Created attachment 24398 [details] reduced testcase (the same as gcc.c-torture/compile/pr36817.c) Compiler output: $ gcc -O -fno-delete-null-pointer-checks -fno-tree-scev-cprop -ftree-vectorize -fno-vect-cost-model testcase.c testcase.c: In function 'foo': testcase.c:9:10: internal compiler error: in int_mode_for_mode, at stor-layout.c:424 Please submit a full bug report, with preprocessed source if appropriate. See <http://gcc.gnu.org/bugs.html> for instructions. (gdb) bt #0 fancy_abort (file=0x11c8af0 "/mnt/svn/gcc-trunk/gcc/stor-layout.c", line=424, function=0x11c8ee0 "int_mode_for_mode") at /mnt/svn/gcc-trunk/gcc/diagnostic.c:892 #1 0x0000000000939533 in int_mode_for_mode (mode=Unhandled dwarf expression opcode 0xf3 ) at /mnt/svn/gcc-trunk/gcc/stor-layout.c:424 #2 0x00000000006bdae2 in emit_move_via_integer (mode=VOIDmode, x=0x7ffff5c345e0, y=0x7ffff7ecc470, force=0 '\000') at /mnt/svn/gcc-trunk/gcc/expr.c:2921 #3 0x00000000006cb8ea in emit_move_insn_1 (x=0x7ffff5c345e0, y=0x7ffff7ecc470) at /mnt/svn/gcc-trunk/gcc/expr.c:3296 #4 0x00000000006cbc53 in emit_move_insn (x=0x7ffff5c345e0, y=0x7ffff7ecc470) at /mnt/svn/gcc-trunk/gcc/expr.c:3359 #5 0x00000000006ad0bb in copy_to_reg (x=0x7ffff7ecc470) at /mnt/svn/gcc-trunk/gcc/explow.c:608 #6 0x00000000006ae10f in memory_address_addr_space (mode=V4SImode, x=0x7ffff5c36438, as=0 '\000') at /mnt/svn/gcc-trunk/gcc/explow.c:480 #7 0x00000000006c49f1 in expand_expr_real_1 (exp=0x7ffff5c319d8, target=0x0, tmode=VOIDmode, modifier=EXPAND_WRITE, alt_rtl=0x0) at /mnt/svn/gcc-trunk/gcc/expr.c:8690 #8 0x00000000006d9cf8 in expand_expr (to=0x7ffff5c319d8, from=0x7ffff5c1fcd8, nontemporal=0 '\000') at /mnt/svn/gcc-trunk/gcc/expr.h:419 #9 expand_assignment (to=0x7ffff5c319d8, from=0x7ffff5c1fcd8, nontemporal=0 '\000') at /mnt/svn/gcc-trunk/gcc/expr.c:4406 #10 expand_assignment (to=0x7ffff5c319d8, from=0x7ffff5c1fcd8, nontemporal=0 '\000') at /mnt/svn/gcc-trunk/gcc/expr.c:4106 #11 0x00000000005e6ce6 in expand_gimple_stmt_1 (stmt=0x7ffff5c37910) at /mnt/svn/gcc-trunk/gcc/cfgexpand.c:1953 #12 expand_gimple_stmt (stmt=0x7ffff5c37910) at /mnt/svn/gcc-trunk/gcc/cfgexpand.c:2050 #13 0x00000000005e86ba in expand_gimple_basic_block (bb=0x7ffff5bb41a0) at /mnt/svn/gcc-trunk/gcc/cfgexpand.c:3633 #14 0x00000000005eec67 in gimple_expand_cfg () at /mnt/svn/gcc-trunk/gcc/cfgexpand.c:4116 #15 0x000000000085e876 in execute_one_pass (pass=0x16adae0) at /mnt/svn/gcc-trunk/gcc/passes.c:1876 #16 0x000000000085eb65 in execute_pass_list (pass=0x16adae0) at /mnt/svn/gcc-trunk/gcc/passes.c:1930 #17 0x00000000009b398d in tree_rest_of_compilation (fndecl=0x7ffff5bcf000) at /mnt/svn/gcc-trunk/gcc/tree-optimize.c:417 #18 0x0000000000614cf5 in cgraph_expand_function (node=0x7ffff5bd4000) at /mnt/svn/gcc-trunk/gcc/cgraphunit.c:1630 #19 0x000000000061697c in cgraph_expand_all_functions () at /mnt/svn/gcc-trunk/gcc/cgraphunit.c:1689 #20 cgraph_optimize () at /mnt/svn/gcc-trunk/gcc/cgraphunit.c:1952 #21 0x0000000000616faa in cgraph_finalize_compilation_unit () at /mnt/svn/gcc-trunk/gcc/cgraphunit.c:1126 #22 0x00000000004fe148 in c_write_global_declarations () at /mnt/svn/gcc-trunk/gcc/c-decl.c:9844 #23 0x0000000000948a74 in compile_file (argc=17, argv=0x7fffffffdb88) at /mnt/svn/gcc-trunk/gcc/toplev.c:586 #24 do_compile (argc=17, argv=0x7fffffffdb88) at /mnt/svn/gcc-trunk/gcc/toplev.c:1923 #25 toplev_main (argc=17, argv=0x7fffffffdb88) at /mnt/svn/gcc-trunk/gcc/toplev.c:1995 #26 0x00007ffff6477cec in __libc_start_main () from /lib64/libc.so.6 #27 0x00000000004e22ed in _start () Tested revisions: r174424 - crash 4.6 r173059 - OK
Confirmed, TARGET_MEM_REF expansion needs fixing. This started with http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=174282
Created attachment 24401 [details] gcc47-pr49235.patch Untested fix. base might be const0_rtx (seems other constants are moved into offset and base stays const0_rtx), and if there is also offset and nothing else, we end up with (plus:DI (const_int 0) (const_int 16)) and similar.
Need to make sure to also backport this then.
(In reply to comment #2) > Created attachment 24401 [details] > gcc47-pr49235.patch > > Untested fix. base might be const0_rtx (seems other constants are moved into > offset and base stays const0_rtx), and if there is also offset and nothing > else, > we end up with (plus:DI (const_int 0) (const_int 16)) and similar. Yes, base has to be of pointer type. But I thought get_address_description would "untangle" this again ... if not it's probably latent pre-MEM_REF as well.
get_address_description only does that if TMR_INDEX2 is present: addr->symbol = NULL_TREE; if (TMR_INDEX2 (op)) { gcc_assert (integer_zerop (TMR_BASE (op))); addr->base = TMR_INDEX2 (op); } else addr->base = TMR_BASE (op); In this case TMR_BASE (op) is (void *) 0 and TMR_OFFSET is INTEGER_CST, the rest is all NULL.
Index: gcc/tree-ssa-address.c =================================================================== --- gcc/tree-ssa-address.c (revision 174469) +++ gcc/tree-ssa-address.c (working copy) @@ -365,7 +365,7 @@ create_mem_ref_raw (tree type, tree alia ??? As IVOPTs does not follow restrictions to where the base pointer may point to create a MEM_REF only if we know that base is valid. */ - if (TREE_CODE (base) == ADDR_EXPR + if ((TREE_CODE (base) == ADDR_EXPR || TREE_CODE (base) == INTEGER_CST) && (!index2 || integer_zerop (index2)) && (!addr->index || integer_zerop (addr->index))) return fold_build2 (MEM_REF, type, base, addr->offset); might hide the issue again (still worth fixing IMHO), and also makes sense.
Author: jakub Date: Tue May 31 12:34:42 2011 New Revision: 174478 URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=174478 Log: PR rtl-optimization/49235 * tree-ssa-address.c (gen_addr_rtx): Ignore base if it is const0_rtx. (create_mem_ref_raw): Create MEM_REF even if base is INTEGER_CST. * gcc.dg/pr49235.c: New test. Added: trunk/gcc/testsuite/gcc.dg/pr49235.c Modified: trunk/gcc/ChangeLog trunk/gcc/testsuite/ChangeLog trunk/gcc/tree-ssa-address.c
Fixed.
Author: rguenth Date: Mon Jun 6 10:13:23 2011 New Revision: 174688 URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=174688 Log: 2011-06-06 Richard Guenther <rguenther@suse.de> PR tree-optimization/48702 * tree-ssa-address.c (create_mem_ref_raw): Create MEM_REFs only when we know the base address is within bounds. * tree-ssa-alias.c (indirect_ref_may_alias_decl_p): Do not assume the base address of TARGET_MEM_REFs is in bounds. (indirect_refs_may_alias_p): Fix TARGET_MEM_REF without index tests. * gcc.dg/torture/pr48702.c: New testcase. Backport from mainline 2011-05-31 Jakub Jelinek <jakub@redhat.com> PR rtl-optimization/49235 * tree-ssa-address.c (gen_addr_rtx): Ignore base if it is const0_rtx. (create_mem_ref_raw): Create MEM_REF even if base is INTEGER_CST. * gcc.dg/pr49235.c: New test. Added: branches/gcc-4_6-branch/gcc/testsuite/gcc.dg/pr49235.c branches/gcc-4_6-branch/gcc/testsuite/gcc.dg/torture/pr48702.c Modified: branches/gcc-4_6-branch/gcc/ChangeLog branches/gcc-4_6-branch/gcc/testsuite/ChangeLog branches/gcc-4_6-branch/gcc/tree-ssa-address.c branches/gcc-4_6-branch/gcc/tree-ssa-alias.c