Bug 24997 - [4.1 regression] ICE with -ftree-vectorize
Summary: [4.1 regression] ICE with -ftree-vectorize
Alias: None
Product: gcc
Classification: Unclassified
Component: target (show other bugs)
Version: 4.1.0
: P3 normal
Target Milestone: 4.1.0
Assignee: Alan Modra
URL: http://gcc.gnu.org/ml/gcc-patches/200...
Keywords: ice-on-valid-code
Depends on: 25212
  Show dependency treegraph
Reported: 2005-11-23 07:11 UTC by James A. Morrison
Modified: 2005-12-12 10:05 UTC (History)
4 users (show)

See Also:
Host: powerpc-linux
Target: powerpc-linux
Build: powerpc-linux
Known to work: 4.0.3 4.1.0 4.2.0
Known to fail:
Last reconfirmed: 2005-11-24 10:21:53

preprocessed source (50.16 KB, text/plain)
2005-11-23 16:05 UTC, James A. Morrison

Note You need to log in before you can comment on or make changes to this bug.
Description James A. Morrison 2005-11-23 07:11:39 UTC
wmadec.c ICE's with:
/home/jim/cvs/mplayer/ffmpeg/libavcodec/wmadec.c:1193: error: unrecognizable insn:
(insn 4809 3267 2762 272 /home/jim/cvs/mplayer/ffmpeg/libavcodec/wmadec.c:1117 (set (reg:SI 9 9)
        (and:SI (plus:SI (reg:SI 27 27)
                (reg:SI 11 11 [orig:284 base_off.641 ] [284]))
            (const_int -16 [0xfffffffffffffff0]))) -1 (nil)
/home/jim/cvs/mplayer/ffmpeg/libavcodec/wmadec.c:1193: internal compiler error: in extract_insn, at recog.c:2084

the command line is:
 /usr/lib/gcc-snapshot/bin/gcc -O3 -g -Wall -Wno-switch  -maltivec -mabi=altivec  -DHAVE_AV_CONFIG_H -I.. -I/home/jim/cvs/mplayer/ffmpeg/libavutil -D_FILE_OFFSET_BITS=64 -D_LARGEFILE_SOURCE -D_GNU_SOURCE  -c -o wmadec.o /home/jim/cvs/mplayer/ffmpeg/libavcodec/wmadec.c
The gcc-snapshot is from the debian package version 20051122-1, the ice also occurs in earlier versions of the package.
Comment 1 Alan Modra 2005-11-23 12:20:53 UTC
Please attach preprocessed source.
Comment 2 Andrew Pinski 2005-11-23 13:15:54 UTC
Hmm, r9, wtf.  Any ways, we need the .i file :).
Comment 3 James A. Morrison 2005-11-23 16:05:16 UTC
Created attachment 10327 [details]
preprocessed source

Here is the preprocessed source.  Sorry, I haven't had a change to run delta on it yet, nor can I reproduce this ICE without using -O3 -ftree-vectorize.
Comment 4 Andrew Pinski 2005-11-23 16:14:45 UTC
Comment 5 Andrew Pinski 2005-11-23 18:54:37 UTC
I will finish reducing this after class (I stopped and started reducing a different testcase anyways).
Comment 6 Andrew Pinski 2005-11-23 21:38:51 UTC
Actually this bug is too funny to reduce, reload is too funny for the life of me.

Compiling with -da with a semi reduced testcases makes it pass.
Comment 7 Alan Modra 2005-11-24 06:51:22 UTC
I'm not surprised that reducing this testcase isn't easy.  Register pressure is needed to trigger the bug.  The problem is that in
(insn:HI 2761 3601 2769 279 wmadec.c:1117 (set (reg:V4SF 280 [ vect_var_.655 ])
        (mem:V4SF (and:SI (plus:SI (reg/f:SI 1824)
                    (reg:SI 284 [ base_off.641 ]))
                (const_int -16 [0xfffffffffffffff0])) [4 S16 A128])) 628 {altivec_lvx_v4sf} (insn_list:REG_DEP_TRUE 2757 (nil))
    (expr_list:REG_DEAD (reg:SI 284 [ base_off.641 ])

reg 1824 doesn't get a hard register, so is replaced with a stack slot

(insn:HI 2761 3601 2769 279 wmadec.c:1117 (set (reg:V4SF 89 12 [orig:280 vect_var_.655 ] [280])
        (mem:V4SF (and:SI (plus:SI (plus:SI (reg/f:SI 1 1)
                        (reg:SI 11 11 [orig:284 base_off.641 ] [284]))
                    (const_int 16512 [0x4080]))
                (const_int -16 [0xfffffffffffffff0])) [4 S16 A128])) 628 {altivec_lvx_v4sf} (insn_list:REG_DEP_TRUE 2757 (nil))
    (expr_list:REG_DEAD (reg:SI 11 11 [orig:284 base_off.641 ] [284])

reload just tries to put the mem address, including the AND, into a reg.  It needs to be taught that the AND should stay with the insn.  The real address is the part inside the AND.
Comment 8 Alan Modra 2005-11-24 07:04:31 UTC
Oops, no, that's not a stack slot.  That's a stack slot gone wrong.  Where's the MEM for the slot?
Comment 9 Alan Modra 2005-11-24 08:59:04 UTC
Indeed it's not a stack slot.  Instead, reg 1824 is an invariant, equal to sfp+16512.  Code in reload1.c:eliminate_regs_1 substitutes r1 for sfp, then reorganises the order of the additions, so we get (r1+r11)+const not (r1+const)+r11.  This of course isn't a valid address on powerpc, where indexed addresses can't take an offset.

In case it isn't obvious, I'm a little lost.  Hopefully my analysis will let someone more familiar with reload sort this out.
Comment 10 Alan Modra 2005-11-24 10:21:53 UTC
I think I've sorted this out after all..  Testing a fix.
Comment 11 Alan Modra 2005-11-28 03:52:05 UTC
Subject: Bug 24997

Author: amodra
Date: Mon Nov 28 03:52:01 2005
New Revision: 107591

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=107591
	PR target/24997
	* config/rs6000/rs6000.c (legitimate_indexed_address_p): Allow pattern
	generated by reload.
	* config/rs6000/predicates.md (indexed_or_indirect_operand): Use
	(indexed_or_indirect_address): Don't test for base reg.  Call
	address_operand last.  Make it a special predicate.


Comment 12 Alan Modra 2005-12-07 05:50:53 UTC
pr25212 has a more complete fix that I propose applying to the 4.1 branch.
Comment 13 Alan Modra 2005-12-12 10:05:56 UTC
Fixed 4.1 by pr25212 commit.