[PATCH] scheduler register pressure excessive stack space

[David Edelsohn]

>>>>> law  writes:

Jeff> Well, you haven't really provided any information to help me track it down,
Jeff> so it's hard to help you.

Jeff> I did manage to find a buglet in flow.c which could cause problems with this
Jeff> code.  Basically flow.c deletes the instruction containing the RETVAL note,
Jeff> thus leaving a libcall sequence unclosed.

	I tried bootstrapping with the GCC from CVS containing the version
of the patch you committed.  It failed with:

./xgcc -B./ -B/gcc/dje/install/powerpc-ibm-aix4.3.3.0-20020529/powerpc-ibm-aix4.
3.3.0/bin/ -isystem /gcc/dje/install/powerpc-ibm-aix4.3.3.0-20020529/powerpc-ibm
-aix4.3.3.0/include -isystem /gcc/dje/install/powerpc-ibm-aix4.3.3.0-20020529/po
werpc-ibm-aix4.3.3.0/sys-include -O2  -DIN_GCC    -W -Wall -Wwrite-strings -Wstr
ict-prototypes -Wmissing-prototypes -isystem ./include   -g -DHAVE_GTHR_DEFAULT 
-DIN_LIBGCC2 -D__GCC_FLOAT_NOT_NEEDED  -I. -I. -I/gcc/dje/src/gcc -I/gcc/dje/src
/gcc/. -I/gcc/dje/src/gcc/config -I/gcc/dje/src/gcc/../include -fexceptions -c /
gcc/dje/src/gcc/unwind-dw2.c -o libgcc/./unwind-dw2.o
In file included from /gcc/dje/src/gcc/unwind-dw2.c:1210:
/gcc/dje/src/gcc/unwind.inc: In function `_Unwind_RaiseException':
/gcc/dje/src/gcc/unwind.inc:132: Attempt to delete prologue/epilogue insn:
(insn/f 420 418 422 (set (mem:SI (plus:SI (reg/f:SI 1 r1)
                (const_int -220 [0xffffffffffffff24])) [1 S4 A8])
        (reg:SI 13 r13)) -1 (nil)
    (expr_list:REG_FRAME_RELATED_EXPR (set (mem:SI (plus:SI (reg/f:SI 1 r1)
                    (const_int -220 [0xffffffffffffff24])) [1 S4 A8])
            (reg:SI 13 r13))
        (nil)))
/gcc/dje/src/gcc/unwind.inc:132: Internal compiler error in propagate_one_insn, at flow.c:1633

	I am pretty sure this is similar to the failure I was seeing.  I
remember it occurring in flow.c, so I was sort of hoping that the latent
bugs being fixed by Honza and Richard might fix my problem as well.  Would
the buglet in flow.c you found affect this?

	I am running c-torture on the compiler I built to see if there are
any other, simpler testcases.

Thanks, David