LoongArch: Use LSX and LASX for block move
Checks
Commit Message
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(-)
Comments
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...
@@ -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)
{
@@ -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.
*/