I tried building a cross to ppc this evening and ran into this
failure. I don't think it's anything serious, but I'm not familiar
with SB-INTROSPECT or the GC bits it's touching here, so I'd
appreciate a double-check on my logic.
The failing part of the test comes about because the code pokes at the
bitfield flags in 'struct page'. The relevant bits are (most comments
trimmed for brevity):
/* 000 free
* 10? boxed data
* 11? boxed code
* 01? unboxed data
* ??1 open region
* If the page is free the following slots are invalid, except
* for the bytes_used which must be zero. */
and the SB-INTROSPECT code tests (logbitp 2 flags) for the page
containing #'CONS. Apparently this is stored in boxed code on x86,
but it's stored in boxed data (?) on ppc, so the test fails.
Is the test wrong, or is there some subtle bug in the GC on ppc?
On 6 April 2010 04:22, Nathan Froyd <froydnj@...> wrote:
> Is the test wrong, or is there some subtle bug in the GC on ppc?
I believe this (the bug): IIRC you enable various GENCGC
sanity-checking mechanisms it keeps working fine on x86oids, but
breaks on PPC.