[PATCH] Add --param max-vartrack-reverse-op-size=N (PR debug/54402)

[PATCH] Add --param max-vartrack-reverse-op-size=N (PR debug/54402)

[Jakub Jelinek]

Hi!

As the PR54402 go testcase shows, sometimes we end up with VALUEs with huge
(many thousands) long locs lists because we add too many reverse ops for the
same value.  Such huge locs lists of equivalences are hardly useful in
practice, reverse ops mainly help when there are just a few locs and most of
them don't have a register or memory location, and make e.g. find_base_term
or get_addr extremely expensive.

The following patch avoids adding further reverse op equivalences for VALUEs
if the locs list is already big enough, the default value of 50 makes the
go1 testcase significantly faster, but debug info/debug_loc size on x86_64-linux as
well as i686-linux was identical (for cc1plus/go1/libstdc++.so.6 I've
checked) between default of 50 and 1000, so I assume it didn't affect debug
info quality for those, and hope it is unlikely it will for other real-world
programs.

Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?

2012-12-12  Jakub Jelinek  <jakub@redhat.com>

	PR debug/54402
	* params.def (PARAM_MAX_VARTRACK_REVERSE_OP_SIZE): New param.
	* var-tracking.c (reverse_op): Don't add reverse ops to
	VALUEs that have already
	PARAM_VALUE (PARAM_MAX_VARTRACK_REVERSE_OP_SIZE) or longer
	locs list.

--- gcc/params.def.jj	2012-11-23 10:01:50.000000000 +0100
+++ gcc/params.def	2012-12-12 16:26:21.063473958 +0100
@@ -869,6 +869,14 @@ DEFPARAM (PARAM_MAX_VARTRACK_EXPR_DEPTH,
 	  "Max. recursion depth for expanding var tracking expressions",
 	  12, 0, 0)
 
+/* Set maximum length of value location list for which var tracking
+   should add reverse operations.  */
+
+DEFPARAM (PARAM_MAX_VARTRACK_REVERSE_OP_SIZE,
+	  "max-vartrack-reverse-op-size",
+	  "Max. size of loc list for which reverse ops should be added",
+	  50, 0, 0)
+
 /* Set minimum insn uid for non-debug insns.  */
 
 DEFPARAM (PARAM_MIN_NONDEBUG_INSN_UID,
--- gcc/var-tracking.c.jj	2012-12-12 15:57:29.000000000 +0100
+++ gcc/var-tracking.c	2012-12-12 16:28:27.089745791 +0100
@@ -5545,6 +5545,7 @@ reverse_op (rtx val, const_rtx expr, rtx
   cselib_val *v;
   struct elt_loc_list *l;
   enum rtx_code code;
+  int count;
 
   if (GET_CODE (expr) != SET)
     return;
@@ -5586,10 +5587,13 @@ reverse_op (rtx val, const_rtx expr, rtx
   /* Adding a reverse op isn't useful if V already has an always valid
      location.  Ignore ENTRY_VALUE, while it is always constant, we should
      prefer non-ENTRY_VALUE locations whenever possible.  */
-  for (l = v->locs; l; l = l->next)
+  for (l = v->locs, count = 0; l; l = l->next, count++)
     if (CONSTANT_P (l->loc)
 	&& (GET_CODE (l->loc) != CONST || !references_value_p (l->loc, 0)))
       return;
+    /* Avoid creating too large locs lists.  */
+    else if (count == PARAM_VALUE (PARAM_MAX_VARTRACK_REVERSE_OP_SIZE))
+      return;
 
   switch (GET_CODE (src))
     {

	Jakub

Re: [PATCH] Add --param max-vartrack-reverse-op-size=N (PR debug/54402)

[Richard Henderson]

On 12/13/2012 01:39 AM, Jakub Jelinek wrote:
> 	PR debug/54402
> 	* params.def (PARAM_MAX_VARTRACK_REVERSE_OP_SIZE): New param.
> 	* var-tracking.c (reverse_op): Don't add reverse ops to
> 	VALUEs that have already
> 	PARAM_VALUE (PARAM_MAX_VARTRACK_REVERSE_OP_SIZE) or longer
> 	locs list.

Ok.


r~