Improve noinline use

- introduce macros NOINLINE and NOINLINE_DECL
- adjust existing code in task.c
- rewrite change for jl_egal using NOINLINE
- add extensive comment explaining jl_egal change

src/builtins.c

@@ -205,12 +205,27 @@ static int bits_equal(void *a, void *b, int sz)
  }
 }
 
-#if defined(_OS_WINDOWS_) && !defined(_COMPILER_MINGW_)
-static int __declspec(noinline)
-#else
-static int __attribute__((noinline))
-#endif
-compare_tuple(jl_value_t *a, jl_value_t *b)
+// jl_egal
+// The frequently used jl_egal function deserves special attention when it
+// comes to performance which is made challenging by the fact that the
+// function has to handle quite a few different cases and because it is
+// called recursively. To optimize performance many special cases are
+// handle with separate comparisons which can dramatically reduce the run
+// time of the function. The compiler can translate these simple tests
+// with little effort, e.g., few registers are used.
+//
+// The complex cases require more effort and more registers to be translated
+// efficiently. The effected cases include comparing tuples and fields. If
+// the code to perform these operation would be inlined in the jl_egal
+// function then the compiler would generate at the or close to the top of
+// the function a prologue which saves all the callee-save registers and at
+// the end the respective epilogue. The result is that even the fast cases
+// are slowed down.
+//
+// The solution is to keep the code in jl_egal simple and split out the
+// (more) complex cases into their own functions which are marked with
+// NOINLINE.
+static int NOINLINE compare_tuple(jl_value_t *a, jl_value_t *b)
 {
  size_t l = jl_tuple_len(a);
  if (l != jl_tuple_len(b))
@@ -222,12 +237,9 @@ compare_tuple(jl_value_t *a, jl_value_t *b)
  return 1;
 }
 
-#if defined(_OS_WINDOWS_) && !defined(_COMPILER_MINGW_)
-static int __declspec(noinline)
-#else
-static int __attribute__((noinline))
-#endif
-compare_fields(jl_value_t *a, jl_value_t *b, jl_datatype_t *dt, size_t nf)
+// See comment above for an explanation of NOINLINE.
+static int NOINLINE compare_fields(jl_value_t *a, jl_value_t *b,
+ jl_datatype_t *dt, size_t nf)
 {
  for (size_t f=0; f < nf; f++) {
  size_t offs = dt->fields[f].offset;

src/support/dtypes.h

@@ -127,8 +127,16 @@
 # define STATIC_INLINE static __inline
 # define INLINE __inline
 #else
-# define STATIC_INLINE static inline
-# define INLINE inline
+# define STATIC_INLINE static inline
+# define INLINE inline
+#endif
+
+#if defined(_OS_WINDOWS_) && !defined(_COMPILER_MINGW_)
+# define NOINLINE __declspec(noinline)
+# define NOINLINE_DECL(f) __declspec(noinline) f
+#else
+# define NOINLINE __attribute__((noinline))
+# define NOINLINE_DECL(f) f __attribute__((noinline))
 #endif
 
 typedef int bool_t;

src/task.c

@@ -74,11 +74,7 @@ static void boundlow(struct _probe_data *p)
 }
 
 // we need this function to exist so we can measure its stack frame!
-#if defined(_OS_WINDOWS_) && !defined(_COMPILER_MINGW_)
-static void __declspec(noinline) fill(struct _probe_data *p);
-#else
-static void fill(struct _probe_data *p) __attribute__ ((noinline));
-#endif
+static void NOINLINE_DECL(fill(struct _probe_data *p));
 
 static void fill(struct _probe_data *p)
 {
@@ -166,12 +162,7 @@ void *jl_stackbase;
 static jl_jmp_buf jl_base_ctx; // base context of stack
 #endif
 
-#if defined(_OS_WINDOWS_) && !defined(_COMPILER_MINGW_)
-static void __declspec(noinline)
-#else
-static void __attribute__((noinline))
-#endif
-save_stack(jl_task_t *t)
+static void NOINLINE save_stack(jl_task_t *t)
 {
  if (t->state == done_sym || t->state == failed_sym)
  return;
@@ -190,12 +181,7 @@ save_stack(jl_task_t *t)
  memcpy(buf, (char*)&_x, nb);
 }
 
-#if defined(_OS_WINDOWS_) && !defined(_COMPILER_MINGW_)
-void __declspec(noinline)
-#else
-void __attribute__((noinline))
-#endif
-restore_stack(jl_task_t *t, jl_jmp_buf *where, char *p)
+void NOINLINE restore_stack(jl_task_t *t, jl_jmp_buf *where, char *p)
 {
  char *_x = (char*)jl_stackbase - t->ssize;
  if (!p) {
@@ -235,12 +221,7 @@ static void NORETURN finish_task(jl_task_t *t, jl_value_t *resultval)
  abort();
 }
 
-#if defined(_OS_WINDOWS_) && !defined(_COMPILER_MINGW_)
-static void __declspec(noinline)
-#else
-static void __attribute__((noinline))
-#endif
-NORETURN start_task()
+static void NOINLINE NORETURN start_task()
 {
  // this runs the first time we switch to a task
  jl_task_t *t = jl_current_task;
@@ -251,12 +232,7 @@ NORETURN start_task()
 }
 
 #ifndef ASM_COPY_STACKS
-#if defined(_OS_WINDOWS_) && !defined(_COMPILER_MINGW_)
-static void __declspec(noinline)
-#else
-static void __attribute__((noinline))
-#endif
-set_base_ctx(char *__stk)
+static void NOINLINE set_base_ctx(char *__stk)
 {
  if (jl_setjmp(jl_base_ctx, 1)) {
  start_task();