This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

[tree-ssa]: Starting explaining various SSAPRE algorithms


I've started to comment the various SSAPRE algorithms in detail, starting with renaming step 1, in the hopes that some human won't actually have to read and attempt to understand the aborted attempt of a useful paper that exists out there.

Hopefully this makes renaming a bit clearer.

2003-01-14 Daniel Berlin <dberlin@dberlin.org>

* tree-ssa-pre.c (rename_1): Add more comments.

Index: tree-ssa-pre.c
===================================================================
RCS file: /cvs/gcc/gcc/gcc/Attic/tree-ssa-pre.c,v
retrieving revision 1.1.4.126
diff -u -3 -p -r1.1.4.126 tree-ssa-pre.c
--- tree-ssa-pre.c	21 Jan 2004 15:58:26 -0000	1.1.4.126
+++ tree-ssa-pre.c	22 Jan 2004 17:47:06 -0000
@@ -1503,7 +1503,15 @@ process_delayed_rename (struct expr_info

 /* Renaming is done like Open64 does it.  Basically as the paper says,
    except that we try to use earlier defined occurrences if they are
-   available in order to keep the number of saves down.  */
+   available in order to keep the number of saves down.
+
+   For the uninitiated, the algorithm is a modified SSA renaming
+   algorithm (working on expressions rather than variables) .  We
+   attempt to determine which expression occurrences have the same
+   ESSA version (we call it class, for equivalence class, which is
+   what the papers call it.  Open64 calls it e-version), and which
+   occurrences are actually operands for an EPHI (since this has to be
+   discovered from the program).  */

 static void
 rename_1 (struct expr_info *ei)
@@ -1515,46 +1523,77 @@ rename_1 (struct expr_info *ei)

VARRAY_GENERIC_PTR_NOGC_INIT (stack, 1, "Stack for renaming");

+  /* Start by inserting the occurrences in preorder, dominator tree
+     order.  */
   insert_occ_in_preorder_dt_order (ei);

+  /* Walk the occurrences.  */
   for (i = 0; i < VARRAY_ACTIVE_SIZE (ei->euses_dt_order); i++)
     {
       occur = VARRAY_TREE (ei->euses_dt_order, i);

+ /* The current occurrence can't have the same version as
+ something on the top of the stack unless it is in a basic
+ block dominated by the basic block of the occurrence on the
+ top of the stack. */
while (VARRAY_ACTIVE_SIZE (stack) > 0
&& !dominated_by_p (CDI_DOMINATORS,
bb_for_stmt (occur),
bb_for_stmt (VARRAY_TOP_TREE (stack))))

VARRAY_POP (stack);
+
+ /* If the stack is empty, we assign a new version since it isn't
+ dominated by any other version. */
if (VARRAY_TOP_TREE (stack) == NULL || VARRAY_ACTIVE_SIZE (stack) == 0)
{
if (TREE_CODE (occur) == EPHI_NODE)
assign_new_class (occur, &stack, NULL);
else if (TREE_CODE (occur) == EUSE_NODE && !EUSE_PHIOP (occur))
assign_new_class (occur, &stack, NULL);
-
}
else
{
if (TREE_CODE (occur) == EUSE_NODE && !EUSE_PHIOP (occur))
{
tree tos = VARRAY_TOP_TREE (stack);
+
if (TREE_CODE (tos) == EUSE_NODE && !EUSE_PHIOP (tos))
{
+ /* If two real occurrences have the same
+ e-version/class, then this occurrence can be
+ defined by the prior occurrence (or whatever
+ the prior occurrence is defined by, if
+ anything).
+ Otherwise, we have to assign a new version.
+ lvalue occurrences always need a new version,
+ since they are definitions. */
if (!EUSE_LVAL (occur)
&& same_e_version_real_occ_real_occ (ei, tos, occur))
{
+
+
tree newdef;
EREF_CLASS (occur) = EREF_CLASS (tos);
newdef = EUSE_DEF (tos) != NULL ? EUSE_DEF (tos) : tos;
EUSE_DEF (occur) = newdef;
- }
+ }
else
assign_new_class (occur, &stack, NULL);
}
else if (TREE_CODE (tos) == EPHI_NODE)
{
+ /* Here we have an ephi occurrence at the top of the
+ stack, and the current occurrence is a real
+ occurrence. So determine if the real occurrence
+ has the same version as the result of the phi.
+ If so, then this real occurrence is defined by the
+ EPHI at the top of the stack.
+ If not, the EPHI isn't downsafe (because something
+ must change in between the ephi result and the next
+ occurrence), and we need a new version for the real
+ occurrence.
+ lvalues always need a new version. */
if (!EUSE_LVAL (occur)
&& same_e_version_phi_result (ei, tos, EREF_STMT (occur),
occur))
@@ -1572,10 +1611,15 @@ rename_1 (struct expr_info *ei)
}
}
}
+ /* EPHI occurrences always get new versions. */
else if (TREE_CODE (occur) == EPHI_NODE)
- {
+ {
assign_new_class (occur, &stack, NULL);
}
+ /* EPHI operands are optimistcally assumed to be whatever is
+ at the top of the stack at the time we hit the ephi
+ operand occurrence. The delayed renaming checks this
+ optimistic assumption for validity. */
else if (TREE_CODE (occur) == EUSE_NODE && EUSE_PHIOP (occur))
{
basic_block pred_bb = bb_for_stmt (occur);
@@ -1593,6 +1637,7 @@ rename_1 (struct expr_info *ei)
}
}
}
+ /* No EPHI can be downsafe past an exit node. */
else if (TREE_CODE (occur) == EEXIT_NODE)
{
if (VARRAY_ACTIVE_SIZE (stack) > 0
@@ -1614,6 +1659,10 @@ rename_1 (struct expr_info *ei)
fprintf (tree_dump_file, "\n");
}
}
+
+ /* Now process the renames for EPHI's that actually have the result
+ valid and used. These will be the EPHI's that have the statement
+ set above. */
FOR_EACH_BB (phi_bb)
{
tree ephi = ephi_at_block (phi_bb);



Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]