diff mbox series

vect: Verify that GET_MODE_NUNITS is greater than one.

Message ID 20230314215256.4153026-1-collison@rivosinc.com
State New
Headers
Series vect: Verify that GET_MODE_NUNITS is greater than one. |

Commit Message

Michael Collison March 14, 2023, 9:52 p.m. UTC 
  While working on autovectorizing for the RISCV port I encountered an issue
where can_duplicate_and_interleave_p assumes that GET_MODE_NUNITS is a
evenly divisible by two. The RISC-V target has vector modes (e.g. VNx1DImode),
where GET_MODE_NUNITS is equal to one.

Tested on RISCV and x86_64-linux-gnu. Okay?

2023-03-09  Michael Collison  <collison@rivosinc.com>

	* tree-vect-slp.cc (can_duplicate_and_interleave_p):
	Check that GET_MODE_NUNITS is greater than one.
---
 gcc/tree-vect-slp.cc | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

Comments

Jeff Law March 19, 2023, 4:58 p.m. UTC | #1 
On 3/14/23 15:52, Michael Collison wrote:
> While working on autovectorizing for the RISCV port I encountered an issue
> where can_duplicate_and_interleave_p assumes that GET_MODE_NUNITS is a
> evenly divisible by two. The RISC-V target has vector modes (e.g. VNx1DImode),
> where GET_MODE_NUNITS is equal to one.
> 
> Tested on RISCV and x86_64-linux-gnu. Okay?
> 
> 2023-03-09  Michael Collison  <collison@rivosinc.com>
> 
> 	* tree-vect-slp.cc (can_duplicate_and_interleave_p):
> 	Check that GET_MODE_NUNITS is greater than one.
As far as I know this doesn't fix a regression so I would defer to 
gc-14.  As release managers, Richi, Jakub or Joseph can gate it in as an 
exception.

jeff
Richard Sandiford March 31, 2023, 9:58 a.m. UTC | #2 
Michael Collison <collison@rivosinc.com> writes:
> While working on autovectorizing for the RISCV port I encountered an issue
> where can_duplicate_and_interleave_p assumes that GET_MODE_NUNITS is a
> evenly divisible by two. The RISC-V target has vector modes (e.g. VNx1DImode),
> where GET_MODE_NUNITS is equal to one.
>
> Tested on RISCV and x86_64-linux-gnu. Okay?
>
> 2023-03-09  Michael Collison  <collison@rivosinc.com>
>
> 	* tree-vect-slp.cc (can_duplicate_and_interleave_p):
> 	Check that GET_MODE_NUNITS is greater than one.
> ---
>  gcc/tree-vect-slp.cc | 3 ++-
>  1 file changed, 2 insertions(+), 1 deletion(-)
>
> diff --git a/gcc/tree-vect-slp.cc b/gcc/tree-vect-slp.cc
> index 9a4e000925e..add58113fa8 100644
> --- a/gcc/tree-vect-slp.cc
> +++ b/gcc/tree-vect-slp.cc
> @@ -426,7 +426,8 @@ can_duplicate_and_interleave_p (vec_info *vinfo, unsigned int count,
>  	  if (vector_type
>  	      && VECTOR_MODE_P (TYPE_MODE (vector_type))
>  	      && known_eq (GET_MODE_SIZE (TYPE_MODE (vector_type)),
> -			   GET_MODE_SIZE (base_vector_mode)))
> +			   GET_MODE_SIZE (base_vector_mode))
> +	      && known_gt (GET_MODE_NUNITS (TYPE_MODE (vector_type)), 1))
>  	    {
>  	      /* Try fusing consecutive sequences of COUNT / NVECTORS elements
>  		 together into elements of type INT_TYPE and using the result

FWIW, I think it'd better to remove:

	      poly_int64 half_nelts = exact_div (nelts, 2);

declare:

	  poly_uint64 half_nelts;

before the if condition, and use:

	    && multiple_p (GET_MODE_NUNITS (TYPE_MODE (vector_type)),
			   2, &half_nelts)

instead of the known_gt.  In other words, now that we can't assert
the exact_div, we should check it (using multiple_p) instead.

Thanks,
Richard
Jeff Law April 22, 2023, 12:23 a.m. UTC | #3 
On 3/14/23 15:52, Michael Collison wrote:
> While working on autovectorizing for the RISCV port I encountered an issue
> where can_duplicate_and_interleave_p assumes that GET_MODE_NUNITS is a
> evenly divisible by two. The RISC-V target has vector modes (e.g. VNx1DImode),
> where GET_MODE_NUNITS is equal to one.
> 
> Tested on RISCV and x86_64-linux-gnu. Okay?
> 
> 2023-03-09  Michael Collison  <collison@rivosinc.com>
> 
> 	* tree-vect-slp.cc (can_duplicate_and_interleave_p):
> 	Check that GET_MODE_NUNITS is greater than one.
Is this still relevant?   I know other changes were made to deal with 
the case where GET_MODE_NUNITS returns 1, but I don't know if they made 
this obsolete.

Any chance we could get a testcase for this?  I realize it might depend 
on unmerged RVV bits.


jeff
diff mbox series

Patch

diff --git a/gcc/tree-vect-slp.cc b/gcc/tree-vect-slp.cc
index 9a4e000925e..add58113fa8 100644
--- a/gcc/tree-vect-slp.cc
+++ b/gcc/tree-vect-slp.cc
@@ -426,7 +426,8 @@  can_duplicate_and_interleave_p (vec_info *vinfo, unsigned int count,
 	  if (vector_type
 	      && VECTOR_MODE_P (TYPE_MODE (vector_type))
 	      && known_eq (GET_MODE_SIZE (TYPE_MODE (vector_type)),
-			   GET_MODE_SIZE (base_vector_mode)))
+			   GET_MODE_SIZE (base_vector_mode))
+	      && known_gt (GET_MODE_NUNITS (TYPE_MODE (vector_type)), 1))
 	    {
 	      /* Try fusing consecutive sequences of COUNT / NVECTORS elements
 		 together into elements of type INT_TYPE and using the result