On Tue, Jul 12, 2022 at 2:07 PM Richard Sandiford via Gcc-patches
<gcc-patches@gcc.gnu.org> wrote:
>
> The PR is about the aarch64 port using an ACLE built-in function
> to vectorise a scalar function call, even though the ECF_* flags for
> the ACLE function didn't match the ECF_* flags for the scalar call.
>
> To some extent that kind of difference is inevitable, since the
> ACLE intrinsics are supposed to follow the behaviour of the
> underlying instruction as closely as possible. Also, using
> target-specific builtins has the drawback of limiting further
> gimple optimisation, since the gimple optimisers won't know what
> the function does.
>
> We handle several other maths functions, including round, floor
> and ceil, by defining directly-mapped internal functions that
> are linked to the associated built-in functions. This has two
> main advantages:
>
> - it means that, internally, we are not restricted to the set of
> scalar types that happen to have associated C/C++ functions
>
> - the functions (and thus the underlying optabs) extend naturally
> to vectors
>
> This patch takes the same approach for the remaining functions
> handled by aarch64_builtin_vectorized_function.
>
> Tested on aarch64-linux-gnu and x86_64-linux-gnu. OK to install?
OK.
Thanks,
Richard.
> Richard
>
>
> gcc/
> PR target/106253
> * predict.h (insn_optimization_type): Declare.
> * predict.cc (insn_optimization_type): New function.
> * internal-fn.def (IFN_ICEIL, IFN_IFLOOR, IFN_IRINT, IFN_IROUND)
> (IFN_LCEIL, IFN_LFLOOR, IFN_LRINT, IFN_LROUND, IFN_LLCEIL)
> (IFN_LLFLOOR, IFN_LLRINT, IFN_LLROUND): New internal functions.
> * internal-fn.cc (unary_convert_direct): New macro.
> (expand_convert_optab_fn): New function.
> (expand_unary_convert_optab_fn): New macro.
> (direct_unary_convert_optab_supported_p): Likewise.
> * optabs.cc (expand_sfix_optab): Pass insn_optimization_type to
> convert_optab_handler.
> * config/aarch64/aarch64-protos.h
> (aarch64_builtin_vectorized_function): Delete.
> * config/aarch64/aarch64-builtins.cc
> (aarch64_builtin_vectorized_function): Delete.
> * config/aarch64/aarch64.cc
> (TARGET_VECTORIZE_BUILTIN_VECTORIZED_FUNCTION): Delete.
> * config/i386/i386.cc (ix86_optab_supported_p): Handle lround_optab.
> * config/i386/i386.md (lround<X87MODEF:mode><SWI248x:mode>2): Remove
> optimize_insn_for_size_p test.
>
> gcc/testsuite/
> PR target/106253
> * gcc.target/aarch64/vect_unary_1.c: Add tests for iroundf,
> llround, iceilf, llceil, ifloorf, llfloor, irintf and llrint.
> * gfortran.dg/vect/pr106253.f: New test.
> ---
> gcc/config/aarch64/aarch64-builtins.cc | 83 -------------------
> gcc/config/aarch64/aarch64-protos.h | 1 -
> gcc/config/aarch64/aarch64.cc | 4 -
> gcc/config/i386/i386.cc | 1 +
> gcc/config/i386/i386.md | 3 -
> gcc/internal-fn.cc | 20 +++++
> gcc/internal-fn.def | 23 +++++
> gcc/optabs.cc | 3 +-
> gcc/predict.cc | 11 +++
> gcc/predict.h | 1 +
> .../gcc.target/aarch64/vect_unary_1.c | 65 ++++++++++++++-
> gcc/testsuite/gfortran.dg/vect/pr106253.f | 35 ++++++++
> 12 files changed, 157 insertions(+), 93 deletions(-)
> create mode 100644 gcc/testsuite/gfortran.dg/vect/pr106253.f
>
> diff --git a/gcc/config/aarch64/aarch64-builtins.cc
> b/gcc/config/aarch64/aarch64-builtins.cc
> index a486321e10f..adfddb8b215 100644
> --- a/gcc/config/aarch64/aarch64-builtins.cc
> +++ b/gcc/config/aarch64/aarch64-builtins.cc
> @@ -2555,89 +2555,6 @@ aarch64_general_expand_builtin (unsigned int fcode,
> tree exp, rtx target,
> gcc_unreachable ();
> }
>
> -tree
> -aarch64_builtin_vectorized_function (unsigned int fn, tree type_out,
> - tree type_in)
> -{
> - machine_mode in_mode, out_mode;
> -
> - if (TREE_CODE (type_out) != VECTOR_TYPE
> - || TREE_CODE (type_in) != VECTOR_TYPE)
> - return NULL_TREE;
> -
> - out_mode = TYPE_MODE (type_out);
> - in_mode = TYPE_MODE (type_in);
> -
> -#undef AARCH64_CHECK_BUILTIN_MODE
> -#define AARCH64_CHECK_BUILTIN_MODE(C, N) 1
> -#define AARCH64_FIND_FRINT_VARIANT(N) \
> - (AARCH64_CHECK_BUILTIN_MODE (2, D) \
> - ? aarch64_builtin_decls[AARCH64_SIMD_BUILTIN_UNOP_##N##v2df] \
> - : (AARCH64_CHECK_BUILTIN_MODE (4, S) \
> - ? aarch64_builtin_decls[AARCH64_SIMD_BUILTIN_UNOP_##N##v4sf] \
> - : (AARCH64_CHECK_BUILTIN_MODE (2, S) \
> - ? aarch64_builtin_decls[AARCH64_SIMD_BUILTIN_UNOP_##N##v2sf] \
> - : NULL_TREE)))
> - switch (fn)
> - {
> -#undef AARCH64_CHECK_BUILTIN_MODE
> -#define AARCH64_CHECK_BUILTIN_MODE(C, N) \
> - (out_mode == V##C##N##Imode && in_mode == V##C##N##Fmode)
> - CASE_CFN_IFLOOR:
> - CASE_CFN_LFLOOR:
> - CASE_CFN_LLFLOOR:
> - {
> - enum aarch64_builtins builtin;
> - if (AARCH64_CHECK_BUILTIN_MODE (2, D))
> - builtin = AARCH64_SIMD_BUILTIN_UNOP_lfloorv2dfv2di;
> - else if (AARCH64_CHECK_BUILTIN_MODE (4, S))
> - builtin = AARCH64_SIMD_BUILTIN_UNOP_lfloorv4sfv4si;
> - else if (AARCH64_CHECK_BUILTIN_MODE (2, S))
> - builtin = AARCH64_SIMD_BUILTIN_UNOP_lfloorv2sfv2si;
> - else
> - return NULL_TREE;
> -
> - return aarch64_builtin_decls[builtin];
> - }
> - CASE_CFN_ICEIL:
> - CASE_CFN_LCEIL:
> - CASE_CFN_LLCEIL:
> - {
> - enum aarch64_builtins builtin;
> - if (AARCH64_CHECK_BUILTIN_MODE (2, D))
> - builtin = AARCH64_SIMD_BUILTIN_UNOP_lceilv2dfv2di;
> - else if (AARCH64_CHECK_BUILTIN_MODE (4, S))
> - builtin = AARCH64_SIMD_BUILTIN_UNOP_lceilv4sfv4si;
> - else if (AARCH64_CHECK_BUILTIN_MODE (2, S))
> - builtin = AARCH64_SIMD_BUILTIN_UNOP_lceilv2sfv2si;
> - else
> - return NULL_TREE;
> -
> - return aarch64_builtin_decls[builtin];
> - }
> - CASE_CFN_IROUND:
> - CASE_CFN_LROUND:
> - CASE_CFN_LLROUND:
> - {
> - enum aarch64_builtins builtin;
> - if (AARCH64_CHECK_BUILTIN_MODE (2, D))
> - builtin = AARCH64_SIMD_BUILTIN_UNOP_lroundv2dfv2di;
> - else if (AARCH64_CHECK_BUILTIN_MODE (4, S))
> - builtin = AARCH64_SIMD_BUILTIN_UNOP_lroundv4sfv4si;
> - else if (AARCH64_CHECK_BUILTIN_MODE (2, S))
> - builtin = AARCH64_SIMD_BUILTIN_UNOP_lroundv2sfv2si;
> - else
> - return NULL_TREE;
> -
> - return aarch64_builtin_decls[builtin];
> - }
> - default:
> - return NULL_TREE;
> - }
> -
> - return NULL_TREE;
> -}
> -
> /* Return builtin for reciprocal square root. */
>
> tree
> diff --git a/gcc/config/aarch64/aarch64-protos.h
> b/gcc/config/aarch64/aarch64-protos.h
> index dabd047d7ba..19c9d3cb179 100644
> --- a/gcc/config/aarch64/aarch64-protos.h
> +++ b/gcc/config/aarch64/aarch64-protos.h
> @@ -986,7 +986,6 @@ gimple *aarch64_general_gimple_fold_builtin (unsigned
> int, gcall *,
> rtx aarch64_general_expand_builtin (unsigned int, tree, rtx, int);
> tree aarch64_general_builtin_decl (unsigned, bool);
> tree aarch64_general_builtin_rsqrt (unsigned int);
> -tree aarch64_builtin_vectorized_function (unsigned int, tree, tree);
> void handle_arm_acle_h (void);
> void handle_arm_neon_h (void);
>
> diff --git a/gcc/config/aarch64/aarch64.cc b/gcc/config/aarch64/aarch64.cc
> index d049f9a9819..25f4cbb466d 100644
> --- a/gcc/config/aarch64/aarch64.cc
> +++ b/gcc/config/aarch64/aarch64.cc
> @@ -27584,10 +27584,6 @@ aarch64_libgcc_floating_mode_supported_p
> #undef TARGET_VECTORIZE_BUILTINS
> #define TARGET_VECTORIZE_BUILTINS
>
> -#undef TARGET_VECTORIZE_BUILTIN_VECTORIZED_FUNCTION
> -#define TARGET_VECTORIZE_BUILTIN_VECTORIZED_FUNCTION \
> - aarch64_builtin_vectorized_function
> -
> #undef TARGET_VECTORIZE_AUTOVECTORIZE_VECTOR_MODES
> #define TARGET_VECTORIZE_AUTOVECTORIZE_VECTOR_MODES \
> aarch64_autovectorize_vector_modes
> diff --git a/gcc/config/i386/i386.cc b/gcc/config/i386/i386.cc
> index 95cb1e2ce70..3a3c7299eb4 100644
> --- a/gcc/config/i386/i386.cc
> +++ b/gcc/config/i386/i386.cc
> @@ -24004,6 +24004,7 @@ ix86_optab_supported_p (int op, machine_mode mode1,
> machine_mode,
> case ldexp_optab:
> case scalb_optab:
> case round_optab:
> + case lround_optab:
> return opt_type == OPTIMIZE_FOR_SPEED;
>
> case rint_optab:
> diff --git a/gcc/config/i386/i386.md b/gcc/config/i386/i386.md
> index 3b02d0cd567..bf29f444382 100644
> --- a/gcc/config/i386/i386.md
> +++ b/gcc/config/i386/i386.md
> @@ -19926,9 +19926,6 @@ (define_expand "lround<X87MODEF:mode><SWI248x:mode>2"
> && ((<SWI248x:MODE>mode != DImode) || TARGET_64BIT)
> && !flag_trapping_math && !flag_rounding_math)"
> {
> - if (optimize_insn_for_size_p ())
> - FAIL;
> -
> if (SSE_FLOAT_MODE_P (<X87MODEF:MODE>mode) && TARGET_SSE_MATH
> && <SWI248x:MODE>mode != HImode
> && ((<SWI248x:MODE>mode != DImode) || TARGET_64BIT)
> diff --git a/gcc/internal-fn.cc b/gcc/internal-fn.cc
> index d666ccccf67..28973d957fb 100644
> --- a/gcc/internal-fn.cc
> +++ b/gcc/internal-fn.cc
> @@ -120,6 +120,7 @@ init_internal_fns ()
> #define len_store_direct { 3, 3, false }
> #define vec_set_direct { 3, 3, false }
> #define unary_direct { 0, 0, true }
> +#define unary_convert_direct { -1, 0, true }
> #define binary_direct { 0, 0, true }
> #define ternary_direct { 0, 0, true }
> #define cond_unary_direct { 1, 1, true }
> @@ -3679,6 +3680,19 @@ expand_while_optab_fn (internal_fn, gcall *stmt,
> convert_optab optab)
> emit_move_insn (lhs_rtx, ops[0].value);
> }
>
> +/* Expand a call to a convert-like optab using the operands in STMT.
> + FN has a single output operand and NARGS input operands. */
> +
> +static void
> +expand_convert_optab_fn (internal_fn fn, gcall *stmt, convert_optab optab,
> + unsigned int nargs)
> +{
> + tree_pair types = direct_internal_fn_types (fn, stmt);
> + insn_code icode = convert_optab_handler (optab, TYPE_MODE (types.first),
> + TYPE_MODE (types.second));
> + expand_fn_using_insn (stmt, icode, 1, nargs);
> +}
> +
> /* Expanders for optabs that can use expand_direct_optab_fn. */
>
> #define expand_unary_optab_fn(FN, STMT, OPTAB) \
> @@ -3711,6 +3725,11 @@ expand_while_optab_fn (internal_fn, gcall *stmt,
> convert_optab optab)
> #define expand_check_ptrs_optab_fn(FN, STMT, OPTAB) \
> expand_direct_optab_fn (FN, STMT, OPTAB, 4)
>
> +/* Expanders for optabs that can use expand_convert_optab_fn. */
> +
> +#define expand_unary_convert_optab_fn(FN, STMT, OPTAB) \
> + expand_convert_optab_fn (FN, STMT, OPTAB, 1)
> +
> /* RETURN_TYPE and ARGS are a return type and argument list that are
> in principle compatible with FN (which satisfies direct_internal_fn_p).
> Return the types that should be used to determine whether the
> @@ -3783,6 +3802,7 @@ multi_vector_optab_supported_p (convert_optab optab,
> tree_pair types,
> }
>
> #define direct_unary_optab_supported_p direct_optab_supported_p
> +#define direct_unary_convert_optab_supported_p convert_optab_supported_p
> #define direct_binary_optab_supported_p direct_optab_supported_p
> #define direct_ternary_optab_supported_p direct_optab_supported_p
> #define direct_cond_unary_optab_supported_p direct_optab_supported_p
> diff --git a/gcc/internal-fn.def b/gcc/internal-fn.def
> index d2d550d3586..7c398baadc8 100644
> --- a/gcc/internal-fn.def
> +++ b/gcc/internal-fn.def
> @@ -61,6 +61,9 @@ along with GCC; see the file COPYING3. If not see
> - binary: a normal binary optab, such as vec_interleave_lo_<mode>
> - ternary: a normal ternary optab, such as fma<mode>4
>
> + - unary_convert: a single-input conversion optab, such as
> + lround<srcmode><dstmode>2.
> +
> - cond_binary: a conditional binary optab, such as cond_add<mode>
> - cond_ternary: a conditional ternary optab, such as cond_fma_rev<mode>
>
> @@ -267,6 +270,26 @@ DEF_INTERNAL_FLT_FLOATN_FN (SQRT, ECF_CONST, sqrt, unary)
> DEF_INTERNAL_FLT_FN (TAN, ECF_CONST, tan, unary)
> DEF_INTERNAL_FLT_FN (TANH, ECF_CONST, tanh, unary)
>
> +/* Floating-point to integer conversions.
> +
> + ??? Here we preserve the I/L/LL prefix convention from the
> + corresponding built-in functions, rather than make the internal
> + functions polymorphic in both the argument and the return types.
> + Perhaps an alternative would be to pass a zero of the required
> + return type as a second parameter. */
> +DEF_INTERNAL_FLT_FN (ICEIL, ECF_CONST, lceil, unary_convert)
> +DEF_INTERNAL_FLT_FN (IFLOOR, ECF_CONST, lfloor, unary_convert)
> +DEF_INTERNAL_FLT_FN (IRINT, ECF_CONST, lrint, unary_convert)
> +DEF_INTERNAL_FLT_FN (IROUND, ECF_CONST, lround, unary_convert)
> +DEF_INTERNAL_FLT_FN (LCEIL, ECF_CONST, lceil, unary_convert)
> +DEF_INTERNAL_FLT_FN (LFLOOR, ECF_CONST, lfloor, unary_convert)
> +DEF_INTERNAL_FLT_FN (LRINT, ECF_CONST, lrint, unary_convert)
> +DEF_INTERNAL_FLT_FN (LROUND, ECF_CONST, lround, unary_convert)
> +DEF_INTERNAL_FLT_FN (LLCEIL, ECF_CONST, lceil, unary_convert)
> +DEF_INTERNAL_FLT_FN (LLFLOOR, ECF_CONST, lfloor, unary_convert)
> +DEF_INTERNAL_FLT_FN (LLRINT, ECF_CONST, lrint, unary_convert)
> +DEF_INTERNAL_FLT_FN (LLROUND, ECF_CONST, lround, unary_convert)
> +
> /* FP rounding. */
> DEF_INTERNAL_FLT_FLOATN_FN (CEIL, ECF_CONST, ceil, unary)
> DEF_INTERNAL_FLT_FLOATN_FN (FLOOR, ECF_CONST, floor, unary)
> diff --git a/gcc/optabs.cc b/gcc/optabs.cc
> index a50dd798f2a..165f8d1fa22 100644
> --- a/gcc/optabs.cc
> +++ b/gcc/optabs.cc
> @@ -5828,7 +5828,8 @@ expand_sfix_optab (rtx to, rtx from, convert_optab tab)
> FOR_EACH_MODE_FROM (fmode, GET_MODE (from))
> FOR_EACH_MODE_FROM (imode, GET_MODE (to))
> {
> - icode = convert_optab_handler (tab, imode, fmode);
> + icode = convert_optab_handler (tab, imode, fmode,
> + insn_optimization_type ());
> if (icode != CODE_FOR_nothing)
> {
> rtx_insn *last = get_last_insn ();
> diff --git a/gcc/predict.cc b/gcc/predict.cc
> index b36caa3ae82..1bc7ab94454 100644
> --- a/gcc/predict.cc
> +++ b/gcc/predict.cc
> @@ -362,6 +362,17 @@ optimize_insn_for_speed_p (void)
> return !optimize_insn_for_size_p ();
> }
>
> +/* Return the optimization type that should be used for the current
> + instruction. */
> +
> +optimization_type
> +insn_optimization_type ()
> +{
> + return (optimize_insn_for_speed_p ()
> + ? OPTIMIZE_FOR_SPEED
> + : OPTIMIZE_FOR_SIZE);
> +}
> +
> /* Return TRUE if LOOP should be optimized for size. */
>
> optimize_size_level
> diff --git a/gcc/predict.h b/gcc/predict.h
> index 864997498ec..25484373769 100644
> --- a/gcc/predict.h
> +++ b/gcc/predict.h
> @@ -68,6 +68,7 @@ extern enum optimize_size_level optimize_edge_for_size_p
> (edge);
> extern bool optimize_edge_for_speed_p (edge);
> extern enum optimize_size_level optimize_insn_for_size_p (void);
> extern bool optimize_insn_for_speed_p (void);
> +extern optimization_type insn_optimization_type ();
> extern enum optimize_size_level optimize_loop_for_size_p (class loop *);
> extern bool optimize_loop_for_speed_p (class loop *);
> extern bool optimize_loop_nest_for_speed_p (class loop *);
> diff --git a/gcc/testsuite/gcc.target/aarch64/vect_unary_1.c
> b/gcc/testsuite/gcc.target/aarch64/vect_unary_1.c
> index 8516808becf..94d9af1a55d 100644
> --- a/gcc/testsuite/gcc.target/aarch64/vect_unary_1.c
> +++ b/gcc/testsuite/gcc.target/aarch64/vect_unary_1.c
> @@ -1,4 +1,4 @@
> -/* { dg-options "-O3 --save-temps" } */
> +/* { dg-options "-O3 -fno-math-errno --save-temps" } */
> /* { dg-final { check-function-bodies "**" "" "" } } */
>
> #include <stdint.h>
> @@ -184,3 +184,66 @@ TEST2 (int, ctz, int)
> ** ret
> */
> TEST4 (int, ctz, int)
> +
> +/*
> +** test2_int_iroundf_float:
> +** fcvtas v0.2s, v1.2s
> +** ret
> +*/
> +TEST2 (int, iroundf, float)
> +
> +/*
> +** test2_int64_t_llround_double:
> +** fcvtas v0.2d, v1.2d
> +** ret
> +*/
> +TEST2 (int64_t, llround, double)
> +
> +/*
> +** test4_int_iroundf_float:
> +** fcvtas v0.4s, v1.4s
> +** ret
> +*/
> +TEST4 (int, iroundf, float)
> +
> +/*
> +** test2_int_ifloorf_float:
> +** fcvtms v0.2s, v1.2s
> +** ret
> +*/
> +TEST2 (int, ifloorf, float)
> +
> +/*
> +** test2_int64_t_llfloor_double:
> +** fcvtms v0.2d, v1.2d
> +** ret
> +*/
> +TEST2 (int64_t, llfloor, double)
> +
> +/*
> +** test4_int_ifloorf_float:
> +** fcvtms v0.4s, v1.4s
> +** ret
> +*/
> +TEST4 (int, ifloorf, float)
> +
> +/*
> +** test2_int_iceilf_float:
> +** fcvtps v0.2s, v1.2s
> +** ret
> +*/
> +TEST2 (int, iceilf, float)
> +
> +/*
> +** test2_int64_t_llceil_double:
> +** fcvtps v0.2d, v1.2d
> +** ret
> +*/
> +TEST2 (int64_t, llceil, double)
> +
> +/*
> +** test4_int_iceilf_float:
> +** fcvtps v0.4s, v1.4s
> +** ret
> +*/
> +TEST4 (int, iceilf, float)
> diff --git a/gcc/testsuite/gfortran.dg/vect/pr106253.f
> b/gcc/testsuite/gfortran.dg/vect/pr106253.f
> new file mode 100644
> index 00000000000..1b6b7e892f2
> --- /dev/null
> +++ b/gcc/testsuite/gfortran.dg/vect/pr106253.f
> @@ -0,0 +1,35 @@
> +! { dg-do compile }
> +
> + SUBROUTINE DGEMV ( TRANS, M, N, ALPHA, A, LDA, X, INCX, &
> + & BETA, Y, INCY )
> + LOGICAL LSAME
> + IF ( .NOT.LSAME( TRANS, 'N' ).AND. &
> + & .NOT.LSAME( TRANS, 'C' ) )THEN
> + END IF
> + END
> + subroutine evlrnf (ptrs0t, nclsm, prnf0t)
> + real, dimension (1:nclsm,1:nclsm), intent (in) :: ptrs0t
> + real, dimension (1:nclsm,1:nclsm), intent (out):: prnf0t
> + real, allocatable, dimension (:,:) :: utrsft ! probas up
> + real, allocatable, dimension (:,:) :: dtrsft ! probas down
> + real, allocatable, dimension (:,:) :: xwrkt ! matrice
> + do icls = 1, nclsm
> + do ival = ipic - 1, 1, -1
> + xwrkt = trs2a2 (ival, ipic, utrsft, dtrsft, ncls)
> + enddo
> + enddo
> + contains
> + function trs2a2 (j, k, u, d, m)
> + real, dimension (1:m,1:m) :: trs2a2 ! resultat
> + real, dimension (1:m,1:m) :: u, d ! matrices utrsft, dtrsft
> + end function trs2a2
> + end
> + program rnflow
> + integer, parameter :: ncls = 256 ! nombre de classes
> + integer, dimension (1:ncls,1:ncls) :: mrnftt ! matrice theorique
> + real, dimension (1:ncls,1:ncls) :: ptrst ! matrice Markov
> + real, dimension (1:ncls,1:ncls) :: prnft ! matrice Rainflow
> + call evlrnf (ptrst, ncls, prnft)
> + mrnftt = nint (real (nsim) * real (npic) * prnft)
> + call cmpmat (mrnftt, mrnfst)
> + end program rnflow
> --
> 2.25.1
>
