//
//===----------------------------------------------------------------------===//
+//===----------------------------------------------------------------------===//
+// ARM Instruction Predicate Definitions.
+//
+def HasFPARMv8 : Predicate<"Subtarget->hasFPARMv8()">,
+ AssemblerPredicate<"FeatureFPARMv8", "fp-armv8">;
+def HasNEON : Predicate<"Subtarget->hasNEON()">,
+ AssemblerPredicate<"FeatureNEON", "neon">;
+def HasCrypto : Predicate<"Subtarget->hasCrypto()">,
+ AssemblerPredicate<"FeatureCrypto","crypto">;
+
+// Use fused MAC if more precision in FP computation is allowed.
+def UseFusedMAC : Predicate<"(TM.Options.AllowFPOpFusion =="
+ " FPOpFusion::Fast)">;
include "AArch64InstrFormats.td"
//===----------------------------------------------------------------------===//
// made for a variable/address at ISelLowering.
// + The output of ISelLowering should be selectable (hence the Wrapper,
// rather than a bare target opcode)
-def SDTAArch64Wrapper : SDTypeProfile<1, 3, [SDTCisSameAs<0, 1>,
- SDTCisSameAs<1, 2>,
- SDTCisVT<3, i32>,
- SDTCisPtrTy<0>]>;
+def SDTAArch64WrapperLarge : SDTypeProfile<1, 4, [SDTCisSameAs<0, 1>,
+ SDTCisSameAs<0, 2>,
+ SDTCisSameAs<0, 3>,
+ SDTCisSameAs<0, 4>,
+ SDTCisPtrTy<0>]>;
+
+def A64WrapperLarge :SDNode<"AArch64ISD::WrapperLarge", SDTAArch64WrapperLarge>;
+
+def SDTAArch64WrapperSmall : SDTypeProfile<1, 3, [SDTCisSameAs<0, 1>,
+ SDTCisSameAs<1, 2>,
+ SDTCisVT<3, i32>,
+ SDTCisPtrTy<0>]>;
-def A64WrapperSmall : SDNode<"AArch64ISD::WrapperSmall", SDTAArch64Wrapper>;
+def A64WrapperSmall :SDNode<"AArch64ISD::WrapperSmall", SDTAArch64WrapperSmall>;
def SDTAArch64GOTLoad : SDTypeProfile<1, 1, [SDTCisPtrTy<0>, SDTCisPtrTy<1>]>;
def A64Ubfx : SDNode<"AArch64ISD::UBFX", SDTA64BFX>;
+class BinOpFrag<dag res> : PatFrag<(ops node:$LHS, node:$RHS), res>;
+
//===----------------------------------------------------------------------===//
// Call sequence pseudo-instructions
//===----------------------------------------------------------------------===//
let DecoderMethod = "DecodeFPZeroOperand";
}
+def fpz64movi : Operand<i64>,
+ ComplexPattern<f64, 1, "SelectFPZeroOperand", [fpimm]> {
+ let ParserMatchClass = fpzero_asmoperand;
+ let PrintMethod = "printFPZeroOperand";
+ let DecoderMethod = "DecodeFPZeroOperand";
+}
+
multiclass A64I_fpcmpSignal<bits<2> type, bit imm, dag ins, dag pattern> {
def _quiet : A64I_fpcmp<0b0, 0b0, type, 0b00, {0b0, imm, 0b0, 0b0, 0b0},
(outs), ins, "fcmp\t$Rn, $Rm", [pattern],
def FNMADDdddd : A64I_fpdp3Impl<"fnmadd", FPR64, f64, 0b01, 0b1, 0b0, fnmadd>;
def FNMSUBdddd : A64I_fpdp3Impl<"fnmsub", FPR64, f64, 0b01, 0b1, 0b1, fnmsub>;
+// Extra patterns for when we're allowed to optimise separate multiplication and
+// addition.
+let Predicates = [HasFPARMv8, UseFusedMAC] in {
+def : Pat<(f32 (fadd FPR32:$Ra, (f32 (fmul FPR32:$Rn, FPR32:$Rm)))),
+ (FMADDssss FPR32:$Rn, FPR32:$Rm, FPR32:$Ra)>;
+def : Pat<(f32 (fsub FPR32:$Ra, (f32 (fmul FPR32:$Rn, FPR32:$Rm)))),
+ (FMSUBssss FPR32:$Rn, FPR32:$Rm, FPR32:$Ra)>;
+def : Pat<(f32 (fsub (f32 (fmul FPR32:$Rn, FPR32:$Rm)), FPR32:$Ra)),
+ (FNMADDssss FPR32:$Rn, FPR32:$Rm, FPR32:$Ra)>;
+def : Pat<(f32 (fsub (f32 (fneg FPR32:$Ra)), (f32 (fmul FPR32:$Rn, FPR32:$Rm)))),
+ (FNMSUBssss FPR32:$Rn, FPR32:$Rm, FPR32:$Ra)>;
+
+def : Pat<(f64 (fadd FPR64:$Ra, (f64 (fmul FPR64:$Rn, FPR64:$Rm)))),
+ (FMADDdddd FPR64:$Rn, FPR64:$Rm, FPR64:$Ra)>;
+def : Pat<(f64 (fsub FPR64:$Ra, (f64 (fmul FPR64:$Rn, FPR64:$Rm)))),
+ (FMSUBdddd FPR64:$Rn, FPR64:$Rm, FPR64:$Ra)>;
+def : Pat<(f64 (fsub (f64 (fmul FPR64:$Rn, FPR64:$Rm)), FPR64:$Ra)),
+ (FNMADDdddd FPR64:$Rn, FPR64:$Rm, FPR64:$Ra)>;
+def : Pat<(f64 (fsub (f64 (fneg FPR64:$Ra)), (f64 (fmul FPR64:$Rn, FPR64:$Rm)))),
+ (FNMSUBdddd FPR64:$Rn, FPR64:$Rm, FPR64:$Ra)>;
+}
+
+
//===----------------------------------------------------------------------===//
// Floating-point <-> fixed-point conversion instructions
//===----------------------------------------------------------------------===//
defm FCVTZ : A64I_fptointRM<0b11, 0b0, "fcvtz">;
defm FCVTA : A64I_fptointRM<0b00, 0b1, "fcvta">;
+let Predicates = [HasFPARMv8] in {
def : Pat<(i32 (fp_to_sint f32:$Rn)), (FCVTZSws $Rn)>;
def : Pat<(i64 (fp_to_sint f32:$Rn)), (FCVTZSxs $Rn)>;
def : Pat<(i32 (fp_to_uint f32:$Rn)), (FCVTZUws $Rn)>;
def : Pat<(i64 (fp_to_sint f64:$Rn)), (FCVTZSxd $Rn)>;
def : Pat<(i32 (fp_to_uint f64:$Rn)), (FCVTZUwd $Rn)>;
def : Pat<(i64 (fp_to_uint f64:$Rn)), (FCVTZUxd $Rn)>;
+}
multiclass A64I_inttofp<bit o0, string asmop> {
def CVTFsw : A64I_fpintI<0b0, 0b00, 0b00, {0, 1, o0}, FPR32, GPR32, asmop>;
defm S : A64I_inttofp<0b0, "scvtf">;
defm U : A64I_inttofp<0b1, "ucvtf">;
+let Predicates = [HasFPARMv8] in {
def : Pat<(f32 (sint_to_fp i32:$Rn)), (SCVTFsw $Rn)>;
def : Pat<(f32 (sint_to_fp i64:$Rn)), (SCVTFsx $Rn)>;
def : Pat<(f64 (sint_to_fp i32:$Rn)), (SCVTFdw $Rn)>;
def : Pat<(f32 (uint_to_fp i64:$Rn)), (UCVTFsx $Rn)>;
def : Pat<(f64 (uint_to_fp i32:$Rn)), (UCVTFdw $Rn)>;
def : Pat<(f64 (uint_to_fp i64:$Rn)), (UCVTFdx $Rn)>;
+}
def FMOVws : A64I_fpintI<0b0, 0b00, 0b00, 0b110, GPR32, FPR32, "fmov">;
def FMOVsw : A64I_fpintI<0b0, 0b00, 0b00, 0b111, FPR32, GPR32, "fmov">;
def FMOVxd : A64I_fpintI<0b1, 0b01, 0b00, 0b110, GPR64, FPR64, "fmov">;
def FMOVdx : A64I_fpintI<0b1, 0b01, 0b00, 0b111, FPR64, GPR64, "fmov">;
+let Predicates = [HasFPARMv8] in {
def : Pat<(i32 (bitconvert f32:$Rn)), (FMOVws $Rn)>;
def : Pat<(f32 (bitconvert i32:$Rn)), (FMOVsw $Rn)>;
def : Pat<(i64 (bitconvert f64:$Rn)), (FMOVxd $Rn)>;
def : Pat<(f64 (bitconvert i64:$Rn)), (FMOVdx $Rn)>;
+}
def lane1_asmoperand : AsmOperandClass {
let Name = "Lane1";
"fmov\t$Rd.d[$Lane], $Rn", [], NoItinerary>;
}
+let Predicates = [HasFPARMv8] in {
def : InstAlias<"fmov $Rd, $Rn.2d[$Lane]",
(FMOVxv GPR64:$Rd, VPR128:$Rn, lane1:$Lane), 0b0>;
def : InstAlias<"fmov $Rd.2d[$Lane], $Rn",
(FMOVvx VPR128:$Rd, GPR64:$Rn, lane1:$Lane), 0b0>;
+}
//===----------------------------------------------------------------------===//
// Floating-point immediate instructions
def LDRx_lit : A64I_LDRlitSimple<0b01, 0b0, GPR64>;
}
+let Predicates = [HasFPARMv8] in {
def LDRs_lit : A64I_LDRlitSimple<0b00, 0b1, FPR32>;
def LDRd_lit : A64I_LDRlitSimple<0b01, 0b1, FPR64>;
+}
let mayLoad = 1 in {
+ let Predicates = [HasFPARMv8] in {
def LDRq_lit : A64I_LDRlitSimple<0b10, 0b1, FPR128>;
+ }
def LDRSWx_lit : A64I_LDRlit<0b10, 0b0,
defm LS64
: A64I_LDRSTR_unsigned<"LS64", 0b11, 0b0, 0b0, "", GPR64, dword_addrparams>;
+let Predicates = [HasFPARMv8] in {
// STR/LDR to/from a B register
defm LSFP8
: A64I_LDRSTR_unsigned<"LSFP8", 0b00, 0b1, 0b0, "", FPR8, byte_addrparams>;
defm LSFP128
: A64I_LDRSTR_unsigned<"LSFP128", 0b00, 0b1, 0b1, "", FPR128,
qword_addrparams>;
+}
//===------------------------------
// 2.3 Signed loads
defm LSPair32 : A64I_LSPsimple<0b00, 0b0, GPR32, word_simm7, "LSPair32">;
defm LSPair64 : A64I_LSPsimple<0b10, 0b0, GPR64, dword_simm7, "LSPair64">;
+
+let Predicates = [HasFPARMv8] in {
defm LSFPPair32 : A64I_LSPsimple<0b00, 0b1, FPR32, word_simm7, "LSFPPair32">;
defm LSFPPair64 : A64I_LSPsimple<0b01, 0b1, FPR64, dword_simm7, "LSFPPair64">;
defm LSFPPair128 : A64I_LSPsimple<0b10, 0b1, FPR128, qword_simm7,
"LSFPPair128">;
+}
def LDPSWx : A64I_LSPoffset<0b01, 0b0, 0b1,
let DiagnosticType = "MOVWUImm16";
}
- def _imm : Operand<i32> {
+ def _imm : Operand<i64> {
let ParserMatchClass = !cast<AsmOperandClass>(prefix # "_imm_asmoperand");
let PrintMethod = "printMoveWideImmOperand";
let EncoderMethod = "getMoveWideImmOpValue";
# "A64Imms::" # immpredicate # ">";
}
- def _movimm : Operand<i32> {
+ def _movimm : Operand<i64> {
let ParserMatchClass = !cast<AsmOperandClass>(prefix # "_asmoperand");
let MIOperandInfo = (ops uimm16:$UImm16, imm:$Shift);
def : movalias<MOVNwii, GPR32, movn32_movimm>;
def : movalias<MOVNxii, GPR64, movn64_movimm>;
+def movw_addressref_g0 : ComplexPattern<i64, 2, "SelectMOVWAddressRef<0>">;
+def movw_addressref_g1 : ComplexPattern<i64, 2, "SelectMOVWAddressRef<1>">;
+def movw_addressref_g2 : ComplexPattern<i64, 2, "SelectMOVWAddressRef<2>">;
+def movw_addressref_g3 : ComplexPattern<i64, 2, "SelectMOVWAddressRef<3>">;
+
+def : Pat<(A64WrapperLarge movw_addressref_g3:$G3, movw_addressref_g2:$G2,
+ movw_addressref_g1:$G1, movw_addressref_g0:$G0),
+ (MOVKxii (MOVKxii (MOVKxii (MOVZxii movw_addressref_g3:$G3),
+ movw_addressref_g2:$G2),
+ movw_addressref_g1:$G1),
+ movw_addressref_g0:$G0)>;
+
//===----------------------------------------------------------------------===//
// PC-relative addressing instructions
//===----------------------------------------------------------------------===//
// GOT access patterns
//===----------------------------------------------------------------------===//
-// FIXME: Wibble
-
class GOTLoadSmall<SDNode addrfrag>
: Pat<(A64GOTLoad (A64WrapperSmall addrfrag:$Hi, addrfrag:$Lo12, 8)),
(LS64_LDR (ADRPxi addrfrag:$Hi), addrfrag:$Lo12)>;
defm : regoff_pats<"Xm", (add i64:$Rn, (shl i64:$Rm, SHIFT)),
(i64 i64:$Rn), (i64 i64:$Rm), (i64 3)>;
+
+//===----------------------------------------------------------------------===//
+// Advanced SIMD (NEON) Support
+//
+
+include "AArch64InstrNEON.td"
projects / oota-llvm.git / blobdiff