[trans-mem] optimize transaction local malloc/alloca memory

[Aldy Hernandez]

On Fri, Dec 04, 2009 at 08:32:13AM -0800, Richard Henderson wrote:
> On 12/04/2009 06:26 AM, Aldy Hernandez wrote:
> >+  p = malloc (sizeof (*p) * 100);
> >+
> >+  __transaction {
> >+    /* Instrument store, since memory pointed to by P is not
> >+       transaction local.  */
> >+    p[5] = 123;
> >+
> >+    if (something())
> >+      __transaction_cancel;
> >+  }
> >+  return p[5];
> >+}
> >+
> >+/* { dg-final { scan-tree-dump-times "ITM_WU" 1 "tmmark" } } */
> 
> Not the best test case to add, since P is thread-private. You could
> instrument this with either a log function or a local copy.

I assume you're going to make me handle dereferencing of thread-private
memory with log functions or a local copy :-).

How about I make sure there are no instrumented calls whatsoever:

	{ scan-tree-dump-times "ITM_WU|ITM_LU|tm_save" 1 "tmmark" }

?

> >+      if (DECL_BUILT_IN_CLASS (fn) == BUILT_IN_NORMAL
> >+	&&  (DECL_FUNCTION_CODE (fn) == BUILT_IN_MALLOC
> >+	      || DECL_FUNCTION_CODE (fn) == BUILT_IN_ALLOCA))
> 
> gimple_call_flags (stmt) & ECF_MALLOC

I thought of doing that, but decided against it because EFC_MALLOC was
set for user defined functions with the "malloc" attribute.  Though, as
you pointed out to Richi, perhaps that is exactly what we want.

> >@@ -1149,6 +1219,8 @@ requires_barrier (basic_block entry_bloc
> >    switch (TREE_CODE (x))
> >      {
> >      case INDIRECT_REF:
> >+      if (transaction_local_dynamic_memory_p (entry_block, TREE_OPERAND (x, 0)))
> >+	return false;
> 
> You're going to do the whole walking and analysis thing for every
> memory reference?  I could imagine it'd be a lot better to cache
> this data so that in the next statement, which might well also use
> the same base SSA_NAME, doesn't have to re-do the analysis.

Fair enough.  Will do.

Aldy