SBCL Internals : SPARC

SBCL Internals
The pages on this CLiki-driven site can be edited by anybody at any time. No warranty of any kind can therefore be made; any implied warranties of merchantability or fitness for a particular purpose are expressly disclaimed

[ Home ]

[ Recent Changes ]

[ About CLiki ]

[ Text Formatting ]

Following Daniel Barlow's port to the Alpha, Christophe Rhodes saw a golden opportunity to learn about SPARC assembler, SBCL internals and programming signal handler?s in parallel. Current status: Merged with mainline SBCL, including Solaris support.

The SPARC port was initially targeted for SPARC/Linux. Currently, unresolved questions are:

The CMUCL SPARC backend doesn't define a DEFINE-FULL-REFFER? macro; instead, it has explicit VOPs for %INSTANCE-REF?. However, in attepmting to compile PCOUNTER-OR-FIXNUM->INTEGER?, these VOPs are not used as the compiler doesn't think that the first type is an INSTANCE?. This has currently been worked around by replacing the INSTANCE? type specifier with *?. This was not all the problem, it turns out; on all platforms, the SLOT-SET?-derived %INSTANCE-SET? VOP was bogusly returning no values, rather than one. Current workaround is not to use this VOP; expect a better solution soon. PPC also has this problem;
As noted in comments in the source, SB-SHOW? aka debug_print? can be broken, and is in the SPARC port; enabling SB-SHOW? allows debug_print to trash the stack. Also, the assembly source needs some reviewing, as even without SB-SHOW the SPARC port dies in first GC with the "correct" value of 0x44 for FRAMESIZE.

SPARC/Linux and `SA_SIGINFO` signal handlers

SBCL uses SA_SIGINFO signal handlers as a way of getting hold of register contents and other important, but unportable, low-level stuff when GCing and at other exceptional moments. POSIX defines SA_SIGINFO-style signal handlers to have a prototype

void sigfoo_handler(int signo, siginfo_t *info, void *context)

where context is castable to a pointer to a ucontext_t.

SPARC Linux doesn't do this, instead it passes a pointer to a struct sigcontext as the context.

For further amusement, note that a sigcontext on SPARC only contains pointers to 19 registers; the SPARC has 32+epsilon (visible to any one procedure - actually, due to the register window architecture, it has 8 + 16 * NWINDOWS). Now, SBCL needs to know the locals and ins as well as the globals and outs, so it's important that we get at them somehow. Magically, CMUCL has in its SPARC assembly:

save_context:
  ta ST_FLUSH_WINDOWS
  ret
  nop

which makes the local and in registers of the caller visible though the saved stack pointer. This works by invoking a trap which flushes all the register windows onto the stack (in the spaces reserved for them by convention in each frame, see The SPARC Architecture Manual, Appendix D.2).

This page is linked from: Christophe Rhodes CPU Garbage Collection Guard index tag bit VOP

CLiki pages can be edited by anyone at any time. Imagine a fearsomely comprehensive disclaimer of liability. Now fear, comprehensively

SPARC/Linux and SA_SIGINFO signal handlers

SPARC/Linux and `SA_SIGINFO` signal handlers