Jakub Jelinek <ja...@redhat.com> writes:
> Hi!
>
> Before combiner added 2 to 2 combinations, the following testcase functions
> have been all compiled into 2 instructions, zero/sign extensions or and
> followed by orr with lsl, e.g. for the first function
> Trying 7 -> 8:
> 7: r96:SI=r94:SI<<0xb
> 8: r95:SI=r96:SI|r94:SI
> REG_DEAD r96:SI
> REG_DEAD r94:SI
> Successfully matched this instruction:
> (set (reg:SI 95)
> (ior:SI (ashift:SI (reg/v:SI 94 [ i ])
> (const_int 11 [0xb]))
> (reg/v:SI 94 [ i ])))
> is the important successful try_combine and so we end up with
> and w0, w0, 255
> orr w0, w0, w0, lsl 11
> in the body.
> With 2 to 2 combination, before that can trigger, another successful
> combination:
> Trying 2 -> 7:
> 2: r94:SI=zero_extend(x0:QI)
> REG_DEAD x0:QI
> 7: r96:SI=r94:SI<<0xb
> is replaced with:
> (set (reg/v:SI 94 [ i ])
> (zero_extend:SI (reg:QI 0 x0 [ i ])))
> and
> (set (reg:SI 96)
> (and:SI (ashift:SI (reg:SI 0 x0 [ i ])
> (const_int 11 [0xb]))
> (const_int 522240 [0x7f800])))
> and in the end results in 3 instructions in the body:
> and w1, w0, 255
> ubfiz w0, w0, 11, 8
> orr w0, w0, w1
> The following combine splitters help undo that when combiner tries to
> combine 3 instructions - the zero/sign extend or and, the other insn
> from the 2 to 2 combination ([us]bfiz) and the logical op, the CPUs
> don't have an insn to do everything in one op, but we can split it
> back into the zero/sign extend or and followed by logical with lsl.
>
> Bootstrapped/regtested on aarch64-linux, ok for trunk?
>
> 2021-04-14 Jakub Jelinek <ja...@redhat.com>
>
> PR target/100056
> * config/aarch64/aarch64.md (*<LOGICAL:optab>_<SHIFT:optab><mode>3):
> Add combine splitters for *<LOGICAL:optab>_ashl<mode>3 with
> ZERO_EXTEND, SIGN_EXTEND or AND.
>
> * gcc.target/aarch64/pr100056.c: New test.
>
> --- gcc/config/aarch64/aarch64.md.jj 2021-04-13 12:40:57.000000000 +0200
> +++ gcc/config/aarch64/aarch64.md 2021-04-13 19:54:17.015764651 +0200
> @@ -4431,6 +4431,59 @@ (define_insn "*<LOGICAL:optab>_<SHIFT:op
> [(set_attr "type" "logic_shift_imm")]
> )
>
> +(define_split
> + [(set (match_operand:GPI 0 "register_operand")
> + (LOGICAL:GPI
> + (and:GPI (ashift:GPI (match_operand:GPI 1 "register_operand")
> + (match_operand:QI 2 "aarch64_shift_imm_<mode>"))
> + (match_operand:GPI 4 "const_int_operand"))
> + (zero_extend:GPI (match_operand 3 "register_operand"))))]
> + "can_create_pseudo_p ()
> + && REG_P (operands[1])
> + && REG_P (operands[3])
> + && REGNO (operands[1]) == REGNO (operands[3])
> + && ((unsigned HOST_WIDE_INT)
> + trunc_int_for_mode (GET_MODE_MASK (GET_MODE (operands[3]))
> + << INTVAL (operands[2]), <MODE>mode)
> + == UINTVAL (operands[4]))"
IMO this would be easier to understand as:
&& (trunc_int_for_mode (GET_MODE_MASK (GET_MODE (operands[3]))
<< INTVAL (operands[2]), <MODE>mode)
== INTVAL (operands[4]))
(At first I thought the cast and UINTVAL were trying to escape the
sign-extension canonicalisation.)
I'm not sure about this one though. The REGNO checks mean that this is
effectively for hard registers only. I thought one of the reasons for
make_more_copies was to avoid combining hard registers like this, so I'm
not sure we should have a pattern that specifically targets them.
Segher, have I misunderstood?
> + [(set (match_dup 4) (zero_extend:GPI (match_dup 3)))
> + (set (match_dup 0) (LOGICAL:GPI (ashift:GPI (match_dup 4) (match_dup 2))
> + (match_dup 4)))]
> + "operands[4] = gen_reg_rtx (<MODE>mode);"
> +)
> +
> +(define_split
> + [(set (match_operand:GPI 0 "register_operand")
> + (LOGICAL:GPI
> + (and:GPI (ashift:GPI (match_operand:GPI 1 "register_operand")
> + (match_operand:QI 2 "aarch64_shift_imm_<mode>"))
> + (match_operand:GPI 4 "const_int_operand"))
> + (and:GPI (match_dup 1) (match_operand:GPI 3 "const_int_operand"))))]
> + "can_create_pseudo_p ()
> + && pow2_or_zerop (UINTVAL (operands[3]) + 1)
> + && ((unsigned HOST_WIDE_INT)
> + trunc_int_for_mode (UINTVAL (operands[3])
> + << INTVAL (operands[2]), <MODE>mode)
> + == UINTVAL (operands[4]))"
> + [(set (match_dup 4) (and:GPI (match_dup 1) (match_dup 3)))
> + (set (match_dup 0) (LOGICAL:GPI (ashift:GPI (match_dup 4) (match_dup 2))
> + (match_dup 4)))]
> + "operands[4] = gen_reg_rtx (<MODE>mode);"
> +)
> +
> +(define_split
> + [(set (match_operand:GPI 0 "register_operand")
> + (LOGICAL:GPI
> + (ashift:GPI (sign_extend:GPI (match_operand 1 "register_operand"))
> + (match_operand:QI 2 "aarch64_shift_imm_<mode>"))
> + (sign_extend:GPI (match_dup 1))))]
> + "can_create_pseudo_p ()"
> + [(set (match_dup 4) (sign_extend:GPI (match_dup 1)))
> + (set (match_dup 0) (LOGICAL:GPI (ashift:GPI (match_dup 4) (match_dup 2))
> + (match_dup 4)))]
> + "operands[4] = gen_reg_rtx (<MODE>mode);"
> +)
These two look good to me apart from the cast nit. The last one feels
like it's more general than just sign_extends though. I guess it would
work for any duplicated operation that can be performed in a single
instruction.
Thanks,
Richard
> +
> (define_insn "*<optab>_rol<mode>3"
> [(set (match_operand:GPI 0 "register_operand" "=r")
> (LOGICAL:GPI (rotate:GPI
> --- gcc/testsuite/gcc.target/aarch64/pr100056.c.jj 2021-04-13
> 14:20:53.334784184 +0200
> +++ gcc/testsuite/gcc.target/aarch64/pr100056.c 2021-04-13
> 19:44:09.358529648 +0200
> @@ -0,0 +1,50 @@
> +/* PR target/100056 */
> +/* { dg-do compile } */
> +/* { dg-options "-O2" } */
> +/* { dg-final { scan-assembler-not {\t[us]bfiz\tw[0-9]+, w[0-9]+, 11} } } */
> +
> +int
> +or_shift_u8 (unsigned char i)
> +{
> + return i | (i << 11);
> +}
> +
> +int
> +or_shift_u3a (unsigned i)
> +{
> + i &= 7;
> + return i | (i << 11);
> +}
> +
> +int
> +or_shift_u3b (unsigned i)
> +{
> + i = (i << 29) >> 29;
> + return i | (i << 11);
> +}
> +
> +int
> +or_shift_s16 (signed short i)
> +{
> + return i | (i << 11);
> +}
> +
> +int
> +or_shift_s8 (signed char i)
> +{
> + return i | (i << 11);
> +}
> +
> +int
> +or_shift_s13 (int i)
> +{
> + i = (i << 19) >> 19;
> + return i | (i << 11);
> +}
> +
> +int
> +or_shift_s3 (int i)
> +{
> + i = (i << 29) >> 29;
> + return i | (i << 11);
> +}
>
> Jakub
