Skip to content

Commit 2d6e6df

Browse files
SpencerAbsonSpencerAbson
committed
Merge new patterns into existing Neon_INS_elt_patterns work
1 parent 5fce5b4 commit 2d6e6df

File tree

2 files changed

+84
-77
lines changed

2 files changed

+84
-77
lines changed

‎llvm/lib/Target/AArch64/AArch64InstrInfo.td‎

Lines changed: 84 additions & 27 deletions
Original file line numberDiff line numberDiff line change
@@ -7222,44 +7222,101 @@ def : Pat<(v2i64 (int_aarch64_neon_vcopy_lane
72227222
V128:$Vd, VectorIndexD:$idx, V128:$Vs, VectorIndexD:$idx2)
72237223
)>;
72247224

7225-
multiclass Neon_INS_elt_pattern<ValueType VT128, ValueType VT64,
7226-
ValueType VTScal, Instruction INS> {
7227-
def : Pat<(VT128 (vector_insert V128:$src,
7228-
(VTScal (vector_extract (VT128 V128:$Rn), (i64 imm:$Immn))),
7225+
// Insert an extracted vector element into a 128-bit Neon vector
7226+
multiclass Neon_INS_elt_pattern_v128<ValueType VT128, ValueType VT64, ValueType VTSVE,
7227+
ValueType VTScal, Operand ExIdxTy, Instruction INS> {
7228+
// Extracting from the lower 128-bits of an SVE vector
7229+
def : Pat<(VT128 (vector_insert VT128:$Rn,
7230+
(VTScal (vector_extract VTSVE:$Rm, (i64 ExIdxTy:$Immn))),
7231+
(i64 imm:$Immd))),
7232+
(INS VT128:$Rn, imm:$Immd, (VT128 (EXTRACT_SUBREG VTSVE:$Rm, zsub)), ExIdxTy:$Immn)>;
7233+
7234+
// Extracting from another Neon vector
7235+
def : Pat<(VT128 (vector_insert V128:$Rn,
7236+
(VTScal (vector_extract (VT128 V128:$Rm), (i64 imm:$Immn))),
72297237
(i64 imm:$Immd))),
7230-
(INS V128:$src, imm:$Immd, V128:$Rn, imm:$Immn)>;
7231-
7232-
def : Pat<(VT128 (vector_insert V128:$src,
7233-
(VTScal (vector_extract (VT64 V64:$Rn), (i64 imm:$Immn))),
7234-
(i64 imm:$Immd))),
7235-
(INS V128:$src, imm:$Immd,
7236-
(SUBREG_TO_REG (i64 0), V64:$Rn, dsub), imm:$Immn)>;
7238+
(INS V128:$Rn, imm:$Immd, V128:$Rm, imm:$Immn)>;
7239+
7240+
def : Pat<(VT128 (vector_insert V128:$Rn,
7241+
(VTScal (vector_extract (VT64 V64:$Rm), (i64 imm:$Immn))),
7242+
(i64 imm:$Immd))),
7243+
(INS V128:$Rn, imm:$Immd,
7244+
(SUBREG_TO_REG (i64 0), V64:$Rm, dsub), imm:$Immn)>;
7245+
}
7246+
7247+
// Insert an extracted vector element into a 64-bit Neon vector
7248+
multiclass Neon_INS_elt_pattern_v64<ValueType VT128, ValueType VT64, ValueType VTSVE,
7249+
ValueType VTScal, Operand ExIdxTy, Instruction INS> {
7250+
// Extracting from the lower 128-bits of an SVE vector
7251+
def : Pat<(VT64 (vector_insert VT64:$Rn,
7252+
(VTScal (vector_extract VTSVE:$Rm, (i64 ExIdxTy:$Immn))),
7253+
(i64 imm:$Immd))),
7254+
(EXTRACT_SUBREG (INS (SUBREG_TO_REG (i64 0), VT64:$Rn, dsub), imm:$Immd,
7255+
(VT128 (EXTRACT_SUBREG VTSVE:$Rm, zsub)), ExIdxTy:$Immn),
7256+
dsub)>;
72377257

7238-
def : Pat<(VT64 (vector_insert V64:$src,
7239-
(VTScal (vector_extract (VT128 V128:$Rn), (i64 imm:$Immn))),
7258+
// Extracting from another Neon vector
7259+
def : Pat<(VT64 (vector_insert V64:$Rn,
7260+
(VTScal (vector_extract (VT128 V128:$Rm), (i64 imm:$Immn))),
72407261
(i64 imm:$Immd))),
7241-
(EXTRACT_SUBREG (INS (SUBREG_TO_REG (i64 0), V64:$src, dsub),
7242-
imm:$Immd, V128:$Rn, imm:$Immn),
7262+
(EXTRACT_SUBREG (INS (SUBREG_TO_REG (i64 0), V64:$Rn, dsub),
7263+
imm:$Immd, V128:$Rm, imm:$Immn),
72437264
dsub)>;
72447265

7245-
def : Pat<(VT64 (vector_insert V64:$src,
7246-
(VTScal (vector_extract (VT64 V64:$Rn), (i64 imm:$Immn))),
7266+
def : Pat<(VT64 (vector_insert V64:$Rn,
7267+
(VTScal (vector_extract (VT64 V64:$Rm), (i64 imm:$Immn))),
72477268
(i64 imm:$Immd))),
72487269
(EXTRACT_SUBREG
7249-
(INS (SUBREG_TO_REG (i64 0), V64:$src, dsub), imm:$Immd,
7250-
(SUBREG_TO_REG (i64 0), V64:$Rn, dsub), imm:$Immn),
7270+
(INS (SUBREG_TO_REG (i64 0), V64:$Rn, dsub), imm:$Immd,
7271+
(SUBREG_TO_REG (i64 0), V64:$Rm, dsub), imm:$Immn),
7272+
dsub)>;
7273+
}
7274+
7275+
// Special case for <1 x double/i64> - insertion may be vector_from_scalar or
7276+
// (vector_insert (vec) 0).
7277+
multiclass Neon_INS_elt_pattern_v64d<ValueType VT128, ValueType VT64, ValueType VTSVE,
7278+
ValueType VTScal> {
7279+
// Extracting from the lower 128-bits of an SVE vector
7280+
def : Pat<(VT64 (vec_ins_or_scal_vec
7281+
(VTScal (vector_extract VTSVE:$Rm, VectorIndexD:$Immn)))),
7282+
(EXTRACT_SUBREG
7283+
(INSvi64lane (IMPLICIT_DEF), 0, (VT128 (EXTRACT_SUBREG VTSVE:$Rm, zsub)),
7284+
VectorIndexD:$Immn),
72517285
dsub)>;
7286+
7287+
def : Pat<(VT64 (vec_ins_or_scal_vec
7288+
(VTScal (vector_extract (VT128 V128:$Rm), (i64 imm:$Immn))))),
7289+
(EXTRACT_SUBREG
7290+
(INSvi64lane (IMPLICIT_DEF), (i64 0), V128:$Rm, imm:$Immn),
7291+
dsub)>;
7292+
7293+
// Extracting from another NEON vector
7294+
def : Pat<(VT64 (vec_ins_or_scal_vec
7295+
(VTScal (vector_extract (VT64 V64:$Rm), (i64 imm:$Immn))))),
7296+
(EXTRACT_SUBREG
7297+
(INSvi64lane (IMPLICIT_DEF), (i64 0),
7298+
(SUBREG_TO_REG (i64 0), V64:$Rm, dsub), imm:$Immn),
7299+
dsub)>;
7300+
}
7301+
7302+
multiclass Neon_INS_elt_pattern<ValueType VT128, ValueType VT64, ValueType SVESrcVT,
7303+
ValueType VTScal, Operand ExIdxTy, Instruction INS> {
7304+
defm : Neon_INS_elt_pattern_v64<VT128, VT64, SVESrcVT, VTScal, ExIdxTy, INS>;
7305+
defm : Neon_INS_elt_pattern_v128<VT128, VT64, SVESrcVT, VTScal, ExIdxTy, INS>;
72527306
}
72537307

7254-
defm : Neon_INS_elt_pattern<v8f16, v4f16, f16, INSvi16lane>;
7255-
defm : Neon_INS_elt_pattern<v8bf16, v4bf16, bf16, INSvi16lane>;
7256-
defm : Neon_INS_elt_pattern<v4f32, v2f32, f32, INSvi32lane>;
7257-
defm : Neon_INS_elt_pattern<v2f64, v1f64, f64, INSvi64lane>;
7308+
defm : Neon_INS_elt_pattern<v4f32, v2f32, nxv4f32, f32, VectorIndexS, INSvi32lane>;
7309+
defm : Neon_INS_elt_pattern<v8f16, v4f16, nxv8f16, f16, VectorIndexH, INSvi16lane>;
7310+
defm : Neon_INS_elt_pattern<v8bf16, v4bf16, nxv8bf16, bf16, VectorIndexH, INSvi16lane>;
7311+
defm : Neon_INS_elt_pattern<v4f32, v2f32, nxv4f32, f32, VectorIndexS, INSvi32lane>;
7312+
defm : Neon_INS_elt_pattern<v16i8, v8i8, nxv16i8, i32, VectorIndexB, INSvi8lane>;
7313+
defm : Neon_INS_elt_pattern<v8i16, v4i16, nxv8i16, i32, VectorIndexH, INSvi16lane>;
7314+
defm : Neon_INS_elt_pattern<v4i32, v2i32, nxv4i32, i32, VectorIndexS, INSvi32lane>;
72587315

7259-
defm : Neon_INS_elt_pattern<v16i8, v8i8, i32, INSvi8lane>;
7260-
defm : Neon_INS_elt_pattern<v8i16, v4i16, i32, INSvi16lane>;
7261-
defm : Neon_INS_elt_pattern<v4i32, v2i32, i32, INSvi32lane>;
7262-
defm : Neon_INS_elt_pattern<v2i64, v1i64, i64, INSvi64lane>;
7316+
defm : Neon_INS_elt_pattern_v128<v2f64, v1f64, nxv2f64, f64, VectorIndexD, INSvi64lane>;
7317+
defm : Neon_INS_elt_pattern_v64d<v2f64, v1f64, nxv2f64, f64>;
7318+
defm : Neon_INS_elt_pattern_v128<v2i64, v1i64, nxv2i64, i64, VectorIndexD, INSvi64lane>;
7319+
defm : Neon_INS_elt_pattern_v64d<v2i64, v1i64, nxv2i64, i64>;
72637320

72647321
// Insert from bitcast
72657322
// vector_insert(bitcast(f32 src), n, lane) -> INSvi32lane(src, lane, INSERT_SUBREG(-, n), 0)

‎llvm/lib/Target/AArch64/AArch64SVEInstrInfo.td‎

Lines changed: 0 additions & 50 deletions
Original file line numberDiff line numberDiff line change
@@ -3367,47 +3367,6 @@ let Predicates = [HasSVEorSME] in {
33673367
(UMOVvi64 (v2i64 (EXTRACT_SUBREG ZPR:$vec, zsub)), VectorIndexD:$index)>;
33683368
} // End HasNEON
33693369

3370-
// Use INS (element) when moving an element from the bottom 128-bits of an SVE type to a NEON vector.
3371-
multiclass Neon_ins_sve_elt_pattern<ValueType NeonTy, ValueType NeonQTy, ValueType SVETy, ValueType ScalTy,
3372-
Operand IdxTy, Operand NarrowIdxTy, Instruction INS> {
3373-
// Insert into 128-bit NEON type from lowest 128-bits of SVE type
3374-
def : Pat<(NeonQTy (vector_insert V128:$src,
3375-
(ScalTy (vector_extract SVETy:$Rn, IdxTy:$idx_extract)),
3376-
(IdxTy:$idx_insert))),
3377-
(INS V128:$src, IdxTy:$idx_insert,
3378-
(NeonQTy (EXTRACT_SUBREG SVETy:$Rn, zsub)), IdxTy:$idx_extract)>;
3379-
3380-
// Insert into 64-bit NEON type from lowest 128-bits of SVE type
3381-
def : Pat<(NeonTy (vector_insert V64:$src,
3382-
(ScalTy (vector_extract SVETy:$Rn, IdxTy:$idx_extract)),
3383-
(NarrowIdxTy:$idx_insert))),
3384-
(EXTRACT_SUBREG
3385-
(INS
3386-
(INSERT_SUBREG (NeonQTy (IMPLICIT_DEF)), V64:$src, dsub), NarrowIdxTy:$idx_insert,
3387-
(NeonQTy (EXTRACT_SUBREG SVETy:$Rn, zsub)), IdxTy:$idx_extract),
3388-
dsub)>;
3389-
}
3390-
3391-
// Inserting into <1 x double/i64> will just create a new vector from the scalar value.
3392-
multiclass Neon_ins_64b_sve_elt_pattern<ValueType NeonTy, ValueType NeonQTy, ValueType SVETy,
3393-
ValueType ScalTy> {
3394-
// Insert into 128-bit NEON type from lowest 128-bits of SVE type
3395-
def : Pat<(NeonQTy (vector_insert V128:$src,
3396-
(ScalTy (vector_extract SVETy:$Rn, VectorIndexD:$idx_extract)),
3397-
(VectorIndexD:$idx_insert))),
3398-
(INSvi64lane
3399-
V128:$src, VectorIndexD:$idx_insert, (NeonQTy (EXTRACT_SUBREG SVETy:$Rn, zsub)),
3400-
VectorIndexD:$idx_extract)>;
3401-
3402-
// Insert into 64-bit NEON type from lowest 128-bits of SVE type
3403-
def : Pat<(NeonTy (scalar_to_vector
3404-
(ScalTy (vector_extract SVETy:$Rn, VectorIndexD:$idx_extract)))),
3405-
(EXTRACT_SUBREG
3406-
(INSvi64lane (IMPLICIT_DEF), 0, (NeonQTy (EXTRACT_SUBREG SVETy:$Rn, zsub)),
3407-
VectorIndexD:$idx_extract),
3408-
dsub)>;
3409-
}
3410-
34113370
let Predicates = [HasNEON] in {
34123371
def : Pat<(sext_inreg (vector_extract nxv16i8:$vec, VectorIndexB:$index), i8),
34133372
(SMOVvi8to32 (v16i8 (EXTRACT_SUBREG ZPR:$vec, zsub)), VectorIndexB:$index)>;
@@ -3421,15 +3380,6 @@ let Predicates = [HasSVEorSME] in {
34213380

34223381
def : Pat<(sext (i32 (vector_extract nxv4i32:$vec, VectorIndexS:$index))),
34233382
(SMOVvi32to64 (v4i32 (EXTRACT_SUBREG ZPR:$vec, zsub)), VectorIndexS:$index)>;
3424-
3425-
defm : Neon_ins_sve_elt_pattern<v8i8, v16i8, nxv16i8, i32, VectorIndexB, VectorIndexH, INSvi8lane>;
3426-
defm : Neon_ins_sve_elt_pattern<v4f16, v8f16, nxv8f16, f16, VectorIndexH, VectorIndexS, INSvi16lane>;
3427-
defm : Neon_ins_sve_elt_pattern<v4bf16, v8bf16, nxv8bf16, bf16, VectorIndexH, VectorIndexS, INSvi16lane>;
3428-
defm : Neon_ins_sve_elt_pattern<v4i16, v8i16, nxv8i16, i32, VectorIndexH, VectorIndexS, INSvi16lane>;
3429-
defm : Neon_ins_sve_elt_pattern<v2f32, v4f32, nxv4f32, f32, VectorIndexS, VectorIndexD, INSvi32lane>;
3430-
defm : Neon_ins_sve_elt_pattern<v2i32, v4i32, nxv4i32, i32, VectorIndexS, VectorIndexD, INSvi32lane>;
3431-
defm : Neon_ins_64b_sve_elt_pattern<v1f64, v2f64, nxv2f64, f64>;
3432-
defm : Neon_ins_64b_sve_elt_pattern<v1i64, v2i64, nxv2i64, i64>;
34333383
} // End HasNEON
34343384

34353385
// Extract first element from vector.

0 commit comments

Comments
 (0)