Fix PR 108874: aarch64 code regression with shift and ands

  After r6-2044-g98e30e515f184b, code like "((x & 0xff00ff00U) >> 8)"
would be optimized like (x >> 8) & 0xff00ffU which is normally better
except on aarch64, the shift right could be combined with another
operation in some cases. So we need to add a few define_splits
to the aarch64 backends that match "((x >> shift) & CST0) OP Y"
and splits it to:
TMP = X & CST1
(TMP >> shift) OP Y

Note this also gets us to matching rev16 back too so I added a
testcase to make sure we don't lose that matching any more.
Note when the generic patch to recognize those as bswap ROT 16,
we might regress again and need to add a few more patterns to
the aarch64 backend but will deal with that once that happens.

OK? Bootstrapped and tested on aarch64 with no regressions.

gcc/ChangeLog:

	* config/aarch64/aarch64.md: Add a new define_split
	to help combine.

gcc/testsuite/ChangeLog:

	* gcc.target/aarch64/rev16_2.c: New test.
	* gcc.target/aarch64/shift_and_operator-1.c: New test.
---
 gcc/config/aarch64/aarch64.md                 | 21 ++++++++++
 gcc/testsuite/gcc.target/aarch64/rev16_2.c    | 39 +++++++++++++++++++
 .../gcc.target/aarch64/shift_and_operator-1.c | 22 +++++++++++
 3 files changed, 82 insertions(+)
 create mode 100644 gcc/testsuite/gcc.target/aarch64/rev16_2.c
 create mode 100644 gcc/testsuite/gcc.target/aarch64/shift_and_operator-1.c
  

Andrew Pinski via Gcc-patches <gcc-patches@gcc.gnu.org> writes:
> After r6-2044-g98e30e515f184b, code like "((x & 0xff00ff00U) >> 8)"
> would be optimized like (x >> 8) & 0xff00ffU which is normally better
> except on aarch64, the shift right could be combined with another
> operation in some cases. So we need to add a few define_splits
> to the aarch64 backends that match "((x >> shift) & CST0) OP Y"
> and splits it to:
> TMP = X & CST1
> (TMP >> shift) OP Y
>
> Note this also gets us to matching rev16 back too so I added a
> testcase to make sure we don't lose that matching any more.
> Note when the generic patch to recognize those as bswap ROT 16,
> we might regress again and need to add a few more patterns to
> the aarch64 backend but will deal with that once that happens.
>
> OK? Bootstrapped and tested on aarch64 with no regressions.
>
> gcc/ChangeLog:
>
> 	* config/aarch64/aarch64.md: Add a new define_split
> 	to help combine.
>
> gcc/testsuite/ChangeLog:
>
> 	* gcc.target/aarch64/rev16_2.c: New test.
> 	* gcc.target/aarch64/shift_and_operator-1.c: New test.
> ---
>  gcc/config/aarch64/aarch64.md                 | 21 ++++++++++
>  gcc/testsuite/gcc.target/aarch64/rev16_2.c    | 39 +++++++++++++++++++
>  .../gcc.target/aarch64/shift_and_operator-1.c | 22 +++++++++++
>  3 files changed, 82 insertions(+)
>  create mode 100644 gcc/testsuite/gcc.target/aarch64/rev16_2.c
>  create mode 100644 gcc/testsuite/gcc.target/aarch64/shift_and_operator-1.c
>
> diff --git a/gcc/config/aarch64/aarch64.md b/gcc/config/aarch64/aarch64.md
> index af9087508ac..41cc563f10c 100644
> --- a/gcc/config/aarch64/aarch64.md
> +++ b/gcc/config/aarch64/aarch64.md
> @@ -4656,6 +4656,27 @@ (define_insn "*<LOGICAL:optab>_<SHIFT:optab><mode>3"
>    [(set_attr "type" "logic_shift_imm")]
>  )
>  
> +(define_split
> +  [(set (match_operand:GPI 0 "register_operand")
> +	(LOGICAL_OR_PLUS:GPI
> +	  (and:GPI
> +	    (lshiftrt:GPI (match_operand:GPI 1 "register_operand")
> +			  (match_operand:QI 2 "aarch64_shift_imm_<mode>"))
> +	    (match_operand:GPI 3 "aarch64_logical_immediate"))
> +	  (match_operand:GPI 4 "register_operand")))]
> +  "can_create_pseudo_p ()
> +   && aarch64_bitmask_imm (UINTVAL (operands[3]) << UINTVAL (operands[2]), <MODE>mode)"

Formatting nit: long line

> +  [(set (match_dup 5) (and:GPI (match_dup 1) (match_dup 6)))
> +   (set (match_dup 0) (match_dup 7))]
> +  {
> +    operands[5] = gen_reg_rtx (<MODE>mode);
> +    operands[6] = gen_int_mode (UINTVAL (operands[3]) << UINTVAL (operands[2]), <MODE>mode);

Here too.

> +    rtx shift = gen_rtx_LSHIFTRT (<MODE>mode, operands[5], operands[2]);
> +    rtx_code new_code = <CODE>;
> +    operands[7] = gen_rtx_fmt_ee (new_code, <MODE>mode, shift, operands[4]);

It should be possible to do the last three statements in the
rtl pattern, e.g. as:

  [(set (match_dup 5) (and:GPI (match_dup 1) (match_dup 6)))
   (set (match_dup 0) (LOGICAL_OR_PLUS:GPI
			(lshiftrt:GPI (match_dup 5) (match_dup 2))
			(match_dup 4)))]

OK with those change, thanks.

Richard

> +  }
> +)
> +
>  (define_split
>    [(set (match_operand:GPI 0 "register_operand")
>  	(LOGICAL_OR_PLUS:GPI
> diff --git a/gcc/testsuite/gcc.target/aarch64/rev16_2.c b/gcc/testsuite/gcc.target/aarch64/rev16_2.c
> new file mode 100644
> index 00000000000..621eb5dfbf0
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/aarch64/rev16_2.c
> @@ -0,0 +1,39 @@
> +/* { dg-options "-O2" } */
> +/* { dg-do compile } */
> +
> +extern void abort (void);
> +
> +typedef unsigned int __u32;
> +
> +__u32
> +__rev16_32_alt (__u32 x)
> +{
> +  return (((__u32)(x) & (__u32)0xff00ff00UL) >> 8)
> +         | (((__u32)(x) & (__u32)0x00ff00ffUL) << 8);
> +}
> +
> +__u32
> +__rev16_32 (__u32 x)
> +{
> +  return (((__u32)(x) & (__u32)0x00ff00ffUL) << 8)
> +         | (((__u32)(x) & (__u32)0xff00ff00UL) >> 8);
> +}
> +
> +typedef unsigned long long __u64;
> +
> +__u64
> +__rev16_64_alt (__u64 x)
> +{
> +  return (((__u64)(x) & (__u64)0xff00ff00ff00ff00UL) >> 8)
> +         | (((__u64)(x) & (__u64)0x00ff00ff00ff00ffUL) << 8);
> +}
> +
> +__u64
> +__rev16_64 (__u64 x)
> +{
> +  return (((__u64)(x) & (__u64)0x00ff00ff00ff00ffUL) << 8)
> +         | (((__u64)(x) & (__u64)0xff00ff00ff00ff00UL) >> 8);
> +}
> +
> +/* { dg-final { scan-assembler-times "rev16\\tx\[0-9\]+" 2 } } */
> +/* { dg-final { scan-assembler-times "rev16\\tw\[0-9\]+" 2 } } */
> diff --git a/gcc/testsuite/gcc.target/aarch64/shift_and_operator-1.c b/gcc/testsuite/gcc.target/aarch64/shift_and_operator-1.c
> new file mode 100644
> index 00000000000..49152c5495a
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/aarch64/shift_and_operator-1.c
> @@ -0,0 +1,22 @@
> +/* { dg-options "-O2" } */
> +/* { dg-do compile } */
> +
> +unsigned f(unsigned x, unsigned b)
> +{
> +  return ((x & 0xff00ff00U) >> 8) | b;
> +}
> +
> +unsigned f0(unsigned x, unsigned b)
> +{
> +  return ((x & 0xff00ff00U) >> 8) ^ b;
> +}
> +unsigned f1(unsigned x, unsigned b)
> +{
> +  return ((x & 0xff00ff00U) >> 8) + b;
> +}
> +
> +/* { dg-final { scan-assembler-times "lsr\\tw\[0-9\]+" 0 } } */
> +/* { dg-final { scan-assembler-times "lsr 8" 3 } } */
> +/* { dg-final { scan-assembler-times "eor\\tw\[0-9\]+" 1 } } */
> +/* { dg-final { scan-assembler-times "add\\tw\[0-9\]+" 1 } } */
> +/* { dg-final { scan-assembler-times "orr\\tw\[0-9\]+" 1 } } */