Created attachment 12056 [details]
small testcase to trigger the bug

-O3 works here while -O2 fails - but, in any case, confirmed.

I was on gcc 4.2 btw.

A regression hunt using an arm-ep93xx-linux-gnueabi cross compiler on powerpc-linux, with the testcase minimal.c, identified the following patch:

    http://gcc.gnu.org/viewcvs?view=rev&rev=105121

    r105121 | kazu | 2005-10-08 18:17:20 +0000 (Sat, 08 Oct 2005)

This is a large merge from the csl-arm-branch to mainline.  The test does not fail for a cross compiler built from revision 104998, the last revision on that branch before the merge, suggesting that his is a problem with the merge itself. 

This is looking like a latent reload bug.
I don't see anything in r105121 that could cause this bug. I can reproduce it on non-linux and non-eabi arm targets.

I am getting exactly the same here on ppc (so it is not just arm specific):

$ c++ -v
Using built-in specs.
Target: powerpc-linux-gnu
Configured with: ../src/configure -v --enable-languages=c,c++,fortran,objc,obj-c++,treelang --prefix=/usr --enable-shared --with-system-zlib --libexecdir=/usr/lib --without-included-gettext --enable-threads=posix --enable-nls --program-suffix=-4.1 --enable-__cxa_atexit --enable-clocale=gnu --enable-libstdcxx-debug --enable-mpfr --disable-softfloat --enable-targets=powerpc-linux,powerpc64-linux --with-cpu=default32 --enable-checking=release powerpc-linux-gnu
Thread model: posix
gcc version 4.1.2 20060901 (prerelease) (Debian 4.1.1-13)

Subject: Re:  [4.1/4.2 regression] ICE in extract_insn,
	at recog.c:2084 (unrecognizable insn) [arm]

On Wed, 2006-09-20 at 16:38 +0000, bugreports at nn7 dot de wrote:
> 
> ------- Comment #6 from bugreports at nn7 dot de  2006-09-20 16:38 -------
> I am getting exactly the same here on ppc (so it is not just arm specific):

Yes it is.  The bug you are getting with PPC is a different bug and
should be filed separately.

Thanks,
Andrew Pinski

gcc/testsuite/gcc.dg/compat/scalar-return-4_y.c fails with this ICE.  Do you
know if this is the same or a different problem?  This is on ARM old ABI, fails with 4.1 and 4.2.  4.0 works.

../../src/gcc/testsuite/gcc.dg/compat/scalar-return-4_y.c:41: error: unrecognizable insn:
(insn 39 20 21 3 (set (reg:HI 0 r0)
        (subreg:HI (mem/c:CQI (plus:SI (reg/f:SI 11 fp)
                    (const_int -16 [0xfffffff0])) [0 S2 A8]) 0)) -1 (nil)
    (nil))
../../src/gcc/testsuite/gcc.dg/compat/scalar-return-4_y.c:41: internal compiler error: in extract_insn, at recog.c:2077

And gcc.c-torture/execute/20050121-1.c

(sid)tbm@ALL6500:~/build/gcc-snapshot-20060922/build/gcc$ ./xgcc -B. -o x -c ../../src/gcc/testsuite/gcc.c-torture/execute/20050121-1.c
../../src/gcc/testsuite/gcc.c-torture/execute/20050121-1.c: In function 'foo_char':
../../src/gcc/testsuite/gcc.c-torture/execute/20050121-1.c:33: error: unrecognizable insn:
(insn 55 36 37 2 (set (reg:HI 0 r0)
        (subreg:HI (mem/c:CQI (plus:SI (reg/f:SI 11 fp)
                    (const_int -24 [0xffffffe8])) [0 S2 A8]) 0)) -1 (nil)
    (nil))
../../src/gcc/testsuite/gcc.c-torture/execute/20050121-1.c:33: internal compiler error: in extract_insn, at recog.c:2077

I ICE in comment #10 looks like it may be a different problem.

Before reload the problematic instruction is:

(set (mem:HI (something))
     (subreg:HI (reg:SI 177)))

Reload decides to reload the second operand into a register.
In find_reloads_subreg_address, reg_equiv_memory_loc tells us that reg 117
is equivalent to

(mem:SI (plus:SI (mult:SI (reg:SI r1) (const_int 16))
                 (reg:SI r6)))

This is a valid SImode address, but not a valid HImode address.
we change the mem:SI to mem:HI, then call find_reloads_address to fixup the
result.  However the comment above find_reloads_address says:

   Note that there is no verification that the address will be valid after
   this routine does its work.  Instead, we rely on the fact that the address
   was valid when reload started.  So we need only undo things that reload
   could have broken.  These are wrong register types, pseudos not allocated
   to a hard register, and frame pointer elimination.  */

No mention of having broken the address by changing the mode of the mem.
The new HImode mem passes through find_reloads_address unmolested, causing
an ICE later when we try to assign it to the reload reg.

My guess would be that we need to be calling something other than 
find_reloads_address after replacing the subreg with the mem, but don't know 
reload well enough to say what.

Created attachment 13445 [details]
Proposed patch

I'm testing this patch on i686-pc-linux-gnu.  If somebody can test it on ARM, that would be great.

Subject: Bug 28675

Author: ian
Date: Fri Apr 27 05:11:47 2007
New Revision: 124211

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=124211
Log:
gcc/:
	PR target/28675
	* reload.c (find_reloads_subreg_address): If the address was valid
	in the original mode but not in the new mode, reload the whole
	address.
testsuite/:
	PR target/28675
	* gcc.c-torture/compile/pr28675.c: New test.

Added:
    trunk/gcc/testsuite/gcc.c-torture/compile/pr28675.c
Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/reload.c
    trunk/gcc/testsuite/ChangeLog

Subject: Bug 28675

Author: ian
Date: Fri Apr 27 05:20:02 2007
New Revision: 124212

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=124212
Log:
./:
	PR target/28675
	* reload.c (find_reloads_subreg_address): If the address was valid
	in the original mode but not in the new mode, reload the whole
	address.
testsuite/:
	PR target/28675
	* gcc.c-torture/compile/pr28675.c: New test.

Added:
    branches/gcc-4_2-branch/gcc/testsuite/gcc.c-torture/compile/pr28675.c
Modified:
    branches/gcc-4_2-branch/gcc/ChangeLog
    branches/gcc-4_2-branch/gcc/reload.c
    branches/gcc-4_2-branch/gcc/testsuite/ChangeLog

Subject: Bug 28675

Author: ian
Date: Fri Apr 27 05:45:53 2007
New Revision: 124213

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=124213
Log:
./:
	PR target/28675
	* reload.c (find_reloads_subreg_address): If the address was valid
	in the original mode but not in the new mode, reload the whole
	address.
testsuite/:
	PR target/28675
	* gcc.c-torture/compile/pr28675.c: New test.

Added:
    branches/gcc-4_1-branch/gcc/testsuite/gcc.c-torture/compile/pr28675.c
Modified:
    branches/gcc-4_1-branch/gcc/ChangeLog
    branches/gcc-4_1-branch/gcc/reload.c
    branches/gcc-4_1-branch/gcc/testsuite/ChangeLog

Fixed on 4.1 branch, 4.2 branch, and mainline.