Re: optimization/10024 [3.3 regression] [HP-PA] inline optimization ICE

[Zack Weinberg <zack@codesourcery.com>]

The following reply was made to PR optimization/10024; it has been noted by GNATS.

From: Zack Weinberg <zack@codesourcery.com>
To: gcc-patches@gcc.gnu.org
Cc: gcc-gnats@gcc.gnu.org, tausq@debian.org
Subject: Re: optimization/10024 [3.3 regression] [HP-PA] inline optimization
 ICE
Date: Wed, 26 Mar 2003 10:27:54 -0800

 [apologies if you get this twice, I mangled the destination addresses
 the first time]
 
 I was able to reproduce this bug with an x86->HPPA cross compiler.
 
 In 3.3, force_nonfallthru_and_redirect is called thus:
 
 Breakpoint 4, force_nonfallthru_and_redirect (e=0x83896b0, target=0x8389110)
     at ../../gcc/cfgrtl.c:936
 936       basic_block jump_block, new_bb = NULL, src = e->src;
 (gdb) p *e
 $3 = {pred_next = 0x8389408, succ_next = 0x0, src = 0x8388f80, 
   dest = 0x8389110, insns = 0x0, aux = 0x0, flags = 1, probability = 10000, 
   count = 0}
 
 Note that e->dest == target.  These are the basic blocks involved:
 
 ;; Basic block 13, loop depth 0, count 0
 ;; Predecessors:  12 [100.0%]  (fallthru) 7 [100.0%]  (fallthru) 4 [100.0%]  (fallthru)
 ;; Registers live at start: 2 [%r2] 20 [%r20] 30 [%r30]
 (code_label 159 266 236 13 13 "" [0 uses])
 (note 236 159 162 13 [bb 13] NOTE_INSN_BASIC_BLOCK)
 (note 162 236 180 13 NOTE_INSN_DELETED)
 (jump_insn:TI 180 162 142 13 0x40331688 (parallel [
             (set (pc)
                 (if_then_else (ge (reg/v:SI 20 %r20 [123])
                         (const_int 0 [0x0]))
                     (label_ref 186)
                     (pc)))
             (set (reg/v:SI 28 %r28 [104])
                 (reg/v:SI 20 %r20 [123]))
         ]) 232 {*pa.md:7406} (nil)
     (expr_list:REG_DEAD (reg/v:SI 20 %r20 [123])
         (expr_list:REG_BR_PROB (const_int 10000 [0x2710])
             (nil))))
 ;; Registers live at end: 2 [%r2] 28 [%r28] 30 [%r30]
 ;; Successors:  14 [100.0%]  (fallthru)
 
 ;; Basic block 14, loop depth 0, count 0
 ;; Predecessors:  13 [100.0%]  (fallthru) 1 [100.0%]  (fallthru)
 ;; Registers live at start: 2 [%r2] 28 [%r28] 30 [%r30]
 (code_label 186 55 241 14 19 "" [1 uses])
 (note 241 186 187 14 [bb 14] NOTE_INSN_BASIC_BLOCK)
 (note 187 241 188 14 NOTE_INSN_DELETED)
 (note 188 187 195 14 NOTE_INSN_DELETED)
 (note 195 188 196 14 NOTE_INSN_DELETED)
 (note 196 195 208 14 NOTE_INSN_DELETED)
 (note 208 196 189 14 NOTE_INSN_FUNCTION_END)
 (insn:TI 189 208 197 14 0x40331688 (set (reg/v:SI 19 %r19 [153])
         (ashiftrt:SI (reg/v:SI 28 %r28 [104])
             (const_int 16 [0x10]))) 189 {*pa.md:5429} (nil)
     (nil))
 (insn 197 189 198 14 0x4033165c (set (reg:SI 28 %r28 [158])
         (zero_extract:SI (reg/v:SI 28 %r28 [104])
             (const_int 1 [0x1])
             (const_int 16 [0x10]))) 225 {extzv_32} 
     (insn_list:REG_DEP_ANTI 189 (nil))
     (nil))
 (insn:TI 198 197 214 14 0x4033165c (set (reg/v:SI 28 %r28 [151])
         (plus:SI (reg/v:SI 19 %r19 [153])
             (reg:SI 28 %r28 [158]))) 121 {addsi3}
     (insn_list 197 (insn_list 189 (nil)))
     (expr_list:REG_DEAD (reg/v:SI 19 %r19 [153])
         (nil)))
 (insn 214 198 269 14 0x40331528 (use (reg/i:SI 28 %r28)) -1
     (insn_list:REG_DEP_ANTI 189 
       (insn_list:REG_DEP_ANTI 197 (insn_list 198 (nil))))
     (nil))
 (note 269 214 270 14 NOTE_INSN_EPILOGUE_BEG)
 (jump_insn:TI 270 269 271 14 (nil) (parallel [
             (return)
             (use (reg:SI 2 %r2))
             (const_int 0 [0x0])
         ]) 197 {return} 
     (insn_list:REG_DEP_ANTI 214 (insn_list 189 (insn_list 198 (nil))))
     (expr_list:REG_DEAD (reg:SI 2 %r2)
         (nil)))
 ;; Registers live at end: 28 [%r28] 30 [%r30]
 ;; Successors:  EXIT [100.0%] 
 
 
 The crash occurs
 
       note = find_reg_note (e->src->end, REG_BR_PROB, NULL_RTX);
       if (note)
 	{
 	  int prob = INTVAL (XEXP (note, 0));
 
 	  b->probability = prob;  /* here */
 
 because make_edge returned NULL, which is what it is documented to do
 when the edge it is asked to create already exists.  I am dubious
 about this interface - the caller presumably wants to get its hands on
 the new edge, so wouldn't it make sense to return the already-existent
 one?  But that's too invasive a change for 3.3.
 
 The minimal fix for this bug is to notice when "e" is the edge we
 want, and use it instead of calling make_edge.  This solves the crash,
 and appears to generate correct code (my ability to read HPPA assembly
 language is limited).  It will not fix a more general bug where the
 edge from e->src to target already exists, but is not e.  I don't know
 if that can happen.
 
 In 3.4 something different is happening -- we get a verify_flow_info
 failure, and this code path is never taken.  I will continue looking.
 I'd appreciate comments on the appended patch for 3.3.
 
 zw
 
         * cfgrtl.c (force_nonfallthru_and_redirect): If e is the edge
         we want, use it.
 
 ===================================================================
 Index: cfgrtl.c
 --- cfgrtl.c	25 Mar 2003 20:31:42 -0000	1.61.2.4
 +++ cfgrtl.c	26 Mar 2003 18:23:55 -0000
 @@ -946,7 +946,11 @@ force_nonfallthru_and_redirect (e, targe
        && JUMP_LABEL (e->src->end) == e->dest->head)
      {
        rtx note;
 -      edge b = make_edge (e->src, target, 0);
 +      edge b;
 +      if (e->dest == target)
 +	b = e;
 +      else
 +	b = make_edge (e->src, target, 0);
  
        if (!redirect_jump (e->src->end, block_label (target), 0))
  	abort ();