On Fri, 6 Nov 2015, Alan Lawrence wrote: > On 28/10/15 13:38, Richard Biener wrote: > > > > Applied as follows. > > > > Bootstrapped / tested on x86_64-unknown-linux-gnu. > > > > Richard. > > > > 2015-10-28 Richard Biener <rguent...@suse.de> > > > > * fold-const.c (negate_expr_p): Adjust the division case to > > properly avoid introducing undefined overflow. > > (fold_negate_expr): Likewise. > > Since this we've been seeing an ICE compiling polynom.c from 254.gap in > SPEC2000 on aarch64-linux-gnu with -O3 -ffast-math -mcpu=cortex-a53 (or -Ofast > -mcpu=cortex-a53), on both native (bootstrapped and --disable-bootstrap) and > cross-linux builds. > > A number of options prevent the ICE, e.g. any of -fno-thread-jumps, > -fno-strict-overflow, -fdump-tree-alias or -fdump-tree-ealias (!). Similarly, > dropping the -mcpu=cortex-a53, or changing to -mcpu=cortex-a57. > > (I have a recent build in a chroot for which -fno-strict-overflow does *not* > fix the ICE but haven't yet figured out exactly what the difference in the > chroot environment is.) > > Moreover, preprocessing in a separate step (i.e. piping preprocessed output > via a file with -E), also avoids the ICE. (This is hindering my efforts to > reduce the testcase!). So my hypothesis is that this is a > front-end/preprocessor bug, rather than anything directly due to this commit. > > The error message in full (line refs from that commit, r229479) is: > ===== > ../spec2000/benchspec/CINT2000/254.gap/src/polynom.c: In function > ‘NormalizeCoeffsListx’: > ../spec2000/benchspec/CINT2000/254.gap/src/polynom.c:358:11: error: > incompatible types in PHI argument 0 > TypHandle NormalizeCoeffsListx ( hdC ) > ^ > long int > > int > > ../spec2000/benchspec/CINT2000/254.gap/src/polynom.c:358:11: error: location > references block not in block tree > l1_279 = PHI <1(28), l1_299(33)>
^^^ this is the error to look at! It means that the GC heap will be corrupted quite easily. > ../spec2000/benchspec/CINT2000/254.gap/src/polynom.c:358:11: error: invalid > PHI argument which means this could be a followup error. We do have a bugreport (or two) about similar issues that were tracked down to different patches. Somebody needs to sit down and debug this properly ;) > ../spec2000/benchspec/CINT2000/254.gap/src/polynom.c:358:11: internal compiler > error: tree check: expected class ‘type’, have ‘declaration’ (namespace_decl) > in useless_type_conversion_p, at gimple-expr.c:84 > 0xd110ef tree_class_check_failed(tree_node const*, tree_code_class, char > const*, int, char const*) > ../../gcc-fsf/gcc/tree.c:9643 > 0x82561b tree_class_check > ../../gcc-fsf/gcc/tree.h:3042 > 0x82561b useless_type_conversion_p(tree_node*, tree_node*) > ../../gcc-fsf/gcc/gimple-expr.c:84 > 0xaca043 verify_gimple_phi > ../../gcc-fsf/gcc/tree-cfg.c:4673 > 0xaca043 verify_gimple_in_cfg(function*, bool) > ../../gcc-fsf/gcc/tree-cfg.c:4967 > 0x9c2e0b execute_function_todo > ../../gcc-fsf/gcc/passes.c:1967 Interesting would be for which pass this happens - just print *pass at this point. > 0x9c360b do_per_function > ../../gcc-fsf/gcc/passes.c:1659 > 0x9c3807 execute_todo > ../../gcc-fsf/gcc/passes.c:2022 > Please submit a full bug report, > with preprocessed source if appropriate. > ===== > which looks like an "incompatible types from PHI argument" from a first call > to verify_gimple_phi, then a second call to verify_gimple_phi prints "invalid > phi argument" and ICEs in the test just before possibly printing a second > incompatible_types message. > > > --Alan > > -- Richard Biener <rguent...@suse.de> SUSE LINUX GmbH, GF: Felix Imendoerffer, Jane Smithard, Graham Norton, HRB 21284 (AG Nuernberg)
