Avoid inlining exact_unwind

On some architectures this can't be inlined. We don't actually care
whether or not it is inlined, only that whether it is or isn't inlined
is consistent regardless of build flags.

Signed-off-by: Matthias Diener <mdiener@illinois.edu>
Signed-off-by: Matt Wozniski <mwozniski@bloomberg.net>

src/memray/_memray/tracking_api.h

@@ -110,6 +110,7 @@ class NativeTrace
         if (size == MAX_SIZE) {
             d_data.resize(0);
             size = exact_unwind();
+            skip += 1;  // skip the non-inlined exact_unwind frame
             MAX_SIZE = MAX_SIZE * 2 > size ? MAX_SIZE * 2 : size;
             d_data.resize(MAX_SIZE);
         }
@@ -143,7 +144,9 @@ class NativeTrace
         return unw_backtrace((void**)data, MAX_SIZE);
     }
 
-    __attribute__((always_inline)) size_t inline exact_unwind()
+    // This can not be inlined as some architectures (e.g. ppc64le) don't
+    // support inlining functions that call setjmp()
+    __attribute__((noinline)) size_t inline exact_unwind()
     {
         unw_context_t context;
         if (unw_getcontext(&context) < 0) {