diff options
Diffstat (limited to 'qt/md')
45 files changed, 4385 insertions, 0 deletions
diff --git a/qt/md/.cvsignore b/qt/md/.cvsignore new file mode 100644 index 000000000..f3c7a7c5d --- /dev/null +++ b/qt/md/.cvsignore @@ -0,0 +1 @@ +Makefile diff --git a/qt/md/Makefile.am b/qt/md/Makefile.am new file mode 100644 index 000000000..aae2b5904 --- /dev/null +++ b/qt/md/Makefile.am @@ -0,0 +1,11 @@ +## Process this file with automake to produce Makefile.in. + +AUTOMAKE_OPTIONS = foreign + +EXTRA_DIST = _sparc.s _sparc_b.s axp.1.Makefile axp.2.Makefile \ +axp.Makefile axp.README axp.c axp.h axp.s axp_b.s default.Makefile \ +hppa-cnx.Makefile hppa.Makefile hppa.h hppa.s hppa_b.s i386.README \ +i386.h i386.s i386_b.s ksr1.Makefile ksr1.h ksr1.s ksr1_b.s \ +m88k.Makefile m88k.c m88k.h m88k.s m88k_b.s mips-irix5.s mips.h mips.s \ +mips_b.s null.README null.c solaris.README sparc.h sparc.s sparc_b.s \ +vax.h vax.s vax_b.s diff --git a/qt/md/Makefile.in b/qt/md/Makefile.in new file mode 100644 index 000000000..eace458d9 --- /dev/null +++ b/qt/md/Makefile.in @@ -0,0 +1,154 @@ +# Makefile.in generated automatically by automake 1.1l from Makefile.am + +# Copyright (C) 1994, 1995, 1996 Free Software Foundation, Inc. +# This Makefile.in is free software; the Free Software Foundation +# gives unlimited permission to copy, distribute and modify it. + + +SHELL = /bin/sh + +srcdir = @srcdir@ +top_srcdir = @top_srcdir@ +VPATH = @srcdir@ +prefix = @prefix@ +exec_prefix = @exec_prefix@ + +bindir = @bindir@ +sbindir = @sbindir@ +libexecdir = @libexecdir@ +datadir = @datadir@ +sysconfdir = @sysconfdir@ +sharedstatedir = @sharedstatedir@ +localstatedir = @localstatedir@ +libdir = @libdir@ +infodir = @infodir@ +mandir = @mandir@ +includedir = @includedir@ +oldincludedir = /usr/include + +pkgdatadir = $(datadir)/@PACKAGE@ +pkglibdir = $(libdir)/@PACKAGE@ +pkgincludedir = $(includedir)/@PACKAGE@ + +top_builddir = .. + +INSTALL = @INSTALL@ +INSTALL_PROGRAM = @INSTALL_PROGRAM@ +INSTALL_DATA = @INSTALL_DATA@ +INSTALL_SCRIPT = @INSTALL_SCRIPT@ +transform = @program_transform_name@ +host_alias = @host_alias@ +host_triplet = @host@ +RANLIB = @RANLIB@ +module = @module@ +qtmd_h = @qtmd_h@ +CC = @CC@ +PACKAGE = @PACKAGE@ +VERSION = @VERSION@ +INSTALL_SCRIPT = @INSTALL_SCRIPT@ +qtmds_o = @qtmds_o@ +qtmdc_o = @qtmdc_o@ +target_libs = @target_libs@ +qtmds_s = @qtmds_s@ +qtmdc_c = @qtmdc_c@ +qtmdb_s = @qtmdb_s@ + +AUTOMAKE_OPTIONS = foreign + +EXTRA_DIST = _sparc.s _sparc_b.s axp.1.Makefile axp.2.Makefile \ +axp.Makefile axp.README axp.c axp.h axp.s axp_b.s default.Makefile \ +hppa-cnx.Makefile hppa.Makefile hppa.h hppa.s hppa_b.s i386.README \ +i386.h i386.s i386_b.s ksr1.Makefile ksr1.h ksr1.s ksr1_b.s \ +m88k.Makefile m88k.c m88k.h m88k.s m88k_b.s mips-irix5.s mips.h mips.s \ +mips_b.s null.README null.c solaris.README sparc.h sparc.s sparc_b.s \ +vax.h vax.s vax_b.s +mkinstalldirs = $(SHELL) $(top_srcdir)/../mkinstalldirs +CONFIG_CLEAN_FILES = +DIST_COMMON = Makefile.am Makefile.in + + +DISTFILES = $(DIST_COMMON) $(SOURCES) $(HEADERS) \ + $(TEXINFOS) $(MANS) $(EXTRA_DIST) + +TAR = tar +default: all + +.SUFFIXES: +$(srcdir)/Makefile.in: Makefile.am $(top_srcdir)/configure.in $(ACLOCAL) + cd $(top_srcdir) && automake --foreign md/Makefile + +Makefile: $(srcdir)/Makefile.in $(top_builddir)/config.status $(BUILT_SOURCES) + cd $(top_builddir) \ + && CONFIG_FILES=$(subdir)/$@ CONFIG_HEADERS= $(SHELL) ./config.status + +tags: TAGS +TAGS: + + +distdir = $(top_builddir)/$(PACKAGE)-$(VERSION)/$(subdir) + +subdir = md +distdir: $(DISTFILES) + @for file in $(DISTFILES); do \ + d=$(srcdir); \ + test -f $(distdir)/$$file \ + || ln $$d/$$file $(distdir)/$$file 2> /dev/null \ + || cp -p $$d/$$file $(distdir)/$$file; \ + done +info: +dvi: +check: all + $(MAKE) +installcheck: +install-exec: + $(NORMAL_INSTALL) + +install-data: + $(NORMAL_INSTALL) + +install: install-exec install-data all + @: + +uninstall: + +all: Makefile + +install-strip: + $(MAKE) INSTALL_PROGRAM='$(INSTALL_PROGRAM) -s' install +installdirs: + + +mostlyclean-generic: + test -z "$(MOSTLYCLEANFILES)" || rm -f $(MOSTLYCLEANFILES) + +clean-generic: + test -z "$(CLEANFILES)" || rm -f $(CLEANFILES) + +distclean-generic: + rm -f Makefile $(DISTCLEANFILES) + rm -f config.cache config.log stamp-h + test -z "$(CONFIG_CLEAN_FILES)" || rm -f $(CONFIG_CLEAN_FILES) + +maintainer-clean-generic: + test -z "$(MAINTAINERCLEANFILES)" || rm -f $(MAINTAINERCLEANFILES) + test -z "$(BUILT_SOURCES)" || rm -f $(BUILT_SOURCES) +mostlyclean: mostlyclean-generic + +clean: clean-generic mostlyclean + +distclean: distclean-generic clean + rm -f config.status + +maintainer-clean: maintainer-clean-generic distclean + @echo "This command is intended for maintainers to use;" + @echo "it deletes files that may require special tools to rebuild." + +.PHONY: default tags distdir info dvi installcheck install-exec \ +install-data install uninstall all installdirs mostlyclean-generic \ +distclean-generic clean-generic maintainer-clean-generic clean \ +mostlyclean distclean maintainer-clean + + +# Tell versions [3.59,3.63) of GNU make to not export all variables. +# Otherwise a system limit (for SysV at least) may be exceeded. +.NOEXPORT: diff --git a/qt/md/_sparc.s b/qt/md/_sparc.s new file mode 100644 index 000000000..1d8adc77e --- /dev/null +++ b/qt/md/_sparc.s @@ -0,0 +1,142 @@ +/* sparc.s -- assembly support for the `qt' thread building kit. */ + +/* + * QuickThreads -- Threads-building toolkit. + * Copyright (c) 1993 by David Keppel + * + * Permission to use, copy, modify and distribute this software and + * its documentation for any purpose and without fee is hereby + * granted, provided that the above copyright notice and this notice + * appear in all copies. This software is provided as a + * proof-of-concept and for demonstration purposes; there is no + * representation about the suitability of this software for any + * purpose. + */ + +/* #include <machine/trap.h> */ + + .text + .align 4 + .global _qt_blocki + .global _qt_block + .global _qt_abort + .global _qt_start + .global _qt_vstart + +/* Register assignment: +// %o0: incoming `helper' function to call after cswap +// also used as outgoing sp of old thread (qt_t *) +// %o1, %o2: +// parameters to `helper' function called after cswap +// %o3: sp of new thread +// %o5: tmp used to save old thread sp, while using %o0 +// to call `helper' f() after cswap. +// +// +// Aborting a thread is easy if there are no cached register window +// frames: just switch to the new stack and away we go. If there are +// cached register window frames they must all be written back to the +// old stack before we move to the new stack. If we fail to do the +// writeback then the old stack memory can be written with register +// window contents e.g., after the stack memory has been freed and +// reused. +// +// If you don't believe this, try setting the frame pointer to zero +// once we're on the new stack. This will not affect correctnes +// otherwise because the frame pointer will eventually get reloaded w/ +// the new thread's frame pointer. But it will be zero briefly before +// the reload. You will eventually (100,000 cswaps later on a small +// SPARC machine that I tried) get an illegal instruction trap from +// the kernel trying to flush a cached window to location 0x0. +// +// Solution: flush windows before switching stacks, which invalidates +// all the other register windows. We could do the trap +// conditionally: if we're in the lowest frame of a thread, the fp is +// zero already so we know there's nothing cached. But we expect most +// aborts will be done from a first function that does a `save', so we +// will rarely save anything and always pay the cost of testing to see +// if we should flush. +// +// All floating-point registers are caller-save, so this routine +// doesn't need to do anything to save and restore them. +// +// `qt_block' and `qt_blocki' return the same value as the value +// returned by the helper function. We get this ``for free'' +// since we don't touch the return value register between the +// return from the helper function and return from qt_block{,i}. +*/ + +_qt_block: +_qt_blocki: + sub %sp, 8, %sp /* Allocate save area for return pc. */ + st %o7, [%sp+64] /* Save return pc. */ +_qt_abort: + ta 0x03 /* Save locals and ins. */ + mov %sp, %o5 /* Remember old sp w/o chng ins/locals. */ + sub %o3, 64, %sp /* Allocate kwsa, switch stacks. */ + call %o0, 0 /* Call `helper' routine. */ + mov %o5, %o0 /* Pass old thread to qt_after_t() */ + /* .. along w/ args in %o1 & %o2. */ + + /* Restore callee-save regs. The kwsa + // is on this stack, so offset all + // loads by sizeof(kwsa), 64 bytes. + */ + ldd [%sp+ 0+64], %l0 + ldd [%sp+ 8+64], %l2 + ldd [%sp+16+64], %l4 + ldd [%sp+24+64], %l6 + ldd [%sp+32+64], %i0 + ldd [%sp+40+64], %i2 + ldd [%sp+48+64], %i4 + ldd [%sp+56+64], %i6 + ld [%sp+64+64], %o7 /* Restore return pc. */ + + retl /* Return to address in %o7. */ + add %sp, 72, %sp /* Deallocate kwsa, ret pc area. */ + + +/* The function calling conventions say there has to be a 1-word area +// in the caller's stack to hold a pointer to space for aggregate +// return values. It also says there should be a 6-word area to hold +// %o0..%o5 if the callee wants to save them (why? I don't know...) +// Round up to 8 words to maintain alignment. +// +// Parameter values were stored in callee-save regs and are moved to +// the parameter registers. +*/ +_qt_start: + mov %i1, %o0 /* `pu': Set up args to `only'. */ + mov %i2, %o1 /* `pt'. */ + mov %i4, %o2 /* `userf'. */ + call %i5, 0 /* Call client function. */ + sub %sp, 32, %sp /* Allocate 6-word callee space. */ + + call _qt_error, 0 /* `only' erroniously returned. */ + nop + + +/* Same comments as `_qt_start' about allocating rounded-up 7-word +// save areas. */ + +_qt_vstart: + sub %sp, 32, %sp /* Allocate 7-word callee space. */ + call %i5, 0 /* call `startup'. */ + mov %i2, %o0 /* .. with argument `pt'. */ + + add %sp, 32, %sp /* Use 7-word space in varargs. */ + ld [%sp+ 4+64], %o0 /* Load arg0 ... */ + ld [%sp+ 8+64], %o1 + ld [%sp+12+64], %o2 + ld [%sp+16+64], %o3 + ld [%sp+20+64], %o4 + call %i4, 0 /* Call `userf'. */ + ld [%sp+24+64], %o5 + + /* Use 6-word space in varargs. */ + mov %o0, %o1 /* Pass return value from userf */ + call %i3, 0 /* .. when call `cleanup. */ + mov %i2, %o0 /* .. along with argument `pt'. */ + + call _qt_error, 0 /* `cleanup' erroniously returned. */ + nop diff --git a/qt/md/_sparc_b.s b/qt/md/_sparc_b.s new file mode 100644 index 000000000..cd26672d7 --- /dev/null +++ b/qt/md/_sparc_b.s @@ -0,0 +1,106 @@ +/* + * QuickThreads -- Threads-building toolkit. + * Copyright (c) 1993 by David Keppel + * + * Permission to use, copy, modify and distribute this software and + * its documentation for any purpose and without fee is hereby + * granted, provided that the above copyright notice and this notice + * appear in all copies. This software is provided as a + * proof-of-concept and for demonstration purposes; there is no + * representation about the suitability of this software for any + * purpose. + */ + + .globl _b_call_reg + .globl _b_call_imm + .globl _b_add + .globl _b_load + +_b_null: + retl + nop + +_b_call_reg: + sethi %hi(_b_null),%o4 + or %o4,%lo(_b_null),%o4 + add %o7,%g0, %o3 +L0: + call %o4 + nop + call %o4 + nop + call %o4 + nop + call %o4 + nop + call %o4 + nop + + subcc %o0,1,%o0 + bg L0 + nop + add %o3,%g0, %o7 + retl + nop + +_b_call_imm: + sethi %hi(_b_null),%o4 + or %o4,%lo(_b_null),%o4 + add %o7,%g0, %o3 +L1: + call _b_null + call _b_null + call _b_null + call _b_null + call _b_null + + subcc %o0,1,%o0 + bg L0 + nop + add %o3,%g0, %o7 + retl + nop + + +_b_add: + add %o0,%g0,%o1 + add %o0,%g0,%o2 + add %o0,%g0,%o3 + add %o0,%g0,%o4 +L2: + sub %o0,5,%o0 + sub %o1,5,%o1 + sub %o2,5,%o2 + sub %o3,5,%o3 + sub %o4,5,%o4 + + subcc %o0,5,%o0 + sub %o1,5,%o1 + sub %o2,5,%o2 + sub %o3,5,%o3 + sub %o4,5,%o4 + + bg L2 + nop + retl + nop + + +_b_load: + ld [%sp+ 0], %g0 +L3: + ld [%sp+ 4],%g0 + ld [%sp+ 8],%g0 + ld [%sp+12],%g0 + ld [%sp+16],%g0 + ld [%sp+20],%g0 + ld [%sp+24],%g0 + ld [%sp+28],%g0 + ld [%sp+32],%g0 + ld [%sp+36],%g0 + + subcc %o0,10,%o0 + bg L3 + ld [%sp+ 0],%g0 + retl + nop diff --git a/qt/md/axp.1.Makefile b/qt/md/axp.1.Makefile new file mode 100644 index 000000000..86ccd8f42 --- /dev/null +++ b/qt/md/axp.1.Makefile @@ -0,0 +1,5 @@ + +# +# Compiling for the DEC AXP (alpha) with GNU CC or version 1.x of OSF. +# +CC = cc -std1 -D__AXP__ -D__OSF1__ diff --git a/qt/md/axp.2.Makefile b/qt/md/axp.2.Makefile new file mode 100644 index 000000000..268636fc9 --- /dev/null +++ b/qt/md/axp.2.Makefile @@ -0,0 +1,5 @@ + +# +# Compiling for the DEC AXP (alpha) with GNU CC or version 2.x of OSF. +# +CC = cc -std1 -D__AXP__ -D__OSF2__ diff --git a/qt/md/axp.Makefile b/qt/md/axp.Makefile new file mode 100644 index 000000000..4e6d74da4 --- /dev/null +++ b/qt/md/axp.Makefile @@ -0,0 +1,5 @@ + +# +# GNU CC +# +CC = gcc -D__AXP__ diff --git a/qt/md/axp.README b/qt/md/axp.README new file mode 100644 index 000000000..b6a705c07 --- /dev/null +++ b/qt/md/axp.README @@ -0,0 +1,10 @@ +The handling of varargs is platform-dependent. Assar Westerlund +stared at the problem for a while and deduces the following table: + +vers / compiler cc gcc +---------------------------------------------------------------------- +1.3 a0, offset __base, __offset +2.0 _a0, _offset __base, __offset + +The current code should handle both cc and gcc versions, provided +you configure for the correct compiler. diff --git a/qt/md/axp.c b/qt/md/axp.c new file mode 100644 index 000000000..26c15c0ea --- /dev/null +++ b/qt/md/axp.c @@ -0,0 +1,133 @@ +/* + * QuickThreads -- Threads-building toolkit. + * Copyright (c) 1993 by David Keppel + * + * Permission to use, copy, modify and distribute this software and + * its documentation for any purpose and without fee is hereby + * granted, provided that the above copyright notice and this notice + * appear in all copies. This software is provided as a + * proof-of-concept and for demonstration purposes; there is no + * representation about the suitability of this software for any + * purpose. + */ + +#include <stdarg.h> +#include "qt.h" + + +/* Varargs is harder on the AXP. Parameters are saved on the stack as + something like (stack grows down to low memory; low at bottom of + picture): + + | : + | arg6 + +--- + | iarg5 + | : + | iarg3 <-- va_list._a0 + va_list._offset + | : + | iarg0 <-- va_list._a0 + +--- + | farg5 + | : + | farg0 + +--- + + When some of the arguments have known type, there is no need to + save all of them in the struct. So, for example, if the routine is + called + + zork (int a0, float a1, int a2, ...) + { + va_list ap; + va_start (ap, a2); + qt_vargs (... &ap ...); + } + + then offset is set to 3 * 8 (8 === sizeof machine word) = 24. + + What this means for us is that the user's routine needs to be + called with an arg list where some of the words in the `any type' + parameter list have to be split and moved up in to the int/fp + region. + + Ways in which this can fail: + - The user might not know the size of the pushed arguments anyway. + - Structures have funny promotion rules. + - Probably lots of other things. + + All in all, we never promised varargs would work reliably. */ + + + +#define QT_VADJ(sp) (((char *)sp) - QT_VSTKBASE) + +#define QT_VARGS_MD0(sp, vabytes) \ + ((qt_t *)(((char *)(sp)) - 6*2*8 - QT_STKROUNDUP(vabytes))) + +extern void qt_vstart(void); +#define QT_VARGS_MD1(sp) (QT_SPUT (sp, QT_R26, qt_vstart)) + + +/* Different machines use different implementations for varargs. + Unfortunately, the code below ``looks in to'' the varargs + structure, `va_list', and thus depends on the conventions. + The following #defines try to deal with it but don't catch + everything. */ + +#ifdef __GNUC__ +#define _a0 __base +#define _offset __offset +#else +#ifdef __OSF1__ +#define _a0 a0 +#define _offset offset +#endif +#endif /* def __GNUC__ */ + + + struct qt_t * +qt_vargs (struct qt_t *qsp, int nbytes, struct va_list *vargs, + void *pt, qt_function_t *startup, + qt_function_t *vuserf, qt_function_t *cleanup) +{ + va_list ap; + int i; + int max; /* Maximum *words* of args to copy. */ + int tmove; /* *Words* of args moved typed->typed. */ + qt_word_t *sp; + + ap = *(va_list *)vargs; + qsp = QT_VARGS_MD0 (qsp, nbytes); + sp = (qt_word_t *)qsp; + + tmove = 6 - ap._offset/sizeof(qt_word_t); + + /* Copy from one typed area to the other. */ + for (i=0; i<tmove; ++i) { + /* Integer args: */ + sp[i+6] = ((qt_word_t *)(ap._a0 + ap._offset))[i]; + /* Fp args: */ + sp[i] = ((qt_word_t *)(ap._a0 + ap._offset))[i-6]; + } + + max = nbytes/sizeof(qt_word_t); + + /* Copy from the untyped area to the typed area. Split each arg. + in to integer and floating-point save areas. */ + for (; i<6 && i<max; ++i) { + sp[i] = sp[i+6] = ((qt_word_t *)(ap._a0 + ap._offset))[i]; + } + + /* Copy from the untyped area to the other untyped area. */ + for (; i<max; ++i) { + sp[i+6] = ((qt_word_t *)(ap._a0 + ap._offset))[i]; + } + + QT_VARGS_MD1 (QT_VADJ(sp)); + QT_SPUT (QT_VADJ(sp), QT_VARGT_INDEX, pt); + QT_SPUT (QT_VADJ(sp), QT_VSTARTUP_INDEX, startup); + QT_SPUT (QT_VADJ(sp), QT_VUSERF_INDEX, vuserf); + QT_SPUT (QT_VADJ(sp), QT_VCLEANUP_INDEX, cleanup); + return ((qt_t *)QT_VADJ(sp)); +} diff --git a/qt/md/axp.h b/qt/md/axp.h new file mode 100644 index 000000000..ff951a0d3 --- /dev/null +++ b/qt/md/axp.h @@ -0,0 +1,160 @@ +/* + * QuickThreads -- Threads-building toolkit. + * Copyright (c) 1993 by David Keppel + * + * Permission to use, copy, modify and distribute this software and + * its documentation for any purpose and without fee is hereby + * granted, provided that the above copyright notice and this notice + * appear in all copies. This software is provided as a + * proof-of-concept and for demonstration purposes; there is no + * representation about the suitability of this software for any + * purpose. + */ + +#ifndef QT_AXP_H +#define QT_AXP_H + +#define QT_GROW_DOWN + +typedef unsigned long qt_word_t; + + +/* Stack layout on the Alpha: + + Integer: + + Caller-save: r0..r8, r22..r25, r27..r29 + argument/caller-save: r16..r21 + callee-save: r9..r15 + return pc *callee-save*: r26 + stack pointer: r30 + zero: r31 + + Floating-point: + + Caller-save: f0..f1, f10..f15 + argument/caller-save: f16..f21, f22..f30 + callee-save: f2..f9 + zero: f31 + + Non-varargs: + + +--- + | padding + | f9 + | f8 + | f7 + | f6 + | f5 + | f4 + | f3 + | f2 + | r26 + +--- + | padding + | r29 + | r15 + | r14 + | r13 + | r12 on startup === `only' + | r11 on startup === `userf' + | r10 on startup === `qt' + | r9 on startup === `qu' + | r26 on startup === qt_start <--- qt.sp + +--- + + Conventions for varargs startup: + + | : + | arg6 + | iarg5 + | : + | iarg0 + | farg5 + | : + | farg0 + +--- + | padding + | r29 + | r15 + | r14 + | r13 + | r12 on startup === `startup' + | r11 on startup === `vuserf' + | r10 on startup === `cleanup' + | r9 on startup === `qt' + | r26 on startup === qt_vstart <--- qt.sp + +--- + + Note: this is a pretty cheap/sleazy way to get things going, + but ``there must be a better way.'' For instance, some varargs + parameters could be loaded in to integer registers, or the return + address could be stored on top of the stack. */ + + +/* Stack must be 16-byte aligned. */ +#define QT_STKALIGN (16) + +/* How much space is allocated to hold all the crud for + initialization: 7 registers times 8 bytes/register. */ + +#define QT_STKBASE (10 * 8) +#define QT_VSTKBASE QT_STKBASE + + +/* Offsets of various registers. */ +#define QT_R26 0 +#define QT_R9 1 +#define QT_R10 2 +#define QT_R11 3 +#define QT_R12 4 + + +/* When a never-before-run thread is restored, the return pc points + to a fragment of code that starts the thread running. For + non-vargs functions, it just calls the client's `only' function. + For varargs functions, it calls the startup, user, and cleanup + functions. + + The varargs startup routine always reads 12 8-byte arguments from + the stack. If fewer argumets were pushed, the startup routine + would read off the top of the stack. To prevent errors we always + allocate enough space. When there are fewer args, the preallocated + words are simply wasted. */ + +extern void qt_start(void); +#define QT_ARGS_MD(sp) (QT_SPUT (sp, QT_R26, qt_start)) + + +/* The AXP uses a struct for `va_list', so pass a pointer to the + struct. This may break some uses of `QT_VARGS', but then we never + claimed it was totally portable. */ + +typedef void (qt_function_t)(void); + +struct qt_t; +struct va_list; +extern struct qt_t *qt_vargs (struct qt_t *sp, int nbytes, + struct va_list *vargs, void *pt, + qt_function_t *startup, + qt_function_t *vuserf, + qt_function_t *cleanup); + +#define QT_VARGS(sp, nbytes, vargs, pt, startup, vuserf, cleanup) \ + (qt_vargs (sp, nbytes, (struct va_list *)(&(vargs)), pt, \ + (qt_function_t *) startup, (qt_function_t *)vuserf, \ + (qt_function_t *)cleanup)); + + +/* The *index* (positive offset) of where to put each value. */ +#define QT_ONLY_INDEX (QT_R12) +#define QT_USER_INDEX (QT_R11) +#define QT_ARGT_INDEX (QT_R10) +#define QT_ARGU_INDEX (QT_R9) + +#define QT_VCLEANUP_INDEX (QT_R10) +#define QT_VUSERF_INDEX (QT_R11) +#define QT_VSTARTUP_INDEX (QT_R12) +#define QT_VARGT_INDEX (QT_R9) + +#endif /* ndef QT_AXP_H */ diff --git a/qt/md/axp.s b/qt/md/axp.s new file mode 100644 index 000000000..a84aab2cc --- /dev/null +++ b/qt/md/axp.s @@ -0,0 +1,160 @@ +/* + * QuickThreads -- Threads-building toolkit. + * Copyright (c) 1993 by David Keppel + * + * Permission to use, copy, modify and distribute this software and + * its documentation for any purpose and without fee is hereby + * granted, provided that the above copyright notice and this notice + * appear in all copies. This software is provided as a + * proof-of-concept and for demonstration purposes; there is no + * representation about the suitability of this software for any + * purpose. + */ + +/* axp.s -- assembly support. */ + + .text + .align 4 + .file 2 "axp.s" + + .globl qt_block + .globl qt_blocki + .globl qt_abort + .globl qt_start + .globl qt_vstart + + /* + ** $16: ptr to function to call once curr is suspended + ** and control is on r19's stack. + ** $17: 1'th arg to (*$16)(...). + ** $18: 2'th arg to (*$16)(...). + ** $19: sp of thread to resume. + ** + ** The helper routine returns a value that is passed on as the + ** return value from the blocking routine. Since we don't + ** touch r0 between the helper's return and the end of + ** function, we get this behavior for free. + */ + + .ent qt_blocki +qt_blocki: + subq $30,80, $30 /* Allocate save area. */ + stq $26, 0($30) /* Save registers. */ + stq $9, 8($30) + stq $10,16($30) + stq $11,24($30) + stq $12,32($30) + stq $13,40($30) + stq $14,48($30) + stq $15,56($30) + stq $29,64($30) + .end qt_blocki + .ent qt_abort +qt_abort: + addq $16,$31, $27 /* Put argument function in PV. */ + addq $30,$31, $16 /* Save stack ptr in outgoing arg. */ + addq $19,$31, $30 /* Set new stack pointer. */ + jsr $26,($27),0 /* Call helper function. */ + + ldq $26, 0($30) /* Restore registers. */ + ldq $9, 8($30) + ldq $10,16($30) + ldq $11,24($30) + ldq $12,32($30) + ldq $13,40($30) + ldq $14,48($30) + ldq $15,56($30) + ldq $29,64($30) + + addq $30,80, $30 /* Deallocate save area. */ + ret $31,($26),1 /* Return, predict===RET. */ + .end qt_abort + + + /* + ** Non-varargs thread startup. + */ + .ent qt_start +qt_start: + addq $9,$31, $16 /* Load up `qu'. */ + addq $10,$31, $17 /* ... user function's `pt'. */ + addq $11,$31, $18 /* ... user function's `userf'. */ + addq $12,$31, $27 /* ... set procedure value to `only'. */ + jsr $26,($27),0 /* Call `only'. */ + + jsr $26,qt_error /* `only' erroniously returned. */ + .end qt_start + + + .ent qt_vstart: +qt_vstart: + /* Call startup function. */ + addq $9,$31, $16 /* Arg0 to `startup'. */ + addq $12,$31, $27 /* Set procedure value. */ + jsr $26,($27),0 /* Call `startup'. */ + + /* Call user function. */ + ldt $f16, 0($30) /* Load fp arg regs. */ + ldt $f17, 8($30) + ldt $f18,16($30) + ldt $f19,24($30) + ldt $f20,32($30) + ldt $f21,40($30) + ldq $16,48($30) /* And integer arg regs. */ + ldq $17,56($30) + ldq $18,64($30) + ldq $19,72($30) + ldq $20,80($30) + ldq $21,88($30) + addq $30,96 $30 /* Pop 6*2*8 saved arg regs. */ + addq $11,$31, $27 /* Set procedure value. */ + jsr $26,($27),0 /* Call `vuserf'. */ + + /* Call cleanup. */ + addq $9,$31, $16 /* Arg0 to `cleanup'. */ + addq $0,$31, $17 /* Users's return value is arg1. */ + addq $10,$31, $27 /* Set procedure value. */ + jsr $26,($27),0 /* Call `cleanup'. */ + + jsr $26,qt_error /* Cleanup erroniously returned. */ + .end qt_start + + + /* + ** Save calle-save floating-point regs $f2..$f9. + ** Also save return pc from whomever called us. + ** + ** Return value from `qt_block' is the same as the return from + ** `qt_blocki'. We get that for free since we don't touch $0 + ** between the return from `qt_blocki' and the return from + ** `qt_block'. + */ + .ent qt_block +qt_block: + subq $30,80, $30 /* Allocate a save space. */ + stq $26, 0($30) /* Save registers. */ + stt $f2, 8($30) + stt $f3,16($30) + stt $f4,24($30) + stt $f5,32($30) + stt $f6,40($30) + stt $f7,48($30) + stt $f8,56($30) + stt $f9,64($30) + + jsr $26,qt_blocki /* Call helper. */ + /* .. who will also restore $gp. */ + + ldq $26, 0($30) /* restore registers. */ + ldt $f2, 8($30) + ldt $f3,16($30) + ldt $f4,24($30) + ldt $f5,32($30) + ldt $f6,40($30) + ldt $f7,48($30) + ldt $f8,56($30) + ldt $f9,64($30) + + addq $30,80, $30 /* Deallcate save space. */ + ret $31,($26),1 /* Return, predict===RET. */ + .end qt_block diff --git a/qt/md/axp_b.s b/qt/md/axp_b.s new file mode 100644 index 000000000..60be726ff --- /dev/null +++ b/qt/md/axp_b.s @@ -0,0 +1,111 @@ +/* + * QuickThreads -- Threads-building toolkit. + * Copyright (c) 1993 by David Keppel + * + * Permission to use, copy, modify and distribute this software and + * its documentation for any purpose and without fee is hereby + * granted, provided that the above copyright notice and this notice + * appear in all copies. This software is provided as a + * proof-of-concept and for demonstration purposes; there is no + * representation about the suitability of this software for any + * purpose. + */ + + .text + .globl b_call_reg + .globl b_call_imm + .globl b_add + .globl b_load + + .ent b_null +b_null: + ret $31,($18),1 + .end b_null + + .ent b_call_reg +b_call_reg: + lda $27,b_null +$L0: + jsr $18,($27) + jsr $18,($27) + jsr $18,($27) + jsr $18,($27) + jsr $18,($27) + + jsr $18,($27) + jsr $18,($27) + jsr $18,($27) + jsr $18,($27) + jsr $18,($27) + + subq $16,1,$16 + bgt $16,$L0 + + ret $31,($26),1 + .end + + + .ent b_call_imm +b_call_imm: +$L1: + jsr $18,b_null + jsr $18,b_null + jsr $18,b_null + jsr $18,b_null + jsr $18,b_null + + jsr $18,b_null + jsr $18,b_null + jsr $18,b_null + jsr $18,b_null + jsr $18,b_null + + subq $16,1,$16 + bgt $16,$L1 + + ret $31,($26),1 + .end + + + .ent b_add +b_add: +$L2: + addq $31,$31,$31 + addq $31,$31,$31 + addq $31,$31,$31 + addq $31,$31,$31 + addq $31,$31,$31 + + addq $31,$31,$31 + addq $31,$31,$31 + addq $31,$31,$31 + addq $31,$31,$31 + addq $31,$31,$31 + + subq $16,1,$16 + bgt $16,$L2 + + ret $31,($26),1 + .end + + + .ent b_load +b_load: +$L3: + ldq $31,0($30) + ldq $31,8($30) + ldq $31,16($30) + ldq $31,24($30) + ldq $31,32($30) + + ldq $31,0($30) + ldq $31,8($30) + ldq $31,16($30) + ldq $31,24($30) + ldq $31,32($30) + + subq $16,1,$16 + bgt $16,$L3 + + ret $31,($26),1 + .end diff --git a/qt/md/default.Makefile b/qt/md/default.Makefile new file mode 100644 index 000000000..e240ca270 --- /dev/null +++ b/qt/md/default.Makefile @@ -0,0 +1,6 @@ + +# +# `Normal' configuration. +# +CC = gcc -ansi -Wall -pedantic + diff --git a/qt/md/hppa-cnx.Makefile b/qt/md/hppa-cnx.Makefile new file mode 100644 index 000000000..bff257d9f --- /dev/null +++ b/qt/md/hppa-cnx.Makefile @@ -0,0 +1,9 @@ +# This file (cnx_spp.Makefile) is part of the port of QuickThreads for +# PA-RISC 1.1 architecture on a Convex SPP. This file is a machine dependent +# makefile for QuickThreads. It was written in 1994 by Uwe Reder +# (`uereder@cip.informatik.uni-erlangen.de') for the Operating Systems +# Department (IMMD4) at the University of Erlangen/Nuernberg Germany. + +# `Normal' configuration. + +CC = /usr/convex/bin/cc diff --git a/qt/md/hppa.Makefile b/qt/md/hppa.Makefile new file mode 100644 index 000000000..a15e28c99 --- /dev/null +++ b/qt/md/hppa.Makefile @@ -0,0 +1,9 @@ +# This file (pa-risc.Makefile) is part of the port of QuickThreads for +# PA-RISC 1.1 architecture. This file is a machine dependent makefile +# for QuickThreads. It was written in 1994 by Uwe Reder +# (`uereder@cip.informatik.uni-erlangen.de') for the Operating Systems +# Department (IMMD4) at the University of Erlangen/Nuernberg Germany. + +# `Normal' configuration. + +CC = cc -Aa diff --git a/qt/md/hppa.h b/qt/md/hppa.h new file mode 100644 index 000000000..0df98de88 --- /dev/null +++ b/qt/md/hppa.h @@ -0,0 +1,194 @@ +/* + * QuickThreads -- Threads-building toolkit. + * Copyright (c) 1993 by David Keppel + * + * Permission to use, copy, modify and distribute this software and + * its documentation for any purpose and without fee is hereby + * granted, provided that the above copyright notice and this notice + * appear in all copies. This software is provided as a + * proof-of-concept and for demonstration purposes; there is no + * representation about the suitability of this software for any + * purpose. + */ + +/* + * This file (pa-risc.h) is part of the port of QuickThreads for the + * PA-RISC 1.1 architecture. This file is a machine dependent header + * file. It was written in 1994 by Uwe Reder + * (`uereder@cip.informatik.uni-erlangen.de') for the Operating Systems + * Department (IMMD4) at the University of Erlangen/Nuernberg Germany. + */ + + +#ifndef QT_PA_RISC_H +#define QT_PA_RISC_H + +#include <qt.h> + +/* size of an integer-register (32 bit) */ +typedef unsigned long qt_word_t; + +/* PA-RISC's stack grows up */ +#define QT_GROW_UP + +/* Stack layout on PA-RISC according to PA-RISC Procedure Calling Conventions: + + Callee-save registers are: gr3-gr18, fr12-fr21. + Also save gr2, return pointer. + + +--- + | fr12 Each floating register is a double word (8 bytes). + | fr13 Floating registers are only saved if `qt_block' is + | fr14 called, in which case it saves the floating-point + | fr15 registers then calls `qt_blocki' to save the integer + | fr16 registers. + | fr17 + | fr18 + | fr19 + | fr20 + | fr21 + | <arg word 3> fixed arguments (must be allocated; may remain unused) + | <arg word 2> + | <arg word 1> + | <arg word 0> + | <LPT> frame marker + | <LPT'> + | <RP'> + | <Current RP> + | <Static Link> + | <Clean Up> + | <RP''> + | <Previous SP> + +--- + | gr3 word each (4 bytes) + | gr4 + | gr5 + | gr6 + | gr7 + | gr8 + | gr9 + | gr10 + | gr11 + | gr12 + | gr13 + | gr14 + | gr15 + | gr16 + | gr17 + | gr18 + | <16 bytes filled in (sp has to be 64-bytes aligned)> + | <arg word 3> fixed arguments (must be allocated; may remain unused) + | <arg word 2> + | <arg word 1> + | <arg word 0> + | <LPT> frame marker + | <LPT'> + | <RP'> + | <Current RP> + | <Static Link> + | <Clean Up> + | <RP''> + | <Previous SP> + +--- <--- sp +*/ + +/* When a never-before-run thread is restored, the return pc points + to a fragment of code that starts the thread running. For + non-vargs functions, it just calls the client's `only' function. + For varargs functions, it calls the startup, user, and cleanup + functions. */ + +/* Note: Procedue Labels on PA-RISC + + <--2--><-------28---------><1-><1-> + ----------------------------------- + | SID | Adress Part | L | X | + ----------------------------------- + + On HP-UX the L field is used to flag wheather the procedure + label (plabel) is a pointer to an LT entry or to the entry point + of the procedure (PA-RISC Procedure Calling Conventions Reference + Manual, 5.3.2 Procedure Labels and Dynamic Calls). */ + +#define QT_PA_RISC_READ_PLABEL(plabel) \ + ( (((int)plabel) & 2) ? \ + ( (*((int *)(((int)plabel) & 0xfffffffc)))) : ((int)plabel) ) + +/* Stack must be 64 bytes aligned. */ +#define QT_STKALIGN (64) + +/* Internal helper for putting stuff on stack (negative index!). */ +#define QT_SPUT(top, at, val) \ + (((qt_word_t *)(top))[-(at)] = (qt_word_t)(val)) + +/* Offsets of various registers which are modified on the stack. + rp (return-pointer) has to be stored in the frame-marker-area + of the "older" stack-segment. */ + +#define QT_crp (12+4+16+5) +#define QT_15 (12+4+4) +#define QT_16 (12+4+3) +#define QT_17 (12+4+2) +#define QT_18 (12+4+1) + + +/** This stuff is for NON-VARARGS. **/ + +/* Stack looks like this (2 stack frames): + + <--- 64-bytes aligned --><------- 64-bytes aligned ------------> + | || | + <--16--><------48-------><----16*4-----><--16-><------48-------> + || | || | | || + ||filler|arg|frame-marker||register-save|filler|arg|frame-marker|| + ------------------------------------------------------------------ + */ + +#define QT_STKBASE (16+48+(16*sizeof(qt_word_t))+16+48) + +/* The index, relative to sp, of where to put each value. */ +#define QT_ONLY_INDEX (QT_15) +#define QT_USER_INDEX (QT_16) +#define QT_ARGT_INDEX (QT_17) +#define QT_ARGU_INDEX (QT_18) + +extern void qt_start(void); +#define QT_ARGS_MD(sp) \ + (QT_SPUT (sp, QT_crp, QT_PA_RISC_READ_PLABEL(qt_start))) + + +/** This is for VARARGS. **/ + +#define QT_VARGS_DEFAULT + +/* Stack looks like this (2 stack frames): + + <------ 64-bytes aligned -------><--------- 64-bytes aligned ----------> + | || | + <---?--><--?---><16><----32-----><----16*4-----><-16--><16><----32-----> + || | | | || | | | || + ||filler|varargs|arg|frame-marker||register-save|filler|arg|frame-marker|| + -------------------------------------------------------------------------- + */ + +/* Sp is moved to the end of the first stack frame. */ +#define QT_VARGS_MD0(sp, vasize) \ + ((qt_t *)(((char *)sp) + QT_STKROUNDUP(vasize + 4*4 + 32))) + +/* To reach the arguments from the end of the first stack frame use 32 + as a negative adjustment. */ +#define QT_VARGS_ADJUST(sp) ((qt_t *)(((char *)sp) - 32)) + +/* Offset to reach the end of the second stack frame. */ +#define QT_VSTKBASE ((16*sizeof(qt_word_t)) + 16 + 4*4 + 32) + +extern void qt_vstart(void); +#define QT_VARGS_MD1(sp) \ + (QT_SPUT (sp, QT_crp, QT_PA_RISC_READ_PLABEL(qt_vstart))) + +#define QT_VARGT_INDEX (QT_15) +#define QT_VSTARTUP_INDEX (QT_16) +#define QT_VUSERF_INDEX (QT_17) +#define QT_VCLEANUP_INDEX (QT_18) + +#endif /* ndef QT_PA_RISC_H */ diff --git a/qt/md/hppa.s b/qt/md/hppa.s new file mode 100644 index 000000000..84d8e875b --- /dev/null +++ b/qt/md/hppa.s @@ -0,0 +1,237 @@ +; pa-risc.s -- assembly support. + +; QuickThreads -- Threads-building toolkit. +; Copyright (c) 1993 by David Keppel +; +; Permission to use, copy, modify and distribute this software and +; its documentation for any purpose and without fee is hereby +; granted, provided that the above copyright notice and this notice +; appear in all copies. This software is provided as a +; proof-of-concept and for demonstration purposes; there is no +; representation about the suitability of this software for any +; purpose. + +; This file (pa-risc.s) is part of the port of QuickThreads for +; PA-RISC 1.1 architecture. This file implements context switches +; and thread startup. It was written in 1994 by Uwe Reder +; (`uereder@cip.informatik.uni-erlangen.de') for the Operating +; Systems Department (IMMD4) at the University of Erlangen/Nuernberg +; Germany. + + +; Callee saves general registers gr3..gr18, +; floating-point registers fr12..fr21. + + .CODE + + .IMPORT $$dyncall, MILLICODE + .IMPORT qt_error, CODE + + .EXPORT qt_blocki, ENTRY + .EXPORT qt_block, ENTRY + .EXPORT qt_abort, ENTRY + .EXPORT qt_start, ENTRY + .EXPORT qt_vstart, ENTRY + + +; arg0: ptr to function (helper) to call once curr is suspended +; and control is on arg3's stack. +; arg1: 1'th arg to *arg0. +; arg2: 2'th arg to *arg0. +; arg3: sp of new thread. + +qt_blocki + .PROC + .CALLINFO CALLER, FRAME=0, SAVE_RP, ENTRY_GR=18 + .ENTRY + + stw %rp,-20(%sp) ; save rp to old frame-marker + + stwm %r3,128(%sp) ; save callee-saves general registers + stw %r4,-124(%sp) + stw %r5,-120(%sp) + stw %r6,-116(%sp) + stw %r7,-112(%sp) + stw %r8,-108(%sp) + stw %r9,-104(%sp) + stw %r10,-100(%sp) + stw %r11,-96(%sp) + stw %r12,-92(%sp) + stw %r13,-88(%sp) + stw %r14,-84(%sp) + stw %r15,-80(%sp) + stw %r16,-76(%sp) + stw %r17,-72(%sp) + stw %r18,-68(%sp) + +qt_abort + copy %arg0,%r22 ; helper to be called by $$dyncall + copy %sp,%arg0 ; pass current sp as arg0 to helper + copy %arg3,%sp ; set new sp + + .CALL + bl $$dyncall,%mrp ; call helper + copy %mrp,%rp + + ldw -68(%sp),%r18 ; restore general registers + ldw -72(%sp),%r17 + ldw -76(%sp),%r16 + ldw -80(%sp),%r15 + ldw -84(%sp),%r14 + ldw -88(%sp),%r13 + ldw -92(%sp),%r12 + ldw -96(%sp),%r11 + ldw -100(%sp),%r10 + ldw -104(%sp),%r9 + ldw -108(%sp),%r8 + ldw -112(%sp),%r7 + ldw -116(%sp),%r6 + ldw -120(%sp),%r5 + ldw -124(%sp),%r4 + + ldw -148(%sp),%rp ; restore return-pointer + + bv %r0(%rp) ; return to caller + ldwm -128(%sp),%r3 + + .EXIT + .PROCEND + + +qt_block + .PROC + .CALLINFO CALLER, FRAME=0, SAVE_RP, ENTRY_FR=21 + .ENTRY + + stw %rp,-20(%sp) ; save rp to old frame-marker + + fstds,ma %fr12,8(%sp) ; save callee-saves float registers + fstds,ma %fr13,8(%sp) + fstds,ma %fr14,8(%sp) + fstds,ma %fr15,8(%sp) + fstds,ma %fr16,8(%sp) + fstds,ma %fr17,8(%sp) + fstds,ma %fr18,8(%sp) + fstds,ma %fr19,8(%sp) + fstds,ma %fr20,8(%sp) + fstds,ma %fr21,8(%sp) + + .CALL + bl qt_blocki,%rp + ldo 48(%sp),%sp + + ldo -48(%sp),%sp + + fldds,mb -8(%sp),%fr21 ; restore callee-saves float registers + fldds,mb -8(%sp),%fr20 + fldds,mb -8(%sp),%fr19 + fldds,mb -8(%sp),%fr18 + fldds,mb -8(%sp),%fr17 + fldds,mb -8(%sp),%fr16 + fldds,mb -8(%sp),%fr15 + fldds,mb -8(%sp),%fr14 + fldds,mb -8(%sp),%fr13 + + ldw -28(%sp),%rp ; restore return-pointer + + bv %r0(%rp) ; return to caller. + fldds,mb -8(%sp),%fr12 + + .EXIT + .PROCEND + + +qt_start + .PROC + .CALLINFO CALLER, FRAME=0 + .ENTRY + + copy %r18,%arg0 ; set user arg `pu'. + copy %r17,%arg1 ; ... user function pt. + copy %r16,%arg2 ; ... user function userf. + ; %r22 is a caller-saves register + copy %r15,%r22 ; function to be called by $$dyncall + + .CALL ; in=%r22 + bl $$dyncall,%mrp ; call `only'. + copy %mrp,%rp + + bl,n qt_error,%r0 ; `only' erroniously returned. + + .EXIT + .PROCEND + + +; Varargs +; +; First, call `startup' with the `pt' argument. +; +; Next, call the user's function with all arguments. +; We don't know whether arguments are integers, 32-bit floating-points or +; even 64-bit floating-points, so we reload all the registers, possibly +; with garbage arguments. The thread creator provided non-garbage for +; the arguments that the callee actually uses, so the callee never gets +; garbage. +; +; -48 -44 -40 -36 -32 +; | arg3 | arg2 | arg1 | arg0 | +; ----------------------------- +; integers: arg3 arg2 arg1 arg0 +; 32-bit fps: farg3 farg2 farg1 farg0 +; 64-bit fps: <---farg3--> <---farg1--> +; +; Finally, call `cleanup' with the `pt' argument and with the return value +; from the user's function. It is an error for `cleanup' to return. + +qt_vstart + .PROC + .CALLINFO CALLER, FRAME=0 + .ENTRY + + ; Because the startup function may damage the fixed arguments + ; on the stack (PA-RISC Procedure Calling Conventions Reference + ; Manual, 2.4 Fixed Arguments Area), we allocate a seperate + ; stack frame for it. + ldo 64(%sp),%sp + + ; call: void startup(void *pt) + + copy %r15,%arg0 ; `pt' is arg0 to `startup'. + copy %r16,%r22 + .CALL + bl $$dyncall,%mrp ; Call `startup'. + copy %mrp,%rp + + ldo -64(%sp),%sp + + ; call: void *qt_vuserf_t(...) + + ldw -36(%sp),%arg0 ; Load args to integer registers. + ldw -40(%sp),%arg1 + ldw -44(%sp),%arg2 + ldw -48(%sp),%arg3 + ; Index of fld[w|d]s only ranges from -16 to 15, so we + ; take r22 to be our new base register. + ldo -32(%sp),%r22 + fldws -4(%r22),%farg0 ; Load args to floating-point registers. + fldds -8(%r22),%farg1 + fldws -12(%r22),%farg2 + fldds -16(%r22),%farg3 + copy %r17,%r22 + .CALL + bl $$dyncall,%mrp ; Call `userf'. + copy %mrp,%rp + + ; call: void cleanup(void *pt, void *vuserf_return) + + copy %r15,%arg0 ; `pt' is arg0 to `cleanup'. + copy %ret0,%arg1 ; Return-value is arg1 to `cleanup'. + copy %r18,%r22 + .CALL + bl $$dyncall,%mrp ; Call `cleanup'. + copy %mrp,%rp + + bl,n qt_error,%r0 + + .EXIT + .PROCEND diff --git a/qt/md/hppa_b.s b/qt/md/hppa_b.s new file mode 100644 index 000000000..1b1e8264e --- /dev/null +++ b/qt/md/hppa_b.s @@ -0,0 +1,203 @@ +; QuickThreads -- Threads-building toolkit. +; Copyright (c) 1993 by David Keppel + +; Permission to use, copy, modify and distribute this software and +; its documentation for any purpose and without fee is hereby +; granted, provided that the above copyright notice and this notice +; appear in all copies. This software is provided as a +; proof-of-concept and for demonstration purposes; there is no +; representation about the suitability of this software for any +; purpose. + +; This file (pa-risc_b.s) is part of the port of QuickThreads for +; PA-RISC 1.1 architecture. It contains assembly-level support for +; raw processor performance measurement. It was written in 1994 by +; Uwe Reder (`uereder@cip.informatik.uni-erlangen.de') +; for the Operating Systems Department (IMMD4) at the +; University of Erlangen/Nuernberg Germany. + + +; Note that the number of instructions in the measurement-loops, differ +; from implementation to implementation. I took eight instructions in a loop +; for every test (execute eight instructions and loop to the start). + + .CODE + + .IMPORT $global$,DATA + .IMPORT $$dyncall,MILLICODE + .EXPORT b_call_reg + .EXPORT b_call_imm + .EXPORT b_add + .EXPORT b_load + +; Just do nothing, only return to caller. This procedure is called by +; `b_call_reg' and `b_call_imm'. + +b_null + .PROC + .CALLINFO NO_CALLS, FRAME=0 + .ENTRY + + bv,n %r0(%rp) ; just return + + .EXIT + .PROCEND + +; Call the procedure `b_null' with function pointer in a register. + +b_call_reg + .PROC + .CALLINFO CALLER, FRAME=0 + .ENTRY + + stwm %r3,64(%sp) ; store r3 (may be used by caller) + stw %rp,-20(%sp) ; save return-pointer to frame-marker + + addil LR'to_call-$global$,%r27 + ldw RR'to_call-$global$(%r1),%r3 + +_loop0 + copy %r3,%r22 ; copy the procedure label to r22, ... + .CALL ; ...this is the input to $$dyncall + bl $$dyncall,%mrp ; call $$dyncall (millicode function) + copy %mrp,%rp ; remember the return-pointer + + copy %r3,%r22 + .CALL + bl $$dyncall,%mrp + copy %mrp,%rp + + copy %r3,%r22 + .CALL + bl $$dyncall,%mrp + copy %mrp,%rp + + copy %r3,%r22 + .CALL + bl $$dyncall,%mrp + copy %mrp,%rp + + copy %r3,%r22 + .CALL + bl $$dyncall,%mrp + copy %mrp,%rp + + copy %r3,%r22 + .CALL + bl $$dyncall,%mrp + copy %mrp,%rp + + copy %r3,%r22 + .CALL + bl $$dyncall,%mrp + copy %mrp,%rp + + copy %r3,%r22 + .CALL + bl $$dyncall,%mrp + copy %mrp,%rp + + addibf,<= -8,%arg0,_loop0 ; decrement counter by 8 and loop + nop + + ldw -20(%sp),%rp ; restore return-pointer + bv %r0(%rp) ; return to caller + ldwm -64(%sp),%r3 ; resore r3 and remove stack frame + + .EXIT + .PROCEND + +; Call the procedure `b_null' immediate. + +b_call_imm + .PROC + .CALLINFO CALLER, FRAME=0, SAVE_RP + .ENTRY + + ldo 64(%sp),%sp ; caller needs a stack-frame + stw %rp,-20(%sp) ; save return-pointer to frame-marker + +_loop1 + bl b_null,%rp ; call `b_null' immediate (8 times) + nop + bl b_null,%rp + nop + bl b_null,%rp + nop + bl b_null,%rp + nop + bl b_null,%rp + nop + bl b_null,%rp + nop + bl b_null,%rp + nop + bl b_null,%rp + nop + + addibf,<= -8,%arg0,_loop1 ; decrement counter by 8 and loop + nop + + ldw -20(%sp),%rp ; restore return-pointer + bv %r0(%rp) ; return to caller + ldo -64(%sp),%sp ; remove stack-frame + + .EXIT + .PROCEND + +; Copy register-to-register. +; On PA-RISC this is implemented with an `or'. +; The `or' is hidden by a pseudo-operation called `copy'. + +b_add + .PROC + .CALLINFO NO_CALLS, FRAME=0 + .ENTRY + +_loop2 + copy %r19,%r20 ; copy register-to-register + copy %r20,%r21 ; use caller-saves registers + copy %r21,%r22 + copy %r22,%r21 + copy %r21,%r20 + copy %r20,%r19 + copy %r19,%r20 + copy %r20,%r21 + + addibf,<= -8,%arg0,_loop2 ; decrement counter by 8 and loop + nop + + bv,n %r0(%rp) + + .EXIT + .PROCEND + +; Load memory to a register. + +b_load + .PROC + .CALLINFO NO_CALLS, FRAME=0 + .ENTRY + +_loop3 + ldw -4(%sp),%r22 ; load data from frame-marker + ldw -8(%sp),%r22 ; use a caller-saves register + ldw -12(%sp),%r22 + ldw -16(%sp),%r22 + ldw -20(%sp),%r22 + ldw -24(%sp),%r22 + ldw -28(%sp),%r22 + ldw -32(%sp),%r22 + + addibf,<= -8,%arg0,_loop3 ; decrement counter by 8 and loop + nop + + bv,n %r0(%rp) + + .EXIT + .PROCEND + + + .ALIGN 8 +to_call + .WORD b_null diff --git a/qt/md/i386.README b/qt/md/i386.README new file mode 100644 index 000000000..8ffb92198 --- /dev/null +++ b/qt/md/i386.README @@ -0,0 +1,7 @@ +Note that some machines want labels to have leading underscores, +while others (e.g. System V) do not. Thus, several labels appear +duplicated except for the leading underscore, e.g. + + _qt_cswap: + qt_cswap: + diff --git a/qt/md/i386.h b/qt/md/i386.h new file mode 100644 index 000000000..158fe2703 --- /dev/null +++ b/qt/md/i386.h @@ -0,0 +1,120 @@ +/* + * QuickThreads -- Threads-building toolkit. + * Copyright (c) 1993 by David Keppel + * + * Permission to use, copy, modify and distribute this software and + * its documentation for any purpose and without fee is hereby + * granted, provided that the above copyright notice and this notice + * appear in all copies. This software is provided as a + * proof-of-concept and for demonstration purposes; there is no + * representation about the suitability of this software for any + * purpose. + */ + +#ifndef QT_386_H +#define QT_386_H + +typedef unsigned long qt_word_t; + +/* Thread's initial stack layout on the i386: + + non-varargs: + + +--- + | arg[2] === `userf' on startup + | arg[1] === `pt' on startup + | arg[0] === `pu' on startup + +--- + | ret pc === qt_error + +--- + | ret pc === `only' on startup + +--- + | %ebp + | %esi + | %edi + | %ebx <--- qt_t.sp + +--- + + When a non-varargs thread is started, it ``returns'' directly to + the client's `only' function. + + varargs: + + +--- + | arg[n-1] + | .. + | arg[0] + +--- + | ret pc === `qt_vstart' + +--- + | %ebp === `startup' + | %esi === `cleanup' + | %edi === `pt' + | %ebx === `vuserf' <--- qt_t.sp + +--- + + When a varargs thread is started, it ``returns'' to the `qt_vstart' + startup code. The startup code calls the appropriate functions. */ + + +/* What to do to start a varargs thread running. */ +extern void qt_vstart (void); + + +/* Hold 4 saved regs plus two return pcs (qt_error, qt_start) plus + three args. */ +#define QT_STKBASE (9 * 4) + +/* Hold 4 saved regs plus one return pc (qt_vstart). */ +#define QT_VSTKBASE (5 * 4) + + +/* Stack must be 4-byte aligned. */ +#define QT_STKALIGN (4) + + +/* Where to place various arguments. */ +#define QT_ONLY_INDEX (QT_PC) +#define QT_USER_INDEX (QT_ARG2) +#define QT_ARGT_INDEX (QT_ARG1) +#define QT_ARGU_INDEX (QT_ARG0) + +#define QT_VSTARTUP_INDEX (QT_EBP) +#define QT_VUSERF_INDEX (QT_EBX) +#define QT_VCLEANUP_INDEX (QT_ESI) +#define QT_VARGT_INDEX (QT_EDI) + + +#define QT_EBX 0 +#define QT_EDI 1 +#define QT_ESI 2 +#define QT_EBP 3 +#define QT_PC 4 +/* The following are defined only for non-varargs. */ +#define QT_RPC 5 +#define QT_ARG0 6 +#define QT_ARG1 7 +#define QT_ARG2 8 + + +/* Stack grows down. The top of the stack is the first thing to + pop off (preincrement, postdecrement). */ +#define QT_GROW_DOWN + +extern void qt_error (void); + +/* Push on the error return address. */ +#define QT_ARGS_MD(sto) \ + (QT_SPUT (sto, QT_RPC, qt_error)) + + +/* When varargs are pushed, allocate space for all the args. */ +#define QT_VARGS_MD0(sto, nbytes) \ + ((qt_t *)(((char *)(sto)) - QT_STKROUNDUP(nbytes))) + +#define QT_VARGS_MD1(sto) \ + (QT_SPUT (sto, QT_PC, qt_vstart)) + +#define QT_VARGS_DEFAULT + +#endif /* QT_386_H */ diff --git a/qt/md/i386.s b/qt/md/i386.s new file mode 100644 index 000000000..ed2c533d1 --- /dev/null +++ b/qt/md/i386.s @@ -0,0 +1,108 @@ +/* i386.s -- assembly support. */ + +/* +// QuickThreads -- Threads-building toolkit. +// Copyright (c) 1993 by David Keppel +// +// Permission to use, copy, modify and distribute this software and +// its documentation for any purpose and without fee is hereby +// granted, provided that the above copyright notice and this notice +// appear in all copies. This software is provided as a +// proof-of-concept and for demonstration purposes; there is no +// representation about the suitability of this software for any +// purpose. */ + +/* NOTE: double-labeled `_name' and `name' for System V compatability. */ +/* NOTE: Comment lines start with '/*' and '//' ONLY. Sorry! */ + +/* Callee-save: %esi, %edi, %ebx, %ebp +// Caller-save: %eax, %ecx +// Can't tell: %edx (seems to work w/o saving it.) +// +// Assignment: +// +// See ``i386.h'' for the somewhat unconventional stack layout. */ + + + .text + .align 2 + + .globl _qt_abort + .globl qt_abort + .globl _qt_block + .globl qt_block + .globl _qt_blocki + .globl qt_blocki + +/* These all have the type signature +// +// void *blocking (helper, arg0, arg1, new) +// +// On procedure entry, the helper is at 4(sp), args at 8(sp) and +// 12(sp) and the new thread's sp at 16(sp). It *appears* that the +// calling convention for the 8X86 requires the caller to save all +// floating-point registers, this makes our life easy. */ + +/* Halt the currently-running thread. Save it's callee-save regs on +// to the stack, 32 bytes. Switch to the new stack (next == 16+32(sp)) +// and call the user function (f == 4+32(sp) with arguments: old sp +// arg1 (8+32(sp)) and arg2 (12+32(sp)). When the user function is +// done, restore the new thread's state and return. +// +// `qt_abort' is (currently) an alias for `qt_block' because most of +// the work is shared. We could save the insns up to `qt_common' by +// replicating, but w/o replicating we need an inital subtract (to +// offset the stack as if it had been a qt_block) and then a jump +// to qt_common. For the cost of a jump, we might as well just do +// all the work. +// +// The helper function (4(sp)) can return a void* that is returned +// by the call to `qt_blockk{,i}'. Since we don't touch %eax in +// between, we get that ``for free''. */ + +_qt_abort: +qt_abort: +_qt_block: +qt_block: +_qt_blocki: +qt_blocki: + pushl %ebp /* Save callee-save, sp-=4. */ + pushl %esi /* Save callee-save, sp-=4. */ + pushl %edi /* Save callee-save, sp-=4. */ + pushl %ebx /* Save callee-save, sp-=4. */ + movl %esp, %eax /* Remember old stack pointer. */ + +qt_common: + movl 32(%esp), %esp /* Move to new thread. */ + pushl 28(%eax) /* Push arg 2. */ + pushl 24(%eax) /* Push arg 1. */ + pushl %eax /* Push arg 0. */ + movl 20(%eax), %ebx /* Get function to call. */ + call *%ebx /* Call f. */ + addl $12, %esp /* Pop args. */ + + popl %ebx /* Restore callee-save, sp+=4. */ + popl %edi /* Restore callee-save, sp+=4. */ + popl %esi /* Restore callee-save, sp+=4. */ + popl %ebp /* Restore callee-save, sp+=4. */ + ret /* Resume the stopped function. */ + hlt + + +/* Start a varargs thread. */ + + .globl _qt_vstart + .globl qt_vstart +_qt_vstart: +qt_vstart: + pushl %edi /* Push `pt' arg to `startup'. */ + call *%ebp /* Call `startup'. */ + popl %eax /* Clean up the stack. */ + + call *%ebx /* Call the user's function. */ + + pushl %eax /* Push return from user's. */ + pushl %edi /* Push `pt' arg to `cleanup'. */ + call *%esi /* Call `cleanup'. */ + + hlt /* `cleanup' never returns. */ diff --git a/qt/md/i386_b.s b/qt/md/i386_b.s new file mode 100644 index 000000000..32129a5d1 --- /dev/null +++ b/qt/md/i386_b.s @@ -0,0 +1,30 @@ +/* +// QuickThreads -- Threads-building toolkit. +// Copyright (c) 1993 by David Keppel +// +// Permission to use, copy, modify and distribute this software and +// its documentation for any purpose and without fee is hereby +// granted, provided that the above copyright notice and this notice +// appear in all copies. This software is provided as a +// proof-of-concept and for demonstration purposes; there is no +// representation about the suitability of this software for any +// purpose. */ + + .globl _b_call_reg + .globl b_call_reg + .globl _b_call_imm + .globl b_call_imm + .globl _b_add + .globl b_add + .globl _b_load + .globl b_load + +_b_call_reg: +b_call_reg: +_b_call_imm: +b_call_imm: +_b_add: +b_add: +_b_load: +b_load: + hlt diff --git a/qt/md/ksr1.Makefile b/qt/md/ksr1.Makefile new file mode 100644 index 000000000..aa195839a --- /dev/null +++ b/qt/md/ksr1.Makefile @@ -0,0 +1,6 @@ + +# +# KSR1 configuration. +# +CC = cc -ansi + diff --git a/qt/md/ksr1.h b/qt/md/ksr1.h new file mode 100644 index 000000000..83537a3c2 --- /dev/null +++ b/qt/md/ksr1.h @@ -0,0 +1,164 @@ +/* + * QuickThreads -- Threads-building toolkit. + * Copyright (c) 1993 by David Keppel + * + * Permission to use, copy, modify and distribute this software and + * its documentation for any purpose and without fee is hereby + * granted, provided that the above copyright notice and this notice + * appear in all copies. This software is provided as a + * proof-of-concept and for demonstration purposes; there is no + * representation about the suitability of this software for any + * purpose. + */ + +#ifndef QT_KSR1_H +#define QT_KSR1_H + +/* + Stack layout: + + Registers are saved in strictly low to high order, FPU regs first + (only if qt_block is called), CEU regs second, IPU regs next, with no + padding between the groups. + + Callee-save: f16..f63; c15..c30; i12..i30. + Args passed in i2..i5. + + Note: c31 is a private data pointer. It is not changed on thread + swaps with the assumption that it represents per-processor rather + than per-thread state. + + Note: i31 is an instruction count register that is updated by the + context switch routines. Like c31, it is not changed on context + switches. + + This is what we want on startup: + + + +------ <-- BOS: Bottom of stack (grows down) + | 80 (128 - 48) bytes of padding to a 128-byte boundary + +--- + | only + | userf + | t + | u + | qt_start$TXT + | (empty) <-- qt.sp + +------ <-- (BOS - 128) + + This is why we want this on startup: + + A thread begins running when the restore procedure switches thread stacks + and pops a return address off of the top of the new stack (see below + for the reason why we explicitly store qt_start$TXT). The + block procedure pushes two jump addresses on a thread's stack before + it switches stacks. The first is the return address for the block + procedure, and the second is a restore address. The return address + is used to jump back to the thread that has been switched to; the + restore address is a jump within the block code to restore the registers. + Normally, this is just a jump to the next address. However, on thread + startup, this is a jump to qt_start$TXT. (The block procedure stores + the restore address at an offset of 8 bytes from the top of the stack, + which is also the offset at which qt_start$TXT is stored on the stacks + of new threads. Hence, when the block procedure switches to a new + thread stack, it will initially jump to qt_start$TXT; thereafter, + it jumps to the restore code.) + + qt_start$TXT, after it has read the initial data on the new thread's + stack and placed it in registers, pops the initial stack frame + and gives the thread the entire stack to use for execution. + + The KSR runtime system has an unusual treatment of pointers to + functions. From C, taking the `name' of a function yields a + pointer to a _constant block_ and *not* the address of the + function. The zero'th entry in the constant block is a pointer to + the function. + + We have to be careful: the restore procedure expects a return + address on the top of the stack (pointed to by qt.sp). This is not + a problem when restoring a thread that has run before, since the + block routine would have stored the return address on top of the + stack. However, when ``faking up'' a thread start (bootstrapping a + thread stack frame), the top of the stack needs to contain a + pointer to the code that will start the thread running. + + The pointer to the startup code is *not* `qt_start'. It is the + word *pointed to* by `qt_start'. Thus, we dereference `qt_start', + see QT_ARGS_MD below. + + On varargs startup (still unimplemented): + + | padding to 128 byte boundary + | varargs <-- padded to a 128-byte-boundary + +--- + | caller's frame, 16 bytes + | 80 bytes of padding (frame padded to a 128-byte boundary) + +--- + | cleanup + | vuserf + | startup + | t + +--- + | qt_start <-- qt.sp + +--- + + Of a suspended thread: + + +--- + | caller's frame, 16 bytes + | fpu registers 47 regs * 8 bytes/reg 376 bytes + | ceu registers 16 regs * 8 bytes/reg 128 bytes + | ipu registers 19 regs * 8 bytes/reg 152 bytes + | : + | 80 bytes of padding + | : + | qt_restore <-- qt.sp + +--- + + */ + + +#define QT_STKALIGN 128 +#define QT_GROW_DOWN +typedef unsigned long qt_word_t; + +#define QT_STKBASE QT_STKALIGN +#define QT_VSTKBASE QT_STKBASE + +extern void qt_start(void); +/* + * See the discussion above for what indexing into a procedure ptr + * does for us (it's lovely, though, isn't it?). + * + * This assumes that the address of a procedure's code is the + * first word in a procedure's constant block. That's how the manual + * says it will be arranged. + */ +#define QT_ARGS_MD(sp) (QT_SPUT (sp, 1, ((qt_word_t *)qt_start)[0])) + +/* + * The *index* (positive offset) of where to put each value. + * See the picture of the stack above that explains the offsets. + */ +#define QT_ONLY_INDEX (5) +#define QT_USER_INDEX (4) +#define QT_ARGT_INDEX (3) +#define QT_ARGU_INDEX (2) + +#define QT_VARGS_DEFAULT +#define QT_VARGS(sp, nb, vargs, pt, startup, vuserf, cleanup) \ + (qt_vargs (sp, nbytes, &vargs, pt, startup, vuserf, cleanup)) + + +#define QT_VARGS_MD0(sp, vabytes) \ + ((qt_t *)(((char *)(sp)) - 4*8 - QT_STKROUNDUP(vabytes))) + +extern void qt_vstart(void); +#define QT_VARGS_MD1(sp) (QT_SPUT (sp, 0, ((qt_word_t *)qt_vstart)[0])) + +#define QT_VCLEANUP_INDEX (4) +#define QT_VUSERF_INDEX (3) +#define QT_VSTARTUP_INDEX (2) +#define QT_VARGT_INDEX (1) + +#endif /* def QT_KSR1_H */ diff --git a/qt/md/ksr1.s b/qt/md/ksr1.s new file mode 100644 index 000000000..d4d51a0a6 --- /dev/null +++ b/qt/md/ksr1.s @@ -0,0 +1,424 @@ +/* + * QuickThreads -- Threads-building toolkit. + * Copyright (c) 1993 by David Keppel + * + * Permission to use, copy, modify and distribute this software and + * its documentation for any purpose and without fee is hereby + * granted, provided that the above copyright notice and this notice + * appear in all copies. This software is provided as a + * proof-of-concept and for demonstration purposes; there is no + * representation about the suitability of this software for any + * purpose. + */ + + .file "ksr1.s" + .def .debug; .endef + + .align 128 + .globl qt_blocki + .globl qt_blocki$TXT + .globl qt_block + .globl qt_block$TXT + .globl qt_start$TXT + .globl qt_start + .globl qt_abort$TXT + .globl qt_abort + .globl qt_vstart + .globl qt_vstart$TXT + +# +# KSR convention: on procedure calls, load both the procedure address +# and a pointer to a constant block. The address of function `f' is +# `f$TXT', and the constant block address is `f'. The constant block +# has several reserved values: +# +# 8 bytes fpu register save mask +# 4 bytes ipu register save mask +# 4 bytes ceu register save mask +# f: f$TXT +# ... whatever you want ... (not quite...read on) +# +# Note, by the way, that a pointer to a function is passed as a +# pointer to the constant area, and the constant area has the text +# address. +# + +# +# Procedures that do not return structures prefix their code with +# +# proc$TXT: +# finop; cxnop +# finop; cxnop +# <proc code> +# +# Calls to those procedures branch to a 16 byte offset (4 instrs) in +# to the procedure to skip those instructions. +# +# Procedures that return structures use a different code prefix: +# +# proc$TXT: +# finop; beq.qt %rc, %rc, 24 # return value entry +# finop; cxnop +# finop; movi8 0, %rc # no return value entry +# <proc code> +# +# Calls that want the returned structure branch directly to the +# procedure address. Callers that don't want (or aren't expecting) a +# return value branche 16 bytes in to the procedure, which will zero +# %rc, telling the called procedure not to return a structure. +# + +# +# On entry: +# %i2 -- control block of helper function to run +# (dereference to get helper) +# %i3 -- a1 +# %i4 -- a2 +# %i5 -- sp of new to run +# + + .data + .half 0x0, 0x0, 0x7ffff000, 0x7fff8000 +qt_blocki: +qt_abort: + .word qt_blocki$TXT + .word qt_restore$TXT + + .text +qt_abort$TXT: +qt_blocki$TXT: + finop ; cxnop # entry prefix + finop ; cxnop # entry prefix + add8.ntr 75,%i31,%i31 ; movi8 512,%c5 # ICR; stk adjust + finop ; ssub8.ntr 0,%sp,%c5,%sp + finop ; st8 %fp,504(%sp) # Save caller's fp + finop ; st8 %cp,496(%sp) # Save caller's cp + finop ; ld8 8(%c10),%c5 # ld qt_restore$TXT + finop ; st8 %c14,0(%sp) # Save special ret addr + finop ; mov8_8 %c10, %cp # Our cp + finop ; sadd8.ntr 0,%sp,%c5,%fp # Our frame ptr + finop ; st8 %c5,8(%sp) # st qt_restore$TXT +# +# CEU registers %c15-%c24, %c26-%c30 (%c14 we restore later) +# + finop ; st8 %c15,456(%sp) + finop ; st8 %c16,448(%sp) + finop ; st8 %c17,440(%sp) + finop ; st8 %c18,432(%sp) + finop ; st8 %c19,424(%sp) + finop ; st8 %c20,416(%sp) + finop ; st8 %c21,408(%sp) + finop ; st8 %c22,400(%sp) + finop ; st8 %c23,392(%sp) + finop ; st8 %c24,384(%sp) +# +# %c25 is the Enclosing Frame Pointer (EFP) -- since C doesn't +# use nested procedures, we ignore it (leaving a gap, though) +# + finop ; st8 %c26,368(%sp) + finop ; st8 %c27,360(%sp) + finop ; st8 %c28,352(%sp) + finop ; st8 %c29,344(%sp) + finop ; st8 %c30,336(%sp) +# +# IPU registers %i12-%i30 +# + finop ; st8 %i12,328(%sp) + finop ; st8 %i13,320(%sp) + finop ; st8 %i14,312(%sp) + finop ; st8 %i15,304(%sp) +# (gap to get alignment for st64) +# -- Doesn't work on version 1.1.3 of the OS +# finop ; st64 %i16,256(%sp) + + finop ; st8 %i16,256(%sp) + finop ; st8 %i17,248(%sp) + finop ; st8 %i18,240(%sp) + finop ; st8 %i19,232(%sp) + finop ; st8 %i20,224(%sp) + finop ; st8 %i21,216(%sp) + finop ; st8 %i22,208(%sp) + finop ; st8 %i23,200(%sp) + finop ; st8 %i24,192(%sp) + finop ; st8 %i25,184(%sp) + finop ; st8 %i26,176(%sp) + finop ; st8 %i27,168(%sp) + finop ; st8 %i28,160(%sp) + finop ; st8 %i29,152(%sp) + finop ; st8 %i30,144(%sp) +# +# FPU already saved, or saving not necessary +# + +# +# Switch to the stack passed in as fourth argument to the block +# routine (%i5) and call the helper routine passed in as the first +# argument (%i2). Note that the address of the helper's constant +# block is passed in, so we must derefence it to get the helper's text +# address. +# + finop ; movb8_8 %i2,%c10 # helper's ConstBlock + finop ; cxnop # Delay slot, fill w/ + finop ; cxnop # .. 2 st8 from above + finop ; ld8 0(%c10),%c4 # load addr of helper + finop ; movb8_8 %sp, %i2 # 1st arg to helper + # is this stack; other + # args remain in regs + finop ; movb8_8 %i5,%sp # switch stacks + finop ; jsr %c14,16(%c4) # call helper + movi8 3, %i0 ; movi8 0,%c8 # nargs brain dmg + finop ; cxnop + finop ; cxnop +# +# Here is where behavior differs for threads being restored and threads +# being started. Blocked threads have a pointer to qt_restore$TXT on +# the top of their stacks; manufactured stacks have a pointer to qt_start$TXT +# on the top of their stacks. With this setup, starting threads +# skip the (unecessary) restore operations. +# +# We jump to an offset of 16 to either (1) skip past the two noop pairs +# at the start of qt_start$TXT, or (2) skip past the two noop pairs +# after qt_restore$TXT. +# + finop ; ld8 8(%sp),%c4 + finop ; cxnop + finop ; cxnop + finop ; jmp 16(%c4) +qt_restore$TXT: + finop ; cxnop + finop ; cxnop +# +# Point of Restore: +# +# The helper funtion will return here. Any result it has placed in +# a return register (most likely %i0) will not get overwritten below +# and will consequently be the return value of the blocking routine. +# + +# +# CEU registers %c15-%c24, %c26-%c30 (%c14 we restore later) +# + finop ; ld8 456(%sp),%c15 + finop ; ld8 448(%sp),%c16 + finop ; ld8 440(%sp),%c17 + finop ; ld8 432(%sp),%c18 + finop ; ld8 424(%sp),%c19 + finop ; ld8 416(%sp),%c20 + finop ; ld8 408(%sp),%c21 + finop ; ld8 400(%sp),%c22 + finop ; ld8 392(%sp),%c23 + finop ; ld8 384(%sp),%c24 +# +# %c25 is the Enclosing Frame Pointer (EFP) -- since C doesn't +# use nested procedures, we ignore it (leaving a gap, though) +# + finop ; ld8 368(%sp),%c26 + finop ; ld8 360(%sp),%c27 + finop ; ld8 352(%sp),%c28 + finop ; ld8 344(%sp),%c29 + finop ; ld8 336(%sp),%c30 +# +# IPU registers %i12-%i30 +# + finop ; ld8 328(%sp),%i12 + finop ; ld8 320(%sp),%i13 + finop ; ld8 312(%sp),%i14 + finop ; ld8 304(%sp),%i15 +# (gap to get alignment for ld64) +# -- Doesn't work on version 1.1.3 of the OS +# finop ; ld64 256(%sp),%i16 + + finop ; ld8 256(%sp),%i16 + finop ; ld8 248(%sp),%i17 + finop ; ld8 240(%sp),%i18 + finop ; ld8 232(%sp),%i19 + finop ; ld8 224(%sp),%i20 + finop ; ld8 216(%sp),%i21 + finop ; ld8 208(%sp),%i22 + finop ; ld8 200(%sp),%i23 + finop ; ld8 192(%sp),%i24 + finop ; ld8 184(%sp),%i25 + finop ; ld8 176(%sp),%i26 + finop ; ld8 168(%sp),%i27 + finop ; ld8 160(%sp),%i28 + finop ; ld8 152(%sp),%i29 + finop ; ld8 144(%sp),%i30 + +# +# FPU registers don't need to be loaded, or will be loaded by an +# enclosing scope (e.g., if this is called by qt_block). +# + +# +# Load the special registers. We don't load the stack ptr because +# the new stack is passed in as an argument, we don't load the EFP +# because we don't use it, and we load the return address specially +# off the top of the stack. +# + finop ; ld8 0(%sp),%c14 # return addr + finop ; ld8 496(%sp),%cp + finop ; ld8 504(%sp),%fp + + finop ; jmp 32(%c14) # jump back to thread + finop ; movi8 512,%c5 # stack adjust + finop ; sadd8.ntr 0,%sp,%c5,%sp + + .data + .half 0x0, 0x0, 0x7ffff000, 0x7fff8000 +qt_block: + .word qt_block$TXT + .word qt_error + .word qt_error$TXT + .word qt_blocki +# +# Handle saving and restoring the FPU regs, relying on qt_blocki +# to save and restore the remaining registers. +# + .text +qt_block$TXT: + finop ; cxnop # entry prefix + finop ; cxnop # entry prefix + + add8.ntr 29,%i31,%i31 ; movi8 512,%c5 # ICR; stk adjust + finop ; ssub8.ntr 0,%sp,%c5,%sp + finop ; st8 %fp,504(%sp) # Save caller's fp + finop ; st8 %cp,496(%sp) # Save caller's cp + finop ; st8 %c14,488(%sp) # store ret addr + finop ; sadd8.ntr 0,%sp,%c5,%fp # Our frame ptr + finop ; mov8_8 %c10, %cp # Our cp + +# +# Store 8 registers at once...destination must be a multiple of 64 +# + finop ; st64 %f16,384(%sp) + finop ; st64 %f24,320(%sp) + finop ; st64 %f32,256(%sp) + finop ; st64 %f40,192(%sp) + finop ; st64 %f48,128(%sp) + finop ; st64 %f56,64(%sp) + +# +# Call the integer blocking routine, passing the arguments passed to us +# + finop ; ld8 24(%cp), %c10 + finop ; cxnop + finop ; jsr %c14, qt_blocki$TXT + finop ; cxnop + finop ; cxnop + movi8 4,%i0 ; movi8 0,%c8 # nargs brain dmg + +# +# Load 8 registers at once...source must be a multiple of 64 +# + finop ; ld64 64(%sp),%f56 + finop ; ld64 128(%sp),%f48 + finop ; ld64 192(%sp),%f40 + finop ; ld64 256(%sp),%f32 + finop ; ld64 320(%sp),%f24 + finop ; ld64 384(%sp),%f16 + + finop ; ld8 488(%sp),%c14 + finop ; ld8 496(%sp),%cp + finop ; ld8 504(%sp),%fp + finop ; jmp 32(%c14) # jump back to thread + finop ; movi8 512,%c5 # stack adjust + finop ; sadd8.ntr 0,%sp,%c5,%sp + + + .data + .half 0x0, 0x0, 0x7ffff000, 0x7fff8000 +qt_start: + .word qt_start$TXT +# +# A new thread is set up to "appear" as if it were executing code at +# the beginning of qt_start and then it called a blocking routine +# (qt_blocki). So when a new thread starts to run, it gets unblocked +# by the code above and "returns" to `qt_start$TXT' in the +# restore step of the switch. Blocked threads jump to 16(qt_restore$TXT), +# and starting threads jump to 16(qt_start$TXT). +# + .text +qt_start$TXT: + finop ; cxnop # + finop ; cxnop # + finop ; ld8 40(%sp),%c10 # `only' constant block + finop ; ld8 32(%sp),%i4 # `userf' arg. + finop ; ld8 24(%sp),%i3 # `t' arg. + finop ; ld8 0(%c10),%c4 # `only' text location + finop ; ld8 16(%sp),%i2 # `u' arg. + finop ; cxnop + finop ; jsr %c14,16(%c4) # call `only' +# +# Pop the frame used to store the thread's initial data +# + finop ; sadd8.ntr 0,%sp,128,%sp + finop ; cxnop + movi8 2,%i0 ; movi8 0,%c8 # nargs brain dmg +# +# If we ever return, it's an error. +# + finop ; jmp qt_error$TXT + finop ; cxnop + finop ; cxnop + movi8 0,%i0 ; movi8 0,%c8 # nargs brain dmg + + +# +# This stuff is broken +# + .data + .half 0x0, 0x0, 0x7ffff000, 0x7fff8000 +qt_vstart: + .word qt_vstart$TXT + + .text +qt_vstart$TXT: + finop ; cxnop # entry prefix + finop ; cxnop # entry prefix + finop ; cxnop + finop ; cxnop + add8.ntr 11,%i31,%i31 ; movi8 512,%c5 + finop ; ssub8.ntr 0,%sp,%c5,%sp # fix stack + finop ; ld8 8(%sp),%i2 # load `t' as arg to + finop ; cxnop # `startup' + finop ; cxnop + finop ; ld8 16(%sp),%c10 # `startup' const block + finop ; cxnop + finop ; cxnop + finop ; ld8 0(%c10),%c4 # `startup' text loc. + finop ; cxnop + finop ; cxnop + finop ; jsr %c14,16(%c4) # call `startup' + finop ; cxnop + finop ; cxnop + movi8 1, %i0 ; movi8 0,%c8 # nargs brain dmg +# +# finop ; sadd 0,%sp,128,%sp # alter stack +# + finop ; ld8 8(%sp),%i2 # load `t' as arg to + finop ; ld8 8(%sp),%i2 # load `t' as arg to + finop ; ld8 8(%sp),%i2 # load `t' as arg to + finop ; ld8 8(%sp),%i2 # load `t' as arg to + + finop ; ld8 32(%sp),%c10 # `only' constant block + finop ; ld8 8(%sp),%i2 # `u' arg. + finop ; ld8 16(%sp),%i3 # `t' arg. + finop ; ld8 0(%c10),%c4 # `only' text location + finop ; ld8 24(%sp),%i4 # `userf' arg. + finop ; cxnop + finop ; jsr %c4,16(%c4) # call `only' + finop ; cxnop + finop ; cxnop +# +# If the callee ever calls `nargs', the following instruction (pair) +# will be executed. However, we don't know when we compile this code +# how many args are being passed. So we give our best guess: 0. +# + movi8 0,%i0 ; movi8 0,%c8 # nargs brain dmg +# +# If we ever return, it's an error. +# + finop ; jmp qt_error$TXT + finop ; cxnop + finop ; cxnop + movi8 0,%i0 ; movi8 0,%c8 # nargs brain dmg diff --git a/qt/md/ksr1_b.s b/qt/md/ksr1_b.s new file mode 100644 index 000000000..80b0c59eb --- /dev/null +++ b/qt/md/ksr1_b.s @@ -0,0 +1,49 @@ +/* + * QuickThreads -- Threads-building toolkit. + * Copyright (c) 1993 by David Keppel + * + * Permission to use, copy, modify and distribute this software and + * its documentation for any purpose and without fee is hereby + * granted, provided that the above copyright notice and this notice + * appear in all copies. This software is provided as a + * proof-of-concept and for demonstration purposes; there is no + * representation about the suitability of this software for any + * purpose. + */ + + .file "ksr1_b.s" + .def .debug; .endef + + .globl b_call_reg$TXT + .globl b_call_reg + .globl b_call_imm$TXT + .globl b_call_imm + .globl b_add$TXT + .globl b_add + .globl b_load$TXT + .globl b_load + + +b_call_reg: +b_call_imm: +b_add: +b_load: + .word b_call_reg$TXT + .word qt_error + .word qt_error$TXT + + +b_call_reg$TXT: +b_call_imm$TXT: +b_add$TXT: +b_load$TXT: + finop ; cxnop + finop ; cxnop + finop ; ld8 16(%cp),%c4 + finop ; ld8 8(%cp),%cp + finop ; cxnop + finop ; cxnop + finop ; jsr %c4,0(%c4) + finop ; cxnop + finop ; cxnop + diff --git a/qt/md/m88k.Makefile b/qt/md/m88k.Makefile new file mode 100644 index 000000000..608c70690 --- /dev/null +++ b/qt/md/m88k.Makefile @@ -0,0 +1,6 @@ + +# +# Hosted compilers for 88k for Meerkat. +# +CC = gcc88 -Dm88k -ansi -pedantic -Wall -fno-builtin +AS = as88 diff --git a/qt/md/m88k.c b/qt/md/m88k.c new file mode 100644 index 000000000..9e3ae8ba8 --- /dev/null +++ b/qt/md/m88k.c @@ -0,0 +1,111 @@ +/* + * QuickThreads -- Threads-building toolkit. + * Copyright (c) 1993 by David Keppel + * + * Permission to use, copy, modify and distribute this software and + * its documentation for any purpose and without fee is hereby + * granted, provided that the above copyright notice and this notice + * appear in all copies. This software is provided as a + * proof-of-concept and for demonstration purposes; there is no + * representation about the suitability of this software for any + * purpose. + */ + +#include <stdarg.h> +#include "qt.h" + +/* Varargs is harder on the m88k. Parameters are saved on the stack as + something like (stack grows down to low memory; low at bottom of + picture): + + | : + | arg8 <-- va_list.__va_stk + +--- + | : + +--- + | arg7 + | : + | iarg0 <-- va_list.__va_reg + +--- + | : + | va_list { __va_arg, __va_stk, __va_reg } + | : + +--- + + Here, `va_list.__va_arg' is the number of word-size arguments + that have already been skipped. Doubles must be double-arligned. + + What this means for us is that the user's routine needs to be + called with an arg list where some of the words in the `__va_stk' + part of the parameter list have to be promoted to registers. + + BUG: doubleword register arguments must be double-aligned. If + something is passed as an even # arg and used as an odd # arg or + vice-versa, the code in the called routine (in the new thread) that + decides how to adjust the index will get it wrong, because it will + be expect it to be, say, doubleword aligned and it will really be + singleword aligned. + + I'm not sure you can solve this without knowing the types of all + the arguments. All in all, we never promised varargs would work + reliably. */ + + + +#define QT_VADJ(sp) (((char *)sp) - QT_VSTKBASE) + +/* Always allocate at least enough space for 8 args; waste some space + at the base of the stack to ensure the startup routine doesn't read + off the end of the stack. */ + +#define QT_VARGS_MD0(sp, vabytes) \ + ((qt_t *)(((char *)(sp)) - 8*4 - QT_STKROUNDUP(vabytes))) + +extern void qt_vstart(void); +#define QT_VARGS_MD1(sp) (QT_SPUT (sp, QT_1, qt_vstart)) + + + struct qt_t * +qt_vargs (struct qt_t *qsp, int nbytes, void *vargs, + void *pt, qt_function_t *startup, + qt_function_t *vuserf, qt_function_t *cleanup) +{ + va_list ap; + int i; + int n; /* Number of words into original arg list. */ + qt_word_t *sp; + int *reg; /* Where to read passed-in-reg args. */ + int *stk; /* Where to read passed-on-stk args. */ + + ap = *(va_list *)vargs; + qsp = QT_VARGS_MD0 (qsp, nbytes); + sp = (qt_word_t *)qsp; + + reg = (ap.__va_arg < 8) + ? &ap.__va_reg[ap.__va_arg] + : 0; + stk = &ap.__va_stk[8]; + n = ap.__va_arg; + for (i=0; i<nbytes/sizeof(qt_word_t) && n<8; ++i,++n) { + sp[i] = *reg++; + } + for (; i<nbytes/sizeof(qt_word_t); ++i) { + sp[i] = *stk++; + } + +#ifdef QT_NDEF + for (i=0; i<nbytes/sizeof(qt_word_t); ++i) { + sp[i] = (n < 8) + ? *reg++ + : *stk++; + ++n; + } +#endif + + QT_VARGS_MD1 (QT_VADJ(sp)); + QT_SPUT (QT_VADJ(sp), QT_VARGT_INDEX, pt); + QT_SPUT (QT_VADJ(sp), QT_VSTARTUP_INDEX, startup); + QT_SPUT (QT_VADJ(sp), QT_VUSERF_INDEX, vuserf); + QT_SPUT (QT_VADJ(sp), QT_VCLEANUP_INDEX, cleanup); + return ((qt_t *)QT_VADJ(sp)); +} diff --git a/qt/md/m88k.h b/qt/md/m88k.h new file mode 100644 index 000000000..df7e07a85 --- /dev/null +++ b/qt/md/m88k.h @@ -0,0 +1,159 @@ +/* + * QuickThreads -- Threads-building toolkit. + * Copyright (c) 1993 by David Keppel + * + * Permission to use, copy, modify and distribute this software and + * its documentation for any purpose and without fee is hereby + * granted, provided that the above copyright notice and this notice + * appear in all copies. This software is provided as a + * proof-of-concept and for demonstration purposes; there is no + * representation about the suitability of this software for any + * purpose. + */ + +#ifndef QT_M88K_H +#define QT_M88K_H + +typedef unsigned long qt_word_t; + +#define QT_GROW_DOWN + +/* Stack layout on the mips: + + Callee-save registers are: $16-$23, $30; $f20-$f30. + Also save $31, return pc. + + Non-varargs: + + +--- + | r30 (fp) on startup === 0 + | r25 + | r24 + | r23 + | r22 + | r21 + | r20 + | r19 + | r18 + | r17 on startup === `only' + | r16 on startup === `userf' + | r15 on startup === `pt' + | r14 on startup === `pu' + | r1 on startup === `qt_start' + | 0 + | 0 + +--- + | 0 + | ... (8 regs worth === 32 bytes of homing area) + | 0 <--- sp + +--- + + Conventions for varargs: + + | : + | arg8 + +--- + | r30 (fp) arg7 + | r25 arg6 + | r24 arg5 + | r23 arg4 + | r22 arg3 + | r21 arg2 + | r20 arg1 + | r19 arg0 + | r18 + | r17 on startup === `startup' + | r16 on startup === `vuserf' + | r15 on startup === `pt' + | r14 on startup === `cleanup' + | r1 on startup === `qt_vstart' + | 0 + | 0 + +--- + | 0 + | ... (8 regs worth === 32 bytes of homing area) + | 0 <--- sp + +--- + + */ + + +/* Stack must be doubleword aligned. */ +#define QT_STKALIGN (16) /* Doubleword aligned. */ + +/* How much space is allocated to hold all the crud for + initialization: saved registers plus padding to keep the stack + aligned plus 8 words of padding to use as a `homing area' (for + r2-r9) when calling helper functions on the stack of the (not yet + started) thread. The varargs save area is small because it gets + overlapped with the top of the parameter list. In case the + parameter list is less than 8 args, QT_ARGS_MD0 adds some dead + space at the top of the stack. */ + +#define QT_STKBASE (16*4 + 8*4) +#define QT_VSTKBASE (8*4 + 8*4) + + +/* Index of various registers. */ +#define QT_1 (8+2) +#define QT_14 (8+3) +#define QT_15 (8+4) +#define QT_16 (8+5) +#define QT_17 (8+6) +#define QT_30 (8+15) + + +/* When a never-before-run thread is restored, the return pc points + to a fragment of code that starts the thread running. For + non-vargs functions, it sets up arguments and calls the client's + `only' function. For varargs functions, the startup code calls the + startup, user, and cleanup functions. + + For non-varargs functions, we set the frame pointer to 0 to + null-terminate the call chain. + + For varargs functions, the frame pointer register is used to hold + one of the arguments, so that all arguments can be laid out in + memory by the conventional `qt_vargs' varargs initialization + routine. + + The varargs startup routine always reads 8 words of arguments from + the stack. If there are less than 8 words of arguments, then the + arg list could call off the top of the stack. To prevent fall-off, + always allocate 8 words. */ + +extern void qt_start(void); +#define QT_ARGS_MD(sp) \ + (QT_SPUT (sp, QT_1, qt_start), \ + QT_SPUT (sp, QT_30, 0)) + + +/* The m88k uses a struct for `va_list', so pass a pointer to the + struct. */ + +typedef void (qt_function_t)(void); + +struct qt_t; +extern struct qt_t *qt_vargs (struct qt_t *sp, int nbytes, + void *vargs, void *pt, + qt_function_t *startup, + qt_function_t *vuserf, + qt_function_t *cleanup); + +#define QT_VARGS(sp, nbytes, vargs, pt, startup, vuserf, cleanup) \ + (qt_vargs (sp, nbytes, &(vargs), pt, (qt_function_t *)startup, \ + (qt_function_t *)vuserf, (qt_function_t *)cleanup)) + + +/* The *index* (positive offset) of where to put each value. */ +#define QT_ONLY_INDEX (QT_17) +#define QT_USER_INDEX (QT_16) +#define QT_ARGT_INDEX (QT_15) +#define QT_ARGU_INDEX (QT_14) + +#define QT_VCLEANUP_INDEX (QT_14) +#define QT_VUSERF_INDEX (QT_16) +#define QT_VSTARTUP_INDEX (QT_17) +#define QT_VARGT_INDEX (QT_15) + +#endif /* ndef QT_M88K_H */ diff --git a/qt/md/m88k.s b/qt/md/m88k.s new file mode 100644 index 000000000..42467e8d5 --- /dev/null +++ b/qt/md/m88k.s @@ -0,0 +1,132 @@ +/* m88k.s -- assembly support. */ + +/* + * QuickThreads -- Threads-building toolkit. + * Copyright (c) 1993 by David Keppel + * + * Permission to use, copy, modify and distribute this software and + * its documentation for any purpose and without fee is hereby + * granted, provided that the above copyright notice and this notice + * appear in all copies. This software is provided as a + * proof-of-concept and for demonstration purposes; there is no + * representation about the suitability of this software for any + * purpose. + */ + +/* Callee-save r14..r25, r31(sp), r30(fp). r1 === return pc. + * Argument registers r2..r9, return value r2..r3. + * + * On startup, restore regs so retpc === call to a function to start. + * + * We're going to call a function (r2) from within the context switch + * routine. Call it on the new thread's stack on behalf of the old + * thread. + */ + + .globl _qt_block + .globl _qt_blocki + .globl _qt_abort + .globl _qt_start + .globl _qt_vstart + + /* + ** r2: ptr to function to call once curr is suspended + ** and control is on r5's stack. + ** r3: 1'th arg to *r2. + ** r4: 2'th arg to *r2. + ** r5: sp of thread to suspend. + ** + ** The helper routine returns a value that is passed on as the + ** return value from the blocking routine. Since we don't + ** touch r2 between the helper's return and the end of + ** function, we get this behavior for free. + ** + ** Same entry for integer-only and floating-point, since there + ** are no separate integer and floating-point registers. + ** + ** Each procedure call sets aside a ``home region'' of 8 regs + ** for r2-r9 for varargs. For context switches we don't use + ** the ``home region'' for varargs so use it to save regs. + ** Allocate 64 bytes of save space -- use 32 bytes of register + ** save area passed in to us plus 32 bytes we allcated, use + ** the other 32 bytes for save area for a save area to call + ** the helper function. + */ +_qt_block: +_qt_blocki: + sub r31, r31,64 /* Allocate reg save space. */ + st r1, r31,8+32 /* Save callee-save registers. */ + st r14, r31,12+32 + st.d r15, r31,16+32 + st.d r17, r31,24+32 + st.d r19, r31,32+32 + st.d r21, r31,40+32 + st.d r23, r31,48+32 + st r25, r31,56+32 + st r30, r31,60+32 + +_qt_abort: + addu r14, r31,0 /* Remember old sp. */ + addu r31, r5,0 /* Set new sp. */ + jsr.n r2 /* Call helper. */ + addu r2, r14,0 /* Pass old sp as an arg0 to helper. */ + + ld r1, r31,8+32 /* Restore callee-save registers. */ + ld r14, r31,12+32 + ld.d r15, r31,16+32 + ld.d r17, r31,24+32 + ld.d r19, r31,32+32 + ld.d r21, r31,40+32 + ld.d r23, r31,48+32 + ld r25, r31,56+32 + ld r30, r31,60+32 + + jmp.n r1 /* Return to new thread's caller. */ + addu r31, r31,64 /* Free register save space. */ + + + /* + ** Non-varargs thread startup. + ** See `m88k.h' for register use conventions. + */ +_qt_start: + addu r2, r14,0 /* Set user arg `pu'. */ + addu r3, r15,0 /* ... user function pt. */ + jsr.n r17 /* Call `only'. */ + addu r4, r16,0 /* ... user function userf. */ + + bsr _qt_error /* `only' erroniously returned. */ + + + /* + ** Varargs thread startup. + ** See `m88k.h' for register use conventions. + ** + ** Call the `startup' function with just argument `pt'. + ** Then call `vuserf' with 8 register args plus any + ** stack args. + ** Then call `cleanup' with `pt' and the return value + ** from `vuserf'. + */ +_qt_vstart: + addu r18, r30,0 /* Remember arg7 to `vuserf'. */ + addu r30, r0,0 /* Null-terminate call chain. */ + + jsr.n r17 /* Call `startup'. */ + addu r2, r15,0 /* `pt' is arg0 to `startup'. */ + + addu r2, r19,0 /* Set arg0. */ + addu r3, r20,0 /* Set arg1. */ + addu r4, r21,0 /* Set arg2. */ + addu r5, r22,0 /* Set arg3. */ + addu r6, r23,0 /* Set arg4. */ + addu r7, r24,0 /* Set arg5. */ + addu r8, r25,0 /* Set arg6. */ + jsr.n r16 /* Call `vuserf'. */ + addu r9, r18,0 /* Set arg7. */ + + addu r3, r2,0 /* Ret. value is arg1 to `cleanup'. */ + jsr.n r14 /* Call `cleanup'. */ + addu r2, r15,0 /* `pt' is arg0 to `cleanup'. */ + + bsr _qt_error /* `cleanup' erroniously returned. */ diff --git a/qt/md/m88k_b.s b/qt/md/m88k_b.s new file mode 100644 index 000000000..1926e6ae8 --- /dev/null +++ b/qt/md/m88k_b.s @@ -0,0 +1,117 @@ +/* + * QuickThreads -- Threads-building toolkit. + * Copyright (c) 1993 by David Keppel + * + * Permission to use, copy, modify and distribute this software and + * its documentation for any purpose and without fee is hereby + * granted, provided that the above copyright notice and this notice + * appear in all copies. This software is provided as a + * proof-of-concept and for demonstration purposes; there is no + * representation about the suitability of this software for any + * purpose. + */ + + .text + .globl _b_call_reg + .globl _b_call_imm + .globl _b_add + .globl _b_load + +_b_null: + jmp r1 + +_b_call_reg: + subu r31, r31,8 /* Alloc ret pc save space. */ + st r1, r31,32 /* Save ret pc. */ + or.u r3, r0,hi16(_b_null) /* Put call addr in a reg. */ + or r3, r3,lo16(_b_null) + jsr r3 +L0: + jsr r3 + jsr r3 + jsr r3 + jsr.n r3 + subu r2, r2,5 /* Decrement #of iter to go. */ + bcnd.n gt0,r2,L0 + jsr r3 + + ld r1, r31,32 + jmp r1 + + +_b_call_imm: + subu r31, r31,8 /* Alloc ret pc save space. */ + st r1, r31,32 /* Save ret pc. */ + bsr _b_null +L1: + bsr _b_null + bsr _b_null + bsr _b_null + bsr.n _b_null + subu r2, r2,5 /* Decrement #of iter to go. */ + bcnd.n gt0,r2,L1 + bsr _b_null + + ld r1, r31,32 + jmp r1 + +_b_add: + add r0, r3,r4 +L2: + add r3, r4,r5 + add r4, r5,r6 + add r5, r6,r7 + add r8, r9,r0 + add r0, r3,r4 + add r3, r4,r5 + add r4, r5,r6 + add r5, r6,r7 + add r8, r9,r0 + + add r0, r3,r4 + add r3, r4,r5 + add r4, r5,r6 + add r5, r6,r7 + add r8, r9,r0 + add r0, r3,r4 + add r3, r4,r5 + add r4, r5,r6 + add r5, r6,r7 + add r8, r9,r0 + + subu r2, r2,20 /* Decrement #of iter to go. */ + bcnd.n gt0,r2,L2 + add r0, r3,r4 + + jmp r1 + + +_b_load: + ld r0, r31,0 +L3: + ld r3, r31,4 + ld r4, r31,8 + ld r5, r31,12 + ld r6, r31,16 + ld r0, r31,0 + ld r3, r31,4 + ld r4, r31,8 + ld r5, r31,12 + ld r6, r31,16 + + ld r0, r31,0 + ld r3, r31,4 + ld r4, r31,8 + ld r5, r31,12 + ld r6, r31,16 + ld r0, r31,0 + ld r3, r31,4 + ld r4, r31,8 + ld r5, r31,12 + ld r6, r31,16 + + subu r2, r2,20 /* Decrement #of iter to go. */ + bcnd.n gt0,r2,L3 + ld r0, r31,0 + + jmp r1 diff --git a/qt/md/mips-irix5.s b/qt/md/mips-irix5.s new file mode 100644 index 000000000..234a953ed --- /dev/null +++ b/qt/md/mips-irix5.s @@ -0,0 +1,182 @@ +/* mips.s -- assembly support. */ + +/* + * QuickThreads -- Threads-building toolkit. + * Copyright (c) 1993 by David Keppel + * + * Permission to use, copy, modify and distribute this software and + * its documentation for any purpose and without fee is hereby + * granted, provided that the above copyright notice and this notice + * appear in all copies. This software is provided as a + * proof-of-concept and for demonstration purposes; there is no + * representation about the suitability of this software for any + * purpose. + */ + +/* Callee-save $16-$23, $30-$31. + * + * $25 is used as a procedure value pointer, used to discover constants + * in a callee. Thus, each caller here sets $25 before the call. + * + * On startup, restore regs so retpc === call to a function to start. + * We're going to call a function ($4) from within this routine. + * We're passing 3 args, therefore need to allocate 12 extra bytes on + * the stack for a save area. The start routine needs a like 16-byte + * save area. Must be doubleword aligned (_mips r3000 risc + * architecture_, gerry kane, pg d-23). + */ + +/* + * Modified by Assar Westerlund <assar@sics.se> to support Irix 5.x + * calling conventions for dynamically-linked code. + */ + + /* Make this position-independent code. */ + .option pic2 + + .globl qt_block + .globl qt_blocki + .globl qt_abort + .globl qt_start + .globl qt_vstart + + /* + ** $4: ptr to function to call once curr is suspended + ** and control is on $7's stack. + ** $5: 1'th arg to $4. + ** $6: 2'th arg to $4 + ** $7: sp of thread to suspend. + ** + ** Totally gross hack: The MIPS calling convention reserves + ** 4 words on the stack for a0..a3. This routine "ought" to + ** allocate space for callee-save registers plus 4 words for + ** the helper function, but instead we use the 4 words + ** provided by the function that called us (we don't need to + ** save our argument registers). So what *appears* to be + ** allocating only 40 bytes is actually allocating 56, by + ** using the caller's 16 bytes. + ** + ** The helper routine returns a value that is passed on as the + ** return value from the blocking routine. Since we don't + ** touch $2 between the helper's return and the end of + ** function, we get this behavior for free. + */ +qt_blocki: + sub $sp,$sp,40 /* Allocate reg save space. */ + sw $16, 0+16($sp) + sw $17, 4+16($sp) + sw $18, 8+16($sp) + sw $19,12+16($sp) + sw $20,16+16($sp) + sw $21,20+16($sp) + sw $22,24+16($sp) + sw $23,28+16($sp) + sw $30,32+16($sp) + sw $31,36+16($sp) + add $2, $sp,$0 /* $2 <= old sp to pass to func@$4. */ +qt_abort: + add $sp, $7,$0 /* $sp <= new sp. */ + .set noreorder + add $25, $4,$0 /* Set helper function procedure value. */ + jal $31,$25 /* Call helper func@$4 . */ + add $4, $2,$0 /* $a0 <= pass old sp as a parameter. */ + .set reorder + lw $31,36+16($sp) /* Restore callee-save regs... */ + lw $30,32+16($sp) + lw $23,28+16($sp) + lw $22,24+16($sp) + lw $21,20+16($sp) + lw $20,16+16($sp) + lw $19,12+16($sp) + lw $18, 8+16($sp) + lw $17, 4+16($sp) + lw $16, 0+16($sp) /* Restore callee-save */ + + add $sp,$sp,40 /* Deallocate reg save space. */ + j $31 /* Return to caller. */ + + /* + ** Non-varargs thread startup. + ** Note: originally, 56 bytes were allocated on the stack. + ** The thread restore routine (_blocki/_abort) removed 40 + ** of them, which means there is still 16 bytes for the + ** argument area required by the MIPS calling convention. + */ +qt_start: + add $4, $16,$0 /* Load up user function pu. */ + add $5, $17,$0 /* ... user function pt. */ + add $6, $18,$0 /* ... user function userf. */ + add $25, $19,$0 /* Set `only' procedure value. */ + jal $31,$25 /* Call `only'. */ + la $25,qt_error /* Set `qt_error' procedure value. */ + j $25 + + + /* + ** Save calle-save floating-point regs $f20-$f30 + ** See comment in `qt_block' about calling conventinos and + ** reserved space. Use the same trick here, but here we + ** actually have to allocate all the bytes since we have to + ** leave 4 words leftover for `qt_blocki'. + ** + ** Return value from `qt_block' is the same as the return from + ** `qt_blocki'. We get that for free since we don't touch $2 + ** between the return from `qt_blocki' and the return from + ** `qt_block'. + */ +qt_block: + sub $sp, $sp,56 /* 6 8-byte regs, saved ret pc, aligned. */ + swc1 $f20, 0+16($sp) + swc1 $f22, 8+16($sp) + swc1 $f24, 16+16($sp) + swc1 $f26, 24+16($sp) + swc1 $f28, 32+16($sp) + swc1 $f30, 40+16($sp) + sw $31, 48+16($sp) + jal qt_blocki + lwc1 $f20, 0+16($sp) + lwc1 $f22, 8+16($sp) + lwc1 $f24, 16+16($sp) + lwc1 $f26, 24+16($sp) + lwc1 $f28, 32+16($sp) + lwc1 $f30, 40+16($sp) + lw $31, 48+16($sp) + add $sp, $sp,56 + j $31 + + + /* + ** First, call `startup' with the `pt' argument. + ** + ** Next, call the user's function with all arguments. + ** Note that we don't know whether args were passed in + ** integer regs, fp regs, or on the stack (See Gerry Kane + ** "MIPS R2000 RISC Architecture" pg D-22), so we reload + ** all the registers, possibly with garbage arguments. + ** + ** Finally, call `cleanup' with the `pt' argument and with + ** the return value from the user's function. It is an error + ** for `cleanup' to return. + */ +qt_vstart: + add $4, $17,$0 /* `pt' is arg0 to `startup'. */ + add $25, $18,$0 /* Set `startup' procedure value. */ + jal $31, $25 /* Call `startup'. */ + + add $sp, $sp,16 /* Free extra save space. */ + lw $4, 0($sp) /* Load up args. */ + lw $5, 4($sp) + lw $6, 8($sp) + lw $7, 12($sp) + lwc1 $f12, 0($sp) /* Load up fp args. */ + lwc1 $f14, 8($sp) + add $25, $19,$0 /* Set `userf' procedure value. */ + jal $31,$25 /* Call `userf'. */ + + add $4, $17,$0 /* `pt' is arg0 to `cleanup'. */ + add $5, $2,$0 /* Ret. val is arg1 to `cleanup'. */ + add $25, $16,$0 /* Set `cleanup' procedure value. */ + jal $31, $25 /* Call `cleanup'. */ + + la $25,qt_error /* Set `qt_error' procedure value. */ + j $25 diff --git a/qt/md/mips.h b/qt/md/mips.h new file mode 100644 index 000000000..c584a681e --- /dev/null +++ b/qt/md/mips.h @@ -0,0 +1,134 @@ +/* + * QuickThreads -- Threads-building toolkit. + * Copyright (c) 1993 by David Keppel + * + * Permission to use, copy, modify and distribute this software and + * its documentation for any purpose and without fee is hereby + * granted, provided that the above copyright notice and this notice + * appear in all copies. This software is provided as a + * proof-of-concept and for demonstration purposes; there is no + * representation about the suitability of this software for any + * purpose. + */ + +#ifndef QT_MIPS_H +#define QT_MIPS_H + +typedef unsigned long qt_word_t; + +#define QT_GROW_DOWN + +/* Stack layout on the mips: + + Callee-save registers are: $16-$23, $30; $f20-$f30. + Also save $31, return pc. + + Non-varargs: + + +--- + | $f30 The first clump is only saved if `qt_block' + | $f28 is called, in which case it saves the fp regs + | $f26 then calls `qt_blocki' to save the int regs. + | $f24 + | $f22 + | $f20 + | $31 === return pc in `qt_block' + +--- + | $31 === return pc; on startup == qt_start + | $30 + | $23 + | $22 + | $21 + | $20 + | $19 on startup === only + | $18 on startup === $a2 === userf + | $17 on startup === $a1 === pt + | $16 on startup === $a0 === pu + | <a3> save area req'd by MIPS calling convention + | <a2> save area req'd by MIPS calling convention + | <a1> save area req'd by MIPS calling convention + | <a0> save area req'd by MIPS calling convention <--- sp + +--- + + Conventions for varargs: + + | args ... + +--- + | : + | : + | $21 + | $20 + | $19 on startup === `userf' + | $18 on startup === `startup' + | $17 on startup === `pt' + | $16 on startup === `cleanup' + | <a3> + | <a2> + | <a1> + | <a0> <--- sp + +--- + + Note: if we wanted to, we could muck about and try to get the 4 + argument registers loaded in to, e.g., $22, $23, $30, and $31, + and the return pc in, say, $20. Then, the first 4 args would + not need to be loaded from memory, they could just use + register-to-register copies. */ + + +/* Stack must be doubleword aligned. */ +#define QT_STKALIGN (8) /* Doubleword aligned. */ + +/* How much space is allocated to hold all the crud for + initialization: $16-$23, $30, $31. Just do an integer restore, + no need to restore floating-point. Four words are needed for the + argument save area for the helper function that will be called for + the old thread, just before the new thread starts to run. */ + +#define QT_STKBASE (14 * 4) +#define QT_VSTKBASE QT_STKBASE + + +/* Offsets of various registers. */ +#define QT_31 (9+4) +#define QT_19 (3+4) +#define QT_18 (2+4) +#define QT_17 (1+4) +#define QT_16 (0+4) + + +/* When a never-before-run thread is restored, the return pc points + to a fragment of code that starts the thread running. For + non-vargs functions, it just calls the client's `only' function. + For varargs functions, it calls the startup, user, and cleanup + functions. + + The varargs startup routine always reads 4 words of arguments from + the stack. If there are less than 4 words of arguments, then the + startup routine can read off the top of the stack. To prevent + errors we always allocate 4 words. If there are more than 3 words + of arguments, the 4 preallocated words are simply wasted. */ + +extern void qt_start(void); +#define QT_ARGS_MD(sp) (QT_SPUT (sp, QT_31, qt_start)) + +#define QT_VARGS_MD0(sp, vabytes) \ + ((qt_t *)(((char *)(sp)) - 4*4 - QT_STKROUNDUP(vabytes))) + +extern void qt_vstart(void); +#define QT_VARGS_MD1(sp) (QT_SPUT (sp, QT_31, qt_vstart)) + +#define QT_VARGS_DEFAULT + + +/* The *index* (positive offset) of where to put each value. */ +#define QT_ONLY_INDEX (QT_19) +#define QT_USER_INDEX (QT_18) +#define QT_ARGT_INDEX (QT_17) +#define QT_ARGU_INDEX (QT_16) + +#define QT_VCLEANUP_INDEX (QT_16) +#define QT_VUSERF_INDEX (QT_19) +#define QT_VSTARTUP_INDEX (QT_18) +#define QT_VARGT_INDEX (QT_17) + +#endif /* ndef QT_MIPS_H */ diff --git a/qt/md/mips.s b/qt/md/mips.s new file mode 100644 index 000000000..b074b98dc --- /dev/null +++ b/qt/md/mips.s @@ -0,0 +1,164 @@ +/* mips.s -- assembly support. */ + +/* + * QuickThreads -- Threads-building toolkit. + * Copyright (c) 1993 by David Keppel + * + * Permission to use, copy, modify and distribute this software and + * its documentation for any purpose and without fee is hereby + * granted, provided that the above copyright notice and this notice + * appear in all copies. This software is provided as a + * proof-of-concept and for demonstration purposes; there is no + * representation about the suitability of this software for any + * purpose. + */ + +/* Callee-save $16-$23, $30-$31. + * + * On startup, restore regs so retpc === call to a function to start. + * We're going to call a function ($4) from within this routine. + * We're passing 3 args, therefore need to allocate 12 extra bytes on + * the stack for a save area. The start routine needs a like 16-byte + * save area. Must be doubleword aligned (_mips r3000 risc + * architecture_, gerry kane, pg d-23). + */ + + .globl qt_block + .globl qt_blocki + .globl qt_abort + .globl qt_start + .globl qt_vstart + + /* + ** $4: ptr to function to call once curr is suspended + ** and control is on $7's stack. + ** $5: 1'th arg to $4. + ** $6: 2'th arg to $4 + ** $7: sp of thread to suspend. + ** + ** Totally gross hack: The MIPS calling convention reserves + ** 4 words on the stack for a0..a3. This routine "ought" to + ** allocate space for callee-save registers plus 4 words for + ** the helper function, but instead we use the 4 words + ** provided by the function that called us (we don't need to + ** save our argument registers). So what *appears* to be + ** allocating only 40 bytes is actually allocating 56, by + ** using the caller's 16 bytes. + ** + ** The helper routine returns a value that is passed on as the + ** return value from the blocking routine. Since we don't + ** touch $2 between the helper's return and the end of + ** function, we get this behavior for free. + */ +qt_blocki: + sub $sp,$sp,40 /* Allocate reg save space. */ + sw $16, 0+16($sp) + sw $17, 4+16($sp) + sw $18, 8+16($sp) + sw $19,12+16($sp) + sw $20,16+16($sp) + sw $21,20+16($sp) + sw $22,24+16($sp) + sw $23,28+16($sp) + sw $30,32+16($sp) + sw $31,36+16($sp) + add $2, $sp,$0 /* $2 <= old sp to pass to func@$4. */ +qt_abort: + add $sp, $7,$0 /* $sp <= new sp. */ + .set noreorder + jal $31,$4 /* Call helper func@$4 . */ + add $4, $2,$0 /* $a0 <= pass old sp as a parameter. */ + .set reorder + lw $31,36+16($sp) /* Restore callee-save regs... */ + lw $30,32+16($sp) + lw $23,28+16($sp) + lw $22,24+16($sp) + lw $21,20+16($sp) + lw $20,16+16($sp) + lw $19,12+16($sp) + lw $18, 8+16($sp) + lw $17, 4+16($sp) + lw $16, 0+16($sp) /* Restore callee-save */ + + add $sp,$sp,40 /* Deallocate reg save space. */ + j $31 /* Return to caller. */ + + /* + ** Non-varargs thread startup. + ** Note: originally, 56 bytes were allocated on the stack. + ** The thread restore routine (_blocki/_abort) removed 40 + ** of them, which means there is still 16 bytes for the + ** argument area required by the MIPS calling convention. + */ +qt_start: + add $4, $16,$0 /* Load up user function pu. */ + add $5, $17,$0 /* ... user function pt. */ + add $6, $18,$0 /* ... user function userf. */ + jal $31,$19 /* Call `only'. */ + j qt_error + + + /* + ** Save calle-save floating-point regs $f20-$f30 + ** See comment in `qt_block' about calling conventinos and + ** reserved space. Use the same trick here, but here we + ** actually have to allocate all the bytes since we have to + ** leave 4 words leftover for `qt_blocki'. + ** + ** Return value from `qt_block' is the same as the return from + ** `qt_blocki'. We get that for free since we don't touch $2 + ** between the return from `qt_blocki' and the return from + ** `qt_block'. + */ +qt_block: + sub $sp, $sp,56 /* 6 8-byte regs, saved ret pc, aligned. */ + swc1 $f20, 0+16($sp) + swc1 $f22, 8+16($sp) + swc1 $f24, 16+16($sp) + swc1 $f26, 24+16($sp) + swc1 $f28, 32+16($sp) + swc1 $f30, 40+16($sp) + sw $31, 48+16($sp) + jal qt_blocki + lwc1 $f20, 0+16($sp) + lwc1 $f22, 8+16($sp) + lwc1 $f24, 16+16($sp) + lwc1 $f26, 24+16($sp) + lwc1 $f28, 32+16($sp) + lwc1 $f30, 40+16($sp) + lw $31, 48+16($sp) + add $sp, $sp,56 + j $31 + + + /* + ** First, call `startup' with the `pt' argument. + ** + ** Next, call the user's function with all arguments. + ** Note that we don't know whether args were passed in + ** integer regs, fp regs, or on the stack (See Gerry Kane + ** "MIPS R2000 RISC Architecture" pg D-22), so we reload + ** all the registers, possibly with garbage arguments. + ** + ** Finally, call `cleanup' with the `pt' argument and with + ** the return value from the user's function. It is an error + ** for `cleanup' to return. + */ +qt_vstart: + add $4, $17,$0 /* `pt' is arg0 to `startup'. */ + jal $31, $18 /* Call `startup'. */ + + add $sp, $sp,16 /* Free extra save space. */ + lw $4, 0($sp) /* Load up args. */ + lw $5, 4($sp) + lw $6, 8($sp) + lw $7, 12($sp) + lwc1 $f12, 0($sp) /* Load up fp args. */ + lwc1 $f14, 8($sp) + jal $31,$19 /* Call `userf'. */ + + add $4, $17,$0 /* `pt' is arg0 to `cleanup'. */ + add $5, $2,$0 /* Ret. val is arg1 to `cleanup'. */ + jal $31, $16 /* Call `cleanup'. */ + + j qt_error diff --git a/qt/md/mips_b.s b/qt/md/mips_b.s new file mode 100644 index 000000000..5b3740843 --- /dev/null +++ b/qt/md/mips_b.s @@ -0,0 +1,99 @@ +/* + * QuickThreads -- Threads-building toolkit. + * Copyright (c) 1993 by David Keppel + * + * Permission to use, copy, modify and distribute this software and + * its documentation for any purpose and without fee is hereby + * granted, provided that the above copyright notice and this notice + * appear in all copies. This software is provided as a + * proof-of-concept and for demonstration purposes; there is no + * representation about the suitability of this software for any + * purpose. + */ + + .globl b_call_reg + .globl b_call_imm + .globl b_add + .globl b_load + + .ent b_null +b_null: + j $31 + .end b_null + + .ent b_call_reg +b_call_reg: + la $5,b_null + add $6, $31,0 +$L0: + jal $5 + jal $5 + jal $5 + jal $5 + jal $5 + + sub $4, $4,5 + bgtz $4,$L0 + j $6 + .end + + + .ent b_call_imm +b_call_imm: + add $6, $31,0 +$L1: + jal b_null + jal b_null + jal b_null + jal b_null + jal b_null + + sub $4, $4,5 + bgtz $4,$L1 + j $6 + .end + + + .ent b_add +b_add: + add $5, $0,$4 + add $6, $0,$4 + add $7, $0,$4 + add $8, $0,$4 +$L2: + sub $4, $4,5 + sub $5, $5,5 + sub $6, $6,5 + sub $7, $7,5 + sub $8, $8,5 + + sub $4, $4,5 + sub $5, $5,5 + sub $6, $6,5 + sub $7, $7,5 + sub $8, $8,5 + + bgtz $4,$L2 + j $31 + .end + + + .ent b_load +b_load: +$L3: + ld $0, 0($sp) + ld $0, 4($sp) + ld $0, 8($sp) + ld $0, 12($sp) + ld $0, 16($sp) + + ld $0, 20($sp) + ld $0, 24($sp) + ld $0, 28($sp) + ld $0, 32($sp) + ld $0, 36($sp) + + sub $4, $4,10 + bgtz $4,$L3 + j $31 + .end diff --git a/qt/md/null.README b/qt/md/null.README new file mode 100644 index 000000000..e69de29bb --- /dev/null +++ b/qt/md/null.README diff --git a/qt/md/null.c b/qt/md/null.c new file mode 100644 index 000000000..775db62be --- /dev/null +++ b/qt/md/null.c @@ -0,0 +1,14 @@ +/* + * QuickThreads -- Threads-building toolkit. + * Copyright (c) 1993 by David Keppel + * + * Permission to use, copy, modify and distribute this software and + * its documentation for any purpose and without fee is hereby + * granted, provided that the above copyright notice and this notice + * appear in all copies. This software is provided as a + * proof-of-concept and for demonstration purposes; there is no + * representation about the suitability of this software for any + * purpose. + */ + +char const qtmd_rcsid[] = "$Header: /home/ludo/src/guile.cvs/gitification/guile-cvs/guile/guile-core/qt/md/null.c,v 1.1 1996-10-01 03:34:16 mdj Exp $"; diff --git a/qt/md/solaris.README b/qt/md/solaris.README new file mode 100644 index 000000000..04f855c44 --- /dev/null +++ b/qt/md/solaris.README @@ -0,0 +1,19 @@ +Solaris 2.x is like System V (maybe it *is* System V?) and is different +from older versions in that it uses no leading underscore for variable +and function names. That is, the old convention was: + + foo(){} + +got compiled as + + .globl _foo + _foo: + +and is now compiled as + + .globl foo + foo: + +The `config' script should fix up the older (leading underscore) versions +of the machine-dependent files to use the newer (no leading underscore) +calling conventions. diff --git a/qt/md/sparc.h b/qt/md/sparc.h new file mode 100644 index 000000000..e2ab281b4 --- /dev/null +++ b/qt/md/sparc.h @@ -0,0 +1,140 @@ +/* + * QuickThreads -- Threads-building toolkit. + * Copyright (c) 1993 by David Keppel + * + * Permission to use, copy, modify and distribute this software and + * its documentation for any purpose and without fee is hereby + * granted, provided that the above copyright notice and this notice + * appear in all copies. This software is provided as a + * proof-of-concept and for demonstration purposes; there is no + * representation about the suitability of this software for any + * purpose. + */ + +#ifndef QT_SPARC_H +#define QT_SPARC_H + +typedef unsigned long qt_word_t; + +/* Stack layout on the sparc: + + non-varargs: + + +--- + | <blank space for alignment> + | %o7 == return address -> qt_start + | %i7 + | %i6 == frame pointer -> 0 (NULL-terminated stack frame chain) + | %i5 -> only + | %i4 -> userf + | %i3 + | %i2 -> pt + | %i1 -> pu + | %i0 + | %l7 + | %l6 + | %l5 + | %l4 + | %l3 + | %l2 + | %l1 + | %l0 <--- qt_t.sp + +--- + + varargs: + + | : + | : + | argument list + | one-word aggregate return pointer + +--- + | <blank space for alignment> + | %o7 == return address -> qt_vstart + | %i7 + | %i6 == frame pointer -> 0 (NULL-terminated stack frame chain) + | %i5 -> startup + | %i4 -> userf + | %i3 -> cleanup + | %i2 -> pt + | %i1 + | %i0 + | %l7 + | %l6 + | %l5 + | %l4 + | %l3 + | %l2 + | %l1 + | %l0 <--- qt_t.sp + +--- + + */ + + +/* What to do to start a thread running. */ +extern void qt_start (void); +extern void qt_vstart (void); + + +/* Hold 17 saved registers + 1 word for alignment. */ +#define QT_STKBASE (18 * 4) +#define QT_VSTKBASE QT_STKBASE + + +/* Stack must be doubleword aligned. */ +#define QT_STKALIGN (8) /* Doubleword aligned. */ + +#define QT_ONLY_INDEX (QT_I5) +#define QT_USER_INDEX (QT_I4) +#define QT_ARGT_INDEX (QT_I2) +#define QT_ARGU_INDEX (QT_I1) + +#define QT_VSTARTUP_INDEX (QT_I5) +#define QT_VUSERF_INDEX (QT_I4) +#define QT_VCLEANUP_INDEX (QT_I3) +#define QT_VARGT_INDEX (QT_I2) + +#define QT_O7 (16) +#define QT_I6 (14) +#define QT_I5 (13) +#define QT_I4 (12) +#define QT_I3 (11) +#define QT_I2 (10) +#define QT_I1 ( 9) + + +/* The thread will ``return'' to the `qt_start' routine to get things + going. The normal return sequence takes us to QT_O7+8, so we + pre-subtract 8. The frame pointer chain is 0-terminated to prevent + the trap handler from chasing off in to random memory when flushing + stack windows. */ + +#define QT_ARGS_MD(top) \ + (QT_SPUT ((top), QT_O7, ((void *)(((int)qt_start)-8))), \ + QT_SPUT ((top), QT_I6, 0)) + + +/* The varargs startup routine always reads 6 words of arguments + (6 argument registers) from the stack, offset by one word to + allow for an aggregate return area pointer. If the varargs + routine actually pushed fewer words than that, qt_vstart could read + off the top of the stack. To prevent errors, we always allocate 8 + words. The space is often just wasted. */ + +#define QT_VARGS_MD0(sp, vabytes) \ + ((qt_t *)(((char *)(sp)) - 8*4 - QT_STKROUNDUP(vabytes))) + +#define QT_VARGS_MD1(sp) \ + (QT_SPUT (sp, QT_O7, ((void *)(((int)qt_vstart)-8)))) + +/* The SPARC has wierdo calling conventions which stores a hidden + parameter for returning aggregate values, so the rest of the + parameters are shoved up the stack by one place. */ +#define QT_VARGS_ADJUST(sp) (((char *)sp)+4) + +#define QT_VARGS_DEFAULT + + +#define QT_GROW_DOWN + +#endif /* ndef QT_SPARC_H */ diff --git a/qt/md/sparc.s b/qt/md/sparc.s new file mode 100644 index 000000000..d9bdf0c58 --- /dev/null +++ b/qt/md/sparc.s @@ -0,0 +1,142 @@ +/* sparc.s -- assembly support for the `qt' thread building kit. */ + +/* + * QuickThreads -- Threads-building toolkit. + * Copyright (c) 1993 by David Keppel + * + * Permission to use, copy, modify and distribute this software and + * its documentation for any purpose and without fee is hereby + * granted, provided that the above copyright notice and this notice + * appear in all copies. This software is provided as a + * proof-of-concept and for demonstration purposes; there is no + * representation about the suitability of this software for any + * purpose. + */ + +/* #include <machine/trap.h> */ + + .text + .align 4 + .global qt_blocki + .global qt_block + .global qt_abort + .global qt_start + .global qt_vstart + +/* Register assignment: +// %o0: incoming `helper' function to call after cswap +// also used as outgoing sp of old thread (qt_t *) +// %o1, %o2: +// parameters to `helper' function called after cswap +// %o3: sp of new thread +// %o5: tmp used to save old thread sp, while using %o0 +// to call `helper' f() after cswap. +// +// +// Aborting a thread is easy if there are no cached register window +// frames: just switch to the new stack and away we go. If there are +// cached register window frames they must all be written back to the +// old stack before we move to the new stack. If we fail to do the +// writeback then the old stack memory can be written with register +// window contents e.g., after the stack memory has been freed and +// reused. +// +// If you don't believe this, try setting the frame pointer to zero +// once we're on the new stack. This will not affect correctnes +// otherwise because the frame pointer will eventually get reloaded w/ +// the new thread's frame pointer. But it will be zero briefly before +// the reload. You will eventually (100,000 cswaps later on a small +// SPARC machine that I tried) get an illegal instruction trap from +// the kernel trying to flush a cached window to location 0x0. +// +// Solution: flush windows before switching stacks, which invalidates +// all the other register windows. We could do the trap +// conditionally: if we're in the lowest frame of a thread, the fp is +// zero already so we know there's nothing cached. But we expect most +// aborts will be done from a first function that does a `save', so we +// will rarely save anything and always pay the cost of testing to see +// if we should flush. +// +// All floating-point registers are caller-save, so this routine +// doesn't need to do anything to save and restore them. +// +// `qt_block' and `qt_blocki' return the same value as the value +// returned by the helper function. We get this ``for free'' +// since we don't touch the return value register between the +// return from the helper function and return from qt_block{,i}. +*/ + +qt_block: +qt_blocki: + sub %sp, 8, %sp /* Allocate save area for return pc. */ + st %o7, [%sp+64] /* Save return pc. */ +qt_abort: + ta 0x03 /* Save locals and ins. */ + mov %sp, %o5 /* Remember old sp w/o chng ins/locals. */ + sub %o3, 64, %sp /* Allocate kwsa, switch stacks. */ + call %o0, 0 /* Call `helper' routine. */ + mov %o5, %o0 /* Pass old thread to qt_after_t() */ + /* .. along w/ args in %o1 & %o2. */ + + /* Restore callee-save regs. The kwsa + // is on this stack, so offset all + // loads by sizeof(kwsa), 64 bytes. + */ + ldd [%sp+ 0+64], %l0 + ldd [%sp+ 8+64], %l2 + ldd [%sp+16+64], %l4 + ldd [%sp+24+64], %l6 + ldd [%sp+32+64], %i0 + ldd [%sp+40+64], %i2 + ldd [%sp+48+64], %i4 + ldd [%sp+56+64], %i6 + ld [%sp+64+64], %o7 /* Restore return pc. */ + + retl /* Return to address in %o7. */ + add %sp, 72, %sp /* Deallocate kwsa, ret pc area. */ + + +/* The function calling conventions say there has to be a 1-word area +// in the caller's stack to hold a pointer to space for aggregate +// return values. It also says there should be a 6-word area to hold +// %o0..%o5 if the callee wants to save them (why? I don't know...) +// Round up to 8 words to maintain alignment. +// +// Parameter values were stored in callee-save regs and are moved to +// the parameter registers. +*/ +qt_start: + mov %i1, %o0 /* `pu': Set up args to `only'. */ + mov %i2, %o1 /* `pt'. */ + mov %i4, %o2 /* `userf'. */ + call %i5, 0 /* Call client function. */ + sub %sp, 32, %sp /* Allocate 6-word callee space. */ + + call qt_error, 0 /* `only' erroniously returned. */ + nop + + +/* Same comments as `qt_start' about allocating rounded-up 7-word +// save areas. */ + +qt_vstart: + sub %sp, 32, %sp /* Allocate 7-word callee space. */ + call %i5, 0 /* call `startup'. */ + mov %i2, %o0 /* .. with argument `pt'. */ + + add %sp, 32, %sp /* Use 7-word space in varargs. */ + ld [%sp+ 4+64], %o0 /* Load arg0 ... */ + ld [%sp+ 8+64], %o1 + ld [%sp+12+64], %o2 + ld [%sp+16+64], %o3 + ld [%sp+20+64], %o4 + call %i4, 0 /* Call `userf'. */ + ld [%sp+24+64], %o5 + + /* Use 6-word space in varargs. */ + mov %o0, %o1 /* Pass return value from userf */ + call %i3, 0 /* .. when call `cleanup. */ + mov %i2, %o0 /* .. along with argument `pt'. */ + + call qt_error, 0 /* `cleanup' erroniously returned. */ + nop diff --git a/qt/md/sparc_b.s b/qt/md/sparc_b.s new file mode 100644 index 000000000..08351d76d --- /dev/null +++ b/qt/md/sparc_b.s @@ -0,0 +1,106 @@ +/* + * QuickThreads -- Threads-building toolkit. + * Copyright (c) 1993 by David Keppel + * + * Permission to use, copy, modify and distribute this software and + * its documentation for any purpose and without fee is hereby + * granted, provided that the above copyright notice and this notice + * appear in all copies. This software is provided as a + * proof-of-concept and for demonstration purposes; there is no + * representation about the suitability of this software for any + * purpose. + */ + + .globl b_call_reg + .globl b_call_imm + .globl b_add + .globl b_load + +b_null: + retl + nop + +b_call_reg: + sethi %hi(b_null),%o4 + or %o4,%lo(b_null),%o4 + add %o7,%g0, %o3 +L0: + call %o4 + nop + call %o4 + nop + call %o4 + nop + call %o4 + nop + call %o4 + nop + + subcc %o0,1,%o0 + bg L0 + nop + add %o3,%g0, %o7 + retl + nop + +b_call_imm: + sethi %hi(b_null),%o4 + or %o4,%lo(b_null),%o4 + add %o7,%g0, %o3 +L1: + call b_null + call b_null + call b_null + call b_null + call b_null + + subcc %o0,1,%o0 + bg L0 + nop + add %o3,%g0, %o7 + retl + nop + + +b_add: + add %o0,%g0,%o1 + add %o0,%g0,%o2 + add %o0,%g0,%o3 + add %o0,%g0,%o4 +L2: + sub %o0,5,%o0 + sub %o1,5,%o1 + sub %o2,5,%o2 + sub %o3,5,%o3 + sub %o4,5,%o4 + + subcc %o0,5,%o0 + sub %o1,5,%o1 + sub %o2,5,%o2 + sub %o3,5,%o3 + sub %o4,5,%o4 + + bg L2 + nop + retl + nop + + +b_load: + ld [%sp+ 0], %g0 +L3: + ld [%sp+ 4],%g0 + ld [%sp+ 8],%g0 + ld [%sp+12],%g0 + ld [%sp+16],%g0 + ld [%sp+20],%g0 + ld [%sp+24],%g0 + ld [%sp+28],%g0 + ld [%sp+32],%g0 + ld [%sp+36],%g0 + + subcc %o0,10,%o0 + bg L3 + ld [%sp+ 0],%g0 + retl + nop diff --git a/qt/md/vax.h b/qt/md/vax.h new file mode 100644 index 000000000..1a5af0f2b --- /dev/null +++ b/qt/md/vax.h @@ -0,0 +1,130 @@ +/* + * QuickThreads -- Threads-building toolkit. + * Copyright (c) 1993 by David Keppel + * + * Permission to use, copy, modify and distribute this software and + * its documentation for any purpose and without fee is hereby + * granted, provided that the above copyright notice and this notice + * appear in all copies. This software is provided as a + * proof-of-concept and for demonstration purposes; there is no + * representation about the suitability of this software for any + * purpose. + */ + +#ifndef QT_VAX_H +#define QT_VAX_H + +typedef unsigned long qt_word_t; + +/* Thread's initial stack layout on the VAX: + + non-varargs: + + +--- + | arg[2] === `userf' on startup + | arg[1] === `pt' on startup + | arg[0] === `pu' on startup + | ... === `only' on startup. + +--- + | ret pc === `qt_start' on startup + | fp === 0 on startup + | ap === 0 on startup + | <mask> + | 0 (handler) <--- qt_t.sp + +--- + + When a non-varargs thread is started, it ``returns'' to the start + routine, which calls the client's `only' function. + + The varargs case is clearly bad code. The various values should be + stored in a save area and snarfed in to callee-save registers on + startup. However, it's too painful to figure out the register + mask (right now), so do it the slow way. + + +--- + | arg[n-1] + | .. + | arg[0] + | nargs + +--- + | === `cleanup' + | === `vuserf' + | === `startup' + | === `pt' + +--- + | ret pc === `qt_start' on startup + | fp === 0 on startup + | ap === 0 on startup + | <mask> + | 0 (handler) <--- qt_t.sp + +--- + + When a varargs thread is started, it ``returns'' to the `qt_vstart' + startup code. The startup code pops all the extra arguments, then + calls the appropriate functions. */ + + +/* What to do to start a thread running. */ +extern void qt_start (void); +extern void qt_vstart (void); + + +/* Initial call frame for non-varargs and varargs cases. */ +#define QT_STKBASE (10 * 4) +#define QT_VSTKBASE (9 * 4) + + +/* Stack "must be" 4-byte aligned. (Actually, no, but it's + easiest and probably fastest to do so.) */ + +#define QT_STKALIGN (4) + + +/* Where to place various arguments. */ +#define QT_ONLY_INDEX (5) +#define QT_USER_INDEX (8) +#define QT_ARGT_INDEX (7) +#define QT_ARGU_INDEX (6) + +#define QT_VSTARTUP_INDEX (6) +#define QT_VUSERF_INDEX (7) +#define QT_VCLEANUP_INDEX (8) +#define QT_VARGT_INDEX (5) + + +/* Stack grows down. The top of the stack is the first thing to + pop off (predecrement, postincrement). */ +#define QT_GROW_DOWN + + +extern void qt_error (void); + +#define QT_VAX_GMASK_NOREGS (0) + +/* Push on the error return address, null termination to call chains, + number of arguments to `only', register save mask (save no + registers). */ + +#define QT_ARGS_MD(sto) \ + (QT_SPUT (sto, 0, 0), \ + QT_SPUT (sto, 1, QT_VAX_GMASK_NOREGS), \ + QT_SPUT (sto, 2, 0), \ + QT_SPUT (sto, 3, 0), \ + QT_SPUT (sto, 4, qt_start)) + +#define QT_VARGS_MD0(sto, nbytes) \ + (QT_SPUT (sto, (-(nbytes)/4)-1, (nbytes)/4), \ + ((char *)(((sto)-4) - QT_STKROUNDUP(nbytes)))) + +#define QT_VARGS_ADJUST(sp) ((char *)sp + 4) + +#define QT_VARGS_MD1(sto) \ + (QT_SPUT (sto, 0, 0), \ + QT_SPUT (sto, 1, QT_VAX_GMASK_NOREGS), \ + QT_SPUT (sto, 2, 0), \ + QT_SPUT (sto, 3, 0), \ + QT_SPUT (sto, 4, qt_vstart)) + +#define QT_VARGS_DEFAULT + +#endif /* QT_VAX_H */ diff --git a/qt/md/vax.s b/qt/md/vax.s new file mode 100644 index 000000000..fed03f043 --- /dev/null +++ b/qt/md/vax.s @@ -0,0 +1,69 @@ +/* + * QuickThreads -- Threads-building toolkit. + * Copyright (c) 1993 by David Keppel + * + * Permission to use, copy, modify and distribute this software and + * its documentation for any purpose and without fee is hereby + * granted, provided that the above copyright notice and this notice + * appear in all copies. This software is provided as a + * proof-of-concept and for demonstration purposes; there is no + * representation about the suitability of this software for any + * purpose. + */ + + .text + + .globl _qt_abort + .globl _qt_block + .globl _qt_blocki + .globl _qt_start + .globl _qt_vstart + + +/* +// Calls to these routines have the signature +// +// void *block (func, arg1, arg2, newsp) +// +// Since the prologue saves 5 registers, nargs, pc, fp, ap, mask, and +// a condition handler (at sp+0), the first argument is 40=4*10 bytes +// offset from the stack pointer. +*/ +_qt_block: +_qt_blocki: +_qt_abort: + .word 0x7c0 /* Callee-save mask: 5 registers. */ + movl 56(sp),r1 /* Get stack pointer of new thread. */ + movl 52(sp),-(r1) /* Push arg2 */ + movl 48(sp),-(r1) /* Push arg1 */ + movl sp,-(r1) /* Push arg0 */ + + movl 44(sp),r0 /* Get helper to call. */ + movl r1,sp /* Move to new thread's stack. */ + addl3 sp,$12,fp /* .. including the frame pointer. */ + calls $3,(r0) /* Call helper. */ + + ret + +_qt_start: + movl (sp)+,r0 /* Get `only'. */ + calls $3,(r0) /* Call `only'. */ + calls $0,_qt_error /* `only' erroniously returned. */ + + +_qt_vstart: + movl (sp)+,r10 /* Get `pt'. */ + movl (sp)+,r9 /* Get `startup'. */ + movl (sp)+,r8 /* Get `vuserf'. */ + movl (sp)+,r7 /* Get `cleanup'. */ + + pushl r10 /* Push `qt'. */ + calls $1,(r9) /* Call `startup', pop `qt' on return. */ + + calls (sp)+,(r8) /* Call user's function. */ + + pushl r0 /* Push `vuserf_retval'. */ + pushl r10 /* Push `qt'. */ + calls $2,(r7) /* Call `cleanup', never return. */ + + calls $0,_qt_error /* `cleanup' erroniously returned. */ diff --git a/qt/md/vax_b.s b/qt/md/vax_b.s new file mode 100644 index 000000000..2db2d4fec --- /dev/null +++ b/qt/md/vax_b.s @@ -0,0 +1,92 @@ +/* + * QuickThreads -- Threads-building toolkit. + * Copyright (c) 1993 by David Keppel + * + * Permission to use, copy, modify and distribute this software and + * its documentation for any purpose and without fee is hereby + * granted, provided that the above copyright notice and this notice + * appear in all copies. This software is provided as a + * proof-of-concept and for demonstration purposes; there is no + * representation about the suitability of this software for any + * purpose. + */ + + .text + .globl _b_call_reg + .globl _b_call_imm + .globl _b_add + .globl _b_load + +_b_null: + .word 0x0 + ret + +_b_call_reg: + .word 0x0 + movl 4(ap),r0 + moval _b_null,r1 +L0: + calls $0,(r1) + calls $0,(r1) + calls $0,(r1) + calls $0,(r1) + calls $0,(r1) + + subl2 $5,r0 + bgtr L0 + ret + + +_b_call_imm: + .word 0x0 + movl 4(ap),r0 +L1: + calls $0,_b_null + calls $0,_b_null + calls $0,_b_null + calls $0,_b_null + calls $0,_b_null + + subl2 $5,r0 + bgtr L1 + ret + + +_b_add: + .word 0x0 + movl 4(ap),r0 +L2: + subl2 $1,r0 + subl2 $1,r0 + subl2 $1,r0 + subl2 $1,r0 + subl2 $1,r0 + + subl2 $1,r0 + subl2 $1,r0 + subl2 $1,r0 + subl2 $1,r0 + subl2 $1,r0 + + bgtr L2 + ret + + +_b_load: + .word 0x0 + movl 4(ap),r0 +L3: + movl 0(sp),r1 + movl 4(sp),r1 + movl 8(sp),r1 + movl 12(sp),r1 + movl 16(sp),r1 + movl 20(sp),r1 + movl 24(sp),r1 + movl 28(sp),r1 + movl 32(sp),r1 + movl 36(sp),r1 + + subl2 $1,r0 + bgtr L3 + ret |