--- libcoro/coro.c	2008/04/24 12:40:38	1.31
+++ libcoro/coro.c	2008/11/05 01:54:34	1.36
@@ -122,43 +122,44 @@
 #endif
 
 #if CORO_ASM
-void __attribute__((__noinline__, __fastcall__))
-coro_transfer (struct coro_context *prev, struct coro_context *next)
-{
-  asm volatile (
+asm (
+     ".text\n"
+     ".globl coro_transfer\n"
+     ".type coro_transfer, @function\n"
+     "coro_transfer:\n"
 #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"
+# define NUM_ALIGN 1
+# define NUM_SAVED 5
+     "\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"
 #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"
+# define NUM_ALIGN 1
+# define NUM_SAVED 4
+     "\tpush %ebx\n"
+     "\tpush %esi\n"
+     "\tpush %edi\n"
+     "\tpush %ebp\n"
+     "\tmov  %esp, (%eax)\n"
+     "\tmov  (%edx), %esp\n"
+     "\tpop  %ebp\n"
+     "\tpop  %edi\n"
+     "\tpop  %esi\n"
+     "\tpop  %ebx\n"
 #else
 # error unsupported architecture
 #endif
-     : "=m" (prev->sp)
-     : "m" (next->sp)
-  );
-}
+     "\tret\n"
+);
 #endif
 
 #if CORO_PTHREAD
@@ -187,6 +188,8 @@
   return 0;
 }
 
+asm("");
+
 void coro_transfer(coro_context *prev, coro_context *next)
 {
   pthread_cond_init (&prev->c, 0);
@@ -208,7 +211,7 @@
 
   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);
@@ -283,17 +286,17 @@
 
   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"
 #endif
@@ -303,16 +306,16 @@
   _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
@@ -321,14 +324,14 @@
 
   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 -= NUM_ALIGN;
   *--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