- Dec 03, 1997
-
-
dtc authored
were being allocated, the frame pointer was pushed then three more slots allocated. At most three slots are needed, so allocate just three. Perhaps only two slot are needed now, for the ofcp and return pc, the code slot is not used.
-
dtc authored
sequence.
-
dtc authored
-
dtc authored
-
dtc authored
not (complex (float -1.0 1.0))). From Raymond Toy.
-
dtc authored
Cleanup the verbose messages, making them all dependent on the verbose level, and disable them all by default.
-
- Dec 02, 1997
-
-
dtc authored
-
- Dec 01, 1997
- Nov 30, 1997
-
-
dtc authored
register but had not declared this register as a temporary.
-
dtc authored
suggested by Raymond Toy.
-
pw authored
-
dtc authored
Slight cleanup of these header files.
-
dtc authored
-
dtc authored
are zero filled and not write protected which should be the case. Make more of the consistency checks conditional and disable them all.
-
- Nov 29, 1997
- Nov 28, 1997
- Nov 27, 1997
-
-
dtc authored
o Fixed bug in real/imag part optimizers which didn't handle union types. o Added derive-type optimizer for complex. o Added deftransforms for complex operations: o complex-complex ops: + - * / o complex-float ops: + - * / o float-complex ops: + - *
-
dtc authored
-
dtc authored
x86 port under CGC and GENCGC.
-
dtc authored
-
- Nov 25, 1997
-
-
dtc authored
early stages so will have to be hand edited.
-
dtc authored
code into the dynamic heap.
-
dtc authored
-
dtc authored
-
dtc authored
improvement in reliability necessary for breakpoint and single stepping to have a chance, and has been tested in the experimental code and works on FreeBSD and Linux. The code object for a given return address in now located with the help of the GC support function component_ptr_from_pc which can reliably locate the code objects. This uses a forward search of the heap so is rather slow - some caching will be needed for better performance. The stack parsing has been re-written using a backtracking search to improve the chance of finding a valid frame path given that the control changes between C and Lisp conventions. The x86-lra hack is no longer necessary (was unreliable anyway) allowing x86 code to be more compact. The return address is now handled as a SAP which is acceptable because the code doesn't move with CGC, and GENCGC ensures that code objects in the dynamic heap with return addresses on the stack are not moved. However there are known problems under GENCGC when return addresses are move off the stack during some breakpoint operations in which case the code object may move.
-
dtc authored
tested in the experimental code and works on both FreeBSD and Linux. Plus a few improvements to the interrupt safety for the x86 port.
-
dtc authored
clear it during a C call-out.
-
dtc authored
of no utility and wasn't even being maintained thoroughly. Can now use pseudo atomic in C code. All interrupt contexts are saved, even those that occur when running foreign C code; this may be necessary for reliable stack parsing on some operating systems, e.g. Linux. When the pending interrupt from a GC trigger is processed by interrupt_handle_pending, any other accumulated pending interrupt is also processed even if *interrupts-enabled* is false, breaking the behaviour of the *interrupts-enabled* flag. The best way around this for the x86 port is to call maybe-gc directly from the allocation routines; remove the recently introduced function set_maybe_gc_pending which also had other problems. Note this problem will likely occur on the other ports.
-
pw authored
-
- Nov 23, 1997
- Nov 22, 1997
-
-
dtc authored
values. The stack was being reset exposing the return address to corruption before it was read - caused the occasional SIGSEGV fault with the MP SIGALRM yield enabled.
-
dtc authored
debugger. Add a few more assertions to check that the current process and stack group are consistent.
-
dtc authored
calling interrupt_handle_now, else when interrupt_handle_now resets signal mask to that in the signal context it will be masking the interrupts where the intention is to re-enable them. This caused the multi-processing to get stuck working on a particular process with the signals masked. Change may have another benefit: the pending_mask is no longer needed so there is no danger from recursive pending interrupts.
-