[PATCH] Fix PR80754

[PATCH] Fix PR80754

[Wilco Dijkstra]

When lra-remat rematerializes an instruction with a clobber, it checks
that the clobber does not kill live registers.  However it fails to check
that the clobber also doesn't overlap with the destination register of the 
final rematerialized instruction.  As a result it is possible to generate
illegal instructions with the same hard register as the destination and a
clobber.  Fix this by also checking for overlaps with the destination
register.

Bootstrap OK on arm-linux-gnueabihf for ARM and Thumb-2, OK for commit?

ChangeLog:
2017-05-16  Wilco Dijkstra  <wdijkstr@arm.com>

	PR rtl-optimization/80754
	* lra-remat.c (do_remat): Add overlap checks for dst_regno.

--
diff --git a/gcc/lra-remat.c b/gcc/lra-remat.c
index 2c51481374ac1f641ac163d345f2865c5536ef21..5cba7412d79a166f36424ada4e22c3acc9a571d5 100644
--- a/gcc/lra-remat.c
+++ b/gcc/lra-remat.c
@@ -1122,6 +1122,7 @@ do_remat (void)
 		  break;
 	    }
 	  int i, hard_regno, nregs;
+	  int dst_hard_regno, dst_nregs;
 	  rtx_insn *remat_insn = NULL;
 	  HOST_WIDE_INT cand_sp_offset = 0;
 	  if (cand != NULL)
@@ -1136,6 +1137,12 @@ do_remat (void)
 	      gcc_assert (REG_P (saved_op));
 	      int ignore_regno = REGNO (saved_op); 
 
+	      dst_hard_regno = dst_regno < FIRST_PSEUDO_REGISTER
+		? dst_regno : reg_renumber[dst_regno];
+	      gcc_assert (dst_hard_regno >= 0);
+	      machine_mode mode = GET_MODE (SET_DEST (set));
+	      dst_nregs = hard_regno_nregs[dst_hard_regno][mode];
+
 	      for (reg = cand_id->regs; reg != NULL; reg = reg->next)
 		if (reg->type != OP_IN && reg->regno != ignore_regno)
 		  {
@@ -1146,6 +1153,10 @@ do_remat (void)
 			break;
 		    if (i < nregs)
 		      break;
+		    /* Ensure the clobber also doesn't overlap dst_regno.  */
+		    if (hard_regno + nregs > dst_hard_regno
+			&& hard_regno < dst_hard_regno + dst_nregs)
+		      break;
 		  }
 
 	      if (reg == NULL)
@@ -1153,9 +1164,14 @@ do_remat (void)
 		  for (reg = static_cand_id->hard_regs;
 		       reg != NULL;
 		       reg = reg->next)
-		    if (reg->type != OP_IN
-			&& TEST_HARD_REG_BIT (live_hard_regs, reg->regno))
-		      break;
+		    if (reg->type != OP_IN)
+		      {
+			if (TEST_HARD_REG_BIT (live_hard_regs, reg->regno))
+			  break;
+			if (reg->regno >= dst_hard_regno
+			    && reg->regno < dst_hard_regno + dst_nregs)
+			  break;
+		      }
 		}
 
 	      if (reg == NULL)

Re: [PATCH] Fix PR80754

[Vladimir Makarov]

On 05/17/2017 07:32 AM, Wilco Dijkstra wrote:
> When lra-remat rematerializes an instruction with a clobber, it checks
> that the clobber does not kill live registers.  However it fails to check
> that the clobber also doesn't overlap with the destination register of the
> final rematerialized instruction.  As a result it is possible to generate
> illegal instructions with the same hard register as the destination and a
> clobber.  Fix this by also checking for overlaps with the destination
> register.
>
> Bootstrap OK on arm-linux-gnueabihf for ARM and Thumb-2, OK for commit?
>
>
Wilco, thank you for working on the PR.

The patch looks OK for me.  You can commit it into the trunk and gcc7 
branch.