[3_0_BRANCH] Backport patch to fix KDE2 blocking bug PR 3387

[Franz Sirl <Franz dot Sirl-kernel at lauterbach dot com>]

Hi,

in an effort to get the "don't use gcc3 with KDE2" label removed I decided to 
sleep with the devil and debug x86 stuff :-). After a short fight with my 
rusty x86 assembly knowledge I was successful in tracking down the patch that 
fixes this PR on the mainline.

I would like to commit the appended backported patch (and the original 
testcase execute/20010403-1.c) to the branch, it fixes the posted testcase 
and also works in the original context producing a working libxml2 on x86.

Jakub, do you by chance remember if this patch relies on anything that may 
not be on the branch yet? Maybe this one:

Wed Apr  4 00:45:38 EDT 2001  John Wehle  (john@feith.com)

        * rtl.h (set_noop_p): Declare.
        * flow.c (set_noop_p): Move from here ...
        * rtlanal.c (set_noop_p): ... to here and enhance.
        * cse.c (delete_trivially_dead_insns): Use it.
        * gcse.c (hash_scan_set): Likewise.
        * jump.c (delete_noop_moves): Likewise.
        * recog.c (split_all_insns): Likewise.

?

Bootstrapped and regtested the branch on x86-linux and ppc-linux (with Jim 
Wilson's fixup patch applied).

OK to commit to the branch?

Franz.

2001-04-20  Jakub Jelinek  <jakub@redhat.com>

	* gcse.c (gcse_main): Fix comment typo.
	(delete_null_pointer_check): Likewise.
	(hash_scan_set): Don't consider sets with REG_EQUIV MEM notes.
	* cse.c (cse_insn): Likewise.
	* function.c (fixup_var_refs_insns_with_hash): The sequence is
	toplevel.

Index: gcc/cse.c
===================================================================
RCS file: /cvs/gcc/egcs/gcc/cse.c,v
retrieving revision 1.174.2.5
diff -u -p -r1.174.2.5 cse.c
--- gcc/cse.c	2001/07/26 18:27:52	1.174.2.5
+++ gcc/cse.c	2001/09/11 08:21:57
@@ -5060,18 +5069,16 @@ cse_insn (insn, libcall_insn)
       sets[i].src_in_memory = hash_arg_in_memory;
 
       /* If SRC is a MEM, there is a REG_EQUIV note for SRC, and DEST is
-	 a pseudo that is set more than once, do not record SRC.  Using
-	 SRC as a replacement for anything else will be incorrect in that
-	 situation.  Note that this usually occurs only for stack slots,
-	 in which case all the RTL would be referring to SRC, so we don't
-	 lose any optimization opportunities by not having SRC in the
-	 hash table.  */
+	 a pseudo, do not record SRC.  Using SRC as a replacement for
+	 anything else will be incorrect in that situation.  Note that
+	 this usually occurs only for stack slots, in which case all the
+	 RTL would be referring to SRC, so we don't lose any optimization
+	 opportunities by not having SRC in the hash table.  */
 
       if (GET_CODE (src) == MEM
-	  && find_reg_note (insn, REG_EQUIV, src) != 0
+	  && find_reg_note (insn, REG_EQUIV, NULL_RTX) != 0
 	  && GET_CODE (dest) == REG
-	  && REGNO (dest) >= FIRST_PSEUDO_REGISTER
-	  && REG_N_SETS (REGNO (dest)) != 1)
+	  && REGNO (dest) >= FIRST_PSEUDO_REGISTER)
 	sets[i].src_volatile = 1;
 
 #if 0
Index: gcc/function.c
===================================================================
RCS file: /cvs/gcc/egcs/gcc/function.c,v
retrieving revision 1.250.2.13
diff -u -p -r1.250.2.13 function.c
--- gcc/function.c	2001/08/06 21:26:24	1.250.2.13
+++ gcc/function.c	2001/09/11 08:22:02
@@ -1680,7 +1680,7 @@ fixup_var_refs_insns_with_hash (ht, var,
       rtx insn = XEXP (insn_list, 0);
 	
       if (INSN_P (insn))
-	fixup_var_refs_insn (insn, var, promoted_mode, unsignedp, 0);
+	fixup_var_refs_insn (insn, var, promoted_mode, unsignedp, 1);
 
       insn_list = XEXP (insn_list, 1);
     }
Index: gcc/gcse.c
===================================================================
RCS file: /cvs/gcc/egcs/gcc/gcse.c,v
retrieving revision 1.113.4.2
diff -u -p -r1.113.4.2 gcse.c
--- gcc/gcse.c	2001/04/11 18:06:52	1.113.4.2
+++ gcc/gcse.c	2001/09/11 08:22:05
@@ -681,7 +681,7 @@ gcse_main (f, file)
      a high connectivity will take a long time and is unlikely to be
      particularly useful.
 
-     In normal circumstances a cfg should have about twice has many edges
+     In normal circumstances a cfg should have about twice as many edges
      as blocks.  But we do not want to punish small functions which have
      a couple switch statements.  So we require a relatively large number
      of basic blocks and the ratio of edges to blocks to be high.  */
@@ -1924,6 +1924,7 @@ hash_scan_set (pat, insn, set_p)
 {
   rtx src = SET_SRC (pat);
   rtx dest = SET_DEST (pat);
+  rtx note;
 
   if (GET_CODE (src) == CALL)
     hash_scan_call (src, insn);
@@ -1939,7 +1940,15 @@ hash_scan_set (pat, insn, set_p)
 	  /* Don't GCSE something if we can't do a reg/reg copy.  */
 	  && can_copy_p [GET_MODE (dest)]
 	  /* Is SET_SRC something we want to gcse?  */
-	  && want_to_gcse_p (src))
+	  && want_to_gcse_p (src)
+	  /* Don't GCSE if it has attached REG_EQUIV note.
+	     At this point this only function parameters should have
+	     REG_EQUIV notes and if the argument slot is used somewhere
+	     explicitely, it means address of parameter has been taken,
+	     so we should not extend the lifetime of the pseudo.  */
+	  && ((note = find_reg_note (insn, REG_EQUIV, NULL_RTX)) == 0
+	      || GET_CODE (XEXP (note, 0)) != MEM))
+
 	{
 	  /* An expression is not anticipatable if its operands are
 	     modified before this insn.  */
@@ -5173,7 +5182,7 @@ delete_null_pointer_checks (f)
      a high connectivity will take a long time and is unlikely to be
      particularly useful.
 
-     In normal circumstances a cfg should have about twice has many edges
+     In normal circumstances a cfg should have about twice as many edges
      as blocks.  But we do not want to punish small functions which have
      a couple switch statements.  So we require a relatively large number
      of basic blocks and the ratio of edges to blocks to be high.  */