Patchwork Reload related segfaults

Some recent patch has exposed a reload bug.  I'm seeing

libtool: compile:  /home/amodra/build/gcc-curr/./gcc/xgcc -B/home/amodra/build/gcc-curr/./gcc/ -B/home/amodra/gnu/powerpc-linux/bin/ -B/home/amodra/gnu/powerpc-linux/lib/ -isystem /home/amodra/gnu/powerpc-linux/include -isystem /home/amodra/gnu/powerpc-linux/sys-include -m64 -fPIC -mstrict-align -DHAVE_CONFIG_H -I. -I/home/amodra/src/gcc-current/libjava/classpath/native/fdlibm -I../../include -fexceptions -fasynchronous-unwind-tables -g -O2 -m64 -fPIC -mstrict-align -MT sf_fabs.lo -MD -MP -MF .deps/sf_fabs.Tpo -c /home/amodra/src/gcc-current/libjava/classpath/native/fdlibm/sf_fabs.c  -fPIC -DPIC -o .libs/sf_fabs.o
/home/amodra/src/gcc-current/libjava/classpath/native/fdlibm/sf_fabs.c: In function 'fabsf':
/home/amodra/src/gcc-current/libjava/classpath/native/fdlibm/sf_fabs.c:33:1: internal compiler error: Segmentation fault

(insn 11 9 15 2 (parallel [
            (set (subreg:SI (reg:SF 123 [ <retval> ]) 0)
                (and:SI (subreg:SI (reg:SF 33 1 [ x ]) 0)
                    (const_int 2147483647 [0x7fffffff])))
            (clobber (scratch:CC))
        ]) /home/amodra/src/gcc-current/libjava/classpath/native/fdlibm/sf_fabs.c:32 146 {andsi3_mc}
     (expr_list:REG_DEAD (reg:SF 33 1 [ x ])
        (nil)))

Reloads for insn # 11
Reload 0: GENERAL_REGS, RELOAD_FOR_INPUT_ADDRESS (opnum = 0), can't combine, secondary_reload_p
	reload_reg_rtx: (reg:SF 8 8)
Reload 1: GENERAL_REGS, RELOAD_FOR_OUTPUT_ADDRESS (opnum = 0), can't combine, secondary_reload_p
	reload_reg_rtx: (reg:SF 8 8)
Reload 2: reload_in (SF) = (reg:SF 33 1 [orig:123 <retval> ] [123])
	reload_out (SF) = (reg:SF 33 1 [orig:123 <retval> ] [123])
	NON_FLOAT_REGS, RELOAD_OTHER (opnum = 0)
	reload_in_reg: (reg:SF 33 1 [orig:123 <retval> ] [123])
	reload_out_reg: (reg:SF 33 1 [orig:123 <retval> ] [123])
	reload_reg_rtx: (reg:SF 10 10)
	secondary_in_reload = 0, secondary_out_reload = 1

Reload 3: reload_out (SI) = (subreg:SI (reg:SF 33 1 [orig:123 <retval> ] [123]) 0)
	GENERAL_REGS, RELOAD_FOR_OUTPUT (opnum = 0)
	reload_out_reg: (subreg:SI (reg:SF 33 1 [orig:123 <retval> ] [123]) 0)
	reload_reg_rtx: (reg:SI 9 9)
Reload 4: reload_in (SF) = (reg:SF 33 1 [ x ])
	GENERAL_REGS, RELOAD_FOR_INPUT (opnum = 1)
	reload_in_reg: (reg:SF 33 1 [ x ])
	reload_reg_rtx: (reg:SF 8 8)

Notice reload 3 has a subreg for reload_out.  We wind up in the
gen_reload code shown below, and try to use REGNO on a subreg, leading
to

(gdb) p debug_rtx(out)
(reg:DI -166922608)

Or at least, that's what you get after fixing print_rtx to not
segfault..  Bootstrap and regression test powerpc-linux in progress.
OK to apply, assuming no regressions?

	* reload1.c (gen_reload): Don't use REGNO on SUBREGs.
	* print-rtl.c (print_rtx): Don't segfault on negative regno.

Note that s/REGNO (in_rtx)/value/ in print_rtx is reasonable given
that REG in rtl.def has a format of "i00", so XINT (in_rtx, i) is
always XINT (in_rtx, 0) for a reg, which is equivalent to REGNO apart
from signedness.

From: Alan Modra <amodra@gmail.com>
Date: Thu, 27 Oct 2011 13:29:56 +1030

> Some recent patch has exposed a reload bug.  I'm seeing

I think this might be a side effect or Eric's recent changes,
CC:'d.

From: David Miller <davem@davemloft.net>
Date: Wed, 26 Oct 2011 23:07:11 -0400 (EDT)

> From: Alan Modra <amodra@gmail.com>
> Date: Thu, 27 Oct 2011 13:29:56 +1030
> 
>> Some recent patch has exposed a reload bug.  I'm seeing
> 
> I think this might be a side effect or Eric's recent changes,
> CC:'d.

Eric, I'm seeing a similar segmentation fault in reload on sparc64.
But it's in a slightly different place than Alan's crash.

Simply compile gcc.target/ultrasp12.c with "-m64 -O2 -mcpu=ultrasparc -mvis"
to see this.

The crash is in find_valid_class() called from push_reload(), via this
code block around line 1184 of reload.c:

      enum reg_class in_out_class
	= find_valid_class (outmode, GET_MODE (SUBREG_REG (out)),
			    subreg_regno_offset (REGNO (SUBREG_REG (out)),
						 GET_MODE (SUBREG_REG (out)),
						 SUBREG_BYTE (out),
						 GET_MODE (out)),
			    REGNO (SUBREG_REG (out)));

'out' is:

(subreg:DI (reg/v:V4QI 50 %f18 [orig:314 s2hi4_ ] [314]) 0)

so subreg_regno_offset() returns -1, and find_valid_class() isn't too happy
about getting "-1" for it's 'n' argument.

I suspect the test that you removed in order to fix rtl-optimization/46603
would guard against this happening.

And indeed, I confirmed that backing out the rtl-optimization/46603
fix makes the segmentation fault go away.

Ping http://gcc.gnu.org/ml/gcc-patches/2011-10/msg02429.html

Eric fixed the bootstrap breakage by another patch, but there is a
fairly obvious bug in gen_reload fixed by my patch.  I gave enough
contect in the diff that you don't even need to look at the file. :)

> OK to apply, assuming no regressions?
>
> 	* reload1.c (gen_reload): Don't use REGNO on SUBREGs.
> 	* print-rtl.c (print_rtx): Don't segfault on negative regno.

OK, thanks.