns_sisf_sse (operands[0], operands[1]);
   else
     x86_emit_floatuns (operands);
--cut here--

but there is another problem for soft-float, where neither signed or unsigned float pattern is provided. Instead of calling __floatunssfsi() directly for soft-float targets, we expand as:

test:
        pushl   %ebp
        movl    %esp, %ebp
        subl    $24, %esp
        movl    8(%ebp), %edx
        testl   %edx, %edx
        js      .L3
        movl    %edx, (%esp)
        call    __floatsisf
        leave
        ret
.L3:
        movl    %edx, %eax
        andl    $1, %eax
        shrl    %edx
        orl     %edx, %eax
        movl    %eax, (%esp)
        call    __floatsisf
        movl    %eax, 4(%esp)
        movl    %eax, (%esp)
        call    __addsf3
        leave
        ret

This is a huge code-size (and performance?) regression from 4.2 where we generated:

test:
        pushl   %ebp
        movl    %esp, %ebp
        subl    $24, %esp
        movl    8(%ebp), %eax
        movl    %eax, (%esp)
        call    __floatunsisf
        leave
        ret

This is probably due to:

	2007-09-05  Janis Johnson  <janis187@us.ibm.com>

	* optabs.c (expand_float): Convert unsigned integer as signed only
	if it provides sufficient accuracy; add mode argument to real_2expN.
	...

optabs.c, line 5150:

--cut here--
  /* Unsigned integer, and no way to convert directly.  Convert as signed,
     then unconditionally adjust the result.  For decimal float values we
     do this only if we have already determined that a signed conversion
     provides sufficient accuracy.  */
  if (unsignedp && (can_do_signed || !DECIMAL_FLOAT_MODE_P (GET_MODE (to))))
--cht here--

Hm? We shortcut with !DECIMAL_FLOAT_MODE_P (...) here. Janis?

Subject: Bug 35083

Author: uros
Date: Tue Feb  5 11:27:41 2008
New Revision: 132114

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=132114
Log:
	PR target/35083
	* config/i386/i386.md (floatunsisf2): Enable for TARGET_SSE_MATH only.
	Call ix86_expand_convert_uns_sisf_sse for TARGET_SSE2.

testsuite/ChangeLog:

	PR target/35083
	* gcc.target/i386/pr35083.c: New test.


Added:
    trunk/gcc/testsuite/gcc.target/i386/pr35083.c
Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/config/i386/i386.md
    trunk/gcc/testsuite/ChangeLog

This is the diff of expand_float() between gcc-4.2 and gcc-4.3. The relevant part is logic at the top of the diff that has changed substantially:

--- 222 2008-02-05 14:52:52.000000000 +0100
+++ 111 2008-02-05 14:52:32.000000000 +0100
@@ -52,11 +52,10 @@
          }
       }
 
-  /* Unsigned integer, and no way to convert directly.  Convert as signed,
-     then unconditionally adjust the result.  For decimal float values we
-     do this only if we have already determined that a signed conversion
-     provides sufficient accuracy.  */
-  if (unsignedp && (can_do_signed || !DECIMAL_FLOAT_MODE_P (GET_MODE (to))))
+  /* Unsigned integer, and no way to convert directly.  For binary
+     floating point modes, convert as signed, then conditionally adjust
+     the result.  */
+  if (unsignedp && can_do_signed && !DECIMAL_FLOAT_MODE_P (GET_MODE (to)))
     {
       rtx label = gen_label_rtx ();
       rtx temp;
@@ -147,7 +146,7 @@
                               0, label);
 
 
-      real_2expN (&offset, GET_MODE_BITSIZE (GET_MODE (from)), fmode);
+      real_2expN (&offset, GET_MODE_BITSIZE (GET_MODE (from)));
       temp = expand_binop (fmode, add_optab, target,
                           CONST_DOUBLE_FROM_REAL_VALUE (offset, fmode),
                           target, 0, OPTAB_LIB_WIDEN);
@@ -169,7 +168,7 @@
       if (GET_MODE_SIZE (GET_MODE (from)) < GET_MODE_SIZE (SImode))
        from = convert_to_mode (SImode, from, unsignedp);
 
-      libfunc = convert_optab_libfunc (tab, GET_MODE (to), GET_MODE (from));
+      libfunc = tab->handlers[GET_MODE (to)][GET_MODE (from)].libfunc;
       gcc_assert (libfunc);
 
       start_sequence ();

Subject: Bug 35083

Author: uros
Date: Wed Feb  6 10:45:29 2008
New Revision: 132144

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=132144
Log:
	PR target/35083
	* optabs.c (expand_float): Do not check for decimal modes when
	expanding unsigned integer through signed conversion.


Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/optabs.c

Fixed.

Thanks Uros, i386-rtems*-gcc now bootstraps again.