$ gcc/xgcc -Bgcc/ ../gcc/testsuite/gcc.dg/graphite/run-id-pr47593.c -O2 -fgraphite-identity -lm -o ./run-id-pr47593.exe -g $ ./run-id-pr47593.exe Segmentation fault Program received signal SIGSEGV, Segmentation fault. 0x800004e6 in bar () at ../gcc/testsuite/gcc.dg/graphite/run-id-pr47593.c:21 21 return ***(int ***)p; (gdb) p p $1 = 0xeffff594 "\200\r\266\b" (gdb) p {int**}p $2 = (int **) 0x800db608 (gdb) p *$ Cannot access memory at address 0x800db608 (gdb) i local f = {p = 0x800db608, q = 0x0} j = 4 p = 0xeffff5c4 "\200\r\266\b" x = <synthetic pointer> y = 0xeffff5c0 (gdb) disass Dump of assembler code for function bar: 0x800004d8 <+0>: linkw %fp,#-12 0x800004dc <+4>: moveq #4,%d0 0x800004de <+6>: movel %d0,%fp@(-12) 0x800004e2 <+10>: moveal %fp@(-4),%a0 => 0x800004e6 <+14>: moveal %a0@,%a0 0x800004e8 <+16>: movel %a0@,%d0 0x800004ea <+18>: unlk %fp 0x800004ec <+20>: rts End of assembler dump. Before r224020: (gdb) disass bar Dump of assembler code for function bar: 0x800004d8 <+0>: linkw %fp,#-24 0x800004dc <+4>: moveq #1,%d0 0x800004de <+6>: unlk %fp 0x800004e0 <+8>: rts End of assembler dump. 05032b10839cf0498c992c819bf2358e86c22bb0 is the first bad commit commit 05032b10839cf0498c992c819bf2358e86c22bb0 Author: amker <amker@138bc75d-0d04-0410-961f-82ee72b054a4> Date: Tue Jun 2 10:19:18 2015 +0000 PR tree-optimization/48052 ... git-svn-id: svn+ssh://gcc.gnu.org/svn/gcc/trunk@224020 138bc75d-0d04-0410-961f-82ee72b054a4
Confirmed.
on x86_64 with -m32 btw
Interesting case. I know nothing about graphite, but I guess the case is fixed in patch for PR47653 by skipping graphite transofromation on loops in specific forms. The skip mechanism is kind of nullified by the overflow change. But as far as I can see, we shouldn't skip graphite transformation on the case in the first place because it looks good to me. See dumps before graphite pass: bar () { _Bool j_lsm.8; int j_lsm.7; int * y; int * x; char * p; int i; int j; struct Foo f; unsigned int j.2_11; int j.1_13; int * * _15; int * _16; int _17; int j.0_31; <bb 2>: i = 1; x = &i; y = &j; f.p = &y; f.q = &x; <bb 3>: # p_27 = PHI <&f(2), p_12(4)> # j.0_31 = PHI <0(2), j.1_13(4)> p_12 = p_27 + 1; j.1_13 = j.0_31 + 1; j.2_11 = (unsigned int) j.1_13; if (j.2_11 <= 3) goto <bb 4>; else goto <bb 5>; <bb 4>: goto <bb 3>; <bb 5>: p_25 = &f + 4; j = 4; _15 = MEM[(int * * *)p_25]; _16 = *_15; _17 = *_16; f ={v} {CLOBBER}; j ={v} {CLOBBER}; i ={v} {CLOBBER}; x ={v} {CLOBBER}; y ={v} {CLOBBER}; return _17; } graphite transforms it infor below shape: bar () { _Bool j_lsm.8; int j_lsm.7; int * y; int * x; char * p; int i; int j; struct Foo f; int * * _15; int * _16; int _17; <bb 2>: i = 1; x = &i; y = &j; f.p = &y; f.q = &x; p_25 = &f + 4; j = 4; _15 = MEM[(int * * *)p_25]; _16 = *_15; _17 = *_16; f ={v} {CLOBBER}; j ={v} {CLOBBER}; i ={v} {CLOBBER}; x ={v} {CLOBBER}; y ={v} {CLOBBER}; return _17; } It is dce4 pass that deletes the bounch of initialization code at beginning of bb 2. dce4.dump is as below: bar () { _Bool j_lsm.8; int j_lsm.7; int * y; int * x; char * p; int i; int j; struct Foo f; int * * _15; int * _16; int _17; <bb 2>: p_25 = &f + 4; j = 4; _15 = MEM[(int * * *)p_25]; _16 = *_15; _17 = *_16; f ={v} {CLOBBER}; j ={v} {CLOBBER}; i ={v} {CLOBBER}; x ={v} {CLOBBER}; y ={v} {CLOBBER}; return _17; } I thought this was a strict-aliasing issue, but specifying "-fno-strict-aliasing" doesn't give any difference.
BTW, it happens on ilp32 platforms I think.
I will have a look.
The issue is corrupted VOPs. <bb 2>: # .MEM_4 = VDEF <.MEM_3(D)> iD.1808@ = 1; # .MEM_5 = VDEF <.MEM_4> xD.1810@ = &iD.1808@; # .MEM_6 = VDEF <.MEM_5> yD.1811@ = &jD.1807@; # .MEM_7 = VDEF <.MEM_6> fD.1806@.pD.1801 = &yD.1811@; # .MEM_8 = VDEF <.MEM_7> fD.1806@.qD.1802 = &xD.1810@; # PT = { D.1806 } p_25 = &fD.1806@ + 4; # .MEM_28 = VDEF <.MEM_3(D)> jD.1807@ = 4; see how that uses the wrong VUSE.
The code in SESE doesn't deal with virtual operands correctly. In other places GRAPHITE tries to paper over this with marking all vops for renaming. But that isn't done when no code is generated, thus this breaks.
Author: rguenth Date: Tue Jun 9 07:51:24 2015 New Revision: 224266 URL: https://gcc.gnu.org/viewcvs?rev=224266&root=gcc&view=rev Log: 2015-06-09 Richard Biener <rguenther@suse.de> PR tree-optimization/66396 * graphite-isl-ast-to-gimple.c (graphite_regenerate_ast_isl): Rename virtual operands. Modified: trunk/gcc/ChangeLog trunk/gcc/graphite-isl-ast-to-gimple.c
Fixed.