// 64-bit value.
def LoReg: OutPatFrag<(ops node:$Rs),
(EXTRACT_SUBREG (i64 $Rs), subreg_loreg)>;
+def HiReg: OutPatFrag<(ops node:$Rs),
+ (EXTRACT_SUBREG (i64 $Rs), subreg_hireg)>;
+
+// SDNode for converting immediate C to C-1.
+def DEC_CONST_SIGNED : SDNodeXForm<imm, [{
+ // Return the byte immediate const-1 as an SDNode.
+ int32_t imm = N->getSExtValue();
+ return XformSToSM1Imm(imm);
+}]>;
+
+// SDNode for converting immediate C to C-2.
+def DEC2_CONST_SIGNED : SDNodeXForm<imm, [{
+ // Return the byte immediate const-2 as an SDNode.
+ int32_t imm = N->getSExtValue();
+ return XformSToSM2Imm(imm);
+}]>;
+
+// SDNode for converting immediate C to C-3.
+def DEC3_CONST_SIGNED : SDNodeXForm<imm, [{
+ // Return the byte immediate const-3 as an SDNode.
+ int32_t imm = N->getSExtValue();
+ return XformSToSM3Imm(imm);
+}]>;
+
+// SDNode for converting immediate C to C-1.
+def DEC_CONST_UNSIGNED : SDNodeXForm<imm, [{
+ // Return the byte immediate const-1 as an SDNode.
+ uint32_t imm = N->getZExtValue();
+ return XformUToUM1Imm(imm);
+}]>;
//===----------------------------------------------------------------------===//
let Inst{1-0} = dst;
}
+let isCodeGenOnly = 0 in {
def C2_cmpeqi : T_CMP <"cmp.eq", 0b00, 0, s10Ext>;
def C2_cmpgti : T_CMP <"cmp.gt", 0b01, 0, s10Ext>;
def C2_cmpgtui : T_CMP <"cmp.gtu", 0b10, 0, u9Ext>;
+}
class T_CMP_pat <InstHexagon MI, PatFrag OpNode, PatLeaf ImmPred>
: Pat<(i1 (OpNode (i32 IntRegs:$src1), ImmPred:$src2)),
let AsmString = "$Rd = "#mnemonic#"($Rs, $Rt)"#suffix;
}
+let isCodeGenOnly = 0 in {
+def A2_svaddh : T_ALU32_3op<"vaddh", 0b110, 0b000, 0, 1>;
+def A2_svsubh : T_ALU32_3op<"vsubh", 0b110, 0b100, 1, 0>;
+}
+
let Defs = [USR_OVF], Itinerary = ALU32_3op_tc_2_SLOT0123,
isCodeGenOnly = 0 in {
+ def A2_svaddhs : T_ALU32_3op_sfx<"vaddh", ":sat", 0b110, 0b001, 0, 1>;
def A2_addsat : T_ALU32_3op_sfx<"add", ":sat", 0b110, 0b010, 0, 1>;
+ def A2_svadduhs : T_ALU32_3op_sfx<"vadduh", ":sat", 0b110, 0b011, 0, 1>;
+ def A2_svsubhs : T_ALU32_3op_sfx<"vsubh", ":sat", 0b110, 0b101, 1, 0>;
def A2_subsat : T_ALU32_3op_sfx<"sub", ":sat", 0b110, 0b110, 1, 0>;
+ def A2_svsubuhs : T_ALU32_3op_sfx<"vsubuh", ":sat", 0b110, 0b111, 1, 0>;
+}
+
+let Itinerary = ALU32_3op_tc_2_SLOT0123, isCodeGenOnly = 0 in
+def A2_svavghs : T_ALU32_3op_sfx<"vavgh", ":rnd", 0b111, 0b001, 0, 1>;
+
+let isCodeGenOnly = 0 in {
+def A2_svavgh : T_ALU32_3op<"vavgh", 0b111, 0b000, 0, 1>;
+def A2_svnavgh : T_ALU32_3op<"vnavgh", 0b111, 0b011, 1, 0>;
}
multiclass T_ALU32_3op_p<string mnemonic, bits<3> MajOp, bits<3> MinOp,
//===----------------------------------------------------------------------===//
// Template class for predicated ADD of a reg and an Immediate value.
//===----------------------------------------------------------------------===//
-let hasNewValue = 1 in
+let hasNewValue = 1, hasSideEffects = 0 in
class T_Addri_Pred <bit PredNot, bit PredNew>
: ALU32_ri <(outs IntRegs:$Rd),
(ins PredRegs:$Pu, IntRegs:$Rs, s8Ext:$s8),
//===----------------------------------------------------------------------===//
// A2_addi: Add a signed immediate to a register.
//===----------------------------------------------------------------------===//
-let hasNewValue = 1 in
+let hasNewValue = 1, hasSideEffects = 0 in
class T_Addri <Operand immOp, list<dag> pattern = [] >
: ALU32_ri <(outs IntRegs:$Rd),
(ins IntRegs:$Rs, immOp:$s16),
def: Pat<(sext_inreg I32:$src1, i8), (A2_sxtb I32:$src1)>;
def: Pat<(sext_inreg I32:$src1, i16), (A2_sxth I32:$src1)>;
-// Mux.
-def VMUX_prr64 : ALU64_rr<(outs DoubleRegs:$dst), (ins PredRegs:$src1,
- DoubleRegs:$src2,
- DoubleRegs:$src3),
- "$dst = vmux($src1, $src2, $src3)",
- []>;
-
-
//===----------------------------------------------------------------------===//
-// ALU32/PERM -
+// Template class for vector add and avg
//===----------------------------------------------------------------------===//
+class T_VectALU_64 <string opc, bits<3> majOp, bits<3> minOp,
+ bit isSat, bit isRnd, bit isCrnd, bit SwapOps >
+ : ALU64_rr < (outs DoubleRegs:$Rdd),
+ (ins DoubleRegs:$Rss, DoubleRegs:$Rtt),
+ "$Rdd = "#opc#"($Rss, $Rtt)"#!if(isRnd, ":rnd", "")
+ #!if(isCrnd,":crnd","")
+ #!if(isSat, ":sat", ""),
+ [], "", ALU64_tc_2_SLOT23 > {
+ bits<5> Rdd;
+ bits<5> Rss;
+ bits<5> Rtt;
+
+ let IClass = 0b1101;
+
+ let Inst{27-24} = 0b0011;
+ let Inst{23-21} = majOp;
+ let Inst{20-16} = !if (SwapOps, Rtt, Rss);
+ let Inst{12-8} = !if (SwapOps, Rss, Rtt);
+ let Inst{7-5} = minOp;
+ let Inst{4-0} = Rdd;
+ }
+
+// ALU64 - Vector add
+// Rdd=vadd[u][bhw](Rss,Rtt)
+let Itinerary = ALU64_tc_1_SLOT23, isCodeGenOnly = 0 in {
+ def A2_vaddub : T_VectALU_64 < "vaddub", 0b000, 0b000, 0, 0, 0, 0>;
+ def A2_vaddh : T_VectALU_64 < "vaddh", 0b000, 0b010, 0, 0, 0, 0>;
+ def A2_vaddw : T_VectALU_64 < "vaddw", 0b000, 0b101, 0, 0, 0, 0>;
+}
+
+// Rdd=vadd[u][bhw](Rss,Rtt):sat
+let Defs = [USR_OVF], isCodeGenOnly = 0 in {
+ def A2_vaddubs : T_VectALU_64 < "vaddub", 0b000, 0b001, 1, 0, 0, 0>;
+ def A2_vaddhs : T_VectALU_64 < "vaddh", 0b000, 0b011, 1, 0, 0, 0>;
+ def A2_vadduhs : T_VectALU_64 < "vadduh", 0b000, 0b100, 1, 0, 0, 0>;
+ def A2_vaddws : T_VectALU_64 < "vaddw", 0b000, 0b110, 1, 0, 0, 0>;
+}
+
+// ALU64 - Vector average
+// Rdd=vavg[u][bhw](Rss,Rtt)
+let Itinerary = ALU64_tc_1_SLOT23, isCodeGenOnly = 0 in {
+ def A2_vavgub : T_VectALU_64 < "vavgub", 0b010, 0b000, 0, 0, 0, 0>;
+ def A2_vavgh : T_VectALU_64 < "vavgh", 0b010, 0b010, 0, 0, 0, 0>;
+ def A2_vavguh : T_VectALU_64 < "vavguh", 0b010, 0b101, 0, 0, 0, 0>;
+ def A2_vavgw : T_VectALU_64 < "vavgw", 0b011, 0b000, 0, 0, 0, 0>;
+ def A2_vavguw : T_VectALU_64 < "vavguw", 0b011, 0b011, 0, 0, 0, 0>;
+}
+
+// Rdd=vavg[u][bhw](Rss,Rtt)[:rnd|:crnd]
+let isCodeGenOnly = 0 in {
+def A2_vavgubr : T_VectALU_64 < "vavgub", 0b010, 0b001, 0, 1, 0, 0>;
+def A2_vavghr : T_VectALU_64 < "vavgh", 0b010, 0b011, 0, 1, 0, 0>;
+def A2_vavghcr : T_VectALU_64 < "vavgh", 0b010, 0b100, 0, 0, 1, 0>;
+def A2_vavguhr : T_VectALU_64 < "vavguh", 0b010, 0b110, 0, 1, 0, 0>;
+}
+
+let isCodeGenOnly = 0 in {
+def A2_vavgwr : T_VectALU_64 < "vavgw", 0b011, 0b001, 0, 1, 0, 0>;
+def A2_vavgwcr : T_VectALU_64 < "vavgw", 0b011, 0b010, 0, 0, 1, 0>;
+def A2_vavguwr : T_VectALU_64 < "vavguw", 0b011, 0b100, 0, 1, 0, 0>;
+}
+
+// Rdd=vnavg[bh](Rss,Rtt)
+let Itinerary = ALU64_tc_1_SLOT23, isCodeGenOnly = 0 in {
+ def A2_vnavgh : T_VectALU_64 < "vnavgh", 0b100, 0b000, 0, 0, 0, 1>;
+ def A2_vnavgw : T_VectALU_64 < "vnavgw", 0b100, 0b011, 0, 0, 0, 1>;
+}
+
+// Rdd=vnavg[bh](Rss,Rtt)[:rnd|:crnd]:sat
+let Defs = [USR_OVF], isCodeGenOnly = 0 in {
+ def A2_vnavghr : T_VectALU_64 < "vnavgh", 0b100, 0b001, 1, 1, 0, 1>;
+ def A2_vnavghcr : T_VectALU_64 < "vnavgh", 0b100, 0b010, 1, 0, 1, 1>;
+ def A2_vnavgwr : T_VectALU_64 < "vnavgw", 0b100, 0b100, 1, 1, 0, 1>;
+ def A2_vnavgwcr : T_VectALU_64 < "vnavgw", 0b100, 0b110, 1, 0, 1, 1>;
+}
+
+// Rdd=vsub[u][bh](Rss,Rtt)
+let Itinerary = ALU64_tc_1_SLOT23, isCodeGenOnly = 0 in {
+ def A2_vsubub : T_VectALU_64 < "vsubub", 0b001, 0b000, 0, 0, 0, 1>;
+ def A2_vsubh : T_VectALU_64 < "vsubh", 0b001, 0b010, 0, 0, 0, 1>;
+ def A2_vsubw : T_VectALU_64 < "vsubw", 0b001, 0b101, 0, 0, 0, 1>;
+}
+
+// Rdd=vsub[u][bh](Rss,Rtt):sat
+let Defs = [USR_OVF], isCodeGenOnly = 0 in {
+ def A2_vsububs : T_VectALU_64 < "vsubub", 0b001, 0b001, 1, 0, 0, 1>;
+ def A2_vsubhs : T_VectALU_64 < "vsubh", 0b001, 0b011, 1, 0, 0, 1>;
+ def A2_vsubuhs : T_VectALU_64 < "vsubuh", 0b001, 0b100, 1, 0, 0, 1>;
+ def A2_vsubws : T_VectALU_64 < "vsubw", 0b001, 0b110, 1, 0, 0, 1>;
+}
+
+// Rdd=vmax[u][bhw](Rss,Rtt)
+let isCodeGenOnly = 0 in {
+def A2_vmaxb : T_VectALU_64 < "vmaxb", 0b110, 0b110, 0, 0, 0, 1>;
+def A2_vmaxub : T_VectALU_64 < "vmaxub", 0b110, 0b000, 0, 0, 0, 1>;
+def A2_vmaxh : T_VectALU_64 < "vmaxh", 0b110, 0b001, 0, 0, 0, 1>;
+def A2_vmaxuh : T_VectALU_64 < "vmaxuh", 0b110, 0b010, 0, 0, 0, 1>;
+def A2_vmaxw : T_VectALU_64 < "vmaxw", 0b110, 0b011, 0, 0, 0, 1>;
+def A2_vmaxuw : T_VectALU_64 < "vmaxuw", 0b101, 0b101, 0, 0, 0, 1>;
+}
+
+// Rdd=vmin[u][bhw](Rss,Rtt)
+let isCodeGenOnly = 0 in {
+def A2_vminb : T_VectALU_64 < "vminb", 0b110, 0b111, 0, 0, 0, 1>;
+def A2_vminub : T_VectALU_64 < "vminub", 0b101, 0b000, 0, 0, 0, 1>;
+def A2_vminh : T_VectALU_64 < "vminh", 0b101, 0b001, 0, 0, 0, 1>;
+def A2_vminuh : T_VectALU_64 < "vminuh", 0b101, 0b010, 0, 0, 0, 1>;
+def A2_vminw : T_VectALU_64 < "vminw", 0b101, 0b011, 0, 0, 0, 1>;
+def A2_vminuw : T_VectALU_64 < "vminuw", 0b101, 0b100, 0, 0, 0, 1>;
+}
//===----------------------------------------------------------------------===//
-// ALU32/PRED +
+// Template class for vector compare
//===----------------------------------------------------------------------===//
+let hasSideEffects = 0 in
+class T_vcmp <string Str, bits<4> minOp>
+ : ALU64_rr <(outs PredRegs:$Pd),
+ (ins DoubleRegs:$Rss, DoubleRegs:$Rtt),
+ "$Pd = "#Str#"($Rss, $Rtt)", [],
+ "", ALU64_tc_2early_SLOT23> {
+ bits<2> Pd;
+ bits<5> Rss;
+ bits<5> Rtt;
-// SDNode for converting immediate C to C-1.
-def DEC_CONST_SIGNED : SDNodeXForm<imm, [{
- // Return the byte immediate const-1 as an SDNode.
- int32_t imm = N->getSExtValue();
- return XformSToSM1Imm(imm);
-}]>;
+ let IClass = 0b1101;
-// SDNode for converting immediate C to C-1.
-def DEC_CONST_UNSIGNED : SDNodeXForm<imm, [{
- // Return the byte immediate const-1 as an SDNode.
- uint32_t imm = N->getZExtValue();
- return XformUToUM1Imm(imm);
-}]>;
+ let Inst{27-23} = 0b00100;
+ let Inst{13} = minOp{3};
+ let Inst{7-5} = minOp{2-0};
+ let Inst{1-0} = Pd;
+ let Inst{20-16} = Rss;
+ let Inst{12-8} = Rtt;
+ }
+
+class T_vcmp_pat<InstHexagon MI, PatFrag Op, ValueType T>
+ : Pat<(i1 (Op (T DoubleRegs:$Rss), (T DoubleRegs:$Rtt))),
+ (i1 (MI DoubleRegs:$Rss, DoubleRegs:$Rtt))>;
+
+// Vector compare bytes
+def A2_vcmpbeq : T_vcmp <"vcmpb.eq", 0b0110>;
+def A2_vcmpbgtu : T_vcmp <"vcmpb.gtu", 0b0111>;
-def CTLZ64_rr : SInst<(outs IntRegs:$dst), (ins DoubleRegs:$src1),
- "$dst = cl0($src1)",
- [(set (i32 IntRegs:$dst), (i32 (trunc (ctlz (i64 DoubleRegs:$src1)))))]>;
+// Vector compare halfwords
+def A2_vcmpheq : T_vcmp <"vcmph.eq", 0b0011>;
+def A2_vcmphgt : T_vcmp <"vcmph.gt", 0b0100>;
+def A2_vcmphgtu : T_vcmp <"vcmph.gtu", 0b0101>;
-def CTTZ64_rr : SInst<(outs IntRegs:$dst), (ins DoubleRegs:$src1),
- "$dst = ct0($src1)",
- [(set (i32 IntRegs:$dst), (i32 (trunc (cttz (i64 DoubleRegs:$src1)))))]>;
+// Vector compare words
+def A2_vcmpweq : T_vcmp <"vcmpw.eq", 0b0000>;
+def A2_vcmpwgt : T_vcmp <"vcmpw.gt", 0b0001>;
+def A2_vcmpwgtu : T_vcmp <"vcmpw.gtu", 0b0010>;
+
+def: T_vcmp_pat<A2_vcmpbeq, seteq, v8i8>;
+def: T_vcmp_pat<A2_vcmpbgtu, setugt, v8i8>;
+def: T_vcmp_pat<A2_vcmpheq, seteq, v4i16>;
+def: T_vcmp_pat<A2_vcmphgt, setgt, v4i16>;
+def: T_vcmp_pat<A2_vcmphgtu, setugt, v4i16>;
+def: T_vcmp_pat<A2_vcmpweq, seteq, v2i32>;
+def: T_vcmp_pat<A2_vcmpwgt, setgt, v2i32>;
+def: T_vcmp_pat<A2_vcmpwgtu, setugt, v2i32>;
+
+//===----------------------------------------------------------------------===//
+// ALU32/PERM -
+//===----------------------------------------------------------------------===//
+
+
+//===----------------------------------------------------------------------===//
+// ALU32/PRED +
+//===----------------------------------------------------------------------===//
//===----------------------------------------------------------------------===//
// ALU32/PRED -
def: T_cmp64_rr_pat<C2_cmpgtp, RevCmp<setlt>>;
def: T_cmp64_rr_pat<C2_cmpgtup, RevCmp<setult>>;
+let isCodeGenOnly = 0 in
+def C2_vmux : ALU64_rr<(outs DoubleRegs:$Rd),
+ (ins PredRegs:$Pu, DoubleRegs:$Rs, DoubleRegs:$Rt),
+ "$Rd = vmux($Pu, $Rs, $Rt)", [], "", ALU64_tc_1_SLOT23> {
+ let hasSideEffects = 0;
+
+ bits<5> Rd;
+ bits<2> Pu;
+ bits<5> Rs;
+ bits<5> Rt;
+
+ let IClass = 0b1101;
+ let Inst{27-24} = 0b0001;
+ let Inst{20-16} = Rs;
+ let Inst{12-8} = Rt;
+ let Inst{6-5} = Pu;
+ let Inst{4-0} = Rd;
+}
+
class T_ALU64_rr<string mnemonic, string suffix, bits<4> RegType,
bits<3> MajOp, bits<3> MinOp, bit OpsRev, bit IsComm,
string Op2Pfx>
let Inst{4-0} = Rd;
}
-def VALIGN_rrp : SInst<(outs DoubleRegs:$dst), (ins DoubleRegs:$src1,
- DoubleRegs:$src2,
- PredRegs:$src3),
- "$dst = valignb($src1, $src2, $src3)",
- []>;
-
-def VSPLICE_rrp : SInst<(outs DoubleRegs:$dst), (ins DoubleRegs:$src1,
- DoubleRegs:$src2,
- PredRegs:$src3),
- "$dst = vspliceb($src1, $src2, $src3)",
- []>;
-
// User control register transfer.
//===----------------------------------------------------------------------===//
// CR -
let accessSize = DoubleWordAccess, opExtentAlign = 3, isCodeGenOnly = 0 in
defm loadrd: LD_Idxd <"memd", "LDrid", DoubleRegs, s11_3Ext, u6_3Ext, 0b1110>;
-def : Pat < (i32 (sextloadi8 ADDRriS11_0:$addr)),
- (L2_loadrb_io AddrFI:$addr, 0) >;
-
-def : Pat < (i32 (zextloadi8 ADDRriS11_0:$addr)),
- (L2_loadrub_io AddrFI:$addr, 0) >;
-
-def : Pat < (i32 (sextloadi16 ADDRriS11_1:$addr)),
- (L2_loadrh_io AddrFI:$addr, 0) >;
-
-def : Pat < (i32 (zextloadi16 ADDRriS11_1:$addr)),
- (L2_loadruh_io AddrFI:$addr, 0) >;
+let accessSize = HalfWordAccess, opExtentAlign = 1, isCodeGenOnly = 0 in {
+ def L2_loadbsw2_io: T_load_io<"membh", IntRegs, 0b0001, s11_1Ext>;
+ def L2_loadbzw2_io: T_load_io<"memubh", IntRegs, 0b0011, s11_1Ext>;
+}
-def : Pat < (i32 (load ADDRriS11_2:$addr)),
- (L2_loadri_io AddrFI:$addr, 0) >;
+let accessSize = WordAccess, opExtentAlign = 2, isCodeGenOnly = 0 in {
+ def L2_loadbzw4_io: T_load_io<"memubh", DoubleRegs, 0b0101, s11_2Ext>;
+ def L2_loadbsw4_io: T_load_io<"membh", DoubleRegs, 0b0111, s11_2Ext>;
+}
-def : Pat < (i64 (load ADDRriS11_3:$addr)),
- (L2_loadrd_io AddrFI:$addr, 0) >;
+// Patterns to select load-indexed (i.e. load from base+offset).
+multiclass Loadx_pat<PatFrag Load, ValueType VT, PatLeaf ImmPred,
+ InstHexagon MI> {
+ def: Pat<(VT (Load AddrFI:$fi)), (VT (MI AddrFI:$fi, 0))>;
+ def: Pat<(VT (Load (add (i32 IntRegs:$Rs), ImmPred:$Off))),
+ (VT (MI IntRegs:$Rs, imm:$Off))>;
+ def: Pat<(VT (Load (i32 IntRegs:$Rs))), (VT (MI IntRegs:$Rs, 0))>;
+}
let AddedComplexity = 20 in {
-def : Pat < (i32 (sextloadi8 (add IntRegs:$src1, s11_0ExtPred:$offset))),
- (L2_loadrb_io IntRegs:$src1, s11_0ExtPred:$offset) >;
-
-def : Pat < (i32 (zextloadi8 (add IntRegs:$src1, s11_0ExtPred:$offset))),
- (L2_loadrub_io IntRegs:$src1, s11_0ExtPred:$offset) >;
-
-def : Pat < (i32 (sextloadi16 (add IntRegs:$src1, s11_1ExtPred:$offset))),
- (L2_loadrh_io IntRegs:$src1, s11_1ExtPred:$offset) >;
-
-def : Pat < (i32 (zextloadi16 (add IntRegs:$src1, s11_1ExtPred:$offset))),
- (L2_loadruh_io IntRegs:$src1, s11_1ExtPred:$offset) >;
-
-def : Pat < (i32 (load (add IntRegs:$src1, s11_2ExtPred:$offset))),
- (L2_loadri_io IntRegs:$src1, s11_2ExtPred:$offset) >;
-
-def : Pat < (i64 (load (add IntRegs:$src1, s11_3ExtPred:$offset))),
- (L2_loadrd_io IntRegs:$src1, s11_3ExtPred:$offset) >;
-}
+ defm: Loadx_pat<load, i32, s11_2ExtPred, L2_loadri_io>;
+ defm: Loadx_pat<load, i64, s11_3ExtPred, L2_loadrd_io>;
+ defm: Loadx_pat<atomic_load_8 , i32, s11_0ExtPred, L2_loadrub_io>;
+ defm: Loadx_pat<atomic_load_16, i32, s11_1ExtPred, L2_loadruh_io>;
+ defm: Loadx_pat<atomic_load_32, i32, s11_2ExtPred, L2_loadri_io>;
+ defm: Loadx_pat<atomic_load_64, i64, s11_3ExtPred, L2_loadrd_io>;
+
+ defm: Loadx_pat<extloadi1, i32, s11_0ExtPred, L2_loadrub_io>;
+ defm: Loadx_pat<extloadi8, i32, s11_0ExtPred, L2_loadrub_io>;
+ defm: Loadx_pat<extloadi16, i32, s11_1ExtPred, L2_loadruh_io>;
+ defm: Loadx_pat<sextloadi8, i32, s11_0ExtPred, L2_loadrb_io>;
+ defm: Loadx_pat<sextloadi16, i32, s11_1ExtPred, L2_loadrh_io>;
+ defm: Loadx_pat<zextloadi1, i32, s11_0ExtPred, L2_loadrub_io>;
+ defm: Loadx_pat<zextloadi8, i32, s11_0ExtPred, L2_loadrub_io>;
+ defm: Loadx_pat<zextloadi16, i32, s11_1ExtPred, L2_loadruh_io>;
+ // No sextloadi1.
+}
+
+// Sign-extending loads of i1 need to replicate the lowest bit throughout
+// the 32-bit value. Since the loaded value can only be 0 or 1, 0-v should
+// do the trick.
+let AddedComplexity = 20 in
+def: Pat<(i32 (sextloadi1 (i32 IntRegs:$Rs))),
+ (SUB_ri 0, (L2_loadrub_io IntRegs:$Rs, 0))>;
//===----------------------------------------------------------------------===//
// Post increment load
let accessSize = DoubleWordAccess, opExtentAlign = 3, isCodeGenOnly = 0 in
defm loadrd : LD_PostInc <"memd", "LDrid", DoubleRegs, s4_3Imm, 0b1110>;
-def : Pat< (i32 (extloadi1 ADDRriS11_0:$addr)),
- (i32 (L2_loadrb_io AddrFI:$addr, 0)) >;
-
-// Load byte any-extend.
-def : Pat < (i32 (extloadi8 ADDRriS11_0:$addr)),
- (i32 (L2_loadrb_io AddrFI:$addr, 0)) >;
-
-// Indexed load byte any-extend.
-let AddedComplexity = 20 in
-def : Pat < (i32 (extloadi8 (add IntRegs:$src1, s11_0ImmPred:$offset))),
- (i32 (L2_loadrb_io IntRegs:$src1, s11_0ImmPred:$offset)) >;
-
-def : Pat < (i32 (extloadi16 ADDRriS11_1:$addr)),
- (i32 (L2_loadrh_io AddrFI:$addr, 0))>;
-
-let AddedComplexity = 20 in
-def : Pat < (i32 (extloadi16 (add IntRegs:$src1, s11_1ImmPred:$offset))),
- (i32 (L2_loadrh_io IntRegs:$src1, s11_1ImmPred:$offset)) >;
-
-let AddedComplexity = 10 in
-def : Pat < (i32 (zextloadi1 ADDRriS11_0:$addr)),
- (i32 (L2_loadrub_io AddrFI:$addr, 0))>;
-
-let AddedComplexity = 20 in
-def : Pat < (i32 (zextloadi1 (add IntRegs:$src1, s11_0ImmPred:$offset))),
- (i32 (L2_loadrub_io IntRegs:$src1, s11_0ImmPred:$offset))>;
+// Rd=memb[u]h(Rx++#s4:1)
+// Rdd=memb[u]h(Rx++#s4:2)
+let accessSize = HalfWordAccess, opExtentAlign = 1, isCodeGenOnly = 0 in {
+ def L2_loadbsw2_pi : T_load_pi <"membh", IntRegs, s4_1Imm, 0b0001>;
+ def L2_loadbzw2_pi : T_load_pi <"memubh", IntRegs, s4_1Imm, 0b0011>;
+}
+let accessSize = WordAccess, opExtentAlign = 2, hasNewValue = 0,
+ isCodeGenOnly = 0 in {
+ def L2_loadbsw4_pi : T_load_pi <"membh", DoubleRegs, s4_2Imm, 0b0111>;
+ def L2_loadbzw4_pi : T_load_pi <"memubh", DoubleRegs, s4_2Imm, 0b0101>;
+}
//===----------------------------------------------------------------------===//
// Template class for post increment loads with register offset.
def L2_loadrh_pr : T_load_pr <"memh", IntRegs, 0b1010, HalfWordAccess>;
def L2_loadruh_pr : T_load_pr <"memuh", IntRegs, 0b1011, HalfWordAccess>;
def L2_loadri_pr : T_load_pr <"memw", IntRegs, 0b1100, WordAccess>;
+
+ def L2_loadbzw2_pr : T_load_pr <"memubh", IntRegs, 0b0011, HalfWordAccess>;
}
-let isCodeGenOnly = 0 in
+let isCodeGenOnly = 0 in {
def L2_loadrd_pr : T_load_pr <"memd", DoubleRegs, 0b1110, DoubleWordAccess>;
+def L2_loadbzw4_pr : T_load_pr <"memubh", DoubleRegs, 0b0101, WordAccess>;
+}
// Load predicate.
let isExtendable = 1, opExtendable = 2, isExtentSigned = 1, opExtentBits = 13,
let accessSize = HalfWordAccess, isCodeGenOnly = 0 in {
def L2_loadrh_pcr : T_load_pcr <"memh", IntRegs, 0b1010>;
def L2_loadruh_pcr : T_load_pcr <"memuh", IntRegs, 0b1011>;
+ def L2_loadbsw2_pcr : T_load_pcr <"membh", IntRegs, 0b0001>;
+ def L2_loadbzw2_pcr : T_load_pcr <"memubh", IntRegs, 0b0011>;
}
let accessSize = WordAccess, isCodeGenOnly = 0 in {
def L2_loadri_pcr : T_load_pcr <"memw", IntRegs, 0b1100>;
+ let hasNewValue = 0 in {
+ def L2_loadbzw4_pcr : T_load_pcr <"memubh", DoubleRegs, 0b0101>;
+ def L2_loadbsw4_pcr : T_load_pcr <"membh", DoubleRegs, 0b0111>;
+ }
}
let accessSize = DoubleWordAccess, isCodeGenOnly = 0 in
let accessSize = HalfWordAccess, isCodeGenOnly = 0 in {
def L2_loadrh_pci : T_load_pci <"memh", IntRegs, s4_1Imm, 0b1010>;
def L2_loadruh_pci : T_load_pci <"memuh", IntRegs, s4_1Imm, 0b1011>;
+ def L2_loadbzw2_pci : T_load_pci <"memubh", IntRegs, s4_1Imm, 0b0011>;
+ def L2_loadbsw2_pci : T_load_pci <"membh", IntRegs, s4_1Imm, 0b0001>;
}
// Word variants of circ load
let accessSize = WordAccess, isCodeGenOnly = 0 in
def L2_loadri_pci : T_load_pci <"memw", IntRegs, s4_2Imm, 0b1100>;
+let accessSize = WordAccess, hasNewValue = 0, isCodeGenOnly = 0 in {
+ def L2_loadbzw4_pci : T_load_pci <"memubh", DoubleRegs, s4_2Imm, 0b0101>;
+ def L2_loadbsw4_pci : T_load_pci <"membh", DoubleRegs, s4_2Imm, 0b0111>;
+}
+
let accessSize = DoubleWordAccess, hasNewValue = 0, isCodeGenOnly = 0 in
def L2_loadrd_pci : T_load_pci <"memd", DoubleRegs, s4_3Imm, 0b1110>;
def L2_loadw_locked : T_load_locked <"memw_locked", IntRegs>;
let accessSize = DoubleWordAccess, isCodeGenOnly = 0 in
def L4_loadd_locked : T_load_locked <"memd_locked", DoubleRegs>;
+
+// S[24]_store[wd]_locked: Store word/double conditionally.
+let isSoloAX = 1, isPredicateLate = 1 in
+class T_store_locked <string mnemonic, RegisterClass RC>
+ : ST0Inst <(outs PredRegs:$Pd), (ins IntRegs:$Rs, RC:$Rt),
+ mnemonic#"($Rs, $Pd) = $Rt"> {
+ bits<2> Pd;
+ bits<5> Rs;
+ bits<5> Rt;
+
+ let IClass = 0b1010;
+ let Inst{27-23} = 0b00001;
+ let Inst{22} = !if (!eq(mnemonic, "memw_locked"), 0b0, 0b1);
+ let Inst{21} = 0b1;
+ let Inst{20-16} = Rs;
+ let Inst{12-8} = Rt;
+ let Inst{1-0} = Pd;
+}
+
+let accessSize = WordAccess, isCodeGenOnly = 0 in
+def S2_storew_locked : T_store_locked <"memw_locked", IntRegs>;
+
+let accessSize = DoubleWordAccess, isCodeGenOnly = 0 in
+def S4_stored_locked : T_store_locked <"memd_locked", DoubleRegs>;
+
//===----------------------------------------------------------------------===//
// Bit-reversed loads with auto-increment register
//===----------------------------------------------------------------------===//
def L2_loadrub_pbr : T_load_pbr <"memub", IntRegs, ByteAccess, 0b1001>;
def L2_loadrh_pbr : T_load_pbr <"memh", IntRegs, HalfWordAccess, 0b1010>;
def L2_loadruh_pbr : T_load_pbr <"memuh", IntRegs, HalfWordAccess, 0b1011>;
+ def L2_loadbsw2_pbr : T_load_pbr <"membh", IntRegs, HalfWordAccess, 0b0001>;
+ def L2_loadbzw2_pbr : T_load_pbr <"memubh", IntRegs, HalfWordAccess, 0b0011>;
def L2_loadri_pbr : T_load_pbr <"memw", IntRegs, WordAccess, 0b1100>;
}
-let isCodeGenOnly = 0 in
+let isCodeGenOnly = 0 in {
+def L2_loadbzw4_pbr : T_load_pbr <"memubh", DoubleRegs, WordAccess, 0b0101>;
+def L2_loadbsw4_pbr : T_load_pbr <"membh", DoubleRegs, WordAccess, 0b0111>;
def L2_loadrd_pbr : T_load_pbr <"memd", DoubleRegs, DoubleWordAccess, 0b1110>;
+}
//===----------------------------------------------------------------------===//
// LD -
def M2_mpyud_nac_ll_s1: T_M2_mpyd_acc <0b00, 1, 1, 1>;
}
+//===----------------------------------------------------------------------===//
+// Template Class -- Vector Multipy
+// Used for complex multiply real or imaginary, dual multiply and even halfwords
+//===----------------------------------------------------------------------===//
+class T_M2_vmpy < string opc, bits<3> MajOp, bits<3> MinOp, bit hasShift,
+ bit isRnd, bit isSat >
+ : MInst <(outs DoubleRegs:$Rdd), (ins DoubleRegs:$Rss, DoubleRegs:$Rtt),
+ "$Rdd = "#opc#"($Rss, $Rtt)"#!if(hasShift,":<<1","")
+ #!if(isRnd,":rnd","")
+ #!if(isSat,":sat",""),
+ [] > {
+ bits<5> Rdd;
+ bits<5> Rss;
+ bits<5> Rtt;
+
+ let IClass = 0b1110;
+
+ let Inst{27-24} = 0b1000;
+ let Inst{23-21} = MajOp;
+ let Inst{7-5} = MinOp;
+ let Inst{4-0} = Rdd;
+ let Inst{20-16} = Rss;
+ let Inst{12-8} = Rtt;
+ }
+
+// Vector complex multiply imaginary: Rdd=vcmpyi(Rss,Rtt)[:<<1]:sat
+let Defs = [USR_OVF], isCodeGenOnly = 0 in {
+def M2_vcmpy_s1_sat_i: T_M2_vmpy <"vcmpyi", 0b110, 0b110, 1, 0, 1>;
+def M2_vcmpy_s0_sat_i: T_M2_vmpy <"vcmpyi", 0b010, 0b110, 0, 0, 1>;
+}
+
+// Vector complex multiply real: Rdd=vcmpyr(Rss,Rtt)[:<<1]:sat
+let isCodeGenOnly = 0 in {
+def M2_vcmpy_s1_sat_r: T_M2_vmpy <"vcmpyr", 0b101, 0b110, 1, 0, 1>;
+def M2_vcmpy_s0_sat_r: T_M2_vmpy <"vcmpyr", 0b001, 0b110, 0, 0, 1>;
+}
+
let hasNewValue = 1, opNewValue = 0 in
class T_MType_mpy <string mnemonic, bits<4> RegTyBits, RegisterClass RC,
bits<3> MajOp, bits<3> MinOp, bit isSat = 0, bit isRnd = 0,
let Inst{4-0} = dst;
}
+class T_MType_vrcmpy <string mnemonic, bits<3> MajOp, bits<3> MinOp, bit isHi>
+ : T_MType_mpy <mnemonic, 0b1001, DoubleRegs, MajOp, MinOp, 1, 1, "", 1, isHi>;
+
class T_MType_dd <string mnemonic, bits<3> MajOp, bits<3> MinOp,
bit isSat = 0, bit isRnd = 0 >
: T_MType_mpy <mnemonic, 0b1001, DoubleRegs, MajOp, MinOp, isSat, isRnd>;
bit isSat = 0, bit isRnd = 0, string op2str = "" >
: T_MType_mpy<mnemonic, 0b1101, IntRegs, MajOp, MinOp, isSat, isRnd, op2str>;
+let isCodeGenOnly = 0 in
+def M2_vradduh : T_MType_dd <"vradduh", 0b000, 0b001, 0, 0>;
+
let CextOpcode = "mpyi", InputType = "reg", isCodeGenOnly = 0 in
def M2_mpyi : T_MType_rr1 <"mpyi", 0b000, 0b000>, ImmRegRel;
def M2_hmmpyl_rs1 : T_MType_rr2 <"mpy", 0b111, 0b100, 1, 1, ".l">;
}
+let isCodeGenOnly = 0 in {
+def M2_cmpyrs_s0 : T_MType_rr2 <"cmpy", 0b001, 0b110, 1, 1>;
+def M2_cmpyrs_s1 : T_MType_rr2 <"cmpy", 0b101, 0b110, 1, 1>;
+def M2_cmpyrsc_s0 : T_MType_rr2 <"cmpy", 0b011, 0b110, 1, 1, "*">;
+def M2_cmpyrsc_s1 : T_MType_rr2 <"cmpy", 0b111, 0b110, 1, 1, "*">;
+}
+
// V4 Instructions
let isCodeGenOnly = 0 in {
+def M2_vraddh : T_MType_dd <"vraddh", 0b001, 0b111, 0>;
def M2_mpysu_up : T_MType_rr1 <"mpysu", 0b011, 0b001, 0>;
+def M2_mpy_up_s1 : T_MType_rr1 <"mpy", 0b101, 0b010, 0>;
def M2_mpy_up_s1_sat : T_MType_rr1 <"mpy", 0b111, 0b000, 1>;
def M2_hmmpyh_s1 : T_MType_rr2 <"mpy", 0b101, 0b000, 1, 0, ".h">;
def : T_MType_acc_pat1 <M2_naccii, add, sub, s8_16ExtPred>;
def : T_MType_acc_pat2 <M2_nacci, add, sub>;
+
+//===----------------------------------------------------------------------===//
+// Template Class -- XType Vector Instructions
+//===----------------------------------------------------------------------===//
+class T_XTYPE_Vect < string opc, bits<3> MajOp, bits<3> MinOp, bit isConj >
+ : MInst <(outs DoubleRegs:$Rdd), (ins DoubleRegs:$Rss, DoubleRegs:$Rtt),
+ "$Rdd = "#opc#"($Rss, $Rtt"#!if(isConj,"*)",")"),
+ [] > {
+ bits<5> Rdd;
+ bits<5> Rss;
+ bits<5> Rtt;
+
+ let IClass = 0b1110;
+
+ let Inst{27-24} = 0b1000;
+ let Inst{23-21} = MajOp;
+ let Inst{7-5} = MinOp;
+ let Inst{4-0} = Rdd;
+ let Inst{20-16} = Rss;
+ let Inst{12-8} = Rtt;
+ }
+
+class T_XTYPE_Vect_acc < string opc, bits<3> MajOp, bits<3> MinOp, bit isConj >
+ : MInst <(outs DoubleRegs:$Rdd),
+ (ins DoubleRegs:$dst2, DoubleRegs:$Rss, DoubleRegs:$Rtt),
+ "$Rdd += "#opc#"($Rss, $Rtt"#!if(isConj,"*)",")"),
+ [], "$dst2 = $Rdd",M_tc_3x_SLOT23 > {
+ bits<5> Rdd;
+ bits<5> Rss;
+ bits<5> Rtt;
+
+ let IClass = 0b1110;
+
+ let Inst{27-24} = 0b1010;
+ let Inst{23-21} = MajOp;
+ let Inst{7-5} = MinOp;
+ let Inst{4-0} = Rdd;
+ let Inst{20-16} = Rss;
+ let Inst{12-8} = Rtt;
+ }
+
+class T_XTYPE_Vect_diff < bits<3> MajOp, string opc >
+ : MInst <(outs DoubleRegs:$Rdd), (ins DoubleRegs:$Rtt, DoubleRegs:$Rss),
+ "$Rdd = "#opc#"($Rtt, $Rss)",
+ [], "",M_tc_2_SLOT23 > {
+ bits<5> Rdd;
+ bits<5> Rss;
+ bits<5> Rtt;
+
+ let IClass = 0b1110;
+
+ let Inst{27-24} = 0b1000;
+ let Inst{23-21} = MajOp;
+ let Inst{7-5} = 0b000;
+ let Inst{4-0} = Rdd;
+ let Inst{20-16} = Rss;
+ let Inst{12-8} = Rtt;
+ }
+
+// Vector reduce add unsigned bytes: Rdd32=vrmpybu(Rss32,Rtt32)
+let isCodeGenOnly = 0 in {
+def A2_vraddub: T_XTYPE_Vect <"vraddub", 0b010, 0b001, 0>;
+def A2_vraddub_acc: T_XTYPE_Vect_acc <"vraddub", 0b010, 0b001, 0>;
+}
+
+// Vector sum of absolute differences unsigned bytes: Rdd=vrsadub(Rss,Rtt)
+let isCodeGenOnly = 0 in {
+def A2_vrsadub: T_XTYPE_Vect <"vrsadub", 0b010, 0b010, 0>;
+def A2_vrsadub_acc: T_XTYPE_Vect_acc <"vrsadub", 0b010, 0b010, 0>;
+}
+
+// Vector absolute difference: Rdd=vabsdiffh(Rtt,Rss)
+let isCodeGenOnly = 0 in
+def M2_vabsdiffh: T_XTYPE_Vect_diff<0b011, "vabsdiffh">;
+
+// Vector reduce complex multiply real or imaginary:
+// Rdd[+]=vrcmpy[ir](Rss,Rtt[*])
+let isCodeGenOnly = 0 in {
+def M2_vrcmpyi_s0: T_XTYPE_Vect <"vrcmpyi", 0b000, 0b000, 0>;
+def M2_vrcmpyi_s0c: T_XTYPE_Vect <"vrcmpyi", 0b010, 0b000, 1>;
+def M2_vrcmaci_s0: T_XTYPE_Vect_acc <"vrcmpyi", 0b000, 0b000, 0>;
+def M2_vrcmaci_s0c: T_XTYPE_Vect_acc <"vrcmpyi", 0b010, 0b000, 1>;
+}
+
+let isCodeGenOnly = 0 in {
+def M2_vrcmpyr_s0: T_XTYPE_Vect <"vrcmpyr", 0b000, 0b001, 0>;
+def M2_vrcmpyr_s0c: T_XTYPE_Vect <"vrcmpyr", 0b011, 0b001, 1>;
+def M2_vrcmacr_s0: T_XTYPE_Vect_acc <"vrcmpyr", 0b000, 0b001, 0>;
+def M2_vrcmacr_s0c: T_XTYPE_Vect_acc <"vrcmpyr", 0b011, 0b001, 1>;
+}
+
+//===----------------------------------------------------------------------===//
+// Template Class -- Vector Multipy with accumulation.
+// Used for complex multiply real or imaginary, dual multiply and even halfwords
+//===----------------------------------------------------------------------===//
+let Defs = [USR_OVF] in
+class T_M2_vmpy_acc_sat < string opc, bits<3> MajOp, bits<3> MinOp,
+ bit hasShift, bit isRnd >
+ : MInst <(outs DoubleRegs:$Rxx),
+ (ins DoubleRegs:$dst2, DoubleRegs:$Rss, DoubleRegs:$Rtt),
+ "$Rxx += "#opc#"($Rss, $Rtt)"#!if(hasShift,":<<1","")
+ #!if(isRnd,":rnd","")#":sat",
+ [], "$dst2 = $Rxx",M_tc_3x_SLOT23 > {
+ bits<5> Rxx;
+ bits<5> Rss;
+ bits<5> Rtt;
+
+ let IClass = 0b1110;
+
+ let Inst{27-24} = 0b1010;
+ let Inst{23-21} = MajOp;
+ let Inst{7-5} = MinOp;
+ let Inst{4-0} = Rxx;
+ let Inst{20-16} = Rss;
+ let Inst{12-8} = Rtt;
+ }
+
+// Vector complex multiply real or imaginary with accumulation
+// Rxx+=vcmpy[ir](Rss,Rtt):sat
+let isCodeGenOnly = 0 in {
+def M2_vcmac_s0_sat_r: T_M2_vmpy_acc_sat <"vcmpyr", 0b001, 0b100, 0, 0>;
+def M2_vcmac_s0_sat_i: T_M2_vmpy_acc_sat <"vcmpyi", 0b010, 0b100, 0, 0>;
+}
+
//===----------------------------------------------------------------------===//
// Template Class -- Multiply signed/unsigned halfwords with and without
// saturation and rounding
def M2_dpmpyuu_acc_s0 : T_XTYPE_mpy64_acc < "mpyu", "+", 0b010, 0b000, 0, 0, 0>;
def M2_dpmpyuu_nac_s0 : T_XTYPE_mpy64_acc < "mpyu", "-", 0b011, 0b000, 0, 0, 0>;
}
+// Complex multiply real or imaginary
+// Rxx=cmpy[ir](Rs,Rt)
+let isCodeGenOnly = 0 in {
+def M2_cmpyi_s0 : T_XTYPE_mpy64 < "cmpyi", 0b000, 0b001, 0, 0, 0>;
+def M2_cmpyr_s0 : T_XTYPE_mpy64 < "cmpyr", 0b000, 0b010, 0, 0, 0>;
+}
+
+// Rxx+=cmpy[ir](Rs,Rt)
+let isCodeGenOnly = 0 in {
+def M2_cmaci_s0 : T_XTYPE_mpy64_acc < "cmpyi", "+", 0b000, 0b001, 0, 0, 0>;
+def M2_cmacr_s0 : T_XTYPE_mpy64_acc < "cmpyr", "+", 0b000, 0b010, 0, 0, 0>;
+}
+
+// Complex multiply
+// Rdd=cmpy(Rs,Rt)[:<<]:sat
+let isCodeGenOnly = 0 in {
+def M2_cmpys_s0 : T_XTYPE_mpy64 < "cmpy", 0b000, 0b110, 1, 0, 0>;
+def M2_cmpys_s1 : T_XTYPE_mpy64 < "cmpy", 0b100, 0b110, 1, 1, 0>;
+}
+
+// Rdd=cmpy(Rs,Rt*)[:<<]:sat
+let isCodeGenOnly = 0 in {
+def M2_cmpysc_s0 : T_XTYPE_mpy64 < "cmpy", 0b010, 0b110, 1, 0, 1>;
+def M2_cmpysc_s1 : T_XTYPE_mpy64 < "cmpy", 0b110, 0b110, 1, 1, 1>;
+}
+
+// Rxx[-+]=cmpy(Rs,Rt)[:<<1]:sat
+let isCodeGenOnly = 0 in {
+def M2_cmacs_s0 : T_XTYPE_mpy64_acc < "cmpy", "+", 0b000, 0b110, 1, 0, 0>;
+def M2_cnacs_s0 : T_XTYPE_mpy64_acc < "cmpy", "-", 0b000, 0b111, 1, 0, 0>;
+def M2_cmacs_s1 : T_XTYPE_mpy64_acc < "cmpy", "+", 0b100, 0b110, 1, 1, 0>;
+def M2_cnacs_s1 : T_XTYPE_mpy64_acc < "cmpy", "-", 0b100, 0b111, 1, 1, 0>;
+}
+
+// Rxx[-+]=cmpy(Rs,Rt*)[:<<1]:sat
+let isCodeGenOnly = 0 in {
+def M2_cmacsc_s0 : T_XTYPE_mpy64_acc < "cmpy", "+", 0b010, 0b110, 1, 0, 1>;
+def M2_cnacsc_s0 : T_XTYPE_mpy64_acc < "cmpy", "-", 0b010, 0b111, 1, 0, 1>;
+def M2_cmacsc_s1 : T_XTYPE_mpy64_acc < "cmpy", "+", 0b110, 0b110, 1, 1, 1>;
+def M2_cnacsc_s1 : T_XTYPE_mpy64_acc < "cmpy", "-", 0b110, 0b111, 1, 1, 1>;
+}
def: Pat<(i64 (mul (i64 (anyext (i32 IntRegs:$src1))),
(i64 (anyext (i32 IntRegs:$src2))))),
let accessSize = HalfWordAccess, isNVStorable = 0, isCodeGenOnly = 0 in
defm storerf: ST_PostInc <"memh", "STrih_H", IntRegs, s4_1Imm, 0b1011, 1>;
+// Patterns for generating stores, where the address takes different forms:
+// - frameindex,,
+// - base + offset,
+// - simple (base address without offset).
+// These would usually be used together (via Storex_pat defined below), but
+// in some cases one may want to apply different properties (such as
+// AddedComplexity) to the individual patterns.
+class Storex_fi_pat<PatFrag Store, PatFrag Value, InstHexagon MI>
+ : Pat<(Store Value:$Rs, AddrFI:$fi), (MI AddrFI:$fi, 0, Value:$Rs)>;
+class Storex_add_pat<PatFrag Store, PatFrag Value, PatFrag ImmPred,
+ InstHexagon MI>
+ : Pat<(Store Value:$Rt, (add (i32 IntRegs:$Rs), ImmPred:$Off)),
+ (MI IntRegs:$Rs, imm:$Off, Value:$Rt)>;
+
+multiclass Storex_pat<PatFrag Store, PatFrag Value, PatLeaf ImmPred,
+ InstHexagon MI> {
+ def: Storex_fi_pat <Store, Value, MI>;
+ def: Storex_add_pat <Store, Value, ImmPred, MI>;
+}
+
def : Pat<(post_truncsti8 (i32 IntRegs:$src1), IntRegs:$src2,
s4_3ImmPred:$offset),
(S2_storerb_pi IntRegs:$src2, s4_0ImmPred:$offset, IntRegs:$src1)>;
u6_1Ext, 0b011, 1>;
}
+class Storex_simple_pat<PatFrag Store, PatFrag Value, InstHexagon MI>
+ : Pat<(Store Value:$Rt, (i32 IntRegs:$Rs)),
+ (MI IntRegs:$Rs, 0, Value:$Rt)>;
+
+// Regular stores in the DAG have two operands: value and address.
+// Atomic stores also have two, but they are reversed: address, value.
+// To use atomic stores with the patterns, they need to have their operands
+// swapped. This relies on the knowledge that the F.Fragment uses names
+// "ptr" and "val".
+class SwapSt<PatFrag F>
+ : PatFrag<(ops node:$val, node:$ptr), F.Fragment>;
+
+def: Storex_simple_pat<SwapSt<atomic_store_8>, I32, S2_storerb_io>;
+def: Storex_simple_pat<SwapSt<atomic_store_16>, I32, S2_storerh_io>;
+def: Storex_simple_pat<SwapSt<atomic_store_32>, I32, S2_storeri_io>;
+def: Storex_simple_pat<SwapSt<atomic_store_64>, I64, S2_storerd_io>;
+
def : Pat<(truncstorei8 (i32 IntRegs:$src1), ADDRriS11_0:$addr),
(S2_storerb_io AddrFI:$addr, 0, (i32 IntRegs:$src1))>;
DoubleWordAccess>;
}
+let Uses = [CS], isNewValue = 1, mayStore = 1, isNVStore = 1, opNewValue = 4 in
+class T_storenew_pci <string mnemonic, Operand Imm,
+ bits<2>MajOp, MemAccessSize AlignSize>
+ : NVInst < (outs IntRegs:$_dst_),
+ (ins IntRegs:$Rz, Imm:$offset, ModRegs:$Mu, IntRegs:$Nt),
+ #mnemonic#"($Rz ++ #$offset:circ($Mu)) = $Nt.new",
+ [],
+ "$Rz = $_dst_"> {
+ bits<5> Rz;
+ bits<6> offset;
+ bits<1> Mu;
+ bits<3> Nt;
+
+ let accessSize = AlignSize;
+
+ let IClass = 0b1010;
+ let Inst{27-21} = 0b1001101;
+ let Inst{20-16} = Rz;
+ let Inst{13} = Mu;
+ let Inst{12-11} = MajOp;
+ let Inst{10-8} = Nt;
+ let Inst{7} = 0b0;
+ let Inst{6-3} =
+ !if (!eq(!cast<string>(AlignSize), "WordAccess"), offset{5-2},
+ !if (!eq(!cast<string>(AlignSize), "HalfWordAccess"), offset{4-1},
+ /* ByteAccess */ offset{3-0}));
+ let Inst{1} = 0b0;
+ }
+let isCodeGenOnly = 0 in {
+def S2_storerbnew_pci : T_storenew_pci <"memb", s4_0Imm, 0b00, ByteAccess>;
+def S2_storerhnew_pci : T_storenew_pci <"memh", s4_1Imm, 0b01, HalfWordAccess>;
+def S2_storerinew_pci : T_storenew_pci <"memw", s4_2Imm, 0b10, WordAccess>;
+}
+
//===----------------------------------------------------------------------===//
// Circular stores with auto-increment register
//===----------------------------------------------------------------------===//
HalfWordAccess, "Rt.h">;
}
+//===----------------------------------------------------------------------===//
+// Circular .new stores with auto-increment register
+//===----------------------------------------------------------------------===//
+let Uses = [CS], isNewValue = 1, mayStore = 1, isNVStore = 1, opNewValue = 3 in
+class T_storenew_pcr <string mnemonic, bits<2>MajOp,
+ MemAccessSize AlignSize>
+ : NVInst <(outs IntRegs:$_dst_),
+ (ins IntRegs:$Rz, ModRegs:$Mu, IntRegs:$Nt),
+ #mnemonic#"($Rz ++ I:circ($Mu)) = $Nt.new" ,
+ [] ,
+ "$Rz = $_dst_"> {
+ bits<5> Rz;
+ bits<1> Mu;
+ bits<3> Nt;
+
+ let accessSize = AlignSize;
+
+ let IClass = 0b1010;
+ let Inst{27-21} = 0b1001101;
+ let Inst{20-16} = Rz;
+ let Inst{13} = Mu;
+ let Inst{12-11} = MajOp;
+ let Inst{10-8} = Nt;
+ let Inst{7} = 0b0;
+ let Inst{1} = 0b1;
+ }
+
+let isCodeGenOnly = 0 in {
+def S2_storerbnew_pcr : T_storenew_pcr <"memb", 0b00, ByteAccess>;
+def S2_storerhnew_pcr : T_storenew_pcr <"memh", 0b01, HalfWordAccess>;
+def S2_storerinew_pcr : T_storenew_pcr <"memw", 0b10, WordAccess>;
+}
+
//===----------------------------------------------------------------------===//
// Bit-reversed stores with auto-increment register
//===----------------------------------------------------------------------===//
def S2_storerd_pbr : T_store_pbr<"memd", DoubleRegs, DoubleWordAccess, 0b110>;
}
+//===----------------------------------------------------------------------===//
+// Bit-reversed .new stores with auto-increment register
+//===----------------------------------------------------------------------===//
+let isNewValue = 1, mayStore = 1, isNVStore = 1, opNewValue = 3,
+ hasSideEffects = 0 in
+class T_storenew_pbr<string mnemonic, MemAccessSize addrSize, bits<2> majOp>
+ : NVInst <(outs IntRegs:$_dst_),
+ (ins IntRegs:$Rz, ModRegs:$Mu, IntRegs:$Nt),
+ #mnemonic#"($Rz ++ $Mu:brev) = $Nt.new", [],
+ "$Rz = $_dst_">, NewValueRel {
+ let accessSize = addrSize;
+ bits<5> Rz;
+ bits<1> Mu;
+ bits<3> Nt;
+
+ let IClass = 0b1010;
+
+ let Inst{27-21} = 0b1111101;
+ let Inst{12-11} = majOp;
+ let Inst{7} = 0b0;
+ let Inst{20-16} = Rz;
+ let Inst{13} = Mu;
+ let Inst{10-8} = Nt;
+ }
+
+let BaseOpcode = "S2_storerb_pbr", isCodeGenOnly = 0 in
+def S2_storerbnew_pbr : T_storenew_pbr<"memb", ByteAccess, 0b00>;
+
+let BaseOpcode = "S2_storerh_pbr", isCodeGenOnly = 0 in
+def S2_storerhnew_pbr : T_storenew_pbr<"memh", HalfWordAccess, 0b01>;
+
+let BaseOpcode = "S2_storeri_pbr", isCodeGenOnly = 0 in
+def S2_storerinew_pbr : T_storenew_pbr<"memw", WordAccess, 0b10>;
+
//===----------------------------------------------------------------------===//
// ST -
//===----------------------------------------------------------------------===//
def A2_satub : T_S2op_1_ii <"satub", 0b11, 0b110>;
def A2_sath : T_S2op_1_ii <"sath", 0b11, 0b100>;
def A2_satuh : T_S2op_1_ii <"satuh", 0b11, 0b101>;
+ def A2_roundsat : T_S2op_1_id <"round", 0b11, 0b001, 0b1>;
}
let Itinerary = S_2op_tc_2_SLOT23, isCodeGenOnly = 0 in {
+ // Bit reverse
+ def S2_brev : T_S2op_1_ii <"brev", 0b01, 0b110>;
+
// Absolute value word
def A2_abs : T_S2op_1_ii <"abs", 0b10, 0b100>;
let Inst{7-5} = MinOp;
let Inst{4-0} = dst;
}
+
+class T_S2op_2_di <string mnemonic, bits<3> MajOp, bits<3> MinOp>
+ : T_S2op_2 <mnemonic, 0b1000, DoubleRegs, IntRegs, MajOp, MinOp, 0, 0>;
+
+let hasNewValue = 1 in
+class T_S2op_2_id <string mnemonic, bits<3> MajOp, bits<3> MinOp>
+ : T_S2op_2 <mnemonic, 0b1000, IntRegs, DoubleRegs, MajOp, MinOp, 0, 0>;
let hasNewValue = 1 in
class T_S2op_2_ii <string mnemonic, bits<3> MajOp, bits<3> MinOp,
let isCodeGenOnly = 0 in
def S2_asr_i_r_rnd : T_S2op_2_ii <"asr", 0b010, 0b000, 0, 1>;
+def S2_asr_i_r_rnd_goodsyntax
+ : SInst <(outs IntRegs:$dst), (ins IntRegs:$src, u5Imm:$u5),
+ "$dst = asrrnd($src, #$u5)",
+ [], "", S_2op_tc_1_SLOT23>;
+
def: Pat<(i32 (sra (i32 (add (i32 (sra I32:$src1, u5ImmPred:$src2)),
(i32 1))),
(i32 1))),
def S2_deinterleave : T_S2op_3 <"deinterleave", 0b11, 0b100>;
}
+// Vector Complex conjugate
+let isCodeGenOnly = 0 in
+def A2_vconj : T_S2op_3 <"vconj", 0b10, 0b111, 1>;
+
+// Vector absolute value halfwords with and without saturation
+// Rdd64=vabsh(Rss64)[:sat]
+let isCodeGenOnly = 0 in {
+def A2_vabsh : T_S2op_3 <"vabsh", 0b01, 0b100>;
+def A2_vabshsat : T_S2op_3 <"vabsh", 0b01, 0b101, 1>;
+}
+
+// Vector absolute value words with and without saturation
+let isCodeGenOnly = 0 in {
+def A2_vabsw : T_S2op_3 <"vabsw", 0b01, 0b110>;
+def A2_vabswsat : T_S2op_3 <"vabsw", 0b01, 0b111, 1>;
+}
+
//===----------------------------------------------------------------------===//
// STYPE/BIT +
//===----------------------------------------------------------------------===//
defm J2_ploop3s : SPLOOP_ri<"3", 0b11>;
}
+
+// if (Rs[!>=<]=#0) jump:[t/nt]
+let Defs = [PC], isPredicated = 1, isBranch = 1, hasSideEffects = 0,
+ hasSideEffects = 0 in
+class J2_jump_0_Base<string compare, bit isTak, bits<2> op>
+ : CRInst <(outs), (ins IntRegs:$Rs, brtarget:$r13_2),
+ "if ($Rs"#compare#"#0) jump"#!if(isTak, ":t", ":nt")#" $r13_2" > {
+ bits<5> Rs;
+ bits<15> r13_2;
+
+ let IClass = 0b0110;
+
+ let Inst{27-24} = 0b0001;
+ let Inst{23-22} = op;
+ let Inst{12} = isTak;
+ let Inst{21} = r13_2{14};
+ let Inst{20-16} = Rs;
+ let Inst{11-1} = r13_2{12-2};
+ let Inst{13} = r13_2{13};
+ }
+
+multiclass J2_jump_compare_0<string compare, bits<2> op> {
+ def NAME : J2_jump_0_Base<compare, 0, op>;
+ def NAME#pt : J2_jump_0_Base<compare, 1, op>;
+}
+let isCodeGenOnly = 0 in {
+defm J2_jumprz : J2_jump_compare_0<"!=", 0b00>;
+defm J2_jumprgtez : J2_jump_compare_0<">=", 0b01>;
+defm J2_jumprnz : J2_jump_compare_0<"==", 0b10>;
+defm J2_jumprltez : J2_jump_compare_0<"<=", 0b11>;
+}
+
// Transfer to/from Control/GPR Guest/GPR
let hasSideEffects = 0 in
class TFR_CR_RS_base<RegisterClass CTRC, RegisterClass RC, bit isDouble>
// Support for generating global address.
// Taken from X86InstrInfo.td.
-def SDTHexagonCONST32 : SDTypeProfile<1, 1, [
- SDTCisVT<0, i32>,
- SDTCisVT<1, i32>,
- SDTCisPtrTy<0>]>;
-def HexagonCONST32 : SDNode<"HexagonISD::CONST32", SDTHexagonCONST32>;
+def SDTHexagonCONST32 : SDTypeProfile<1, 1, [SDTCisVT<0, i32>,
+ SDTCisVT<1, i32>,
+ SDTCisPtrTy<0>]>;
+def HexagonCONST32 : SDNode<"HexagonISD::CONST32", SDTHexagonCONST32>;
def HexagonCONST32_GP : SDNode<"HexagonISD::CONST32_GP", SDTHexagonCONST32>;
// HI/LO Instructions
"$dst = xor($dst, $dst)",
[(set (i1 PredRegs:$dst), 0)]>;
-def MPY_trsext : MInst<(outs IntRegs:$dst), (ins IntRegs:$src1, IntRegs:$src2),
- "$dst = mpy($src1, $src2)",
- [(set (i32 IntRegs:$dst),
- (trunc (i64 (srl (i64 (mul (i64 (sext (i32 IntRegs:$src1))),
- (i64 (sext (i32 IntRegs:$src2))))),
- (i32 32)))))]>;
-
// Pseudo instructions.
def SDT_SPCallSeqStart : SDCallSeqStart<[ SDTCisVT<0, i32> ]>;
-
-def SDT_SPCallSeqEnd : SDCallSeqEnd<[ SDTCisVT<0, i32>,
+def SDT_SPCallSeqEnd : SDCallSeqEnd<[ SDTCisVT<0, i32>,
SDTCisVT<1, i32> ]>;
-def callseq_end : SDNode<"ISD::CALLSEQ_END", SDT_SPCallSeqEnd,
- [SDNPHasChain, SDNPOptInGlue, SDNPOutGlue]>;
-
def callseq_start : SDNode<"ISD::CALLSEQ_START", SDT_SPCallSeqStart,
[SDNPHasChain, SDNPOutGlue]>;
+def callseq_end : SDNode<"ISD::CALLSEQ_END", SDT_SPCallSeqEnd,
+ [SDNPHasChain, SDNPOptInGlue, SDNPOutGlue]>;
-def SDT_SPCall : SDTypeProfile<0, 1, [SDTCisVT<0, i32>]>;
-
-def call : SDNode<"HexagonISD::CALL", SDT_SPCall,
- [SDNPHasChain, SDNPOptInGlue, SDNPOutGlue, SDNPVariadic]>;
+def SDT_SPCall : SDTypeProfile<0, 1, [SDTCisVT<0, i32>]>;
// For tailcalls a HexagonTCRet SDNode has 3 SDNode Properties - a chain,
// Optional Flag and Variable Arguments.
[], "", J_tc_2early_SLOT23>;
}
-// Map call instruction.
-def : Pat<(call (i32 IntRegs:$dst)),
- (J2_callr (i32 IntRegs:$dst))>, Requires<[HasV2TOnly]>;
-def : Pat<(call tglobaladdr:$dst),
- (CALL tglobaladdr:$dst)>, Requires<[HasV2TOnly]>;
-def : Pat<(call texternalsym:$dst),
- (CALL texternalsym:$dst)>, Requires<[HasV2TOnly]>;
//Tail calls.
def : Pat<(HexagonTCRet tglobaladdr:$dst),
(TCRETURNtg tglobaladdr:$dst)>;
def : Pat<(HexagonTCRet (i32 IntRegs:$dst)),
(TCRETURNR (i32 IntRegs:$dst))>;
-// Atomic load and store support
-// 8 bit atomic load
-def : Pat<(atomic_load_8 ADDRriS11_0:$src1),
- (i32 (L2_loadrub_io AddrFI:$src1, 0))>;
-
-def : Pat<(atomic_load_8 (add (i32 IntRegs:$src1), s11_0ImmPred:$offset)),
- (i32 (L2_loadrub_io (i32 IntRegs:$src1), s11_0ImmPred:$offset))>;
-
-// 16 bit atomic load
-def : Pat<(atomic_load_16 ADDRriS11_1:$src1),
- (i32 (L2_loadruh_io AddrFI:$src1, 0))>;
-
-def : Pat<(atomic_load_16 (add (i32 IntRegs:$src1), s11_1ImmPred:$offset)),
- (i32 (L2_loadruh_io (i32 IntRegs:$src1), s11_1ImmPred:$offset))>;
-
-def : Pat<(atomic_load_32 ADDRriS11_2:$src1),
- (i32 (L2_loadri_io AddrFI:$src1, 0))>;
-
-def : Pat<(atomic_load_32 (add (i32 IntRegs:$src1), s11_2ImmPred:$offset)),
- (i32 (L2_loadri_io (i32 IntRegs:$src1), s11_2ImmPred:$offset))>;
-
-// 64 bit atomic load
-def : Pat<(atomic_load_64 ADDRriS11_3:$src1),
- (i64 (L2_loadrd_io AddrFI:$src1, 0))>;
-
-def : Pat<(atomic_load_64 (add (i32 IntRegs:$src1), s11_3ImmPred:$offset)),
- (i64 (L2_loadrd_io (i32 IntRegs:$src1), s11_3ImmPred:$offset))>;
-
-
-def : Pat<(atomic_store_8 ADDRriS11_0:$src2, (i32 IntRegs:$src1)),
- (S2_storerb_io AddrFI:$src2, 0, (i32 IntRegs:$src1))>;
-
-def : Pat<(atomic_store_8 (add (i32 IntRegs:$src2), s11_0ImmPred:$offset),
- (i32 IntRegs:$src1)),
- (S2_storerb_io (i32 IntRegs:$src2), s11_0ImmPred:$offset,
- (i32 IntRegs:$src1))>;
-
-
-def : Pat<(atomic_store_16 ADDRriS11_1:$src2, (i32 IntRegs:$src1)),
- (S2_storerh_io AddrFI:$src2, 0, (i32 IntRegs:$src1))>;
-
-def : Pat<(atomic_store_16 (i32 IntRegs:$src1),
- (add (i32 IntRegs:$src2), s11_1ImmPred:$offset)),
- (S2_storerh_io (i32 IntRegs:$src2), s11_1ImmPred:$offset,
- (i32 IntRegs:$src1))>;
-
-def : Pat<(atomic_store_32 ADDRriS11_2:$src2, (i32 IntRegs:$src1)),
- (S2_storeri_io AddrFI:$src2, 0, (i32 IntRegs:$src1))>;
-
-def : Pat<(atomic_store_32 (add (i32 IntRegs:$src2), s11_2ImmPred:$offset),
- (i32 IntRegs:$src1)),
- (S2_storeri_io (i32 IntRegs:$src2), s11_2ImmPred:$offset,
- (i32 IntRegs:$src1))>;
-
-
-
-
-def : Pat<(atomic_store_64 ADDRriS11_3:$src2, (i64 DoubleRegs:$src1)),
- (S2_storerd_io AddrFI:$src2, 0, (i64 DoubleRegs:$src1))>;
-
-def : Pat<(atomic_store_64 (add (i32 IntRegs:$src2), s11_3ImmPred:$offset),
- (i64 DoubleRegs:$src1)),
- (S2_storerd_io (i32 IntRegs:$src2), s11_3ImmPred:$offset,
- (i64 DoubleRegs:$src1))>;
-
// Map from r0 = and(r1, 65535) to r0 = zxth(r1)
def : Pat <(and (i32 IntRegs:$src1), 65535),
(A2_zxth (i32 IntRegs:$src1))>;
let isCodeGenOnly = 0 in
def S2_lfsp : T_S3op_64 < "lfs", 0b10, 0b110, 0>;
+let hasSideEffects = 0 in
+class T_S3op_2 <string mnemonic, bits<3> MajOp, bit SwapOps>
+ : SInst < (outs DoubleRegs:$Rdd),
+ (ins DoubleRegs:$Rss, DoubleRegs:$Rtt, PredRegs:$Pu),
+ "$Rdd = "#mnemonic#"($Rss, $Rtt, $Pu)",
+ [], "", S_3op_tc_1_SLOT23 > {
+ bits<5> Rdd;
+ bits<5> Rss;
+ bits<5> Rtt;
+ bits<2> Pu;
+
+ let IClass = 0b1100;
+
+ let Inst{27-24} = 0b0010;
+ let Inst{23-21} = MajOp;
+ let Inst{20-16} = !if (SwapOps, Rtt, Rss);
+ let Inst{12-8} = !if (SwapOps, Rss, Rtt);
+ let Inst{6-5} = Pu;
+ let Inst{4-0} = Rdd;
+ }
+
+let isCodeGenOnly = 0 in {
+def S2_valignrb : T_S3op_2 < "valignb", 0b000, 1>;
+def S2_vsplicerb : T_S3op_2 < "vspliceb", 0b100, 0>;
+}
+
//===----------------------------------------------------------------------===//
// Template class used by vector shift, vector rotate, vector neg,
// 32-bit shift, 64-bit shifts, etc.
def S2_asl_r_r_sat : T_S3op_shift32_Sat<"asl", 0b10>;
}
+let hasNewValue = 1, hasSideEffects = 0 in
+class T_S3op_8 <string opc, bits<3> MinOp, bit isSat, bit isRnd, bit hasShift, bit hasSplat = 0>
+ : SInst < (outs IntRegs:$Rd),
+ (ins DoubleRegs:$Rss, IntRegs:$Rt),
+ "$Rd = "#opc#"($Rss, $Rt"#!if(hasSplat, "*", "")#")"
+ #!if(hasShift, ":<<1", "")
+ #!if(isRnd, ":rnd", "")
+ #!if(isSat, ":sat", ""),
+ [], "", S_3op_tc_1_SLOT23 > {
+ bits<5> Rd;
+ bits<5> Rss;
+ bits<5> Rt;
+
+ let IClass = 0b1100;
+
+ let Inst{27-24} = 0b0101;
+ let Inst{20-16} = Rss;
+ let Inst{12-8} = Rt;
+ let Inst{7-5} = MinOp;
+ let Inst{4-0} = Rd;
+ }
+
+let Defs = [USR_OVF], Itinerary = S_3op_tc_2_SLOT23, isCodeGenOnly = 0 in
+def S2_vcrotate : T_S3op_shiftVect < "vcrotate", 0b11, 0b00>;
+
//===----------------------------------------------------------------------===//
// Template class for 'insert bitfield' instructions
//===----------------------------------------------------------------------===//
projects / oota-llvm.git / blobdiff