rt: More work on morestack

mk/rt.mk

@@ -67,7 +67,8 @@ RUNTIME_CS_$(1) := \
 
 RUNTIME_S_$(1) := rt/arch/$$(HOST_$(1))/_context.S \
                   rt/arch/$$(HOST_$(1))/ccall.S \
-                  rt/arch/$$(HOST_$(1))/morestack.S
+                  rt/arch/$$(HOST_$(1))/morestack.S \
+                  rt/arch/$$(HOST_$(1))/record_sp.S
 
 RUNTIME_HDR_$(1) := rt/globals.h \
                rt/rust.h \

src/rt/arch/i386/morestack.S

@@ -2,40 +2,117 @@
 
 // __morestack
 //
-// LLVM generates a call to this to allocate more stack space in a functiono
+// LLVM generates a call to this to allocate more stack space in a function
 // prolog when we run out.
 
 #if defined(__APPLE__) || defined(_WIN32)
 #define RUST_NEW_STACK          _rust_new_stack
 #define RUST_DEL_STACK          _rust_del_stack
+#define RUST_GET_PREV_STACK     _rust_get_prev_stack
+#define RUST_GET_TASK           _rust_get_task
+#define UPCALL_ALLOC_C_STACK    _upcall_alloc_c_stack
+#define UPCALL_CALL_C_STACK     _upcall_call_c_stack
 #define MORESTACK               ___morestack
 #else
 #define RUST_NEW_STACK          rust_new_stack
 #define RUST_DEL_STACK          rust_del_stack
+#define RUST_GET_PREV_STACK     rust_get_prev_stack
+#define RUST_GET_TASK           rust_get_task
+#define UPCALL_ALLOC_C_STACK    upcall_alloc_c_stack
+#define UPCALL_CALL_C_STACK     upcall_call_c_stack
 #define MORESTACK               __morestack
 #endif
 
+#ifdef __APPLE__
+#define ALIGNMENT               4
+#else
+#define ALIGNMENT               8
+#endif
+
+#define RETURN_OFFSET   7
+
 .globl RUST_NEW_STACK
 .globl RUST_DEL_STACK
-
+.globl RUST_GET_PREV_STACK
+.globl RUST_GET_TASK
+.globl UPCALL_ALLOC_C_STACK
+.globl UPCALL_CALL_C_STACK
 .globl MORESTACK
 
 MORESTACK:
-    pushl 8(%esp)           //              argsz > ra stksz argsz x x ra args
-    leal 28(%esp),%eax      //                argsz ra stksz argsz x x ra args
-    pushl %eax              //         argp > argsz ra stksz argsz x x ra args
-    pushl 12(%esp)          //   stksz > argp argsz ra stksz argsz x x ra args
-    calll RUST_NEW_STACK
-    addl $12,%esp           //                      ra stksz argsz x x ra args
 
-    movl (%esp),%edx        // Grab the return pointer.
+    // Sanity check to make sure that there is a currently-running task.
-    incl %edx               // Skip past the `ret`.
+    subl $12,%esp
-    movl %eax,%esp          // Switch to the new stack.
+    calll RUST_GET_TASK
+    testl %eax,%eax
+    jz L$bail
+
+    subl $12,%esp
+    pushl $12
+    calll UPCALL_ALLOC_C_STACK
+    movl %eax,%edx
+
+                            // C stack              | esp+12
+                            // ---------------------+-------------------------
+    movl 20(%esp),%eax      //                      | ra stksz argsz x ra args
+    movl %eax,8(%edx)       //              argsz > | ra stksz argsz x ra args
+    leal 32(%esp),%eax      //                argsz | ra stksz argsz x ra args
+    movl %eax,4(%edx)       //         argp > argsz | ra stksz argsz x ra args
+    movl 16(%esp),%eax      //           argp argsz | ra stksz argsz x ra args
+    movl %eax,(%edx)        //   stksz > argp argsz | ra stksz argsz x ra args
+
+    calll L$pic_ref_pt_0
+L$pic_ref_pt_0:
+    popl %eax
+
+    movl rust_new_stack_sym-L$pic_ref_pt_0(%eax),%eax
+    movl %eax,(%esp)
+    movl %edx,4(%esp)
+    calll UPCALL_CALL_C_STACK
+
+    movl 16(%esp),%edx          // Grab the return pointer.
+    addl $RETURN_OFFSET,%edx    // Skip past the `add esp,4` and the `ret`.
+
+    movl %eax,%esp          // Switch stacks.
+    subl $12,%esp           // Align the stack.
     calll *%edx             // Re-enter the function that called us.
 
     // Now the function that called us has returned, so we need to delete the
     // old stack space.
-    calll RUST_DEL_STACK
+
-    movl %eax,%esp
+    calll RUST_GET_PREV_STACK
-    retl $8                 //                      ra stksz argsz x x ra args
+    movl %eax,%esp          // Switch back to the old stack.
+
+    movl $0,(%esp)
+    calll UPCALL_ALLOC_C_STACK
+
+    calll L$pic_ref_pt_1
+L$pic_ref_pt_1:
+    popl %edx
+
+    movl rust_del_stack_sym-L$pic_ref_pt_1(%edx),%edx
+    movl %edx,(%esp)
+    movl %eax,4(%esp)
+    calll UPCALL_CALL_C_STACK
+
+    addl $16,%esp
+    retl $16 + ALIGNMENT    //                      ra stksz argsz x ra args
+
+L$bail:
+    movl 12(%esp),%edx
+    addl $RETURN_OFFSET,%edx
+    addl $12+16,%esp
+    jmpl *%edx
+
+#ifdef __APPLE__
+
+	.section	__IMPORT,__pointers,non_lazy_symbol_pointers
+rust_new_stack_sym:
+    .indirect_symbol RUST_NEW_STACK
+    .long 0
+rust_del_stack_sym:
+    .indirect_symbol RUST_DEL_STACK
+    .long 0
+
+#endif
 

src/rt/rust.cpp

@@ -1,4 +1,5 @@
 #include "rust_internal.h"
+#include <cstdio>
 
 struct
 command_line_args : public kernel_owned<command_line_args>
@@ -75,6 +76,8 @@ int check_claims = 0;
 
 extern "C" CDECL int
 rust_start(uintptr_t main_fn, int argc, char **argv, void* crate_map) {
+    fprintf(stderr, "rust_start, argc=%d argv=%p\n", argc, argv);
+
     rust_env *env = load_env();
 
     update_log_settings(crate_map, env->logspec);

src/rt/rust_scheduler.cpp

@@ -1,6 +1,7 @@
 
 #include <stdarg.h>
 #include <cassert>
+#include <pthread.h>
 #include "rust_internal.h"
 #include "globals.h"
 
@@ -12,6 +13,9 @@ DWORD rust_scheduler::task_key;
 
 bool rust_scheduler::tls_initialized = false;
 
+// Defined in arch/*/record_sp.S.
+extern "C" void rust_record_sp(uintptr_t sp);
+
 rust_scheduler::rust_scheduler(rust_kernel *kernel,
                                rust_srv *srv,
                                int id) :
@@ -285,6 +289,8 @@ rust_scheduler::start_main_loop() {
              scheduled_task->state->name);
 
         place_task_in_tls(scheduled_task);
+        rust_record_sp(scheduled_task->stk->limit);
+        //pthread_setspecific(89, (void *)scheduled_task->stk->limit);
 
         interrupt_flag = 0;
 
@@ -374,6 +380,8 @@ rust_scheduler::place_task_in_tls(rust_task *task) {
 
 rust_task *
 rust_scheduler::get_task() {
+    if (!tls_initialized)
+        return NULL;
     rust_task *task = reinterpret_cast<rust_task *>
         (pthread_getspecific(task_key));
     assert(task && "Couldn't get the task from TLS!");
@@ -395,6 +403,8 @@ rust_scheduler::place_task_in_tls(rust_task *task) {
 
 rust_task *
 rust_scheduler::get_task() {
+    if (!tls_initialized)
+        return NULL;
     rust_task *task = reinterpret_cast<rust_task *>(TlsGetValue(task_key));
     assert(task && "Couldn't get the task from TLS!");
     return task;

src/rt/rust_task.cpp

@@ -8,11 +8,14 @@
 #ifndef __WIN32__
 #include <execinfo.h>
 #endif
+#include <iostream>
 #include <cassert>
 #include <cstring>
 
 #include "globals.h"
 
+#define RED_ZONE_SIZE   128
+
 // Stack size
 size_t g_custom_min_stack_size = 0;
 
@@ -34,16 +37,16 @@ new_stk(rust_scheduler *sched, rust_task *task, size_t minsz)
     size_t min_stk_bytes = get_min_stk_size(sched->min_stack_size);
     if (minsz < min_stk_bytes)
         minsz = min_stk_bytes;
-    size_t sz = sizeof(stk_seg) + minsz;
+    size_t sz = sizeof(stk_seg) + minsz + RED_ZONE_SIZE;
     stk_seg *stk = (stk_seg *)task->malloc(sz, "stack");
     LOGPTR(task->sched, "new stk", (uintptr_t)stk);
     memset(stk, 0, sizeof(stk_seg));
     stk->next = task->stk;
-    stk->limit = (uintptr_t) &stk->data[minsz];
+    stk->limit = (uintptr_t) &stk->data[minsz + RED_ZONE_SIZE];
     LOGPTR(task->sched, "stk limit", stk->limit);
     stk->valgrind_id =
         VALGRIND_STACK_REGISTER(&stk->data[0],
-                                &stk->data[minsz]);
+                                &stk->data[minsz + RED_ZONE_SIZE]);
     task->stk = stk;
     return stk;
 }
@@ -63,18 +66,32 @@ del_stk(rust_task *task, stk_seg *stk)
 // Entry points for `__morestack` (see arch/*/morestack.S).
 extern "C" void *
 rust_new_stack(size_t stk_sz, void *args_addr, size_t args_sz) {
+    std::cerr << "*** New stack!\n";
+
     rust_task *task = rust_scheduler::get_task();
+    if (!task)
+        return NULL;
+
     stk_seg *stk_seg = new_stk(task->sched, task, stk_sz);
     memcpy(stk_seg->data, args_addr, args_sz);
     return stk_seg->data;
 }
 
-extern "C" void *
+extern "C" void
 rust_del_stack() {
     rust_task *task = rust_scheduler::get_task();
-    stk_seg *next_seg = task->stk->next;
     del_stk(task, task->stk);
-    return next_seg->data;
+}
+
+extern "C" void *
+rust_get_prev_stack() {
+    rust_task *task = rust_scheduler::get_task();
+    return task->stk->next;
+}
+
+extern "C" rust_task *
+rust_get_task() {
+    return rust_scheduler::get_task();
 }
 
 

src/rt/rustrt.def.in

@@ -1,3 +1,4 @@
+__morestack
 align_of
 chan_id_send
 check_claims
@@ -28,7 +29,10 @@ rand_free
 rand_new
 rand_next
 refcount
+rust_del_stack
 rust_file_is_dir
+rust_getcwd
+rust_get_prev_stack
 rust_get_stdin
 rust_get_stdout
 rust_get_stderr
@@ -36,12 +40,14 @@ rust_str_push
 rust_list_files
 rust_port_detach
 rust_port_size
+rust_new_stack
 rust_process_wait
 rust_ptr_eq
 rust_run_program
 rust_start
 rust_getcwd
 rust_task_sleep
+rust_get_task
 set_min_stack
 sched_threads
 size_of
@@ -63,10 +69,10 @@ upcall_dynastack_mark
 upcall_fail
 upcall_free
 upcall_get_type_desc
-upcall_vec_grow
-upcall_vec_push
 upcall_log_type
 upcall_malloc
 upcall_rust_personality
 upcall_shared_malloc
 upcall_shared_free
+upcall_vec_grow
+upcall_vec_push