1 //===----- AArch64InstrInfo.td - AArch64 Instruction Info ----*- tablegen -*-=//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file describes the AArch64 scalar instructions in TableGen format.
12 //===----------------------------------------------------------------------===//
14 //===----------------------------------------------------------------------===//
15 // ARM Instruction Predicate Definitions.
17 def HasNEON : Predicate<"Subtarget->hasNEON()">,
18 AssemblerPredicate<"FeatureNEON", "neon">;
19 def HasCrypto : Predicate<"Subtarget->hasCrypto()">,
20 AssemblerPredicate<"FeatureCrypto","crypto">;
22 // Use fused MAC if more precision in FP computation is allowed.
23 def UseFusedMAC : Predicate<"(TM.Options.AllowFPOpFusion =="
24 " FPOpFusion::Fast)">;
25 include "AArch64InstrFormats.td"
27 //===----------------------------------------------------------------------===//
28 // Target-specific ISD nodes and profiles
29 //===----------------------------------------------------------------------===//
31 def SDT_A64ret : SDTypeProfile<0, 0, []>;
32 def A64ret : SDNode<"AArch64ISD::Ret", SDT_A64ret, [SDNPHasChain,
36 // (ins NZCV, Condition, Dest)
37 def SDT_A64br_cc : SDTypeProfile<0, 3, [SDTCisVT<0, i32>]>;
38 def A64br_cc : SDNode<"AArch64ISD::BR_CC", SDT_A64br_cc, [SDNPHasChain]>;
40 // (outs Result), (ins NZCV, IfTrue, IfFalse, Condition)
41 def SDT_A64select_cc : SDTypeProfile<1, 4, [SDTCisVT<1, i32>,
44 def A64select_cc : SDNode<"AArch64ISD::SELECT_CC", SDT_A64select_cc>;
46 // (outs NZCV), (ins LHS, RHS, Condition)
47 def SDT_A64setcc : SDTypeProfile<1, 3, [SDTCisVT<0, i32>,
49 def A64setcc : SDNode<"AArch64ISD::SETCC", SDT_A64setcc>;
52 // (outs GPR64), (ins)
53 def A64threadpointer : SDNode<"AArch64ISD::THREAD_POINTER", SDTPtrLeaf>;
55 // A64 compares don't care about the cond really (they set all flags) so a
56 // simple binary operator is useful.
57 def A64cmp : PatFrag<(ops node:$lhs, node:$rhs),
58 (A64setcc node:$lhs, node:$rhs, cond)>;
61 // When matching a notional (CMP op1, (sub 0, op2)), we'd like to use a CMN
62 // instruction on the grounds that "op1 - (-op2) == op1 + op2". However, the C
63 // and V flags can be set differently by this operation. It comes down to
64 // whether "SInt(~op2)+1 == SInt(~op2+1)" (and the same for UInt). If they are
65 // then everything is fine. If not then the optimization is wrong. Thus general
66 // comparisons are only valid if op2 != 0.
68 // So, finally, the only LLVM-native comparisons that don't mention C and V are
69 // SETEQ and SETNE. They're the only ones we can safely use CMN for in the
70 // absence of information about op2.
71 def equality_cond : PatLeaf<(cond), [{
72 return N->get() == ISD::SETEQ || N->get() == ISD::SETNE;
75 def A64cmn : PatFrag<(ops node:$lhs, node:$rhs),
76 (A64setcc node:$lhs, (sub 0, node:$rhs), equality_cond)>;
78 // There are two layers of indirection here, driven by the following
80 // + TableGen does not know CodeModel or Reloc so that decision should be
81 // made for a variable/address at ISelLowering.
82 // + The output of ISelLowering should be selectable (hence the Wrapper,
83 // rather than a bare target opcode)
84 def SDTAArch64WrapperLarge : SDTypeProfile<1, 4, [SDTCisSameAs<0, 1>,
90 def A64WrapperLarge :SDNode<"AArch64ISD::WrapperLarge", SDTAArch64WrapperLarge>;
92 def SDTAArch64WrapperSmall : SDTypeProfile<1, 3, [SDTCisSameAs<0, 1>,
97 def A64WrapperSmall :SDNode<"AArch64ISD::WrapperSmall", SDTAArch64WrapperSmall>;
100 def SDTAArch64GOTLoad : SDTypeProfile<1, 1, [SDTCisPtrTy<0>, SDTCisPtrTy<1>]>;
101 def A64GOTLoad : SDNode<"AArch64ISD::GOTLoad", SDTAArch64GOTLoad,
105 // (A64BFI LHS, RHS, LSB, Width)
106 def SDTA64BFI : SDTypeProfile<1, 4, [SDTCisSameAs<0, 1>,
111 def A64Bfi : SDNode<"AArch64ISD::BFI", SDTA64BFI>;
113 // (A64EXTR HiReg, LoReg, LSB)
114 def SDTA64EXTR : SDTypeProfile<1, 3, [SDTCisSameAs<0, 1>, SDTCisSameAs<1, 2>,
116 def A64Extr : SDNode<"AArch64ISD::EXTR", SDTA64EXTR>;
118 // (A64[SU]BFX Field, ImmR, ImmS).
120 // Note that ImmR and ImmS are already encoded for the actual instructions. The
121 // more natural LSB and Width mix together to form ImmR and ImmS, something
122 // which TableGen can't handle.
123 def SDTA64BFX : SDTypeProfile<1, 3, [SDTCisVT<2, i64>, SDTCisVT<3, i64>]>;
124 def A64Sbfx : SDNode<"AArch64ISD::SBFX", SDTA64BFX>;
126 def A64Ubfx : SDNode<"AArch64ISD::UBFX", SDTA64BFX>;
128 class BinOpFrag<dag res> : PatFrag<(ops node:$LHS, node:$RHS), res>;
130 //===----------------------------------------------------------------------===//
131 // Call sequence pseudo-instructions
132 //===----------------------------------------------------------------------===//
135 def SDT_AArch64Call : SDTypeProfile<0, -1, [SDTCisPtrTy<0>]>;
136 def AArch64Call : SDNode<"AArch64ISD::Call", SDT_AArch64Call,
137 [SDNPHasChain, SDNPOptInGlue, SDNPOutGlue, SDNPVariadic]>;
139 def AArch64tcret : SDNode<"AArch64ISD::TC_RETURN", SDT_AArch64Call,
140 [SDNPHasChain, SDNPOptInGlue, SDNPVariadic]>;
142 // The TLSDESCCALL node is a variant call which goes to an indirectly calculated
143 // destination but needs a relocation against a fixed symbol. As such it has two
144 // certain operands: the callee and the relocated variable.
146 // The TLS ABI only allows it to be selected to a BLR instructin (with
147 // appropriate relocation).
148 def SDTTLSDescCall : SDTypeProfile<0, -2, [SDTCisPtrTy<0>, SDTCisPtrTy<1>]>;
150 def A64tlsdesc_blr : SDNode<"AArch64ISD::TLSDESCCALL", SDTTLSDescCall,
151 [SDNPInGlue, SDNPOutGlue, SDNPHasChain,
155 def SDT_AArch64CallSeqStart : SDCallSeqStart<[ SDTCisPtrTy<0> ]>;
156 def AArch64callseq_start : SDNode<"ISD::CALLSEQ_START", SDT_AArch64CallSeqStart,
157 [SDNPHasChain, SDNPOutGlue]>;
159 def SDT_AArch64CallSeqEnd : SDCallSeqEnd<[ SDTCisPtrTy<0>, SDTCisPtrTy<1> ]>;
160 def AArch64callseq_end : SDNode<"ISD::CALLSEQ_END", SDT_AArch64CallSeqEnd,
161 [SDNPHasChain, SDNPOptInGlue, SDNPOutGlue]>;
165 // These pseudo-instructions have special semantics by virtue of being passed to
166 // the InstrInfo constructor. CALLSEQ_START/CALLSEQ_END are produced by
167 // LowerCall to (in our case) tell the back-end about stack adjustments for
168 // arguments passed on the stack. Here we select those markers to
169 // pseudo-instructions which explicitly set the stack, and finally in the
170 // RegisterInfo we convert them to a true stack adjustment.
171 let Defs = [XSP], Uses = [XSP] in {
172 def ADJCALLSTACKDOWN : PseudoInst<(outs), (ins i64imm:$amt),
173 [(AArch64callseq_start timm:$amt)]>;
175 def ADJCALLSTACKUP : PseudoInst<(outs), (ins i64imm:$amt1, i64imm:$amt2),
176 [(AArch64callseq_end timm:$amt1, timm:$amt2)]>;
179 //===----------------------------------------------------------------------===//
180 // Atomic operation pseudo-instructions
181 //===----------------------------------------------------------------------===//
183 // These get selected from C++ code as a pretty much direct translation from the
184 // generic DAG nodes. The one exception is the AtomicOrdering is added as an
185 // operand so that the eventual lowering can make use of it and choose
186 // acquire/release operations when required.
188 let usesCustomInserter = 1, hasCtrlDep = 1, mayLoad = 1, mayStore = 1 in {
189 multiclass AtomicSizes {
190 def _I8 : PseudoInst<(outs GPR32:$dst),
191 (ins GPR64xsp:$ptr, GPR32:$incr, i32imm:$ordering), []>;
192 def _I16 : PseudoInst<(outs GPR32:$dst),
193 (ins GPR64xsp:$ptr, GPR32:$incr, i32imm:$ordering), []>;
194 def _I32 : PseudoInst<(outs GPR32:$dst),
195 (ins GPR64xsp:$ptr, GPR32:$incr, i32imm:$ordering), []>;
196 def _I64 : PseudoInst<(outs GPR64:$dst),
197 (ins GPR64xsp:$ptr, GPR64:$incr, i32imm:$ordering), []>;
201 defm ATOMIC_LOAD_ADD : AtomicSizes;
202 defm ATOMIC_LOAD_SUB : AtomicSizes;
203 defm ATOMIC_LOAD_AND : AtomicSizes;
204 defm ATOMIC_LOAD_OR : AtomicSizes;
205 defm ATOMIC_LOAD_XOR : AtomicSizes;
206 defm ATOMIC_LOAD_NAND : AtomicSizes;
207 defm ATOMIC_SWAP : AtomicSizes;
208 let Defs = [NZCV] in {
209 // These operations need a CMP to calculate the correct value
210 defm ATOMIC_LOAD_MIN : AtomicSizes;
211 defm ATOMIC_LOAD_MAX : AtomicSizes;
212 defm ATOMIC_LOAD_UMIN : AtomicSizes;
213 defm ATOMIC_LOAD_UMAX : AtomicSizes;
216 class AtomicCmpSwap<RegisterClass GPRData>
217 : PseudoInst<(outs GPRData:$dst),
218 (ins GPR64xsp:$ptr, GPRData:$old, GPRData:$new,
219 i32imm:$ordering), []> {
220 let usesCustomInserter = 1;
227 def ATOMIC_CMP_SWAP_I8 : AtomicCmpSwap<GPR32>;
228 def ATOMIC_CMP_SWAP_I16 : AtomicCmpSwap<GPR32>;
229 def ATOMIC_CMP_SWAP_I32 : AtomicCmpSwap<GPR32>;
230 def ATOMIC_CMP_SWAP_I64 : AtomicCmpSwap<GPR64>;
232 //===----------------------------------------------------------------------===//
233 // Add-subtract (extended register) instructions
234 //===----------------------------------------------------------------------===//
235 // Contains: ADD, ADDS, SUB, SUBS + aliases CMN, CMP
237 // The RHS of these operations is conceptually a sign/zero-extended
238 // register, optionally shifted left by 1-4. The extension can be a
239 // NOP (e.g. "sxtx" sign-extending a 64-bit register to 64-bits) but
240 // must be specified with one exception:
242 // If one of the registers is sp/wsp then LSL is an alias for UXTW in
243 // 32-bit instructions and UXTX in 64-bit versions, the shift amount
244 // is not optional in that case (but can explicitly be 0), and the
245 // entire suffix can be skipped (e.g. "add sp, x3, x2").
247 multiclass extend_operands<string PREFIX, string Diag> {
248 def _asmoperand : AsmOperandClass {
250 let RenderMethod = "addRegExtendOperands";
251 let PredicateMethod = "isRegExtend<A64SE::" # PREFIX # ">";
252 let DiagnosticType = "AddSubRegExtend" # Diag;
255 def _operand : Operand<i64>,
256 ImmLeaf<i64, [{ return Imm >= 0 && Imm <= 4; }]> {
257 let PrintMethod = "printRegExtendOperand<A64SE::" # PREFIX # ">";
258 let DecoderMethod = "DecodeRegExtendOperand";
259 let ParserMatchClass = !cast<AsmOperandClass>(PREFIX # "_asmoperand");
263 defm UXTB : extend_operands<"UXTB", "Small">;
264 defm UXTH : extend_operands<"UXTH", "Small">;
265 defm UXTW : extend_operands<"UXTW", "Small">;
266 defm UXTX : extend_operands<"UXTX", "Large">;
267 defm SXTB : extend_operands<"SXTB", "Small">;
268 defm SXTH : extend_operands<"SXTH", "Small">;
269 defm SXTW : extend_operands<"SXTW", "Small">;
270 defm SXTX : extend_operands<"SXTX", "Large">;
272 def LSL_extasmoperand : AsmOperandClass {
273 let Name = "RegExtendLSL";
274 let RenderMethod = "addRegExtendOperands";
275 let DiagnosticType = "AddSubRegExtendLarge";
278 def LSL_extoperand : Operand<i64> {
279 let ParserMatchClass = LSL_extasmoperand;
283 // The patterns for various sign-extensions are a little ugly and
284 // non-uniform because everything has already been promoted to the
285 // legal i64 and i32 types. We'll wrap the various variants up in a
286 // class for use later.
288 dag uxtb; dag uxth; dag uxtw; dag uxtx;
289 dag sxtb; dag sxth; dag sxtw; dag sxtx;
294 def extends_to_i64 : extend_types {
295 let uxtb = (and (anyext i32:$Rm), 255);
296 let uxth = (and (anyext i32:$Rm), 65535);
297 let uxtw = (zext i32:$Rm);
298 let uxtx = (i64 $Rm);
300 let sxtb = (sext_inreg (anyext i32:$Rm), i8);
301 let sxth = (sext_inreg (anyext i32:$Rm), i16);
302 let sxtw = (sext i32:$Rm);
303 let sxtx = (i64 $Rm);
310 def extends_to_i32 : extend_types {
311 let uxtb = (and i32:$Rm, 255);
312 let uxth = (and i32:$Rm, 65535);
313 let uxtw = (i32 i32:$Rm);
314 let uxtx = (i32 i32:$Rm);
316 let sxtb = (sext_inreg i32:$Rm, i8);
317 let sxth = (sext_inreg i32:$Rm, i16);
318 let sxtw = (i32 i32:$Rm);
319 let sxtx = (i32 i32:$Rm);
325 // Now, six of the extensions supported are easy and uniform: if the source size
326 // is 32-bits or less, then Rm is always a 32-bit register. We'll instantiate
327 // those instructions in one block.
329 // The uxtx/sxtx could potentially be merged in, but three facts dissuaded me:
330 // + It would break the naming scheme: either ADDxx_uxtx or ADDww_uxtx would
332 // + Patterns are very different as well.
333 // + Passing different registers would be ugly (more fields in extend_types
334 // would probably be the best option).
335 multiclass addsub_exts<bit sf, bit op, bit S, string asmop,
336 SDPatternOperator opfrag,
337 dag outs, extend_types exts> {
338 def w_uxtb : A64I_addsubext<sf, op, S, 0b00, 0b000,
339 outs, (ins exts.GPR:$Rn, GPR32:$Rm, UXTB_operand:$Imm3),
340 !strconcat(asmop, "$Rn, $Rm, $Imm3"),
341 [(opfrag exts.ty:$Rn, (shl exts.uxtb, UXTB_operand:$Imm3))],
343 def w_uxth : A64I_addsubext<sf, op, S, 0b00, 0b001,
344 outs, (ins exts.GPR:$Rn, GPR32:$Rm, UXTH_operand:$Imm3),
345 !strconcat(asmop, "$Rn, $Rm, $Imm3"),
346 [(opfrag exts.ty:$Rn, (shl exts.uxth, UXTH_operand:$Imm3))],
348 def w_uxtw : A64I_addsubext<sf, op, S, 0b00, 0b010,
349 outs, (ins exts.GPR:$Rn, GPR32:$Rm, UXTW_operand:$Imm3),
350 !strconcat(asmop, "$Rn, $Rm, $Imm3"),
351 [(opfrag exts.ty:$Rn, (shl exts.uxtw, UXTW_operand:$Imm3))],
354 def w_sxtb : A64I_addsubext<sf, op, S, 0b00, 0b100,
355 outs, (ins exts.GPR:$Rn, GPR32:$Rm, SXTB_operand:$Imm3),
356 !strconcat(asmop, "$Rn, $Rm, $Imm3"),
357 [(opfrag exts.ty:$Rn, (shl exts.sxtb, SXTB_operand:$Imm3))],
359 def w_sxth : A64I_addsubext<sf, op, S, 0b00, 0b101,
360 outs, (ins exts.GPR:$Rn, GPR32:$Rm, SXTH_operand:$Imm3),
361 !strconcat(asmop, "$Rn, $Rm, $Imm3"),
362 [(opfrag exts.ty:$Rn, (shl exts.sxth, SXTH_operand:$Imm3))],
364 def w_sxtw : A64I_addsubext<sf, op, S, 0b00, 0b110,
365 outs, (ins exts.GPR:$Rn, GPR32:$Rm, SXTW_operand:$Imm3),
366 !strconcat(asmop, "$Rn, $Rm, $Imm3"),
367 [(opfrag exts.ty:$Rn, (shl exts.sxtw, SXTW_operand:$Imm3))],
371 // These two could be merge in with the above, but their patterns aren't really
372 // necessary and the naming-scheme would necessarily break:
373 multiclass addsub_xxtx<bit op, bit S, string asmop, SDPatternOperator opfrag,
375 def x_uxtx : A64I_addsubext<0b1, op, S, 0b00, 0b011,
377 (ins GPR64xsp:$Rn, GPR64:$Rm, UXTX_operand:$Imm3),
378 !strconcat(asmop, "$Rn, $Rm, $Imm3"),
379 [(opfrag i64:$Rn, (shl i64:$Rm, UXTX_operand:$Imm3))],
382 def x_sxtx : A64I_addsubext<0b1, op, S, 0b00, 0b111,
384 (ins GPR64xsp:$Rn, GPR64:$Rm, SXTX_operand:$Imm3),
385 !strconcat(asmop, "$Rn, $Rm, $Imm3"),
386 [/* No Pattern: same as uxtx */],
390 multiclass addsub_wxtx<bit op, bit S, string asmop, dag outs> {
391 def w_uxtx : A64I_addsubext<0b0, op, S, 0b00, 0b011,
393 (ins GPR32wsp:$Rn, GPR32:$Rm, UXTX_operand:$Imm3),
394 !strconcat(asmop, "$Rn, $Rm, $Imm3"),
395 [/* No pattern: probably same as uxtw */],
398 def w_sxtx : A64I_addsubext<0b0, op, S, 0b00, 0b111,
400 (ins GPR32wsp:$Rn, GPR32:$Rm, SXTX_operand:$Imm3),
401 !strconcat(asmop, "$Rn, $Rm, $Imm3"),
402 [/* No Pattern: probably same as uxtw */],
406 class SetRD<RegisterClass RC, SDPatternOperator op>
407 : PatFrag<(ops node:$lhs, node:$rhs), (set RC:$Rd, (op node:$lhs, node:$rhs))>;
408 class SetNZCV<SDPatternOperator op>
409 : PatFrag<(ops node:$lhs, node:$rhs), (set NZCV, (op node:$lhs, node:$rhs))>;
411 defm ADDxx :addsub_exts<0b1, 0b0, 0b0, "add\t$Rd, ", SetRD<GPR64xsp, add>,
412 (outs GPR64xsp:$Rd), extends_to_i64>,
413 addsub_xxtx< 0b0, 0b0, "add\t$Rd, ", SetRD<GPR64xsp, add>,
414 (outs GPR64xsp:$Rd)>;
415 defm ADDww :addsub_exts<0b0, 0b0, 0b0, "add\t$Rd, ", SetRD<GPR32wsp, add>,
416 (outs GPR32wsp:$Rd), extends_to_i32>,
417 addsub_wxtx< 0b0, 0b0, "add\t$Rd, ",
418 (outs GPR32wsp:$Rd)>;
419 defm SUBxx :addsub_exts<0b1, 0b1, 0b0, "sub\t$Rd, ", SetRD<GPR64xsp, sub>,
420 (outs GPR64xsp:$Rd), extends_to_i64>,
421 addsub_xxtx< 0b1, 0b0, "sub\t$Rd, ", SetRD<GPR64xsp, sub>,
422 (outs GPR64xsp:$Rd)>;
423 defm SUBww :addsub_exts<0b0, 0b1, 0b0, "sub\t$Rd, ", SetRD<GPR32wsp, sub>,
424 (outs GPR32wsp:$Rd), extends_to_i32>,
425 addsub_wxtx< 0b1, 0b0, "sub\t$Rd, ",
426 (outs GPR32wsp:$Rd)>;
428 let Defs = [NZCV] in {
429 defm ADDSxx :addsub_exts<0b1, 0b0, 0b1, "adds\t$Rd, ", SetRD<GPR64, addc>,
430 (outs GPR64:$Rd), extends_to_i64>,
431 addsub_xxtx< 0b0, 0b1, "adds\t$Rd, ", SetRD<GPR64, addc>,
433 defm ADDSww :addsub_exts<0b0, 0b0, 0b1, "adds\t$Rd, ", SetRD<GPR32, addc>,
434 (outs GPR32:$Rd), extends_to_i32>,
435 addsub_wxtx< 0b0, 0b1, "adds\t$Rd, ",
437 defm SUBSxx :addsub_exts<0b1, 0b1, 0b1, "subs\t$Rd, ", SetRD<GPR64, subc>,
438 (outs GPR64:$Rd), extends_to_i64>,
439 addsub_xxtx< 0b1, 0b1, "subs\t$Rd, ", SetRD<GPR64, subc>,
441 defm SUBSww :addsub_exts<0b0, 0b1, 0b1, "subs\t$Rd, ", SetRD<GPR32, subc>,
442 (outs GPR32:$Rd), extends_to_i32>,
443 addsub_wxtx< 0b1, 0b1, "subs\t$Rd, ",
447 let Rd = 0b11111, isCompare = 1 in {
448 defm CMNx : addsub_exts<0b1, 0b0, 0b1, "cmn\t", SetNZCV<A64cmn>,
449 (outs), extends_to_i64>,
450 addsub_xxtx< 0b0, 0b1, "cmn\t", SetNZCV<A64cmn>, (outs)>;
451 defm CMNw : addsub_exts<0b0, 0b0, 0b1, "cmn\t", SetNZCV<A64cmn>,
452 (outs), extends_to_i32>,
453 addsub_wxtx< 0b0, 0b1, "cmn\t", (outs)>;
454 defm CMPx : addsub_exts<0b1, 0b1, 0b1, "cmp\t", SetNZCV<A64cmp>,
455 (outs), extends_to_i64>,
456 addsub_xxtx< 0b1, 0b1, "cmp\t", SetNZCV<A64cmp>, (outs)>;
457 defm CMPw : addsub_exts<0b0, 0b1, 0b1, "cmp\t", SetNZCV<A64cmp>,
458 (outs), extends_to_i32>,
459 addsub_wxtx< 0b1, 0b1, "cmp\t", (outs)>;
463 // Now patterns for the operation without a shift being needed. No patterns are
464 // created for uxtx/sxtx since they're non-uniform and it's expected that
465 // add/sub (shifted register) will handle those cases anyway.
466 multiclass addsubext_noshift_patterns<string prefix, SDPatternOperator nodeop,
468 def : Pat<(nodeop exts.ty:$Rn, exts.uxtb),
469 (!cast<Instruction>(prefix # "w_uxtb") $Rn, $Rm, 0)>;
470 def : Pat<(nodeop exts.ty:$Rn, exts.uxth),
471 (!cast<Instruction>(prefix # "w_uxth") $Rn, $Rm, 0)>;
472 def : Pat<(nodeop exts.ty:$Rn, exts.uxtw),
473 (!cast<Instruction>(prefix # "w_uxtw") $Rn, $Rm, 0)>;
475 def : Pat<(nodeop exts.ty:$Rn, exts.sxtb),
476 (!cast<Instruction>(prefix # "w_sxtb") $Rn, $Rm, 0)>;
477 def : Pat<(nodeop exts.ty:$Rn, exts.sxth),
478 (!cast<Instruction>(prefix # "w_sxth") $Rn, $Rm, 0)>;
479 def : Pat<(nodeop exts.ty:$Rn, exts.sxtw),
480 (!cast<Instruction>(prefix # "w_sxtw") $Rn, $Rm, 0)>;
483 defm : addsubext_noshift_patterns<"ADDxx", add, extends_to_i64>;
484 defm : addsubext_noshift_patterns<"ADDww", add, extends_to_i32>;
485 defm : addsubext_noshift_patterns<"SUBxx", sub, extends_to_i64>;
486 defm : addsubext_noshift_patterns<"SUBww", sub, extends_to_i32>;
488 defm : addsubext_noshift_patterns<"CMNx", A64cmn, extends_to_i64>;
489 defm : addsubext_noshift_patterns<"CMNw", A64cmn, extends_to_i32>;
490 defm : addsubext_noshift_patterns<"CMPx", A64cmp, extends_to_i64>;
491 defm : addsubext_noshift_patterns<"CMPw", A64cmp, extends_to_i32>;
493 // An extend of "lsl #imm" is valid if and only if one of Rn and Rd is
494 // sp/wsp. It is synonymous with uxtx/uxtw depending on the size of the
495 // operation. Also permitted in this case is complete omission of the argument,
496 // which implies "lsl #0".
497 multiclass lsl_aliases<string asmop, Instruction inst, RegisterClass GPR_Rd,
498 RegisterClass GPR_Rn, RegisterClass GPR_Rm> {
499 def : InstAlias<!strconcat(asmop, " $Rd, $Rn, $Rm"),
500 (inst GPR_Rd:$Rd, GPR_Rn:$Rn, GPR_Rm:$Rm, 0)>;
502 def : InstAlias<!strconcat(asmop, " $Rd, $Rn, $Rm, $LSL"),
503 (inst GPR_Rd:$Rd, GPR_Rn:$Rn, GPR_Rm:$Rm, LSL_extoperand:$LSL)>;
507 defm : lsl_aliases<"add", ADDxxx_uxtx, Rxsp, GPR64xsp, GPR64>;
508 defm : lsl_aliases<"add", ADDxxx_uxtx, GPR64xsp, Rxsp, GPR64>;
509 defm : lsl_aliases<"add", ADDwww_uxtw, Rwsp, GPR32wsp, GPR32>;
510 defm : lsl_aliases<"add", ADDwww_uxtw, GPR32wsp, Rwsp, GPR32>;
511 defm : lsl_aliases<"sub", SUBxxx_uxtx, Rxsp, GPR64xsp, GPR64>;
512 defm : lsl_aliases<"sub", SUBxxx_uxtx, GPR64xsp, Rxsp, GPR64>;
513 defm : lsl_aliases<"sub", SUBwww_uxtw, Rwsp, GPR32wsp, GPR32>;
514 defm : lsl_aliases<"sub", SUBwww_uxtw, GPR32wsp, Rwsp, GPR32>;
516 // Rd cannot be sp for flag-setting variants so only half of the aliases are
518 defm : lsl_aliases<"adds", ADDSxxx_uxtx, GPR64, Rxsp, GPR64>;
519 defm : lsl_aliases<"adds", ADDSwww_uxtw, GPR32, Rwsp, GPR32>;
520 defm : lsl_aliases<"subs", SUBSxxx_uxtx, GPR64, Rxsp, GPR64>;
521 defm : lsl_aliases<"subs", SUBSwww_uxtw, GPR32, Rwsp, GPR32>;
523 // CMP unfortunately has to be different because the instruction doesn't have a
525 multiclass cmp_lsl_aliases<string asmop, Instruction inst,
526 RegisterClass GPR_Rn, RegisterClass GPR_Rm> {
527 def : InstAlias<!strconcat(asmop, " $Rn, $Rm"),
528 (inst GPR_Rn:$Rn, GPR_Rm:$Rm, 0)>;
530 def : InstAlias<!strconcat(asmop, " $Rn, $Rm, $LSL"),
531 (inst GPR_Rn:$Rn, GPR_Rm:$Rm, LSL_extoperand:$LSL)>;
534 defm : cmp_lsl_aliases<"cmp", CMPxx_uxtx, Rxsp, GPR64>;
535 defm : cmp_lsl_aliases<"cmp", CMPww_uxtw, Rwsp, GPR32>;
536 defm : cmp_lsl_aliases<"cmn", CMNxx_uxtx, Rxsp, GPR64>;
537 defm : cmp_lsl_aliases<"cmn", CMNww_uxtw, Rwsp, GPR32>;
539 //===----------------------------------------------------------------------===//
540 // Add-subtract (immediate) instructions
541 //===----------------------------------------------------------------------===//
542 // Contains: ADD, ADDS, SUB, SUBS + aliases CMN, CMP, MOV
544 // These instructions accept a 12-bit unsigned immediate, optionally shifted
545 // left by 12 bits. Official assembly format specifies a 12 bit immediate with
546 // one of "", "LSL #0", "LSL #12" supplementary operands.
548 // There are surprisingly few ways to make this work with TableGen, so this
549 // implementation has separate instructions for the "LSL #0" and "LSL #12"
552 // If the MCInst retained a single combined immediate (which could be 0x123000,
553 // for example) then both components (imm & shift) would have to be delegated to
554 // a single assembly operand. This would entail a separate operand parser
555 // (because the LSL would have to live in the same AArch64Operand as the
556 // immediate to be accessible); assembly parsing is rather complex and
557 // error-prone C++ code.
559 // By splitting the immediate, we can delegate handling this optional operand to
560 // an InstAlias. Supporting functions to generate the correct MCInst are still
561 // required, but these are essentially trivial and parsing can remain generic.
563 // Rejected plans with rationale:
564 // ------------------------------
566 // In an ideal world you'de have two first class immediate operands (in
567 // InOperandList, specifying imm12 and shift). Unfortunately this is not
568 // selectable by any means I could discover.
570 // An Instruction with two MCOperands hidden behind a single entry in
571 // InOperandList (expanded by ComplexPatterns and MIOperandInfo) was functional,
572 // but required more C++ code to handle encoding/decoding. Parsing (the intended
573 // main beneficiary) ended up equally complex because of the optional nature of
576 // Attempting to circumvent the need for a custom OperandParser above by giving
577 // InstAliases without the "lsl #0" failed. add/sub could be accommodated but
578 // the cmp/cmn aliases didn't use the MIOperandInfo to determine how operands
579 // should be parsed: there was no way to accommodate an "lsl #12".
581 let ParserMethod = "ParseImmWithLSLOperand",
582 RenderMethod = "addImmWithLSLOperands" in {
583 // Derived PredicateMethod fields are different for each
584 def addsubimm_lsl0_asmoperand : AsmOperandClass {
585 let Name = "AddSubImmLSL0";
586 // If an error is reported against this operand, instruction could also be a
588 let DiagnosticType = "AddSubSecondSource";
591 def addsubimm_lsl12_asmoperand : AsmOperandClass {
592 let Name = "AddSubImmLSL12";
593 let DiagnosticType = "AddSubSecondSource";
597 def shr_12_XFORM : SDNodeXForm<imm, [{
598 return CurDAG->getTargetConstant(N->getSExtValue() >> 12, MVT::i32);
601 def shr_12_neg_XFORM : SDNodeXForm<imm, [{
602 return CurDAG->getTargetConstant((-N->getSExtValue()) >> 12, MVT::i32);
605 def neg_XFORM : SDNodeXForm<imm, [{
606 return CurDAG->getTargetConstant(-N->getSExtValue(), MVT::i32);
610 multiclass addsub_imm_operands<ValueType ty> {
611 let PrintMethod = "printAddSubImmLSL0Operand",
612 EncoderMethod = "getAddSubImmOpValue",
613 ParserMatchClass = addsubimm_lsl0_asmoperand in {
614 def _posimm_lsl0 : Operand<ty>,
615 ImmLeaf<ty, [{ return Imm >= 0 && (Imm & ~0xfff) == 0; }]>;
616 def _negimm_lsl0 : Operand<ty>,
617 ImmLeaf<ty, [{ return Imm < 0 && (-Imm & ~0xfff) == 0; }],
621 let PrintMethod = "printAddSubImmLSL12Operand",
622 EncoderMethod = "getAddSubImmOpValue",
623 ParserMatchClass = addsubimm_lsl12_asmoperand in {
624 def _posimm_lsl12 : Operand<ty>,
625 ImmLeaf<ty, [{ return Imm >= 0 && (Imm & ~0xfff000) == 0; }],
628 def _negimm_lsl12 : Operand<ty>,
629 ImmLeaf<ty, [{ return Imm < 0 && (-Imm & ~0xfff000) == 0; }],
634 // The add operands don't need any transformation
635 defm addsubimm_operand_i32 : addsub_imm_operands<i32>;
636 defm addsubimm_operand_i64 : addsub_imm_operands<i64>;
638 multiclass addsubimm_varieties<string prefix, bit sf, bit op, bits<2> shift,
639 string asmop, string cmpasmop,
640 Operand imm_operand, Operand cmp_imm_operand,
641 RegisterClass GPR, RegisterClass GPRsp,
642 AArch64Reg ZR, ValueType Ty> {
643 // All registers for non-S variants allow SP
644 def _s : A64I_addsubimm<sf, op, 0b0, shift,
646 (ins GPRsp:$Rn, imm_operand:$Imm12),
647 !strconcat(asmop, "\t$Rd, $Rn, $Imm12"),
648 [(set Ty:$Rd, (add Ty:$Rn, imm_operand:$Imm12))],
652 // S variants can read SP but would write to ZR
653 def _S : A64I_addsubimm<sf, op, 0b1, shift,
655 (ins GPRsp:$Rn, imm_operand:$Imm12),
656 !strconcat(asmop, "s\t$Rd, $Rn, $Imm12"),
657 [(set Ty:$Rd, (addc Ty:$Rn, imm_operand:$Imm12))],
662 // Note that the pattern here for ADDS is subtle. Canonically CMP
663 // a, b becomes SUBS a, b. If b < 0 then this is equivalent to
664 // ADDS a, (-b). This is not true in general.
665 def _cmp : A64I_addsubimm<sf, op, 0b1, shift,
666 (outs), (ins GPRsp:$Rn, imm_operand:$Imm12),
667 !strconcat(cmpasmop, " $Rn, $Imm12"),
669 (A64cmp Ty:$Rn, cmp_imm_operand:$Imm12))],
678 multiclass addsubimm_shifts<string prefix, bit sf, bit op,
679 string asmop, string cmpasmop, string operand, string cmpoperand,
680 RegisterClass GPR, RegisterClass GPRsp, AArch64Reg ZR,
682 defm _lsl0 : addsubimm_varieties<prefix # "_lsl0", sf, op, 0b00,
684 !cast<Operand>(operand # "_lsl0"),
685 !cast<Operand>(cmpoperand # "_lsl0"),
688 defm _lsl12 : addsubimm_varieties<prefix # "_lsl12", sf, op, 0b01,
690 !cast<Operand>(operand # "_lsl12"),
691 !cast<Operand>(cmpoperand # "_lsl12"),
695 defm ADDwwi : addsubimm_shifts<"ADDwi", 0b0, 0b0, "add", "cmn",
696 "addsubimm_operand_i32_posimm",
697 "addsubimm_operand_i32_negimm",
698 GPR32, GPR32wsp, WZR, i32>;
699 defm ADDxxi : addsubimm_shifts<"ADDxi", 0b1, 0b0, "add", "cmn",
700 "addsubimm_operand_i64_posimm",
701 "addsubimm_operand_i64_negimm",
702 GPR64, GPR64xsp, XZR, i64>;
703 defm SUBwwi : addsubimm_shifts<"SUBwi", 0b0, 0b1, "sub", "cmp",
704 "addsubimm_operand_i32_negimm",
705 "addsubimm_operand_i32_posimm",
706 GPR32, GPR32wsp, WZR, i32>;
707 defm SUBxxi : addsubimm_shifts<"SUBxi", 0b1, 0b1, "sub", "cmp",
708 "addsubimm_operand_i64_negimm",
709 "addsubimm_operand_i64_posimm",
710 GPR64, GPR64xsp, XZR, i64>;
712 multiclass MOVsp<RegisterClass GPRsp, RegisterClass SP, Instruction addop> {
713 def _fromsp : InstAlias<"mov $Rd, $Rn",
714 (addop GPRsp:$Rd, SP:$Rn, 0),
717 def _tosp : InstAlias<"mov $Rd, $Rn",
718 (addop SP:$Rd, GPRsp:$Rn, 0),
722 // Recall Rxsp is a RegisterClass containing *just* xsp.
723 defm MOVxx : MOVsp<GPR64xsp, Rxsp, ADDxxi_lsl0_s>;
724 defm MOVww : MOVsp<GPR32wsp, Rwsp, ADDwwi_lsl0_s>;
726 //===----------------------------------------------------------------------===//
727 // Add-subtract (shifted register) instructions
728 //===----------------------------------------------------------------------===//
729 // Contains: ADD, ADDS, SUB, SUBS + aliases CMN, CMP, NEG, NEGS
731 //===-------------------------------
732 // 1. The "shifed register" operands. Shared with logical insts.
733 //===-------------------------------
735 multiclass shift_operands<string prefix, string form> {
736 def _asmoperand_i32 : AsmOperandClass {
737 let Name = "Shift" # form # "i32";
738 let RenderMethod = "addShiftOperands";
739 let PredicateMethod = "isShift<A64SE::" # form # ", false>";
740 let DiagnosticType = "AddSubRegShift32";
743 // Note that the operand type is intentionally i64 because the DAGCombiner
744 // puts these into a canonical form.
745 def _i32 : Operand<i64>, ImmLeaf<i64, [{ return Imm >= 0 && Imm <= 31; }]> {
747 = !cast<AsmOperandClass>(prefix # "_asmoperand_i32");
748 let PrintMethod = "printShiftOperand<A64SE::" # form # ">";
749 let DecoderMethod = "Decode32BitShiftOperand";
752 def _asmoperand_i64 : AsmOperandClass {
753 let Name = "Shift" # form # "i64";
754 let RenderMethod = "addShiftOperands";
755 let PredicateMethod = "isShift<A64SE::" # form # ", true>";
756 let DiagnosticType = "AddSubRegShift64";
759 def _i64 : Operand<i64>, ImmLeaf<i64, [{ return Imm >= 0 && Imm <= 63; }]> {
761 = !cast<AsmOperandClass>(prefix # "_asmoperand_i64");
762 let PrintMethod = "printShiftOperand<A64SE::" # form # ">";
766 defm lsl_operand : shift_operands<"lsl_operand", "LSL">;
767 defm lsr_operand : shift_operands<"lsr_operand", "LSR">;
768 defm asr_operand : shift_operands<"asr_operand", "ASR">;
770 // Not used for add/sub, but defined here for completeness. The "logical
771 // (shifted register)" instructions *do* have an ROR variant.
772 defm ror_operand : shift_operands<"ror_operand", "ROR">;
774 //===-------------------------------
775 // 2. The basic 3.5-operand ADD/SUB/ADDS/SUBS instructions.
776 //===-------------------------------
778 // N.b. the commutable parameter is just !N. It will be first against the wall
779 // when the revolution comes.
780 multiclass addsub_shifts<string prefix, bit sf, bit op, bit s, bit commutable,
781 string asmop, SDPatternOperator opfrag, ValueType ty,
782 RegisterClass GPR, list<Register> defs> {
783 let isCommutable = commutable, Defs = defs in {
784 def _lsl : A64I_addsubshift<sf, op, s, 0b00,
786 (ins GPR:$Rn, GPR:$Rm,
787 !cast<Operand>("lsl_operand_" # ty):$Imm6),
788 !strconcat(asmop, "\t$Rd, $Rn, $Rm, $Imm6"),
789 [(set GPR:$Rd, (opfrag ty:$Rn, (shl ty:$Rm,
790 !cast<Operand>("lsl_operand_" # ty):$Imm6))
794 def _lsr : A64I_addsubshift<sf, op, s, 0b01,
796 (ins GPR:$Rn, GPR:$Rm,
797 !cast<Operand>("lsr_operand_" # ty):$Imm6),
798 !strconcat(asmop, "\t$Rd, $Rn, $Rm, $Imm6"),
799 [(set ty:$Rd, (opfrag ty:$Rn, (srl ty:$Rm,
800 !cast<Operand>("lsr_operand_" # ty):$Imm6))
804 def _asr : A64I_addsubshift<sf, op, s, 0b10,
806 (ins GPR:$Rn, GPR:$Rm,
807 !cast<Operand>("asr_operand_" # ty):$Imm6),
808 !strconcat(asmop, "\t$Rd, $Rn, $Rm, $Imm6"),
809 [(set ty:$Rd, (opfrag ty:$Rn, (sra ty:$Rm,
810 !cast<Operand>("asr_operand_" # ty):$Imm6))
816 : InstAlias<!strconcat(asmop, " $Rd, $Rn, $Rm"),
817 (!cast<Instruction>(prefix # "_lsl") GPR:$Rd, GPR:$Rn,
820 def : Pat<(opfrag ty:$Rn, ty:$Rm),
821 (!cast<Instruction>(prefix # "_lsl") $Rn, $Rm, 0)>;
824 multiclass addsub_sizes<string prefix, bit op, bit s, bit commutable,
825 string asmop, SDPatternOperator opfrag,
826 list<Register> defs> {
827 defm xxx : addsub_shifts<prefix # "xxx", 0b1, op, s,
828 commutable, asmop, opfrag, i64, GPR64, defs>;
829 defm www : addsub_shifts<prefix # "www", 0b0, op, s,
830 commutable, asmop, opfrag, i32, GPR32, defs>;
834 defm ADD : addsub_sizes<"ADD", 0b0, 0b0, 0b1, "add", add, []>;
835 defm SUB : addsub_sizes<"SUB", 0b1, 0b0, 0b0, "sub", sub, []>;
837 defm ADDS : addsub_sizes<"ADDS", 0b0, 0b1, 0b1, "adds", addc, [NZCV]>;
838 defm SUBS : addsub_sizes<"SUBS", 0b1, 0b1, 0b0, "subs", subc, [NZCV]>;
840 //===-------------------------------
841 // 1. The NEG/NEGS aliases
842 //===-------------------------------
844 multiclass neg_alias<Instruction INST, RegisterClass GPR, Register ZR,
845 ValueType ty, Operand shift_operand, SDNode shiftop> {
846 def : InstAlias<"neg $Rd, $Rm, $Imm6",
847 (INST GPR:$Rd, ZR, GPR:$Rm, shift_operand:$Imm6)>;
849 def : Pat<(sub 0, (shiftop ty:$Rm, shift_operand:$Imm6)),
850 (INST ZR, $Rm, shift_operand:$Imm6)>;
853 defm : neg_alias<SUBwww_lsl, GPR32, WZR, i32, lsl_operand_i32, shl>;
854 defm : neg_alias<SUBwww_lsr, GPR32, WZR, i32, lsr_operand_i32, srl>;
855 defm : neg_alias<SUBwww_asr, GPR32, WZR, i32, asr_operand_i32, sra>;
856 def : InstAlias<"neg $Rd, $Rm", (SUBwww_lsl GPR32:$Rd, WZR, GPR32:$Rm, 0)>;
857 def : Pat<(sub 0, i32:$Rm), (SUBwww_lsl WZR, $Rm, 0)>;
859 defm : neg_alias<SUBxxx_lsl, GPR64, XZR, i64, lsl_operand_i64, shl>;
860 defm : neg_alias<SUBxxx_lsr, GPR64, XZR, i64, lsr_operand_i64, srl>;
861 defm : neg_alias<SUBxxx_asr, GPR64, XZR, i64, asr_operand_i64, sra>;
862 def : InstAlias<"neg $Rd, $Rm", (SUBxxx_lsl GPR64:$Rd, XZR, GPR64:$Rm, 0)>;
863 def : Pat<(sub 0, i64:$Rm), (SUBxxx_lsl XZR, $Rm, 0)>;
865 // NEGS doesn't get any patterns yet: defining multiple outputs means C++ has to
867 class negs_alias<Instruction INST, RegisterClass GPR,
868 Register ZR, Operand shift_operand, SDNode shiftop>
869 : InstAlias<"negs $Rd, $Rm, $Imm6",
870 (INST GPR:$Rd, ZR, GPR:$Rm, shift_operand:$Imm6)>;
872 def : negs_alias<SUBSwww_lsl, GPR32, WZR, lsl_operand_i32, shl>;
873 def : negs_alias<SUBSwww_lsr, GPR32, WZR, lsr_operand_i32, srl>;
874 def : negs_alias<SUBSwww_asr, GPR32, WZR, asr_operand_i32, sra>;
875 def : InstAlias<"negs $Rd, $Rm", (SUBSwww_lsl GPR32:$Rd, WZR, GPR32:$Rm, 0)>;
877 def : negs_alias<SUBSxxx_lsl, GPR64, XZR, lsl_operand_i64, shl>;
878 def : negs_alias<SUBSxxx_lsr, GPR64, XZR, lsr_operand_i64, srl>;
879 def : negs_alias<SUBSxxx_asr, GPR64, XZR, asr_operand_i64, sra>;
880 def : InstAlias<"negs $Rd, $Rm", (SUBSxxx_lsl GPR64:$Rd, XZR, GPR64:$Rm, 0)>;
882 //===-------------------------------
883 // 1. The CMP/CMN aliases
884 //===-------------------------------
886 multiclass cmp_shifts<string prefix, bit sf, bit op, bit commutable,
887 string asmop, SDPatternOperator opfrag, ValueType ty,
889 let isCommutable = commutable, Rd = 0b11111, Defs = [NZCV] in {
890 def _lsl : A64I_addsubshift<sf, op, 0b1, 0b00,
892 (ins GPR:$Rn, GPR:$Rm,
893 !cast<Operand>("lsl_operand_" # ty):$Imm6),
894 !strconcat(asmop, "\t$Rn, $Rm, $Imm6"),
895 [(set NZCV, (opfrag ty:$Rn, (shl ty:$Rm,
896 !cast<Operand>("lsl_operand_" # ty):$Imm6))
900 def _lsr : A64I_addsubshift<sf, op, 0b1, 0b01,
902 (ins GPR:$Rn, GPR:$Rm,
903 !cast<Operand>("lsr_operand_" # ty):$Imm6),
904 !strconcat(asmop, "\t$Rn, $Rm, $Imm6"),
905 [(set NZCV, (opfrag ty:$Rn, (srl ty:$Rm,
906 !cast<Operand>("lsr_operand_" # ty):$Imm6))
910 def _asr : A64I_addsubshift<sf, op, 0b1, 0b10,
912 (ins GPR:$Rn, GPR:$Rm,
913 !cast<Operand>("asr_operand_" # ty):$Imm6),
914 !strconcat(asmop, "\t$Rn, $Rm, $Imm6"),
915 [(set NZCV, (opfrag ty:$Rn, (sra ty:$Rm,
916 !cast<Operand>("asr_operand_" # ty):$Imm6))
922 : InstAlias<!strconcat(asmop, " $Rn, $Rm"),
923 (!cast<Instruction>(prefix # "_lsl") GPR:$Rn, GPR:$Rm, 0)>;
925 def : Pat<(opfrag ty:$Rn, ty:$Rm),
926 (!cast<Instruction>(prefix # "_lsl") $Rn, $Rm, 0)>;
929 defm CMPww : cmp_shifts<"CMPww", 0b0, 0b1, 0b0, "cmp", A64cmp, i32, GPR32>;
930 defm CMPxx : cmp_shifts<"CMPxx", 0b1, 0b1, 0b0, "cmp", A64cmp, i64, GPR64>;
932 defm CMNww : cmp_shifts<"CMNww", 0b0, 0b0, 0b1, "cmn", A64cmn, i32, GPR32>;
933 defm CMNxx : cmp_shifts<"CMNxx", 0b1, 0b0, 0b1, "cmn", A64cmn, i64, GPR64>;
935 //===----------------------------------------------------------------------===//
936 // Add-subtract (with carry) instructions
937 //===----------------------------------------------------------------------===//
938 // Contains: ADC, ADCS, SBC, SBCS + aliases NGC, NGCS
940 multiclass A64I_addsubcarrySizes<bit op, bit s, string asmop> {
941 let Uses = [NZCV] in {
942 def www : A64I_addsubcarry<0b0, op, s, 0b000000,
943 (outs GPR32:$Rd), (ins GPR32:$Rn, GPR32:$Rm),
944 !strconcat(asmop, "\t$Rd, $Rn, $Rm"),
947 def xxx : A64I_addsubcarry<0b1, op, s, 0b000000,
948 (outs GPR64:$Rd), (ins GPR64:$Rn, GPR64:$Rm),
949 !strconcat(asmop, "\t$Rd, $Rn, $Rm"),
954 let isCommutable = 1 in {
955 defm ADC : A64I_addsubcarrySizes<0b0, 0b0, "adc">;
958 defm SBC : A64I_addsubcarrySizes<0b1, 0b0, "sbc">;
960 let Defs = [NZCV] in {
961 let isCommutable = 1 in {
962 defm ADCS : A64I_addsubcarrySizes<0b0, 0b1, "adcs">;
965 defm SBCS : A64I_addsubcarrySizes<0b1, 0b1, "sbcs">;
968 def : InstAlias<"ngc $Rd, $Rm", (SBCwww GPR32:$Rd, WZR, GPR32:$Rm)>;
969 def : InstAlias<"ngc $Rd, $Rm", (SBCxxx GPR64:$Rd, XZR, GPR64:$Rm)>;
970 def : InstAlias<"ngcs $Rd, $Rm", (SBCSwww GPR32:$Rd, WZR, GPR32:$Rm)>;
971 def : InstAlias<"ngcs $Rd, $Rm", (SBCSxxx GPR64:$Rd, XZR, GPR64:$Rm)>;
973 // Note that adde and sube can form a chain longer than two (e.g. for 256-bit
974 // addition). So the flag-setting instructions are appropriate.
975 def : Pat<(adde i32:$Rn, i32:$Rm), (ADCSwww $Rn, $Rm)>;
976 def : Pat<(adde i64:$Rn, i64:$Rm), (ADCSxxx $Rn, $Rm)>;
977 def : Pat<(sube i32:$Rn, i32:$Rm), (SBCSwww $Rn, $Rm)>;
978 def : Pat<(sube i64:$Rn, i64:$Rm), (SBCSxxx $Rn, $Rm)>;
980 //===----------------------------------------------------------------------===//
982 //===----------------------------------------------------------------------===//
983 // Contains: SBFM, BFM, UBFM, [SU]XT[BHW], ASR, LSR, LSL, SBFI[ZX], BFI, BFXIL,
986 // Because of the rather complicated nearly-overlapping aliases, the decoding of
987 // this range of instructions is handled manually. The architectural
988 // instructions are BFM, SBFM and UBFM but a disassembler should never produce
991 // In the end, the best option was to use BFM instructions for decoding under
992 // almost all circumstances, but to create aliasing *Instructions* for each of
993 // the canonical forms and specify a completely custom decoder which would
994 // substitute the correct MCInst as needed.
996 // This also simplifies instruction selection, parsing etc because the MCInsts
997 // have a shape that's closer to their use in code.
999 //===-------------------------------
1000 // 1. The architectural BFM instructions
1001 //===-------------------------------
1003 def uimm5_asmoperand : AsmOperandClass {
1005 let PredicateMethod = "isUImm<5>";
1006 let RenderMethod = "addImmOperands";
1007 let DiagnosticType = "UImm5";
1010 def uimm6_asmoperand : AsmOperandClass {
1012 let PredicateMethod = "isUImm<6>";
1013 let RenderMethod = "addImmOperands";
1014 let DiagnosticType = "UImm6";
1017 def bitfield32_imm : Operand<i64>,
1018 ImmLeaf<i64, [{ return Imm >= 0 && Imm < 32; }]> {
1019 let ParserMatchClass = uimm5_asmoperand;
1021 let DecoderMethod = "DecodeBitfield32ImmOperand";
1025 def bitfield64_imm : Operand<i64>,
1026 ImmLeaf<i64, [{ return Imm >= 0 && Imm < 64; }]> {
1027 let ParserMatchClass = uimm6_asmoperand;
1029 // Default decoder works in 64-bit case: the 6-bit field can take any value.
1032 multiclass A64I_bitfieldSizes<bits<2> opc, string asmop> {
1033 def wwii : A64I_bitfield<0b0, opc, 0b0, (outs GPR32:$Rd),
1034 (ins GPR32:$Rn, bitfield32_imm:$ImmR, bitfield32_imm:$ImmS),
1035 !strconcat(asmop, "\t$Rd, $Rn, $ImmR, $ImmS"),
1037 let DecoderMethod = "DecodeBitfieldInstruction";
1040 def xxii : A64I_bitfield<0b1, opc, 0b1, (outs GPR64:$Rd),
1041 (ins GPR64:$Rn, bitfield64_imm:$ImmR, bitfield64_imm:$ImmS),
1042 !strconcat(asmop, "\t$Rd, $Rn, $ImmR, $ImmS"),
1044 let DecoderMethod = "DecodeBitfieldInstruction";
1048 defm SBFM : A64I_bitfieldSizes<0b00, "sbfm">;
1049 defm UBFM : A64I_bitfieldSizes<0b10, "ubfm">;
1051 // BFM instructions modify the destination register rather than defining it
1054 A64I_bitfield<0b0, 0b01, 0b0, (outs GPR32:$Rd),
1055 (ins GPR32:$src, GPR32:$Rn, bitfield32_imm:$ImmR, bitfield32_imm:$ImmS),
1056 "bfm\t$Rd, $Rn, $ImmR, $ImmS", [], NoItinerary> {
1057 let DecoderMethod = "DecodeBitfieldInstruction";
1058 let Constraints = "$src = $Rd";
1062 A64I_bitfield<0b1, 0b01, 0b1, (outs GPR64:$Rd),
1063 (ins GPR64:$src, GPR64:$Rn, bitfield64_imm:$ImmR, bitfield64_imm:$ImmS),
1064 "bfm\t$Rd, $Rn, $ImmR, $ImmS", [], NoItinerary> {
1065 let DecoderMethod = "DecodeBitfieldInstruction";
1066 let Constraints = "$src = $Rd";
1070 //===-------------------------------
1071 // 2. Extend aliases to 64-bit dest
1072 //===-------------------------------
1074 // Unfortunately the extensions that end up as 64-bits cannot be handled by an
1075 // instruction alias: their syntax is (for example) "SXTB x0, w0", which needs
1076 // to be mapped to "SBFM x0, x0, #0, 7" (changing the class of Rn). InstAlias is
1077 // not capable of such a map as far as I'm aware
1079 // Note that these instructions are strictly more specific than the
1080 // BFM ones (in ImmR) so they can handle their own decoding.
1081 class A64I_bf_ext<bit sf, bits<2> opc, RegisterClass GPRDest, ValueType dty,
1082 string asmop, bits<6> imms, dag pattern>
1083 : A64I_bitfield<sf, opc, sf,
1084 (outs GPRDest:$Rd), (ins GPR32:$Rn),
1085 !strconcat(asmop, "\t$Rd, $Rn"),
1086 [(set dty:$Rd, pattern)], NoItinerary> {
1087 let ImmR = 0b000000;
1091 // Signed extensions
1092 def SXTBxw : A64I_bf_ext<0b1, 0b00, GPR64, i64, "sxtb", 7,
1093 (sext_inreg (anyext i32:$Rn), i8)>;
1094 def SXTBww : A64I_bf_ext<0b0, 0b00, GPR32, i32, "sxtb", 7,
1095 (sext_inreg i32:$Rn, i8)>;
1096 def SXTHxw : A64I_bf_ext<0b1, 0b00, GPR64, i64, "sxth", 15,
1097 (sext_inreg (anyext i32:$Rn), i16)>;
1098 def SXTHww : A64I_bf_ext<0b0, 0b00, GPR32, i32, "sxth", 15,
1099 (sext_inreg i32:$Rn, i16)>;
1100 def SXTWxw : A64I_bf_ext<0b1, 0b00, GPR64, i64, "sxtw", 31, (sext i32:$Rn)>;
1102 // Unsigned extensions
1103 def UXTBww : A64I_bf_ext<0b0, 0b10, GPR32, i32, "uxtb", 7,
1104 (and i32:$Rn, 255)>;
1105 def UXTHww : A64I_bf_ext<0b0, 0b10, GPR32, i32, "uxth", 15,
1106 (and i32:$Rn, 65535)>;
1108 // The 64-bit unsigned variants are not strictly architectural but recommended
1110 let isAsmParserOnly = 1 in {
1111 def UXTBxw : A64I_bf_ext<0b0, 0b10, GPR64, i64, "uxtb", 7,
1112 (and (anyext i32:$Rn), 255)>;
1113 def UXTHxw : A64I_bf_ext<0b0, 0b10, GPR64, i64, "uxth", 15,
1114 (and (anyext i32:$Rn), 65535)>;
1117 // Extra patterns for when the source register is actually 64-bits
1118 // too. There's no architectural difference here, it's just LLVM
1119 // shinanigans. There's no need for equivalent zero-extension patterns
1120 // because they'll already be caught by logical (immediate) matching.
1121 def : Pat<(sext_inreg i64:$Rn, i8),
1122 (SXTBxw (EXTRACT_SUBREG $Rn, sub_32))>;
1123 def : Pat<(sext_inreg i64:$Rn, i16),
1124 (SXTHxw (EXTRACT_SUBREG $Rn, sub_32))>;
1125 def : Pat<(sext_inreg i64:$Rn, i32),
1126 (SXTWxw (EXTRACT_SUBREG $Rn, sub_32))>;
1129 //===-------------------------------
1130 // 3. Aliases for ASR and LSR (the simple shifts)
1131 //===-------------------------------
1133 // These also handle their own decoding because ImmS being set makes
1134 // them take precedence over BFM.
1135 multiclass A64I_shift<bits<2> opc, string asmop, SDNode opnode> {
1136 def wwi : A64I_bitfield<0b0, opc, 0b0,
1137 (outs GPR32:$Rd), (ins GPR32:$Rn, bitfield32_imm:$ImmR),
1138 !strconcat(asmop, "\t$Rd, $Rn, $ImmR"),
1139 [(set i32:$Rd, (opnode i32:$Rn, bitfield32_imm:$ImmR))],
1144 def xxi : A64I_bitfield<0b1, opc, 0b1,
1145 (outs GPR64:$Rd), (ins GPR64:$Rn, bitfield64_imm:$ImmR),
1146 !strconcat(asmop, "\t$Rd, $Rn, $ImmR"),
1147 [(set i64:$Rd, (opnode i64:$Rn, bitfield64_imm:$ImmR))],
1154 defm ASR : A64I_shift<0b00, "asr", sra>;
1155 defm LSR : A64I_shift<0b10, "lsr", srl>;
1157 //===-------------------------------
1158 // 4. Aliases for LSL
1159 //===-------------------------------
1161 // Unfortunately LSL and subsequent aliases are much more complicated. We need
1162 // to be able to say certain output instruction fields depend in a complex
1163 // manner on combinations of input assembly fields).
1165 // MIOperandInfo *might* have been able to do it, but at the cost of
1166 // significantly more C++ code.
1168 // N.b. contrary to usual practice these operands store the shift rather than
1169 // the machine bits in an MCInst. The complexity overhead of consistency
1170 // outweighed the benefits in this case (custom asmparser, printer and selection
1171 // vs custom encoder).
1172 def bitfield32_lsl_imm : Operand<i64>,
1173 ImmLeaf<i64, [{ return Imm >= 0 && Imm <= 31; }]> {
1174 let ParserMatchClass = uimm5_asmoperand;
1175 let EncoderMethod = "getBitfield32LSLOpValue";
1178 def bitfield64_lsl_imm : Operand<i64>,
1179 ImmLeaf<i64, [{ return Imm >= 0 && Imm <= 63; }]> {
1180 let ParserMatchClass = uimm6_asmoperand;
1181 let EncoderMethod = "getBitfield64LSLOpValue";
1184 class A64I_bitfield_lsl<bit sf, RegisterClass GPR, ValueType ty,
1186 : A64I_bitfield<sf, 0b10, sf, (outs GPR:$Rd), (ins GPR:$Rn, operand:$FullImm),
1187 "lsl\t$Rd, $Rn, $FullImm",
1188 [(set ty:$Rd, (shl ty:$Rn, operand:$FullImm))],
1191 let ImmR = FullImm{5-0};
1192 let ImmS = FullImm{11-6};
1194 // No disassembler allowed because it would overlap with BFM which does the
1196 let isAsmParserOnly = 1;
1199 def LSLwwi : A64I_bitfield_lsl<0b0, GPR32, i32, bitfield32_lsl_imm>;
1200 def LSLxxi : A64I_bitfield_lsl<0b1, GPR64, i64, bitfield64_lsl_imm>;
1202 //===-------------------------------
1203 // 5. Aliases for bitfield extract instructions
1204 //===-------------------------------
1206 def bfx32_width_asmoperand : AsmOperandClass {
1207 let Name = "BFX32Width";
1208 let PredicateMethod = "isBitfieldWidth<32>";
1209 let RenderMethod = "addBFXWidthOperands";
1210 let DiagnosticType = "Width32";
1213 def bfx32_width : Operand<i64>, ImmLeaf<i64, [{ return true; }]> {
1214 let PrintMethod = "printBFXWidthOperand";
1215 let ParserMatchClass = bfx32_width_asmoperand;
1218 def bfx64_width_asmoperand : AsmOperandClass {
1219 let Name = "BFX64Width";
1220 let PredicateMethod = "isBitfieldWidth<64>";
1221 let RenderMethod = "addBFXWidthOperands";
1222 let DiagnosticType = "Width64";
1225 def bfx64_width : Operand<i64> {
1226 let PrintMethod = "printBFXWidthOperand";
1227 let ParserMatchClass = bfx64_width_asmoperand;
1231 multiclass A64I_bitfield_extract<bits<2> opc, string asmop, SDNode op> {
1232 def wwii : A64I_bitfield<0b0, opc, 0b0, (outs GPR32:$Rd),
1233 (ins GPR32:$Rn, bitfield32_imm:$ImmR, bfx32_width:$ImmS),
1234 !strconcat(asmop, "\t$Rd, $Rn, $ImmR, $ImmS"),
1235 [(set i32:$Rd, (op i32:$Rn, imm:$ImmR, imm:$ImmS))],
1237 // As above, no disassembler allowed.
1238 let isAsmParserOnly = 1;
1241 def xxii : A64I_bitfield<0b1, opc, 0b1, (outs GPR64:$Rd),
1242 (ins GPR64:$Rn, bitfield64_imm:$ImmR, bfx64_width:$ImmS),
1243 !strconcat(asmop, "\t$Rd, $Rn, $ImmR, $ImmS"),
1244 [(set i64:$Rd, (op i64:$Rn, imm:$ImmR, imm:$ImmS))],
1246 // As above, no disassembler allowed.
1247 let isAsmParserOnly = 1;
1251 defm SBFX : A64I_bitfield_extract<0b00, "sbfx", A64Sbfx>;
1252 defm UBFX : A64I_bitfield_extract<0b10, "ubfx", A64Ubfx>;
1254 // Again, variants based on BFM modify Rd so need it as an input too.
1255 def BFXILwwii : A64I_bitfield<0b0, 0b01, 0b0, (outs GPR32:$Rd),
1256 (ins GPR32:$src, GPR32:$Rn, bitfield32_imm:$ImmR, bfx32_width:$ImmS),
1257 "bfxil\t$Rd, $Rn, $ImmR, $ImmS", [], NoItinerary> {
1258 // As above, no disassembler allowed.
1259 let isAsmParserOnly = 1;
1260 let Constraints = "$src = $Rd";
1263 def BFXILxxii : A64I_bitfield<0b1, 0b01, 0b1, (outs GPR64:$Rd),
1264 (ins GPR64:$src, GPR64:$Rn, bitfield64_imm:$ImmR, bfx64_width:$ImmS),
1265 "bfxil\t$Rd, $Rn, $ImmR, $ImmS", [], NoItinerary> {
1266 // As above, no disassembler allowed.
1267 let isAsmParserOnly = 1;
1268 let Constraints = "$src = $Rd";
1271 // SBFX instructions can do a 1-instruction sign-extension of boolean values.
1272 def : Pat<(sext_inreg i64:$Rn, i1), (SBFXxxii $Rn, 0, 0)>;
1273 def : Pat<(sext_inreg i32:$Rn, i1), (SBFXwwii $Rn, 0, 0)>;
1274 def : Pat<(i64 (sext_inreg (anyext i32:$Rn), i1)),
1275 (SBFXxxii (SUBREG_TO_REG (i64 0), $Rn, sub_32), 0, 0)>;
1277 // UBFX makes sense as an implementation of a 64-bit zero-extension too. Could
1278 // use either 64-bit or 32-bit variant, but 32-bit might be more efficient.
1279 def : Pat<(zext i32:$Rn), (SUBREG_TO_REG (i64 0), (UBFXwwii $Rn, 0, 31),
1282 //===-------------------------------
1283 // 6. Aliases for bitfield insert instructions
1284 //===-------------------------------
1286 def bfi32_lsb_asmoperand : AsmOperandClass {
1287 let Name = "BFI32LSB";
1288 let PredicateMethod = "isUImm<5>";
1289 let RenderMethod = "addBFILSBOperands<32>";
1290 let DiagnosticType = "UImm5";
1293 def bfi32_lsb : Operand<i64>,
1294 ImmLeaf<i64, [{ return Imm >= 0 && Imm <= 31; }]> {
1295 let PrintMethod = "printBFILSBOperand<32>";
1296 let ParserMatchClass = bfi32_lsb_asmoperand;
1299 def bfi64_lsb_asmoperand : AsmOperandClass {
1300 let Name = "BFI64LSB";
1301 let PredicateMethod = "isUImm<6>";
1302 let RenderMethod = "addBFILSBOperands<64>";
1303 let DiagnosticType = "UImm6";
1306 def bfi64_lsb : Operand<i64>,
1307 ImmLeaf<i64, [{ return Imm >= 0 && Imm <= 63; }]> {
1308 let PrintMethod = "printBFILSBOperand<64>";
1309 let ParserMatchClass = bfi64_lsb_asmoperand;
1312 // Width verification is performed during conversion so width operand can be
1313 // shared between 32/64-bit cases. Still needed for the print method though
1314 // because ImmR encodes "width - 1".
1315 def bfi32_width_asmoperand : AsmOperandClass {
1316 let Name = "BFI32Width";
1317 let PredicateMethod = "isBitfieldWidth<32>";
1318 let RenderMethod = "addBFIWidthOperands";
1319 let DiagnosticType = "Width32";
1322 def bfi32_width : Operand<i64>,
1323 ImmLeaf<i64, [{ return Imm >= 1 && Imm <= 32; }]> {
1324 let PrintMethod = "printBFIWidthOperand";
1325 let ParserMatchClass = bfi32_width_asmoperand;
1328 def bfi64_width_asmoperand : AsmOperandClass {
1329 let Name = "BFI64Width";
1330 let PredicateMethod = "isBitfieldWidth<64>";
1331 let RenderMethod = "addBFIWidthOperands";
1332 let DiagnosticType = "Width64";
1335 def bfi64_width : Operand<i64>,
1336 ImmLeaf<i64, [{ return Imm >= 1 && Imm <= 64; }]> {
1337 let PrintMethod = "printBFIWidthOperand";
1338 let ParserMatchClass = bfi64_width_asmoperand;
1341 multiclass A64I_bitfield_insert<bits<2> opc, string asmop> {
1342 def wwii : A64I_bitfield<0b0, opc, 0b0, (outs GPR32:$Rd),
1343 (ins GPR32:$Rn, bfi32_lsb:$ImmR, bfi32_width:$ImmS),
1344 !strconcat(asmop, "\t$Rd, $Rn, $ImmR, $ImmS"),
1346 // As above, no disassembler allowed.
1347 let isAsmParserOnly = 1;
1350 def xxii : A64I_bitfield<0b1, opc, 0b1, (outs GPR64:$Rd),
1351 (ins GPR64:$Rn, bfi64_lsb:$ImmR, bfi64_width:$ImmS),
1352 !strconcat(asmop, "\t$Rd, $Rn, $ImmR, $ImmS"),
1354 // As above, no disassembler allowed.
1355 let isAsmParserOnly = 1;
1359 defm SBFIZ : A64I_bitfield_insert<0b00, "sbfiz">;
1360 defm UBFIZ : A64I_bitfield_insert<0b10, "ubfiz">;
1363 def BFIwwii : A64I_bitfield<0b0, 0b01, 0b0, (outs GPR32:$Rd),
1364 (ins GPR32:$src, GPR32:$Rn, bfi32_lsb:$ImmR, bfi32_width:$ImmS),
1365 "bfi\t$Rd, $Rn, $ImmR, $ImmS", [], NoItinerary> {
1366 // As above, no disassembler allowed.
1367 let isAsmParserOnly = 1;
1368 let Constraints = "$src = $Rd";
1371 def BFIxxii : A64I_bitfield<0b1, 0b01, 0b1, (outs GPR64:$Rd),
1372 (ins GPR64:$src, GPR64:$Rn, bfi64_lsb:$ImmR, bfi64_width:$ImmS),
1373 "bfi\t$Rd, $Rn, $ImmR, $ImmS", [], NoItinerary> {
1374 // As above, no disassembler allowed.
1375 let isAsmParserOnly = 1;
1376 let Constraints = "$src = $Rd";
1379 //===----------------------------------------------------------------------===//
1380 // Compare and branch (immediate)
1381 //===----------------------------------------------------------------------===//
1382 // Contains: CBZ, CBNZ
1384 class label_asmoperand<int width, int scale> : AsmOperandClass {
1385 let Name = "Label" # width # "_" # scale;
1386 let PredicateMethod = "isLabel<" # width # "," # scale # ">";
1387 let RenderMethod = "addLabelOperands<" # width # ", " # scale # ">";
1388 let DiagnosticType = "Label";
1391 def label_wid19_scal4_asmoperand : label_asmoperand<19, 4>;
1393 // All conditional immediate branches are the same really: 19 signed bits scaled
1394 // by the instruction-size (4).
1395 def bcc_target : Operand<OtherVT> {
1396 // This label is a 19-bit offset from PC, scaled by the instruction-width: 4.
1397 let ParserMatchClass = label_wid19_scal4_asmoperand;
1398 let PrintMethod = "printLabelOperand<19, 4>";
1399 let EncoderMethod = "getLabelOpValue<AArch64::fixup_a64_condbr>";
1400 let OperandType = "OPERAND_PCREL";
1403 multiclass cmpbr_sizes<bit op, string asmop, ImmLeaf SETOP> {
1404 let isBranch = 1, isTerminator = 1 in {
1405 def x : A64I_cmpbr<0b1, op,
1407 (ins GPR64:$Rt, bcc_target:$Label),
1408 !strconcat(asmop,"\t$Rt, $Label"),
1409 [(A64br_cc (A64cmp i64:$Rt, 0), SETOP, bb:$Label)],
1412 def w : A64I_cmpbr<0b0, op,
1414 (ins GPR32:$Rt, bcc_target:$Label),
1415 !strconcat(asmop,"\t$Rt, $Label"),
1416 [(A64br_cc (A64cmp i32:$Rt, 0), SETOP, bb:$Label)],
1421 defm CBZ : cmpbr_sizes<0b0, "cbz", ImmLeaf<i32, [{
1422 return Imm == A64CC::EQ;
1424 defm CBNZ : cmpbr_sizes<0b1, "cbnz", ImmLeaf<i32, [{
1425 return Imm == A64CC::NE;
1428 //===----------------------------------------------------------------------===//
1429 // Conditional branch (immediate) instructions
1430 //===----------------------------------------------------------------------===//
1433 def cond_code_asmoperand : AsmOperandClass {
1434 let Name = "CondCode";
1435 let DiagnosticType = "CondCode";
1438 def cond_code : Operand<i32>, ImmLeaf<i32, [{
1439 return Imm >= 0 && Imm <= 15;
1441 let PrintMethod = "printCondCodeOperand";
1442 let ParserMatchClass = cond_code_asmoperand;
1445 def Bcc : A64I_condbr<0b0, 0b0, (outs),
1446 (ins cond_code:$Cond, bcc_target:$Label),
1447 "b.$Cond $Label", [(A64br_cc NZCV, (i32 imm:$Cond), bb:$Label)],
1451 let isTerminator = 1;
1454 //===----------------------------------------------------------------------===//
1455 // Conditional compare (immediate) instructions
1456 //===----------------------------------------------------------------------===//
1457 // Contains: CCMN, CCMP
1459 def uimm4_asmoperand : AsmOperandClass {
1461 let PredicateMethod = "isUImm<4>";
1462 let RenderMethod = "addImmOperands";
1463 let DiagnosticType = "UImm4";
1466 def uimm4 : Operand<i32> {
1467 let ParserMatchClass = uimm4_asmoperand;
1470 def uimm5 : Operand<i32> {
1471 let ParserMatchClass = uimm5_asmoperand;
1474 // The only difference between this operand and the one for instructions like
1475 // B.cc is that it's parsed manually. The other get parsed implicitly as part of
1476 // the mnemonic handling.
1477 def cond_code_op_asmoperand : AsmOperandClass {
1478 let Name = "CondCodeOp";
1479 let RenderMethod = "addCondCodeOperands";
1480 let PredicateMethod = "isCondCode";
1481 let ParserMethod = "ParseCondCodeOperand";
1482 let DiagnosticType = "CondCode";
1485 def cond_code_op : Operand<i32> {
1486 let PrintMethod = "printCondCodeOperand";
1487 let ParserMatchClass = cond_code_op_asmoperand;
1490 class A64I_condcmpimmImpl<bit sf, bit op, RegisterClass GPR, string asmop>
1491 : A64I_condcmpimm<sf, op, 0b0, 0b0, 0b1, (outs),
1492 (ins GPR:$Rn, uimm5:$UImm5, uimm4:$NZCVImm, cond_code_op:$Cond),
1493 !strconcat(asmop, "\t$Rn, $UImm5, $NZCVImm, $Cond"),
1498 def CCMNwi : A64I_condcmpimmImpl<0b0, 0b0, GPR32, "ccmn">;
1499 def CCMNxi : A64I_condcmpimmImpl<0b1, 0b0, GPR64, "ccmn">;
1500 def CCMPwi : A64I_condcmpimmImpl<0b0, 0b1, GPR32, "ccmp">;
1501 def CCMPxi : A64I_condcmpimmImpl<0b1, 0b1, GPR64, "ccmp">;
1503 //===----------------------------------------------------------------------===//
1504 // Conditional compare (register) instructions
1505 //===----------------------------------------------------------------------===//
1506 // Contains: CCMN, CCMP
1508 class A64I_condcmpregImpl<bit sf, bit op, RegisterClass GPR, string asmop>
1509 : A64I_condcmpreg<sf, op, 0b0, 0b0, 0b1,
1511 (ins GPR:$Rn, GPR:$Rm, uimm4:$NZCVImm, cond_code_op:$Cond),
1512 !strconcat(asmop, "\t$Rn, $Rm, $NZCVImm, $Cond"),
1517 def CCMNww : A64I_condcmpregImpl<0b0, 0b0, GPR32, "ccmn">;
1518 def CCMNxx : A64I_condcmpregImpl<0b1, 0b0, GPR64, "ccmn">;
1519 def CCMPww : A64I_condcmpregImpl<0b0, 0b1, GPR32, "ccmp">;
1520 def CCMPxx : A64I_condcmpregImpl<0b1, 0b1, GPR64, "ccmp">;
1522 //===----------------------------------------------------------------------===//
1523 // Conditional select instructions
1524 //===----------------------------------------------------------------------===//
1525 // Contains: CSEL, CSINC, CSINV, CSNEG + aliases CSET, CSETM, CINC, CINV, CNEG
1527 // Condition code which is encoded as the inversion (semantically rather than
1528 // bitwise) in the instruction.
1529 def inv_cond_code_op_asmoperand : AsmOperandClass {
1530 let Name = "InvCondCodeOp";
1531 let RenderMethod = "addInvCondCodeOperands";
1532 let PredicateMethod = "isCondCode";
1533 let ParserMethod = "ParseCondCodeOperand";
1534 let DiagnosticType = "CondCode";
1537 def inv_cond_code_op : Operand<i32> {
1538 let ParserMatchClass = inv_cond_code_op_asmoperand;
1541 // Having a separate operand for the selectable use-case is debatable, but gives
1542 // consistency with cond_code.
1543 def inv_cond_XFORM : SDNodeXForm<imm, [{
1544 A64CC::CondCodes CC = static_cast<A64CC::CondCodes>(N->getZExtValue());
1545 return CurDAG->getTargetConstant(A64InvertCondCode(CC), MVT::i32);
1549 : ImmLeaf<i32, [{ return Imm >= 0 && Imm <= 15; }], inv_cond_XFORM>;
1552 multiclass A64I_condselSizes<bit op, bits<2> op2, string asmop,
1553 SDPatternOperator select> {
1554 let Uses = [NZCV] in {
1555 def wwwc : A64I_condsel<0b0, op, 0b0, op2,
1557 (ins GPR32:$Rn, GPR32:$Rm, cond_code_op:$Cond),
1558 !strconcat(asmop, "\t$Rd, $Rn, $Rm, $Cond"),
1559 [(set i32:$Rd, (select i32:$Rn, i32:$Rm))],
1563 def xxxc : A64I_condsel<0b1, op, 0b0, op2,
1565 (ins GPR64:$Rn, GPR64:$Rm, cond_code_op:$Cond),
1566 !strconcat(asmop, "\t$Rd, $Rn, $Rm, $Cond"),
1567 [(set i64:$Rd, (select i64:$Rn, i64:$Rm))],
1573 : PatFrag<(ops node:$lhs, node:$rhs),
1574 (A64select_cc NZCV, node:$lhs, node:$rhs, (i32 imm:$Cond))>;
1576 class complex_select<SDPatternOperator opnode>
1577 : PatFrag<(ops node:$lhs, node:$rhs),
1578 (A64select_cc NZCV, node:$lhs, (opnode node:$rhs), (i32 imm:$Cond))>;
1581 defm CSEL : A64I_condselSizes<0b0, 0b00, "csel", simple_select>;
1582 defm CSINC : A64I_condselSizes<0b0, 0b01, "csinc",
1583 complex_select<PatFrag<(ops node:$val),
1584 (add node:$val, 1)>>>;
1585 defm CSINV : A64I_condselSizes<0b1, 0b00, "csinv", complex_select<not>>;
1586 defm CSNEG : A64I_condselSizes<0b1, 0b01, "csneg", complex_select<ineg>>;
1588 // Now the instruction aliases, which fit nicely into LLVM's model:
1590 def : InstAlias<"cset $Rd, $Cond",
1591 (CSINCwwwc GPR32:$Rd, WZR, WZR, inv_cond_code_op:$Cond)>;
1592 def : InstAlias<"cset $Rd, $Cond",
1593 (CSINCxxxc GPR64:$Rd, XZR, XZR, inv_cond_code_op:$Cond)>;
1594 def : InstAlias<"csetm $Rd, $Cond",
1595 (CSINVwwwc GPR32:$Rd, WZR, WZR, inv_cond_code_op:$Cond)>;
1596 def : InstAlias<"csetm $Rd, $Cond",
1597 (CSINVxxxc GPR64:$Rd, XZR, XZR, inv_cond_code_op:$Cond)>;
1598 def : InstAlias<"cinc $Rd, $Rn, $Cond",
1599 (CSINCwwwc GPR32:$Rd, GPR32:$Rn, GPR32:$Rn, inv_cond_code_op:$Cond)>;
1600 def : InstAlias<"cinc $Rd, $Rn, $Cond",
1601 (CSINCxxxc GPR64:$Rd, GPR64:$Rn, GPR64:$Rn, inv_cond_code_op:$Cond)>;
1602 def : InstAlias<"cinv $Rd, $Rn, $Cond",
1603 (CSINVwwwc GPR32:$Rd, GPR32:$Rn, GPR32:$Rn, inv_cond_code_op:$Cond)>;
1604 def : InstAlias<"cinv $Rd, $Rn, $Cond",
1605 (CSINVxxxc GPR64:$Rd, GPR64:$Rn, GPR64:$Rn, inv_cond_code_op:$Cond)>;
1606 def : InstAlias<"cneg $Rd, $Rn, $Cond",
1607 (CSNEGwwwc GPR32:$Rd, GPR32:$Rn, GPR32:$Rn, inv_cond_code_op:$Cond)>;
1608 def : InstAlias<"cneg $Rd, $Rn, $Cond",
1609 (CSNEGxxxc GPR64:$Rd, GPR64:$Rn, GPR64:$Rn, inv_cond_code_op:$Cond)>;
1611 // Finally some helper patterns.
1613 // For CSET (a.k.a. zero-extension of icmp)
1614 def : Pat<(A64select_cc NZCV, 0, 1, cond_code:$Cond),
1615 (CSINCwwwc WZR, WZR, cond_code:$Cond)>;
1616 def : Pat<(A64select_cc NZCV, 1, 0, inv_cond_code:$Cond),
1617 (CSINCwwwc WZR, WZR, inv_cond_code:$Cond)>;
1619 def : Pat<(A64select_cc NZCV, 0, 1, cond_code:$Cond),
1620 (CSINCxxxc XZR, XZR, cond_code:$Cond)>;
1621 def : Pat<(A64select_cc NZCV, 1, 0, inv_cond_code:$Cond),
1622 (CSINCxxxc XZR, XZR, inv_cond_code:$Cond)>;
1624 // For CSETM (a.k.a. sign-extension of icmp)
1625 def : Pat<(A64select_cc NZCV, 0, -1, cond_code:$Cond),
1626 (CSINVwwwc WZR, WZR, cond_code:$Cond)>;
1627 def : Pat<(A64select_cc NZCV, -1, 0, inv_cond_code:$Cond),
1628 (CSINVwwwc WZR, WZR, inv_cond_code:$Cond)>;
1630 def : Pat<(A64select_cc NZCV, 0, -1, cond_code:$Cond),
1631 (CSINVxxxc XZR, XZR, cond_code:$Cond)>;
1632 def : Pat<(A64select_cc NZCV, -1, 0, inv_cond_code:$Cond),
1633 (CSINVxxxc XZR, XZR, inv_cond_code:$Cond)>;
1635 // CINC, CINV and CNEG get dealt with automatically, which leaves the issue of
1636 // commutativity. The instructions are to complex for isCommutable to be used,
1637 // so we have to create the patterns manually:
1639 // No commutable pattern for CSEL since the commuted version is isomorphic.
1642 def :Pat<(A64select_cc NZCV, (add i32:$Rm, 1), i32:$Rn, inv_cond_code:$Cond),
1643 (CSINCwwwc $Rn, $Rm, inv_cond_code:$Cond)>;
1644 def :Pat<(A64select_cc NZCV, (add i64:$Rm, 1), i64:$Rn, inv_cond_code:$Cond),
1645 (CSINCxxxc $Rn, $Rm, inv_cond_code:$Cond)>;
1648 def :Pat<(A64select_cc NZCV, (not i32:$Rm), i32:$Rn, inv_cond_code:$Cond),
1649 (CSINVwwwc $Rn, $Rm, inv_cond_code:$Cond)>;
1650 def :Pat<(A64select_cc NZCV, (not i64:$Rm), i64:$Rn, inv_cond_code:$Cond),
1651 (CSINVxxxc $Rn, $Rm, inv_cond_code:$Cond)>;
1654 def :Pat<(A64select_cc NZCV, (ineg i32:$Rm), i32:$Rn, inv_cond_code:$Cond),
1655 (CSNEGwwwc $Rn, $Rm, inv_cond_code:$Cond)>;
1656 def :Pat<(A64select_cc NZCV, (ineg i64:$Rm), i64:$Rn, inv_cond_code:$Cond),
1657 (CSNEGxxxc $Rn, $Rm, inv_cond_code:$Cond)>;
1659 //===----------------------------------------------------------------------===//
1660 // Data Processing (1 source) instructions
1661 //===----------------------------------------------------------------------===//
1662 // Contains: RBIT, REV16, REV, REV32, CLZ, CLS.
1664 // We define an unary operator which always fails. We will use this to
1665 // define unary operators that cannot be matched.
1667 class A64I_dp_1src_impl<bit sf, bits<6> opcode, string asmop,
1668 list<dag> patterns, RegisterClass GPRrc,
1669 InstrItinClass itin>:
1674 !strconcat(asmop, "\t$Rd, $Rn"),
1680 multiclass A64I_dp_1src <bits<6> opcode, string asmop> {
1681 let hasSideEffects = 0 in {
1682 def ww : A64I_dp_1src_impl<0b0, opcode, asmop, [], GPR32, NoItinerary>;
1683 def xx : A64I_dp_1src_impl<0b1, opcode, asmop, [], GPR64, NoItinerary>;
1687 defm RBIT : A64I_dp_1src<0b000000, "rbit">;
1688 defm CLS : A64I_dp_1src<0b000101, "cls">;
1689 defm CLZ : A64I_dp_1src<0b000100, "clz">;
1691 def : Pat<(ctlz i32:$Rn), (CLZww $Rn)>;
1692 def : Pat<(ctlz i64:$Rn), (CLZxx $Rn)>;
1693 def : Pat<(ctlz_zero_undef i32:$Rn), (CLZww $Rn)>;
1694 def : Pat<(ctlz_zero_undef i64:$Rn), (CLZxx $Rn)>;
1696 def : Pat<(cttz i32:$Rn), (CLZww (RBITww $Rn))>;
1697 def : Pat<(cttz i64:$Rn), (CLZxx (RBITxx $Rn))>;
1698 def : Pat<(cttz_zero_undef i32:$Rn), (CLZww (RBITww $Rn))>;
1699 def : Pat<(cttz_zero_undef i64:$Rn), (CLZxx (RBITxx $Rn))>;
1702 def REVww : A64I_dp_1src_impl<0b0, 0b000010, "rev",
1703 [(set i32:$Rd, (bswap i32:$Rn))],
1704 GPR32, NoItinerary>;
1705 def REVxx : A64I_dp_1src_impl<0b1, 0b000011, "rev",
1706 [(set i64:$Rd, (bswap i64:$Rn))],
1707 GPR64, NoItinerary>;
1708 def REV32xx : A64I_dp_1src_impl<0b1, 0b000010, "rev32",
1709 [(set i64:$Rd, (bswap (rotr i64:$Rn, (i64 32))))],
1710 GPR64, NoItinerary>;
1711 def REV16ww : A64I_dp_1src_impl<0b0, 0b000001, "rev16",
1712 [(set i32:$Rd, (bswap (rotr i32:$Rn, (i64 16))))],
1715 def REV16xx : A64I_dp_1src_impl<0b1, 0b000001, "rev16", [], GPR64, NoItinerary>;
1717 //===----------------------------------------------------------------------===//
1718 // Data Processing (2 sources) instructions
1719 //===----------------------------------------------------------------------===//
1720 // Contains: CRC32C?[BHWX], UDIV, SDIV, LSLV, LSRV, ASRV, RORV + aliases LSL,
1724 class dp_2src_impl<bit sf, bits<6> opcode, string asmop, list<dag> patterns,
1725 RegisterClass GPRsp,
1726 InstrItinClass itin>:
1730 !strconcat(asmop, "\t$Rd, $Rn, $Rm"),
1732 (ins GPRsp:$Rn, GPRsp:$Rm),
1736 multiclass dp_2src_crc<bit c, string asmop> {
1737 def B_www : dp_2src_impl<0b0, {0, 1, 0, c, 0, 0},
1738 !strconcat(asmop, "b"), [], GPR32, NoItinerary>;
1739 def H_www : dp_2src_impl<0b0, {0, 1, 0, c, 0, 1},
1740 !strconcat(asmop, "h"), [], GPR32, NoItinerary>;
1741 def W_www : dp_2src_impl<0b0, {0, 1, 0, c, 1, 0},
1742 !strconcat(asmop, "w"), [], GPR32, NoItinerary>;
1743 def X_wwx : A64I_dp_2src<0b1, {0, 1, 0, c, 1, 1}, 0b0,
1744 !strconcat(asmop, "x\t$Rd, $Rn, $Rm"),
1745 (outs GPR32:$Rd), (ins GPR32:$Rn, GPR64:$Rm), [],
1749 multiclass dp_2src_zext <bits<6> opcode, string asmop, SDPatternOperator op> {
1750 def www : dp_2src_impl<0b0,
1754 (op i32:$Rn, (i64 (zext i32:$Rm))))],
1757 def xxx : dp_2src_impl<0b1,
1760 [(set i64:$Rd, (op i64:$Rn, i64:$Rm))],
1766 multiclass dp_2src <bits<6> opcode, string asmop, SDPatternOperator op> {
1767 def www : dp_2src_impl<0b0,
1770 [(set i32:$Rd, (op i32:$Rn, i32:$Rm))],
1773 def xxx : dp_2src_impl<0b1,
1776 [(set i64:$Rd, (op i64:$Rn, i64:$Rm))],
1781 // Here we define the data processing 2 source instructions.
1782 defm CRC32 : dp_2src_crc<0b0, "crc32">;
1783 defm CRC32C : dp_2src_crc<0b1, "crc32c">;
1785 defm UDIV : dp_2src<0b000010, "udiv", udiv>;
1786 defm SDIV : dp_2src<0b000011, "sdiv", sdiv>;
1788 defm LSLV : dp_2src_zext<0b001000, "lsl", shl>;
1789 defm LSRV : dp_2src_zext<0b001001, "lsr", srl>;
1790 defm ASRV : dp_2src_zext<0b001010, "asr", sra>;
1791 defm RORV : dp_2src_zext<0b001011, "ror", rotr>;
1793 // Extra patterns for an incoming 64-bit value for a 32-bit
1794 // operation. Since the LLVM operations are undefined (as in C) if the
1795 // RHS is out of range, it's perfectly permissible to discard the high
1796 // bits of the GPR64.
1797 def : Pat<(shl i32:$Rn, i64:$Rm),
1798 (LSLVwww $Rn, (EXTRACT_SUBREG $Rm, sub_32))>;
1799 def : Pat<(srl i32:$Rn, i64:$Rm),
1800 (LSRVwww $Rn, (EXTRACT_SUBREG $Rm, sub_32))>;
1801 def : Pat<(sra i32:$Rn, i64:$Rm),
1802 (ASRVwww $Rn, (EXTRACT_SUBREG $Rm, sub_32))>;
1803 def : Pat<(rotr i32:$Rn, i64:$Rm),
1804 (RORVwww $Rn, (EXTRACT_SUBREG $Rm, sub_32))>;
1806 // Here we define the aliases for the data processing 2 source instructions.
1807 def LSL_mnemonic : MnemonicAlias<"lslv", "lsl">;
1808 def LSR_mnemonic : MnemonicAlias<"lsrv", "lsr">;
1809 def ASR_menmonic : MnemonicAlias<"asrv", "asr">;
1810 def ROR_menmonic : MnemonicAlias<"rorv", "ror">;
1812 //===----------------------------------------------------------------------===//
1813 // Data Processing (3 sources) instructions
1814 //===----------------------------------------------------------------------===//
1815 // Contains: MADD, MSUB, SMADDL, SMSUBL, SMULH, UMADDL, UMSUBL, UMULH
1816 // + aliases MUL, MNEG, SMULL, SMNEGL, UMULL, UMNEGL
1818 class A64I_dp3_4operand<bit sf, bits<6> opcode, RegisterClass AccReg,
1819 ValueType AccTy, RegisterClass SrcReg,
1820 string asmop, dag pattern>
1821 : A64I_dp3<sf, opcode,
1822 (outs AccReg:$Rd), (ins SrcReg:$Rn, SrcReg:$Rm, AccReg:$Ra),
1823 !strconcat(asmop, "\t$Rd, $Rn, $Rm, $Ra"),
1824 [(set AccTy:$Rd, pattern)], NoItinerary> {
1825 RegisterClass AccGPR = AccReg;
1826 RegisterClass SrcGPR = SrcReg;
1829 def MADDwwww : A64I_dp3_4operand<0b0, 0b000000, GPR32, i32, GPR32, "madd",
1830 (add i32:$Ra, (mul i32:$Rn, i32:$Rm))>;
1831 def MADDxxxx : A64I_dp3_4operand<0b1, 0b000000, GPR64, i64, GPR64, "madd",
1832 (add i64:$Ra, (mul i64:$Rn, i64:$Rm))>;
1834 def MSUBwwww : A64I_dp3_4operand<0b0, 0b000001, GPR32, i32, GPR32, "msub",
1835 (sub i32:$Ra, (mul i32:$Rn, i32:$Rm))>;
1836 def MSUBxxxx : A64I_dp3_4operand<0b1, 0b000001, GPR64, i64, GPR64, "msub",
1837 (sub i64:$Ra, (mul i64:$Rn, i64:$Rm))>;
1839 def SMADDLxwwx : A64I_dp3_4operand<0b1, 0b000010, GPR64, i64, GPR32, "smaddl",
1840 (add i64:$Ra, (mul (i64 (sext i32:$Rn)), (sext i32:$Rm)))>;
1841 def SMSUBLxwwx : A64I_dp3_4operand<0b1, 0b000011, GPR64, i64, GPR32, "smsubl",
1842 (sub i64:$Ra, (mul (i64 (sext i32:$Rn)), (sext i32:$Rm)))>;
1844 def UMADDLxwwx : A64I_dp3_4operand<0b1, 0b001010, GPR64, i64, GPR32, "umaddl",
1845 (add i64:$Ra, (mul (i64 (zext i32:$Rn)), (zext i32:$Rm)))>;
1846 def UMSUBLxwwx : A64I_dp3_4operand<0b1, 0b001011, GPR64, i64, GPR32, "umsubl",
1847 (sub i64:$Ra, (mul (i64 (zext i32:$Rn)), (zext i32:$Rm)))>;
1849 let isCommutable = 1, PostEncoderMethod = "fixMulHigh" in {
1850 def UMULHxxx : A64I_dp3<0b1, 0b001100, (outs GPR64:$Rd),
1851 (ins GPR64:$Rn, GPR64:$Rm),
1852 "umulh\t$Rd, $Rn, $Rm",
1853 [(set i64:$Rd, (mulhu i64:$Rn, i64:$Rm))],
1856 def SMULHxxx : A64I_dp3<0b1, 0b000100, (outs GPR64:$Rd),
1857 (ins GPR64:$Rn, GPR64:$Rm),
1858 "smulh\t$Rd, $Rn, $Rm",
1859 [(set i64:$Rd, (mulhs i64:$Rn, i64:$Rm))],
1863 multiclass A64I_dp3_3operand<string asmop, A64I_dp3_4operand INST,
1864 Register ZR, dag pattern> {
1865 def : InstAlias<asmop # " $Rd, $Rn, $Rm",
1866 (INST INST.AccGPR:$Rd, INST.SrcGPR:$Rn, INST.SrcGPR:$Rm, ZR)>;
1868 def : Pat<pattern, (INST $Rn, $Rm, ZR)>;
1871 defm : A64I_dp3_3operand<"mul", MADDwwww, WZR, (mul i32:$Rn, i32:$Rm)>;
1872 defm : A64I_dp3_3operand<"mul", MADDxxxx, XZR, (mul i64:$Rn, i64:$Rm)>;
1874 defm : A64I_dp3_3operand<"mneg", MSUBwwww, WZR,
1875 (sub 0, (mul i32:$Rn, i32:$Rm))>;
1876 defm : A64I_dp3_3operand<"mneg", MSUBxxxx, XZR,
1877 (sub 0, (mul i64:$Rn, i64:$Rm))>;
1879 defm : A64I_dp3_3operand<"smull", SMADDLxwwx, XZR,
1880 (mul (i64 (sext i32:$Rn)), (sext i32:$Rm))>;
1881 defm : A64I_dp3_3operand<"smnegl", SMSUBLxwwx, XZR,
1882 (sub 0, (mul (i64 (sext i32:$Rn)), (sext i32:$Rm)))>;
1884 defm : A64I_dp3_3operand<"umull", UMADDLxwwx, XZR,
1885 (mul (i64 (zext i32:$Rn)), (zext i32:$Rm))>;
1886 defm : A64I_dp3_3operand<"umnegl", UMSUBLxwwx, XZR,
1887 (sub 0, (mul (i64 (zext i32:$Rn)), (zext i32:$Rm)))>;
1890 //===----------------------------------------------------------------------===//
1891 // Exception generation
1892 //===----------------------------------------------------------------------===//
1893 // Contains: SVC, HVC, SMC, BRK, HLT, DCPS1, DCPS2, DCPS3
1895 def uimm16_asmoperand : AsmOperandClass {
1896 let Name = "UImm16";
1897 let PredicateMethod = "isUImm<16>";
1898 let RenderMethod = "addImmOperands";
1899 let DiagnosticType = "UImm16";
1902 def uimm16 : Operand<i32> {
1903 let ParserMatchClass = uimm16_asmoperand;
1906 class A64I_exceptImpl<bits<3> opc, bits<2> ll, string asmop>
1907 : A64I_exception<opc, 0b000, ll, (outs), (ins uimm16:$UImm16),
1908 !strconcat(asmop, "\t$UImm16"), [], NoItinerary> {
1910 let isTerminator = 1;
1913 def SVCi : A64I_exceptImpl<0b000, 0b01, "svc">;
1914 def HVCi : A64I_exceptImpl<0b000, 0b10, "hvc">;
1915 def SMCi : A64I_exceptImpl<0b000, 0b11, "smc">;
1916 def BRKi : A64I_exceptImpl<0b001, 0b00, "brk">;
1917 def HLTi : A64I_exceptImpl<0b010, 0b00, "hlt">;
1919 def DCPS1i : A64I_exceptImpl<0b101, 0b01, "dcps1">;
1920 def DCPS2i : A64I_exceptImpl<0b101, 0b10, "dcps2">;
1921 def DCPS3i : A64I_exceptImpl<0b101, 0b11, "dcps3">;
1923 // The immediate is optional for the DCPS instructions, defaulting to 0.
1924 def : InstAlias<"dcps1", (DCPS1i 0)>;
1925 def : InstAlias<"dcps2", (DCPS2i 0)>;
1926 def : InstAlias<"dcps3", (DCPS3i 0)>;
1928 //===----------------------------------------------------------------------===//
1929 // Extract (immediate)
1930 //===----------------------------------------------------------------------===//
1931 // Contains: EXTR + alias ROR
1933 def EXTRwwwi : A64I_extract<0b0, 0b000, 0b0,
1935 (ins GPR32:$Rn, GPR32:$Rm, bitfield32_imm:$LSB),
1936 "extr\t$Rd, $Rn, $Rm, $LSB",
1938 (A64Extr i32:$Rn, i32:$Rm, imm:$LSB))],
1940 def EXTRxxxi : A64I_extract<0b1, 0b000, 0b1,
1942 (ins GPR64:$Rn, GPR64:$Rm, bitfield64_imm:$LSB),
1943 "extr\t$Rd, $Rn, $Rm, $LSB",
1945 (A64Extr i64:$Rn, i64:$Rm, imm:$LSB))],
1948 def : InstAlias<"ror $Rd, $Rs, $LSB",
1949 (EXTRwwwi GPR32:$Rd, GPR32:$Rs, GPR32:$Rs, bitfield32_imm:$LSB)>;
1950 def : InstAlias<"ror $Rd, $Rs, $LSB",
1951 (EXTRxxxi GPR64:$Rd, GPR64:$Rs, GPR64:$Rs, bitfield64_imm:$LSB)>;
1953 def : Pat<(rotr i32:$Rn, bitfield32_imm:$LSB),
1954 (EXTRwwwi $Rn, $Rn, bitfield32_imm:$LSB)>;
1955 def : Pat<(rotr i64:$Rn, bitfield64_imm:$LSB),
1956 (EXTRxxxi $Rn, $Rn, bitfield64_imm:$LSB)>;
1958 //===----------------------------------------------------------------------===//
1959 // Floating-point compare instructions
1960 //===----------------------------------------------------------------------===//
1961 // Contains: FCMP, FCMPE
1963 def fpzero_asmoperand : AsmOperandClass {
1964 let Name = "FPZero";
1965 let ParserMethod = "ParseFPImmOperand";
1966 let DiagnosticType = "FPZero";
1969 def fpz32 : Operand<f32>,
1970 ComplexPattern<f32, 1, "SelectFPZeroOperand", [fpimm]> {
1971 let ParserMatchClass = fpzero_asmoperand;
1972 let PrintMethod = "printFPZeroOperand";
1973 let DecoderMethod = "DecodeFPZeroOperand";
1976 def fpz64 : Operand<f64>,
1977 ComplexPattern<f64, 1, "SelectFPZeroOperand", [fpimm]> {
1978 let ParserMatchClass = fpzero_asmoperand;
1979 let PrintMethod = "printFPZeroOperand";
1980 let DecoderMethod = "DecodeFPZeroOperand";
1983 def fpz64movi : Operand<i64>,
1984 ComplexPattern<f64, 1, "SelectFPZeroOperand", [fpimm]> {
1985 let ParserMatchClass = fpzero_asmoperand;
1986 let PrintMethod = "printFPZeroOperand";
1987 let DecoderMethod = "DecodeFPZeroOperand";
1990 multiclass A64I_fpcmpSignal<bits<2> type, bit imm, dag ins, dag pattern> {
1991 def _quiet : A64I_fpcmp<0b0, 0b0, type, 0b00, {0b0, imm, 0b0, 0b0, 0b0},
1992 (outs), ins, "fcmp\t$Rn, $Rm", [pattern],
1997 def _sig : A64I_fpcmp<0b0, 0b0, type, 0b00, {0b1, imm, 0b0, 0b0, 0b0},
1998 (outs), ins, "fcmpe\t$Rn, $Rm", [], NoItinerary> {
2003 defm FCMPss : A64I_fpcmpSignal<0b00, 0b0, (ins FPR32:$Rn, FPR32:$Rm),
2004 (set NZCV, (A64cmp f32:$Rn, f32:$Rm))>;
2005 defm FCMPdd : A64I_fpcmpSignal<0b01, 0b0, (ins FPR64:$Rn, FPR64:$Rm),
2006 (set NZCV, (A64cmp f64:$Rn, f64:$Rm))>;
2008 // What would be Rm should be written as 0; note that even though it's called
2009 // "$Rm" here to fit in with the InstrFormats, it's actually an immediate.
2010 defm FCMPsi : A64I_fpcmpSignal<0b00, 0b1, (ins FPR32:$Rn, fpz32:$Rm),
2011 (set NZCV, (A64cmp f32:$Rn, fpz32:$Rm))>;
2013 defm FCMPdi : A64I_fpcmpSignal<0b01, 0b1, (ins FPR64:$Rn, fpz64:$Rm),
2014 (set NZCV, (A64cmp f64:$Rn, fpz64:$Rm))>;
2017 //===----------------------------------------------------------------------===//
2018 // Floating-point conditional compare instructions
2019 //===----------------------------------------------------------------------===//
2020 // Contains: FCCMP, FCCMPE
2022 class A64I_fpccmpImpl<bits<2> type, bit op, RegisterClass FPR, string asmop>
2023 : A64I_fpccmp<0b0, 0b0, type, op,
2025 (ins FPR:$Rn, FPR:$Rm, uimm4:$NZCVImm, cond_code_op:$Cond),
2026 !strconcat(asmop, "\t$Rn, $Rm, $NZCVImm, $Cond"),
2031 def FCCMPss : A64I_fpccmpImpl<0b00, 0b0, FPR32, "fccmp">;
2032 def FCCMPEss : A64I_fpccmpImpl<0b00, 0b1, FPR32, "fccmpe">;
2033 def FCCMPdd : A64I_fpccmpImpl<0b01, 0b0, FPR64, "fccmp">;
2034 def FCCMPEdd : A64I_fpccmpImpl<0b01, 0b1, FPR64, "fccmpe">;
2036 //===----------------------------------------------------------------------===//
2037 // Floating-point conditional select instructions
2038 //===----------------------------------------------------------------------===//
2041 let Uses = [NZCV] in {
2042 def FCSELsssc : A64I_fpcondsel<0b0, 0b0, 0b00, (outs FPR32:$Rd),
2043 (ins FPR32:$Rn, FPR32:$Rm, cond_code_op:$Cond),
2044 "fcsel\t$Rd, $Rn, $Rm, $Cond",
2046 (simple_select f32:$Rn, f32:$Rm))],
2050 def FCSELdddc : A64I_fpcondsel<0b0, 0b0, 0b01, (outs FPR64:$Rd),
2051 (ins FPR64:$Rn, FPR64:$Rm, cond_code_op:$Cond),
2052 "fcsel\t$Rd, $Rn, $Rm, $Cond",
2054 (simple_select f64:$Rn, f64:$Rm))],
2058 //===----------------------------------------------------------------------===//
2059 // Floating-point data-processing (1 source)
2060 //===----------------------------------------------------------------------===//
2061 // Contains: FMOV, FABS, FNEG, FSQRT, FCVT, FRINT[NPMZAXI].
2063 def FPNoUnop : PatFrag<(ops node:$val), (fneg node:$val),
2064 [{ (void)N; return false; }]>;
2066 // First we do the fairly trivial bunch with uniform "OP s, s" and "OP d, d"
2067 // syntax. Default to no pattern because most are odd enough not to have one.
2068 multiclass A64I_fpdp1sizes<bits<6> opcode, string asmstr,
2069 SDPatternOperator opnode = FPNoUnop> {
2070 def ss : A64I_fpdp1<0b0, 0b0, 0b00, opcode, (outs FPR32:$Rd), (ins FPR32:$Rn),
2071 !strconcat(asmstr, "\t$Rd, $Rn"),
2072 [(set f32:$Rd, (opnode f32:$Rn))],
2075 def dd : A64I_fpdp1<0b0, 0b0, 0b01, opcode, (outs FPR64:$Rd), (ins FPR64:$Rn),
2076 !strconcat(asmstr, "\t$Rd, $Rn"),
2077 [(set f64:$Rd, (opnode f64:$Rn))],
2081 defm FMOV : A64I_fpdp1sizes<0b000000, "fmov">;
2082 defm FABS : A64I_fpdp1sizes<0b000001, "fabs", fabs>;
2083 defm FNEG : A64I_fpdp1sizes<0b000010, "fneg", fneg>;
2084 defm FSQRT : A64I_fpdp1sizes<0b000011, "fsqrt", fsqrt>;
2086 defm FRINTN : A64I_fpdp1sizes<0b001000, "frintn">;
2087 defm FRINTP : A64I_fpdp1sizes<0b001001, "frintp", fceil>;
2088 defm FRINTM : A64I_fpdp1sizes<0b001010, "frintm", ffloor>;
2089 defm FRINTZ : A64I_fpdp1sizes<0b001011, "frintz", ftrunc>;
2090 defm FRINTA : A64I_fpdp1sizes<0b001100, "frinta">;
2091 defm FRINTX : A64I_fpdp1sizes<0b001110, "frintx", frint>;
2092 defm FRINTI : A64I_fpdp1sizes<0b001111, "frinti", fnearbyint>;
2094 // The FCVT instrucitons have different source and destination register-types,
2095 // but the fields are uniform everywhere a D-register (say) crops up. Package
2096 // this information in a Record.
2097 class FCVTRegType<RegisterClass rc, bits<2> fld, ValueType vt> {
2098 RegisterClass Class = rc;
2104 def FCVT16 : FCVTRegType<FPR16, 0b11, f16>;
2105 def FCVT32 : FCVTRegType<FPR32, 0b00, f32>;
2106 def FCVT64 : FCVTRegType<FPR64, 0b01, f64>;
2108 class A64I_fpdp1_fcvt<FCVTRegType DestReg, FCVTRegType SrcReg, SDNode opnode>
2109 : A64I_fpdp1<0b0, 0b0, {SrcReg.t1, SrcReg.t0},
2110 {0,0,0,1, DestReg.t1, DestReg.t0},
2111 (outs DestReg.Class:$Rd), (ins SrcReg.Class:$Rn),
2113 [(set DestReg.VT:$Rd, (opnode SrcReg.VT:$Rn))], NoItinerary>;
2115 def FCVTds : A64I_fpdp1_fcvt<FCVT64, FCVT32, fextend>;
2116 def FCVThs : A64I_fpdp1_fcvt<FCVT16, FCVT32, fround>;
2117 def FCVTsd : A64I_fpdp1_fcvt<FCVT32, FCVT64, fround>;
2118 def FCVThd : A64I_fpdp1_fcvt<FCVT16, FCVT64, fround>;
2119 def FCVTsh : A64I_fpdp1_fcvt<FCVT32, FCVT16, fextend>;
2120 def FCVTdh : A64I_fpdp1_fcvt<FCVT64, FCVT16, fextend>;
2123 //===----------------------------------------------------------------------===//
2124 // Floating-point data-processing (2 sources) instructions
2125 //===----------------------------------------------------------------------===//
2126 // Contains: FMUL, FDIV, FADD, FSUB, FMAX, FMIN, FMAXNM, FMINNM, FNMUL
2128 def FPNoBinop : PatFrag<(ops node:$lhs, node:$rhs), (fadd node:$lhs, node:$rhs),
2129 [{ (void)N; return false; }]>;
2131 multiclass A64I_fpdp2sizes<bits<4> opcode, string asmstr,
2132 SDPatternOperator opnode> {
2133 def sss : A64I_fpdp2<0b0, 0b0, 0b00, opcode,
2135 (ins FPR32:$Rn, FPR32:$Rm),
2136 !strconcat(asmstr, "\t$Rd, $Rn, $Rm"),
2137 [(set f32:$Rd, (opnode f32:$Rn, f32:$Rm))],
2140 def ddd : A64I_fpdp2<0b0, 0b0, 0b01, opcode,
2142 (ins FPR64:$Rn, FPR64:$Rm),
2143 !strconcat(asmstr, "\t$Rd, $Rn, $Rm"),
2144 [(set f64:$Rd, (opnode f64:$Rn, f64:$Rm))],
2148 let isCommutable = 1 in {
2149 defm FMUL : A64I_fpdp2sizes<0b0000, "fmul", fmul>;
2150 defm FADD : A64I_fpdp2sizes<0b0010, "fadd", fadd>;
2152 // No patterns for these.
2153 defm FMAX : A64I_fpdp2sizes<0b0100, "fmax", FPNoBinop>;
2154 defm FMIN : A64I_fpdp2sizes<0b0101, "fmin", FPNoBinop>;
2155 defm FMAXNM : A64I_fpdp2sizes<0b0110, "fmaxnm", FPNoBinop>;
2156 defm FMINNM : A64I_fpdp2sizes<0b0111, "fminnm", FPNoBinop>;
2158 defm FNMUL : A64I_fpdp2sizes<0b1000, "fnmul",
2159 PatFrag<(ops node:$lhs, node:$rhs),
2160 (fneg (fmul node:$lhs, node:$rhs))> >;
2163 defm FDIV : A64I_fpdp2sizes<0b0001, "fdiv", fdiv>;
2164 defm FSUB : A64I_fpdp2sizes<0b0011, "fsub", fsub>;
2166 //===----------------------------------------------------------------------===//
2167 // Floating-point data-processing (3 sources) instructions
2168 //===----------------------------------------------------------------------===//
2169 // Contains: FMADD, FMSUB, FNMADD, FNMSUB
2171 def fmsub : PatFrag<(ops node:$Rn, node:$Rm, node:$Ra),
2172 (fma (fneg node:$Rn), node:$Rm, node:$Ra)>;
2173 def fnmadd : PatFrag<(ops node:$Rn, node:$Rm, node:$Ra),
2174 (fma node:$Rn, node:$Rm, (fneg node:$Ra))>;
2175 def fnmsub : PatFrag<(ops node:$Rn, node:$Rm, node:$Ra),
2176 (fma (fneg node:$Rn), node:$Rm, (fneg node:$Ra))>;
2178 class A64I_fpdp3Impl<string asmop, RegisterClass FPR, ValueType VT,
2179 bits<2> type, bit o1, bit o0, SDPatternOperator fmakind>
2180 : A64I_fpdp3<0b0, 0b0, type, o1, o0, (outs FPR:$Rd),
2181 (ins FPR:$Rn, FPR:$Rm, FPR:$Ra),
2182 !strconcat(asmop,"\t$Rd, $Rn, $Rm, $Ra"),
2183 [(set VT:$Rd, (fmakind VT:$Rn, VT:$Rm, VT:$Ra))],
2186 def FMADDssss : A64I_fpdp3Impl<"fmadd", FPR32, f32, 0b00, 0b0, 0b0, fma>;
2187 def FMSUBssss : A64I_fpdp3Impl<"fmsub", FPR32, f32, 0b00, 0b0, 0b1, fmsub>;
2188 def FNMADDssss : A64I_fpdp3Impl<"fnmadd", FPR32, f32, 0b00, 0b1, 0b0, fnmadd>;
2189 def FNMSUBssss : A64I_fpdp3Impl<"fnmsub", FPR32, f32, 0b00, 0b1, 0b1, fnmsub>;
2191 def FMADDdddd : A64I_fpdp3Impl<"fmadd", FPR64, f64, 0b01, 0b0, 0b0, fma>;
2192 def FMSUBdddd : A64I_fpdp3Impl<"fmsub", FPR64, f64, 0b01, 0b0, 0b1, fmsub>;
2193 def FNMADDdddd : A64I_fpdp3Impl<"fnmadd", FPR64, f64, 0b01, 0b1, 0b0, fnmadd>;
2194 def FNMSUBdddd : A64I_fpdp3Impl<"fnmsub", FPR64, f64, 0b01, 0b1, 0b1, fnmsub>;
2196 // Extra patterns for when we're allowed to optimise separate multiplication and
2198 let Predicates = [UseFusedMAC] in {
2199 def : Pat<(f32 (fadd FPR32:$Ra, (f32 (fmul FPR32:$Rn, FPR32:$Rm)))),
2200 (FMADDssss FPR32:$Rn, FPR32:$Rm, FPR32:$Ra)>;
2201 def : Pat<(f32 (fsub FPR32:$Ra, (f32 (fmul FPR32:$Rn, FPR32:$Rm)))),
2202 (FMSUBssss FPR32:$Rn, FPR32:$Rm, FPR32:$Ra)>;
2203 def : Pat<(f32 (fsub (f32 (fmul FPR32:$Rn, FPR32:$Rm)), FPR32:$Ra)),
2204 (FNMADDssss FPR32:$Rn, FPR32:$Rm, FPR32:$Ra)>;
2205 def : Pat<(f32 (fsub (f32 (fneg FPR32:$Ra)), (f32 (fmul FPR32:$Rn, FPR32:$Rm)))),
2206 (FNMSUBssss FPR32:$Rn, FPR32:$Rm, FPR32:$Ra)>;
2208 def : Pat<(f64 (fadd FPR64:$Ra, (f64 (fmul FPR64:$Rn, FPR64:$Rm)))),
2209 (FMADDdddd FPR64:$Rn, FPR64:$Rm, FPR64:$Ra)>;
2210 def : Pat<(f64 (fsub FPR64:$Ra, (f64 (fmul FPR64:$Rn, FPR64:$Rm)))),
2211 (FMSUBdddd FPR64:$Rn, FPR64:$Rm, FPR64:$Ra)>;
2212 def : Pat<(f64 (fsub (f64 (fmul FPR64:$Rn, FPR64:$Rm)), FPR64:$Ra)),
2213 (FNMADDdddd FPR64:$Rn, FPR64:$Rm, FPR64:$Ra)>;
2214 def : Pat<(f64 (fsub (f64 (fneg FPR64:$Ra)), (f64 (fmul FPR64:$Rn, FPR64:$Rm)))),
2215 (FNMSUBdddd FPR64:$Rn, FPR64:$Rm, FPR64:$Ra)>;
2219 //===----------------------------------------------------------------------===//
2220 // Floating-point <-> fixed-point conversion instructions
2221 //===----------------------------------------------------------------------===//
2222 // Contains: FCVTZS, FCVTZU, SCVTF, UCVTF
2224 // #1-#32 allowed, encoded as "64 - <specified imm>
2225 def fixedpos_asmoperand_i32 : AsmOperandClass {
2226 let Name = "CVTFixedPos32";
2227 let RenderMethod = "addCVTFixedPosOperands";
2228 let PredicateMethod = "isCVTFixedPos<32>";
2229 let DiagnosticType = "CVTFixedPos32";
2232 // Also encoded as "64 - <specified imm>" but #1-#64 allowed.
2233 def fixedpos_asmoperand_i64 : AsmOperandClass {
2234 let Name = "CVTFixedPos64";
2235 let RenderMethod = "addCVTFixedPosOperands";
2236 let PredicateMethod = "isCVTFixedPos<64>";
2237 let DiagnosticType = "CVTFixedPos64";
2240 // We need the cartesian product of f32/f64 i32/i64 operands for
2242 // + Selection needs to use operands of correct floating type
2243 // + Assembly parsing and decoding depend on integer width
2244 class cvtfix_i32_op<ValueType FloatVT>
2246 ComplexPattern<FloatVT, 1, "SelectCVTFixedPosOperand<32>", [fpimm]> {
2247 let ParserMatchClass = fixedpos_asmoperand_i32;
2248 let DecoderMethod = "DecodeCVT32FixedPosOperand";
2249 let PrintMethod = "printCVTFixedPosOperand";
2252 class cvtfix_i64_op<ValueType FloatVT>
2254 ComplexPattern<FloatVT, 1, "SelectCVTFixedPosOperand<64>", [fpimm]> {
2255 let ParserMatchClass = fixedpos_asmoperand_i64;
2256 let PrintMethod = "printCVTFixedPosOperand";
2259 // Because of the proliferation of weird operands, it's not really
2260 // worth going for a multiclass here. Oh well.
2262 class A64I_fptofix<bit sf, bits<2> type, bits<3> opcode,
2263 RegisterClass GPR, RegisterClass FPR,
2264 ValueType DstTy, ValueType SrcTy,
2265 Operand scale_op, string asmop, SDNode cvtop>
2266 : A64I_fpfixed<sf, 0b0, type, 0b11, opcode,
2267 (outs GPR:$Rd), (ins FPR:$Rn, scale_op:$Scale),
2268 !strconcat(asmop, "\t$Rd, $Rn, $Scale"),
2269 [(set DstTy:$Rd, (cvtop (fmul SrcTy:$Rn, scale_op:$Scale)))],
2272 def FCVTZSwsi : A64I_fptofix<0b0, 0b00, 0b000, GPR32, FPR32, i32, f32,
2273 cvtfix_i32_op<f32>, "fcvtzs", fp_to_sint>;
2274 def FCVTZSxsi : A64I_fptofix<0b1, 0b00, 0b000, GPR64, FPR32, i64, f32,
2275 cvtfix_i64_op<f32>, "fcvtzs", fp_to_sint>;
2276 def FCVTZUwsi : A64I_fptofix<0b0, 0b00, 0b001, GPR32, FPR32, i32, f32,
2277 cvtfix_i32_op<f32>, "fcvtzu", fp_to_uint>;
2278 def FCVTZUxsi : A64I_fptofix<0b1, 0b00, 0b001, GPR64, FPR32, i64, f32,
2279 cvtfix_i64_op<f32>, "fcvtzu", fp_to_uint>;
2281 def FCVTZSwdi : A64I_fptofix<0b0, 0b01, 0b000, GPR32, FPR64, i32, f64,
2282 cvtfix_i32_op<f64>, "fcvtzs", fp_to_sint>;
2283 def FCVTZSxdi : A64I_fptofix<0b1, 0b01, 0b000, GPR64, FPR64, i64, f64,
2284 cvtfix_i64_op<f64>, "fcvtzs", fp_to_sint>;
2285 def FCVTZUwdi : A64I_fptofix<0b0, 0b01, 0b001, GPR32, FPR64, i32, f64,
2286 cvtfix_i32_op<f64>, "fcvtzu", fp_to_uint>;
2287 def FCVTZUxdi : A64I_fptofix<0b1, 0b01, 0b001, GPR64, FPR64, i64, f64,
2288 cvtfix_i64_op<f64>, "fcvtzu", fp_to_uint>;
2291 class A64I_fixtofp<bit sf, bits<2> type, bits<3> opcode,
2292 RegisterClass FPR, RegisterClass GPR,
2293 ValueType DstTy, ValueType SrcTy,
2294 Operand scale_op, string asmop, SDNode cvtop>
2295 : A64I_fpfixed<sf, 0b0, type, 0b00, opcode,
2296 (outs FPR:$Rd), (ins GPR:$Rn, scale_op:$Scale),
2297 !strconcat(asmop, "\t$Rd, $Rn, $Scale"),
2298 [(set DstTy:$Rd, (fdiv (cvtop SrcTy:$Rn), scale_op:$Scale))],
2301 def SCVTFswi : A64I_fixtofp<0b0, 0b00, 0b010, FPR32, GPR32, f32, i32,
2302 cvtfix_i32_op<f32>, "scvtf", sint_to_fp>;
2303 def SCVTFsxi : A64I_fixtofp<0b1, 0b00, 0b010, FPR32, GPR64, f32, i64,
2304 cvtfix_i64_op<f32>, "scvtf", sint_to_fp>;
2305 def UCVTFswi : A64I_fixtofp<0b0, 0b00, 0b011, FPR32, GPR32, f32, i32,
2306 cvtfix_i32_op<f32>, "ucvtf", uint_to_fp>;
2307 def UCVTFsxi : A64I_fixtofp<0b1, 0b00, 0b011, FPR32, GPR64, f32, i64,
2308 cvtfix_i64_op<f32>, "ucvtf", uint_to_fp>;
2309 def SCVTFdwi : A64I_fixtofp<0b0, 0b01, 0b010, FPR64, GPR32, f64, i32,
2310 cvtfix_i32_op<f64>, "scvtf", sint_to_fp>;
2311 def SCVTFdxi : A64I_fixtofp<0b1, 0b01, 0b010, FPR64, GPR64, f64, i64,
2312 cvtfix_i64_op<f64>, "scvtf", sint_to_fp>;
2313 def UCVTFdwi : A64I_fixtofp<0b0, 0b01, 0b011, FPR64, GPR32, f64, i32,
2314 cvtfix_i32_op<f64>, "ucvtf", uint_to_fp>;
2315 def UCVTFdxi : A64I_fixtofp<0b1, 0b01, 0b011, FPR64, GPR64, f64, i64,
2316 cvtfix_i64_op<f64>, "ucvtf", uint_to_fp>;
2318 //===----------------------------------------------------------------------===//
2319 // Floating-point <-> integer conversion instructions
2320 //===----------------------------------------------------------------------===//
2321 // Contains: FCVTZS, FCVTZU, SCVTF, UCVTF
2323 class A64I_fpintI<bit sf, bits<2> type, bits<2> rmode, bits<3> opcode,
2324 RegisterClass DestPR, RegisterClass SrcPR, string asmop>
2325 : A64I_fpint<sf, 0b0, type, rmode, opcode, (outs DestPR:$Rd), (ins SrcPR:$Rn),
2326 !strconcat(asmop, "\t$Rd, $Rn"), [], NoItinerary>;
2328 multiclass A64I_fptointRM<bits<2> rmode, bit o2, string asmop> {
2329 def Sws : A64I_fpintI<0b0, 0b00, rmode, {o2, 0, 0},
2330 GPR32, FPR32, asmop # "s">;
2331 def Sxs : A64I_fpintI<0b1, 0b00, rmode, {o2, 0, 0},
2332 GPR64, FPR32, asmop # "s">;
2333 def Uws : A64I_fpintI<0b0, 0b00, rmode, {o2, 0, 1},
2334 GPR32, FPR32, asmop # "u">;
2335 def Uxs : A64I_fpintI<0b1, 0b00, rmode, {o2, 0, 1},
2336 GPR64, FPR32, asmop # "u">;
2338 def Swd : A64I_fpintI<0b0, 0b01, rmode, {o2, 0, 0},
2339 GPR32, FPR64, asmop # "s">;
2340 def Sxd : A64I_fpintI<0b1, 0b01, rmode, {o2, 0, 0},
2341 GPR64, FPR64, asmop # "s">;
2342 def Uwd : A64I_fpintI<0b0, 0b01, rmode, {o2, 0, 1},
2343 GPR32, FPR64, asmop # "u">;
2344 def Uxd : A64I_fpintI<0b1, 0b01, rmode, {o2, 0, 1},
2345 GPR64, FPR64, asmop # "u">;
2348 defm FCVTN : A64I_fptointRM<0b00, 0b0, "fcvtn">;
2349 defm FCVTP : A64I_fptointRM<0b01, 0b0, "fcvtp">;
2350 defm FCVTM : A64I_fptointRM<0b10, 0b0, "fcvtm">;
2351 defm FCVTZ : A64I_fptointRM<0b11, 0b0, "fcvtz">;
2352 defm FCVTA : A64I_fptointRM<0b00, 0b1, "fcvta">;
2354 def : Pat<(i32 (fp_to_sint f32:$Rn)), (FCVTZSws $Rn)>;
2355 def : Pat<(i64 (fp_to_sint f32:$Rn)), (FCVTZSxs $Rn)>;
2356 def : Pat<(i32 (fp_to_uint f32:$Rn)), (FCVTZUws $Rn)>;
2357 def : Pat<(i64 (fp_to_uint f32:$Rn)), (FCVTZUxs $Rn)>;
2358 def : Pat<(i32 (fp_to_sint f64:$Rn)), (FCVTZSwd $Rn)>;
2359 def : Pat<(i64 (fp_to_sint f64:$Rn)), (FCVTZSxd $Rn)>;
2360 def : Pat<(i32 (fp_to_uint f64:$Rn)), (FCVTZUwd $Rn)>;
2361 def : Pat<(i64 (fp_to_uint f64:$Rn)), (FCVTZUxd $Rn)>;
2363 multiclass A64I_inttofp<bit o0, string asmop> {
2364 def CVTFsw : A64I_fpintI<0b0, 0b00, 0b00, {0, 1, o0}, FPR32, GPR32, asmop>;
2365 def CVTFsx : A64I_fpintI<0b1, 0b00, 0b00, {0, 1, o0}, FPR32, GPR64, asmop>;
2366 def CVTFdw : A64I_fpintI<0b0, 0b01, 0b00, {0, 1, o0}, FPR64, GPR32, asmop>;
2367 def CVTFdx : A64I_fpintI<0b1, 0b01, 0b00, {0, 1, o0}, FPR64, GPR64, asmop>;
2370 defm S : A64I_inttofp<0b0, "scvtf">;
2371 defm U : A64I_inttofp<0b1, "ucvtf">;
2373 def : Pat<(f32 (sint_to_fp i32:$Rn)), (SCVTFsw $Rn)>;
2374 def : Pat<(f32 (sint_to_fp i64:$Rn)), (SCVTFsx $Rn)>;
2375 def : Pat<(f64 (sint_to_fp i32:$Rn)), (SCVTFdw $Rn)>;
2376 def : Pat<(f64 (sint_to_fp i64:$Rn)), (SCVTFdx $Rn)>;
2377 def : Pat<(f32 (uint_to_fp i32:$Rn)), (UCVTFsw $Rn)>;
2378 def : Pat<(f32 (uint_to_fp i64:$Rn)), (UCVTFsx $Rn)>;
2379 def : Pat<(f64 (uint_to_fp i32:$Rn)), (UCVTFdw $Rn)>;
2380 def : Pat<(f64 (uint_to_fp i64:$Rn)), (UCVTFdx $Rn)>;
2382 def FMOVws : A64I_fpintI<0b0, 0b00, 0b00, 0b110, GPR32, FPR32, "fmov">;
2383 def FMOVsw : A64I_fpintI<0b0, 0b00, 0b00, 0b111, FPR32, GPR32, "fmov">;
2384 def FMOVxd : A64I_fpintI<0b1, 0b01, 0b00, 0b110, GPR64, FPR64, "fmov">;
2385 def FMOVdx : A64I_fpintI<0b1, 0b01, 0b00, 0b111, FPR64, GPR64, "fmov">;
2387 def : Pat<(i32 (bitconvert f32:$Rn)), (FMOVws $Rn)>;
2388 def : Pat<(f32 (bitconvert i32:$Rn)), (FMOVsw $Rn)>;
2389 def : Pat<(i64 (bitconvert f64:$Rn)), (FMOVxd $Rn)>;
2390 def : Pat<(f64 (bitconvert i64:$Rn)), (FMOVdx $Rn)>;
2392 def lane1_asmoperand : AsmOperandClass {
2394 let RenderMethod = "addImmOperands";
2395 let DiagnosticType = "Lane1";
2398 def lane1 : Operand<i32> {
2399 let ParserMatchClass = lane1_asmoperand;
2400 let PrintMethod = "printBareImmOperand";
2403 let DecoderMethod = "DecodeFMOVLaneInstruction" in {
2404 def FMOVxv : A64I_fpint<0b1, 0b0, 0b10, 0b01, 0b110,
2405 (outs GPR64:$Rd), (ins VPR128:$Rn, lane1:$Lane),
2406 "fmov\t$Rd, $Rn.d[$Lane]", [], NoItinerary>;
2408 def FMOVvx : A64I_fpint<0b1, 0b0, 0b10, 0b01, 0b111,
2409 (outs VPR128:$Rd), (ins GPR64:$Rn, lane1:$Lane),
2410 "fmov\t$Rd.d[$Lane], $Rn", [], NoItinerary>;
2413 def : InstAlias<"fmov $Rd, $Rn.2d[$Lane]",
2414 (FMOVxv GPR64:$Rd, VPR128:$Rn, lane1:$Lane), 0b0>;
2416 def : InstAlias<"fmov $Rd.2d[$Lane], $Rn",
2417 (FMOVvx VPR128:$Rd, GPR64:$Rn, lane1:$Lane), 0b0>;
2419 //===----------------------------------------------------------------------===//
2420 // Floating-point immediate instructions
2421 //===----------------------------------------------------------------------===//
2424 def fpimm_asmoperand : AsmOperandClass {
2425 let Name = "FMOVImm";
2426 let ParserMethod = "ParseFPImmOperand";
2427 let DiagnosticType = "FPImm";
2430 // The MCOperand for these instructions are the encoded 8-bit values.
2431 def SDXF_fpimm : SDNodeXForm<fpimm, [{
2433 A64Imms::isFPImm(N->getValueAPF(), Imm8);
2434 return CurDAG->getTargetConstant(Imm8, MVT::i32);
2437 class fmov_operand<ValueType FT>
2439 PatLeaf<(FT fpimm), [{ return A64Imms::isFPImm(N->getValueAPF()); }],
2441 let PrintMethod = "printFPImmOperand";
2442 let ParserMatchClass = fpimm_asmoperand;
2445 def fmov32_operand : fmov_operand<f32>;
2446 def fmov64_operand : fmov_operand<f64>;
2448 class A64I_fpimm_impl<bits<2> type, RegisterClass Reg, ValueType VT,
2449 Operand fmov_operand>
2450 : A64I_fpimm<0b0, 0b0, type, 0b00000,
2452 (ins fmov_operand:$Imm8),
2454 [(set VT:$Rd, fmov_operand:$Imm8)],
2457 def FMOVsi : A64I_fpimm_impl<0b00, FPR32, f32, fmov32_operand>;
2458 def FMOVdi : A64I_fpimm_impl<0b01, FPR64, f64, fmov64_operand>;
2460 //===----------------------------------------------------------------------===//
2461 // Load-register (literal) instructions
2462 //===----------------------------------------------------------------------===//
2463 // Contains: LDR, LDRSW, PRFM
2465 def ldrlit_label_asmoperand : AsmOperandClass {
2466 let Name = "LoadLitLabel";
2467 let RenderMethod = "addLabelOperands<19, 4>";
2468 let DiagnosticType = "Label";
2471 def ldrlit_label : Operand<i64> {
2472 let EncoderMethod = "getLoadLitLabelOpValue";
2474 // This label is a 19-bit offset from PC, scaled by the instruction-width: 4.
2475 let PrintMethod = "printLabelOperand<19, 4>";
2476 let ParserMatchClass = ldrlit_label_asmoperand;
2477 let OperandType = "OPERAND_PCREL";
2480 // Various instructions take an immediate value (which can always be used),
2481 // where some numbers have a symbolic name to make things easier. These operands
2482 // and the associated functions abstract away the differences.
2483 multiclass namedimm<string prefix, string mapper> {
2484 def _asmoperand : AsmOperandClass {
2485 let Name = "NamedImm" # prefix;
2486 let PredicateMethod = "isUImm";
2487 let RenderMethod = "addImmOperands";
2488 let ParserMethod = "ParseNamedImmOperand<" # mapper # ">";
2489 let DiagnosticType = "NamedImm_" # prefix;
2492 def _op : Operand<i32> {
2493 let ParserMatchClass = !cast<AsmOperandClass>(prefix # "_asmoperand");
2494 let PrintMethod = "printNamedImmOperand<" # mapper # ">";
2495 let DecoderMethod = "DecodeNamedImmOperand<" # mapper # ">";
2499 defm prefetch : namedimm<"prefetch", "A64PRFM::PRFMMapper">;
2501 class A64I_LDRlitSimple<bits<2> opc, bit v, RegisterClass OutReg,
2502 list<dag> patterns = []>
2503 : A64I_LDRlit<opc, v, (outs OutReg:$Rt), (ins ldrlit_label:$Imm19),
2504 "ldr\t$Rt, $Imm19", patterns, NoItinerary>;
2506 let mayLoad = 1 in {
2507 def LDRw_lit : A64I_LDRlitSimple<0b00, 0b0, GPR32>;
2508 def LDRx_lit : A64I_LDRlitSimple<0b01, 0b0, GPR64>;
2511 def LDRs_lit : A64I_LDRlitSimple<0b00, 0b1, FPR32>;
2512 def LDRd_lit : A64I_LDRlitSimple<0b01, 0b1, FPR64>;
2514 let mayLoad = 1 in {
2515 def LDRq_lit : A64I_LDRlitSimple<0b10, 0b1, FPR128>;
2518 def LDRSWx_lit : A64I_LDRlit<0b10, 0b0,
2520 (ins ldrlit_label:$Imm19),
2521 "ldrsw\t$Rt, $Imm19",
2524 def PRFM_lit : A64I_LDRlit<0b11, 0b0,
2525 (outs), (ins prefetch_op:$Rt, ldrlit_label:$Imm19),
2526 "prfm\t$Rt, $Imm19",
2530 //===----------------------------------------------------------------------===//
2531 // Load-store exclusive instructions
2532 //===----------------------------------------------------------------------===//
2533 // Contains: STXRB, STXRH, STXR, LDXRB, LDXRH, LDXR. STXP, LDXP, STLXRB,
2534 // STLXRH, STLXR, LDAXRB, LDAXRH, LDAXR, STLXP, LDAXP, STLRB,
2535 // STLRH, STLR, LDARB, LDARH, LDAR
2537 // Since these instructions have the undefined register bits set to 1 in
2538 // their canonical form, we need a post encoder method to set those bits
2539 // to 1 when encoding these instructions. We do this using the
2540 // fixLoadStoreExclusive function. This function has template parameters:
2542 // fixLoadStoreExclusive<int hasRs, int hasRt2>
2544 // hasRs indicates that the instruction uses the Rs field, so we won't set
2545 // it to 1 (and the same for Rt2). We don't need template parameters for
2546 // the other register fiels since Rt and Rn are always used.
2548 // This operand parses a GPR64xsp register, followed by an optional immediate
2550 def GPR64xsp0_asmoperand : AsmOperandClass {
2551 let Name = "GPR64xsp0";
2552 let PredicateMethod = "isWrappedReg";
2553 let RenderMethod = "addRegOperands";
2554 let ParserMethod = "ParseLSXAddressOperand";
2555 // Diagnostics are provided by ParserMethod
2558 def GPR64xsp0 : RegisterOperand<GPR64xsp> {
2559 let ParserMatchClass = GPR64xsp0_asmoperand;
2562 //===----------------------------------
2563 // Store-exclusive (releasing & normal)
2564 //===----------------------------------
2566 class A64I_SRexs_impl<bits<2> size, bits<3> opcode, string asm, dag outs,
2567 dag ins, list<dag> pat,
2568 InstrItinClass itin> :
2569 A64I_LDSTex_stn <size,
2570 opcode{2}, 0, opcode{1}, opcode{0},
2572 !strconcat(asm, "\t$Rs, $Rt, [$Rn]"),
2575 let PostEncoderMethod = "fixLoadStoreExclusive<1,0>";
2578 multiclass A64I_SRex<string asmstr, bits<3> opcode, string prefix> {
2579 def _byte: A64I_SRexs_impl<0b00, opcode, !strconcat(asmstr, "b"),
2580 (outs GPR32:$Rs), (ins GPR32:$Rt, GPR64xsp0:$Rn),
2583 def _hword: A64I_SRexs_impl<0b01, opcode, !strconcat(asmstr, "h"),
2584 (outs GPR32:$Rs), (ins GPR32:$Rt, GPR64xsp0:$Rn),
2587 def _word: A64I_SRexs_impl<0b10, opcode, asmstr,
2588 (outs GPR32:$Rs), (ins GPR32:$Rt, GPR64xsp0:$Rn),
2591 def _dword: A64I_SRexs_impl<0b11, opcode, asmstr,
2592 (outs GPR32:$Rs), (ins GPR64:$Rt, GPR64xsp0:$Rn),
2596 defm STXR : A64I_SRex<"stxr", 0b000, "STXR">;
2597 defm STLXR : A64I_SRex<"stlxr", 0b001, "STLXR">;
2599 //===----------------------------------
2601 //===----------------------------------
2603 class A64I_LRexs_impl<bits<2> size, bits<3> opcode, string asm, dag outs,
2604 dag ins, list<dag> pat,
2605 InstrItinClass itin> :
2606 A64I_LDSTex_tn <size,
2607 opcode{2}, 1, opcode{1}, opcode{0},
2609 !strconcat(asm, "\t$Rt, [$Rn]"),
2612 let PostEncoderMethod = "fixLoadStoreExclusive<0,0>";
2615 multiclass A64I_LRex<string asmstr, bits<3> opcode> {
2616 def _byte: A64I_LRexs_impl<0b00, opcode, !strconcat(asmstr, "b"),
2617 (outs GPR32:$Rt), (ins GPR64xsp0:$Rn),
2620 def _hword: A64I_LRexs_impl<0b01, opcode, !strconcat(asmstr, "h"),
2621 (outs GPR32:$Rt), (ins GPR64xsp0:$Rn),
2624 def _word: A64I_LRexs_impl<0b10, opcode, asmstr,
2625 (outs GPR32:$Rt), (ins GPR64xsp0:$Rn),
2628 def _dword: A64I_LRexs_impl<0b11, opcode, asmstr,
2629 (outs GPR64:$Rt), (ins GPR64xsp0:$Rn),
2633 defm LDXR : A64I_LRex<"ldxr", 0b000>;
2634 defm LDAXR : A64I_LRex<"ldaxr", 0b001>;
2635 defm LDAR : A64I_LRex<"ldar", 0b101>;
2637 class acquiring_load<PatFrag base>
2638 : PatFrag<(ops node:$ptr), (base node:$ptr), [{
2639 AtomicOrdering Ordering = cast<AtomicSDNode>(N)->getOrdering();
2640 return Ordering == Acquire || Ordering == SequentiallyConsistent;
2643 def atomic_load_acquire_8 : acquiring_load<atomic_load_8>;
2644 def atomic_load_acquire_16 : acquiring_load<atomic_load_16>;
2645 def atomic_load_acquire_32 : acquiring_load<atomic_load_32>;
2646 def atomic_load_acquire_64 : acquiring_load<atomic_load_64>;
2648 def : Pat<(atomic_load_acquire_8 i64:$Rn), (LDAR_byte $Rn)>;
2649 def : Pat<(atomic_load_acquire_16 i64:$Rn), (LDAR_hword $Rn)>;
2650 def : Pat<(atomic_load_acquire_32 i64:$Rn), (LDAR_word $Rn)>;
2651 def : Pat<(atomic_load_acquire_64 i64:$Rn), (LDAR_dword $Rn)>;
2653 //===----------------------------------
2654 // Store-release (no exclusivity)
2655 //===----------------------------------
2657 class A64I_SLexs_impl<bits<2> size, bits<3> opcode, string asm, dag outs,
2658 dag ins, list<dag> pat,
2659 InstrItinClass itin> :
2660 A64I_LDSTex_tn <size,
2661 opcode{2}, 0, opcode{1}, opcode{0},
2663 !strconcat(asm, "\t$Rt, [$Rn]"),
2666 let PostEncoderMethod = "fixLoadStoreExclusive<0,0>";
2669 class releasing_store<PatFrag base>
2670 : PatFrag<(ops node:$ptr, node:$val), (base node:$ptr, node:$val), [{
2671 AtomicOrdering Ordering = cast<AtomicSDNode>(N)->getOrdering();
2672 return Ordering == Release || Ordering == SequentiallyConsistent;
2675 def atomic_store_release_8 : releasing_store<atomic_store_8>;
2676 def atomic_store_release_16 : releasing_store<atomic_store_16>;
2677 def atomic_store_release_32 : releasing_store<atomic_store_32>;
2678 def atomic_store_release_64 : releasing_store<atomic_store_64>;
2680 multiclass A64I_SLex<string asmstr, bits<3> opcode, string prefix> {
2681 def _byte: A64I_SLexs_impl<0b00, opcode, !strconcat(asmstr, "b"),
2682 (outs), (ins GPR32:$Rt, GPR64xsp0:$Rn),
2683 [(atomic_store_release_8 i64:$Rn, i32:$Rt)],
2686 def _hword: A64I_SLexs_impl<0b01, opcode, !strconcat(asmstr, "h"),
2687 (outs), (ins GPR32:$Rt, GPR64xsp0:$Rn),
2688 [(atomic_store_release_16 i64:$Rn, i32:$Rt)],
2691 def _word: A64I_SLexs_impl<0b10, opcode, asmstr,
2692 (outs), (ins GPR32:$Rt, GPR64xsp0:$Rn),
2693 [(atomic_store_release_32 i64:$Rn, i32:$Rt)],
2696 def _dword: A64I_SLexs_impl<0b11, opcode, asmstr,
2697 (outs), (ins GPR64:$Rt, GPR64xsp0:$Rn),
2698 [(atomic_store_release_64 i64:$Rn, i64:$Rt)],
2702 defm STLR : A64I_SLex<"stlr", 0b101, "STLR">;
2704 //===----------------------------------
2705 // Store-exclusive pair (releasing & normal)
2706 //===----------------------------------
2708 class A64I_SPexs_impl<bits<2> size, bits<3> opcode, string asm, dag outs,
2709 dag ins, list<dag> pat,
2710 InstrItinClass itin> :
2711 A64I_LDSTex_stt2n <size,
2712 opcode{2}, 0, opcode{1}, opcode{0},
2714 !strconcat(asm, "\t$Rs, $Rt, $Rt2, [$Rn]"),
2720 multiclass A64I_SPex<string asmstr, bits<3> opcode> {
2721 def _word: A64I_SPexs_impl<0b10, opcode, asmstr, (outs),
2722 (ins GPR32:$Rs, GPR32:$Rt, GPR32:$Rt2,
2726 def _dword: A64I_SPexs_impl<0b11, opcode, asmstr, (outs),
2727 (ins GPR32:$Rs, GPR64:$Rt, GPR64:$Rt2,
2732 defm STXP : A64I_SPex<"stxp", 0b010>;
2733 defm STLXP : A64I_SPex<"stlxp", 0b011>;
2735 //===----------------------------------
2736 // Load-exclusive pair (acquiring & normal)
2737 //===----------------------------------
2739 class A64I_LPexs_impl<bits<2> size, bits<3> opcode, string asm, dag outs,
2740 dag ins, list<dag> pat,
2741 InstrItinClass itin> :
2742 A64I_LDSTex_tt2n <size,
2743 opcode{2}, 1, opcode{1}, opcode{0},
2745 !strconcat(asm, "\t$Rt, $Rt2, [$Rn]"),
2748 let DecoderMethod = "DecodeLoadPairExclusiveInstruction";
2749 let PostEncoderMethod = "fixLoadStoreExclusive<0,1>";
2752 multiclass A64I_LPex<string asmstr, bits<3> opcode> {
2753 def _word: A64I_LPexs_impl<0b10, opcode, asmstr,
2754 (outs GPR32:$Rt, GPR32:$Rt2),
2755 (ins GPR64xsp0:$Rn),
2758 def _dword: A64I_LPexs_impl<0b11, opcode, asmstr,
2759 (outs GPR64:$Rt, GPR64:$Rt2),
2760 (ins GPR64xsp0:$Rn),
2764 defm LDXP : A64I_LPex<"ldxp", 0b010>;
2765 defm LDAXP : A64I_LPex<"ldaxp", 0b011>;
2767 //===----------------------------------------------------------------------===//
2768 // Load-store register (unscaled immediate) instructions
2769 //===----------------------------------------------------------------------===//
2770 // Contains: LDURB, LDURH, LDRUSB, LDRUSH, LDRUSW, STUR, STURB, STURH and PRFUM
2774 //===----------------------------------------------------------------------===//
2775 // Load-store register (register offset) instructions
2776 //===----------------------------------------------------------------------===//
2777 // Contains: LDRB, LDRH, LDRSB, LDRSH, LDRSW, STR, STRB, STRH and PRFM
2781 //===----------------------------------------------------------------------===//
2782 // Load-store register (unsigned immediate) instructions
2783 //===----------------------------------------------------------------------===//
2784 // Contains: LDRB, LDRH, LDRSB, LDRSH, LDRSW, STR, STRB, STRH and PRFM
2788 //===----------------------------------------------------------------------===//
2789 // Load-store register (immediate post-indexed) instructions
2790 //===----------------------------------------------------------------------===//
2791 // Contains: STRB, STRH, STR, LDRB, LDRH, LDR, LDRSB, LDRSH, LDRSW
2795 //===----------------------------------------------------------------------===//
2796 // Load-store register (immediate pre-indexed) instructions
2797 //===----------------------------------------------------------------------===//
2798 // Contains: STRB, STRH, STR, LDRB, LDRH, LDR, LDRSB, LDRSH, LDRSW
2800 // Note that patterns are much later on in a completely separate section (they
2801 // need ADRPxi to be defined).
2803 //===-------------------------------
2804 // 1. Various operands needed
2805 //===-------------------------------
2807 //===-------------------------------
2808 // 1.1 Unsigned 12-bit immediate operands
2809 //===-------------------------------
2810 // The addressing mode for these instructions consists of an unsigned 12-bit
2811 // immediate which is scaled by the size of the memory access.
2813 // We represent this in the MC layer by two operands:
2814 // 1. A base register.
2815 // 2. A 12-bit immediate: not multiplied by access size, so "LDR x0,[x0,#8]"
2816 // would have '1' in this field.
2817 // This means that separate functions are needed for converting representations
2818 // which *are* aware of the intended access size.
2820 // Anything that creates an MCInst (Decoding, selection and AsmParsing) has to
2821 // know the access size via some means. An isolated operand does not have this
2822 // information unless told from here, which means we need separate tablegen
2823 // Operands for each access size. This multiclass takes care of instantiating
2824 // the correct template functions in the rest of the backend.
2826 //===-------------------------------
2827 // 1.1 Unsigned 12-bit immediate operands
2828 //===-------------------------------
2830 multiclass offsets_uimm12<int MemSize, string prefix> {
2831 def uimm12_asmoperand : AsmOperandClass {
2832 let Name = "OffsetUImm12_" # MemSize;
2833 let PredicateMethod = "isOffsetUImm12<" # MemSize # ">";
2834 let RenderMethod = "addOffsetUImm12Operands<" # MemSize # ">";
2835 let DiagnosticType = "LoadStoreUImm12_" # MemSize;
2838 // Pattern is really no more than an ImmLeaf, but predicated on MemSize which
2839 // complicates things beyond TableGen's ken.
2840 def uimm12 : Operand<i64>,
2841 ComplexPattern<i64, 1, "SelectOffsetUImm12<" # MemSize # ">"> {
2842 let ParserMatchClass
2843 = !cast<AsmOperandClass>(prefix # uimm12_asmoperand);
2845 let PrintMethod = "printOffsetUImm12Operand<" # MemSize # ">";
2846 let EncoderMethod = "getOffsetUImm12OpValue<" # MemSize # ">";
2850 defm byte_ : offsets_uimm12<1, "byte_">;
2851 defm hword_ : offsets_uimm12<2, "hword_">;
2852 defm word_ : offsets_uimm12<4, "word_">;
2853 defm dword_ : offsets_uimm12<8, "dword_">;
2854 defm qword_ : offsets_uimm12<16, "qword_">;
2856 //===-------------------------------
2857 // 1.1 Signed 9-bit immediate operands
2858 //===-------------------------------
2860 // The MCInst is expected to store the bit-wise encoding of the value,
2861 // which amounts to lopping off the extended sign bits.
2862 def SDXF_simm9 : SDNodeXForm<imm, [{
2863 return CurDAG->getTargetConstant(N->getZExtValue() & 0x1ff, MVT::i32);
2866 def simm9_asmoperand : AsmOperandClass {
2868 let PredicateMethod = "isSImm<9>";
2869 let RenderMethod = "addSImmOperands<9>";
2870 let DiagnosticType = "LoadStoreSImm9";
2873 def simm9 : Operand<i64>,
2874 ImmLeaf<i64, [{ return Imm >= -0x100 && Imm <= 0xff; }],
2876 let PrintMethod = "printOffsetSImm9Operand";
2877 let ParserMatchClass = simm9_asmoperand;
2881 //===-------------------------------
2882 // 1.3 Register offset extensions
2883 //===-------------------------------
2885 // The assembly-syntax for these addressing-modes is:
2886 // [<Xn|SP>, <R><m> {, <extend> {<amount>}}]
2888 // The essential semantics are:
2889 // + <amount> is a shift: #<log(transfer size)> or #0
2890 // + <R> can be W or X.
2891 // + If <R> is W, <extend> can be UXTW or SXTW
2892 // + If <R> is X, <extend> can be LSL or SXTX
2894 // The trickiest of those constraints is that Rm can be either GPR32 or GPR64,
2895 // which will need separate instructions for LLVM type-consistency. We'll also
2896 // need separate operands, of course.
2897 multiclass regexts<int MemSize, int RmSize, RegisterClass GPR,
2898 string Rm, string prefix> {
2899 def regext_asmoperand : AsmOperandClass {
2900 let Name = "AddrRegExtend_" # MemSize # "_" # Rm;
2901 let PredicateMethod = "isAddrRegExtend<" # MemSize # "," # RmSize # ">";
2902 let RenderMethod = "addAddrRegExtendOperands<" # MemSize # ">";
2903 let DiagnosticType = "LoadStoreExtend" # RmSize # "_" # MemSize;
2906 def regext : Operand<i64> {
2908 = "printAddrRegExtendOperand<" # MemSize # ", " # RmSize # ">";
2910 let DecoderMethod = "DecodeAddrRegExtendOperand";
2911 let ParserMatchClass
2912 = !cast<AsmOperandClass>(prefix # regext_asmoperand);
2916 multiclass regexts_wx<int MemSize, string prefix> {
2917 // Rm is an X-register if LSL or SXTX are specified as the shift.
2918 defm Xm_ : regexts<MemSize, 64, GPR64, "Xm", prefix # "Xm_">;
2920 // Rm is a W-register if UXTW or SXTW are specified as the shift.
2921 defm Wm_ : regexts<MemSize, 32, GPR32, "Wm", prefix # "Wm_">;
2924 defm byte_ : regexts_wx<1, "byte_">;
2925 defm hword_ : regexts_wx<2, "hword_">;
2926 defm word_ : regexts_wx<4, "word_">;
2927 defm dword_ : regexts_wx<8, "dword_">;
2928 defm qword_ : regexts_wx<16, "qword_">;
2931 //===------------------------------
2932 // 2. The instructions themselves.
2933 //===------------------------------
2935 // We have the following instructions to implement:
2936 // | | B | H | W | X |
2937 // |-----------------+-------+-------+-------+--------|
2938 // | unsigned str | STRB | STRH | STR | STR |
2939 // | unsigned ldr | LDRB | LDRH | LDR | LDR |
2940 // | signed ldr to W | LDRSB | LDRSH | - | - |
2941 // | signed ldr to X | LDRSB | LDRSH | LDRSW | (PRFM) |
2943 // This will instantiate the LDR/STR instructions you'd expect to use for an
2944 // unsigned datatype (first two rows above) or floating-point register, which is
2945 // reasonably uniform across all access sizes.
2948 //===------------------------------
2949 // 2.1 Regular instructions
2950 //===------------------------------
2952 // This class covers the basic unsigned or irrelevantly-signed loads and stores,
2953 // to general-purpose and floating-point registers.
2955 class AddrParams<string prefix> {
2956 Operand uimm12 = !cast<Operand>(prefix # "_uimm12");
2958 Operand regextWm = !cast<Operand>(prefix # "_Wm_regext");
2959 Operand regextXm = !cast<Operand>(prefix # "_Xm_regext");
2962 def byte_addrparams : AddrParams<"byte">;
2963 def hword_addrparams : AddrParams<"hword">;
2964 def word_addrparams : AddrParams<"word">;
2965 def dword_addrparams : AddrParams<"dword">;
2966 def qword_addrparams : AddrParams<"qword">;
2968 multiclass A64I_LDRSTR_unsigned<string prefix, bits<2> size, bit v,
2969 bit high_opc, string asmsuffix,
2970 RegisterClass GPR, AddrParams params> {
2971 // Unsigned immediate
2972 def _STR : A64I_LSunsigimm<size, v, {high_opc, 0b0},
2973 (outs), (ins GPR:$Rt, GPR64xsp:$Rn, params.uimm12:$UImm12),
2974 "str" # asmsuffix # "\t$Rt, [$Rn, $UImm12]",
2978 def : InstAlias<"str" # asmsuffix # " $Rt, [$Rn]",
2979 (!cast<Instruction>(prefix # "_STR") GPR:$Rt, GPR64xsp:$Rn, 0)>;
2981 def _LDR : A64I_LSunsigimm<size, v, {high_opc, 0b1},
2982 (outs GPR:$Rt), (ins GPR64xsp:$Rn, params.uimm12:$UImm12),
2983 "ldr" # asmsuffix # "\t$Rt, [$Rn, $UImm12]",
2987 def : InstAlias<"ldr" # asmsuffix # " $Rt, [$Rn]",
2988 (!cast<Instruction>(prefix # "_LDR") GPR:$Rt, GPR64xsp:$Rn, 0)>;
2990 // Register offset (four of these: load/store and Wm/Xm).
2991 let mayLoad = 1 in {
2992 def _Wm_RegOffset_LDR : A64I_LSregoff<size, v, {high_opc, 0b1}, 0b0,
2994 (ins GPR64xsp:$Rn, GPR32:$Rm, params.regextWm:$Ext),
2995 "ldr" # asmsuffix # "\t$Rt, [$Rn, $Rm, $Ext]",
2998 def _Xm_RegOffset_LDR : A64I_LSregoff<size, v, {high_opc, 0b1}, 0b1,
3000 (ins GPR64xsp:$Rn, GPR64:$Rm, params.regextXm:$Ext),
3001 "ldr" # asmsuffix # "\t$Rt, [$Rn, $Rm, $Ext]",
3004 def : InstAlias<"ldr" # asmsuffix # " $Rt, [$Rn, $Rm]",
3005 (!cast<Instruction>(prefix # "_Xm_RegOffset_LDR") GPR:$Rt, GPR64xsp:$Rn,
3008 let mayStore = 1 in {
3009 def _Wm_RegOffset_STR : A64I_LSregoff<size, v, {high_opc, 0b0}, 0b0,
3010 (outs), (ins GPR:$Rt, GPR64xsp:$Rn, GPR32:$Rm,
3011 params.regextWm:$Ext),
3012 "str" # asmsuffix # "\t$Rt, [$Rn, $Rm, $Ext]",
3015 def _Xm_RegOffset_STR : A64I_LSregoff<size, v, {high_opc, 0b0}, 0b1,
3016 (outs), (ins GPR:$Rt, GPR64xsp:$Rn, GPR64:$Rm,
3017 params.regextXm:$Ext),
3018 "str" # asmsuffix # "\t$Rt, [$Rn, $Rm, $Ext]",
3021 def : InstAlias<"str" # asmsuffix # " $Rt, [$Rn, $Rm]",
3022 (!cast<Instruction>(prefix # "_Xm_RegOffset_STR") GPR:$Rt, GPR64xsp:$Rn,
3025 // Unaligned immediate
3026 def _STUR : A64I_LSunalimm<size, v, {high_opc, 0b0},
3027 (outs), (ins GPR:$Rt, GPR64xsp:$Rn, simm9:$SImm9),
3028 "stur" # asmsuffix # "\t$Rt, [$Rn, $SImm9]",
3032 def : InstAlias<"stur" # asmsuffix # " $Rt, [$Rn]",
3033 (!cast<Instruction>(prefix # "_STUR") GPR:$Rt, GPR64xsp:$Rn, 0)>;
3035 def _LDUR : A64I_LSunalimm<size, v, {high_opc, 0b1},
3036 (outs GPR:$Rt), (ins GPR64xsp:$Rn, simm9:$SImm9),
3037 "ldur" # asmsuffix # "\t$Rt, [$Rn, $SImm9]",
3041 def : InstAlias<"ldur" # asmsuffix # " $Rt, [$Rn]",
3042 (!cast<Instruction>(prefix # "_LDUR") GPR:$Rt, GPR64xsp:$Rn, 0)>;
3045 def _PostInd_STR : A64I_LSpostind<size, v, {high_opc, 0b0},
3046 (outs GPR64xsp:$Rn_wb),
3047 (ins GPR:$Rt, GPR64xsp:$Rn, simm9:$SImm9),
3048 "str" # asmsuffix # "\t$Rt, [$Rn], $SImm9",
3050 let Constraints = "$Rn = $Rn_wb";
3053 // Decoder only needed for unpredictability checking (FIXME).
3054 let DecoderMethod = "DecodeSingleIndexedInstruction";
3057 def _PostInd_LDR : A64I_LSpostind<size, v, {high_opc, 0b1},
3058 (outs GPR:$Rt, GPR64xsp:$Rn_wb),
3059 (ins GPR64xsp:$Rn, simm9:$SImm9),
3060 "ldr" # asmsuffix # "\t$Rt, [$Rn], $SImm9",
3063 let Constraints = "$Rn = $Rn_wb";
3064 let DecoderMethod = "DecodeSingleIndexedInstruction";
3068 def _PreInd_STR : A64I_LSpreind<size, v, {high_opc, 0b0},
3069 (outs GPR64xsp:$Rn_wb),
3070 (ins GPR:$Rt, GPR64xsp:$Rn, simm9:$SImm9),
3071 "str" # asmsuffix # "\t$Rt, [$Rn, $SImm9]!",
3073 let Constraints = "$Rn = $Rn_wb";
3076 // Decoder only needed for unpredictability checking (FIXME).
3077 let DecoderMethod = "DecodeSingleIndexedInstruction";
3080 def _PreInd_LDR : A64I_LSpreind<size, v, {high_opc, 0b1},
3081 (outs GPR:$Rt, GPR64xsp:$Rn_wb),
3082 (ins GPR64xsp:$Rn, simm9:$SImm9),
3083 "ldr" # asmsuffix # "\t$Rt, [$Rn, $SImm9]!",
3086 let Constraints = "$Rn = $Rn_wb";
3087 let DecoderMethod = "DecodeSingleIndexedInstruction";
3092 // STRB/LDRB: First define the instructions
3094 : A64I_LDRSTR_unsigned<"LS8", 0b00, 0b0, 0b0, "b", GPR32, byte_addrparams>;
3098 : A64I_LDRSTR_unsigned<"LS16", 0b01, 0b0, 0b0, "h", GPR32, hword_addrparams>;
3101 // STR/LDR to/from a W register
3103 : A64I_LDRSTR_unsigned<"LS32", 0b10, 0b0, 0b0, "", GPR32, word_addrparams>;
3105 // STR/LDR to/from an X register
3107 : A64I_LDRSTR_unsigned<"LS64", 0b11, 0b0, 0b0, "", GPR64, dword_addrparams>;
3109 // STR/LDR to/from a B register
3111 : A64I_LDRSTR_unsigned<"LSFP8", 0b00, 0b1, 0b0, "", FPR8, byte_addrparams>;
3113 // STR/LDR to/from an H register
3115 : A64I_LDRSTR_unsigned<"LSFP16", 0b01, 0b1, 0b0, "", FPR16, hword_addrparams>;
3117 // STR/LDR to/from an S register
3119 : A64I_LDRSTR_unsigned<"LSFP32", 0b10, 0b1, 0b0, "", FPR32, word_addrparams>;
3120 // STR/LDR to/from a D register
3122 : A64I_LDRSTR_unsigned<"LSFP64", 0b11, 0b1, 0b0, "", FPR64, dword_addrparams>;
3123 // STR/LDR to/from a Q register
3125 : A64I_LDRSTR_unsigned<"LSFP128", 0b00, 0b1, 0b1, "", FPR128,
3128 //===------------------------------
3130 //===------------------------------
3132 // Byte and half-word signed loads can both go into either an X or a W register,
3133 // so it's worth factoring out. Signed word loads don't fit because there is no
3135 multiclass A64I_LDR_signed<bits<2> size, string asmopcode, AddrParams params,
3138 def w : A64I_LSunsigimm<size, 0b0, 0b11,
3140 (ins GPR64xsp:$Rn, params.uimm12:$UImm12),
3141 "ldrs" # asmopcode # "\t$Rt, [$Rn, $UImm12]",
3145 def : InstAlias<"ldrs" # asmopcode # " $Rt, [$Rn]",
3146 (!cast<Instruction>(prefix # w) GPR32:$Rt, GPR64xsp:$Rn, 0)>;
3148 def x : A64I_LSunsigimm<size, 0b0, 0b10,
3150 (ins GPR64xsp:$Rn, params.uimm12:$UImm12),
3151 "ldrs" # asmopcode # "\t$Rt, [$Rn, $UImm12]",
3155 def : InstAlias<"ldrs" # asmopcode # " $Rt, [$Rn]",
3156 (!cast<Instruction>(prefix # x) GPR64:$Rt, GPR64xsp:$Rn, 0)>;
3159 let mayLoad = 1 in {
3160 def w_Wm_RegOffset : A64I_LSregoff<size, 0b0, 0b11, 0b0,
3162 (ins GPR64xsp:$Rn, GPR32:$Rm, params.regextWm:$Ext),
3163 "ldrs" # asmopcode # "\t$Rt, [$Rn, $Rm, $Ext]",
3166 def w_Xm_RegOffset : A64I_LSregoff<size, 0b0, 0b11, 0b1,
3168 (ins GPR64xsp:$Rn, GPR64:$Rm, params.regextXm:$Ext),
3169 "ldrs" # asmopcode # "\t$Rt, [$Rn, $Rm, $Ext]",
3172 def x_Wm_RegOffset : A64I_LSregoff<size, 0b0, 0b10, 0b0,
3174 (ins GPR64xsp:$Rn, GPR32:$Rm, params.regextWm:$Ext),
3175 "ldrs" # asmopcode # "\t$Rt, [$Rn, $Rm, $Ext]",
3178 def x_Xm_RegOffset : A64I_LSregoff<size, 0b0, 0b10, 0b1,
3180 (ins GPR64xsp:$Rn, GPR64:$Rm, params.regextXm:$Ext),
3181 "ldrs" # asmopcode # "\t$Rt, [$Rn, $Rm, $Ext]",
3184 def : InstAlias<"ldrs" # asmopcode # " $Rt, [$Rn, $Rm]",
3185 (!cast<Instruction>(prefix # "w_Xm_RegOffset") GPR32:$Rt, GPR64xsp:$Rn,
3188 def : InstAlias<"ldrs" # asmopcode # " $Rt, [$Rn, $Rm]",
3189 (!cast<Instruction>(prefix # "x_Xm_RegOffset") GPR64:$Rt, GPR64xsp:$Rn,
3193 let mayLoad = 1 in {
3195 def w_U : A64I_LSunalimm<size, 0b0, 0b11,
3197 (ins GPR64xsp:$Rn, simm9:$SImm9),
3198 "ldurs" # asmopcode # "\t$Rt, [$Rn, $SImm9]",
3201 def x_U : A64I_LSunalimm<size, 0b0, 0b10,
3203 (ins GPR64xsp:$Rn, simm9:$SImm9),
3204 "ldurs" # asmopcode # "\t$Rt, [$Rn, $SImm9]",
3209 def w_PostInd : A64I_LSpostind<size, 0b0, 0b11,
3210 (outs GPR32:$Rt, GPR64xsp:$Rn_wb),
3211 (ins GPR64xsp:$Rn, simm9:$SImm9),
3212 "ldrs" # asmopcode # "\t$Rt, [$Rn], $SImm9",
3214 let Constraints = "$Rn = $Rn_wb";
3215 let DecoderMethod = "DecodeSingleIndexedInstruction";
3218 def x_PostInd : A64I_LSpostind<size, 0b0, 0b10,
3219 (outs GPR64:$Rt, GPR64xsp:$Rn_wb),
3220 (ins GPR64xsp:$Rn, simm9:$SImm9),
3221 "ldrs" # asmopcode # "\t$Rt, [$Rn], $SImm9",
3223 let Constraints = "$Rn = $Rn_wb";
3224 let DecoderMethod = "DecodeSingleIndexedInstruction";
3228 def w_PreInd : A64I_LSpreind<size, 0b0, 0b11,
3229 (outs GPR32:$Rt, GPR64xsp:$Rn_wb),
3230 (ins GPR64xsp:$Rn, simm9:$SImm9),
3231 "ldrs" # asmopcode # "\t$Rt, [$Rn, $SImm9]!",
3233 let Constraints = "$Rn = $Rn_wb";
3234 let DecoderMethod = "DecodeSingleIndexedInstruction";
3237 def x_PreInd : A64I_LSpreind<size, 0b0, 0b10,
3238 (outs GPR64:$Rt, GPR64xsp:$Rn_wb),
3239 (ins GPR64xsp:$Rn, simm9:$SImm9),
3240 "ldrs" # asmopcode # "\t$Rt, [$Rn, $SImm9]!",
3242 let Constraints = "$Rn = $Rn_wb";
3243 let DecoderMethod = "DecodeSingleIndexedInstruction";
3245 } // let mayLoad = 1
3249 defm LDRSB : A64I_LDR_signed<0b00, "b", byte_addrparams, "LDRSB">;
3251 defm LDRSH : A64I_LDR_signed<0b01, "h", hword_addrparams, "LDRSH">;
3253 // LDRSW: load a 32-bit register, sign-extending to 64-bits.
3255 : A64I_LSunsigimm<0b10, 0b0, 0b10,
3257 (ins GPR64xsp:$Rn, word_uimm12:$UImm12),
3258 "ldrsw\t$Rt, [$Rn, $UImm12]",
3262 def : InstAlias<"ldrsw $Rt, [$Rn]", (LDRSWx GPR64:$Rt, GPR64xsp:$Rn, 0)>;
3264 let mayLoad = 1 in {
3265 def LDRSWx_Wm_RegOffset : A64I_LSregoff<0b10, 0b0, 0b10, 0b0,
3267 (ins GPR64xsp:$Rn, GPR32:$Rm, word_Wm_regext:$Ext),
3268 "ldrsw\t$Rt, [$Rn, $Rm, $Ext]",
3271 def LDRSWx_Xm_RegOffset : A64I_LSregoff<0b10, 0b0, 0b10, 0b1,
3273 (ins GPR64xsp:$Rn, GPR64:$Rm, word_Xm_regext:$Ext),
3274 "ldrsw\t$Rt, [$Rn, $Rm, $Ext]",
3277 def : InstAlias<"ldrsw $Rt, [$Rn, $Rm]",
3278 (LDRSWx_Xm_RegOffset GPR64:$Rt, GPR64xsp:$Rn, GPR64:$Rm, 2)>;
3282 : A64I_LSunalimm<0b10, 0b0, 0b10,
3284 (ins GPR64xsp:$Rn, simm9:$SImm9),
3285 "ldursw\t$Rt, [$Rn, $SImm9]",
3289 def : InstAlias<"ldursw $Rt, [$Rn]", (LDURSWx GPR64:$Rt, GPR64xsp:$Rn, 0)>;
3292 : A64I_LSpostind<0b10, 0b0, 0b10,
3293 (outs GPR64:$Rt, GPR64xsp:$Rn_wb),
3294 (ins GPR64xsp:$Rn, simm9:$SImm9),
3295 "ldrsw\t$Rt, [$Rn], $SImm9",
3298 let Constraints = "$Rn = $Rn_wb";
3299 let DecoderMethod = "DecodeSingleIndexedInstruction";
3302 def LDRSWx_PreInd : A64I_LSpreind<0b10, 0b0, 0b10,
3303 (outs GPR64:$Rt, GPR64xsp:$Rn_wb),
3304 (ins GPR64xsp:$Rn, simm9:$SImm9),
3305 "ldrsw\t$Rt, [$Rn, $SImm9]!",
3308 let Constraints = "$Rn = $Rn_wb";
3309 let DecoderMethod = "DecodeSingleIndexedInstruction";
3312 //===------------------------------
3313 // 2.4 Prefetch operations
3314 //===------------------------------
3316 def PRFM : A64I_LSunsigimm<0b11, 0b0, 0b10, (outs),
3317 (ins prefetch_op:$Rt, GPR64xsp:$Rn, dword_uimm12:$UImm12),
3318 "prfm\t$Rt, [$Rn, $UImm12]",
3322 def : InstAlias<"prfm $Rt, [$Rn]",
3323 (PRFM prefetch_op:$Rt, GPR64xsp:$Rn, 0)>;
3325 let mayLoad = 1 in {
3326 def PRFM_Wm_RegOffset : A64I_LSregoff<0b11, 0b0, 0b10, 0b0, (outs),
3327 (ins prefetch_op:$Rt, GPR64xsp:$Rn,
3328 GPR32:$Rm, dword_Wm_regext:$Ext),
3329 "prfm\t$Rt, [$Rn, $Rm, $Ext]",
3331 def PRFM_Xm_RegOffset : A64I_LSregoff<0b11, 0b0, 0b10, 0b1, (outs),
3332 (ins prefetch_op:$Rt, GPR64xsp:$Rn,
3333 GPR64:$Rm, dword_Xm_regext:$Ext),
3334 "prfm\t$Rt, [$Rn, $Rm, $Ext]",
3338 def : InstAlias<"prfm $Rt, [$Rn, $Rm]",
3339 (PRFM_Xm_RegOffset prefetch_op:$Rt, GPR64xsp:$Rn,
3343 def PRFUM : A64I_LSunalimm<0b11, 0b0, 0b10, (outs),
3344 (ins prefetch_op:$Rt, GPR64xsp:$Rn, simm9:$SImm9),
3345 "prfum\t$Rt, [$Rn, $SImm9]",
3349 def : InstAlias<"prfum $Rt, [$Rn]",
3350 (PRFUM prefetch_op:$Rt, GPR64xsp:$Rn, 0)>;
3352 //===----------------------------------------------------------------------===//
3353 // Load-store register (unprivileged) instructions
3354 //===----------------------------------------------------------------------===//
3355 // Contains: LDTRB, LDTRH, LDTRSB, LDTRSH, LDTRSW, STTR, STTRB and STTRH
3357 // These instructions very much mirror the "unscaled immediate" loads, but since
3358 // there are no floating-point variants we need to split them out into their own
3359 // section to avoid instantiation of "ldtr d0, [sp]" etc.
3361 multiclass A64I_LDTRSTTR<bits<2> size, string asmsuffix, RegisterClass GPR,
3363 def _UnPriv_STR : A64I_LSunpriv<size, 0b0, 0b00,
3364 (outs), (ins GPR:$Rt, GPR64xsp:$Rn, simm9:$SImm9),
3365 "sttr" # asmsuffix # "\t$Rt, [$Rn, $SImm9]",
3370 def : InstAlias<"sttr" # asmsuffix # " $Rt, [$Rn]",
3371 (!cast<Instruction>(prefix # "_UnPriv_STR") GPR:$Rt, GPR64xsp:$Rn, 0)>;
3373 def _UnPriv_LDR : A64I_LSunpriv<size, 0b0, 0b01,
3374 (outs GPR:$Rt), (ins GPR64xsp:$Rn, simm9:$SImm9),
3375 "ldtr" # asmsuffix # "\t$Rt, [$Rn, $SImm9]",
3380 def : InstAlias<"ldtr" # asmsuffix # " $Rt, [$Rn]",
3381 (!cast<Instruction>(prefix # "_UnPriv_LDR") GPR:$Rt, GPR64xsp:$Rn, 0)>;
3385 // STTRB/LDTRB: First define the instructions
3386 defm LS8 : A64I_LDTRSTTR<0b00, "b", GPR32, "LS8">;
3389 defm LS16 : A64I_LDTRSTTR<0b01, "h", GPR32, "LS16">;
3391 // STTR/LDTR to/from a W register
3392 defm LS32 : A64I_LDTRSTTR<0b10, "", GPR32, "LS32">;
3394 // STTR/LDTR to/from an X register
3395 defm LS64 : A64I_LDTRSTTR<0b11, "", GPR64, "LS64">;
3397 // Now a class for the signed instructions that can go to either 32 or 64
3399 multiclass A64I_LDTR_signed<bits<2> size, string asmopcode, string prefix> {
3400 let mayLoad = 1 in {
3401 def w : A64I_LSunpriv<size, 0b0, 0b11,
3403 (ins GPR64xsp:$Rn, simm9:$SImm9),
3404 "ldtrs" # asmopcode # "\t$Rt, [$Rn, $SImm9]",
3407 def x : A64I_LSunpriv<size, 0b0, 0b10,
3409 (ins GPR64xsp:$Rn, simm9:$SImm9),
3410 "ldtrs" # asmopcode # "\t$Rt, [$Rn, $SImm9]",
3414 def : InstAlias<"ldtrs" # asmopcode # " $Rt, [$Rn]",
3415 (!cast<Instruction>(prefix # "w") GPR32:$Rt, GPR64xsp:$Rn, 0)>;
3417 def : InstAlias<"ldtrs" # asmopcode # " $Rt, [$Rn]",
3418 (!cast<Instruction>(prefix # "x") GPR64:$Rt, GPR64xsp:$Rn, 0)>;
3423 defm LDTRSB : A64I_LDTR_signed<0b00, "b", "LDTRSB">;
3425 defm LDTRSH : A64I_LDTR_signed<0b01, "h", "LDTRSH">;
3427 // And finally LDTRSW which only goes to 64 bits.
3428 def LDTRSWx : A64I_LSunpriv<0b10, 0b0, 0b10,
3430 (ins GPR64xsp:$Rn, simm9:$SImm9),
3431 "ldtrsw\t$Rt, [$Rn, $SImm9]",
3435 def : InstAlias<"ldtrsw $Rt, [$Rn]", (LDTRSWx GPR64:$Rt, GPR64xsp:$Rn, 0)>;
3437 //===----------------------------------------------------------------------===//
3438 // Load-store register pair (offset) instructions
3439 //===----------------------------------------------------------------------===//
3443 //===----------------------------------------------------------------------===//
3444 // Load-store register pair (post-indexed) instructions
3445 //===----------------------------------------------------------------------===//
3446 // Contains: STP, LDP, LDPSW
3450 //===----------------------------------------------------------------------===//
3451 // Load-store register pair (pre-indexed) instructions
3452 //===----------------------------------------------------------------------===//
3453 // Contains: STP, LDP, LDPSW
3457 //===----------------------------------------------------------------------===//
3458 // Load-store non-temporal register pair (offset) instructions
3459 //===----------------------------------------------------------------------===//
3460 // Contains: STNP, LDNP
3463 // Anything that creates an MCInst (Decoding, selection and AsmParsing) has to
3464 // know the access size via some means. An isolated operand does not have this
3465 // information unless told from here, which means we need separate tablegen
3466 // Operands for each access size. This multiclass takes care of instantiating
3467 // the correct template functions in the rest of the backend.
3469 multiclass offsets_simm7<string MemSize, string prefix> {
3470 // The bare signed 7-bit immediate is used in post-indexed instructions, but
3471 // because of the scaling performed a generic "simm7" operand isn't
3472 // appropriate here either.
3473 def simm7_asmoperand : AsmOperandClass {
3474 let Name = "SImm7_Scaled" # MemSize;
3475 let PredicateMethod = "isSImm7Scaled<" # MemSize # ">";
3476 let RenderMethod = "addSImm7ScaledOperands<" # MemSize # ">";
3477 let DiagnosticType = "LoadStoreSImm7_" # MemSize;
3480 def simm7 : Operand<i64> {
3481 let PrintMethod = "printSImm7ScaledOperand<" # MemSize # ">";
3482 let ParserMatchClass = !cast<AsmOperandClass>(prefix # "simm7_asmoperand");
3486 defm word_ : offsets_simm7<"4", "word_">;
3487 defm dword_ : offsets_simm7<"8", "dword_">;
3488 defm qword_ : offsets_simm7<"16", "qword_">;
3490 multiclass A64I_LSPsimple<bits<2> opc, bit v, RegisterClass SomeReg,
3491 Operand simm7, string prefix> {
3492 def _STR : A64I_LSPoffset<opc, v, 0b0, (outs),
3493 (ins SomeReg:$Rt, SomeReg:$Rt2, GPR64xsp:$Rn, simm7:$SImm7),
3494 "stp\t$Rt, $Rt2, [$Rn, $SImm7]", [], NoItinerary> {
3496 let DecoderMethod = "DecodeLDSTPairInstruction";
3498 def : InstAlias<"stp $Rt, $Rt2, [$Rn]",
3499 (!cast<Instruction>(prefix # "_STR") SomeReg:$Rt,
3500 SomeReg:$Rt2, GPR64xsp:$Rn, 0)>;
3502 def _LDR : A64I_LSPoffset<opc, v, 0b1,
3503 (outs SomeReg:$Rt, SomeReg:$Rt2),
3504 (ins GPR64xsp:$Rn, simm7:$SImm7),
3505 "ldp\t$Rt, $Rt2, [$Rn, $SImm7]", [], NoItinerary> {
3507 let DecoderMethod = "DecodeLDSTPairInstruction";
3509 def : InstAlias<"ldp $Rt, $Rt2, [$Rn]",
3510 (!cast<Instruction>(prefix # "_LDR") SomeReg:$Rt,
3511 SomeReg:$Rt2, GPR64xsp:$Rn, 0)>;
3513 def _PostInd_STR : A64I_LSPpostind<opc, v, 0b0,
3514 (outs GPR64xsp:$Rn_wb),
3515 (ins SomeReg:$Rt, SomeReg:$Rt2,
3518 "stp\t$Rt, $Rt2, [$Rn], $SImm7",
3521 let Constraints = "$Rn = $Rn_wb";
3523 // Decoder only needed for unpredictability checking (FIXME).
3524 let DecoderMethod = "DecodeLDSTPairInstruction";
3527 def _PostInd_LDR : A64I_LSPpostind<opc, v, 0b1,
3528 (outs SomeReg:$Rt, SomeReg:$Rt2, GPR64xsp:$Rn_wb),
3529 (ins GPR64xsp:$Rn, simm7:$SImm7),
3530 "ldp\t$Rt, $Rt2, [$Rn], $SImm7",
3533 let Constraints = "$Rn = $Rn_wb";
3534 let DecoderMethod = "DecodeLDSTPairInstruction";
3537 def _PreInd_STR : A64I_LSPpreind<opc, v, 0b0, (outs GPR64xsp:$Rn_wb),
3538 (ins SomeReg:$Rt, SomeReg:$Rt2, GPR64xsp:$Rn, simm7:$SImm7),
3539 "stp\t$Rt, $Rt2, [$Rn, $SImm7]!",
3542 let Constraints = "$Rn = $Rn_wb";
3543 let DecoderMethod = "DecodeLDSTPairInstruction";
3546 def _PreInd_LDR : A64I_LSPpreind<opc, v, 0b1,
3547 (outs SomeReg:$Rt, SomeReg:$Rt2, GPR64xsp:$Rn_wb),
3548 (ins GPR64xsp:$Rn, simm7:$SImm7),
3549 "ldp\t$Rt, $Rt2, [$Rn, $SImm7]!",
3552 let Constraints = "$Rn = $Rn_wb";
3553 let DecoderMethod = "DecodeLDSTPairInstruction";
3556 def _NonTemp_STR : A64I_LSPnontemp<opc, v, 0b0, (outs),
3557 (ins SomeReg:$Rt, SomeReg:$Rt2, GPR64xsp:$Rn, simm7:$SImm7),
3558 "stnp\t$Rt, $Rt2, [$Rn, $SImm7]", [], NoItinerary> {
3560 let DecoderMethod = "DecodeLDSTPairInstruction";
3562 def : InstAlias<"stnp $Rt, $Rt2, [$Rn]",
3563 (!cast<Instruction>(prefix # "_NonTemp_STR") SomeReg:$Rt,
3564 SomeReg:$Rt2, GPR64xsp:$Rn, 0)>;
3566 def _NonTemp_LDR : A64I_LSPnontemp<opc, v, 0b1,
3567 (outs SomeReg:$Rt, SomeReg:$Rt2),
3568 (ins GPR64xsp:$Rn, simm7:$SImm7),
3569 "ldnp\t$Rt, $Rt2, [$Rn, $SImm7]", [], NoItinerary> {
3571 let DecoderMethod = "DecodeLDSTPairInstruction";
3573 def : InstAlias<"ldnp $Rt, $Rt2, [$Rn]",
3574 (!cast<Instruction>(prefix # "_NonTemp_LDR") SomeReg:$Rt,
3575 SomeReg:$Rt2, GPR64xsp:$Rn, 0)>;
3580 defm LSPair32 : A64I_LSPsimple<0b00, 0b0, GPR32, word_simm7, "LSPair32">;
3581 defm LSPair64 : A64I_LSPsimple<0b10, 0b0, GPR64, dword_simm7, "LSPair64">;
3582 defm LSFPPair32 : A64I_LSPsimple<0b00, 0b1, FPR32, word_simm7, "LSFPPair32">;
3583 defm LSFPPair64 : A64I_LSPsimple<0b01, 0b1, FPR64, dword_simm7, "LSFPPair64">;
3584 defm LSFPPair128 : A64I_LSPsimple<0b10, 0b1, FPR128, qword_simm7,
3588 def LDPSWx : A64I_LSPoffset<0b01, 0b0, 0b1,
3589 (outs GPR64:$Rt, GPR64:$Rt2),
3590 (ins GPR64xsp:$Rn, word_simm7:$SImm7),
3591 "ldpsw\t$Rt, $Rt2, [$Rn, $SImm7]", [], NoItinerary> {
3593 let DecoderMethod = "DecodeLDSTPairInstruction";
3595 def : InstAlias<"ldpsw $Rt, $Rt2, [$Rn]",
3596 (LDPSWx GPR64:$Rt, GPR64:$Rt2, GPR64xsp:$Rn, 0)>;
3598 def LDPSWx_PostInd : A64I_LSPpostind<0b01, 0b0, 0b1,
3599 (outs GPR64:$Rt, GPR64:$Rt2, GPR64:$Rn_wb),
3600 (ins GPR64xsp:$Rn, word_simm7:$SImm7),
3601 "ldpsw\t$Rt, $Rt2, [$Rn], $SImm7",
3604 let Constraints = "$Rn = $Rn_wb";
3605 let DecoderMethod = "DecodeLDSTPairInstruction";
3608 def LDPSWx_PreInd : A64I_LSPpreind<0b01, 0b0, 0b1,
3609 (outs GPR64:$Rt, GPR64:$Rt2, GPR64:$Rn_wb),
3610 (ins GPR64xsp:$Rn, word_simm7:$SImm7),
3611 "ldpsw\t$Rt, $Rt2, [$Rn, $SImm7]!",
3614 let Constraints = "$Rn = $Rn_wb";
3615 let DecoderMethod = "DecodeLDSTPairInstruction";
3618 //===----------------------------------------------------------------------===//
3619 // Logical (immediate) instructions
3620 //===----------------------------------------------------------------------===//
3621 // Contains: AND, ORR, EOR, ANDS, + aliases TST, MOV
3623 multiclass logical_imm_operands<string prefix, string note,
3624 int size, ValueType VT> {
3625 def _asmoperand : AsmOperandClass {
3626 let Name = "LogicalImm" # note # size;
3627 let PredicateMethod = "isLogicalImm" # note # "<" # size # ">";
3628 let RenderMethod = "addLogicalImmOperands<" # size # ">";
3629 let DiagnosticType = "LogicalSecondSource";
3633 : Operand<VT>, ComplexPattern<VT, 1, "SelectLogicalImm", [imm]> {
3634 let ParserMatchClass = !cast<AsmOperandClass>(prefix # "_asmoperand");
3635 let PrintMethod = "printLogicalImmOperand<" # size # ">";
3636 let DecoderMethod = "DecodeLogicalImmOperand<" # size # ">";
3640 defm logical_imm32 : logical_imm_operands<"logical_imm32", "", 32, i32>;
3641 defm logical_imm64 : logical_imm_operands<"logical_imm64", "", 64, i64>;
3643 // The mov versions only differ in assembly parsing, where they
3644 // exclude values representable with either MOVZ or MOVN.
3645 defm logical_imm32_mov
3646 : logical_imm_operands<"logical_imm32_mov", "MOV", 32, i32>;
3647 defm logical_imm64_mov
3648 : logical_imm_operands<"logical_imm64_mov", "MOV", 64, i64>;
3651 multiclass A64I_logimmSizes<bits<2> opc, string asmop, SDNode opnode> {
3652 def wwi : A64I_logicalimm<0b0, opc, (outs GPR32wsp:$Rd),
3653 (ins GPR32:$Rn, logical_imm32_operand:$Imm),
3654 !strconcat(asmop, "\t$Rd, $Rn, $Imm"),
3656 (opnode i32:$Rn, logical_imm32_operand:$Imm))],
3659 def xxi : A64I_logicalimm<0b1, opc, (outs GPR64xsp:$Rd),
3660 (ins GPR64:$Rn, logical_imm64_operand:$Imm),
3661 !strconcat(asmop, "\t$Rd, $Rn, $Imm"),
3663 (opnode i64:$Rn, logical_imm64_operand:$Imm))],
3667 defm AND : A64I_logimmSizes<0b00, "and", and>;
3668 defm ORR : A64I_logimmSizes<0b01, "orr", or>;
3669 defm EOR : A64I_logimmSizes<0b10, "eor", xor>;
3671 let Defs = [NZCV] in {
3672 def ANDSwwi : A64I_logicalimm<0b0, 0b11, (outs GPR32:$Rd),
3673 (ins GPR32:$Rn, logical_imm32_operand:$Imm),
3674 "ands\t$Rd, $Rn, $Imm",
3677 def ANDSxxi : A64I_logicalimm<0b1, 0b11, (outs GPR64:$Rd),
3678 (ins GPR64:$Rn, logical_imm64_operand:$Imm),
3679 "ands\t$Rd, $Rn, $Imm",
3684 def : InstAlias<"tst $Rn, $Imm",
3685 (ANDSwwi WZR, GPR32:$Rn, logical_imm32_operand:$Imm)>;
3686 def : InstAlias<"tst $Rn, $Imm",
3687 (ANDSxxi XZR, GPR64:$Rn, logical_imm64_operand:$Imm)>;
3688 def : InstAlias<"mov $Rd, $Imm",
3689 (ORRwwi GPR32wsp:$Rd, WZR, logical_imm32_mov_operand:$Imm)>;
3690 def : InstAlias<"mov $Rd, $Imm",
3691 (ORRxxi GPR64xsp:$Rd, XZR, logical_imm64_mov_operand:$Imm)>;
3693 //===----------------------------------------------------------------------===//
3694 // Logical (shifted register) instructions
3695 //===----------------------------------------------------------------------===//
3696 // Contains: AND, BIC, ORR, ORN, EOR, EON, ANDS, BICS + aliases TST, MVN, MOV
3698 // Operand for optimizing (icmp (and LHS, RHS), 0, SomeCode). In theory "ANDS"
3699 // behaves differently for unsigned comparisons, so we defensively only allow
3700 // signed or n/a as the operand. In practice "unsigned greater than 0" is "not
3701 // equal to 0" and LLVM gives us this.
3702 def signed_cond : PatLeaf<(cond), [{
3703 return !isUnsignedIntSetCC(N->get());
3707 // These instructions share their "shift" operands with add/sub (shifted
3708 // register instructions). They are defined there.
3710 // N.b. the commutable parameter is just !N. It will be first against the wall
3711 // when the revolution comes.
3712 multiclass logical_shifts<string prefix, bit sf, bits<2> opc,
3713 bit N, bit commutable,
3714 string asmop, SDPatternOperator opfrag, ValueType ty,
3715 RegisterClass GPR, list<Register> defs> {
3716 let isCommutable = commutable, Defs = defs in {
3717 def _lsl : A64I_logicalshift<sf, opc, 0b00, N,
3719 (ins GPR:$Rn, GPR:$Rm,
3720 !cast<Operand>("lsl_operand_" # ty):$Imm6),
3721 !strconcat(asmop, "\t$Rd, $Rn, $Rm, $Imm6"),
3722 [(set ty:$Rd, (opfrag ty:$Rn, (shl ty:$Rm,
3723 !cast<Operand>("lsl_operand_" # ty):$Imm6))
3727 def _lsr : A64I_logicalshift<sf, opc, 0b01, N,
3729 (ins GPR:$Rn, GPR:$Rm,
3730 !cast<Operand>("lsr_operand_" # ty):$Imm6),
3731 !strconcat(asmop, "\t$Rd, $Rn, $Rm, $Imm6"),
3732 [(set ty:$Rd, (opfrag ty:$Rn, (srl ty:$Rm,
3733 !cast<Operand>("lsr_operand_" # ty):$Imm6))
3737 def _asr : A64I_logicalshift<sf, opc, 0b10, N,
3739 (ins GPR:$Rn, GPR:$Rm,
3740 !cast<Operand>("asr_operand_" # ty):$Imm6),
3741 !strconcat(asmop, "\t$Rd, $Rn, $Rm, $Imm6"),
3742 [(set ty:$Rd, (opfrag ty:$Rn, (sra ty:$Rm,
3743 !cast<Operand>("asr_operand_" # ty):$Imm6))
3747 def _ror : A64I_logicalshift<sf, opc, 0b11, N,
3749 (ins GPR:$Rn, GPR:$Rm,
3750 !cast<Operand>("ror_operand_" # ty):$Imm6),
3751 !strconcat(asmop, "\t$Rd, $Rn, $Rm, $Imm6"),
3752 [(set ty:$Rd, (opfrag ty:$Rn, (rotr ty:$Rm,
3753 !cast<Operand>("ror_operand_" # ty):$Imm6))
3759 : InstAlias<!strconcat(asmop, " $Rd, $Rn, $Rm"),
3760 (!cast<Instruction>(prefix # "_lsl") GPR:$Rd, GPR:$Rn,
3763 def : Pat<(opfrag ty:$Rn, ty:$Rm),
3764 (!cast<Instruction>(prefix # "_lsl") $Rn, $Rm, 0)>;
3767 multiclass logical_sizes<string prefix, bits<2> opc, bit N, bit commutable,
3768 string asmop, SDPatternOperator opfrag,
3769 list<Register> defs> {
3770 defm xxx : logical_shifts<prefix # "xxx", 0b1, opc, N,
3771 commutable, asmop, opfrag, i64, GPR64, defs>;
3772 defm www : logical_shifts<prefix # "www", 0b0, opc, N,
3773 commutable, asmop, opfrag, i32, GPR32, defs>;
3777 defm AND : logical_sizes<"AND", 0b00, 0b0, 0b1, "and", and, []>;
3778 defm ORR : logical_sizes<"ORR", 0b01, 0b0, 0b1, "orr", or, []>;
3779 defm EOR : logical_sizes<"EOR", 0b10, 0b0, 0b1, "eor", xor, []>;
3780 defm ANDS : logical_sizes<"ANDS", 0b11, 0b0, 0b1, "ands",
3781 PatFrag<(ops node:$lhs, node:$rhs), (and node:$lhs, node:$rhs),
3782 [{ (void)N; return false; }]>,
3785 defm BIC : logical_sizes<"BIC", 0b00, 0b1, 0b0, "bic",
3786 PatFrag<(ops node:$lhs, node:$rhs),
3787 (and node:$lhs, (not node:$rhs))>, []>;
3788 defm ORN : logical_sizes<"ORN", 0b01, 0b1, 0b0, "orn",
3789 PatFrag<(ops node:$lhs, node:$rhs),
3790 (or node:$lhs, (not node:$rhs))>, []>;
3791 defm EON : logical_sizes<"EON", 0b10, 0b1, 0b0, "eon",
3792 PatFrag<(ops node:$lhs, node:$rhs),
3793 (xor node:$lhs, (not node:$rhs))>, []>;
3794 defm BICS : logical_sizes<"BICS", 0b11, 0b1, 0b0, "bics",
3795 PatFrag<(ops node:$lhs, node:$rhs),
3796 (and node:$lhs, (not node:$rhs)),
3797 [{ (void)N; return false; }]>,
3800 multiclass tst_shifts<string prefix, bit sf, ValueType ty, RegisterClass GPR> {
3801 let isCommutable = 1, Rd = 0b11111, Defs = [NZCV] in {
3802 def _lsl : A64I_logicalshift<sf, 0b11, 0b00, 0b0,
3804 (ins GPR:$Rn, GPR:$Rm,
3805 !cast<Operand>("lsl_operand_" # ty):$Imm6),
3806 "tst\t$Rn, $Rm, $Imm6",
3807 [(set NZCV, (A64setcc (and ty:$Rn, (shl ty:$Rm,
3808 !cast<Operand>("lsl_operand_" # ty):$Imm6)),
3813 def _lsr : A64I_logicalshift<sf, 0b11, 0b01, 0b0,
3815 (ins GPR:$Rn, GPR:$Rm,
3816 !cast<Operand>("lsr_operand_" # ty):$Imm6),
3817 "tst\t$Rn, $Rm, $Imm6",
3818 [(set NZCV, (A64setcc (and ty:$Rn, (srl ty:$Rm,
3819 !cast<Operand>("lsr_operand_" # ty):$Imm6)),
3823 def _asr : A64I_logicalshift<sf, 0b11, 0b10, 0b0,
3825 (ins GPR:$Rn, GPR:$Rm,
3826 !cast<Operand>("asr_operand_" # ty):$Imm6),
3827 "tst\t$Rn, $Rm, $Imm6",
3828 [(set NZCV, (A64setcc (and ty:$Rn, (sra ty:$Rm,
3829 !cast<Operand>("asr_operand_" # ty):$Imm6)),
3833 def _ror : A64I_logicalshift<sf, 0b11, 0b11, 0b0,
3835 (ins GPR:$Rn, GPR:$Rm,
3836 !cast<Operand>("ror_operand_" # ty):$Imm6),
3837 "tst\t$Rn, $Rm, $Imm6",
3838 [(set NZCV, (A64setcc (and ty:$Rn, (rotr ty:$Rm,
3839 !cast<Operand>("ror_operand_" # ty):$Imm6)),
3844 def _noshift : InstAlias<"tst $Rn, $Rm",
3845 (!cast<Instruction>(prefix # "_lsl") GPR:$Rn, GPR:$Rm, 0)>;
3847 def : Pat<(A64setcc (and ty:$Rn, ty:$Rm), 0, signed_cond),
3848 (!cast<Instruction>(prefix # "_lsl") $Rn, $Rm, 0)>;
3851 defm TSTxx : tst_shifts<"TSTxx", 0b1, i64, GPR64>;
3852 defm TSTww : tst_shifts<"TSTww", 0b0, i32, GPR32>;
3855 multiclass mvn_shifts<string prefix, bit sf, ValueType ty, RegisterClass GPR> {
3856 let isCommutable = 0, Rn = 0b11111 in {
3857 def _lsl : A64I_logicalshift<sf, 0b01, 0b00, 0b1,
3860 !cast<Operand>("lsl_operand_" # ty):$Imm6),
3861 "mvn\t$Rd, $Rm, $Imm6",
3862 [(set ty:$Rd, (not (shl ty:$Rm,
3863 !cast<Operand>("lsl_operand_" # ty):$Imm6)))],
3867 def _lsr : A64I_logicalshift<sf, 0b01, 0b01, 0b1,
3870 !cast<Operand>("lsr_operand_" # ty):$Imm6),
3871 "mvn\t$Rd, $Rm, $Imm6",
3872 [(set ty:$Rd, (not (srl ty:$Rm,
3873 !cast<Operand>("lsr_operand_" # ty):$Imm6)))],
3876 def _asr : A64I_logicalshift<sf, 0b01, 0b10, 0b1,
3879 !cast<Operand>("asr_operand_" # ty):$Imm6),
3880 "mvn\t$Rd, $Rm, $Imm6",
3881 [(set ty:$Rd, (not (sra ty:$Rm,
3882 !cast<Operand>("asr_operand_" # ty):$Imm6)))],
3885 def _ror : A64I_logicalshift<sf, 0b01, 0b11, 0b1,
3888 !cast<Operand>("ror_operand_" # ty):$Imm6),
3889 "mvn\t$Rd, $Rm, $Imm6",
3890 [(set ty:$Rd, (not (rotr ty:$Rm,
3891 !cast<Operand>("lsl_operand_" # ty):$Imm6)))],
3895 def _noshift : InstAlias<"mvn $Rn, $Rm",
3896 (!cast<Instruction>(prefix # "_lsl") GPR:$Rn, GPR:$Rm, 0)>;
3898 def : Pat<(not ty:$Rm),
3899 (!cast<Instruction>(prefix # "_lsl") $Rm, 0)>;
3902 defm MVNxx : mvn_shifts<"MVNxx", 0b1, i64, GPR64>;
3903 defm MVNww : mvn_shifts<"MVNww", 0b0, i32, GPR32>;
3905 def MOVxx :InstAlias<"mov $Rd, $Rm", (ORRxxx_lsl GPR64:$Rd, XZR, GPR64:$Rm, 0)>;
3906 def MOVww :InstAlias<"mov $Rd, $Rm", (ORRwww_lsl GPR32:$Rd, WZR, GPR32:$Rm, 0)>;
3908 //===----------------------------------------------------------------------===//
3909 // Move wide (immediate) instructions
3910 //===----------------------------------------------------------------------===//
3911 // Contains: MOVN, MOVZ, MOVK + MOV aliases
3913 // A wide variety of different relocations are needed for variants of these
3914 // instructions, so it turns out that we need a different operand for all of
3916 multiclass movw_operands<string prefix, string instname, int width> {
3917 def _imm_asmoperand : AsmOperandClass {
3918 let Name = instname # width # "Shifted" # shift;
3919 let PredicateMethod = "is" # instname # width # "Imm";
3920 let RenderMethod = "addMoveWideImmOperands";
3921 let ParserMethod = "ParseImmWithLSLOperand";
3922 let DiagnosticType = "MOVWUImm16";
3925 def _imm : Operand<i64> {
3926 let ParserMatchClass = !cast<AsmOperandClass>(prefix # "_imm_asmoperand");
3927 let PrintMethod = "printMoveWideImmOperand";
3928 let EncoderMethod = "getMoveWideImmOpValue";
3929 let DecoderMethod = "DecodeMoveWideImmOperand<" # width # ">";
3931 let MIOperandInfo = (ops uimm16:$UImm16, imm:$Shift);
3935 defm movn32 : movw_operands<"movn32", "MOVN", 32>;
3936 defm movn64 : movw_operands<"movn64", "MOVN", 64>;
3937 defm movz32 : movw_operands<"movz32", "MOVZ", 32>;
3938 defm movz64 : movw_operands<"movz64", "MOVZ", 64>;
3939 defm movk32 : movw_operands<"movk32", "MOVK", 32>;
3940 defm movk64 : movw_operands<"movk64", "MOVK", 64>;
3942 multiclass A64I_movwSizes<bits<2> opc, string asmop, dag ins32bit,
3945 def wii : A64I_movw<0b0, opc, (outs GPR32:$Rd), ins32bit,
3946 !strconcat(asmop, "\t$Rd, $FullImm"),
3949 let UImm16 = FullImm{15-0};
3950 let Shift = FullImm{17-16};
3953 def xii : A64I_movw<0b1, opc, (outs GPR64:$Rd), ins64bit,
3954 !strconcat(asmop, "\t$Rd, $FullImm"),
3957 let UImm16 = FullImm{15-0};
3958 let Shift = FullImm{17-16};
3962 let isMoveImm = 1, isReMaterializable = 1,
3963 isAsCheapAsAMove = 1, hasSideEffects = 0 in {
3964 defm MOVN : A64I_movwSizes<0b00, "movn",
3965 (ins movn32_imm:$FullImm),
3966 (ins movn64_imm:$FullImm)>;
3968 // Some relocations are able to convert between a MOVZ and a MOVN. If these
3969 // are applied the instruction must be emitted with the corresponding bits as
3970 // 0, which means a MOVZ needs to override that bit from the default.
3971 let PostEncoderMethod = "fixMOVZ" in
3972 defm MOVZ : A64I_movwSizes<0b10, "movz",
3973 (ins movz32_imm:$FullImm),
3974 (ins movz64_imm:$FullImm)>;
3977 let Constraints = "$src = $Rd" in
3978 defm MOVK : A64I_movwSizes<0b11, "movk",
3979 (ins GPR32:$src, movk32_imm:$FullImm),
3980 (ins GPR64:$src, movk64_imm:$FullImm)>;
3983 // And now the "MOV" aliases. These also need their own operands because what
3984 // they accept is completely different to what the base instructions accept.
3985 multiclass movalias_operand<string prefix, string basename,
3986 string immpredicate, int width> {
3987 def _asmoperand : AsmOperandClass {
3988 let Name = basename # width # "MovAlias";
3990 = "isMoveWideMovAlias<" # width # ", A64Imms::" # immpredicate # ">";
3992 = "addMoveWideMovAliasOperands<" # width # ", "
3993 # "A64Imms::" # immpredicate # ">";
3996 def _movimm : Operand<i64> {
3997 let ParserMatchClass = !cast<AsmOperandClass>(prefix # "_asmoperand");
3999 let MIOperandInfo = (ops uimm16:$UImm16, imm:$Shift);
4003 defm movz32 : movalias_operand<"movz32", "MOVZ", "isMOVZImm", 32>;
4004 defm movz64 : movalias_operand<"movz64", "MOVZ", "isMOVZImm", 64>;
4005 defm movn32 : movalias_operand<"movn32", "MOVN", "isOnlyMOVNImm", 32>;
4006 defm movn64 : movalias_operand<"movn64", "MOVN", "isOnlyMOVNImm", 64>;
4008 // FIXME: these are officially canonical aliases, but TableGen is too limited to
4009 // print them at the moment. I believe in this case an "AliasPredicate" method
4010 // will need to be implemented. to allow it, as well as the more generally
4011 // useful handling of non-register, non-constant operands.
4012 class movalias<Instruction INST, RegisterClass GPR, Operand operand>
4013 : InstAlias<"mov $Rd, $FullImm", (INST GPR:$Rd, operand:$FullImm)>;
4015 def : movalias<MOVZwii, GPR32, movz32_movimm>;
4016 def : movalias<MOVZxii, GPR64, movz64_movimm>;
4017 def : movalias<MOVNwii, GPR32, movn32_movimm>;
4018 def : movalias<MOVNxii, GPR64, movn64_movimm>;
4020 def movw_addressref_g0 : ComplexPattern<i64, 2, "SelectMOVWAddressRef<0>">;
4021 def movw_addressref_g1 : ComplexPattern<i64, 2, "SelectMOVWAddressRef<1>">;
4022 def movw_addressref_g2 : ComplexPattern<i64, 2, "SelectMOVWAddressRef<2>">;
4023 def movw_addressref_g3 : ComplexPattern<i64, 2, "SelectMOVWAddressRef<3>">;
4025 def : Pat<(A64WrapperLarge movw_addressref_g3:$G3, movw_addressref_g2:$G2,
4026 movw_addressref_g1:$G1, movw_addressref_g0:$G0),
4027 (MOVKxii (MOVKxii (MOVKxii (MOVZxii movw_addressref_g3:$G3),
4028 movw_addressref_g2:$G2),
4029 movw_addressref_g1:$G1),
4030 movw_addressref_g0:$G0)>;
4032 //===----------------------------------------------------------------------===//
4033 // PC-relative addressing instructions
4034 //===----------------------------------------------------------------------===//
4035 // Contains: ADR, ADRP
4037 def adr_label : Operand<i64> {
4038 let EncoderMethod = "getLabelOpValue<AArch64::fixup_a64_adr_prel>";
4040 // This label is a 21-bit offset from PC, unscaled
4041 let PrintMethod = "printLabelOperand<21, 1>";
4042 let ParserMatchClass = label_asmoperand<21, 1>;
4043 let OperandType = "OPERAND_PCREL";
4046 def adrp_label_asmoperand : AsmOperandClass {
4047 let Name = "AdrpLabel";
4048 let RenderMethod = "addLabelOperands<21, 4096>";
4049 let DiagnosticType = "Label";
4052 def adrp_label : Operand<i64> {
4053 let EncoderMethod = "getAdrpLabelOpValue";
4055 // This label is a 21-bit offset from PC, scaled by the page-size: 4096.
4056 let PrintMethod = "printLabelOperand<21, 4096>";
4057 let ParserMatchClass = adrp_label_asmoperand;
4058 let OperandType = "OPERAND_PCREL";
4061 let hasSideEffects = 0 in {
4062 def ADRxi : A64I_PCADR<0b0, (outs GPR64:$Rd), (ins adr_label:$Label),
4063 "adr\t$Rd, $Label", [], NoItinerary>;
4065 def ADRPxi : A64I_PCADR<0b1, (outs GPR64:$Rd), (ins adrp_label:$Label),
4066 "adrp\t$Rd, $Label", [], NoItinerary>;
4069 //===----------------------------------------------------------------------===//
4070 // System instructions
4071 //===----------------------------------------------------------------------===//
4072 // Contains: HINT, CLREX, DSB, DMB, ISB, MSR, SYS, SYSL, MRS
4073 // + aliases IC, DC, AT, TLBI, NOP, YIELD, WFE, WFI, SEV, SEVL
4075 // Op1 and Op2 fields are sometimes simple 3-bit unsigned immediate values.
4076 def uimm3_asmoperand : AsmOperandClass {
4078 let PredicateMethod = "isUImm<3>";
4079 let RenderMethod = "addImmOperands";
4080 let DiagnosticType = "UImm3";
4083 def uimm3 : Operand<i32> {
4084 let ParserMatchClass = uimm3_asmoperand;
4087 // The HINT alias can accept a simple unsigned 7-bit immediate.
4088 def uimm7_asmoperand : AsmOperandClass {
4090 let PredicateMethod = "isUImm<7>";
4091 let RenderMethod = "addImmOperands";
4092 let DiagnosticType = "UImm7";
4095 def uimm7 : Operand<i32> {
4096 let ParserMatchClass = uimm7_asmoperand;
4099 // Multiclass namedimm is defined with the prefetch operands. Most of these fit
4100 // into the NamedImmMapper scheme well: they either accept a named operand or
4101 // any immediate under a particular value (which may be 0, implying no immediate
4103 defm dbarrier : namedimm<"dbarrier", "A64DB::DBarrierMapper">;
4104 defm isb : namedimm<"isb", "A64ISB::ISBMapper">;
4105 defm ic : namedimm<"ic", "A64IC::ICMapper">;
4106 defm dc : namedimm<"dc", "A64DC::DCMapper">;
4107 defm at : namedimm<"at", "A64AT::ATMapper">;
4108 defm tlbi : namedimm<"tlbi", "A64TLBI::TLBIMapper">;
4110 // However, MRS and MSR are more complicated for a few reasons:
4111 // * There are ~1000 generic names S3_<op1>_<CRn>_<CRm>_<Op2> which have an
4112 // implementation-defined effect
4113 // * Most registers are shared, but some are read-only or write-only.
4114 // * There is a variant of MSR which accepts the same register name (SPSel),
4115 // but which would have a different encoding.
4117 // In principle these could be resolved in with more complicated subclasses of
4118 // NamedImmMapper, however that imposes an overhead on other "named
4119 // immediates". Both in concrete terms with virtual tables and in unnecessary
4122 // The solution adopted here is to take the MRS/MSR Mappers out of the usual
4123 // hierarchy (they're not derived from NamedImmMapper) and to add logic for
4124 // their special situation.
4125 def mrs_asmoperand : AsmOperandClass {
4127 let ParserMethod = "ParseSysRegOperand";
4128 let DiagnosticType = "MRS";
4131 def mrs_op : Operand<i32> {
4132 let ParserMatchClass = mrs_asmoperand;
4133 let PrintMethod = "printMRSOperand";
4134 let DecoderMethod = "DecodeMRSOperand";
4137 def msr_asmoperand : AsmOperandClass {
4138 let Name = "MSRWithReg";
4140 // Note that SPSel is valid for both this and the pstate operands, but with
4141 // different immediate encodings. This is why these operands provide a string
4142 // AArch64Operand rather than an immediate. The overlap is small enough that
4143 // it could be resolved with hackery now, but who can say in future?
4144 let ParserMethod = "ParseSysRegOperand";
4145 let DiagnosticType = "MSR";
4148 def msr_op : Operand<i32> {
4149 let ParserMatchClass = msr_asmoperand;
4150 let PrintMethod = "printMSROperand";
4151 let DecoderMethod = "DecodeMSROperand";
4154 def pstate_asmoperand : AsmOperandClass {
4155 let Name = "MSRPState";
4156 // See comment above about parser.
4157 let ParserMethod = "ParseSysRegOperand";
4158 let DiagnosticType = "MSR";
4161 def pstate_op : Operand<i32> {
4162 let ParserMatchClass = pstate_asmoperand;
4163 let PrintMethod = "printNamedImmOperand<A64PState::PStateMapper>";
4164 let DecoderMethod = "DecodeNamedImmOperand<A64PState::PStateMapper>";
4167 // When <CRn> is specified, an assembler should accept something like "C4", not
4168 // the usual "#4" immediate.
4169 def CRx_asmoperand : AsmOperandClass {
4171 let PredicateMethod = "isUImm<4>";
4172 let RenderMethod = "addImmOperands";
4173 let ParserMethod = "ParseCRxOperand";
4174 // Diagnostics are handled in all cases by ParseCRxOperand.
4177 def CRx : Operand<i32> {
4178 let ParserMatchClass = CRx_asmoperand;
4179 let PrintMethod = "printCRxOperand";
4183 // Finally, we can start defining the instructions.
4185 // HINT is straightforward, with a few aliases.
4186 def HINTi : A64I_system<0b0, (outs), (ins uimm7:$UImm7), "hint\t$UImm7",
4189 let CRm = UImm7{6-3};
4190 let Op2 = UImm7{2-0};
4198 def : InstAlias<"nop", (HINTi 0)>;
4199 def : InstAlias<"yield", (HINTi 1)>;
4200 def : InstAlias<"wfe", (HINTi 2)>;
4201 def : InstAlias<"wfi", (HINTi 3)>;
4202 def : InstAlias<"sev", (HINTi 4)>;
4203 def : InstAlias<"sevl", (HINTi 5)>;
4205 // Quite a few instructions then follow a similar pattern of fixing common
4206 // fields in the bitpattern, we'll define a helper-class for them.
4207 class simple_sys<bits<2> op0, bits<3> op1, bits<4> crn, bits<3> op2,
4208 Operand operand, string asmop>
4209 : A64I_system<0b0, (outs), (ins operand:$CRm), !strconcat(asmop, "\t$CRm"),
4219 def CLREXi : simple_sys<0b00, 0b011, 0b0011, 0b010, uimm4, "clrex">;
4220 def DSBi : simple_sys<0b00, 0b011, 0b0011, 0b100, dbarrier_op, "dsb">;
4221 def DMBi : simple_sys<0b00, 0b011, 0b0011, 0b101, dbarrier_op, "dmb">;
4222 def ISBi : simple_sys<0b00, 0b011, 0b0011, 0b110, isb_op, "isb">;
4224 def : InstAlias<"clrex", (CLREXi 0b1111)>;
4225 def : InstAlias<"isb", (ISBi 0b1111)>;
4227 // (DMBi 0xb) is a "DMB ISH" instruciton, appropriate for Linux SMP
4228 // configurations at least.
4229 def : Pat<(atomic_fence imm, imm), (DMBi 0xb)>;
4231 // Any SYS bitpattern can be represented with a complex and opaque "SYS"
4233 def SYSiccix : A64I_system<0b0, (outs),
4234 (ins uimm3:$Op1, CRx:$CRn, CRx:$CRm,
4235 uimm3:$Op2, GPR64:$Rt),
4236 "sys\t$Op1, $CRn, $CRm, $Op2, $Rt",
4241 // You can skip the Xt argument whether it makes sense or not for the generic
4243 def : InstAlias<"sys $Op1, $CRn, $CRm, $Op2",
4244 (SYSiccix uimm3:$Op1, CRx:$CRn, CRx:$CRm, uimm3:$Op2, XZR)>;
4247 // But many have aliases, which obviously don't fit into
4248 class SYSalias<dag ins, string asmstring>
4249 : A64I_system<0b0, (outs), ins, asmstring, [], NoItinerary> {
4250 let isAsmParserOnly = 1;
4254 let Op1 = SysOp{13-11};
4255 let CRn = SysOp{10-7};
4256 let CRm = SysOp{6-3};
4257 let Op2 = SysOp{2-0};
4260 def ICix : SYSalias<(ins ic_op:$SysOp, GPR64:$Rt), "ic\t$SysOp, $Rt">;
4262 def ICi : SYSalias<(ins ic_op:$SysOp), "ic\t$SysOp"> {
4266 def DCix : SYSalias<(ins dc_op:$SysOp, GPR64:$Rt), "dc\t$SysOp, $Rt">;
4267 def ATix : SYSalias<(ins at_op:$SysOp, GPR64:$Rt), "at\t$SysOp, $Rt">;
4269 def TLBIix : SYSalias<(ins tlbi_op:$SysOp, GPR64:$Rt), "tlbi\t$SysOp, $Rt">;
4271 def TLBIi : SYSalias<(ins tlbi_op:$SysOp), "tlbi\t$SysOp"> {
4276 def SYSLxicci : A64I_system<0b1, (outs GPR64:$Rt),
4277 (ins uimm3:$Op1, CRx:$CRn, CRx:$CRm, uimm3:$Op2),
4278 "sysl\t$Rt, $Op1, $CRn, $CRm, $Op2",
4283 // The instructions themselves are rather simple for MSR and MRS.
4284 def MSRix : A64I_system<0b0, (outs), (ins msr_op:$SysReg, GPR64:$Rt),
4285 "msr\t$SysReg, $Rt", [], NoItinerary> {
4287 let Op0 = SysReg{15-14};
4288 let Op1 = SysReg{13-11};
4289 let CRn = SysReg{10-7};
4290 let CRm = SysReg{6-3};
4291 let Op2 = SysReg{2-0};
4294 def MRSxi : A64I_system<0b1, (outs GPR64:$Rt), (ins mrs_op:$SysReg),
4295 "mrs\t$Rt, $SysReg", [], NoItinerary> {
4297 let Op0 = SysReg{15-14};
4298 let Op1 = SysReg{13-11};
4299 let CRn = SysReg{10-7};
4300 let CRm = SysReg{6-3};
4301 let Op2 = SysReg{2-0};
4304 def MSRii : A64I_system<0b0, (outs), (ins pstate_op:$PState, uimm4:$CRm),
4305 "msr\t$PState, $CRm", [], NoItinerary> {
4309 let Op1 = PState{5-3};
4311 let Op2 = PState{2-0};
4315 //===----------------------------------------------------------------------===//
4316 // Test & branch (immediate) instructions
4317 //===----------------------------------------------------------------------===//
4318 // Contains: TBZ, TBNZ
4320 // The bit to test is a simple unsigned 6-bit immediate in the X-register
4322 def uimm6 : Operand<i64> {
4323 let ParserMatchClass = uimm6_asmoperand;
4326 def label_wid14_scal4_asmoperand : label_asmoperand<14, 4>;
4328 def tbimm_target : Operand<OtherVT> {
4329 let EncoderMethod = "getLabelOpValue<AArch64::fixup_a64_tstbr>";
4331 // This label is a 14-bit offset from PC, scaled by the instruction-width: 4.
4332 let PrintMethod = "printLabelOperand<14, 4>";
4333 let ParserMatchClass = label_wid14_scal4_asmoperand;
4335 let OperandType = "OPERAND_PCREL";
4338 def A64eq : ImmLeaf<i32, [{ return Imm == A64CC::EQ; }]>;
4339 def A64ne : ImmLeaf<i32, [{ return Imm == A64CC::NE; }]>;
4341 // These instructions correspond to patterns involving "and" with a power of
4342 // two, which we need to be able to select.
4343 def tstb64_pat : ComplexPattern<i64, 1, "SelectTSTBOperand<64>">;
4344 def tstb32_pat : ComplexPattern<i32, 1, "SelectTSTBOperand<32>">;
4346 let isBranch = 1, isTerminator = 1 in {
4347 def TBZxii : A64I_TBimm<0b0, (outs),
4348 (ins GPR64:$Rt, uimm6:$Imm, tbimm_target:$Label),
4349 "tbz\t$Rt, $Imm, $Label",
4350 [(A64br_cc (A64cmp (and i64:$Rt, tstb64_pat:$Imm), 0),
4354 def TBNZxii : A64I_TBimm<0b1, (outs),
4355 (ins GPR64:$Rt, uimm6:$Imm, tbimm_target:$Label),
4356 "tbnz\t$Rt, $Imm, $Label",
4357 [(A64br_cc (A64cmp (and i64:$Rt, tstb64_pat:$Imm), 0),
4362 // Note, these instructions overlap with the above 64-bit patterns. This is
4363 // intentional, "tbz x3, #1, somewhere" and "tbz w3, #1, somewhere" would both
4364 // do the same thing and are both permitted assembly. They also both have
4365 // sensible DAG patterns.
4366 def TBZwii : A64I_TBimm<0b0, (outs),
4367 (ins GPR32:$Rt, uimm5:$Imm, tbimm_target:$Label),
4368 "tbz\t$Rt, $Imm, $Label",
4369 [(A64br_cc (A64cmp (and i32:$Rt, tstb32_pat:$Imm), 0),
4375 def TBNZwii : A64I_TBimm<0b1, (outs),
4376 (ins GPR32:$Rt, uimm5:$Imm, tbimm_target:$Label),
4377 "tbnz\t$Rt, $Imm, $Label",
4378 [(A64br_cc (A64cmp (and i32:$Rt, tstb32_pat:$Imm), 0),
4385 //===----------------------------------------------------------------------===//
4386 // Unconditional branch (immediate) instructions
4387 //===----------------------------------------------------------------------===//
4390 def label_wid26_scal4_asmoperand : label_asmoperand<26, 4>;
4392 def bimm_target : Operand<OtherVT> {
4393 let EncoderMethod = "getLabelOpValue<AArch64::fixup_a64_uncondbr>";
4395 // This label is a 26-bit offset from PC, scaled by the instruction-width: 4.
4396 let PrintMethod = "printLabelOperand<26, 4>";
4397 let ParserMatchClass = label_wid26_scal4_asmoperand;
4399 let OperandType = "OPERAND_PCREL";
4402 def blimm_target : Operand<i64> {
4403 let EncoderMethod = "getLabelOpValue<AArch64::fixup_a64_call>";
4405 // This label is a 26-bit offset from PC, scaled by the instruction-width: 4.
4406 let PrintMethod = "printLabelOperand<26, 4>";
4407 let ParserMatchClass = label_wid26_scal4_asmoperand;
4409 let OperandType = "OPERAND_PCREL";
4412 class A64I_BimmImpl<bit op, string asmop, list<dag> patterns, Operand lbl_type>
4413 : A64I_Bimm<op, (outs), (ins lbl_type:$Label),
4414 !strconcat(asmop, "\t$Label"), patterns,
4417 let isBranch = 1 in {
4418 def Bimm : A64I_BimmImpl<0b0, "b", [(br bb:$Label)], bimm_target> {
4419 let isTerminator = 1;
4423 def BLimm : A64I_BimmImpl<0b1, "bl",
4424 [(AArch64Call tglobaladdr:$Label)], blimm_target> {
4430 def : Pat<(AArch64Call texternalsym:$Label), (BLimm texternalsym:$Label)>;
4432 //===----------------------------------------------------------------------===//
4433 // Unconditional branch (register) instructions
4434 //===----------------------------------------------------------------------===//
4435 // Contains: BR, BLR, RET, ERET, DRP.
4437 // Most of the notional opcode fields in the A64I_Breg format are fixed in A64
4439 class A64I_BregImpl<bits<4> opc,
4440 dag outs, dag ins, string asmstr, list<dag> patterns,
4441 InstrItinClass itin = NoItinerary>
4442 : A64I_Breg<opc, 0b11111, 0b000000, 0b00000,
4443 outs, ins, asmstr, patterns, itin> {
4445 let isIndirectBranch = 1;
4448 // Note that these are not marked isCall or isReturn because as far as LLVM is
4449 // concerned they're not. "ret" is just another jump unless it has been selected
4450 // by LLVM as the function's return.
4452 let isBranch = 1 in {
4453 def BRx : A64I_BregImpl<0b0000,(outs), (ins GPR64:$Rn),
4454 "br\t$Rn", [(brind i64:$Rn)]> {
4456 let isTerminator = 1;
4459 def BLRx : A64I_BregImpl<0b0001, (outs), (ins GPR64:$Rn),
4460 "blr\t$Rn", [(AArch64Call i64:$Rn)]> {
4466 def RETx : A64I_BregImpl<0b0010, (outs), (ins GPR64:$Rn),
4469 let isTerminator = 1;
4473 // Create a separate pseudo-instruction for codegen to use so that we don't
4474 // flag x30 as used in every function. It'll be restored before the RET by the
4475 // epilogue if it's legitimately used.
4476 def RET : A64PseudoExpand<(outs), (ins), [(A64ret)], (RETx (ops X30))> {
4477 let isTerminator = 1;
4482 def ERET : A64I_BregImpl<0b0100, (outs), (ins), "eret", []> {
4485 let isTerminator = 1;
4489 def DRPS : A64I_BregImpl<0b0101, (outs), (ins), "drps", []> {
4495 def RETAlias : InstAlias<"ret", (RETx X30)>;
4498 //===----------------------------------------------------------------------===//
4499 // Address generation patterns
4500 //===----------------------------------------------------------------------===//
4502 // Primary method of address generation for the small/absolute memory model is
4503 // an ADRP/ADR pair:
4504 // ADRP x0, some_variable
4505 // ADD x0, x0, #:lo12:some_variable
4507 // The load/store elision of the ADD is accomplished when selecting
4508 // addressing-modes. This just mops up the cases where that doesn't work and we
4509 // really need an address in some register.
4511 // This wrapper applies a LO12 modifier to the address. Otherwise we could just
4512 // use the same address.
4514 class ADRP_ADD<SDNode Wrapper, SDNode addrop>
4515 : Pat<(Wrapper addrop:$Hi, addrop:$Lo12, (i32 imm)),
4516 (ADDxxi_lsl0_s (ADRPxi addrop:$Hi), addrop:$Lo12)>;
4518 def : ADRP_ADD<A64WrapperSmall, tblockaddress>;
4519 def : ADRP_ADD<A64WrapperSmall, texternalsym>;
4520 def : ADRP_ADD<A64WrapperSmall, tglobaladdr>;
4521 def : ADRP_ADD<A64WrapperSmall, tglobaltlsaddr>;
4522 def : ADRP_ADD<A64WrapperSmall, tjumptable>;
4524 //===----------------------------------------------------------------------===//
4525 // GOT access patterns
4526 //===----------------------------------------------------------------------===//
4528 class GOTLoadSmall<SDNode addrfrag>
4529 : Pat<(A64GOTLoad (A64WrapperSmall addrfrag:$Hi, addrfrag:$Lo12, 8)),
4530 (LS64_LDR (ADRPxi addrfrag:$Hi), addrfrag:$Lo12)>;
4532 def : GOTLoadSmall<texternalsym>;
4533 def : GOTLoadSmall<tglobaladdr>;
4534 def : GOTLoadSmall<tglobaltlsaddr>;
4536 //===----------------------------------------------------------------------===//
4537 // Tail call handling
4538 //===----------------------------------------------------------------------===//
4540 let isCall = 1, isTerminator = 1, isReturn = 1, isBarrier = 1, Uses = [XSP] in {
4542 : PseudoInst<(outs), (ins i64imm:$dst, i32imm:$FPDiff),
4543 [(AArch64tcret tglobaladdr:$dst, (i32 timm:$FPDiff))]>;
4546 : PseudoInst<(outs), (ins tcGPR64:$dst, i32imm:$FPDiff),
4547 [(AArch64tcret i64:$dst, (i32 timm:$FPDiff))]>;
4550 let isCall = 1, isTerminator = 1, isReturn = 1, isBarrier = 1,
4552 def TAIL_Bimm : A64PseudoExpand<(outs), (ins bimm_target:$Label), [],
4553 (Bimm bimm_target:$Label)>;
4555 def TAIL_BRx : A64PseudoExpand<(outs), (ins tcGPR64:$Rd), [],
4560 def : Pat<(AArch64tcret texternalsym:$dst, (i32 timm:$FPDiff)),
4561 (TC_RETURNdi texternalsym:$dst, imm:$FPDiff)>;
4563 //===----------------------------------------------------------------------===//
4564 // Thread local storage
4565 //===----------------------------------------------------------------------===//
4567 // This is a pseudo-instruction representing the ".tlsdesccall" directive in
4568 // assembly. Its effect is to insert an R_AARCH64_TLSDESC_CALL relocation at the
4569 // current location. It should always be immediately followed by a BLR
4570 // instruction, and is intended solely for relaxation by the linker.
4572 def : Pat<(A64threadpointer), (MRSxi 0xde82)>;
4574 def TLSDESCCALL : PseudoInst<(outs), (ins i64imm:$Lbl), []> {
4575 let hasSideEffects = 1;
4578 def TLSDESC_BLRx : PseudoInst<(outs), (ins GPR64:$Rn, i64imm:$Var),
4579 [(A64tlsdesc_blr i64:$Rn, tglobaltlsaddr:$Var)]> {
4584 def : Pat<(A64tlsdesc_blr i64:$Rn, texternalsym:$Var),
4585 (TLSDESC_BLRx $Rn, texternalsym:$Var)>;
4587 //===----------------------------------------------------------------------===//
4588 // Bitfield patterns
4589 //===----------------------------------------------------------------------===//
4591 def bfi32_lsb_to_immr : SDNodeXForm<imm, [{
4592 return CurDAG->getTargetConstant((32 - N->getZExtValue()) % 32, MVT::i64);
4595 def bfi64_lsb_to_immr : SDNodeXForm<imm, [{
4596 return CurDAG->getTargetConstant((64 - N->getZExtValue()) % 64, MVT::i64);
4599 def bfi_width_to_imms : SDNodeXForm<imm, [{
4600 return CurDAG->getTargetConstant(N->getZExtValue() - 1, MVT::i64);
4604 // The simpler patterns deal with cases where no AND mask is actually needed
4605 // (either all bits are used or the low 32 bits are used).
4606 let AddedComplexity = 10 in {
4608 def : Pat<(A64Bfi i64:$src, i64:$Rn, imm:$ImmR, imm:$ImmS),
4610 (bfi64_lsb_to_immr (i64 imm:$ImmR)),
4611 (bfi_width_to_imms (i64 imm:$ImmS)))>;
4613 def : Pat<(A64Bfi i32:$src, i32:$Rn, imm:$ImmR, imm:$ImmS),
4615 (bfi32_lsb_to_immr (i64 imm:$ImmR)),
4616 (bfi_width_to_imms (i64 imm:$ImmS)))>;
4619 def : Pat<(and (A64Bfi i64:$src, i64:$Rn, imm:$ImmR, imm:$ImmS),
4621 (SUBREG_TO_REG (i64 0),
4622 (BFIwwii (EXTRACT_SUBREG $src, sub_32),
4623 (EXTRACT_SUBREG $Rn, sub_32),
4624 (bfi32_lsb_to_immr (i64 imm:$ImmR)),
4625 (bfi_width_to_imms (i64 imm:$ImmS))),
4630 //===----------------------------------------------------------------------===//
4631 // Miscellaneous patterns
4632 //===----------------------------------------------------------------------===//
4634 // Truncation from 64 to 32-bits just involves renaming your register.
4635 def : Pat<(i32 (trunc i64:$val)), (EXTRACT_SUBREG $val, sub_32)>;
4637 // Similarly, extension where we don't care about the high bits is
4639 def : Pat<(i64 (anyext i32:$val)),
4640 (INSERT_SUBREG (IMPLICIT_DEF), $val, sub_32)>;
4642 // SELECT instructions providing f128 types need to be handled by a
4643 // pseudo-instruction since the eventual code will need to introduce basic
4644 // blocks and control flow.
4645 def F128CSEL : PseudoInst<(outs FPR128:$Rd),
4646 (ins FPR128:$Rn, FPR128:$Rm, cond_code_op:$Cond),
4647 [(set f128:$Rd, (simple_select f128:$Rn, f128:$Rm))]> {
4649 let usesCustomInserter = 1;
4652 //===----------------------------------------------------------------------===//
4653 // Load/store patterns
4654 //===----------------------------------------------------------------------===//
4656 // There are lots of patterns here, because we need to allow at least three
4657 // parameters to vary independently.
4658 // 1. Instruction: "ldrb w9, [sp]", "ldrh w9, [sp]", ...
4659 // 2. LLVM source: zextloadi8, anyextloadi8, ...
4660 // 3. Address-generation: A64Wrapper, (add BASE, OFFSET), ...
4662 // The biggest problem turns out to be the address-generation variable. At the
4663 // point of instantiation we need to produce two DAGs, one for the pattern and
4664 // one for the instruction. Doing this at the lowest level of classes doesn't
4667 // Consider the simple uimm12 addressing mode, and the desire to match both (add
4668 // GPR64xsp:$Rn, uimm12:$Offset) and GPR64xsp:$Rn, particularly on the
4669 // instruction side. We'd need to insert either "GPR64xsp" and "uimm12" or
4670 // "GPR64xsp" and "0" into an unknown dag. !subst is not capable of this
4671 // operation, and PatFrags are for selection not output.
4673 // As a result, the address-generation patterns are the final
4674 // instantiations. However, we do still need to vary the operand for the address
4675 // further down (At the point we're deciding A64WrapperSmall, we don't know
4676 // the memory width of the operation).
4678 //===------------------------------
4679 // 1. Basic infrastructural defs
4680 //===------------------------------
4682 // First, some simple classes for !foreach and !subst to use:
4693 // You can't use !subst on an actual immediate, but you *can* use it on an
4694 // operand record that happens to match a single immediate. So we do.
4695 def imm_eq0 : ImmLeaf<i64, [{ return Imm == 0; }]>;
4696 def imm_eq1 : ImmLeaf<i64, [{ return Imm == 1; }]>;
4697 def imm_eq2 : ImmLeaf<i64, [{ return Imm == 2; }]>;
4698 def imm_eq3 : ImmLeaf<i64, [{ return Imm == 3; }]>;
4699 def imm_eq4 : ImmLeaf<i64, [{ return Imm == 4; }]>;
4701 // If the low bits of a pointer are known to be 0 then an "or" is just as good
4702 // as addition for computing an offset. This fragment forwards that check for
4704 def add_like_or : PatFrag<(ops node:$lhs, node:$rhs), (or node:$lhs, node:$rhs),
4706 return CurDAG->isBaseWithConstantOffset(SDValue(N, 0));
4709 // Load/store (unsigned immediate) operations with relocations against global
4710 // symbols (for lo12) are only valid if those symbols have correct alignment
4711 // (since the immediate offset is divided by the access scale, it can't have a
4714 // The guaranteed alignment is provided as part of the WrapperSmall
4715 // operation, and checked against one of these.
4716 def any_align : ImmLeaf<i32, [{ (void)Imm; return true; }]>;
4717 def min_align2 : ImmLeaf<i32, [{ return Imm >= 2; }]>;
4718 def min_align4 : ImmLeaf<i32, [{ return Imm >= 4; }]>;
4719 def min_align8 : ImmLeaf<i32, [{ return Imm >= 8; }]>;
4720 def min_align16 : ImmLeaf<i32, [{ return Imm >= 16; }]>;
4722 // "Normal" load/store instructions can be used on atomic operations, provided
4723 // the ordering parameter is at most "monotonic". Anything above that needs
4724 // special handling with acquire/release instructions.
4725 class simple_load<PatFrag base>
4726 : PatFrag<(ops node:$ptr), (base node:$ptr), [{
4727 return cast<AtomicSDNode>(N)->getOrdering() <= Monotonic;
4730 def atomic_load_simple_i8 : simple_load<atomic_load_8>;
4731 def atomic_load_simple_i16 : simple_load<atomic_load_16>;
4732 def atomic_load_simple_i32 : simple_load<atomic_load_32>;
4733 def atomic_load_simple_i64 : simple_load<atomic_load_64>;
4735 class simple_store<PatFrag base>
4736 : PatFrag<(ops node:$ptr, node:$val), (base node:$ptr, node:$val), [{
4737 return cast<AtomicSDNode>(N)->getOrdering() <= Monotonic;
4740 def atomic_store_simple_i8 : simple_store<atomic_store_8>;
4741 def atomic_store_simple_i16 : simple_store<atomic_store_16>;
4742 def atomic_store_simple_i32 : simple_store<atomic_store_32>;
4743 def atomic_store_simple_i64 : simple_store<atomic_store_64>;
4745 //===------------------------------
4746 // 2. UImm12 and SImm9
4747 //===------------------------------
4749 // These instructions have two operands providing the address so they can be
4750 // treated similarly for most purposes.
4752 //===------------------------------
4753 // 2.1 Base patterns covering extend/truncate semantics
4754 //===------------------------------
4756 // Atomic patterns can be shared between integer operations of all sizes, a
4757 // quick multiclass here allows reuse.
4758 multiclass ls_atomic_pats<Instruction LOAD, Instruction STORE, dag Base,
4759 dag Offset, dag address, ValueType transty,
4761 def : Pat<(!cast<PatFrag>("atomic_load_simple_" # sty) address),
4762 (LOAD Base, Offset)>;
4764 def : Pat<(!cast<PatFrag>("atomic_store_simple_" # sty) address, transty:$Rt),
4765 (STORE $Rt, Base, Offset)>;
4768 // Instructions accessing a memory chunk smaller than a register (or, in a
4769 // pinch, the same size) have a characteristic set of patterns they want to
4770 // match: extending loads and truncating stores. This class deals with the
4771 // sign-neutral version of those patterns.
4773 // It will be instantiated across multiple addressing-modes.
4774 multiclass ls_small_pats<Instruction LOAD, Instruction STORE,
4775 dag Base, dag Offset,
4776 dag address, ValueType sty>
4777 : ls_atomic_pats<LOAD, STORE, Base, Offset, address, i32, sty> {
4778 def : Pat<(!cast<SDNode>(zextload # sty) address), (LOAD Base, Offset)>;
4780 def : Pat<(!cast<SDNode>(extload # sty) address), (LOAD Base, Offset)>;
4782 // For zero-extension to 64-bits we have to tell LLVM that the whole 64-bit
4783 // register was actually set.
4784 def : Pat<(i64 (!cast<SDNode>(zextload # sty) address)),
4785 (SUBREG_TO_REG (i64 0), (LOAD Base, Offset), sub_32)>;
4787 def : Pat<(i64 (!cast<SDNode>(extload # sty) address)),
4788 (SUBREG_TO_REG (i64 0), (LOAD Base, Offset), sub_32)>;
4790 def : Pat<(!cast<SDNode>(truncstore # sty) i32:$Rt, address),
4791 (STORE $Rt, Base, Offset)>;
4793 // For truncating store from 64-bits, we have to manually tell LLVM to
4794 // ignore the high bits of the x register.
4795 def : Pat<(!cast<SDNode>(truncstore # sty) i64:$Rt, address),
4796 (STORE (EXTRACT_SUBREG $Rt, sub_32), Base, Offset)>;
4799 // Next come patterns for sign-extending loads.
4800 multiclass load_signed_pats<string T, string U, dag Base, dag Offset,
4801 dag address, ValueType sty> {
4802 def : Pat<(i32 (!cast<SDNode>("sextload" # sty) address)),
4803 (!cast<Instruction>("LDRS" # T # "w" # U) Base, Offset)>;
4805 def : Pat<(i64 (!cast<SDNode>("sextload" # sty) address)),
4806 (!cast<Instruction>("LDRS" # T # "x" # U) Base, Offset)>;
4810 // and finally "natural-width" loads and stores come next.
4811 multiclass ls_neutral_pats<Instruction LOAD, Instruction STORE, dag Base,
4812 dag Offset, dag address, ValueType sty> {
4813 def : Pat<(sty (load address)), (LOAD Base, Offset)>;
4814 def : Pat<(store sty:$Rt, address), (STORE $Rt, Base, Offset)>;
4817 // Integer operations also get atomic instructions to select for.
4818 multiclass ls_int_neutral_pats<Instruction LOAD, Instruction STORE, dag Base,
4819 dag Offset, dag address, ValueType sty>
4820 : ls_neutral_pats<LOAD, STORE, Base, Offset, address, sty>,
4821 ls_atomic_pats<LOAD, STORE, Base, Offset, address, sty, sty>;
4823 //===------------------------------
4824 // 2.2. Addressing-mode instantiations
4825 //===------------------------------
4827 multiclass uimm12_pats<dag address, dag Base, dag Offset> {
4828 defm : ls_small_pats<LS8_LDR, LS8_STR, Base,
4829 !foreach(decls.pattern, Offset,
4830 !subst(OFFSET, byte_uimm12, decls.pattern)),
4831 !foreach(decls.pattern, address,
4832 !subst(OFFSET, byte_uimm12,
4833 !subst(ALIGN, any_align, decls.pattern))),
4835 defm : ls_small_pats<LS16_LDR, LS16_STR, Base,
4836 !foreach(decls.pattern, Offset,
4837 !subst(OFFSET, hword_uimm12, decls.pattern)),
4838 !foreach(decls.pattern, address,
4839 !subst(OFFSET, hword_uimm12,
4840 !subst(ALIGN, min_align2, decls.pattern))),
4842 defm : ls_small_pats<LS32_LDR, LS32_STR, Base,
4843 !foreach(decls.pattern, Offset,
4844 !subst(OFFSET, word_uimm12, decls.pattern)),
4845 !foreach(decls.pattern, address,
4846 !subst(OFFSET, word_uimm12,
4847 !subst(ALIGN, min_align4, decls.pattern))),
4850 defm : ls_int_neutral_pats<LS32_LDR, LS32_STR, Base,
4851 !foreach(decls.pattern, Offset,
4852 !subst(OFFSET, word_uimm12, decls.pattern)),
4853 !foreach(decls.pattern, address,
4854 !subst(OFFSET, word_uimm12,
4855 !subst(ALIGN, min_align4, decls.pattern))),
4858 defm : ls_int_neutral_pats<LS64_LDR, LS64_STR, Base,
4859 !foreach(decls.pattern, Offset,
4860 !subst(OFFSET, dword_uimm12, decls.pattern)),
4861 !foreach(decls.pattern, address,
4862 !subst(OFFSET, dword_uimm12,
4863 !subst(ALIGN, min_align8, decls.pattern))),
4866 defm : ls_neutral_pats<LSFP16_LDR, LSFP16_STR, Base,
4867 !foreach(decls.pattern, Offset,
4868 !subst(OFFSET, hword_uimm12, decls.pattern)),
4869 !foreach(decls.pattern, address,
4870 !subst(OFFSET, hword_uimm12,
4871 !subst(ALIGN, min_align2, decls.pattern))),
4874 defm : ls_neutral_pats<LSFP32_LDR, LSFP32_STR, Base,
4875 !foreach(decls.pattern, Offset,
4876 !subst(OFFSET, word_uimm12, decls.pattern)),
4877 !foreach(decls.pattern, address,
4878 !subst(OFFSET, word_uimm12,
4879 !subst(ALIGN, min_align4, decls.pattern))),
4882 defm : ls_neutral_pats<LSFP64_LDR, LSFP64_STR, Base,
4883 !foreach(decls.pattern, Offset,
4884 !subst(OFFSET, dword_uimm12, decls.pattern)),
4885 !foreach(decls.pattern, address,
4886 !subst(OFFSET, dword_uimm12,
4887 !subst(ALIGN, min_align8, decls.pattern))),
4890 defm : ls_neutral_pats<LSFP128_LDR, LSFP128_STR, Base,
4891 !foreach(decls.pattern, Offset,
4892 !subst(OFFSET, qword_uimm12, decls.pattern)),
4893 !foreach(decls.pattern, address,
4894 !subst(OFFSET, qword_uimm12,
4895 !subst(ALIGN, min_align16, decls.pattern))),
4898 defm : load_signed_pats<"B", "", Base,
4899 !foreach(decls.pattern, Offset,
4900 !subst(OFFSET, byte_uimm12, decls.pattern)),
4901 !foreach(decls.pattern, address,
4902 !subst(OFFSET, byte_uimm12,
4903 !subst(ALIGN, any_align, decls.pattern))),
4906 defm : load_signed_pats<"H", "", Base,
4907 !foreach(decls.pattern, Offset,
4908 !subst(OFFSET, hword_uimm12, decls.pattern)),
4909 !foreach(decls.pattern, address,
4910 !subst(OFFSET, hword_uimm12,
4911 !subst(ALIGN, min_align2, decls.pattern))),
4914 def : Pat<(sextloadi32 !foreach(decls.pattern, address,
4915 !subst(OFFSET, word_uimm12,
4916 !subst(ALIGN, min_align4, decls.pattern)))),
4917 (LDRSWx Base, !foreach(decls.pattern, Offset,
4918 !subst(OFFSET, word_uimm12, decls.pattern)))>;
4921 // Straightforward patterns of last resort: a pointer with or without an
4922 // appropriate offset.
4923 defm : uimm12_pats<(i64 i64:$Rn), (i64 i64:$Rn), (i64 0)>;
4924 defm : uimm12_pats<(add i64:$Rn, OFFSET:$UImm12),
4925 (i64 i64:$Rn), (i64 OFFSET:$UImm12)>;
4927 // The offset could be hidden behind an "or", of course:
4928 defm : uimm12_pats<(add_like_or i64:$Rn, OFFSET:$UImm12),
4929 (i64 i64:$Rn), (i64 OFFSET:$UImm12)>;
4931 // Global addresses under the small-absolute model should use these
4932 // instructions. There are ELF relocations specifically for it.
4933 defm : uimm12_pats<(A64WrapperSmall tglobaladdr:$Hi, tglobaladdr:$Lo12, ALIGN),
4934 (ADRPxi tglobaladdr:$Hi), (i64 tglobaladdr:$Lo12)>;
4936 defm : uimm12_pats<(A64WrapperSmall tglobaltlsaddr:$Hi, tglobaltlsaddr:$Lo12,
4938 (ADRPxi tglobaltlsaddr:$Hi), (i64 tglobaltlsaddr:$Lo12)>;
4940 // External symbols that make it this far should also get standard relocations.
4941 defm : uimm12_pats<(A64WrapperSmall texternalsym:$Hi, texternalsym:$Lo12,
4943 (ADRPxi texternalsym:$Hi), (i64 texternalsym:$Lo12)>;
4945 defm : uimm12_pats<(A64WrapperSmall tconstpool:$Hi, tconstpool:$Lo12, ALIGN),
4946 (ADRPxi tconstpool:$Hi), (i64 tconstpool:$Lo12)>;
4948 // We also want to use uimm12 instructions for local variables at the moment.
4949 def tframeindex_XFORM : SDNodeXForm<frameindex, [{
4950 int FI = cast<FrameIndexSDNode>(N)->getIndex();
4951 return CurDAG->getTargetFrameIndex(FI, MVT::i64);
4954 defm : uimm12_pats<(i64 frameindex:$Rn),
4955 (tframeindex_XFORM tframeindex:$Rn), (i64 0)>;
4957 // These can be much simpler than uimm12 because we don't to change the operand
4958 // type (e.g. LDURB and LDURH take the same operands).
4959 multiclass simm9_pats<dag address, dag Base, dag Offset> {
4960 defm : ls_small_pats<LS8_LDUR, LS8_STUR, Base, Offset, address, i8>;
4961 defm : ls_small_pats<LS16_LDUR, LS16_STUR, Base, Offset, address, i16>;
4963 defm : ls_int_neutral_pats<LS32_LDUR, LS32_STUR, Base, Offset, address, i32>;
4964 defm : ls_int_neutral_pats<LS64_LDUR, LS64_STUR, Base, Offset, address, i64>;
4966 defm : ls_neutral_pats<LSFP16_LDUR, LSFP16_STUR, Base, Offset, address, f16>;
4967 defm : ls_neutral_pats<LSFP32_LDUR, LSFP32_STUR, Base, Offset, address, f32>;
4968 defm : ls_neutral_pats<LSFP64_LDUR, LSFP64_STUR, Base, Offset, address, f64>;
4969 defm : ls_neutral_pats<LSFP128_LDUR, LSFP128_STUR, Base, Offset, address,
4972 def : Pat<(i64 (zextloadi32 address)),
4973 (SUBREG_TO_REG (i64 0), (LS32_LDUR Base, Offset), sub_32)>;
4975 def : Pat<(truncstorei32 i64:$Rt, address),
4976 (LS32_STUR (EXTRACT_SUBREG $Rt, sub_32), Base, Offset)>;
4978 defm : load_signed_pats<"B", "_U", Base, Offset, address, i8>;
4979 defm : load_signed_pats<"H", "_U", Base, Offset, address, i16>;
4980 def : Pat<(sextloadi32 address), (LDURSWx Base, Offset)>;
4983 defm : simm9_pats<(add i64:$Rn, simm9:$SImm9),
4984 (i64 $Rn), (SDXF_simm9 simm9:$SImm9)>;
4986 defm : simm9_pats<(add_like_or i64:$Rn, simm9:$SImm9),
4987 (i64 $Rn), (SDXF_simm9 simm9:$SImm9)>;
4990 //===------------------------------
4991 // 3. Register offset patterns
4992 //===------------------------------
4994 // Atomic patterns can be shared between integer operations of all sizes, a
4995 // quick multiclass here allows reuse.
4996 multiclass ro_atomic_pats<Instruction LOAD, Instruction STORE, dag Base,
4997 dag Offset, dag Extend, dag address,
4998 ValueType transty, ValueType sty> {
4999 def : Pat<(!cast<PatFrag>("atomic_load_simple_" # sty) address),
5000 (LOAD Base, Offset, Extend)>;
5002 def : Pat<(!cast<PatFrag>("atomic_store_simple_" # sty) address, transty:$Rt),
5003 (STORE $Rt, Base, Offset, Extend)>;
5006 // The register offset instructions take three operands giving the instruction,
5007 // and have an annoying split between instructions where Rm is 32-bit and
5008 // 64-bit. So we need a special hierarchy to describe them. Other than that the
5009 // same operations should be supported as for simm9 and uimm12 addressing.
5011 multiclass ro_small_pats<Instruction LOAD, Instruction STORE,
5012 dag Base, dag Offset, dag Extend,
5013 dag address, ValueType sty>
5014 : ro_atomic_pats<LOAD, STORE, Base, Offset, Extend, address, i32, sty> {
5015 def : Pat<(!cast<SDNode>(zextload # sty) address),
5016 (LOAD Base, Offset, Extend)>;
5018 def : Pat<(!cast<SDNode>(extload # sty) address),
5019 (LOAD Base, Offset, Extend)>;
5021 // For zero-extension to 64-bits we have to tell LLVM that the whole 64-bit
5022 // register was actually set.
5023 def : Pat<(i64 (!cast<SDNode>(zextload # sty) address)),
5024 (SUBREG_TO_REG (i64 0), (LOAD Base, Offset, Extend), sub_32)>;
5026 def : Pat<(i64 (!cast<SDNode>(extload # sty) address)),
5027 (SUBREG_TO_REG (i64 0), (LOAD Base, Offset, Extend), sub_32)>;
5029 def : Pat<(!cast<SDNode>(truncstore # sty) i32:$Rt, address),
5030 (STORE $Rt, Base, Offset, Extend)>;
5032 // For truncating store from 64-bits, we have to manually tell LLVM to
5033 // ignore the high bits of the x register.
5034 def : Pat<(!cast<SDNode>(truncstore # sty) i64:$Rt, address),
5035 (STORE (EXTRACT_SUBREG $Rt, sub_32), Base, Offset, Extend)>;
5039 // Next come patterns for sign-extending loads.
5040 multiclass ro_signed_pats<string T, string Rm, dag Base, dag Offset, dag Extend,
5041 dag address, ValueType sty> {
5042 def : Pat<(i32 (!cast<SDNode>("sextload" # sty) address)),
5043 (!cast<Instruction>("LDRS" # T # "w_" # Rm # "_RegOffset")
5044 Base, Offset, Extend)>;
5046 def : Pat<(i64 (!cast<SDNode>("sextload" # sty) address)),
5047 (!cast<Instruction>("LDRS" # T # "x_" # Rm # "_RegOffset")
5048 Base, Offset, Extend)>;
5051 // and finally "natural-width" loads and stores come next.
5052 multiclass ro_neutral_pats<Instruction LOAD, Instruction STORE,
5053 dag Base, dag Offset, dag Extend, dag address,
5055 def : Pat<(sty (load address)), (LOAD Base, Offset, Extend)>;
5056 def : Pat<(store sty:$Rt, address),
5057 (STORE $Rt, Base, Offset, Extend)>;
5060 multiclass ro_int_neutral_pats<Instruction LOAD, Instruction STORE,
5061 dag Base, dag Offset, dag Extend, dag address,
5063 : ro_neutral_pats<LOAD, STORE, Base, Offset, Extend, address, sty>,
5064 ro_atomic_pats<LOAD, STORE, Base, Offset, Extend, address, sty, sty>;
5066 multiclass regoff_pats<string Rm, dag address, dag Base, dag Offset,
5068 defm : ro_small_pats<!cast<Instruction>("LS8_" # Rm # "_RegOffset_LDR"),
5069 !cast<Instruction>("LS8_" # Rm # "_RegOffset_STR"),
5070 Base, Offset, Extend,
5071 !foreach(decls.pattern, address,
5072 !subst(SHIFT, imm_eq0, decls.pattern)),
5074 defm : ro_small_pats<!cast<Instruction>("LS16_" # Rm # "_RegOffset_LDR"),
5075 !cast<Instruction>("LS16_" # Rm # "_RegOffset_STR"),
5076 Base, Offset, Extend,
5077 !foreach(decls.pattern, address,
5078 !subst(SHIFT, imm_eq1, decls.pattern)),
5080 defm : ro_small_pats<!cast<Instruction>("LS32_" # Rm # "_RegOffset_LDR"),
5081 !cast<Instruction>("LS32_" # Rm # "_RegOffset_STR"),
5082 Base, Offset, Extend,
5083 !foreach(decls.pattern, address,
5084 !subst(SHIFT, imm_eq2, decls.pattern)),
5087 defm : ro_int_neutral_pats<
5088 !cast<Instruction>("LS32_" # Rm # "_RegOffset_LDR"),
5089 !cast<Instruction>("LS32_" # Rm # "_RegOffset_STR"),
5090 Base, Offset, Extend,
5091 !foreach(decls.pattern, address,
5092 !subst(SHIFT, imm_eq2, decls.pattern)),
5095 defm : ro_int_neutral_pats<
5096 !cast<Instruction>("LS64_" # Rm # "_RegOffset_LDR"),
5097 !cast<Instruction>("LS64_" # Rm # "_RegOffset_STR"),
5098 Base, Offset, Extend,
5099 !foreach(decls.pattern, address,
5100 !subst(SHIFT, imm_eq3, decls.pattern)),
5103 defm : ro_neutral_pats<!cast<Instruction>("LSFP16_" # Rm # "_RegOffset_LDR"),
5104 !cast<Instruction>("LSFP16_" # Rm # "_RegOffset_STR"),
5105 Base, Offset, Extend,
5106 !foreach(decls.pattern, address,
5107 !subst(SHIFT, imm_eq1, decls.pattern)),
5110 defm : ro_neutral_pats<!cast<Instruction>("LSFP32_" # Rm # "_RegOffset_LDR"),
5111 !cast<Instruction>("LSFP32_" # Rm # "_RegOffset_STR"),
5112 Base, Offset, Extend,
5113 !foreach(decls.pattern, address,
5114 !subst(SHIFT, imm_eq2, decls.pattern)),
5117 defm : ro_neutral_pats<!cast<Instruction>("LSFP64_" # Rm # "_RegOffset_LDR"),
5118 !cast<Instruction>("LSFP64_" # Rm # "_RegOffset_STR"),
5119 Base, Offset, Extend,
5120 !foreach(decls.pattern, address,
5121 !subst(SHIFT, imm_eq3, decls.pattern)),
5124 defm : ro_neutral_pats<!cast<Instruction>("LSFP128_" # Rm # "_RegOffset_LDR"),
5125 !cast<Instruction>("LSFP128_" # Rm # "_RegOffset_STR"),
5126 Base, Offset, Extend,
5127 !foreach(decls.pattern, address,
5128 !subst(SHIFT, imm_eq4, decls.pattern)),
5131 defm : ro_signed_pats<"B", Rm, Base, Offset, Extend,
5132 !foreach(decls.pattern, address,
5133 !subst(SHIFT, imm_eq0, decls.pattern)),
5136 defm : ro_signed_pats<"H", Rm, Base, Offset, Extend,
5137 !foreach(decls.pattern, address,
5138 !subst(SHIFT, imm_eq1, decls.pattern)),
5141 def : Pat<(sextloadi32 !foreach(decls.pattern, address,
5142 !subst(SHIFT, imm_eq2, decls.pattern))),
5143 (!cast<Instruction>("LDRSWx_" # Rm # "_RegOffset")
5144 Base, Offset, Extend)>;
5148 // Finally we're in a position to tell LLVM exactly what addresses are reachable
5149 // using register-offset instructions. Essentially a base plus a possibly
5150 // extended, possibly shifted (by access size) offset.
5152 defm : regoff_pats<"Wm", (add i64:$Rn, (sext i32:$Rm)),
5153 (i64 i64:$Rn), (i32 i32:$Rm), (i64 6)>;
5155 defm : regoff_pats<"Wm", (add i64:$Rn, (shl (sext i32:$Rm), SHIFT)),
5156 (i64 i64:$Rn), (i32 i32:$Rm), (i64 7)>;
5158 defm : regoff_pats<"Wm", (add i64:$Rn, (zext i32:$Rm)),
5159 (i64 i64:$Rn), (i32 i32:$Rm), (i64 2)>;
5161 defm : regoff_pats<"Wm", (add i64:$Rn, (shl (zext i32:$Rm), SHIFT)),
5162 (i64 i64:$Rn), (i32 i32:$Rm), (i64 3)>;
5164 defm : regoff_pats<"Xm", (add i64:$Rn, i64:$Rm),
5165 (i64 i64:$Rn), (i64 i64:$Rm), (i64 2)>;
5167 defm : regoff_pats<"Xm", (add i64:$Rn, (shl i64:$Rm, SHIFT)),
5168 (i64 i64:$Rn), (i64 i64:$Rm), (i64 3)>;
5170 //===----------------------------------------------------------------------===//
5171 // Advanced SIMD (NEON) Support
5174 include "AArch64InstrNEON.td"