(ins addrmode6:$addr, am6offset:$offset, QQQQPR:$src), itin,
"$addr.addr = $wb, $src = $dst">;
-let mayLoad = 1, neverHasSideEffects = 1, hasExtraDefRegAllocReq = 1 in {
+let mayLoad = 1, hasSideEffects = 0, hasExtraDefRegAllocReq = 1 in {
// VLD1 : Vector Load (multiple single elements)
class VLD1D<bits<4> op7_4, string Dt, Operand AddrMode>
def VLD4q16oddPseudo_UPD : VLDQQQQWBPseudo<IIC_VLD4u>;
def VLD4q32oddPseudo_UPD : VLDQQQQWBPseudo<IIC_VLD4u>;
-} // mayLoad = 1, neverHasSideEffects = 1, hasExtraDefRegAllocReq = 1
+} // mayLoad = 1, hasSideEffects = 0, hasExtraDefRegAllocReq = 1
// Classes for VLD*LN pseudo-instructions with multi-register operands.
// These are expanded to real instructions after register allocation.
(f32 (load addrmode6:$addr)), imm:$lane),
(VLD1LNq32Pseudo addrmode6:$addr, QPR:$src, imm:$lane)>;
-let mayLoad = 1, neverHasSideEffects = 1, hasExtraDefRegAllocReq = 1 in {
+let mayLoad = 1, hasSideEffects = 0, hasExtraDefRegAllocReq = 1 in {
// ...with address register writeback:
class VLD1LNWB<bits<4> op11_8, bits<4> op7_4, string Dt>
def VLD4LNq16Pseudo_UPD : VLDQQQQLNWBPseudo<IIC_VLD4lnu>;
def VLD4LNq32Pseudo_UPD : VLDQQQQLNWBPseudo<IIC_VLD4lnu>;
-} // mayLoad = 1, neverHasSideEffects = 1, hasExtraDefRegAllocReq = 1
+} // mayLoad = 1, hasSideEffects = 0, hasExtraDefRegAllocReq = 1
// VLD1DUP : Vector Load (single element to all lanes)
class VLD1DUP<bits<4> op7_4, string Dt, ValueType Ty, PatFrag LoadOp,
def : Pat<(v4f32 (NEONvdup (f32 (load addrmode6dup:$addr)))),
(VLD1DUPq32 addrmode6:$addr)>;
-let mayLoad = 1, neverHasSideEffects = 1, hasExtraDefRegAllocReq = 1 in {
+let mayLoad = 1, hasSideEffects = 0, hasExtraDefRegAllocReq = 1 in {
// ...with address register writeback:
multiclass VLD1DUPWB<bits<4> op7_4, string Dt, Operand AddrMode> {
def _fixed : NLdSt<1, 0b10, 0b1100, op7_4,
def VLD4DUPd16Pseudo_UPD : VLDQQWBPseudo<IIC_VLD4dupu>;
def VLD4DUPd32Pseudo_UPD : VLDQQWBPseudo<IIC_VLD4dupu>;
-} // mayLoad = 1, neverHasSideEffects = 1, hasExtraDefRegAllocReq = 1
+} // mayLoad = 1, hasSideEffects = 0, hasExtraDefRegAllocReq = 1
-let mayStore = 1, neverHasSideEffects = 1, hasExtraSrcRegAllocReq = 1 in {
+let mayStore = 1, hasSideEffects = 0, hasExtraSrcRegAllocReq = 1 in {
// Classes for VST* pseudo-instructions with multi-register operands.
// These are expanded to real instructions after register allocation.
def VST4q16oddPseudo_UPD : VSTQQQQWBPseudo<IIC_VST4u>;
def VST4q32oddPseudo_UPD : VSTQQQQWBPseudo<IIC_VST4u>;
-} // mayStore = 1, neverHasSideEffects = 1, hasExtraSrcRegAllocReq = 1
+} // mayStore = 1, hasSideEffects = 0, hasExtraSrcRegAllocReq = 1
// Classes for VST*LN pseudo-instructions with multi-register operands.
// These are expanded to real instructions after register allocation.
def VST1LNq16Pseudo_UPD : VST1QLNWBPseudo<v8i16, post_truncsti16,NEONvgetlaneu>;
def VST1LNq32Pseudo_UPD : VST1QLNWBPseudo<v4i32, post_store, extractelt>;
-let mayStore = 1, neverHasSideEffects = 1, hasExtraSrcRegAllocReq = 1 in {
+let mayStore = 1, hasSideEffects = 0, hasExtraSrcRegAllocReq = 1 in {
// VST2LN : Vector Store (single 2-element structure from one lane)
class VST2LN<bits<4> op11_8, bits<4> op7_4, string Dt>
def VST4LNq16Pseudo_UPD : VSTQQQQLNWBPseudo<IIC_VST4lnu>;
def VST4LNq32Pseudo_UPD : VSTQQQQLNWBPseudo<IIC_VST4lnu>;
-} // mayStore = 1, neverHasSideEffects = 1, hasExtraSrcRegAllocReq = 1
+} // mayStore = 1, hasSideEffects = 0, hasExtraSrcRegAllocReq = 1
// Use vld1/vst1 for unaligned f64 load / store
def : Pat<(f64 (hword_alignedload addrmode6:$addr)),
imm:$lane)))))))]>;
class N3VDMulOpSL16<bits<2> op21_20, bits<4> op11_8, InstrItinClass itin,
string OpcodeStr, string Dt,
- ValueType Ty, SDNode MulOp, SDNode ShOp>
+ ValueType Ty, SDPatternOperator MulOp, SDPatternOperator ShOp>
: N3VLane16<0, 1, op21_20, op11_8, 1, 0,
(outs DPR:$Vd),
(ins DPR:$src1, DPR:$Vn, DPR_8:$Vm, VectorIndex16:$lane),
class N3VQMulOpSL16<bits<2> op21_20, bits<4> op11_8, InstrItinClass itin,
string OpcodeStr, string Dt,
ValueType ResTy, ValueType OpTy,
- SDNode MulOp, SDNode ShOp>
+ SDPatternOperator MulOp, SDPatternOperator ShOp>
: N3VLane16<1, 1, op21_20, op11_8, 1, 0,
(outs QPR:$Vd),
(ins QPR:$src1, QPR:$Vn, DPR_8:$Vm, VectorIndex16:$lane),
multiclass N3VMulOpSL_HS<bits<4> op11_8,
InstrItinClass itinD16, InstrItinClass itinD32,
InstrItinClass itinQ16, InstrItinClass itinQ32,
- string OpcodeStr, string Dt, SDNode ShOp> {
+ string OpcodeStr, string Dt, SDPatternOperator ShOp> {
def v4i16 : N3VDMulOpSL16<0b01, op11_8, itinD16,
OpcodeStr, !strconcat(Dt, "16"), v4i16, mul, ShOp>;
def v2i32 : N3VDMulOpSL<0b10, op11_8, itinD32,
}
// Neon 3-argument intrinsics,
-// element sizes of 8, 16 and 32 bits:
-multiclass N3VInt3_QHS<bit op24, bit op23, bits<4> op11_8, bit op4,
- InstrItinClass itinD, InstrItinClass itinQ,
+// element sizes of 16 and 32 bits:
+multiclass N3VInt3_HS<bit op24, bit op23, bits<4> op11_8, bit op4,
+ InstrItinClass itinD16, InstrItinClass itinD32,
+ InstrItinClass itinQ16, InstrItinClass itinQ32,
string OpcodeStr, string Dt, SDPatternOperator IntOp> {
// 64-bit vector types.
- def v8i8 : N3VDInt3<op24, op23, 0b00, op11_8, op4, itinD,
- OpcodeStr, !strconcat(Dt, "8"), v8i8, v8i8, IntOp>;
- def v4i16 : N3VDInt3<op24, op23, 0b01, op11_8, op4, itinD,
+ def v4i16 : N3VDInt3<op24, op23, 0b01, op11_8, op4, itinD16,
OpcodeStr, !strconcat(Dt, "16"), v4i16, v4i16, IntOp>;
- def v2i32 : N3VDInt3<op24, op23, 0b10, op11_8, op4, itinD,
+ def v2i32 : N3VDInt3<op24, op23, 0b10, op11_8, op4, itinD32,
OpcodeStr, !strconcat(Dt, "32"), v2i32, v2i32, IntOp>;
// 128-bit vector types.
- def v16i8 : N3VQInt3<op24, op23, 0b00, op11_8, op4, itinQ,
- OpcodeStr, !strconcat(Dt, "8"), v16i8, v16i8, IntOp>;
- def v8i16 : N3VQInt3<op24, op23, 0b01, op11_8, op4, itinQ,
+ def v8i16 : N3VQInt3<op24, op23, 0b01, op11_8, op4, itinQ16,
OpcodeStr, !strconcat(Dt, "16"), v8i16, v8i16, IntOp>;
- def v4i32 : N3VQInt3<op24, op23, 0b10, op11_8, op4, itinQ,
+ def v4i32 : N3VQInt3<op24, op23, 0b10, op11_8, op4, itinQ32,
OpcodeStr, !strconcat(Dt, "32"), v4i32, v4i32, IntOp>;
}
+// element sizes of 8, 16 and 32 bits:
+multiclass N3VInt3_QHS<bit op24, bit op23, bits<4> op11_8, bit op4,
+ InstrItinClass itinD16, InstrItinClass itinD32,
+ InstrItinClass itinQ16, InstrItinClass itinQ32,
+ string OpcodeStr, string Dt, SDPatternOperator IntOp>
+ :N3VInt3_HS <op24, op23, op11_8, op4, itinD16, itinD32,
+ itinQ16, itinQ32, OpcodeStr, Dt, IntOp>{
+ // 64-bit vector types.
+ def v8i8 : N3VDInt3<op24, op23, 0b00, op11_8, op4, itinD16,
+ OpcodeStr, !strconcat(Dt, "8"), v8i8, v8i8, IntOp>;
+ // 128-bit vector types.
+ def v16i8 : N3VQInt3<op24, op23, 0b00, op11_8, op4, itinQ16,
+ OpcodeStr, !strconcat(Dt, "8"), v16i8, v16i8, IntOp>;
+}
// Neon Long Multiply-Op vector operations,
// element sizes of 8, 16 and 32 bits:
defm VMLALsls : N3VLMulOpSL_HS<0, 0b0010, "vmlal", "s", NEONvmulls, add>;
defm VMLALslu : N3VLMulOpSL_HS<1, 0b0010, "vmlal", "u", NEONvmullu, add>;
+let Predicates = [HasNEON, HasV8_1a] in {
+ // v8.1a Neon Rounding Double Multiply-Op vector operations,
+ // VQRDMLAH : Vector Saturating Rounding Doubling Multiply Accumulate Long
+ // (Q += D * D)
+ defm VQRDMLAH : N3VInt3_HS<1, 0, 0b1011, 1, IIC_VMACi16D, IIC_VMACi32D,
+ IIC_VMACi16Q, IIC_VMACi32Q, "vqrdmlah", "s",
+ null_frag>;
+ def : Pat<(v4i16 (int_arm_neon_vqadds
+ (v4i16 DPR:$src1),
+ (v4i16 (int_arm_neon_vqrdmulh (v4i16 DPR:$Vn),
+ (v4i16 DPR:$Vm))))),
+ (v4i16 (VQRDMLAHv4i16 DPR:$src1, DPR:$Vn, DPR:$Vm))>;
+ def : Pat<(v2i32 (int_arm_neon_vqadds
+ (v2i32 DPR:$src1),
+ (v2i32 (int_arm_neon_vqrdmulh (v2i32 DPR:$Vn),
+ (v2i32 DPR:$Vm))))),
+ (v2i32 (VQRDMLAHv2i32 DPR:$src1, DPR:$Vn, DPR:$Vm))>;
+ def : Pat<(v8i16 (int_arm_neon_vqadds
+ (v8i16 QPR:$src1),
+ (v8i16 (int_arm_neon_vqrdmulh (v8i16 QPR:$Vn),
+ (v8i16 QPR:$Vm))))),
+ (v8i16 (VQRDMLAHv8i16 QPR:$src1, QPR:$Vn, QPR:$Vm))>;
+ def : Pat<(v4i32 (int_arm_neon_vqadds
+ (v4i32 QPR:$src1),
+ (v4i32 (int_arm_neon_vqrdmulh (v4i32 QPR:$Vn),
+ (v4i32 QPR:$Vm))))),
+ (v4i32 (VQRDMLAHv4i32 QPR:$src1, QPR:$Vn, QPR:$Vm))>;
+
+ defm VQRDMLAHsl : N3VMulOpSL_HS<0b1110, IIC_VMACi16D, IIC_VMACi32D,
+ IIC_VMACi16Q, IIC_VMACi32Q, "vqrdmlah", "s",
+ null_frag>;
+ def : Pat<(v4i16 (int_arm_neon_vqadds
+ (v4i16 DPR:$src1),
+ (v4i16 (int_arm_neon_vqrdmulh
+ (v4i16 DPR:$Vn),
+ (v4i16 (NEONvduplane (v4i16 DPR_8:$Vm),
+ imm:$lane)))))),
+ (v4i16 (VQRDMLAHslv4i16 DPR:$src1, DPR:$Vn, DPR_8:$Vm,
+ imm:$lane))>;
+ def : Pat<(v2i32 (int_arm_neon_vqadds
+ (v2i32 DPR:$src1),
+ (v2i32 (int_arm_neon_vqrdmulh
+ (v2i32 DPR:$Vn),
+ (v2i32 (NEONvduplane (v2i32 DPR_VFP2:$Vm),
+ imm:$lane)))))),
+ (v2i32 (VQRDMLAHslv2i32 DPR:$src1, DPR:$Vn, DPR_VFP2:$Vm,
+ imm:$lane))>;
+ def : Pat<(v8i16 (int_arm_neon_vqadds
+ (v8i16 QPR:$src1),
+ (v8i16 (int_arm_neon_vqrdmulh
+ (v8i16 QPR:$src2),
+ (v8i16 (NEONvduplane (v8i16 QPR:$src3),
+ imm:$lane)))))),
+ (v8i16 (VQRDMLAHslv8i16 (v8i16 QPR:$src1),
+ (v8i16 QPR:$src2),
+ (v4i16 (EXTRACT_SUBREG
+ QPR:$src3,
+ (DSubReg_i16_reg imm:$lane))),
+ (SubReg_i16_lane imm:$lane)))>;
+ def : Pat<(v4i32 (int_arm_neon_vqadds
+ (v4i32 QPR:$src1),
+ (v4i32 (int_arm_neon_vqrdmulh
+ (v4i32 QPR:$src2),
+ (v4i32 (NEONvduplane (v4i32 QPR:$src3),
+ imm:$lane)))))),
+ (v4i32 (VQRDMLAHslv4i32 (v4i32 QPR:$src1),
+ (v4i32 QPR:$src2),
+ (v2i32 (EXTRACT_SUBREG
+ QPR:$src3,
+ (DSubReg_i32_reg imm:$lane))),
+ (SubReg_i32_lane imm:$lane)))>;
+
+ // VQRDMLSH : Vector Saturating Rounding Doubling Multiply Subtract Long
+ // (Q -= D * D)
+ defm VQRDMLSH : N3VInt3_HS<1, 0, 0b1100, 1, IIC_VMACi16D, IIC_VMACi32D,
+ IIC_VMACi16Q, IIC_VMACi32Q, "vqrdmlsh", "s",
+ null_frag>;
+ def : Pat<(v4i16 (int_arm_neon_vqsubs
+ (v4i16 DPR:$src1),
+ (v4i16 (int_arm_neon_vqrdmulh (v4i16 DPR:$Vn),
+ (v4i16 DPR:$Vm))))),
+ (v4i16 (VQRDMLSHv4i16 DPR:$src1, DPR:$Vn, DPR:$Vm))>;
+ def : Pat<(v2i32 (int_arm_neon_vqsubs
+ (v2i32 DPR:$src1),
+ (v2i32 (int_arm_neon_vqrdmulh (v2i32 DPR:$Vn),
+ (v2i32 DPR:$Vm))))),
+ (v2i32 (VQRDMLSHv2i32 DPR:$src1, DPR:$Vn, DPR:$Vm))>;
+ def : Pat<(v8i16 (int_arm_neon_vqsubs
+ (v8i16 QPR:$src1),
+ (v8i16 (int_arm_neon_vqrdmulh (v8i16 QPR:$Vn),
+ (v8i16 QPR:$Vm))))),
+ (v8i16 (VQRDMLSHv8i16 QPR:$src1, QPR:$Vn, QPR:$Vm))>;
+ def : Pat<(v4i32 (int_arm_neon_vqsubs
+ (v4i32 QPR:$src1),
+ (v4i32 (int_arm_neon_vqrdmulh (v4i32 QPR:$Vn),
+ (v4i32 QPR:$Vm))))),
+ (v4i32 (VQRDMLSHv4i32 QPR:$src1, QPR:$Vn, QPR:$Vm))>;
+
+ defm VQRDMLSHsl : N3VMulOpSL_HS<0b1111, IIC_VMACi16D, IIC_VMACi32D,
+ IIC_VMACi16Q, IIC_VMACi32Q, "vqrdmlsh", "s",
+ null_frag>;
+ def : Pat<(v4i16 (int_arm_neon_vqsubs
+ (v4i16 DPR:$src1),
+ (v4i16 (int_arm_neon_vqrdmulh
+ (v4i16 DPR:$Vn),
+ (v4i16 (NEONvduplane (v4i16 DPR_8:$Vm),
+ imm:$lane)))))),
+ (v4i16 (VQRDMLSHslv4i16 DPR:$src1, DPR:$Vn, DPR_8:$Vm, imm:$lane))>;
+ def : Pat<(v2i32 (int_arm_neon_vqsubs
+ (v2i32 DPR:$src1),
+ (v2i32 (int_arm_neon_vqrdmulh
+ (v2i32 DPR:$Vn),
+ (v2i32 (NEONvduplane (v2i32 DPR_VFP2:$Vm),
+ imm:$lane)))))),
+ (v2i32 (VQRDMLSHslv2i32 DPR:$src1, DPR:$Vn, DPR_VFP2:$Vm,
+ imm:$lane))>;
+ def : Pat<(v8i16 (int_arm_neon_vqsubs
+ (v8i16 QPR:$src1),
+ (v8i16 (int_arm_neon_vqrdmulh
+ (v8i16 QPR:$src2),
+ (v8i16 (NEONvduplane (v8i16 QPR:$src3),
+ imm:$lane)))))),
+ (v8i16 (VQRDMLSHslv8i16 (v8i16 QPR:$src1),
+ (v8i16 QPR:$src2),
+ (v4i16 (EXTRACT_SUBREG
+ QPR:$src3,
+ (DSubReg_i16_reg imm:$lane))),
+ (SubReg_i16_lane imm:$lane)))>;
+ def : Pat<(v4i32 (int_arm_neon_vqsubs
+ (v4i32 QPR:$src1),
+ (v4i32 (int_arm_neon_vqrdmulh
+ (v4i32 QPR:$src2),
+ (v4i32 (NEONvduplane (v4i32 QPR:$src3),
+ imm:$lane)))))),
+ (v4i32 (VQRDMLSHslv4i32 (v4i32 QPR:$src1),
+ (v4i32 QPR:$src2),
+ (v2i32 (EXTRACT_SUBREG
+ QPR:$src3,
+ (DSubReg_i32_reg imm:$lane))),
+ (SubReg_i32_lane imm:$lane)))>;
+}
// VQDMLAL : Vector Saturating Doubling Multiply Accumulate Long (Q += D * D)
defm VQDMLAL : N3VLInt3_HS<0, 1, 0b1001, 0, IIC_VMACi16D, IIC_VMACi32D,
"vqdmlal", "s", null_frag>;
IIC_VMOVSI, "vmov", "32", "$R, $V$lane",
[(set GPR:$R, (extractelt (v2i32 DPR:$V),
imm:$lane))]>,
- Requires<[HasNEON, HasFastVGETLNi32]> {
+ Requires<[HasVFP2, HasFastVGETLNi32]> {
let Inst{21} = lane{0};
}
// def VGETLNf32: see FMRDH and FMRDL in ARMInstrVFP.td
(ins DPR:$src1, GPR:$R, VectorIndex32:$lane),
IIC_VMOVISL, "vmov", "32", "$V$lane, $R",
[(set DPR:$V, (insertelt (v2i32 DPR:$src1),
- GPR:$R, imm:$lane))]> {
+ GPR:$R, imm:$lane))]>,
+ Requires<[HasVFP2]> {
let Inst{21} = lane{0};
// This instruction is equivalent as
// $V = INSERT_SUBREG $src1, $R, translateImmToSubIdx($imm)
(v2f32 (COPY_TO_REGCLASS (v2f32 (IMPLICIT_DEF)), DPR_VFP2)),
SPR:$b, ssub_0)), DPR_VFP2)), ssub_0)>;
+class NVCVTIFPat<SDNode OpNode, NeonI Inst>
+ : NEONFPPat<(f32 (OpNode GPR:$a)),
+ (f32 (EXTRACT_SUBREG
+ (v2f32 (Inst
+ (INSERT_SUBREG
+ (v2f32 (IMPLICIT_DEF)),
+ (i32 (COPY_TO_REGCLASS GPR:$a, SPR)), ssub_0))),
+ ssub_0))>;
+class NVCVTFIPat<SDNode OpNode, NeonI Inst>
+ : NEONFPPat<(i32 (OpNode SPR:$a)),
+ (i32 (EXTRACT_SUBREG
+ (v2f32 (Inst (INSERT_SUBREG (v2f32 (IMPLICIT_DEF)),
+ SPR:$a, ssub_0))),
+ ssub_0))>;
+
def : N3VSPat<fadd, VADDfd>;
def : N3VSPat<fsub, VSUBfd>;
def : N3VSPat<fmul, VMULfd>;
def : N2VSPat<fneg, VNEGfd>;
def : N3VSPat<NEONfmax, VMAXfd>;
def : N3VSPat<NEONfmin, VMINfd>;
-def : N2VSPat<arm_ftosi, VCVTf2sd>;
-def : N2VSPat<arm_ftoui, VCVTf2ud>;
-def : N2VSPat<arm_sitof, VCVTs2fd>;
-def : N2VSPat<arm_uitof, VCVTu2fd>;
+def : NVCVTFIPat<fp_to_sint, VCVTf2sd>;
+def : NVCVTFIPat<fp_to_uint, VCVTf2ud>;
+def : NVCVTIFPat<sint_to_fp, VCVTs2fd>;
+def : NVCVTIFPat<uint_to_fp, VCVTu2fd>;
+
+// NEON doesn't have any f64 conversions, so provide patterns to make
+// sure the VFP conversions match when extracting from a vector.
+def : VFPPat<(f64 (sint_to_fp (extractelt (v2i32 DPR:$src), imm:$lane))),
+ (VSITOD (EXTRACT_SUBREG DPR:$src, (SSubReg_f32_reg imm:$lane)))>;
+def : VFPPat<(f64 (sint_to_fp (extractelt (v4i32 QPR:$src), imm:$lane))),
+ (VSITOD (EXTRACT_SUBREG QPR:$src, (SSubReg_f32_reg imm:$lane)))>;
+def : VFPPat<(f64 (uint_to_fp (extractelt (v2i32 DPR:$src), imm:$lane))),
+ (VUITOD (EXTRACT_SUBREG DPR:$src, (SSubReg_f32_reg imm:$lane)))>;
+def : VFPPat<(f64 (uint_to_fp (extractelt (v4i32 QPR:$src), imm:$lane))),
+ (VUITOD (EXTRACT_SUBREG QPR:$src, (SSubReg_f32_reg imm:$lane)))>;
+
// Prefer VMOVDRR for i32 -> f32 bitcasts, it can write all DPR registers.
def : Pat<(f32 (bitconvert GPR:$a)),
projects / oota-llvm.git / blobdiff