[PATCH] Fix MIPS-specific ICE in gcc.dg/pr77834.c (PR rtl-optimization/79150).

[Toma Tabacu <Toma dot Tabacu at imgtec dot com>]

Hi,

This ICE is caused by "gcc_assert (!JUMP_P (last))" in
commit_one_edge_insertion (gcc/cfgrtl.c):

  if (returnjump_p (last))
    {
      /* ??? Remove all outgoing edges from BB and add one for EXIT.
	 This is not currently a problem because this only happens
	 for the (single) epilogue, which already has a fallthru edge
	 to EXIT.  */

      e = single_succ_edge (bb);
      gcc_assert (e->dest == EXIT_BLOCK_PTR_FOR_FN (cfun)
		  && single_succ_p (bb) && (e->flags & EDGE_FALLTHRU));

      e->flags &= ~EDGE_FALLTHRU;
      emit_barrier_after (last);

      if (before)
	delete_insn (before);
    }
  else
    gcc_assert (!JUMP_P (last));

The assert is triggered because we're removing an edge which ends on
a conditional jump, which is part of a loop.

The reason for the existence of this loop is the size of the vector used in
gcc.dg/pr77834.c.
When compiling code which uses vectors for MIPS, a looping memcpy is generated
if the vectors are big enough (>32 bytes for MIPS32 or >64 bytes for MIPS64).
This takes place in mips_expand_block_move (gcc/config/mips.c).

In our case, a looping memcpy gets generated for a partition copy which is
inserted on an edge (in insert_partition_copy_on_edge, gcc/tree-outof-ssa.c).
This happens during PHI node elimination, which is done by eliminate_phi
(gcc/tree-outof-ssa.c), as a part of expand_phi_nodes, which is called by the
expand pass (gcc/cfgexpand.c).

My original fix was to update the CFG by calling find_many_sub_basic_blocks
with an all-one block bitmap (which also happens in cfgexpand.c, after edge
removal) whenever an edge contains an insn which satisfies control_flow_insn_p.
However, Segher Boessenkool said that this would be too risky for stage 4
and suggested inserting a NOP after the conditional jump in order to fool the
assert. This assumes that it is safe to delay the CFG update for this
particular case.

This patch changes mips_block_move_loop to emit a NOP after the conditional
jump, if the jump is the last insn of the loop. This prevents
"gcc_assert (!JUMP_P (last))" from triggering.

The NOP will never make it into the final assembly code because it is removed
during the cse1 pass through DCE for -O1, -O2, and -O3, and it's not even
emitted for -O0 and -Os because looping memcpy's are not generated for those
optimization levels, as can be seen in mips_expand_block_move from mips.c:

      if (INTVAL (length) <= MIPS_MAX_MOVE_BYTES_STRAIGHT)
	{
	  mips_block_move_straight (dest, src, INTVAL (length));
	  return true;
	}
      else if (optimize)
	{
	  mips_block_move_loop (dest, src, INTVAL (length),
				MIPS_MAX_MOVE_BYTES_PER_LOOP_ITER);
	  return true;
	}

Tested with mips-mti-elf.

Regards,
Toma Tabacu

gcc/

	* config/mips/mips.c (mips_block_move_loop): Emit a NOP after the
	conditional jump, if the jump is the last insn of the loop.

diff --git a/gcc/config/mips/mips.c b/gcc/config/mips/mips.c
index 4e13fbe..43e719f 100644
--- a/gcc/config/mips/mips.c
+++ b/gcc/config/mips/mips.c
@@ -8098,6 +8098,9 @@ mips_block_move_loop (rtx dest, rtx src, HOST_WIDE_INT length,
   /* Mop up any left-over bytes.  */
   if (leftover)
     mips_block_move_straight (dest, src, leftover);
+  else
+    /* Temporary fix for PR79150.  */
+    emit_insn (gen_nop ());
 }
 
 /* Expand a movmemsi instruction, which copies LENGTH bytes from