RE: [Neon intrinsics] Literal vector construction through vcombine is poor

[Michael Collison <Michael dot Collison at arm dot com>]

Richard,

I reworked the patch using an assert as you suggested. Bootstrapped and retested. Okay for trunk?


-----Original Message-----
From: Richard Earnshaw (lists) [mailto:Richard.Earnshaw@arm.com] 
Sent: Friday, June 23, 2017 2:09 AM
To: Michael Collison <Michael.Collison@arm.com>; GCC Patches <gcc-patches@gcc.gnu.org>
Cc: nd <nd@arm.com>
Subject: Re: [Neon intrinsics] Literal vector construction through vcombine is poor

On 23/06/17 00:10, Michael Collison wrote:
> Richard,
> 
> I reworked the patch and retested on big endian as well as little. The original code was performing two swaps in the big endian case which works out to no swaps at all.
> 
> I also updated the ChangeLog per your comments. Okay for trunk?
> 
> 2017-06-19  Michael Collison  <michael.collison@arm.com>
> 
> 	* config/aarch64/aarch64-simd.md (aarch64_combine<mode>): Directly
> 	call aarch64_split_simd_combine.
> 	* (aarch64_combine_internal<mode>): Delete pattern.
> 	* config/aarch64/aarch64.c (aarch64_split_simd_combine):
> 	Allow register and subreg operands.
> 
> -----Original Message-----
> From: Richard Earnshaw (lists) [mailto:Richard.Earnshaw@arm.com]
> Sent: Monday, June 19, 2017 6:37 AM
> To: Michael Collison <Michael.Collison@arm.com>; GCC Patches 
> <gcc-patches@gcc.gnu.org>
> Cc: nd <nd@arm.com>
> Subject: Re: [Neon intrinsics] Literal vector construction through 
> vcombine is poor
> 
> On 16/06/17 22:08, Michael Collison wrote:
>> This patch improves code generation for literal vector construction by expanding and exposing the pattern to rtl optimization earlier. The current implementation delays splitting the pattern until after reload which results in poor code generation for the following code:
>>
>>
>> #include "arm_neon.h"
>>
>> int16x8_t
>> foo ()
>> {
>>   return vcombine_s16 (vdup_n_s16 (0), vdup_n_s16 (8)); }
>>
>> Trunk generates:
>>
>> foo:
>> 	movi	v1.2s, 0
>> 	movi	v0.4h, 0x8
>> 	dup	d2, v1.d[0]
>> 	ins	v2.d[1], v0.d[0]
>> 	orr	v0.16b, v2.16b, v2.16b
>> 	ret
>>
>> With the patch we now generate:
>>
>> foo:
>> 	movi	v1.4h, 0x8
>> 	movi	v0.4s, 0
>> 	ins	v0.d[1], v1.d[0]
>> 	ret
>>
>> Bootstrapped and tested on aarch64-linux-gnu. Okay for trunk.
>>
>> 2017-06-15  Michael Collison  <michael.collison@arm.com>
>>
>> 	* config/aarch64/aarch64-simd.md(aarch64_combine_internal<mode>):
>> 	Convert from define_insn_and_split into define_expand
>> 	* config/aarch64/aarch64.c(aarch64_split_simd_combine):
>> 	Allow register and subreg operands.
>>
> 
> Your changelog entry is confusing.  You've deleted the aarch64_combine_internal<mode> pattern entirely, having merged some of its functionality directly into its caller (aarch64_combine<mode>).
> 
> So I think it should read:
> 
> * config/aarch64/aarch64-simd.md (aarch64_combine<mode>): Directly call aarch64_split_simd_combine.
> (aarch64_combine_internal<mode>): Delete pattern.
> * ...
> 
> Note also there should be a space between the file name and the open bracket for the first function name.
> 
> Why don't you need the big-endian code path any more?
> 
> R.
> 
>>
>> pr7057.patch
>>
>>
>> diff --git a/gcc/config/aarch64/aarch64-simd.md
>> b/gcc/config/aarch64/aarch64-simd.md
>> index c462164..4a253a9 100644
>> --- a/gcc/config/aarch64/aarch64-simd.md
>> +++ b/gcc/config/aarch64/aarch64-simd.md
>> @@ -2807,27 +2807,11 @@
>>        op1 = operands[1];
>>        op2 = operands[2];
>>      }
>> -  emit_insn (gen_aarch64_combine_internal<mode> (operands[0], op1, 
>> op2));
>> -  DONE;
>> -}
>> -)
>>  
>> -(define_insn_and_split "aarch64_combine_internal<mode>"
>> -  [(set (match_operand:<VDBL> 0 "register_operand" "=&w")
>> -        (vec_concat:<VDBL> (match_operand:VDC 1 "register_operand" "w")
>> -			   (match_operand:VDC 2 "register_operand" "w")))]
>> -  "TARGET_SIMD"
>> -  "#"
>> -  "&& reload_completed"
>> -  [(const_int 0)]
>> -{
>> -  if (BYTES_BIG_ENDIAN)
>> -    aarch64_split_simd_combine (operands[0], operands[2], operands[1]);
>> -  else
>> -    aarch64_split_simd_combine (operands[0], operands[1], operands[2]);
>> +  aarch64_split_simd_combine (operands[0], op1, op2);
>> +
>>    DONE;
>>  }
>> -[(set_attr "type" "multiple")]
>>  )
>>  
>>  (define_expand "aarch64_simd_combine<mode>"
>> diff --git a/gcc/config/aarch64/aarch64.c 
>> b/gcc/config/aarch64/aarch64.c index 2e385c4..46bd78b 100644
>> --- a/gcc/config/aarch64/aarch64.c
>> +++ b/gcc/config/aarch64/aarch64.c
>> @@ -1650,7 +1650,8 @@ aarch64_split_simd_combine (rtx dst, rtx src1, 
>> rtx src2)
>>  
>>    gcc_assert (VECTOR_MODE_P (dst_mode));
>>  
>> -  if (REG_P (dst) && REG_P (src1) && REG_P (src2))
>> +  if (register_operand (dst, dst_mode) && register_operand (src1, src_mode)
>> +      && register_operand (src2, src_mode))
>>      {
>>        rtx (*gen) (rtx, rtx, rtx);
>>  
>>
> 
> 
> pr7057v4.patch
> 
> 
> diff --git a/gcc/config/aarch64/aarch64-simd.md 
> b/gcc/config/aarch64/aarch64-simd.md
> index c462164..3043f81 100644
> --- a/gcc/config/aarch64/aarch64-simd.md
> +++ b/gcc/config/aarch64/aarch64-simd.md
> @@ -2796,38 +2796,10 @@
>     (match_operand:VDC 2 "register_operand")]
>    "TARGET_SIMD"
>  {
> -  rtx op1, op2;
> -  if (BYTES_BIG_ENDIAN)
> -    {
> -      op1 = operands[2];
> -      op2 = operands[1];
> -    }
> -  else
> -    {
> -      op1 = operands[1];
> -      op2 = operands[2];
> -    }
> -  emit_insn (gen_aarch64_combine_internal<mode> (operands[0], op1, 
> op2));
> -  DONE;
> -}
> -)
> +  aarch64_split_simd_combine (operands[0], operands[1], operands[2]);
>  
> -(define_insn_and_split "aarch64_combine_internal<mode>"
> -  [(set (match_operand:<VDBL> 0 "register_operand" "=&w")
> -        (vec_concat:<VDBL> (match_operand:VDC 1 "register_operand" "w")
> -			   (match_operand:VDC 2 "register_operand" "w")))]
> -  "TARGET_SIMD"
> -  "#"
> -  "&& reload_completed"
> -  [(const_int 0)]
> -{
> -  if (BYTES_BIG_ENDIAN)
> -    aarch64_split_simd_combine (operands[0], operands[2], operands[1]);
> -  else
> -    aarch64_split_simd_combine (operands[0], operands[1], operands[2]);
>    DONE;
>  }
> -[(set_attr "type" "multiple")]
>  )
>  
>  (define_expand "aarch64_simd_combine<mode>"
> diff --git a/gcc/config/aarch64/aarch64.c 
> b/gcc/config/aarch64/aarch64.c index 2e385c4..46bd78b 100644
> --- a/gcc/config/aarch64/aarch64.c
> +++ b/gcc/config/aarch64/aarch64.c
> @@ -1650,7 +1650,8 @@ aarch64_split_simd_combine (rtx dst, rtx src1, 
> rtx src2)
>  
>    gcc_assert (VECTOR_MODE_P (dst_mode));
>  
> -  if (REG_P (dst) && REG_P (src1) && REG_P (src2))
> +  if (register_operand (dst, dst_mode) && register_operand (src1, src_mode)
> +      && register_operand (src2, src_mode))
>      {
>        rtx (*gen) (rtx, rtx, rtx);
>  
> 

As far as I can see aarch64_split_simd_combine is only called from this one expand pattern and the predicates on the pattern enforce all the operands being registers.  Furthermore, aarch64_split_simd_combine does nothing if they aren't all registers, which would obviously result in wrong code generation.

So convert

> +  if (register_operand (dst, dst_mode) && register_operand (src1,
src_mode)
> +      && register_operand (src2, src_mode))

into an assertion and make the following code unconditional.

OK with that change.

R.

Attachment: pr7057v5.patch
Description: pr7057v5.patch