https://gcc.gnu.org/g:04a69e68bb4c6334595f0e7da837ea0944ae0dcd
commit r17-6229-g04a69e68bb4c6334595f0e7da837ea0944ae0dcd Author: Dhruv Chawla <[email protected]> Date: Thu Nov 5 15:43:57 2026 +0530 aarch64: Port NEON widening shifts by an immediate to the pragma-based framework Port the following families of NEON intrinsics to the pragma-based framework: * vshll_n * vshll_high_n Bootstrapped and regtested on aarch64-linux-gnu. Signed-off-by: Dhruv Chawla <[email protected]> gcc/ChangeLog: * config/aarch64/aarch64-builtins.cc (aarch64_types_unsigned_shift_qualifiers, TYPES_USHIFTIMM): Delete. * config/aarch64/aarch64-neon-builtins-base.cc (vshll_n, vshll_high_n): New function bases. * config/aarch64/aarch64-neon-builtins-base.def: Add the corresponding function group. * config/aarch64/aarch64-simd-builtins.def: Delete the sshll_n, ushll_n, sshll2_n and ushll2_n builtins, which no longer have any caller. * config/aarch64/arm_neon.h (vshll_high_n_s16, vshll_high_n_s32, vshll_high_n_s8, vshll_high_n_u16, vshll_high_n_u32, vshll_high_n_u8, vshll_n_s16, vshll_n_s32, vshll_n_s8, vshll_n_u16, vshll_n_u32, vshll_n_u8): Delete. gcc/testsuite/ChangeLog: * gcc.target/aarch64/neon/vshll_n.c: New test. * gcc.target/aarch64/neon/shift_imm_range.c: Likewise. Diff: --- gcc/config/aarch64/aarch64-builtins.cc | 4 - gcc/config/aarch64/aarch64-neon-builtins-base.cc | 3 + gcc/config/aarch64/aarch64-neon-builtins-base.def | 6 ++ gcc/config/aarch64/aarch64-simd-builtins.def | 6 -- gcc/config/aarch64/arm_neon.h | 86 ---------------------- .../gcc.target/aarch64/neon/shift_imm_range.c | 48 ++++++++++++ gcc/testsuite/gcc.target/aarch64/neon/vshll_n.c | 39 ++++++++++ 7 files changed, 96 insertions(+), 96 deletions(-) diff --git a/gcc/config/aarch64/aarch64-builtins.cc b/gcc/config/aarch64/aarch64-builtins.cc index 87e588d19fee..840963be8dab 100644 --- a/gcc/config/aarch64/aarch64-builtins.cc +++ b/gcc/config/aarch64/aarch64-builtins.cc @@ -344,10 +344,6 @@ static enum aarch64_type_qualifiers aarch64_types_fcvt_from_unsigned_qualifiers[SIMD_MAX_BUILTIN_ARGS] = { qualifier_none, qualifier_unsigned, qualifier_immediate }; #define TYPES_FCVTIMM_SUS (aarch64_types_fcvt_from_unsigned_qualifiers) -static enum aarch64_type_qualifiers -aarch64_types_unsigned_shift_qualifiers[SIMD_MAX_BUILTIN_ARGS] - = { qualifier_unsigned, qualifier_unsigned, qualifier_immediate }; -#define TYPES_USHIFTIMM (aarch64_types_unsigned_shift_qualifiers) static enum aarch64_type_qualifiers aarch64_types_load1_qualifiers[SIMD_MAX_BUILTIN_ARGS] diff --git a/gcc/config/aarch64/aarch64-neon-builtins-base.cc b/gcc/config/aarch64/aarch64-neon-builtins-base.cc index cdd46dbeb863..e57015d52df9 100644 --- a/gcc/config/aarch64/aarch64-neon-builtins-base.cc +++ b/gcc/config/aarch64/aarch64-neon-builtins-base.cc @@ -1701,6 +1701,9 @@ NEON_FUNCTION (vqrshrun_n, vqrshrunh_n, vqrshruns_n, vqrshrund_n, NEON_FUNCTION (vqshrun_high_n, gimple_code_for_mode_function, (code_for_aarch64_sqshrun2_n, 1)) NEON_FUNCTION (vqrshrun_high_n, gimple_code_for_mode_function, (code_for_aarch64_sqrshrun2_n, 1)) +// Widening shifts by an immediate +NEON_FUNCTION (vshll_n, gimple_unspec_based_function, (code_for_aarch64_shll_n, UNSPEC_SSHLL, UNSPEC_USHLL)) +NEON_FUNCTION (vshll_high_n, gimple_unspec_based_function, (code_for_aarch64_shll2_n, UNSPEC_SSHLL, UNSPEC_USHLL)) // Bitwise operations NEON_FUNCTION (vand, vandq, gimple_expr, (BIT_AND_EXPR)) diff --git a/gcc/config/aarch64/aarch64-neon-builtins-base.def b/gcc/config/aarch64/aarch64-neon-builtins-base.def index 2684ae281799..981e101cfdb2 100644 --- a/gcc/config/aarch64/aarch64-neon-builtins-base.def +++ b/gcc/config/aarch64/aarch64-neon-builtins-base.def @@ -491,6 +491,12 @@ DEF_NEON_FUNCTION (vqrshruns_n, s_signed, ("shu0,s0,ss32", shift_rig DEF_NEON_FUNCTION (vqrshrund_n, d_signed, ("shu0,s0,ss32", shift_right_narrow<1>)) #undef REQUIRED_EXTENSIONS +// Widening shifts by an immediate +#define REQUIRED_EXTENSIONS nonstreaming_only (AARCH64_FL_SIMD) +DEF_NEON_FUNCTION (vshll_n, bhs_integer, ("Qd0,D0,ss32", shift_left_long<1>)) +DEF_NEON_FUNCTION (vshll_high_n, bhs_integer, ("Qd0,Q0,ss32", shift_left_long<1>)) +#undef REQUIRED_EXTENSIONS + // Bitwise operations (FEAT_SHA3) #define REQUIRED_EXTENSIONS nonstreaming_only (AARCH64_FL_SHA3) DEF_NEON_FUNCTION (vbcaxq, all_integer, ("Q0,Q0,Q0,Q0")) diff --git a/gcc/config/aarch64/aarch64-simd-builtins.def b/gcc/config/aarch64/aarch64-simd-builtins.def index c7c9ac4466b3..0288c4c3e482 100644 --- a/gcc/config/aarch64/aarch64-simd-builtins.def +++ b/gcc/config/aarch64/aarch64-simd-builtins.def @@ -396,12 +396,6 @@ BUILTIN_VQ_HSF (QUADOP_LANE_PAIR, fcmlaq_lane90, 0, FP) BUILTIN_VQ_HSF (QUADOP_LANE_PAIR, fcmlaq_lane180, 0, FP) BUILTIN_VQ_HSF (QUADOP_LANE_PAIR, fcmlaq_lane270, 0, FP) - /* Implemented by aarch64_<sur>shll_n<mode>. */ - BUILTIN_VD_BHSI (SHIFTIMM, sshll_n, 0, DEFAULT) - BUILTIN_VD_BHSI (USHIFTIMM, ushll_n, 0, DEFAULT) - /* Implemented by aarch64_<sur>shll2_n<mode>. */ - BUILTIN_VQW (SHIFTIMM, sshll2_n, 0, DEFAULT) - BUILTIN_VQW (SHIFTIMM, ushll2_n, 0, DEFAULT) /* Implemented by aarch64_xtn2<mode>. */ BUILTIN_VQN (BINOP, xtn2, 0, DEFAULT) diff --git a/gcc/config/aarch64/arm_neon.h b/gcc/config/aarch64/arm_neon.h index 24213304be03..94aca0706c77 100644 --- a/gcc/config/aarch64/arm_neon.h +++ b/gcc/config/aarch64/arm_neon.h @@ -11176,92 +11176,6 @@ vrsqrtsq_f64 (float64x2_t __a, float64x2_t __b) return __builtin_aarch64_rsqrtsv2df (__a, __b); } -/* vshl */ - -__extension__ extern __inline int16x8_t -__attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) -vshll_high_n_s8 (int8x16_t __a, const int __b) -{ - return __builtin_aarch64_sshll2_nv16qi (__a, __b); -} - -__extension__ extern __inline int32x4_t -__attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) -vshll_high_n_s16 (int16x8_t __a, const int __b) -{ - return __builtin_aarch64_sshll2_nv8hi (__a, __b); -} - -__extension__ extern __inline int64x2_t -__attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) -vshll_high_n_s32 (int32x4_t __a, const int __b) -{ - return __builtin_aarch64_sshll2_nv4si (__a, __b); -} - -__extension__ extern __inline uint16x8_t -__attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) -vshll_high_n_u8 (uint8x16_t __a, const int __b) -{ - return (uint16x8_t) __builtin_aarch64_ushll2_nv16qi ((int8x16_t) __a, __b); -} - -__extension__ extern __inline uint32x4_t -__attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) -vshll_high_n_u16 (uint16x8_t __a, const int __b) -{ - return (uint32x4_t) __builtin_aarch64_ushll2_nv8hi ((int16x8_t) __a, __b); -} - -__extension__ extern __inline uint64x2_t -__attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) -vshll_high_n_u32 (uint32x4_t __a, const int __b) -{ - return (uint64x2_t) __builtin_aarch64_ushll2_nv4si ((int32x4_t) __a, __b); -} - -__extension__ extern __inline int16x8_t -__attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) -vshll_n_s8 (int8x8_t __a, const int __b) -{ - return __builtin_aarch64_sshll_nv8qi (__a, __b); -} - -__extension__ extern __inline int32x4_t -__attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) -vshll_n_s16 (int16x4_t __a, const int __b) -{ - return __builtin_aarch64_sshll_nv4hi (__a, __b); -} - -__extension__ extern __inline int64x2_t -__attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) -vshll_n_s32 (int32x2_t __a, const int __b) -{ - return __builtin_aarch64_sshll_nv2si (__a, __b); -} - -__extension__ extern __inline uint16x8_t -__attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) -vshll_n_u8 (uint8x8_t __a, const int __b) -{ - return __builtin_aarch64_ushll_nv8qi_uus (__a, __b); -} - -__extension__ extern __inline uint32x4_t -__attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) -vshll_n_u16 (uint16x4_t __a, const int __b) -{ - return __builtin_aarch64_ushll_nv4hi_uus (__a, __b); -} - -__extension__ extern __inline uint64x2_t -__attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) -vshll_n_u32 (uint32x2_t __a, const int __b) -{ - return __builtin_aarch64_ushll_nv2si_uus (__a, __b); -} - /* vsqadd */ __extension__ extern __inline uint8x8_t diff --git a/gcc/testsuite/gcc.target/aarch64/neon/shift_imm_range.c b/gcc/testsuite/gcc.target/aarch64/neon/shift_imm_range.c new file mode 100644 index 000000000000..c41bb56b4a7e --- /dev/null +++ b/gcc/testsuite/gcc.target/aarch64/neon/shift_imm_range.c @@ -0,0 +1,48 @@ +/* { dg-do compile } */ + +#include "arm_neon_test.h" + +void +test_left (int8x8_t a, int16x8_t b) +{ + /* Left shifts take 0 to one less than the element width. */ + vshl_n_s8 (a, 0); + vshl_n_s8 (a, 7); + vshl_n_s8 (a, 8); /* { dg-error {passing 8 to argument 2 of 'vshl_n_s8', which expects a value in the range \[0, 7\]} } */ + vqshl_n_s8 (a, 8); /* { dg-error {passing 8 to argument 2 of 'vqshl_n_s8', which expects a value in the range \[0, 7\]} } */ + vqshlu_n_s8 (a, 8); /* { dg-error {passing 8 to argument 2 of 'vqshlu_n_s8', which expects a value in the range \[0, 7\]} } */ + vsli_n_s8 (a, a, 8); /* { dg-error {passing 8 to argument 3 of 'vsli_n_s8', which expects a value in the range \[0, 7\]} } */ +} + +void +test_right (int8x8_t a, int16x8_t b) +{ + /* Right shifts take 1 to the element width. */ + vshr_n_s8 (a, 1); + vshr_n_s8 (a, 8); + vshr_n_s8 (a, 0); /* { dg-error {passing 0 to argument 2 of 'vshr_n_s8', which expects a value in the range \[1, 8\]} } */ + vshr_n_s8 (a, 9); /* { dg-error {passing 9 to argument 2 of 'vshr_n_s8', which expects a value in the range \[1, 8\]} } */ + vrshr_n_s8 (a, 0); /* { dg-error {passing 0 to argument 2 of 'vrshr_n_s8', which expects a value in the range \[1, 8\]} } */ + vsra_n_s8 (a, a, 0); /* { dg-error {passing 0 to argument 3 of 'vsra_n_s8', which expects a value in the range \[1, 8\]} } */ + vsri_n_s8 (a, a, 0); /* { dg-error {passing 0 to argument 3 of 'vsri_n_s8', which expects a value in the range \[1, 8\]} } */ +} + +void +test_narrow (int16x8_t a) +{ + /* Narrowing shifts take 1 to the width of the result elements, which is + half the width of the argument's. */ + vshrn_n_s16 (a, 8); + vshrn_n_s16 (a, 9); /* { dg-error {passing 9 to argument 2 of 'vshrn_n_s16', which expects a value in the range \[1, 8\]} } */ + vqshrn_n_s16 (a, 0); /* { dg-error {passing 0 to argument 2 of 'vqshrn_n_s16', which expects a value in the range \[1, 8\]} } */ + vqshrun_n_s16 (a, 9); /* { dg-error {passing 9 to argument 2 of 'vqshrun_n_s16', which expects a value in the range \[1, 8\]} } */ +} + +void +test_long (int8x8_t a, int8x16_t b) +{ + /* Widening shifts may name the element width itself, which selects SHLL. */ + vshll_n_s8 (a, 8); + vshll_n_s8 (a, 9); /* { dg-error {passing 9 to argument 2 of 'vshll_n_s8', which expects a value in the range \[0, 8\]} } */ + vshll_high_n_s8 (b, 9); /* { dg-error {passing 9 to argument 2 of 'vshll_high_n_s8', which expects a value in the range \[0, 8\]} } */ +} diff --git a/gcc/testsuite/gcc.target/aarch64/neon/vshll_n.c b/gcc/testsuite/gcc.target/aarch64/neon/vshll_n.c new file mode 100644 index 000000000000..ddd5a7367e36 --- /dev/null +++ b/gcc/testsuite/gcc.target/aarch64/neon/vshll_n.c @@ -0,0 +1,39 @@ +/* { dg-do compile } */ +/* { dg-final { check-function-bodies "**" "" } } */ + +#include "arm_neon_test.h" + +/* +** test_vshll_n_s8_3: +** sshll v0\.8h, v0\.8b, 3 +** ret +*/ +TEST_UNARY_N (vshll_n_s8, int16x8_t, int8x8_t, 3) + +/* +** test_vshll_n_u16_0: +** uxtl v0\.4s, v0\.4h +** ret +*/ +TEST_UNARY_N (vshll_n_u16, uint32x4_t, uint16x4_t, 0) + +/* +** test_vshll_n_u8_8: +** shll v0\.8h, v0\.8b, 8 +** ret +*/ +TEST_UNARY_N (vshll_n_u8, uint16x8_t, uint8x8_t, 8) + +/* +** test_vshll_high_n_s16_5: +** sshll2 v0\.4s, v0\.8h, 5 +** ret +*/ +TEST_UNARY_N (vshll_high_n_s16, int32x4_t, int16x8_t, 5) + +/* +** test_vshll_high_n_u32_32: +** shll2 v0\.2d, v0\.4s, 32 +** ret +*/ +TEST_UNARY_N (vshll_high_n_u32, uint64x2_t, uint32x4_t, 32)
