marovich@hplabsz.HPL.HP.COM (Scott Marovich) (06/30/90)
In article <7556@ccncsu.ColoState.EDU>, haines@debussy.cs.colostate.edu (Matt Haines) writes: > Hi. > > I am writing some context-switching assembly code for the HPPA architecture, > and I am having some problems. Programs that do not use doubles are > switching fine, but those with doubles are having some problems. I don't quite understand what you're trying to do; implement a coroutine package, perhaps? If so, you should become familiar with HP's "Procedure Calling Conventions Reference Manual", Order No. 09740-90015, which describes the layout of an HP-PA stack frame. Among other constraints, all stack frames must be a multiple of 8 Bytes long, aligned on an 8-Byte storage boundary, precisely so that Floating-point Coprocessor registers can be saved/restored without storage-boundary alignment problems. The architected positions within a stack frame of floating-point register save areas were allocated with this idea in mind, so adhering to the convention is strongly recommended. If this isn't the source of your problem, here's another idea: Does your machine have a Floating-point Coprocessor at all? If not, these instructions will be done by software emulation via a trap into the HP-UX kernel, which as a side effect, can provide an occasion for signal delivery. What does your context-switching software do in the event that a signal is delivered in the middle of a switch? And, oh, by the way: This also implies that user-mode coroutine-switching on a Coprocessor-less machine may not be as fast as you expect it to be, if the floating-point registers are included in coroutine "state" information. Now, here's a third possibility: Be careful of exactly how you code a floating- point "Load"/"Store" instruction; in particular, be aware that the displacement field of an indexed memory-reference instruction is only 5 bits for floating- point, instead of 14 for a General Register (equivalently, "Store Double" is roughly comparable to "Store Short"). Are you possibly exceeding the [-16,+15] Byte addressing span of these instructions? If none of these ideas correctly diagnoses your problem, send me E-mail and perhaps we can figure it out . . .. -- Scott Marovich Hewlett-Packard Laboratories M/S 1U15 CSNet : marovich@hplabs.HP.COM P.O. Box 10490 (1501 Page Mill Road) HP-DESK: Scott_Marovich/1900/UX Palo Alto CA 94303-0971 U.S.A. HP TELNET 857-2998 (415)857-2998