[PATCH-3v2,rs6000] Enable 16-byte by pieces move [PR111449]
Checks
Commit Message
Hi,
The patch 2 enables 16-byte by pieces move on rs6000. This patch fixes
the regression cases caused by previous patch. For sra-17/18, the long
array with 4 elements can be loaded by one 16-byte by pieces move on 32-bit
platform. So the array is not be constructed in LC0 and SRA optimization
is unable to be taken. "no-vsx" option is added for 32-bit platform, as
it sets the MOVE_MAX_PIECES to 4-byte on 32-bit platform and the array
can't be loaded by one by pieces move.
Another regression is on P8 LE. The 16-byte memory to memory is
implemented by two TImode load/store. The TImode load/store is finally
split to two DImode load/store on P8 LE as it doesn't have unaligned
vector load/store instructions. Actually, 16-byte memory to memory move
can be implement by two V2DI reversed load/store on P8 LE. The patch
creates a insn_and_split pattern for this optimization.
Compared to previous version, it fixes the syntax errors in test cases.
Bootstrapped and tested on x86 and powerpc64-linux BE and LE with no
regressions. Is this OK for trunk?
Thanks
Gui Haochen
ChangeLog
rs6000: Enable 16-byte by pieces move
This patch enables 16-byte by pieces move. The 16-byte move is generated
with TImode and finally implemented by vector instructions. There are
several regression cases after the enablement. 16-byte TImode memory to
memory move is originally implemented by two pairs of DImode load/store on
P8 LE as there is no unaligned vsx load/store on it. The patch fixes
the problem by creating an insn_and_split pattern and converts it to one
pair of reversed load/store. Two SRA cases lost the SRA optimization as
the array can be loaded by one 16-byte move so that not be initialized in
LC0 on 32-bit platform. So fixes them by adding no-vsx option.
gcc/
PR target/111449
* config/rs6000/vsx.md (*vsx_le_mem_to_mem_mov_ti): New.
gcc/testsuite/
PR target/111449
* gcc.dg/tree-ssa/sra-17.c: Add no-vsx option for powerpc ilp32.
* gcc.dg/tree-ssa/sra-18.c: Likewise.
* gcc.target/powerpc/pr111449-1.c: New.
patch.diff
Comments
Hi,
on 2023/11/6 17:47, HAO CHEN GUI wrote:
> Hi,
> The patch 2 enables 16-byte by pieces move on rs6000. This patch fixes
> the regression cases caused by previous patch. For sra-17/18, the long
> array with 4 elements can be loaded by one 16-byte by pieces move on 32-bit
> platform. So the array is not be constructed in LC0 and SRA optimization
> is unable to be taken. "no-vsx" option is added for 32-bit platform, as
> it sets the MOVE_MAX_PIECES to 4-byte on 32-bit platform and the array
> can't be loaded by one by pieces move.
I'd expect the adjustments on these two affected sra-17/18 test cases are
moved to the previous related patch, as that patch caused them to fail and
your new define_insn_and_split doesn't actually affect them?
>
> Another regression is on P8 LE. The 16-byte memory to memory is
> implemented by two TImode load/store. The TImode load/store is finally
> split to two DImode load/store on P8 LE as it doesn't have unaligned
> vector load/store instructions. Actually, 16-byte memory to memory move
> can be implement by two V2DI reversed load/store on P8 LE. The patch
> creates a insn_and_split pattern for this optimization.
>
> Compared to previous version, it fixes the syntax errors in test cases.
>
> Bootstrapped and tested on x86 and powerpc64-linux BE and LE with no
> regressions. Is this OK for trunk?
>
> Thanks
> Gui Haochen
>
> ChangeLog
> rs6000: Enable 16-byte by pieces move
>
> This patch enables 16-byte by pieces move. The 16-byte move is generated
> with TImode and finally implemented by vector instructions. There are
> several regression cases after the enablement. 16-byte TImode memory to
> memory move is originally implemented by two pairs of DImode load/store on
> P8 LE as there is no unaligned vsx load/store on it. The patch fixes
> the problem by creating an insn_and_split pattern and converts it to one
> pair of reversed load/store. Two SRA cases lost the SRA optimization as
> the array can be loaded by one 16-byte move so that not be initialized in
> LC0 on 32-bit platform. So fixes them by adding no-vsx option.
>
> gcc/
> PR target/111449
> * config/rs6000/vsx.md (*vsx_le_mem_to_mem_mov_ti): New.
>
> gcc/testsuite/
> PR target/111449
> * gcc.dg/tree-ssa/sra-17.c: Add no-vsx option for powerpc ilp32.
> * gcc.dg/tree-ssa/sra-18.c: Likewise.
> * gcc.target/powerpc/pr111449-1.c: New.
Wrong test case name.
>
>
> patch.diff
> diff --git a/gcc/config/rs6000/vsx.md b/gcc/config/rs6000/vsx.md
> index f3b40229094..9f6bc49998a 100644
> --- a/gcc/config/rs6000/vsx.md
> +++ b/gcc/config/rs6000/vsx.md
> @@ -414,6 +414,27 @@ (define_mode_attr VM3_char [(V2DI "d")
>
> ;; VSX moves
>
> +;; TImode memory to memory move optimization on LE with p8vector
> +(define_insn_and_split "*vsx_le_mem_to_mem_mov_ti"
> + [(set (match_operand:TI 0 "indexed_or_indirect_operand" "=Z")
> + (match_operand:TI 1 "indexed_or_indirect_operand" "Z"))]
> + "!BYTES_BIG_ENDIAN && TARGET_VSX && !TARGET_P9_VECTOR
Nit: Make each one in its separated line since the below ones are
separated.
> + && !MEM_VOLATILE_P (operands[0])
> + && !MEM_VOLATILE_P (operands[1])
> + && !reload_completed"
!reload_completed is for required pseudo? It would be better with
can_create_pseudo_p ().
BR,
Kewen
> + "#"
> + "&& 1"
> + [(const_int 0)]
> +{
> + rtx tmp = gen_reg_rtx (V2DImode);
> + rtx src = adjust_address (operands[1], V2DImode, 0);
> + emit_insn (gen_vsx_ld_elemrev_v2di (tmp, src));
> + rtx dest = adjust_address (operands[0], V2DImode, 0);
> + emit_insn (gen_vsx_st_elemrev_v2di (dest, tmp));
> + DONE;
> +}
> + [(set_attr "length" "16")])
> +
> ;; The patterns for LE permuted loads and stores come before the general
> ;; VSX moves so they match first.
> (define_insn_and_split "*vsx_le_perm_load_<mode>"
> diff --git a/gcc/testsuite/gcc.dg/tree-ssa/sra-17.c b/gcc/testsuite/gcc.dg/tree-ssa/sra-17.c
> index 221d96b6cd9..b0d4811e77b 100644
> --- a/gcc/testsuite/gcc.dg/tree-ssa/sra-17.c
> +++ b/gcc/testsuite/gcc.dg/tree-ssa/sra-17.c
> @@ -1,6 +1,7 @@
> /* { dg-do run { target { aarch64*-*-* alpha*-*-* arm*-*-* hppa*-*-* powerpc*-*-* s390*-*-* } } } */
> /* { dg-options "-O2 -fdump-tree-esra --param sra-max-scalarization-size-Ospeed=32" } */
> /* { dg-additional-options "-mcpu=ev4" { target alpha*-*-* } } */
> +/* { dg-additional-options "-mno-vsx" { target { powerpc*-*-* && ilp32 } } } */
>
> extern void abort (void);
>
> diff --git a/gcc/testsuite/gcc.dg/tree-ssa/sra-18.c b/gcc/testsuite/gcc.dg/tree-ssa/sra-18.c
> index f5e6a21c2ae..2cdeae6e9e7 100644
> --- a/gcc/testsuite/gcc.dg/tree-ssa/sra-18.c
> +++ b/gcc/testsuite/gcc.dg/tree-ssa/sra-18.c
> @@ -1,6 +1,7 @@
> /* { dg-do run { target { aarch64*-*-* alpha*-*-* arm*-*-* hppa*-*-* powerpc*-*-* s390*-*-* } } } */
> /* { dg-options "-O2 -fdump-tree-esra --param sra-max-scalarization-size-Ospeed=32" } */
> /* { dg-additional-options "-mcpu=ev4" { target alpha*-*-* } } */
> +/* { dg-additional-options "-mno-vsx" { target { powerpc*-*-* && ilp32 } } } */
>
> extern void abort (void);
> struct foo { long x; };
> diff --git a/gcc/testsuite/gcc.target/powerpc/pr111449-2.c b/gcc/testsuite/gcc.target/powerpc/pr111449-2.c
> new file mode 100644
> index 00000000000..7003bdc0208
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/powerpc/pr111449-2.c
> @@ -0,0 +1,18 @@
> +/* { dg-do compile { target { has_arch_pwr8 } } } */
> +/* { dg-require-effective-target powerpc_p8vector_ok } */
> +/* { dg-options "-mvsx -O2" } */
> +
> +/* Ensure 16-byte by pieces move is enabled. */
> +
> +void move1 (void *s1, void *s2)
> +{
> + __builtin_memcpy (s1, s2, 16);
> +}
> +
> +void move2 (void *s1)
> +{
> + __builtin_memcpy (s1, "0123456789012345", 16);
> +}
> +
> +/* { dg-final { scan-assembler-times {\mlxvd2x\M|\mp?lxv\M} 2 } } */
> +/* { dg-final { scan-assembler-times {\mstxvd2x\M|\mstxv\M} 2 } } */
@@ -414,6 +414,27 @@ (define_mode_attr VM3_char [(V2DI "d")
;; VSX moves
+;; TImode memory to memory move optimization on LE with p8vector
+(define_insn_and_split "*vsx_le_mem_to_mem_mov_ti"
+ [(set (match_operand:TI 0 "indexed_or_indirect_operand" "=Z")
+ (match_operand:TI 1 "indexed_or_indirect_operand" "Z"))]
+ "!BYTES_BIG_ENDIAN && TARGET_VSX && !TARGET_P9_VECTOR
+ && !MEM_VOLATILE_P (operands[0])
+ && !MEM_VOLATILE_P (operands[1])
+ && !reload_completed"
+ "#"
+ "&& 1"
+ [(const_int 0)]
+{
+ rtx tmp = gen_reg_rtx (V2DImode);
+ rtx src = adjust_address (operands[1], V2DImode, 0);
+ emit_insn (gen_vsx_ld_elemrev_v2di (tmp, src));
+ rtx dest = adjust_address (operands[0], V2DImode, 0);
+ emit_insn (gen_vsx_st_elemrev_v2di (dest, tmp));
+ DONE;
+}
+ [(set_attr "length" "16")])
+
;; The patterns for LE permuted loads and stores come before the general
;; VSX moves so they match first.
(define_insn_and_split "*vsx_le_perm_load_<mode>"
@@ -1,6 +1,7 @@
/* { dg-do run { target { aarch64*-*-* alpha*-*-* arm*-*-* hppa*-*-* powerpc*-*-* s390*-*-* } } } */
/* { dg-options "-O2 -fdump-tree-esra --param sra-max-scalarization-size-Ospeed=32" } */
/* { dg-additional-options "-mcpu=ev4" { target alpha*-*-* } } */
+/* { dg-additional-options "-mno-vsx" { target { powerpc*-*-* && ilp32 } } } */
extern void abort (void);
@@ -1,6 +1,7 @@
/* { dg-do run { target { aarch64*-*-* alpha*-*-* arm*-*-* hppa*-*-* powerpc*-*-* s390*-*-* } } } */
/* { dg-options "-O2 -fdump-tree-esra --param sra-max-scalarization-size-Ospeed=32" } */
/* { dg-additional-options "-mcpu=ev4" { target alpha*-*-* } } */
+/* { dg-additional-options "-mno-vsx" { target { powerpc*-*-* && ilp32 } } } */
extern void abort (void);
struct foo { long x; };
new file mode 100644
@@ -0,0 +1,18 @@
+/* { dg-do compile { target { has_arch_pwr8 } } } */
+/* { dg-require-effective-target powerpc_p8vector_ok } */
+/* { dg-options "-mvsx -O2" } */
+
+/* Ensure 16-byte by pieces move is enabled. */
+
+void move1 (void *s1, void *s2)
+{
+ __builtin_memcpy (s1, s2, 16);
+}
+
+void move2 (void *s1)
+{
+ __builtin_memcpy (s1, "0123456789012345", 16);
+}
+
+/* { dg-final { scan-assembler-times {\mlxvd2x\M|\mp?lxv\M} 2 } } */
+/* { dg-final { scan-assembler-times {\mstxvd2x\M|\mstxv\M} 2 } } */