aboutsummaryrefslogtreecommitdiffstats
path: root/lib/Target/ARM/ARMInstrNEON.td
diff options
context:
space:
mode:
Diffstat (limited to 'lib/Target/ARM/ARMInstrNEON.td')
-rw-r--r--lib/Target/ARM/ARMInstrNEON.td321
1 files changed, 263 insertions, 58 deletions
diff --git a/lib/Target/ARM/ARMInstrNEON.td b/lib/Target/ARM/ARMInstrNEON.td
index f61eb2b..fd8ac0b 100644
--- a/lib/Target/ARM/ARMInstrNEON.td
+++ b/lib/Target/ARM/ARMInstrNEON.td
@@ -530,16 +530,16 @@ def NEONimmAllOnesV: PatLeaf<(NEONvmovImm (i32 timm)), [{
// Use VLDM to load a Q register as a D register pair.
// This is a pseudo instruction that is expanded to VLDMD after reg alloc.
def VLDMQIA
- : PseudoVFPLdStM<(outs QPR:$dst), (ins GPR:$Rn),
+ : PseudoVFPLdStM<(outs DPair:$dst), (ins GPR:$Rn),
IIC_fpLoad_m, "",
- [(set QPR:$dst, (v2f64 (load GPR:$Rn)))]>;
+ [(set DPair:$dst, (v2f64 (load GPR:$Rn)))]>;
// Use VSTM to store a Q register as a D register pair.
// This is a pseudo instruction that is expanded to VSTMD after reg alloc.
def VSTMQIA
- : PseudoVFPLdStM<(outs), (ins QPR:$src, GPR:$Rn),
+ : PseudoVFPLdStM<(outs), (ins DPair:$src, GPR:$Rn),
IIC_fpStore_m, "",
- [(store (v2f64 QPR:$src), GPR:$Rn)]>;
+ [(store (v2f64 DPair:$src), GPR:$Rn)]>;
// Classes for VLD* pseudo-instructions with multi-register operands.
// These are expanded to real instructions after register allocation.
@@ -1938,20 +1938,11 @@ class VSTQQQQLNWBPseudo<InstrItinClass itin>
// VST1LN : Vector Store (single element from one lane)
class VST1LN<bits<4> op11_8, bits<4> op7_4, string Dt, ValueType Ty,
- PatFrag StoreOp, SDNode ExtractOp>
+ PatFrag StoreOp, SDNode ExtractOp, Operand AddrMode>
: NLdStLn<1, 0b00, op11_8, op7_4, (outs),
- (ins addrmode6:$Rn, DPR:$Vd, nohash_imm:$lane),
+ (ins AddrMode:$Rn, DPR:$Vd, nohash_imm:$lane),
IIC_VST1ln, "vst1", Dt, "\\{$Vd[$lane]\\}, $Rn", "",
- [(StoreOp (ExtractOp (Ty DPR:$Vd), imm:$lane), addrmode6:$Rn)]> {
- let Rm = 0b1111;
- let DecoderMethod = "DecodeVST1LN";
-}
-class VST1LN32<bits<4> op11_8, bits<4> op7_4, string Dt, ValueType Ty,
- PatFrag StoreOp, SDNode ExtractOp>
- : NLdStLn<1, 0b00, op11_8, op7_4, (outs),
- (ins addrmode6oneL32:$Rn, DPR:$Vd, nohash_imm:$lane),
- IIC_VST1ln, "vst1", Dt, "\\{$Vd[$lane]\\}, $Rn", "",
- [(StoreOp (ExtractOp (Ty DPR:$Vd), imm:$lane), addrmode6oneL32:$Rn)]>{
+ [(StoreOp (ExtractOp (Ty DPR:$Vd), imm:$lane), AddrMode:$Rn)]> {
let Rm = 0b1111;
let DecoderMethod = "DecodeVST1LN";
}
@@ -1962,16 +1953,17 @@ class VST1QLNPseudo<ValueType Ty, PatFrag StoreOp, SDNode ExtractOp>
}
def VST1LNd8 : VST1LN<0b0000, {?,?,?,0}, "8", v8i8, truncstorei8,
- NEONvgetlaneu> {
+ NEONvgetlaneu, addrmode6> {
let Inst{7-5} = lane{2-0};
}
def VST1LNd16 : VST1LN<0b0100, {?,?,0,?}, "16", v4i16, truncstorei16,
- NEONvgetlaneu> {
+ NEONvgetlaneu, addrmode6> {
let Inst{7-6} = lane{1-0};
let Inst{4} = Rn{5};
}
-def VST1LNd32 : VST1LN32<0b1000, {?,0,?,?}, "32", v2i32, store, extractelt> {
+def VST1LNd32 : VST1LN<0b1000, {?,0,?,?}, "32", v2i32, store, extractelt,
+ addrmode6oneL32> {
let Inst{7} = lane{0};
let Inst{5-4} = Rn{5-4};
}
@@ -1987,14 +1979,14 @@ def : Pat<(store (extractelt (v4f32 QPR:$src), imm:$lane), addrmode6:$addr),
// ...with address register writeback:
class VST1LNWB<bits<4> op11_8, bits<4> op7_4, string Dt, ValueType Ty,
- PatFrag StoreOp, SDNode ExtractOp>
+ PatFrag StoreOp, SDNode ExtractOp, Operand AdrMode>
: NLdStLn<1, 0b00, op11_8, op7_4, (outs GPR:$wb),
- (ins addrmode6:$Rn, am6offset:$Rm,
+ (ins AdrMode:$Rn, am6offset:$Rm,
DPR:$Vd, nohash_imm:$lane), IIC_VST1lnu, "vst1", Dt,
"\\{$Vd[$lane]\\}, $Rn$Rm",
"$Rn.addr = $wb",
[(set GPR:$wb, (StoreOp (ExtractOp (Ty DPR:$Vd), imm:$lane),
- addrmode6:$Rn, am6offset:$Rm))]> {
+ AdrMode:$Rn, am6offset:$Rm))]> {
let DecoderMethod = "DecodeVST1LN";
}
class VST1QLNWBPseudo<ValueType Ty, PatFrag StoreOp, SDNode ExtractOp>
@@ -2004,16 +1996,16 @@ class VST1QLNWBPseudo<ValueType Ty, PatFrag StoreOp, SDNode ExtractOp>
}
def VST1LNd8_UPD : VST1LNWB<0b0000, {?,?,?,0}, "8", v8i8, post_truncsti8,
- NEONvgetlaneu> {
+ NEONvgetlaneu, addrmode6> {
let Inst{7-5} = lane{2-0};
}
def VST1LNd16_UPD : VST1LNWB<0b0100, {?,?,0,?}, "16", v4i16, post_truncsti16,
- NEONvgetlaneu> {
+ NEONvgetlaneu, addrmode6> {
let Inst{7-6} = lane{1-0};
let Inst{4} = Rn{5};
}
def VST1LNd32_UPD : VST1LNWB<0b1000, {?,0,?,?}, "32", v2i32, post_store,
- extractelt> {
+ extractelt, addrmode6oneL32> {
let Inst{7} = lane{0};
let Inst{5-4} = Rn{5-4};
}
@@ -3642,7 +3634,7 @@ multiclass N2VShL_QHSD<bit op24, bit op23, bits<4> op11_8, bit op4,
}
multiclass N2VShR_QHSD<bit op24, bit op23, bits<4> op11_8, bit op4,
InstrItinClass itin, string OpcodeStr, string Dt,
- SDNode OpNode> {
+ string baseOpc, SDNode OpNode> {
// 64-bit vector types.
def v8i8 : N2VDSh<op24, op23, op11_8, 0, op4, N2RegVShRFrm, itin, shr_imm8,
OpcodeStr, !strconcat(Dt, "8"), v8i8, OpNode> {
@@ -3676,6 +3668,33 @@ multiclass N2VShR_QHSD<bit op24, bit op23, bits<4> op11_8, bit op4,
def v2i64 : N2VQSh<op24, op23, op11_8, 1, op4, N2RegVShRFrm, itin, shr_imm64,
OpcodeStr, !strconcat(Dt, "64"), v2i64, OpNode>;
// imm6 = xxxxxx
+
+ // Aliases for two-operand forms (source and dest regs the same).
+ def : NEONInstAlias<!strconcat(OpcodeStr, "${p}.", Dt, "8 $Vdn, $imm"),
+ (!cast<Instruction>(!strconcat(baseOpc, "v8i8"))
+ DPR:$Vdn, DPR:$Vdn, shr_imm8:$imm, pred:$p)>;
+ def : NEONInstAlias<!strconcat(OpcodeStr, "${p}.", Dt, "16 $Vdn, $imm"),
+ (!cast<Instruction>(!strconcat(baseOpc, "v4i16"))
+ DPR:$Vdn, DPR:$Vdn, shr_imm16:$imm, pred:$p)>;
+ def : NEONInstAlias<!strconcat(OpcodeStr, "${p}.", Dt, "32 $Vdn, $imm"),
+ (!cast<Instruction>(!strconcat(baseOpc, "v2i32"))
+ DPR:$Vdn, DPR:$Vdn, shr_imm32:$imm, pred:$p)>;
+ def : NEONInstAlias<!strconcat(OpcodeStr, "${p}.", Dt, "64 $Vdn, $imm"),
+ (!cast<Instruction>(!strconcat(baseOpc, "v1i64"))
+ DPR:$Vdn, DPR:$Vdn, shr_imm64:$imm, pred:$p)>;
+
+ def : NEONInstAlias<!strconcat(OpcodeStr, "${p}.", Dt, "8 $Vdn, $imm"),
+ (!cast<Instruction>(!strconcat(baseOpc, "v16i8"))
+ QPR:$Vdn, QPR:$Vdn, shr_imm8:$imm, pred:$p)>;
+ def : NEONInstAlias<!strconcat(OpcodeStr, "${p}.", Dt, "16 $Vdn, $imm"),
+ (!cast<Instruction>(!strconcat(baseOpc, "v8i16"))
+ QPR:$Vdn, QPR:$Vdn, shr_imm16:$imm, pred:$p)>;
+ def : NEONInstAlias<!strconcat(OpcodeStr, "${p}.", Dt, "32 $Vdn, $imm"),
+ (!cast<Instruction>(!strconcat(baseOpc, "v4i32"))
+ QPR:$Vdn, QPR:$Vdn, shr_imm32:$imm, pred:$p)>;
+ def : NEONInstAlias<!strconcat(OpcodeStr, "${p}.", Dt, "64 $Vdn, $imm"),
+ (!cast<Instruction>(!strconcat(baseOpc, "v2i64"))
+ QPR:$Vdn, QPR:$Vdn, shr_imm64:$imm, pred:$p)>;
}
// Neon Shift-Accumulate vector operations,
@@ -3986,10 +4005,10 @@ defm VMLA : N3VMulOp_QHS<0, 0, 0b1001, 0, IIC_VMACi16D, IIC_VMACi32D,
IIC_VMACi16Q, IIC_VMACi32Q, "vmla", "i", add>;
def VMLAfd : N3VDMulOp<0, 0, 0b00, 0b1101, 1, IIC_VMACD, "vmla", "f32",
v2f32, fmul_su, fadd_mlx>,
- Requires<[HasNEON, UseFPVMLx, NoNEON2]>;
+ Requires<[HasNEON, UseFPVMLx, DontUseFusedMAC]>;
def VMLAfq : N3VQMulOp<0, 0, 0b00, 0b1101, 1, IIC_VMACQ, "vmla", "f32",
v4f32, fmul_su, fadd_mlx>,
- Requires<[HasNEON, UseFPVMLx, NoNEON2]>;
+ Requires<[HasNEON, UseFPVMLx, DontUseFusedMAC]>;
defm VMLAsl : N3VMulOpSL_HS<0b0000, IIC_VMACi16D, IIC_VMACi32D,
IIC_VMACi16Q, IIC_VMACi32Q, "vmla", "i", add>;
def VMLAslfd : N3VDMulOpSL<0b10, 0b0001, IIC_VMACD, "vmla", "f32",
@@ -4044,10 +4063,10 @@ defm VMLS : N3VMulOp_QHS<1, 0, 0b1001, 0, IIC_VMACi16D, IIC_VMACi32D,
IIC_VMACi16Q, IIC_VMACi32Q, "vmls", "i", sub>;
def VMLSfd : N3VDMulOp<0, 0, 0b10, 0b1101, 1, IIC_VMACD, "vmls", "f32",
v2f32, fmul_su, fsub_mlx>,
- Requires<[HasNEON, UseFPVMLx, NoNEON2]>;
+ Requires<[HasNEON, UseFPVMLx, DontUseFusedMAC]>;
def VMLSfq : N3VQMulOp<0, 0, 0b10, 0b1101, 1, IIC_VMACQ, "vmls", "f32",
v4f32, fmul_su, fsub_mlx>,
- Requires<[HasNEON, UseFPVMLx, NoNEON2]>;
+ Requires<[HasNEON, UseFPVMLx, DontUseFusedMAC]>;
defm VMLSsl : N3VMulOpSL_HS<0b0100, IIC_VMACi16D, IIC_VMACi32D,
IIC_VMACi16Q, IIC_VMACi32Q, "vmls", "i", sub>;
def VMLSslfd : N3VDMulOpSL<0b10, 0b0101, IIC_VMACD, "vmls", "f32",
@@ -4096,23 +4115,36 @@ defm VQDMLSL : N3VLInt3_HS<0, 1, 0b1011, 0, IIC_VMACi16D, IIC_VMACi32D,
"vqdmlsl", "s", int_arm_neon_vqdmlsl>;
defm VQDMLSLsl: N3VLInt3SL_HS<0, 0b111, "vqdmlsl", "s", int_arm_neon_vqdmlsl>;
-
// Fused Vector Multiply-Accumulate and Fused Multiply-Subtract Operations.
def VFMAfd : N3VDMulOp<0, 0, 0b00, 0b1100, 1, IIC_VFMACD, "vfma", "f32",
v2f32, fmul_su, fadd_mlx>,
- Requires<[HasNEON2,FPContractions]>;
+ Requires<[HasVFP4,UseFusedMAC]>;
def VFMAfq : N3VQMulOp<0, 0, 0b00, 0b1100, 1, IIC_VFMACQ, "vfma", "f32",
v4f32, fmul_su, fadd_mlx>,
- Requires<[HasNEON2,FPContractions]>;
+ Requires<[HasVFP4,UseFusedMAC]>;
// Fused Vector Multiply Subtract (floating-point)
def VFMSfd : N3VDMulOp<0, 0, 0b10, 0b1100, 1, IIC_VFMACD, "vfms", "f32",
v2f32, fmul_su, fsub_mlx>,
- Requires<[HasNEON2,FPContractions]>;
+ Requires<[HasVFP4,UseFusedMAC]>;
def VFMSfq : N3VQMulOp<0, 0, 0b10, 0b1100, 1, IIC_VFMACQ, "vfms", "f32",
v4f32, fmul_su, fsub_mlx>,
- Requires<[HasNEON2,FPContractions]>;
+ Requires<[HasVFP4,UseFusedMAC]>;
+
+// Match @llvm.fma.* intrinsics
+def : Pat<(v2f32 (fma DPR:$src1, DPR:$Vn, DPR:$Vm)),
+ (VFMAfd DPR:$src1, DPR:$Vn, DPR:$Vm)>,
+ Requires<[HasVFP4]>;
+def : Pat<(v4f32 (fma QPR:$src1, QPR:$Vn, QPR:$Vm)),
+ (VFMAfq QPR:$src1, QPR:$Vn, QPR:$Vm)>,
+ Requires<[HasVFP4]>;
+def : Pat<(v2f32 (fma (fneg DPR:$src1), DPR:$Vn, DPR:$Vm)),
+ (VFMSfd DPR:$src1, DPR:$Vn, DPR:$Vm)>,
+ Requires<[HasVFP4]>;
+def : Pat<(v4f32 (fma (fneg QPR:$src1), QPR:$Vn, QPR:$Vm)),
+ (VFMSfq QPR:$src1, QPR:$Vn, QPR:$Vm)>,
+ Requires<[HasVFP4]>;
// Vector Subtract Operations.
@@ -4614,8 +4646,10 @@ defm VSHLu : N3VInt_QHSDSh<1, 0, 0b0100, 0, N3RegVShFrm,
defm VSHLi : N2VShL_QHSD<0, 1, 0b0101, 1, IIC_VSHLiD, "vshl", "i", NEONvshl>;
// VSHR : Vector Shift Right (Immediate)
-defm VSHRs : N2VShR_QHSD<0, 1, 0b0000, 1, IIC_VSHLiD, "vshr", "s",NEONvshrs>;
-defm VSHRu : N2VShR_QHSD<1, 1, 0b0000, 1, IIC_VSHLiD, "vshr", "u",NEONvshru>;
+defm VSHRs : N2VShR_QHSD<0, 1, 0b0000, 1, IIC_VSHLiD, "vshr", "s", "VSHRs",
+ NEONvshrs>;
+defm VSHRu : N2VShR_QHSD<1, 1, 0b0000, 1, IIC_VSHLiD, "vshr", "u", "VSHRu",
+ NEONvshru>;
// VSHLL : Vector Shift Left Long
defm VSHLLs : N2VLSh_QHS<0, 1, 0b1010, 0, 0, 1, "vshll", "s", NEONvshlls>;
@@ -4649,8 +4683,10 @@ defm VRSHLu : N3VInt_QHSDSh<1, 0, 0b0101, 0, N3RegVShFrm,
IIC_VSHLi4D, IIC_VSHLi4D, IIC_VSHLi4Q, IIC_VSHLi4Q,
"vrshl", "u", int_arm_neon_vrshiftu>;
// VRSHR : Vector Rounding Shift Right
-defm VRSHRs : N2VShR_QHSD<0,1,0b0010,1, IIC_VSHLi4D, "vrshr", "s",NEONvrshrs>;
-defm VRSHRu : N2VShR_QHSD<1,1,0b0010,1, IIC_VSHLi4D, "vrshr", "u",NEONvrshru>;
+defm VRSHRs : N2VShR_QHSD<0,1,0b0010,1, IIC_VSHLi4D, "vrshr", "s", "VRSHRs",
+ NEONvrshrs>;
+defm VRSHRu : N2VShR_QHSD<1,1,0b0010,1, IIC_VSHLi4D, "vrshr", "u", "VRSHRu",
+ NEONvrshru>;
// VRSHRN : Vector Rounding Shift Right and Narrow
defm VRSHRN : N2VNSh_HSD<0, 1, 0b1000, 0, 1, 1, IIC_VSHLi4D, "vrshrn", "i",
@@ -4795,12 +4831,12 @@ def VCNTq : N2VQInt<0b11, 0b11, 0b00, 0b00, 0b01010, 0,
// Vector Swap
def VSWPd : N2VX<0b11, 0b11, 0b00, 0b10, 0b00000, 0, 0,
- (outs DPR:$Vd, DPR:$Vd1), (ins DPR:$Vm, DPR:$Vm1),
- NoItinerary, "vswp", "$Vd, $Vd1", "$Vm = $Vd, $Vm1 = $Vd1",
+ (outs DPR:$Vd, DPR:$Vm), (ins DPR:$in1, DPR:$in2),
+ NoItinerary, "vswp", "$Vd, $Vm", "$in1 = $Vd, $in2 = $Vm",
[]>;
def VSWPq : N2VX<0b11, 0b11, 0b00, 0b10, 0b00000, 1, 0,
- (outs QPR:$Vd, QPR:$Vd1), (ins QPR:$Vm, QPR:$Vm1),
- NoItinerary, "vswp", "$Vd, $Vd1", "$Vm = $Vd, $Vm1 = $Vd1",
+ (outs QPR:$Vd, QPR:$Vm), (ins QPR:$in1, QPR:$in2),
+ NoItinerary, "vswp", "$Vd, $Vm", "$in1 = $Vd, $in2 = $Vm",
[]>;
// Vector Move Operations.
@@ -5342,7 +5378,9 @@ def VTRNq32 : N2VQShuffle<0b10, 0b00001, IIC_VPERMQ, "vtrn", "32">;
def VUZPd8 : N2VDShuffle<0b00, 0b00010, "vuzp", "8">;
def VUZPd16 : N2VDShuffle<0b01, 0b00010, "vuzp", "16">;
-def VUZPd32 : N2VDShuffle<0b10, 0b00010, "vuzp", "32">;
+// vuzp.32 Dd, Dm is a pseudo-instruction expanded to vtrn.32 Dd, Dm.
+def : NEONInstAlias<"vuzp${p}.32 $Dd, $Dm",
+ (VTRNd32 DPR:$Dd, DPR:$Dm, pred:$p)>;
def VUZPq8 : N2VQShuffle<0b00, 0b00010, IIC_VPERMQ3, "vuzp", "8">;
def VUZPq16 : N2VQShuffle<0b01, 0b00010, IIC_VPERMQ3, "vuzp", "16">;
@@ -5352,7 +5390,9 @@ def VUZPq32 : N2VQShuffle<0b10, 0b00010, IIC_VPERMQ3, "vuzp", "32">;
def VZIPd8 : N2VDShuffle<0b00, 0b00011, "vzip", "8">;
def VZIPd16 : N2VDShuffle<0b01, 0b00011, "vzip", "16">;
-def VZIPd32 : N2VDShuffle<0b10, 0b00011, "vzip", "32">;
+// vzip.32 Dd, Dm is a pseudo-instruction expanded to vtrn.32 Dd, Dm.
+def : NEONInstAlias<"vzip${p}.32 $Dd, $Dm",
+ (VTRNd32 DPR:$Dd, DPR:$Dm, pred:$p)>;
def VZIPq8 : N2VQShuffle<0b00, 0b00011, IIC_VPERMQ3, "vzip", "8">;
def VZIPq16 : N2VQShuffle<0b01, 0b00011, IIC_VPERMQ3, "vzip", "16">;
@@ -5462,13 +5502,13 @@ def : N3VSPat<fadd, VADDfd>;
def : N3VSPat<fsub, VSUBfd>;
def : N3VSPat<fmul, VMULfd>;
def : N3VSMulOpPat<fmul, fadd, VMLAfd>,
- Requires<[HasNEON, UseNEONForFP, UseFPVMLx, NoNEON2]>;
+ Requires<[HasNEON, UseNEONForFP, UseFPVMLx, DontUseFusedMAC]>;
def : N3VSMulOpPat<fmul, fsub, VMLSfd>,
- Requires<[HasNEON, UseNEONForFP, UseFPVMLx, NoNEON2]>;
+ Requires<[HasNEON, UseNEONForFP, UseFPVMLx, DontUseFusedMAC]>;
def : N3VSMulOpPat<fmul, fadd, VFMAfd>,
- Requires<[HasNEON2, UseNEONForFP,FPContractions]>;
+ Requires<[HasVFP4, UseNEONForFP, UseFusedMAC]>;
def : N3VSMulOpPat<fmul, fsub, VFMSfd>,
- Requires<[HasNEON2, UseNEONForFP,FPContractions]>;
+ Requires<[HasVFP4, UseNEONForFP, UseFusedMAC]>;
def : N2VSPat<fabs, VABSfd>;
def : N2VSPat<fneg, VNEGfd>;
def : N3VSPat<NEONfmax, VMAXfd>;
@@ -5594,6 +5634,7 @@ multiclass Lengthen_HalfSingle<string DestLanes, string DestTy, string SrcTy,
// extload, zextload and sextload for a lengthening load followed by another
// lengthening load, to quadruple the initial length.
+//
// Lengthen_Double<"4", "i32", "i8", "8", "i16", "4", "i32", qsub_0> =
// Pat<(v4i32 (extloadvi8 addrmode5:$addr))
// (EXTRACT_SUBREG (VMOVLuv4i32
@@ -5604,28 +5645,63 @@ multiclass Lengthen_HalfSingle<string DestLanes, string DestTy, string SrcTy,
// qsub_0)>;
multiclass Lengthen_Double<string DestLanes, string DestTy, string SrcTy,
string Insn1Lanes, string Insn1Ty, string Insn2Lanes,
- string Insn2Ty, SubRegIndex RegType> {
+ string Insn2Ty> {
+ def _Any : Pat<(!cast<ValueType>("v" # DestLanes # DestTy)
+ (!cast<PatFrag>("extloadv" # SrcTy) addrmode5:$addr)),
+ (!cast<Instruction>("VMOVLuv" # Insn2Lanes # Insn2Ty)
+ (EXTRACT_SUBREG (!cast<Instruction>("VMOVLuv" # Insn1Lanes # Insn1Ty)
+ (INSERT_SUBREG (f64 (IMPLICIT_DEF)), (VLDRS addrmode5:$addr),
+ ssub_0)), dsub_0))>;
+ def _Z : Pat<(!cast<ValueType>("v" # DestLanes # DestTy)
+ (!cast<PatFrag>("zextloadv" # SrcTy) addrmode5:$addr)),
+ (!cast<Instruction>("VMOVLuv" # Insn2Lanes # Insn2Ty)
+ (EXTRACT_SUBREG (!cast<Instruction>("VMOVLuv" # Insn1Lanes # Insn1Ty)
+ (INSERT_SUBREG (f64 (IMPLICIT_DEF)), (VLDRS addrmode5:$addr),
+ ssub_0)), dsub_0))>;
+ def _S : Pat<(!cast<ValueType>("v" # DestLanes # DestTy)
+ (!cast<PatFrag>("sextloadv" # SrcTy) addrmode5:$addr)),
+ (!cast<Instruction>("VMOVLsv" # Insn2Lanes # Insn2Ty)
+ (EXTRACT_SUBREG (!cast<Instruction>("VMOVLsv" # Insn1Lanes # Insn1Ty)
+ (INSERT_SUBREG (f64 (IMPLICIT_DEF)), (VLDRS addrmode5:$addr),
+ ssub_0)), dsub_0))>;
+}
+
+// extload, zextload and sextload for a lengthening load followed by another
+// lengthening load, to quadruple the initial length, but which ends up only
+// requiring half the available lanes (a 64-bit outcome instead of a 128-bit).
+//
+// Lengthen_HalfDouble<"2", "i32", "i8", "8", "i16", "4", "i32"> =
+// Pat<(v4i32 (extloadvi8 addrmode5:$addr))
+// (EXTRACT_SUBREG (VMOVLuv4i32
+// (EXTRACT_SUBREG (VMOVLuv8i16 (INSERT_SUBREG (f64 (IMPLICIT_DEF)),
+// (VLDRS addrmode5:$addr),
+// ssub_0)),
+// dsub_0)),
+// dsub_0)>;
+multiclass Lengthen_HalfDouble<string DestLanes, string DestTy, string SrcTy,
+ string Insn1Lanes, string Insn1Ty, string Insn2Lanes,
+ string Insn2Ty> {
def _Any : Pat<(!cast<ValueType>("v" # DestLanes # DestTy)
(!cast<PatFrag>("extloadv" # SrcTy) addrmode5:$addr)),
(EXTRACT_SUBREG (!cast<Instruction>("VMOVLuv" # Insn2Lanes # Insn2Ty)
(EXTRACT_SUBREG (!cast<Instruction>("VMOVLuv" # Insn1Lanes # Insn1Ty)
(INSERT_SUBREG (f64 (IMPLICIT_DEF)), (VLDRS addrmode5:$addr),
ssub_0)), dsub_0)),
- RegType)>;
+ dsub_0)>;
def _Z : Pat<(!cast<ValueType>("v" # DestLanes # DestTy)
(!cast<PatFrag>("zextloadv" # SrcTy) addrmode5:$addr)),
(EXTRACT_SUBREG (!cast<Instruction>("VMOVLuv" # Insn2Lanes # Insn2Ty)
(EXTRACT_SUBREG (!cast<Instruction>("VMOVLuv" # Insn1Lanes # Insn1Ty)
(INSERT_SUBREG (f64 (IMPLICIT_DEF)), (VLDRS addrmode5:$addr),
ssub_0)), dsub_0)),
- RegType)>;
+ dsub_0)>;
def _S : Pat<(!cast<ValueType>("v" # DestLanes # DestTy)
(!cast<PatFrag>("sextloadv" # SrcTy) addrmode5:$addr)),
(EXTRACT_SUBREG (!cast<Instruction>("VMOVLsv" # Insn2Lanes # Insn2Ty)
(EXTRACT_SUBREG (!cast<Instruction>("VMOVLsv" # Insn1Lanes # Insn1Ty)
(INSERT_SUBREG (f64 (IMPLICIT_DEF)), (VLDRS addrmode5:$addr),
ssub_0)), dsub_0)),
- RegType)>;
+ dsub_0)>;
}
defm : Lengthen_Single<"8", "i16", "i8">; // v8i8 -> v8i16
@@ -5636,12 +5712,12 @@ defm : Lengthen_HalfSingle<"4", "i16", "i8", "8", "i16">; // v4i8 -> v4i16
defm : Lengthen_HalfSingle<"2", "i16", "i8", "8", "i16">; // v2i8 -> v2i16
defm : Lengthen_HalfSingle<"2", "i32", "i16", "4", "i32">; // v2i16 -> v2i32
-// Double lengthening - v4i8 -> v4i16 -> v4i32
-defm : Lengthen_Double<"4", "i32", "i8", "8", "i16", "4", "i32", qsub_0>;
+// Double lengthening - v4i8 -> v4i16 -> v4i32
+defm : Lengthen_Double<"4", "i32", "i8", "8", "i16", "4", "i32">;
// v2i8 -> v2i16 -> v2i32
-defm : Lengthen_Double<"2", "i32", "i8", "8", "i16", "4", "i32", dsub_0>;
+defm : Lengthen_HalfDouble<"2", "i32", "i8", "8", "i16", "4", "i32">;
// v2i16 -> v2i32 -> v2i64
-defm : Lengthen_Double<"2", "i64", "i16", "4", "i32", "2", "i64", qsub_0>;
+defm : Lengthen_Double<"2", "i64", "i16", "4", "i32", "2", "i64">;
// Triple lengthening - v2i8 -> v2i16 -> v2i32 -> v2i64
def : Pat<(v2i64 (extloadvi8 addrmode5:$addr)),
@@ -5911,7 +5987,7 @@ def : NEONInstAlias<"vshl${p}.u32 $Vdn, $Vm",
def : NEONInstAlias<"vshl${p}.u64 $Vdn, $Vm",
(VSHLuv2i64 QPR:$Vdn, QPR:$Vdn, QPR:$Vm, pred:$p)>;
-// VSHL (immediate) two-operand aliases.
+// VSHR (immediate) two-operand aliases.
def : NEONInstAlias<"vshr${p}.s8 $Vdn, $imm",
(VSHRsv8i8 DPR:$Vdn, DPR:$Vdn, shr_imm8:$imm, pred:$p)>;
def : NEONInstAlias<"vshr${p}.s16 $Vdn, $imm",
@@ -5948,6 +6024,41 @@ def : NEONInstAlias<"vshr${p}.u32 $Vdn, $imm",
def : NEONInstAlias<"vshr${p}.u64 $Vdn, $imm",
(VSHRuv2i64 QPR:$Vdn, QPR:$Vdn, shr_imm64:$imm, pred:$p)>;
+// VRSHL two-operand aliases.
+def : NEONInstAlias<"vrshl${p}.s8 $Vdn, $Vm",
+ (VRSHLsv8i8 DPR:$Vdn, DPR:$Vdn, DPR:$Vm, pred:$p)>;
+def : NEONInstAlias<"vrshl${p}.s16 $Vdn, $Vm",
+ (VRSHLsv4i16 DPR:$Vdn, DPR:$Vdn, DPR:$Vm, pred:$p)>;
+def : NEONInstAlias<"vrshl${p}.s32 $Vdn, $Vm",
+ (VRSHLsv2i32 DPR:$Vdn, DPR:$Vdn, DPR:$Vm, pred:$p)>;
+def : NEONInstAlias<"vrshl${p}.s64 $Vdn, $Vm",
+ (VRSHLsv1i64 DPR:$Vdn, DPR:$Vdn, DPR:$Vm, pred:$p)>;
+def : NEONInstAlias<"vrshl${p}.u8 $Vdn, $Vm",
+ (VRSHLuv8i8 DPR:$Vdn, DPR:$Vdn, DPR:$Vm, pred:$p)>;
+def : NEONInstAlias<"vrshl${p}.u16 $Vdn, $Vm",
+ (VRSHLuv4i16 DPR:$Vdn, DPR:$Vdn, DPR:$Vm, pred:$p)>;
+def : NEONInstAlias<"vrshl${p}.u32 $Vdn, $Vm",
+ (VRSHLuv2i32 DPR:$Vdn, DPR:$Vdn, DPR:$Vm, pred:$p)>;
+def : NEONInstAlias<"vrshl${p}.u64 $Vdn, $Vm",
+ (VRSHLuv1i64 DPR:$Vdn, DPR:$Vdn, DPR:$Vm, pred:$p)>;
+
+def : NEONInstAlias<"vrshl${p}.s8 $Vdn, $Vm",
+ (VRSHLsv16i8 QPR:$Vdn, QPR:$Vdn, QPR:$Vm, pred:$p)>;
+def : NEONInstAlias<"vrshl${p}.s16 $Vdn, $Vm",
+ (VRSHLsv8i16 QPR:$Vdn, QPR:$Vdn, QPR:$Vm, pred:$p)>;
+def : NEONInstAlias<"vrshl${p}.s32 $Vdn, $Vm",
+ (VRSHLsv4i32 QPR:$Vdn, QPR:$Vdn, QPR:$Vm, pred:$p)>;
+def : NEONInstAlias<"vrshl${p}.s64 $Vdn, $Vm",
+ (VRSHLsv2i64 QPR:$Vdn, QPR:$Vdn, QPR:$Vm, pred:$p)>;
+def : NEONInstAlias<"vrshl${p}.u8 $Vdn, $Vm",
+ (VRSHLuv16i8 QPR:$Vdn, QPR:$Vdn, QPR:$Vm, pred:$p)>;
+def : NEONInstAlias<"vrshl${p}.u16 $Vdn, $Vm",
+ (VRSHLuv8i16 QPR:$Vdn, QPR:$Vdn, QPR:$Vm, pred:$p)>;
+def : NEONInstAlias<"vrshl${p}.u32 $Vdn, $Vm",
+ (VRSHLuv4i32 QPR:$Vdn, QPR:$Vdn, QPR:$Vm, pred:$p)>;
+def : NEONInstAlias<"vrshl${p}.u64 $Vdn, $Vm",
+ (VRSHLuv2i64 QPR:$Vdn, QPR:$Vdn, QPR:$Vm, pred:$p)>;
+
// VLD1 single-lane pseudo-instructions. These need special handling for
// the lane index that an InstAlias can't handle, so we use these instead.
def VLD1LNdAsm_8 : NEONDataTypeAsmPseudoInst<"vld1${p}", ".8", "$list, $addr",
@@ -6911,6 +7022,100 @@ def : NEONInstAlias<"vsli${p}.32 $Vdm, $imm",
def : NEONInstAlias<"vsli${p}.64 $Vdm, $imm",
(VSLIv2i64 QPR:$Vdm, QPR:$Vdm, shr_imm64:$imm, pred:$p)>;
+// Two-operand variants for VHSUB.
+ // Signed.
+def : NEONInstAlias<"vhsub${p}.s8 $Vdn, $Vm",
+ (VHSUBsv8i8 DPR:$Vdn, DPR:$Vdn, DPR:$Vm, pred:$p)>;
+def : NEONInstAlias<"vhsub${p}.s16 $Vdn, $Vm",
+ (VHSUBsv4i16 DPR:$Vdn, DPR:$Vdn, DPR:$Vm, pred:$p)>;
+def : NEONInstAlias<"vhsub${p}.s32 $Vdn, $Vm",
+ (VHSUBsv2i32 DPR:$Vdn, DPR:$Vdn, DPR:$Vm, pred:$p)>;
+
+def : NEONInstAlias<"vhsub${p}.s8 $Vdn, $Vm",
+ (VHSUBsv16i8 QPR:$Vdn, QPR:$Vdn, QPR:$Vm, pred:$p)>;
+def : NEONInstAlias<"vhsub${p}.s16 $Vdn, $Vm",
+ (VHSUBsv8i16 QPR:$Vdn, QPR:$Vdn, QPR:$Vm, pred:$p)>;
+def : NEONInstAlias<"vhsub${p}.s32 $Vdn, $Vm",
+ (VHSUBsv4i32 QPR:$Vdn, QPR:$Vdn, QPR:$Vm, pred:$p)>;
+
+ // Unsigned.
+def : NEONInstAlias<"vhsub${p}.u8 $Vdn, $Vm",
+ (VHSUBuv8i8 DPR:$Vdn, DPR:$Vdn, DPR:$Vm, pred:$p)>;
+def : NEONInstAlias<"vhsub${p}.u16 $Vdn, $Vm",
+ (VHSUBuv4i16 DPR:$Vdn, DPR:$Vdn, DPR:$Vm, pred:$p)>;
+def : NEONInstAlias<"vhsub${p}.u32 $Vdn, $Vm",
+ (VHSUBuv2i32 DPR:$Vdn, DPR:$Vdn, DPR:$Vm, pred:$p)>;
+
+def : NEONInstAlias<"vhsub${p}.u8 $Vdn, $Vm",
+ (VHSUBuv16i8 QPR:$Vdn, QPR:$Vdn, QPR:$Vm, pred:$p)>;
+def : NEONInstAlias<"vhsub${p}.u16 $Vdn, $Vm",
+ (VHSUBuv8i16 QPR:$Vdn, QPR:$Vdn, QPR:$Vm, pred:$p)>;
+def : NEONInstAlias<"vhsub${p}.u32 $Vdn, $Vm",
+ (VHSUBuv4i32 QPR:$Vdn, QPR:$Vdn, QPR:$Vm, pred:$p)>;
+
+
+// Two-operand variants for VHADD.
+ // Signed.
+def : NEONInstAlias<"vhadd${p}.s8 $Vdn, $Vm",
+ (VHADDsv8i8 DPR:$Vdn, DPR:$Vdn, DPR:$Vm, pred:$p)>;
+def : NEONInstAlias<"vhadd${p}.s16 $Vdn, $Vm",
+ (VHADDsv4i16 DPR:$Vdn, DPR:$Vdn, DPR:$Vm, pred:$p)>;
+def : NEONInstAlias<"vhadd${p}.s32 $Vdn, $Vm",
+ (VHADDsv2i32 DPR:$Vdn, DPR:$Vdn, DPR:$Vm, pred:$p)>;
+
+def : NEONInstAlias<"vhadd${p}.s8 $Vdn, $Vm",
+ (VHADDsv16i8 QPR:$Vdn, QPR:$Vdn, QPR:$Vm, pred:$p)>;
+def : NEONInstAlias<"vhadd${p}.s16 $Vdn, $Vm",
+ (VHADDsv8i16 QPR:$Vdn, QPR:$Vdn, QPR:$Vm, pred:$p)>;
+def : NEONInstAlias<"vhadd${p}.s32 $Vdn, $Vm",
+ (VHADDsv4i32 QPR:$Vdn, QPR:$Vdn, QPR:$Vm, pred:$p)>;
+
+ // Unsigned.
+def : NEONInstAlias<"vhadd${p}.u8 $Vdn, $Vm",
+ (VHADDuv8i8 DPR:$Vdn, DPR:$Vdn, DPR:$Vm, pred:$p)>;
+def : NEONInstAlias<"vhadd${p}.u16 $Vdn, $Vm",
+ (VHADDuv4i16 DPR:$Vdn, DPR:$Vdn, DPR:$Vm, pred:$p)>;
+def : NEONInstAlias<"vhadd${p}.u32 $Vdn, $Vm",
+ (VHADDuv2i32 DPR:$Vdn, DPR:$Vdn, DPR:$Vm, pred:$p)>;
+
+def : NEONInstAlias<"vhadd${p}.u8 $Vdn, $Vm",
+ (VHADDuv16i8 QPR:$Vdn, QPR:$Vdn, QPR:$Vm, pred:$p)>;
+def : NEONInstAlias<"vhadd${p}.u16 $Vdn, $Vm",
+ (VHADDuv8i16 QPR:$Vdn, QPR:$Vdn, QPR:$Vm, pred:$p)>;
+def : NEONInstAlias<"vhadd${p}.u32 $Vdn, $Vm",
+ (VHADDuv4i32 QPR:$Vdn, QPR:$Vdn, QPR:$Vm, pred:$p)>;
+
+// Two-operand variants for VRHADD.
+ // Signed.
+def : NEONInstAlias<"vrhadd${p}.s8 $Vdn, $Rm",
+ (VRHADDsv8i8 DPR:$Vdn, DPR:$Vdn, DPR:$Rm, pred:$p)>;
+def : NEONInstAlias<"vrhadd${p}.s16 $Vdn, $Rm",
+ (VRHADDsv4i16 DPR:$Vdn, DPR:$Vdn, DPR:$Rm, pred:$p)>;
+def : NEONInstAlias<"vrhadd${p}.s32 $Vdn, $Rm",
+ (VRHADDsv2i32 DPR:$Vdn, DPR:$Vdn, DPR:$Rm, pred:$p)>;
+
+def : NEONInstAlias<"vrhadd${p}.s8 $Vdn, $Rm",
+ (VRHADDsv16i8 QPR:$Vdn, QPR:$Vdn, QPR:$Rm, pred:$p)>;
+def : NEONInstAlias<"vrhadd${p}.s16 $Vdn, $Rm",
+ (VRHADDsv8i16 QPR:$Vdn, QPR:$Vdn, QPR:$Rm, pred:$p)>;
+def : NEONInstAlias<"vrhadd${p}.s32 $Vdn, $Rm",
+ (VRHADDsv4i32 QPR:$Vdn, QPR:$Vdn, QPR:$Rm, pred:$p)>;
+
+ // Unsigned.
+def : NEONInstAlias<"vrhadd${p}.u8 $Vdn, $Rm",
+ (VRHADDuv8i8 DPR:$Vdn, DPR:$Vdn, DPR:$Rm, pred:$p)>;
+def : NEONInstAlias<"vrhadd${p}.u16 $Vdn, $Rm",
+ (VRHADDuv4i16 DPR:$Vdn, DPR:$Vdn, DPR:$Rm, pred:$p)>;
+def : NEONInstAlias<"vrhadd${p}.u32 $Vdn, $Rm",
+ (VRHADDuv2i32 DPR:$Vdn, DPR:$Vdn, DPR:$Rm, pred:$p)>;
+
+def : NEONInstAlias<"vrhadd${p}.u8 $Vdn, $Rm",
+ (VRHADDuv16i8 QPR:$Vdn, QPR:$Vdn, QPR:$Rm, pred:$p)>;
+def : NEONInstAlias<"vrhadd${p}.u16 $Vdn, $Rm",
+ (VRHADDuv8i16 QPR:$Vdn, QPR:$Vdn, QPR:$Rm, pred:$p)>;
+def : NEONInstAlias<"vrhadd${p}.u32 $Vdn, $Rm",
+ (VRHADDuv4i32 QPR:$Vdn, QPR:$Vdn, QPR:$Rm, pred:$p)>;
+
// VSWP allows, but does not require, a type suffix.
defm : NEONDTAnyInstAlias<"vswp${p}", "$Vd, $Vm",
(VSWPd DPR:$Vd, DPR:$Vm, pred:$p)>;
generated by cgit v1.1 at 2025-08-15 16:13:11 +0000