https://gcc.gnu.org/g:7d33a477b794f77d18d32c5a65256400291276cf
commit r17-6221-g7d33a477b794f77d18d32c5a65256400291276cf Author: Karl Meakin <[email protected]> Date: Thu Nov 12 17:12:08 2026 +0000 aarch64: Port NEON FP min/max reduction intrinsics to pragma-based framework The floating point variants of `vminv` and `vmaxv` were not ported to the pragma-based framework when the integer variants were, because the semantics AArch64's `fminv`/`fmaxv` instructions don't match the semantics of GCC's `IFN_REDUC_MIN`/`IFN_REDUC_MAX`. This commit ports them to the pragma framework, by expanding them to `UNSPEC_FMINV` and `UNSPEC_FMAXV`. There are no test changes, as tests for FP variants of vminv/vmaxv were already included in vminv.c and vmaxv.c. The absence of any changes to the tests does show that there were no changes in codegen (ie no regressions). gcc/ChangeLog: * config/aarch64/aarch64-builtins.cc (aarch64_general_gimple_fold_builtin): Delete cases for handling `reduc_{u,s}{min,max}_scal_`. * config/aarch64/aarch64-neon-builtins-base.cc (struct gimple_reduce_minmax): New struct. * config/aarch64/aarch64-neon-builtins-base.def (vminv, vminvq, vmaxv, vmaxvq): New intrinsic function declarations. * config/aarch64/aarch64-simd-builtins.def: Delete `reduc_{us}{min,max}{,_nan}_scal_` builtin functions. * config/aarch64/aarch64-simd.md (reduc_<optab>_scal_<mode>): Rename to ... (@reduc_<optab>_scal_<mode>): ...this, so that it is callable from `gimple_reduce_minmax::expand ()`. * config/aarch64/arm_neon.h (vpmaxqd_f64, vpmaxs_f32, vpmaxnmqd_f64, vpmaxnms_f32, vpminqd_f64, vpmins_f32, vpminnmqd_f64): Call the intrinsic (v{min,max}{,nm}v) functions rather than the now-deleted `__builtin_aarch64_reduc_s{min,max}_scal` functions. (vmaxv_f32, vmaxvq_f32, vmaxvq_f64, vminv_f32, vminvq_f32, vminvq_f64, vmaxv_f16, vmaxvq_f16, vminv_f16, vminvq_f16): Delete intrinsic function definitions. Diff: --- gcc/config/aarch64/aarch64-builtins.cc | 12 --- gcc/config/aarch64/aarch64-neon-builtins-base.cc | 32 +++++++- gcc/config/aarch64/aarch64-neon-builtins-base.def | 12 ++- gcc/config/aarch64/aarch64-simd-builtins.def | 8 -- gcc/config/aarch64/aarch64-simd.md | 2 +- gcc/config/aarch64/arm_neon.h | 92 ++--------------------- 6 files changed, 48 insertions(+), 110 deletions(-) diff --git a/gcc/config/aarch64/aarch64-builtins.cc b/gcc/config/aarch64/aarch64-builtins.cc index 1d8baaef5c41..337ecb1088a6 100644 --- a/gcc/config/aarch64/aarch64-builtins.cc +++ b/gcc/config/aarch64/aarch64-builtins.cc @@ -5268,18 +5268,6 @@ aarch64_general_gimple_fold_builtin (unsigned int fcode, gcall *stmt, break; } - BUILTIN_VDQIF (UNOP, reduc_smax_scal_, 10, ALL) - BUILTIN_VDQ_BHSI (UNOPU, reduc_umax_scal_, 10, ALL) - new_stmt = gimple_build_call_internal (IFN_REDUC_MAX, - 1, args[0]); - gimple_call_set_lhs (new_stmt, gimple_call_lhs (stmt)); - break; - BUILTIN_VDQIF (UNOP, reduc_smin_scal_, 10, ALL) - BUILTIN_VDQ_BHSI (UNOPU, reduc_umin_scal_, 10, ALL) - new_stmt = gimple_build_call_internal (IFN_REDUC_MIN, - 1, args[0]); - gimple_call_set_lhs (new_stmt, gimple_call_lhs (stmt)); - break; BUILTIN_VSDQ_I_DI (BINOP, ashl, 3, DEFAULT) if (TREE_CODE (args[1]) == INTEGER_CST && wi::ltu_p (wi::to_wide (args[1]), element_precision (args[0]))) diff --git a/gcc/config/aarch64/aarch64-neon-builtins-base.cc b/gcc/config/aarch64/aarch64-neon-builtins-base.cc index f7b2843d01ba..9b77445334b5 100644 --- a/gcc/config/aarch64/aarch64-neon-builtins-base.cc +++ b/gcc/config/aarch64/aarch64-neon-builtins-base.cc @@ -189,7 +189,7 @@ build_var (tree type) Handles singleton vector arguments specially: they are cast to their inner element type before the function is called, and then the return value is cast back to the expected singleton vector type. */ -tree +static tree build_ifn_call (gimple_folder &f, internal_fn ifn, tree ret_type, vec<tree> &args) { @@ -1291,6 +1291,32 @@ struct gimple_minmax : public gimple_function_base } }; +template <bool min_p> +struct gimple_reduce_minmax : public gimple_function_base +{ + rtx expand (function_expander &e) const override + { + auto arg_mode = e.args[0]->mode; + auto unspec = min_p ? UNSPEC_FMINV : UNSPEC_FMAXV; + auto icode = code_for_reduc_scal (unspec, arg_mode); + gcc_assert (icode != CODE_FOR_nothing); + return e.use_exact_insn (icode); + } + + gimple *fold (gimple_folder &f) const override + { + auto arg = gimple_call_arg (f.call, 0); + auto ret_type = TREE_TYPE (f.lhs); + if (FLOAT_TYPE_P (ret_type)) + return nullptr; + + auto_vec<tree> args{}; + args.safe_push (arg); + auto ifn = min_p ? IFN_REDUC_MIN : IFN_REDUC_MAX; + return gimple_build_assign (f.lhs, build_ifn_call (f, ifn, ret_type, args)); + } +}; + // Reinterpret NEON_FUNCTION (vreinterpret, vreinterpretq, gimple_reinterpret,) @@ -1363,8 +1389,8 @@ NEON_FUNCTION (vminnm, vminnmq, gimple_ifn, (IFN_FMIN)) // Reductions NEON_FUNCTION (vaddv, vaddvq, gimple_ifn, (IFN_REDUC_PLUS)) -NEON_FUNCTION (vmaxv, vmaxvq, gimple_ifn, (IFN_REDUC_MAX)) -NEON_FUNCTION (vminv, vminvq, gimple_ifn, (IFN_REDUC_MIN)) +NEON_FUNCTION (vmaxv, vmaxvq, gimple_reduce_minmax<false>,) +NEON_FUNCTION (vminv, vminvq, gimple_reduce_minmax<true>,) NEON_FUNCTION (vmaxnmv, vmaxnmvq, gimple_ifn, (IFN_REDUC_FMAX)) NEON_FUNCTION (vminnmv, vminnmvq, gimple_ifn, (IFN_REDUC_FMIN)) diff --git a/gcc/config/aarch64/aarch64-neon-builtins-base.def b/gcc/config/aarch64/aarch64-neon-builtins-base.def index 9e9dea1064c9..5b05d2f40348 100644 --- a/gcc/config/aarch64/aarch64-neon-builtins-base.def +++ b/gcc/config/aarch64/aarch64-neon-builtins-base.def @@ -211,8 +211,12 @@ DEF_NEON_FUNCTION (vaddv, s_float_bhs_integer, ("s0,D0")) DEF_NEON_FUNCTION (vaddvq, sd_float_all_integer, ("s0,Q0")) DEF_NEON_FUNCTION (vmaxv, bhs_integer, ("s0,D0")) DEF_NEON_FUNCTION (vmaxvq, bhs_integer, ("s0,Q0")) -DEF_NEON_FUNCTION (vminv, bhs_integer, ("s0,D0")) -DEF_NEON_FUNCTION (vminvq, bhs_integer, ("s0,Q0")) +DEF_NEON_FUNCTION (vmaxv, sd_float, ("s0,D0")) +DEF_NEON_FUNCTION (vmaxvq, sd_float, ("s0,Q0")) +DEF_NEON_FUNCTION (vminv, bhs_integer, ("s0,D0")) +DEF_NEON_FUNCTION (vminvq, bhs_integer, ("s0,Q0")) +DEF_NEON_FUNCTION (vminv, sd_float, ("s0,D0")) +DEF_NEON_FUNCTION (vminvq, sd_float, ("s0,Q0")) DEF_NEON_FUNCTION (vmaxnmv, s_float, ("s0,D0")) DEF_NEON_FUNCTION (vmaxnmvq, sd_float, ("s0,Q0")) DEF_NEON_FUNCTION (vminnmv, s_float, ("s0,D0")) @@ -221,6 +225,10 @@ DEF_NEON_FUNCTION (vminnmvq, sd_float, ("s0,Q0")) // Reductions (FP16) #define REQUIRED_EXTENSIONS nonstreaming_only (AARCH64_FL_SIMD | AARCH64_FL_F16) +DEF_NEON_FUNCTION (vmaxv, h_float, ("s0,D0")) +DEF_NEON_FUNCTION (vmaxvq, h_float, ("s0,Q0")) +DEF_NEON_FUNCTION (vminv, h_float, ("s0,D0")) +DEF_NEON_FUNCTION (vminvq, h_float, ("s0,Q0")) DEF_NEON_FUNCTION (vmaxnmv, h_float, ("s0,D0")) DEF_NEON_FUNCTION (vmaxnmvq, h_float, ("s0,Q0")) DEF_NEON_FUNCTION (vminnmv, h_float, ("s0,D0")) diff --git a/gcc/config/aarch64/aarch64-simd-builtins.def b/gcc/config/aarch64/aarch64-simd-builtins.def index caef2a5bb1c8..aa7db1005191 100644 --- a/gcc/config/aarch64/aarch64-simd-builtins.def +++ b/gcc/config/aarch64/aarch64-simd-builtins.def @@ -481,14 +481,6 @@ BUILTIN_VALL (UNOP, reduc_plus_scal_, 10, DEFAULT) BUILTIN_VDQ_I (UNOPU, reduc_plus_scal_, 10, DEFAULT) - /* Implemented by reduc_<maxmin_uns>_scal_<mode> (producing scalar). */ - BUILTIN_VDQIF_F16 (UNOP, reduc_smax_scal_, 10, DEFAULT) - BUILTIN_VDQIF_F16 (UNOP, reduc_smin_scal_, 10, DEFAULT) - BUILTIN_VDQ_BHSI (UNOPU, reduc_umax_scal_, 10, DEFAULT) - BUILTIN_VDQ_BHSI (UNOPU, reduc_umin_scal_, 10, DEFAULT) - BUILTIN_VHSDF (UNOP, reduc_smax_nan_scal_, 10, DEFAULT) - BUILTIN_VHSDF (UNOP, reduc_smin_nan_scal_, 10, DEFAULT) - /* Implemented by <fmaxmin><mode>3. */ VAR1 (BINOP, fmax, 3, FP, hf) VAR1 (BINOP, fmin, 3, FP, hf) diff --git a/gcc/config/aarch64/aarch64-simd.md b/gcc/config/aarch64/aarch64-simd.md index 8f626a6d42e9..b76e01d81db9 100644 --- a/gcc/config/aarch64/aarch64-simd.md +++ b/gcc/config/aarch64/aarch64-simd.md @@ -4175,7 +4175,7 @@ ;; Template for outputting a scalar, so we can create __builtins which can be ;; gimple_fold'd to the IFN_REDUC_(MAX|MIN) function. (This is FP smax/smin). -(define_expand "reduc_<optab>_scal_<mode>" +(define_expand "@reduc_<optab>_scal_<mode>" [(match_operand:<VEL> 0 "register_operand") (unspec:<VEL> [(match_operand:VHSDF 1 "register_operand")] FMAXMINV)] diff --git a/gcc/config/aarch64/arm_neon.h b/gcc/config/aarch64/arm_neon.h index 0d885743413b..b3a4080bb9ec 100644 --- a/gcc/config/aarch64/arm_neon.h +++ b/gcc/config/aarch64/arm_neon.h @@ -10204,14 +10204,14 @@ __extension__ extern __inline float64_t __attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) vpmaxqd_f64 (float64x2_t __a) { - return __builtin_aarch64_reduc_smax_nan_scal_v2df (__a); + return vmaxvq_f64 (__a); } __extension__ extern __inline float32_t __attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) vpmaxs_f32 (float32x2_t __a) { - return __builtin_aarch64_reduc_smax_nan_scal_v2sf (__a); + return vmaxv_f32 (__a); } /* vpmaxnm */ @@ -10241,14 +10241,14 @@ __extension__ extern __inline float64_t __attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) vpmaxnmqd_f64 (float64x2_t __a) { - return __builtin_aarch64_reduc_smax_scal_v2df (__a); + return vmaxnmvq_f64 (__a); } __extension__ extern __inline float32_t __attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) vpmaxnms_f32 (float32x2_t __a) { - return __builtin_aarch64_reduc_smax_scal_v2sf (__a); + return vmaxnmv_f32 (__a); } /* vpmin */ @@ -10368,14 +10368,14 @@ __extension__ extern __inline float64_t __attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) vpminqd_f64 (float64x2_t __a) { - return __builtin_aarch64_reduc_smin_nan_scal_v2df (__a); + return vminvq_f64 (__a); } __extension__ extern __inline float32_t __attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) vpmins_f32 (float32x2_t __a) { - return __builtin_aarch64_reduc_smin_nan_scal_v2sf (__a); + return vminv_f32 (__a); } /* vpminnm */ @@ -10405,60 +10405,14 @@ __extension__ extern __inline float64_t __attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) vpminnmqd_f64 (float64x2_t __a) { - return __builtin_aarch64_reduc_smin_scal_v2df (__a); + return vminnmvq_f64 (__a); } __extension__ extern __inline float32_t __attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) vpminnms_f32 (float32x2_t __a) { - return __builtin_aarch64_reduc_smin_scal_v2sf (__a); -} - -/* vmaxv */ - -__extension__ extern __inline float32_t -__attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) -vmaxv_f32 (float32x2_t __a) -{ - return __builtin_aarch64_reduc_smax_nan_scal_v2sf (__a); -} - -__extension__ extern __inline float32_t -__attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) -vmaxvq_f32 (float32x4_t __a) -{ - return __builtin_aarch64_reduc_smax_nan_scal_v4sf (__a); -} - -__extension__ extern __inline float64_t -__attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) -vmaxvq_f64 (float64x2_t __a) -{ - return __builtin_aarch64_reduc_smax_nan_scal_v2df (__a); -} - -/* vminv */ - -__extension__ extern __inline float32_t -__attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) -vminv_f32 (float32x2_t __a) -{ - return __builtin_aarch64_reduc_smin_nan_scal_v2sf (__a); -} - -__extension__ extern __inline float32_t -__attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) -vminvq_f32 (float32x4_t __a) -{ - return __builtin_aarch64_reduc_smin_nan_scal_v4sf (__a); -} - -__extension__ extern __inline float64_t -__attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) -vminvq_f64 (float64x2_t __a) -{ - return __builtin_aarch64_reduc_smin_nan_scal_v2df (__a); + return vminnmv_f32 (__a); } /* vpadd */ @@ -16227,36 +16181,6 @@ vmulxq_n_f16 (float16x8_t __a, float16_t __b) return vmulxq_f16 (__a, vdupq_n_f16 (__b)); } -/* ARMv8.2-A FP16 reduction vector intrinsics. */ - -__extension__ extern __inline float16_t -__attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) -vmaxv_f16 (float16x4_t __a) -{ - return __builtin_aarch64_reduc_smax_nan_scal_v4hf (__a); -} - -__extension__ extern __inline float16_t -__attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) -vmaxvq_f16 (float16x8_t __a) -{ - return __builtin_aarch64_reduc_smax_nan_scal_v8hf (__a); -} - -__extension__ extern __inline float16_t -__attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) -vminv_f16 (float16x4_t __a) -{ - return __builtin_aarch64_reduc_smin_nan_scal_v4hf (__a); -} - -__extension__ extern __inline float16_t -__attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) -vminvq_f16 (float16x8_t __a) -{ - return __builtin_aarch64_reduc_smin_nan_scal_v8hf (__a); -} - #pragma GCC pop_options /* AdvSIMD Dot Product intrinsics. */
