"H.J. Lu" <[email protected]> writes:
> 1. Update vec_duplicate to allow to fail so that backend can only allow
> broadcasting an integer constant to a vector when broadcast instruction
> is available. This can be used by memset expander to avoid vec_duplicate
> when loading from constant pool is more efficient.
I don't see any changes in target-independent code though, other than
the doc update. It's still the case that (existing) uses of
vec_duplicate_optab do not allow it to fail.
Thanks,
Richard
> 2. Add vec_duplicate<mode> expander and enable vec_duplicate from a
> non-standard SSE constant integer only if vector broadcast is available.
>
> * config/i386/i386-expand.c (ix86_expand_integer_vec_duplicate):
> New function.
> * config/i386/i386-protos.h (ix86_expand_integer_vec_duplicat):
> New prototype.
> * config/i386/sse.md (INT_BROADCAST_MODE): New mode iterator.
> (vec_duplicate<mode>): New expander.
> * doc/md.texi: Update vec_duplicate.
> ---
> gcc/config/i386/i386-expand.c | 24 ++++++++++++++++++++++++
> gcc/config/i386/i386-protos.h | 1 +
> gcc/config/i386/sse.md | 28 ++++++++++++++++++++++++++++
> gcc/doc/md.texi | 2 --
> 4 files changed, 53 insertions(+), 2 deletions(-)
>
> diff --git a/gcc/config/i386/i386-expand.c b/gcc/config/i386/i386-expand.c
> index e9e89c82764..75c160d4349 100644
> --- a/gcc/config/i386/i386-expand.c
> +++ b/gcc/config/i386/i386-expand.c
> @@ -15742,6 +15742,30 @@ ix86_expand_vector_extract (bool mmx_ok, rtx target,
> rtx vec, int elt)
> }
> }
>
> +/* Expand integer vec_duplicate. Return true if successful. */
> +
> +bool
> +ix86_expand_integer_vec_duplicate (rtx *operands)
> +{
> + /* Enable VEC_DUPLICATE from a non-standard SSE constant integer only
> + if vector broadcast is available. */
> + machine_mode mode = GET_MODE (operands[0]);
> + if (CONST_INT_P (operands[1])
> + && (!(TARGET_AVX2
> + || (TARGET_AVX
> + && (GET_MODE_INNER (mode) == SImode
> + || GET_MODE_INNER (mode) == DImode)))
> + || standard_sse_constant_p (operands[1], mode)))
> + return false;
> +
> + bool ok = ix86_expand_vector_init_duplicate (false, mode,
> + operands[0],
> + operands[1]);
> + gcc_assert (ok);
> +
> + return true;
> +}
> +
> /* Generate code to copy vector bits i / 2 ... i - 1 from vector SRC
> to bits 0 ... i / 2 - 1 of vector DEST, which has the same mode.
> The upper bits of DEST are undefined, though they shouldn't cause
> diff --git a/gcc/config/i386/i386-protos.h b/gcc/config/i386/i386-protos.h
> index 71745b9a1ea..a6cc09bb75b 100644
> --- a/gcc/config/i386/i386-protos.h
> +++ b/gcc/config/i386/i386-protos.h
> @@ -258,6 +258,7 @@ extern void ix86_expand_mul_widen_hilo (rtx, rtx, rtx,
> bool, bool);
> extern void ix86_expand_sse2_mulv4si3 (rtx, rtx, rtx);
> extern void ix86_expand_sse2_mulvxdi3 (rtx, rtx, rtx);
> extern void ix86_expand_sse2_abs (rtx, rtx);
> +extern bool ix86_expand_integer_vec_duplicate (rtx *);
>
> /* In i386-c.c */
> extern void ix86_target_macros (void);
> diff --git a/gcc/config/i386/sse.md b/gcc/config/i386/sse.md
> index e4f01e64bc1..53a703fb466 100644
> --- a/gcc/config/i386/sse.md
> +++ b/gcc/config/i386/sse.md
> @@ -24640,3 +24640,31 @@ (define_insn "*aes<aeswideklvariant>u8"
> "TARGET_WIDEKL"
> "aes<aeswideklvariant>\t{%0}"
> [(set_attr "type" "other")])
> +
> +;; Modes handled by broadcast patterns. NB: Allow V64QI and V32HI with
> +;; TARGET_AVX512F since ix86_expand_integer_vec_duplicate can expand
> +;; without TARGET_AVX512BW which is used by memset vector broadcast
> +;; expander to XI with:
> +;; vmovd %edi, %xmm15
> +;; vpbroadcastb %xmm15, %ymm15
> +;; vinserti64x4 $0x1, %ymm15, %zmm15, %zmm15
> +
> +(define_mode_iterator INT_BROADCAST_MODE
> + [(V64QI "TARGET_AVX512F") (V32QI "TARGET_AVX") V16QI
> + (V32HI "TARGET_AVX512F") (V16HI "TARGET_AVX") V8HI
> + (V16SI "TARGET_AVX512F") (V8SI "TARGET_AVX") V4SI
> + (V8DI "TARGET_AVX512F && TARGET_64BIT")
> + (V4DI "TARGET_AVX && TARGET_64BIT") (V2DI "TARGET_64BIT")])
> +
> +;; Broadcast from an integer. NB: Enable broadcast only if we can move
> +;; from GPR to SSE register directly.
> +(define_expand "vec_duplicate<mode>"
> + [(set (match_operand:INT_BROADCAST_MODE 0 "register_operand")
> + (vec_duplicate:INT_BROADCAST_MODE
> + (match_operand:<ssescalarmode> 1 "general_operand")))]
> + "TARGET_SSE2 && TARGET_INTER_UNIT_MOVES_TO_VEC"
> +{
> + if (!ix86_expand_integer_vec_duplicate (operands))
> + FAIL;
> + DONE;
> +})
> diff --git a/gcc/doc/md.texi b/gcc/doc/md.texi
> index 1b918144330..a892c94d163 100644
> --- a/gcc/doc/md.texi
> +++ b/gcc/doc/md.texi
> @@ -5077,8 +5077,6 @@ the mode appropriate for one element of @var{m}.
> This pattern only handles duplicates of non-constant inputs. Constant
> vectors go through the @code{mov@var{m}} pattern instead.
>
> -This pattern is not allowed to @code{FAIL}.
> -
> @cindex @code{vec_series@var{m}} instruction pattern
> @item @samp{vec_series@var{m}}
> Initialize vector output operand 0 so that element @var{i} is equal to
