[PATCH] veclower: Don't ICE on .VEC_CONVERT calls with no lhs [PR96426]

[PATCH] veclower: Don't ICE on .VEC_CONVERT calls with no lhs [PR96426]

[Jakub Jelinek]

Hi!

.VEC_CONVERT is a const internal call, so normally if the lhs is not used,
we'd DCE it far before getting to veclower, but with -O0 (or perhaps
-fno-tree-dce and some other -fno-* options) it can happen.
But as the internal fn needs the lhs to know the type to which the
conversion is done (and I think that is a reasonable representation, having
some magic another argument and having to create constants with that type
looks overkill to me), we just should DCE those calls ourselves.
During veclower, we can't really remove insns, as the callers would be
upset, so this just replaces it with a GIMPLE_NOP.

Bootstrapped/regtested on x86_64-linux, ok for trunk?

2020-08-04  Jakub Jelinek  <jakub@redhat.com>

	PR middle-end/96426
	* tree-vect-generic.c (expand_vector_conversion): Replace .VEC_CONVERT
	call with GIMPLE_NOP if there is no lhs.

	* gcc.c-torture/compile/pr96426.c: New test.

--- gcc/tree-vect-generic.c.jj	2020-07-28 15:39:10.081755224 +0200
+++ gcc/tree-vect-generic.c	2020-08-03 12:34:32.193423693 +0200
@@ -1775,6 +1775,12 @@ expand_vector_conversion (gimple_stmt_it
   gimple *stmt = gsi_stmt (*gsi);
   gimple *g;
   tree lhs = gimple_call_lhs (stmt);
+  if (lhs == NULL_TREE)
+    {
+      g = gimple_build_nop ();
+      gsi_replace (gsi, g, false);
+      return;
+    }
   tree arg = gimple_call_arg (stmt, 0);
   tree ret_type = TREE_TYPE (lhs);
   tree arg_type = TREE_TYPE (arg);
--- gcc/testsuite/gcc.c-torture/compile/pr96426.c.jj	2020-08-03 12:40:23.442449729 +0200
+++ gcc/testsuite/gcc.c-torture/compile/pr96426.c	2020-08-03 12:40:09.458647750 +0200
@@ -0,0 +1,10 @@
+/* PR middle-end/96426 */
+
+typedef long long V __attribute__((vector_size(16)));
+typedef double W __attribute__((vector_size(16)));
+
+void
+foo (V *v)
+{
+  __builtin_convertvector (*v, W);
+}

	Jakub

Re: [PATCH] veclower: Don't ICE on .VEC_CONVERT calls with no lhs [PR96426]

[Richard Sandiford]

Jakub Jelinek via Gcc-patches <gcc-patches@gcc.gnu.org> writes:
> Hi!
>
> .VEC_CONVERT is a const internal call, so normally if the lhs is not used,
> we'd DCE it far before getting to veclower, but with -O0 (or perhaps
> -fno-tree-dce and some other -fno-* options) it can happen.
> But as the internal fn needs the lhs to know the type to which the
> conversion is done (and I think that is a reasonable representation, having
> some magic another argument and having to create constants with that type
> looks overkill to me), we just should DCE those calls ourselves.

FWIW, the reason we took that approach for .WHILE_ULT was for things
like value numbering.  It seems reasonable to expect that one const call
to internal function F with arguments A produces the same result as
another const call to internal function F with arguments A, without
having to differentiate based on lhs type as well.

(Not a comment on the patch itself btw.)

Thanks,
Richard

Re: [PATCH] veclower: Don't ICE on .VEC_CONVERT calls with no lhs [PR96426]

[Richard Biener]

On Tue, 4 Aug 2020, Jakub Jelinek wrote:

> Hi!
> 
> .VEC_CONVERT is a const internal call, so normally if the lhs is not used,
> we'd DCE it far before getting to veclower, but with -O0 (or perhaps
> -fno-tree-dce and some other -fno-* options) it can happen.
> But as the internal fn needs the lhs to know the type to which the
> conversion is done (and I think that is a reasonable representation, having
> some magic another argument and having to create constants with that type
> looks overkill to me), we just should DCE those calls ourselves.
> During veclower, we can't really remove insns, as the callers would be
> upset, so this just replaces it with a GIMPLE_NOP.
> 
> Bootstrapped/regtested on x86_64-linux, ok for trunk?

OK.

Richard.

> 2020-08-04  Jakub Jelinek  <jakub@redhat.com>
> 
> 	PR middle-end/96426
> 	* tree-vect-generic.c (expand_vector_conversion): Replace .VEC_CONVERT
> 	call with GIMPLE_NOP if there is no lhs.
> 
> 	* gcc.c-torture/compile/pr96426.c: New test.
> 
> --- gcc/tree-vect-generic.c.jj	2020-07-28 15:39:10.081755224 +0200
> +++ gcc/tree-vect-generic.c	2020-08-03 12:34:32.193423693 +0200
> @@ -1775,6 +1775,12 @@ expand_vector_conversion (gimple_stmt_it
>    gimple *stmt = gsi_stmt (*gsi);
>    gimple *g;
>    tree lhs = gimple_call_lhs (stmt);
> +  if (lhs == NULL_TREE)
> +    {
> +      g = gimple_build_nop ();
> +      gsi_replace (gsi, g, false);
> +      return;
> +    }
>    tree arg = gimple_call_arg (stmt, 0);
>    tree ret_type = TREE_TYPE (lhs);
>    tree arg_type = TREE_TYPE (arg);
> --- gcc/testsuite/gcc.c-torture/compile/pr96426.c.jj	2020-08-03 12:40:23.442449729 +0200
> +++ gcc/testsuite/gcc.c-torture/compile/pr96426.c	2020-08-03 12:40:09.458647750 +0200
> @@ -0,0 +1,10 @@
> +/* PR middle-end/96426 */
> +
> +typedef long long V __attribute__((vector_size(16)));
> +typedef double W __attribute__((vector_size(16)));
> +
> +void
> +foo (V *v)
> +{
> +  __builtin_convertvector (*v, W);
> +}
> 
> 	Jakub
> 
> 

-- 
Richard Biener <rguenther@suse.de>
SUSE Software Solutions Germany GmbH, Maxfeldstrasse 5, 90409 Nuernberg,
Germany; GF: Felix Imendörffer; HRB 36809 (AG Nuernberg)