The error is easily reproducible outside of the conditions of the kernel code, and, as I'm sure you'd appreciate, I have written an example that successfully duplicates the exact ICE as below. Attached is the significantly smaller preprocessed example, compressed at 3k, in lieu of the 46k compressed kernel driver :) System: i686-pc-linux-gnu (Fedora 2 release, kernel 2.6.7-1.494.2.2, athlon-xp 1800+) GCC Version: 3.5.0 20040801 (experimental) GCC ./configure Options: --enable-threads --with-cpu=athlon-xp --with-arch=athlon-xp --enable-languages=c,c++ Command: [root@epithumadze linux-2.6.8]# make modules V=1 ... make -f scripts/Makefile.build obj=sound/synth/emux gcc -Wp,-MD,sound/synth/emux/.emux_oss.o.d -nostdinc -iwithprefix include -D__KERNEL__ -Iinclude -Wall -Wstrict-prototypes -Wno-trigraphs -fno-strict-aliasing -fno-common -pipe -msoft-float -mpreferred-stack-boundary=2 -fno-unit-at-a-time -march=athlon -mregparm=3 -Iinclude/asm-i386/mach-default -O2 -fomit-frame-pointer -Wdeclaration-after-statement -DMODULE -DKBUILD_BASENAME=emux_oss -DKBUILD_MODNAME=snd_emux_synth -c -o sound/synth/emux/emux_oss.o sound/synth/emux/emux_oss.c Compiler Output: sound/synth/emux/emux_oss.c: In function `emuspec_control': sound/synth/emux/emux_oss.c:400: internal compiler error: tree check: expected integer_cst, have var_decl in fold, at fold-const.c:8855
Created attachment 6932 [details] Preprocessed bug example source
Confirmed : Search converges between 2004-06-20-trunk (#469) and 2004-06-22-trunk (#470). Reduced to: int main( int argc, char **arg ) { int R , N = 4; unsigned int A = 2; signed int B = 2; ((B >> N) & 1) ? 1 : 0; ((A >> N) & 1) ? 1 : 0; return 0; } Looks like someone forgets to check to make sure that the N is really a constant.
Subject: Re: [3.5 Regression] ICE: (A >> N) & 1 ? (1 << N) : 0 where A of type unsigned int pinskia at gcc dot gnu dot org wrote: > ------- Additional Comments From pinskia at gcc dot gnu dot org 2004-08-15 17:05 ------- > Confirmed > : Search converges between 2004-06-20-trunk (#469) and 2004-06-22-trunk (#470). > Reduced to: > int main( int argc, char **arg ) > { > int R , N = 4; > unsigned int A = 2; > signed int B = 2; > ((B >> N) & 1) ? 1 : 0; > ((A >> N) & 1) ? 1 : 0; > return 0; > } > > > Looks like someone forgets to check to make sure that the N is really a constant. > This patch should fixit but I don't have time to do a full bootstrap and regression test on the patch. ------------------------------------------------------------------- Index: fold-const.c =================================================================== RCS file: /cvs/gcc/gcc/gcc/fold-const.c,v retrieving revision 1.435 diff -c -p -r1.435 fold-const.c *** fold-const.c 12 Aug 2004 04:01:04 -0000 1.435 --- fold-const.c 15 Aug 2004 17:53:10 -0000 *************** fold (tree expr) *** 8874,8879 **** --- 8874,8880 ---- tree tem = TREE_OPERAND (arg0, 0); STRIP_NOPS (tem); if (TREE_CODE (tem) == RSHIFT_EXPR + && TREE_CODE (TREE_OPERAND (tem, 1)) == INTEGER_CST && (unsigned HOST_WIDE_INT) tree_log2 (arg1) == TREE_INT_CST_LOW (TREE_OPERAND (tem, 1))) return fold (build2 (BIT_AND_EXPR, type, -------------------------------------------------------------------
> This patch should fixit but I don't have time to do a full bootstrap and > regression test on the patch. > > ------------------------------------------------------------------- > Index: fold-const.c > =================================================================== > RCS file: /cvs/gcc/gcc/gcc/fold-const.c,v > retrieving revision 1.435 > diff -c -p -r1.435 fold-const.c > *** fold-const.c 12 Aug 2004 04:01:04 -0000 1.435 > --- fold-const.c 15 Aug 2004 17:53:10 -0000 > *************** fold (tree expr) > *** 8874,8879 **** > --- 8874,8880 ---- > tree tem = TREE_OPERAND (arg0, 0); > STRIP_NOPS (tem); > if (TREE_CODE (tem) == RSHIFT_EXPR > + && TREE_CODE (TREE_OPERAND (tem, 1)) == INTEGER_CST > && (unsigned HOST_WIDE_INT) tree_log2 (arg1) == > TREE_INT_CST_LOW (TREE_OPERAND (tem, 1))) > return fold (build2 (BIT_AND_EXPR, type, > ------------------------------------------------------------------- > When patched against the 20040801 tree that I had initially generated the error on, it successfully recompiled and properly generated the code. Thanks to both of you for the quick response time :)
Subject: Re: [3.5 Regression] ICE: (A >> N) & 1 ? (1 << N) : 0 where A of type unsigned int >>This patch should fixit but I don't have time to do a full bootstrap and >>regression test on the patch. Starting a bootstrap and regression test now, will commit as obvious if it succeeds. Thanks for investigating this. Paolo
Subject: Bug 17036 CVSROOT: /cvs/gcc Module name: gcc Changes by: bonzini@gcc.gnu.org 2004-08-17 07:40:23 Modified files: gcc : ChangeLog fold-const.c gcc/testsuite : ChangeLog Added files: gcc/testsuite/gcc.dg: pr17036-1.c Log message: gcc/ChangeLog: 2004-08-17 Paolo Bonzini <bonzini@gnu.org> PR middle-end/17036 * fold-const.c (fold): Check for integer operand when folding (A >> N) & 1 ? (1 << N) : 0. Fix from Andrew Pinski <pinskia@physics.uc.edu>. gcc/testsuite/ChangeLog: 2004-08-17 Paolo Bonzini <bonzini@gnu.org> * gcc.dg/pr17036-1.c: New test. Patches: http://gcc.gnu.org/cgi-bin/cvsweb.cgi/gcc/gcc/ChangeLog.diff?cvsroot=gcc&r1=2.4931&r2=2.4932 http://gcc.gnu.org/cgi-bin/cvsweb.cgi/gcc/gcc/fold-const.c.diff?cvsroot=gcc&r1=1.436&r2=1.437 http://gcc.gnu.org/cgi-bin/cvsweb.cgi/gcc/gcc/testsuite/ChangeLog.diff?cvsroot=gcc&r1=1.4145&r2=1.4146 http://gcc.gnu.org/cgi-bin/cvsweb.cgi/gcc/gcc/testsuite/gcc.dg/pr17036-1.c.diff?cvsroot=gcc&r1=NONE&r2=1.1
Fixed.