Hi Christophe,
> -----Original Message-----
> From: Gcc-patches <gcc-patches-
> [email protected]> On Behalf Of Christophe
> Lyon via Gcc-patches
> Sent: Thursday, January 13, 2022 2:56 PM
> To: [email protected]
> Subject: [PATCH v3 07/15] arm: Implement MVE predicates as vectors of
> booleans
>
> This patch implements support for vectors of booleans to support MVE
> predicates, instead of HImode. Since the ABI mandates pred16_t (aka
> uint16_t) to represent predicates in intrinsics prototypes, we
> introduce a new "predicate" type qualifier so that we can map relevant
> builtins HImode arguments and return value to the appropriate vector
> of booleans (VxBI).
>
> We have to update test_vector_ops_duplicate, because it iterates using
> an offset in bytes, where we would need to iterate in bits: we stop
> iterating when we reach the end of the vector of booleans.
>
> In addition, we have to fix the underlying definition of vectors of
> booleans because ARM/MVE needs a different representation than
> AArch64/SVE. With ARM/MVE the 'true' bit is duplicated over the
> element size, so that a true element of V4BI is represented by
> '0b1111'. This patch updates the aarch64 definition of VNx*BI as
> needed.
>
> 2022-01-13 Christophe Lyon <[email protected]>
> Richard Sandiford <[email protected]>
>
> gcc/
> PR target/100757
> PR target/101325
> * config/aarch64/aarch64-modes.def (VNx16BI, VNx8BI, VNx4BI,
> VNx2BI): Update definition.
> * config/arm/arm-builtins.c (arm_init_simd_builtin_types): Add new
> simd types.
> (arm_init_builtin): Map predicate vectors arguments to HImode.
> (arm_expand_builtin_args): Move HImode predicate arguments to
> VxBI
> rtx. Move return value to HImode rtx.
> * config/arm/arm-builtins.h (arm_type_qualifiers): Add
> qualifier_predicate.
> * config/arm/arm-modes.def (B2I, B4I, V16BI, V8BI, V4BI): New
> modes.
> * config/arm/arm-simd-builtin-types.def (Pred1x16_t,
> Pred2x8_t,Pred4x4_t): New.
> * emit-rtl.c (init_emit_once): Handle all boolean modes.
> * genmodes.c (mode_data): Add boolean field.
> (blank_mode): Initialize it.
> (make_complex_modes): Fix handling of boolean modes.
> (make_vector_modes): Likewise.
> (VECTOR_BOOL_MODE): Use new COMPONENT parameter.
> (make_vector_bool_mode): Likewise.
> (BOOL_MODE): New.
> (make_bool_mode): New.
> (emit_insn_modes_h): Fix generation of boolean modes.
> (emit_class_narrowest_mode): Likewise.
> * machmode.def: Use new BOOL_MODE instead of
> FRACTIONAL_INT_MODE
> to define BImode.
> * rtx-vector-builder.c (rtx_vector_builder::find_cached_value):
> Fix handling of constm1_rtx for VECTOR_BOOL.
> * simplify-rtx.c (native_encode_rtx): Fix support for VECTOR_BOOL.
> (native_decode_vector_rtx): Likewise.
> (test_vector_ops_duplicate): Skip vec_merge test
> with vectors of booleans.
> * varasm.c (output_constant_pool_2): Likewise.
The arm parts look ok. I guess Richard is best placed to approve the midend
parts, but I see he's on the ChangeLog so maybe he needs others to review them.
But then again Richard is maintainer of the gen* machinery that's the most
complicated part of the patch so he can self-approve 😊
Thanks,
Kyrill
>
> diff --git a/gcc/config/aarch64/aarch64-modes.def
> b/gcc/config/aarch64/aarch64-modes.def
> index 976bf9b42be..8f399225a80 100644
> --- a/gcc/config/aarch64/aarch64-modes.def
> +++ b/gcc/config/aarch64/aarch64-modes.def
> @@ -47,10 +47,10 @@ ADJUST_FLOAT_FORMAT (HF, &ieee_half_format);
>
> /* Vector modes. */
>
> -VECTOR_BOOL_MODE (VNx16BI, 16, 2);
> -VECTOR_BOOL_MODE (VNx8BI, 8, 2);
> -VECTOR_BOOL_MODE (VNx4BI, 4, 2);
> -VECTOR_BOOL_MODE (VNx2BI, 2, 2);
> +VECTOR_BOOL_MODE (VNx16BI, 16, BI, 2);
> +VECTOR_BOOL_MODE (VNx8BI, 8, BI, 2);
> +VECTOR_BOOL_MODE (VNx4BI, 4, BI, 2);
> +VECTOR_BOOL_MODE (VNx2BI, 2, BI, 2);
>
> ADJUST_NUNITS (VNx16BI, aarch64_sve_vg * 8);
> ADJUST_NUNITS (VNx8BI, aarch64_sve_vg * 4);
> diff --git a/gcc/config/arm/arm-builtins.c b/gcc/config/arm/arm-builtins.c
> index 9c645722230..2ccfa37c302 100644
> --- a/gcc/config/arm/arm-builtins.c
> +++ b/gcc/config/arm/arm-builtins.c
> @@ -1548,6 +1548,13 @@ arm_init_simd_builtin_types (void)
> arm_simd_types[Bfloat16x4_t].eltype = arm_bf16_type_node;
> arm_simd_types[Bfloat16x8_t].eltype = arm_bf16_type_node;
>
> + if (TARGET_HAVE_MVE)
> + {
> + arm_simd_types[Pred1x16_t].eltype = unsigned_intHI_type_node;
> + arm_simd_types[Pred2x8_t].eltype = unsigned_intHI_type_node;
> + arm_simd_types[Pred4x4_t].eltype = unsigned_intHI_type_node;
> + }
> +
> for (i = 0; i < nelts; i++)
> {
> tree eltype = arm_simd_types[i].eltype;
> @@ -1695,6 +1702,11 @@ arm_init_builtin (unsigned int fcode,
> arm_builtin_datum *d,
> if (qualifiers & qualifier_map_mode)
> op_mode = d->mode;
>
> + /* MVE Predicates use HImode as mandated by the ABI: pred16_t is
> unsigned
> + short. */
> + if (qualifiers & qualifier_predicate)
> + op_mode = HImode;
> +
> /* For pointers, we want a pointer to the basic type
> of the vector. */
> if (qualifiers & qualifier_pointer && VECTOR_MODE_P (op_mode))
> @@ -2939,6 +2951,11 @@ arm_expand_builtin_args (rtx target,
> machine_mode map_mode, int fcode,
> case ARG_BUILTIN_COPY_TO_REG:
> if (POINTER_TYPE_P (TREE_TYPE (arg[argc])))
> op[argc] = convert_memory_address (Pmode, op[argc]);
> +
> + /* MVE uses mve_pred16_t (aka HImode) for vectors of
> predicates. */
> + if (GET_MODE_CLASS (mode[argc]) == MODE_VECTOR_BOOL)
> + op[argc] = gen_lowpart (mode[argc], op[argc]);
> +
> /*gcc_assert (GET_MODE (op[argc]) == mode[argc]); */
> if (!(*insn_data[icode].operand[opno].predicate)
> (op[argc], mode[argc]))
> @@ -3144,6 +3161,13 @@ constant_arg:
> else
> emit_insn (insn);
>
> + if (GET_MODE_CLASS (tmode) == MODE_VECTOR_BOOL)
> + {
> + rtx HItarget = gen_reg_rtx (HImode);
> + emit_move_insn (HItarget, gen_lowpart (HImode, target));
> + return HItarget;
> + }
> +
> return target;
> }
>
> diff --git a/gcc/config/arm/arm-builtins.h b/gcc/config/arm/arm-builtins.h
> index e5130d6d286..a8ef8aef82d 100644
> --- a/gcc/config/arm/arm-builtins.h
> +++ b/gcc/config/arm/arm-builtins.h
> @@ -84,7 +84,9 @@ enum arm_type_qualifiers
> qualifier_lane_pair_index = 0x1000,
> /* Lane indices selected in quadtuplets - must be within range of previous
> argument = a vector. */
> - qualifier_lane_quadtup_index = 0x2000
> + qualifier_lane_quadtup_index = 0x2000,
> + /* MVE vector predicates. */
> + qualifier_predicate = 0x4000
> };
>
> struct arm_simd_type_info
> diff --git a/gcc/config/arm/arm-modes.def b/gcc/config/arm/arm-modes.def
> index de689c8b45e..9ed0cd042c5 100644
> --- a/gcc/config/arm/arm-modes.def
> +++ b/gcc/config/arm/arm-modes.def
> @@ -84,6 +84,14 @@ VECTOR_MODE (FLOAT, BF, 2); /* V2BF. */
> VECTOR_MODE (FLOAT, BF, 4); /* V4BF. */
> VECTOR_MODE (FLOAT, BF, 8); /* V8BF. */
>
> +/* Predicates for MVE. */
> +BOOL_MODE (B2I, 2, 1);
> +BOOL_MODE (B4I, 4, 1);
> +
> +VECTOR_BOOL_MODE (V16BI, 16, BI, 2);
> +VECTOR_BOOL_MODE (V8BI, 8, B2I, 2);
> +VECTOR_BOOL_MODE (V4BI, 4, B4I, 2);
> +
> /* Fraction and accumulator vector modes. */
> VECTOR_MODES (FRACT, 4); /* V4QQ V2HQ */
> VECTOR_MODES (UFRACT, 4); /* V4UQQ V2UHQ */
> diff --git a/gcc/config/arm/arm-simd-builtin-types.def
> b/gcc/config/arm/arm-simd-builtin-types.def
> index 6ba6f211531..920c2a68e4c 100644
> --- a/gcc/config/arm/arm-simd-builtin-types.def
> +++ b/gcc/config/arm/arm-simd-builtin-types.def
> @@ -51,3 +51,7 @@
> ENTRY (Bfloat16x2_t, V2BF, none, 32, bfloat16, 20)
> ENTRY (Bfloat16x4_t, V4BF, none, 64, bfloat16, 20)
> ENTRY (Bfloat16x8_t, V8BF, none, 128, bfloat16, 20)
> +
> + ENTRY (Pred1x16_t, V16BI, unsigned, 16, uint16, 21)
> + ENTRY (Pred2x8_t, V8BI, unsigned, 8, uint16, 21)
> + ENTRY (Pred4x4_t, V4BI, unsigned, 4, uint16, 21)
> diff --git a/gcc/emit-rtl.c b/gcc/emit-rtl.c
> index feeee16d320..5f559f8fd93 100644
> --- a/gcc/emit-rtl.c
> +++ b/gcc/emit-rtl.c
> @@ -6239,9 +6239,14 @@ init_emit_once (void)
>
> /* For BImode, 1 and -1 are unsigned and signed interpretations
> of the same value. */
> - const_tiny_rtx[0][(int) BImode] = const0_rtx;
> - const_tiny_rtx[1][(int) BImode] = const_true_rtx;
> - const_tiny_rtx[3][(int) BImode] = const_true_rtx;
> + for (mode = MIN_MODE_BOOL;
> + mode <= MAX_MODE_BOOL;
> + mode = (machine_mode)((int)(mode) + 1))
> + {
> + const_tiny_rtx[0][(int) mode] = const0_rtx;
> + const_tiny_rtx[1][(int) mode] = const_true_rtx;
> + const_tiny_rtx[3][(int) mode] = const_true_rtx;
> + }
>
> for (mode = MIN_MODE_PARTIAL_INT;
> mode <= MAX_MODE_PARTIAL_INT;
> @@ -6260,13 +6265,16 @@ init_emit_once (void)
> const_tiny_rtx[0][(int) mode] = gen_rtx_CONCAT (mode, inner, inner);
> }
>
> - /* As for BImode, "all 1" and "all -1" are unsigned and signed
> - interpretations of the same value. */
> FOR_EACH_MODE_IN_CLASS (mode, MODE_VECTOR_BOOL)
> {
> const_tiny_rtx[0][(int) mode] = gen_const_vector (mode, 0);
> const_tiny_rtx[3][(int) mode] = gen_const_vector (mode, 3);
> - const_tiny_rtx[1][(int) mode] = const_tiny_rtx[3][(int) mode];
> + if (GET_MODE_INNER (mode) == BImode)
> + /* As for BImode, "all 1" and "all -1" are unsigned and signed
> + interpretations of the same value. */
> + const_tiny_rtx[1][(int) mode] = const_tiny_rtx[3][(int) mode];
> + else
> + const_tiny_rtx[1][(int) mode] = gen_const_vector (mode, 1);
> }
>
> FOR_EACH_MODE_IN_CLASS (mode, MODE_VECTOR_INT)
> diff --git a/gcc/genmodes.c b/gcc/genmodes.c
> index 6001b854547..0bb1a7c0b48 100644
> --- a/gcc/genmodes.c
> +++ b/gcc/genmodes.c
> @@ -78,6 +78,7 @@ struct mode_data
> bool need_bytesize_adj; /* true if this mode needs dynamic size
> adjustment */
> unsigned int int_n; /* If nonzero, then __int<INT_N> will be
> defined */
> + bool boolean;
> };
>
> static struct mode_data *modes[MAX_MODE_CLASS];
> @@ -88,7 +89,8 @@ static const struct mode_data blank_mode = {
> 0, "<unknown>", MAX_MODE_CLASS,
> 0, -1U, -1U, -1U, -1U,
> 0, 0, 0, 0, 0, 0,
> - "<unknown>", 0, 0, 0, 0, false, false, 0
> + "<unknown>", 0, 0, 0, 0, false, false, 0,
> + false
> };
>
> static htab_t modes_by_name;
> @@ -456,7 +458,7 @@ make_complex_modes (enum mode_class cl,
> size_t m_len;
>
> /* Skip BImode. FIXME: BImode probably shouldn't be MODE_INT. */
> - if (m->precision == 1)
> + if (m->boolean)
> continue;
>
> m_len = strlen (m->name);
> @@ -528,7 +530,7 @@ make_vector_modes (enum mode_class cl, const
> char *prefix, unsigned int width,
> not be necessary. */
> if (cl == MODE_FLOAT && m->bytesize == 1)
> continue;
> - if (cl == MODE_INT && m->precision == 1)
> + if (m->boolean)
> continue;
>
> if ((size_t) snprintf (buf, sizeof buf, "%s%u%s", prefix,
> @@ -548,17 +550,18 @@ make_vector_modes (enum mode_class cl, const
> char *prefix, unsigned int width,
>
> /* Create a vector of booleans called NAME with COUNT elements and
> BYTESIZE bytes in total. */
> -#define VECTOR_BOOL_MODE(NAME, COUNT, BYTESIZE) \
> - make_vector_bool_mode (#NAME, COUNT, BYTESIZE, __FILE__, __LINE__)
> +#define VECTOR_BOOL_MODE(NAME, COUNT, COMPONENT, BYTESIZE)
> \
> + make_vector_bool_mode (#NAME, COUNT, #COMPONENT, BYTESIZE,
> \
> + __FILE__, __LINE__)
> static void ATTRIBUTE_UNUSED
> make_vector_bool_mode (const char *name, unsigned int count,
> - unsigned int bytesize, const char *file,
> - unsigned int line)
> + const char *component, unsigned int bytesize,
> + const char *file, unsigned int line)
> {
> - struct mode_data *m = find_mode ("BI");
> + struct mode_data *m = find_mode (component);
> if (!m)
> {
> - error ("%s:%d: no mode \"BI\"", file, line);
> + error ("%s:%d: no mode \"%s\"", file, line, component);
> return;
> }
>
> @@ -596,6 +599,20 @@ make_int_mode (const char *name,
> m->precision = precision;
> }
>
> +#define BOOL_MODE(N, B, Y) \
> + make_bool_mode (#N, B, Y, __FILE__, __LINE__)
> +
> +static void
> +make_bool_mode (const char *name,
> + unsigned int precision, unsigned int bytesize,
> + const char *file, unsigned int line)
> +{
> + struct mode_data *m = new_mode (MODE_INT, name, file, line);
> + m->bytesize = bytesize;
> + m->precision = precision;
> + m->boolean = true;
> +}
> +
> #define OPAQUE_MODE(N, B) \
> make_opaque_mode (#N, -1U, B, __FILE__, __LINE__)
>
> @@ -1298,9 +1315,21 @@ enum machine_mode\n{");
> /* Don't use BImode for MIN_MODE_INT, since otherwise the middle
> end will try to use it for bitfields in structures and the
> like, which we do not want. Only the target md file should
> - generate BImode widgets. */
> - if (first && first->precision == 1 && c == MODE_INT)
> - first = first->next;
> + generate BImode widgets. Since some targets such as ARM/MVE
> + define boolean modes with multiple bits, handle those too. */
> + if (first && first->boolean)
> + {
> + struct mode_data *last_bool = first;
> + printf (" MIN_MODE_BOOL = E_%smode,\n", first->name);
> +
> + while (first && first->boolean)
> + {
> + last_bool = first;
> + first = first->next;
> + }
> +
> + printf (" MAX_MODE_BOOL = E_%smode,\n\n", last_bool->name);
> + }
>
> if (first && last)
> printf (" MIN_%s = E_%smode,\n MAX_%s = E_%smode,\n\n",
> @@ -1679,15 +1708,25 @@ emit_class_narrowest_mode (void)
> print_decl ("unsigned char", "class_narrowest_mode",
> "MAX_MODE_CLASS");
>
> for (c = 0; c < MAX_MODE_CLASS; c++)
> - /* Bleah, all this to get the comment right for MIN_MODE_INT. */
> - tagged_printf ("MIN_%s", mode_class_names[c],
> - modes[c]
> - ? ((c != MODE_INT || modes[c]->precision != 1)
> - ? modes[c]->name
> - : (modes[c]->next
> - ? modes[c]->next->name
> - : void_mode->name))
> - : void_mode->name);
> + {
> + /* Bleah, all this to get the comment right for MIN_MODE_INT. */
> + const char *comment_name = void_mode->name;
> +
> + if (modes[c])
> + if (c != MODE_INT || !modes[c]->boolean)
> + comment_name = modes[c]->name;
> + else
> + {
> + struct mode_data *m = modes[c];
> + while (m->boolean)
> + m = m->next;
> + if (m)
> + comment_name = m->name;
> + else
> + comment_name = void_mode->name;
> + }
> + tagged_printf ("MIN_%s", mode_class_names[c], comment_name);
> + }
>
> print_closer ();
> }
> diff --git a/gcc/machmode.def b/gcc/machmode.def
> index 866a2082d01..eb7905ea23d 100644
> --- a/gcc/machmode.def
> +++ b/gcc/machmode.def
> @@ -196,7 +196,7 @@ RANDOM_MODE (VOID);
> RANDOM_MODE (BLK);
>
> /* Single bit mode used for booleans. */
> -FRACTIONAL_INT_MODE (BI, 1, 1);
> +BOOL_MODE (BI, 1, 1);
>
> /* Basic integer modes. We go up to TI in generic code (128 bits).
> TImode is needed here because the some front ends now genericly
> diff --git a/gcc/rtx-vector-builder.c b/gcc/rtx-vector-builder.c
> index e36aba010a0..55ffe0d5a76 100644
> --- a/gcc/rtx-vector-builder.c
> +++ b/gcc/rtx-vector-builder.c
> @@ -90,8 +90,10 @@ rtx_vector_builder::find_cached_value ()
>
> if (GET_MODE_CLASS (m_mode) == MODE_VECTOR_BOOL)
> {
> - if (elt == const1_rtx || elt == constm1_rtx)
> + if (elt == const1_rtx)
> return CONST1_RTX (m_mode);
> + else if (elt == constm1_rtx)
> + return CONSTM1_RTX (m_mode);
> else if (elt == const0_rtx)
> return CONST0_RTX (m_mode);
> else
> diff --git a/gcc/simplify-rtx.c b/gcc/simplify-rtx.c
> index c36c825f958..532537ea48d 100644
> --- a/gcc/simplify-rtx.c
> +++ b/gcc/simplify-rtx.c
> @@ -6876,12 +6876,13 @@ native_encode_rtx (machine_mode mode, rtx x,
> vec<target_unit> &bytes,
> /* This is the only case in which elements can be smaller than
> a byte. */
> gcc_assert (GET_MODE_CLASS (mode) == MODE_VECTOR_BOOL);
> + auto mask = GET_MODE_MASK (GET_MODE_INNER (mode));
> for (unsigned int i = 0; i < num_bytes; ++i)
> {
> target_unit value = 0;
> for (unsigned int j = 0; j < BITS_PER_UNIT; j += elt_bits)
> {
> - value |= (INTVAL (CONST_VECTOR_ELT (x, elt)) & 1) << j;
> + value |= (INTVAL (CONST_VECTOR_ELT (x, elt)) & mask) << j;
> elt += 1;
> }
> bytes.quick_push (value);
> @@ -7025,9 +7026,8 @@ native_decode_vector_rtx (machine_mode mode,
> const vec<target_unit> &bytes,
> unsigned int bit_index = first_byte * BITS_PER_UNIT + i * elt_bits;
> unsigned int byte_index = bit_index / BITS_PER_UNIT;
> unsigned int lsb = bit_index % BITS_PER_UNIT;
> - builder.quick_push (bytes[byte_index] & (1 << lsb)
> - ? CONST1_RTX (BImode)
> - : CONST0_RTX (BImode));
> + unsigned int value = bytes[byte_index] >> lsb;
> + builder.quick_push (gen_int_mode (value, GET_MODE_INNER
> (mode)));
> }
> }
> else
> @@ -7994,17 +7994,23 @@ test_vector_ops_duplicate (machine_mode
> mode, rtx scalar_reg)
> duplicate, last_par));
>
> /* Test a scalar subreg of a VEC_MERGE of a VEC_DUPLICATE. */
> - rtx vector_reg = make_test_reg (mode);
> - for (unsigned HOST_WIDE_INT i = 0; i < const_nunits; i++)
> + /* Skip this test for vectors of booleans, because offset is in bytes,
> + while vec_merge indices are in elements (usually bits). */
> + if (GET_MODE_CLASS (mode) != MODE_VECTOR_BOOL)
> {
> - if (i >= HOST_BITS_PER_WIDE_INT)
> - break;
> - rtx mask = GEN_INT ((HOST_WIDE_INT_1U << i) | (i + 1));
> - rtx vm = gen_rtx_VEC_MERGE (mode, duplicate, vector_reg, mask);
> - poly_uint64 offset = i * GET_MODE_SIZE (inner_mode);
> - ASSERT_RTX_EQ (scalar_reg,
> - simplify_gen_subreg (inner_mode, vm,
> - mode, offset));
> + rtx vector_reg = make_test_reg (mode);
> + for (unsigned HOST_WIDE_INT i = 0; i < const_nunits; i++)
> + {
> + if (i >= HOST_BITS_PER_WIDE_INT)
> + break;
> + rtx mask = GEN_INT ((HOST_WIDE_INT_1U << i) | (i + 1));
> + rtx vm = gen_rtx_VEC_MERGE (mode, duplicate, vector_reg,
> mask);
> + poly_uint64 offset = i * GET_MODE_SIZE (inner_mode);
> +
> + ASSERT_RTX_EQ (scalar_reg,
> + simplify_gen_subreg (inner_mode, vm,
> + mode, offset));
> + }
> }
> }
>
> diff --git a/gcc/varasm.c b/gcc/varasm.c
> index 76574be191f..5f59b6ace15 100644
> --- a/gcc/varasm.c
> +++ b/gcc/varasm.c
> @@ -4085,6 +4085,7 @@ output_constant_pool_2 (fixed_size_mode mode,
> rtx x, unsigned int align)
> unsigned int elt_bits = GET_MODE_BITSIZE (mode) / nelts;
> unsigned int int_bits = MAX (elt_bits, BITS_PER_UNIT);
> scalar_int_mode int_mode = int_mode_for_size (int_bits, 0).require
> ();
> + unsigned int mask = GET_MODE_MASK (GET_MODE_INNER (mode));
>
> /* Build the constant up one integer at a time. */
> unsigned int elts_per_int = int_bits / elt_bits;
> @@ -4093,8 +4094,10 @@ output_constant_pool_2 (fixed_size_mode
> mode, rtx x, unsigned int align)
> unsigned HOST_WIDE_INT value = 0;
> unsigned int limit = MIN (nelts - i, elts_per_int);
> for (unsigned int j = 0; j < limit; ++j)
> - if (INTVAL (CONST_VECTOR_ELT (x, i + j)) != 0)
> - value |= 1 << (j * elt_bits);
> + {
> + auto elt = INTVAL (CONST_VECTOR_ELT (x, i + j));
> + value |= (elt & mask) << (j * elt_bits);
> + }
> output_constant_pool_2 (int_mode, gen_int_mode (value,
> int_mode),
> i != 0 ? MIN (align, int_bits) : align);
> }
> --
> 2.25.1
