diff --git a/boehmgc-coroutine-sp-fallback.diff b/boehmgc-coroutine-sp-fallback.diff
new file mode 100644
index 000000000..8fdafbecb
--- /dev/null
+++ b/boehmgc-coroutine-sp-fallback.diff
@@ -0,0 +1,77 @@
+diff --git a/darwin_stop_world.c b/darwin_stop_world.c
+index 3dbaa3fb..36a1d1f7 100644
+--- a/darwin_stop_world.c
++++ b/darwin_stop_world.c
+@@ -352,6 +352,7 @@ GC_INNER void GC_push_all_stacks(void)
+   int nthreads = 0;
+   word total_size = 0;
+   mach_msg_type_number_t listcount = (mach_msg_type_number_t)THREAD_TABLE_SZ;
++  size_t stack_limit;
+   if (!EXPECT(GC_thr_initialized, TRUE))
+     GC_thr_init();
+ 
+@@ -407,6 +408,19 @@ GC_INNER void GC_push_all_stacks(void)
+             GC_push_all_stack_sections(lo, hi, p->traced_stack_sect);
+           }
+           if (altstack_lo) {
++            // When a thread goes into a coroutine, we lose its original sp until
++            // control flow returns to the thread.
++            // While in the coroutine, the sp points outside the thread stack,
++            // so we can detect this and push the entire thread stack instead,
++            // as an approximation.
++            // We assume that the coroutine has similarly added its entire stack.
++            // This could be made accurate by cooperating with the application
++            // via new functions and/or callbacks.
++            stack_limit = pthread_get_stacksize_np(p->id);
++            if (altstack_lo >= altstack_hi || altstack_lo < altstack_hi - stack_limit) { // sp outside stack
++              altstack_lo = altstack_hi - stack_limit;
++            }
++
+             total_size += altstack_hi - altstack_lo;
+             GC_push_all_stack(altstack_lo, altstack_hi);
+           }
+diff --git a/pthread_stop_world.c b/pthread_stop_world.c
+index 4b2c429..1fb4c52 100644
+--- a/pthread_stop_world.c
++++ b/pthread_stop_world.c
+@@ -673,6 +673,8 @@ GC_INNER void GC_push_all_stacks(void)
+     struct GC_traced_stack_sect_s *traced_stack_sect;
+     pthread_t self = pthread_self();
+     word total_size = 0;
++    size_t stack_limit;
++    pthread_attr_t pattr;
+ 
+     if (!EXPECT(GC_thr_initialized, TRUE))
+       GC_thr_init();
+@@ -722,6 +724,31 @@ GC_INNER void GC_push_all_stacks(void)
+           hi = p->altstack + p->altstack_size;
+           /* FIXME: Need to scan the normal stack too, but how ? */
+           /* FIXME: Assume stack grows down */
++        } else {
++          if (pthread_getattr_np(p->id, &pattr)) {
++            ABORT("GC_push_all_stacks: pthread_getattr_np failed!");
++          }
++          if (pthread_attr_getstacksize(&pattr, &stack_limit)) {
++            ABORT("GC_push_all_stacks: pthread_attr_getstacksize failed!");
++          }
++          if (pthread_attr_destroy(&pattr)) {
++            ABORT("GC_push_all_stacks: pthread_attr_destroy failed!");
++          }
++          // When a thread goes into a coroutine, we lose its original sp until
++          // control flow returns to the thread.
++          // While in the coroutine, the sp points outside the thread stack,
++          // so we can detect this and push the entire thread stack instead,
++          // as an approximation.
++          // We assume that the coroutine has similarly added its entire stack.
++          // This could be made accurate by cooperating with the application
++          // via new functions and/or callbacks.
++          #ifndef STACK_GROWS_UP
++            if (lo >= hi || lo < hi - stack_limit) { // sp outside stack
++              lo = hi - stack_limit;
++            }
++          #else
++          #error "STACK_GROWS_UP not supported in boost_coroutine2 (as of june 2021), so we don't support it in Nix."
++          #endif
+         }
+         GC_push_all_stack_sections(lo, hi, traced_stack_sect);
+ #       ifdef STACK_GROWS_UP