LoongArch: Use LSX and LASX for block move

gcc/ChangeLog:

	* config/loongarch/loongarch.h (LARCH_MAX_MOVE_PER_INSN):
	Define to the maximum amount of bytes able to be loaded or
	stored with one machine instruction.
	* config/loongarch/loongarch.cc (loongarch_mode_for_move_size):
	New static function.
	(loongarch_block_move_straight): Call
	loongarch_mode_for_move_size for machine_mode to be moved.
	(loongarch_expand_block_move): Use LARCH_MAX_MOVE_PER_INSN
	instead of UNITS_PER_WORD.
---

Bootstrapped and regtested on loongarch64-linux-gnu, with PR110939 patch
applied, the "lib_build_self_spec = %<..." line in t-linux commented out
(because it's silently making -mlasx in BOOT_CFLAGS ineffective, Yujie
is working on a proper fix), and BOOT_CFLAGS="-O3 -mlasx".  Ok for trunk?

 gcc/config/loongarch/loongarch.cc | 22 ++++++++++++++++++----
 gcc/config/loongarch/loongarch.h  |  3 +++
 2 files changed, 21 insertions(+), 4 deletions(-)

在 2023/9/8 上午12:14, Xi Ruoyao 写道:
> gcc/ChangeLog:
>
> 	* config/loongarch/loongarch.h (LARCH_MAX_MOVE_PER_INSN):
> 	Define to the maximum amount of bytes able to be loaded or
> 	stored with one machine instruction.
> 	* config/loongarch/loongarch.cc (loongarch_mode_for_move_size):
> 	New static function.
> 	(loongarch_block_move_straight): Call
> 	loongarch_mode_for_move_size for machine_mode to be moved.
> 	(loongarch_expand_block_move): Use LARCH_MAX_MOVE_PER_INSN
> 	instead of UNITS_PER_WORD.
> ---
>
> Bootstrapped and regtested on loongarch64-linux-gnu, with PR110939 patch
> applied, the "lib_build_self_spec = %<..." line in t-linux commented out
> (because it's silently making -mlasx in BOOT_CFLAGS ineffective, Yujie
> is working on a proper fix), and BOOT_CFLAGS="-O3 -mlasx".  Ok for trunk?

I think test cases need to be added here.

Otherwise OK, thanks!

>   gcc/config/loongarch/loongarch.cc | 22 ++++++++++++++++++----
>   gcc/config/loongarch/loongarch.h  |  3 +++
>   2 files changed, 21 insertions(+), 4 deletions(-)
>
> diff --git a/gcc/config/loongarch/loongarch.cc b/gcc/config/loongarch/loongarch.cc
> index 6698414281e..509ef2b97f1 100644
> --- a/gcc/config/loongarch/loongarch.cc
> +++ b/gcc/config/loongarch/loongarch.cc
> @@ -5191,6 +5191,20 @@ loongarch_function_ok_for_sibcall (tree decl ATTRIBUTE_UNUSED,
>     return true;
>   }
>   
> +static machine_mode
> +loongarch_mode_for_move_size (HOST_WIDE_INT size)
> +{
> +  switch (size)
> +    {
> +    case 32:
> +      return V32QImode;
> +    case 16:
> +      return V16QImode;
> +    }
> +
> +  return int_mode_for_size (size * BITS_PER_UNIT, 0).require ();
> +}
> +
>   /* Emit straight-line code to move LENGTH bytes from SRC to DEST.
>      Assume that the areas do not overlap.  */
>   
> @@ -5220,7 +5234,7 @@ loongarch_block_move_straight (rtx dest, rtx src, HOST_WIDE_INT length,
>   
>     for (delta_cur = delta, i = 0, offs = 0; offs < length; delta_cur /= 2)
>       {
> -      mode = int_mode_for_size (delta_cur * BITS_PER_UNIT, 0).require ();
> +      mode = loongarch_mode_for_move_size (delta_cur);
>   
>         for (; offs + delta_cur <= length; offs += delta_cur, i++)
>   	{
> @@ -5231,7 +5245,7 @@ loongarch_block_move_straight (rtx dest, rtx src, HOST_WIDE_INT length,
>   
>     for (delta_cur = delta, i = 0, offs = 0; offs < length; delta_cur /= 2)
>       {
> -      mode = int_mode_for_size (delta_cur * BITS_PER_UNIT, 0).require ();
> +      mode = loongarch_mode_for_move_size (delta_cur);
>   
>         for (; offs + delta_cur <= length; offs += delta_cur, i++)
>   	loongarch_emit_move (adjust_address (dest, mode, offs), regs[i]);
> @@ -5326,8 +5340,8 @@ loongarch_expand_block_move (rtx dest, rtx src, rtx r_length, rtx r_align)
>   
>     HOST_WIDE_INT align = INTVAL (r_align);
>   
> -  if (!TARGET_STRICT_ALIGN || align > UNITS_PER_WORD)
> -    align = UNITS_PER_WORD;
> +  if (!TARGET_STRICT_ALIGN || align > LARCH_MAX_MOVE_PER_INSN)
> +    align = LARCH_MAX_MOVE_PER_INSN;
>   
>     if (length <= align * LARCH_MAX_MOVE_OPS_STRAIGHT)
>       {
> diff --git a/gcc/config/loongarch/loongarch.h b/gcc/config/loongarch/loongarch.h
> index 3fc9dc43ab1..7e391205583 100644
> --- a/gcc/config/loongarch/loongarch.h
> +++ b/gcc/config/loongarch/loongarch.h
> @@ -1181,6 +1181,9 @@ typedef struct {
>      least twice.  */
>   #define LARCH_MAX_MOVE_OPS_STRAIGHT (LARCH_MAX_MOVE_OPS_PER_LOOP_ITER * 2)
>   
> +#define LARCH_MAX_MOVE_PER_INSN \
> +  (ISA_HAS_LASX ? 32 : (ISA_HAS_LSX ? 16 : UNITS_PER_WORD))
> +
>   /* The base cost of a memcpy call, for MOVE_RATIO and friends.  These
>      values were determined experimentally by benchmarking with CSiBE.
>   */

Hi,RuoYao:

  I think the test example memcpy-vec-3.c submitted in r14-3818 is 
implemented incorrectly.

The 16-byte length in this test example will cause can_move_by_pieces to 
return true when with '-mstrict-align', so no vector load instructions 
will be generated.

在 2023/9/8 上午12:14, Xi Ruoyao 写道:
> gcc/ChangeLog:
>
> 	* config/loongarch/loongarch.h (LARCH_MAX_MOVE_PER_INSN):
> 	Define to the maximum amount of bytes able to be loaded or
> 	stored with one machine instruction.
> 	* config/loongarch/loongarch.cc (loongarch_mode_for_move_size):
> 	New static function.
> 	(loongarch_block_move_straight): Call
> 	loongarch_mode_for_move_size for machine_mode to be moved.
> 	(loongarch_expand_block_move): Use LARCH_MAX_MOVE_PER_INSN
> 	instead of UNITS_PER_WORD.
> ---
>
> Bootstrapped and regtested on loongarch64-linux-gnu, with PR110939 patch
> applied, the "lib_build_self_spec = %<..." line in t-linux commented out
> (because it's silently making -mlasx in BOOT_CFLAGS ineffective, Yujie
> is working on a proper fix), and BOOT_CFLAGS="-O3 -mlasx".  Ok for trunk?
>
>   gcc/config/loongarch/loongarch.cc | 22 ++++++++++++++++++----
>   gcc/config/loongarch/loongarch.h  |  3 +++
>   2 files changed, 21 insertions(+), 4 deletions(-)
>
> diff --git a/gcc/config/loongarch/loongarch.cc b/gcc/config/loongarch/loongarch.cc
> index 6698414281e..509ef2b97f1 100644
> --- a/gcc/config/loongarch/loongarch.cc
> +++ b/gcc/config/loongarch/loongarch.cc
> @@ -5191,6 +5191,20 @@ loongarch_function_ok_for_sibcall (tree decl ATTRIBUTE_UNUSED,
>     return true;
>   }
>   
> +static machine_mode
> +loongarch_mode_for_move_size (HOST_WIDE_INT size)
> +{
> +  switch (size)
> +    {
> +    case 32:
> +      return V32QImode;
> +    case 16:
> +      return V16QImode;
> +    }
> +
> +  return int_mode_for_size (size * BITS_PER_UNIT, 0).require ();
> +}
> +
>   /* Emit straight-line code to move LENGTH bytes from SRC to DEST.
>      Assume that the areas do not overlap.  */
>   
> @@ -5220,7 +5234,7 @@ loongarch_block_move_straight (rtx dest, rtx src, HOST_WIDE_INT length,
>   
>     for (delta_cur = delta, i = 0, offs = 0; offs < length; delta_cur /= 2)
>       {
> -      mode = int_mode_for_size (delta_cur * BITS_PER_UNIT, 0).require ();
> +      mode = loongarch_mode_for_move_size (delta_cur);
>   
>         for (; offs + delta_cur <= length; offs += delta_cur, i++)
>   	{
> @@ -5231,7 +5245,7 @@ loongarch_block_move_straight (rtx dest, rtx src, HOST_WIDE_INT length,
>   
>     for (delta_cur = delta, i = 0, offs = 0; offs < length; delta_cur /= 2)
>       {
> -      mode = int_mode_for_size (delta_cur * BITS_PER_UNIT, 0).require ();
> +      mode = loongarch_mode_for_move_size (delta_cur);
>   
>         for (; offs + delta_cur <= length; offs += delta_cur, i++)
>   	loongarch_emit_move (adjust_address (dest, mode, offs), regs[i]);
> @@ -5326,8 +5340,8 @@ loongarch_expand_block_move (rtx dest, rtx src, rtx r_length, rtx r_align)
>   
>     HOST_WIDE_INT align = INTVAL (r_align);
>   
> -  if (!TARGET_STRICT_ALIGN || align > UNITS_PER_WORD)
> -    align = UNITS_PER_WORD;
> +  if (!TARGET_STRICT_ALIGN || align > LARCH_MAX_MOVE_PER_INSN)
> +    align = LARCH_MAX_MOVE_PER_INSN;
>   
>     if (length <= align * LARCH_MAX_MOVE_OPS_STRAIGHT)
>       {
> diff --git a/gcc/config/loongarch/loongarch.h b/gcc/config/loongarch/loongarch.h
> index 3fc9dc43ab1..7e391205583 100644
> --- a/gcc/config/loongarch/loongarch.h
> +++ b/gcc/config/loongarch/loongarch.h
> @@ -1181,6 +1181,9 @@ typedef struct {
>      least twice.  */
>   #define LARCH_MAX_MOVE_OPS_STRAIGHT (LARCH_MAX_MOVE_OPS_PER_LOOP_ITER * 2)
>   
> +#define LARCH_MAX_MOVE_PER_INSN \
> +  (ISA_HAS_LASX ? 32 : (ISA_HAS_LSX ? 16 : UNITS_PER_WORD))
> +
>   /* The base cost of a memcpy call, for MOVE_RATIO and friends.  These
>      values were determined experimentally by benchmarking with CSiBE.
>   */

On Sat, 2023-09-09 at 15:04 +0800, chenglulu wrote:
> Hi,RuoYao:
> 
>   I think the test example memcpy-vec-3.c submitted in r14-3818 is 
> implemented incorrectly.
> 
> The 16-byte length in this test example will cause can_move_by_pieces to 
> return true when with '-mstrict-align', so no vector load instructions
> will be generated.

Yes, in this case we cannot use vst because we don't know if b is
aligned.  Thus a { scan-assembler-not "vst" } guarantees that.

Or am I understanding something wrongly here?

在 2023/9/9 下午3:06, Xi Ruoyao 写道:
> On Sat, 2023-09-09 at 15:04 +0800, chenglulu wrote:
>> Hi,RuoYao:
>>
>>    I think the test example memcpy-vec-3.c submitted in r14-3818 is
>> implemented incorrectly.
>>
>> The 16-byte length in this test example will cause can_move_by_pieces to
>> return true when with '-mstrict-align', so no vector load instructions
>> will be generated.
> Yes, in this case we cannot use vst because we don't know if b is
> aligned.  Thus a { scan-assembler-not "vst" } guarantees that.
>
> Or am I understanding something wrongly here?
>
Well, what I mean is that even if '-mno-strict-align' is used here, 
vst/vld will not be used,

so this test example cannot test what we want to test.

On Sat, 2023-09-09 at 15:14 +0800, chenglulu wrote:
> 
> 在 2023/9/9 下午3:06, Xi Ruoyao 写道:
> > On Sat, 2023-09-09 at 15:04 +0800, chenglulu wrote:
> > > Hi,RuoYao:
> > > 
> > >    I think the test example memcpy-vec-3.c submitted in r14-3818 is
> > > implemented incorrectly.
> > > 
> > > The 16-byte length in this test example will cause can_move_by_pieces to
> > > return true when with '-mstrict-align', so no vector load instructions
> > > will be generated.
> > Yes, in this case we cannot use vst because we don't know if b is
> > aligned.  Thus a { scan-assembler-not "vst" } guarantees that.
> > 
> > Or am I understanding something wrongly here?
> > 
> Well, what I mean is that even if '-mno-strict-align' is used here, 
> vst/vld will not be used,
> 
> so this test example cannot test what we want to test.

Let me revise it...