On 31/10/16 17:15, Segher Boessenkool wrote:
If maybe_record_trace_start fails because the CFI is inconsistent on two paths into a block it currently just ICEs. This changes it to also dump the CFI on those two paths in the dump file; debugging it without that information is hopeless.Tested on powerpc64-linux {-m32,-m64}. Is this okay for trunk?
+1 from me as a variant of this patch has helped me debug issues with the CFI. Kyrill
Segher 2016-10-31 Segher Boessenkool <seg...@kernel.crashing.org> * dwarf2cfi.c (dump_cfi_row): Add forward declaration. (maybe_record_trace_start): If the CFI is different on the new and old paths, print out both to the dump file before ICEing. --- gcc/dwarf2cfi.c | 18 +++++++++++++++++- 1 file changed, 17 insertions(+), 1 deletion(-) diff --git a/gcc/dwarf2cfi.c b/gcc/dwarf2cfi.c index da9da52..19edc28 100644 --- a/gcc/dwarf2cfi.c +++ b/gcc/dwarf2cfi.c @@ -2239,6 +2239,8 @@ add_cfis_to_fde (void) } }+static void dump_cfi_row (FILE *f, dw_cfi_row *row);+ /* If LABEL is the start of a trace, then initialize the state of that trace from CUR_TRACE and CUR_ROW. */@@ -2282,7 +2284,21 @@ maybe_record_trace_start (rtx_insn *start, rtx_insn *origin)/* We ought to have the same state incoming to a given trace no matter how we arrive at the trace. Anything else means we've got some kind of optimization error. */ - gcc_checking_assert (cfi_row_equal_p (cur_row, ti->beg_row)); +#if CHECKING_P + if (!cfi_row_equal_p (cur_row, ti->beg_row)) + { + if (dump_file) + { + fprintf (dump_file, "Inconsistent CFI state!\n"); + fprintf (dump_file, "SHOULD have:\n"); + dump_cfi_row (dump_file, ti->beg_row); + fprintf (dump_file, "DO have:\n"); + dump_cfi_row (dump_file, cur_row); + } + + gcc_unreachable (); + } +#endif/* The args_size is allowed to conflict if it isn't actually used. */if (ti->beg_true_args_size != args_size)
