Hi Haochen,
on 2022/9/21 13:13, HAO CHEN GUI wrote:
> Hi,
> This patch adds a new insn for vector splat with small V2DI constants on P8.
> If the value of constant is in RANGE (-16, 15) and not 0 or -1, it can be
> loaded
> with vspltisw and vupkhsw on P8. It should be efficient than loading vector
> from
> TOC.
Nice.
>
> Bootstrapped and tested on powerpc64-linux BE and LE with no regressions.
> Is this okay for trunk? Any recommendations? Thanks a lot.
>
> ChangeLog
> 2022-09-21 Haochen Gui <guih...@linux.ibm.com>
>
> gcc/
> PR target/104124
> * config/rs6000/altivec.md (*altivec_vupkhs<VU_char>_direct): Renamed
s/Renamed/Rename/
> to...
> (altivec_vupkhs<VU_char>_direct): ...this.
> * config/rs6000/constraints.md (wT constraint): New constant for a
> vector constraint that can be loaded with vspltisw and vupkhsw.
> * config/rs6000/predicates.md (vspltisw_constant_split): New
> predicate for wT constraint.
> * config/rs6000/rs6000-protos.h (vspltisw_constant_p): Add declaration.
> * config/rs6000/rs6000.cc (easy_altivec_constant): Call
> vspltisw_constant_p to judge if a V2DI constant can be synthesized with
> a vspltisw and a vupkhsw.
> * (vspltisw_constant_p): New function to return true if OP mode is
> V2DI and can be synthesized with ISA 2.07 instruction vupkhsw and
> vspltisw.
> * gcc/config/rs6000/vsx.md (*vspltisw_v2di_split): New insn to load up
> constants with vspltisw and vupkhsw.
>
> gcc/testsuite/
> PR target/104124
> * gcc.target/powerpc/p8-splat.c: New.
>
> patch.diff
> diff --git a/gcc/config/rs6000/altivec.md b/gcc/config/rs6000/altivec.md
> index 2c4940f2e21..185414df021 100644
> --- a/gcc/config/rs6000/altivec.md
> +++ b/gcc/config/rs6000/altivec.md
> @@ -2542,7 +2542,7 @@ (define_insn "altivec_vupkhs<VU_char>"
> }
> [(set_attr "type" "vecperm")])
>
> -(define_insn "*altivec_vupkhs<VU_char>_direct"
> +(define_insn "altivec_vupkhs<VU_char>_direct"
> [(set (match_operand:VP 0 "register_operand" "=v")
> (unspec:VP [(match_operand:<VP_small> 1 "register_operand" "v")]
> UNSPEC_VUNPACK_HI_SIGN_DIRECT))]
> diff --git a/gcc/config/rs6000/constraints.md
> b/gcc/config/rs6000/constraints.md
> index 5a44a92142e..f65dea6e0c7 100644
> --- a/gcc/config/rs6000/constraints.md
> +++ b/gcc/config/rs6000/constraints.md
> @@ -150,6 +150,10 @@ (define_constraint "wS"
> "@internal Vector constant that can be loaded with XXSPLTIB & sign
> extension."
> (match_test "xxspltib_constant_split (op, mode)"))
>
> +(define_constraint "wT"
> + "@internal Vector constant that can be loaded with vspltisw & vupkhsw."
> + (match_test "vspltisw_constant_split (op, mode)"))
> +
> ;; ISA 3.0 DS-form instruction that has the bottom 2 bits 0 and no update
> form.
> ;; Used by LXSD/STXSD/LXSSP/STXSSP. In contrast to "Y", the multiple-of-four
> ;; offset is enforced for 32-bit too.
> diff --git a/gcc/config/rs6000/predicates.md b/gcc/config/rs6000/predicates.md
> index b1fcc69bb60..00cf60bbe58 100644
> --- a/gcc/config/rs6000/predicates.md
> +++ b/gcc/config/rs6000/predicates.md
> @@ -694,6 +694,19 @@ (define_predicate "xxspltib_constant_split"
> return num_insns > 1;
> })
>
> +;; Return true if the operand is a constant that can be loaded with a
> vspltisw
> +;; instruction and then a vupkhsw instruction.
> +
> +(define_predicate "vspltisw_constant_split"
> + (match_code "const_vector,vec_duplicate")
> +{
> + int value = 32;
> +
> + if (!vspltisw_constant_p (op, mode, &value))
> + return false;
> +
> + return true;
> +})
This part can be made simpler once we update the prototype for function
vspltisw_constant_p as below comments.
>
> ;; Return 1 if the operand is constant that can loaded directly with a
> XXSPLTIB
> ;; instruction.
> diff --git a/gcc/config/rs6000/rs6000-protos.h
> b/gcc/config/rs6000/rs6000-protos.h
> index b3c16e7448d..45f3d044eee 100644
> --- a/gcc/config/rs6000/rs6000-protos.h
> +++ b/gcc/config/rs6000/rs6000-protos.h
> @@ -32,6 +32,7 @@ extern void init_cumulative_args (CUMULATIVE_ARGS *, tree,
> rtx, int, int, int,
>
> extern int easy_altivec_constant (rtx, machine_mode);
> extern bool xxspltib_constant_p (rtx, machine_mode, int *, int *);
> +extern bool vspltisw_constant_p (rtx, machine_mode, int *);
> extern int vspltis_shifted (rtx);
> extern HOST_WIDE_INT const_vector_elt_as_int (rtx, unsigned int);
> extern bool macho_lo_sum_memory_operand (rtx, machine_mode);
> diff --git a/gcc/config/rs6000/rs6000.cc b/gcc/config/rs6000/rs6000.cc
> index df491bee2ea..984624026c2 100644
> --- a/gcc/config/rs6000/rs6000.cc
> +++ b/gcc/config/rs6000/rs6000.cc
> @@ -6292,6 +6292,12 @@ easy_altivec_constant (rtx op, machine_mode mode)
> && INTVAL (CONST_VECTOR_ELT (op, 1)) == -1)
> return 8;
>
> + /* If V2DI constant is within RANGE (-16, 15), it can be synthesized
> with
> + a vspltisw and a vupkhsw. */
> + int value = 32;
> + if (vspltisw_constant_p (op, mode, &value))
> + return 8;
> +
It seems to be a bad idea to put this check here, since the comment of this
function easy_altivec_constant says it's for "can be synthesized with
a vspltisb, vspltish or vspltisw.", it's consistent with what function
gen_easy_altivec_constant does. Though the existence of vspltis_shifted
already makes it different, we don't make it worse, so could you move this
check to easy_vector_constant instead?
> return 0;
> }
>
> @@ -6494,6 +6500,69 @@ xxspltib_constant_p (rtx op,
> return true;
> }
>
> +/* Return true if OP mode is V2DI and can be synthesized with ISA 2.07
> + instructions vupkhsw and vspltisw.
> +
> + Return the constant that is being split via CONSTANT_PTR. */
> +
> +bool
> +vspltisw_constant_p (rtx op, machine_mode mode, int *constant_ptr)
This function name is bad, as it implies that it's to check if the constant
can be synthesized with vspltisw, but actually it isn't. So how about
"vspltisw_vupkhsw_constant_p"?
> +{
> + HOST_WIDE_INT value;
> + rtx element;
> +
> + *constant_ptr = 32;
This assignment is useless, could we make the prototype as:
... constant_p (rtx op, machine_mode mode, int *constant_ptr = nullptr)
and only do the final assignment when constant_ptr isn't nullptr
(and it return true obviously), since I noticed that several uses
of this function don't requires the value for *constant_ptr.
> +
> + if (!TARGET_P8_VECTOR)
> + return false;
> +
> + if (mode == VOIDmode)
> + mode = GET_MODE (op);
> + else if (mode != GET_MODE (op) && GET_MODE (op) != VOIDmode)
> + return false;
Why we need these checks?
> +
> + if (mode != V2DImode)
> + return false;
> +
> + if (GET_CODE (op) == VEC_DUPLICATE)
> + {
I'd expect that VEC_DUPLICATE constant will be normalized into
CONST_VECTOR, so this hunk for VEC_DUPLICATE is useless.
> + element = XEXP (op, 0);
> +
> + if (!CONST_INT_P (element))
> + return false;
> +
> + value = INTVAL (element);
> + if (value == 0 || value == 1
> + || !EASY_VECTOR_15 (value))
> + return false;
> + }
> +
> + else if (GET_CODE (op) == CONST_VECTOR)
> + {
> + element = CONST_VECTOR_ELT (op, 0);
You can use const_vec_duplicate_p here, like:
rtx elt;
if (!const_vec_duplicate_p (op, &elt))
return false;
> +
> + if (!CONST_INT_P (element))
> + return false;
> +
I think it's always true, you can just assert CONST_INT_P (elt))
since you have early returned for mode != V2DImode.
> + value = INTVAL (element);
> + if (value == 0 || value == 1
> + || !EASY_VECTOR_15 (value))
> + return false;
> +
> + element = CONST_VECTOR_ELT (op, 1);
> + if (!CONST_INT_P (element))
> + return false;
> +
> + if (value != INTVAL (element))
> + return false;
> + }
> + else
> + return false;
> +
> + *constant_ptr = (int) value;
> + return true;
> +}
> +
> const char *
> output_vec_const_move (rtx *operands)
> {
> diff --git a/gcc/config/rs6000/vsx.md b/gcc/config/rs6000/vsx.md
> index e226a93bbe5..6805f794848 100644
> --- a/gcc/config/rs6000/vsx.md
> +++ b/gcc/config/rs6000/vsx.md
> @@ -1174,6 +1174,32 @@ (define_insn_and_split "*xxspltib_<mode>_split"
> [(set_attr "type" "vecperm")
> (set_attr "length" "8")])
>
> +(define_insn_and_split "*vspltisw_v2di_split"
> + [(set (match_operand:V2DI 0 "altivec_register_operand" "=v")
> + (match_operand:V2DI 1 "vspltisw_constant_split" "wT"))]
> + "TARGET_P8_VECTOR"
> + "#"
> + "&& 1"
> + [(const_int 0)]
> +{
> + int value = 32;
> + rtx op0 = operands[0];
> + rtx op1 = operands[1];
> + rtx tmp = can_create_pseudo_p ()
> + ? gen_reg_rtx (V4SImode)
> + : gen_lowpart (V4SImode, op0);
> +
> + if (!vspltisw_constant_p (op1, V2DImode, &value))
> + gcc_unreachable ();
Shouldn't predicate vspltisw_constant_split guarantee this?
> +
> + emit_insn (gen_altivec_vspltisw (tmp, GEN_INT (value)));
> + emit_insn (gen_altivec_vupkhsw_direct (op0, tmp));
> +
> + DONE;
> +}
> + [(set_attr "type" "vecperm")
> + (set_attr "length" "8")])
> +
>
> ;; Prefer using vector registers over GPRs. Prefer using ISA 3.0's XXSPLTISB
> ;; or Altivec VSPLITW 0/-1 over XXLXOR/XXLORC to set a register to all 0's or
> diff --git a/gcc/testsuite/gcc.target/powerpc/p8-splat.c
> b/gcc/testsuite/gcc.target/powerpc/p8-splat.c
> new file mode 100644
> index 00000000000..aec0f20edb9
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/powerpc/p8-splat.c
Please use pr104124 for test case name.
BR,
Kewen
