--- libcoro/coro.c 2008/03/02 16:10:22 1.29 +++ libcoro/coro.c 2008/11/07 20:12:26 1.38 @@ -40,15 +40,17 @@ #include "coro.h" +#include + #if !defined(STACK_ADJUST_PTR) /* IRIX is decidedly NON-unix */ # if __sgi # define STACK_ADJUST_PTR(sp,ss) ((char *)(sp) + (ss) - 8) # define STACK_ADJUST_SIZE(sp,ss) ((ss) - 8) -# elif __i386__ && CORO_LINUX +# elif (__i386__ && CORO_LINUX) || (_M_IX86 && CORO_LOSER) # define STACK_ADJUST_PTR(sp,ss) ((char *)(sp) + (ss)) # define STACK_ADJUST_SIZE(sp,ss) (ss) -# elif __amd64__ && CORO_LINUX +# elif (__amd64__ && CORO_LINUX) || ((_M_AMD64 || _M_IA64) && CORO_LOSER) # define STACK_ADJUST_PTR(sp,ss) ((char *)(sp) + (ss) - 8) # define STACK_ADJUST_SIZE(sp,ss) (ss) # else @@ -63,13 +65,13 @@ #if CORO_SJLJ || CORO_LOSER || CORO_LINUX || CORO_IRIX || CORO_ASM -#include +# include -#if CORO_SJLJ -# include -# include -# include -#endif +# if CORO_SJLJ +# include +# include +# include +# endif static volatile coro_func coro_init_func; static volatile void *coro_init_arg; @@ -77,9 +79,9 @@ /* what we really want to detect here is wether we use a new-enough version of GAS */ /* instead, check for gcc 3, ELF and GNU/Linux and hope for the best */ -#if __GNUC__ >= 3 && __ELF__ && __linux__ -# define HAVE_CFI 1 -#endif +# if __GNUC__ >= 3 && __ELF__ && __linux__ +# define HAVE_CFI 1 +# endif static void coro_init (void) @@ -105,13 +107,13 @@ { if (setjmp (((coro_context *)new_coro)->env)) { -#if HAVE_CFI +# if HAVE_CFI asm (".cfi_startproc"); -#endif +# endif coro_init (); /* start it */ -#if HAVE_CFI +# if HAVE_CFI asm (".cfi_endproc"); -#endif +# endif } else trampoline_count++; @@ -122,60 +124,127 @@ #endif #if CORO_ASM -void __attribute__((__noinline__, __fastcall__)) -coro_transfer (struct coro_context *prev, struct coro_context *next) -{ - asm volatile ( -#if __amd64 -# define NUM_CLOBBERED 5 - "push %%rbx\n\t" - "push %%r12\n\t" - "push %%r13\n\t" - "push %%r14\n\t" - "push %%r15\n\t" - "mov %%rsp, %0\n\t" - "mov %1, %%rsp\n\t" - "pop %%r15\n\t" - "pop %%r14\n\t" - "pop %%r13\n\t" - "pop %%r12\n\t" - "pop %%rbx\n\t" -#elif __i386 -# define NUM_CLOBBERED 4 - "push %%ebx\n\t" - "push %%esi\n\t" - "push %%edi\n\t" - "push %%ebp\n\t" - "mov %%esp, %0\n\t" - "mov %1, %%esp\n\t" - "pop %%ebp\n\t" - "pop %%edi\n\t" - "pop %%esi\n\t" - "pop %%ebx\n\t" -#else -# error unsupported architecture -#endif - : "=m" (prev->sp) - : "m" (next->sp) + + asm ( + ".text\n" + ".globl coro_transfer\n" + ".type coro_transfer, @function\n" + "coro_transfer:\n" +# if __amd64 +# define NUM_SAVED 6 + "\tpush %rbp\n" + "\tpush %rbx\n" + "\tpush %r12\n" + "\tpush %r13\n" + "\tpush %r14\n" + "\tpush %r15\n" + "\tmov %rsp, (%rdi)\n" + "\tmov (%rsi), %rsp\n" + "\tpop %r15\n" + "\tpop %r14\n" + "\tpop %r13\n" + "\tpop %r12\n" + "\tpop %rbx\n" + "\tpop %rbp\n" +# elif __i386 +# define NUM_SAVED 4 + "\tpush %ebp\n" + "\tpush %ebx\n" + "\tpush %esi\n" + "\tpush %edi\n" + "\tmov %esp, (%eax)\n" + "\tmov (%edx), %esp\n" + "\tpop %edi\n" + "\tpop %esi\n" + "\tpop %ebx\n" + "\tpop %ebp\n" +# else +# error unsupported architecture +# endif + "\tret\n" ); + +#endif + +#if CORO_PTHREAD + +/* this mutex will be locked by the running coroutine */ +pthread_mutex_t coro_mutex = PTHREAD_MUTEX_INITIALIZER; + +struct coro_init_args +{ + coro_func func; + void *arg; + coro_context *self, *main; +}; + +static pthread_t null_tid; + +/* I'd so love to cast pthread_mutex_unlock to void (*)(void *)... */ +static void +mutex_unlock_wrapper (void *arg) +{ + pthread_mutex_unlock ((pthread_mutex_t *)arg); } + +static void * +trampoline (void *args_) +{ + struct coro_init_args *args = (struct coro_init_args *)args_; + coro_func func = args->func; + void *arg = args->arg; + + pthread_mutex_lock (&coro_mutex); + + /* we try to be good citizens and use deferred cancellation and cleanup handlers */ + pthread_cleanup_push (mutex_unlock_wrapper, &coro_mutex); + coro_transfer (args->self, args->main); + func (arg); + pthread_cleanup_pop (1); + + return 0; +} + +void +coro_transfer (coro_context *prev, coro_context *next) +{ + pthread_cond_signal (&next->cv); + pthread_cond_wait (&prev->cv, &coro_mutex); +} + +void +coro_destroy (coro_context *ctx) +{ + if (!pthread_equal (ctx->id, null_tid)) + { + pthread_cancel (ctx->id); + pthread_mutex_unlock (&coro_mutex); + pthread_join (ctx->id, 0); + pthread_mutex_lock (&coro_mutex); + } + + pthread_cond_destroy (&ctx->cv); +} + #endif /* initialize a machine state */ -void coro_create (coro_context *ctx, - coro_func coro, void *arg, - void *sptr, long ssize) +void +coro_create (coro_context *ctx, coro_func coro, void *arg, void *sptr, long ssize) { #if CORO_UCONTEXT + if (!coro) + return; + getcontext (&(ctx->uc)); ctx->uc.uc_link = 0; ctx->uc.uc_stack.ss_sp = STACK_ADJUST_PTR (sptr,ssize); - ctx->uc.uc_stack.ss_size = (size_t) STACK_ADJUST_SIZE (sptr,ssize); + ctx->uc.uc_stack.ss_size = (size_t)STACK_ADJUST_SIZE (sptr,ssize); ctx->uc.uc_stack.ss_flags = 0; - makecontext (&(ctx->uc), (void (*)()) coro, 1, arg); + makecontext (&(ctx->uc), (void (*)())coro, 1, arg); #elif CORO_SJLJ || CORO_LOSER || CORO_LINUX || CORO_IRIX || CORO_ASM @@ -186,6 +255,9 @@ # endif coro_context nctx; + if (!coro) + return; + coro_init_func = coro; coro_init_arg = arg; @@ -247,19 +319,19 @@ setjmp (ctx->env); #if __CYGWIN__ - ctx->env[7] = (long)((char *)sptr + ssize); + ctx->env[7] = (long)((char *)sptr + ssize) - sizeof (long); ctx->env[8] = (long)coro_init; #elif defined(_M_IX86) ((_JUMP_BUFFER *)&ctx->env)->Eip = (long)coro_init; - ((_JUMP_BUFFER *)&ctx->env)->Esp = (long)STACK_ADJUST_PTR (sptr,ssize); + ((_JUMP_BUFFER *)&ctx->env)->Esp = (long)STACK_ADJUST_PTR (sptr, ssize) - sizeof (long); #elif defined(_M_AMD64) ((_JUMP_BUFFER *)&ctx->env)->Rip = (__int64)coro_init; - ((_JUMP_BUFFER *)&ctx->env)->Rsp = (__int64)STACK_ADJUST_PTR (sptr,ssize); + ((_JUMP_BUFFER *)&ctx->env)->Rsp = (__int64)STACK_ADJUST_PTR (sptr, ssize) - sizeof (long); #elif defined(_M_IA64) ((_JUMP_BUFFER *)&ctx->env)->StIIP = (__int64)coro_init; - ((_JUMP_BUFFER *)&ctx->env)->IntSp = (__int64)STACK_ADJUST_PTR (sptr,ssize); + ((_JUMP_BUFFER *)&ctx->env)->IntSp = (__int64)STACK_ADJUST_PTR (sptr, ssize) - sizeof (long); #else -#error "microsoft libc or architecture not supported" +# error "microsoft libc or architecture not supported" #endif # elif CORO_LINUX @@ -267,16 +339,16 @@ _setjmp (ctx->env); #if __GLIBC__ >= 2 && __GLIBC_MINOR__ >= 0 && defined (JB_PC) && defined (JB_SP) ctx->env[0].__jmpbuf[JB_PC] = (long)coro_init; - ctx->env[0].__jmpbuf[JB_SP] = (long)STACK_ADJUST_PTR (sptr, ssize); + ctx->env[0].__jmpbuf[JB_SP] = (long)STACK_ADJUST_PTR (sptr, ssize) - sizeof (long); #elif __GLIBC__ >= 2 && __GLIBC_MINOR__ >= 0 && defined (__mc68000__) ctx->env[0].__jmpbuf[0].__aregs[0] = (long int)coro_init; - ctx->env[0].__jmpbuf[0].__sp = (int *)((char *)sptr + ssize); + ctx->env[0].__jmpbuf[0].__sp = (int *)((char *)sptr + ssize) - sizeof (long); #elif defined (__GNU_LIBRARY__) && defined (__i386__) ctx->env[0].__jmpbuf[0].__pc = (char *)coro_init; - ctx->env[0].__jmpbuf[0].__sp = (void *)((char *)sptr + ssize); + ctx->env[0].__jmpbuf[0].__sp = (void *)((char *)sptr + ssize) - sizeof (long); #elif defined (__GNU_LIBRARY__) && defined (__amd64__) ctx->env[0].__jmpbuf[JB_PC] = (long)coro_init; - ctx->env[0].__jmpbuf[JB_RSP] = (long)STACK_ADJUST_PTR (sptr, ssize); + ctx->env[0].__jmpbuf[0].__sp = (void *)((char *)sptr + ssize) - sizeof (long); #else # error "linux libc or architecture not supported" #endif @@ -285,21 +357,57 @@ setjmp (ctx->env); ctx->env[JB_PC] = (__uint64_t)coro_init; - ctx->env[JB_SP] = (__uint64_t)STACK_ADJUST_PTR (sptr, ssize); + ctx->env[JB_SP] = (__uint64_t)STACK_ADJUST_PTR (sptr, ssize) - sizeof (long); # elif CORO_ASM ctx->sp = (volatile void **)(ssize + (char *)sptr); + *--ctx->sp = (void *)abort; /* needed for alignment only */ *--ctx->sp = (void *)coro_init; - *--ctx->sp = (void *)coro_init; // this is needed when the prologue saves ebp - ctx->sp -= NUM_CLOBBERED; + ctx->sp -= NUM_SAVED; # endif coro_transfer ((coro_context *)create_coro, (coro_context *)new_coro); +# elif CORO_PTHREAD + + static coro_context nctx; + static int once; + + if (!once) + { + once = 1; + + pthread_mutex_lock (&coro_mutex); + pthread_cond_init (&nctx.cv, 0); + null_tid = pthread_self (); + } + + pthread_cond_init (&ctx->cv, 0); + + if (coro) + { + pthread_attr_t attr; + struct coro_init_args args; + + args.func = coro; + args.arg = arg; + args.self = ctx; + args.main = &nctx; + + pthread_attr_init (&attr); + pthread_attr_setstack (&attr, sptr, (size_t)ssize); + pthread_attr_setscope (&attr, PTHREAD_SCOPE_PROCESS); + pthread_create (&ctx->id, &attr, trampoline, &args); + + coro_transfer (args.main, args.self); + } + else + ctx->id = null_tid; + #else -# error unsupported architecture +# error unsupported backend #endif }