[PATCH] Fix recent store-merging ICE (PR tree-optimization/82838)

[PATCH] Fix recent store-merging ICE (PR tree-optimization/82838)

[Jakub Jelinek]

Hi!

force_gimple_operand_1 clears the *seq first and then adds statements
there if any are needed.  So calling force_gimple_operand_1 twice on the
same seq is throwing away the earlier statements if any, rather than
appending new statements to those.

Fixed thusly, bootstrapped/regtested on x86_64-linux and i686-linux, ok for
trunk?

2017-11-06  Jakub Jelinek  <jakub@redhat.com>

	PR tree-optimization/82838
	* gimple-ssa-store-merging.c
	(imm_store_chain_info::output_merged_store): Call force_gimple_operand_1
	on a separate gimple_seq which is then appended to seq.

	* gcc.c-torture/compile/pr82838.c: New test.

--- gcc/gimple-ssa-store-merging.c.jj	2017-11-04 19:59:28.000000000 +0100
+++ gcc/gimple-ssa-store-merging.c	2017-11-04 22:59:07.410232483 +0100
@@ -1670,10 +1670,14 @@ imm_store_chain_info::output_merged_stor
       else if (operand_equal_p (base_addr, op.base_addr, 0))
 	load_addr[j] = addr;
       else
-	load_addr[j]
-	  = force_gimple_operand_1 (unshare_expr (op.base_addr),
-				    &seq, is_gimple_mem_ref_addr,
-				    NULL_TREE);
+	{
+	  gimple_seq this_seq;
+	  load_addr[j]
+	    = force_gimple_operand_1 (unshare_expr (op.base_addr),
+				      &this_seq, is_gimple_mem_ref_addr,
+				      NULL_TREE);
+	  gimple_seq_add_seq_without_update (&seq, this_seq);
+	}
     }
 
   FOR_EACH_VEC_ELT (split_stores, i, split_store)
--- gcc/testsuite/gcc.c-torture/compile/pr82838.c.jj	2017-11-04 23:02:09.522012666 +0100
+++ gcc/testsuite/gcc.c-torture/compile/pr82838.c	2017-11-04 23:01:54.000000000 +0100
@@ -0,0 +1,12 @@
+/* PR tree-optimization/82838 */
+
+struct S { unsigned short a, b, c; };
+struct S f[10];
+
+void
+foo (int e)
+{
+  struct S *x;
+  f[e].b = x[e].a;
+  f[e].c = x[e].b;
+}

	Jakub

Re: [PATCH] Fix recent store-merging ICE (PR tree-optimization/82838)

[Richard Biener]

On Mon, 6 Nov 2017, Jakub Jelinek wrote:

> Hi!
> 
> force_gimple_operand_1 clears the *seq first and then adds statements
> there if any are needed.  So calling force_gimple_operand_1 twice on the
> same seq is throwing away the earlier statements if any, rather than
> appending new statements to those.
> 
> Fixed thusly, bootstrapped/regtested on x86_64-linux and i686-linux, ok for
> trunk?

Ok.

Richard.

> 2017-11-06  Jakub Jelinek  <jakub@redhat.com>
> 
> 	PR tree-optimization/82838
> 	* gimple-ssa-store-merging.c
> 	(imm_store_chain_info::output_merged_store): Call force_gimple_operand_1
> 	on a separate gimple_seq which is then appended to seq.
> 
> 	* gcc.c-torture/compile/pr82838.c: New test.
> 
> --- gcc/gimple-ssa-store-merging.c.jj	2017-11-04 19:59:28.000000000 +0100
> +++ gcc/gimple-ssa-store-merging.c	2017-11-04 22:59:07.410232483 +0100
> @@ -1670,10 +1670,14 @@ imm_store_chain_info::output_merged_stor
>        else if (operand_equal_p (base_addr, op.base_addr, 0))
>  	load_addr[j] = addr;
>        else
> -	load_addr[j]
> -	  = force_gimple_operand_1 (unshare_expr (op.base_addr),
> -				    &seq, is_gimple_mem_ref_addr,
> -				    NULL_TREE);
> +	{
> +	  gimple_seq this_seq;
> +	  load_addr[j]
> +	    = force_gimple_operand_1 (unshare_expr (op.base_addr),
> +				      &this_seq, is_gimple_mem_ref_addr,
> +				      NULL_TREE);
> +	  gimple_seq_add_seq_without_update (&seq, this_seq);
> +	}
>      }
>  
>    FOR_EACH_VEC_ELT (split_stores, i, split_store)
> --- gcc/testsuite/gcc.c-torture/compile/pr82838.c.jj	2017-11-04 23:02:09.522012666 +0100
> +++ gcc/testsuite/gcc.c-torture/compile/pr82838.c	2017-11-04 23:01:54.000000000 +0100
> @@ -0,0 +1,12 @@
> +/* PR tree-optimization/82838 */
> +
> +struct S { unsigned short a, b, c; };
> +struct S f[10];
> +
> +void
> +foo (int e)
> +{
> +  struct S *x;
> +  f[e].b = x[e].a;
> +  f[e].c = x[e].b;
> +}
> 
> 	Jakub
> 
> 

-- 
Richard Biener <rguenther@suse.de>
SUSE LINUX GmbH, GF: Felix Imendoerffer, Jane Smithard, Graham Norton, HRB 21284 (AG Nuernberg)