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 HasFPARMv8 : Predicate<"Subtarget->hasFPARMv8()">,
18 AssemblerPredicate<"FeatureFPARMv8", "fp-armv8">;
19 def HasNEON : Predicate<"Subtarget->hasNEON()">,
20 AssemblerPredicate<"FeatureNEON", "neon">;
21 def HasCrypto : Predicate<"Subtarget->hasCrypto()">,
22 AssemblerPredicate<"FeatureCrypto","crypto">;
24 // Use fused MAC if more precision in FP computation is allowed.
25 def UseFusedMAC : Predicate<"(TM.Options.AllowFPOpFusion =="
26 " FPOpFusion::Fast)">;
27 include "AArch64InstrFormats.td"
29 //===----------------------------------------------------------------------===//
30 // AArch64 specific pattern fragments.
32 // An 'fmul' node with a single use.
33 def fmul_su : PatFrag<(ops node:$lhs, node:$rhs), (fmul node:$lhs, node:$rhs),[{
34 return N->hasOneUse();
38 //===----------------------------------------------------------------------===//
39 // Target-specific ISD nodes and profiles
40 //===----------------------------------------------------------------------===//
42 def SDT_A64ret : SDTypeProfile<0, 0, []>;
43 def A64ret : SDNode<"AArch64ISD::Ret", SDT_A64ret, [SDNPHasChain,
47 // (ins NZCV, Condition, Dest)
48 def SDT_A64br_cc : SDTypeProfile<0, 3, [SDTCisVT<0, i32>]>;
49 def A64br_cc : SDNode<"AArch64ISD::BR_CC", SDT_A64br_cc, [SDNPHasChain]>;
51 // (outs Result), (ins NZCV, IfTrue, IfFalse, Condition)
52 def SDT_A64select_cc : SDTypeProfile<1, 4, [SDTCisVT<1, i32>,
55 def A64select_cc : SDNode<"AArch64ISD::SELECT_CC", SDT_A64select_cc>;
57 // (outs NZCV), (ins LHS, RHS, Condition)
58 def SDT_A64setcc : SDTypeProfile<1, 3, [SDTCisVT<0, i32>,
60 def A64setcc : SDNode<"AArch64ISD::SETCC", SDT_A64setcc>;
63 // (outs GPR64), (ins)
64 def A64threadpointer : SDNode<"AArch64ISD::THREAD_POINTER", SDTPtrLeaf>;
66 // A64 compares don't care about the cond really (they set all flags) so a
67 // simple binary operator is useful.
68 def A64cmp : PatFrag<(ops node:$lhs, node:$rhs),
69 (A64setcc node:$lhs, node:$rhs, cond)>;
72 // When matching a notional (CMP op1, (sub 0, op2)), we'd like to use a CMN
73 // instruction on the grounds that "op1 - (-op2) == op1 + op2". However, the C
74 // and V flags can be set differently by this operation. It comes down to
75 // whether "SInt(~op2)+1 == SInt(~op2+1)" (and the same for UInt). If they are
76 // then everything is fine. If not then the optimization is wrong. Thus general
77 // comparisons are only valid if op2 != 0.
79 // So, finally, the only LLVM-native comparisons that don't mention C and V are
80 // SETEQ and SETNE. They're the only ones we can safely use CMN for in the
81 // absence of information about op2.
82 def equality_cond : PatLeaf<(cond), [{
83 return N->get() == ISD::SETEQ || N->get() == ISD::SETNE;
86 def A64cmn : PatFrag<(ops node:$lhs, node:$rhs),
87 (A64setcc node:$lhs, (sub 0, node:$rhs), equality_cond)>;
89 // There are two layers of indirection here, driven by the following
91 // + TableGen does not know CodeModel or Reloc so that decision should be
92 // made for a variable/address at ISelLowering.
93 // + The output of ISelLowering should be selectable (hence the Wrapper,
94 // rather than a bare target opcode)
95 def SDTAArch64WrapperLarge : SDTypeProfile<1, 4, [SDTCisSameAs<0, 1>,
101 def A64WrapperLarge :SDNode<"AArch64ISD::WrapperLarge", SDTAArch64WrapperLarge>;
103 def SDTAArch64WrapperSmall : SDTypeProfile<1, 3, [SDTCisSameAs<0, 1>,
108 def A64WrapperSmall :SDNode<"AArch64ISD::WrapperSmall", SDTAArch64WrapperSmall>;
111 def SDTAArch64GOTLoad : SDTypeProfile<1, 1, [SDTCisPtrTy<0>, SDTCisPtrTy<1>]>;
112 def A64GOTLoad : SDNode<"AArch64ISD::GOTLoad", SDTAArch64GOTLoad,
116 // (A64BFI LHS, RHS, LSB, Width)
117 def SDTA64BFI : SDTypeProfile<1, 4, [SDTCisSameAs<0, 1>,
122 def A64Bfi : SDNode<"AArch64ISD::BFI", SDTA64BFI>;
124 // (A64EXTR HiReg, LoReg, LSB)
125 def SDTA64EXTR : SDTypeProfile<1, 3, [SDTCisSameAs<0, 1>, SDTCisSameAs<1, 2>,
127 def A64Extr : SDNode<"AArch64ISD::EXTR", SDTA64EXTR>;
129 // (A64[SU]BFX Field, ImmR, ImmS).
131 // Note that ImmR and ImmS are already encoded for the actual instructions. The
132 // more natural LSB and Width mix together to form ImmR and ImmS, something
133 // which TableGen can't handle.
134 def SDTA64BFX : SDTypeProfile<1, 3, [SDTCisVT<2, i64>, SDTCisVT<3, i64>]>;
135 def A64Sbfx : SDNode<"AArch64ISD::SBFX", SDTA64BFX>;
137 def A64Ubfx : SDNode<"AArch64ISD::UBFX", SDTA64BFX>;
139 class BinOpFrag<dag res> : PatFrag<(ops node:$LHS, node:$RHS), res>;
141 //===----------------------------------------------------------------------===//
142 // Call sequence pseudo-instructions
143 //===----------------------------------------------------------------------===//
146 def SDT_AArch64Call : SDTypeProfile<0, -1, [SDTCisPtrTy<0>]>;
147 def AArch64Call : SDNode<"AArch64ISD::Call", SDT_AArch64Call,
148 [SDNPHasChain, SDNPOptInGlue, SDNPOutGlue, SDNPVariadic]>;
150 def AArch64tcret : SDNode<"AArch64ISD::TC_RETURN", SDT_AArch64Call,
151 [SDNPHasChain, SDNPOptInGlue, SDNPVariadic]>;
153 // The TLSDESCCALL node is a variant call which goes to an indirectly calculated
154 // destination but needs a relocation against a fixed symbol. As such it has two
155 // certain operands: the callee and the relocated variable.
157 // The TLS ABI only allows it to be selected to a BLR instructin (with
158 // appropriate relocation).
159 def SDTTLSDescCall : SDTypeProfile<0, -2, [SDTCisPtrTy<0>, SDTCisPtrTy<1>]>;
161 def A64tlsdesc_blr : SDNode<"AArch64ISD::TLSDESCCALL", SDTTLSDescCall,
162 [SDNPInGlue, SDNPOutGlue, SDNPHasChain,
166 def SDT_AArch64CallSeqStart : SDCallSeqStart<[ SDTCisPtrTy<0> ]>;
167 def AArch64callseq_start : SDNode<"ISD::CALLSEQ_START", SDT_AArch64CallSeqStart,
168 [SDNPHasChain, SDNPOutGlue]>;
170 def SDT_AArch64CallSeqEnd : SDCallSeqEnd<[ SDTCisPtrTy<0>, SDTCisPtrTy<1> ]>;
171 def AArch64callseq_end : SDNode<"ISD::CALLSEQ_END", SDT_AArch64CallSeqEnd,
172 [SDNPHasChain, SDNPOptInGlue, SDNPOutGlue]>;
176 // These pseudo-instructions have special semantics by virtue of being passed to
177 // the InstrInfo constructor. CALLSEQ_START/CALLSEQ_END are produced by
178 // LowerCall to (in our case) tell the back-end about stack adjustments for
179 // arguments passed on the stack. Here we select those markers to
180 // pseudo-instructions which explicitly set the stack, and finally in the
181 // RegisterInfo we convert them to a true stack adjustment.
182 let Defs = [XSP], Uses = [XSP] in {
183 def ADJCALLSTACKDOWN : PseudoInst<(outs), (ins i64imm:$amt),
184 [(AArch64callseq_start timm:$amt)]>;
186 def ADJCALLSTACKUP : PseudoInst<(outs), (ins i64imm:$amt1, i64imm:$amt2),
187 [(AArch64callseq_end timm:$amt1, timm:$amt2)]>;
190 //===----------------------------------------------------------------------===//
191 // Atomic operation pseudo-instructions
192 //===----------------------------------------------------------------------===//
194 // These get selected from C++ code as a pretty much direct translation from the
195 // generic DAG nodes. The one exception is the AtomicOrdering is added as an
196 // operand so that the eventual lowering can make use of it and choose
197 // acquire/release operations when required.
199 let usesCustomInserter = 1, hasCtrlDep = 1, mayLoad = 1, mayStore = 1 in {
200 multiclass AtomicSizes {
201 def _I8 : PseudoInst<(outs GPR32:$dst),
202 (ins GPR64xsp:$ptr, GPR32:$incr, i32imm:$ordering), []>;
203 def _I16 : PseudoInst<(outs GPR32:$dst),
204 (ins GPR64xsp:$ptr, GPR32:$incr, i32imm:$ordering), []>;
205 def _I32 : PseudoInst<(outs GPR32:$dst),
206 (ins GPR64xsp:$ptr, GPR32:$incr, i32imm:$ordering), []>;
207 def _I64 : PseudoInst<(outs GPR64:$dst),
208 (ins GPR64xsp:$ptr, GPR64:$incr, i32imm:$ordering), []>;
212 defm ATOMIC_LOAD_ADD : AtomicSizes;
213 defm ATOMIC_LOAD_SUB : AtomicSizes;
214 defm ATOMIC_LOAD_AND : AtomicSizes;
215 defm ATOMIC_LOAD_OR : AtomicSizes;
216 defm ATOMIC_LOAD_XOR : AtomicSizes;
217 defm ATOMIC_LOAD_NAND : AtomicSizes;
218 defm ATOMIC_SWAP : AtomicSizes;
219 let Defs = [NZCV] in {
220 // These operations need a CMP to calculate the correct value
221 defm ATOMIC_LOAD_MIN : AtomicSizes;
222 defm ATOMIC_LOAD_MAX : AtomicSizes;
223 defm ATOMIC_LOAD_UMIN : AtomicSizes;
224 defm ATOMIC_LOAD_UMAX : AtomicSizes;
227 class AtomicCmpSwap<RegisterClass GPRData>
228 : PseudoInst<(outs GPRData:$dst),
229 (ins GPR64xsp:$ptr, GPRData:$old, GPRData:$new,
230 i32imm:$ordering), []> {
231 let usesCustomInserter = 1;
238 def ATOMIC_CMP_SWAP_I8 : AtomicCmpSwap<GPR32>;
239 def ATOMIC_CMP_SWAP_I16 : AtomicCmpSwap<GPR32>;
240 def ATOMIC_CMP_SWAP_I32 : AtomicCmpSwap<GPR32>;
241 def ATOMIC_CMP_SWAP_I64 : AtomicCmpSwap<GPR64>;
243 //===----------------------------------------------------------------------===//
244 // Add-subtract (extended register) instructions
245 //===----------------------------------------------------------------------===//
246 // Contains: ADD, ADDS, SUB, SUBS + aliases CMN, CMP
248 // The RHS of these operations is conceptually a sign/zero-extended
249 // register, optionally shifted left by 1-4. The extension can be a
250 // NOP (e.g. "sxtx" sign-extending a 64-bit register to 64-bits) but
251 // must be specified with one exception:
253 // If one of the registers is sp/wsp then LSL is an alias for UXTW in
254 // 32-bit instructions and UXTX in 64-bit versions, the shift amount
255 // is not optional in that case (but can explicitly be 0), and the
256 // entire suffix can be skipped (e.g. "add sp, x3, x2").
258 multiclass extend_operands<string PREFIX, string Diag> {
259 def _asmoperand : AsmOperandClass {
261 let RenderMethod = "addRegExtendOperands";
262 let PredicateMethod = "isRegExtend<A64SE::" # PREFIX # ">";
263 let DiagnosticType = "AddSubRegExtend" # Diag;
266 def _operand : Operand<i64>,
267 ImmLeaf<i64, [{ return Imm >= 0 && Imm <= 4; }]> {
268 let PrintMethod = "printRegExtendOperand<A64SE::" # PREFIX # ">";
269 let DecoderMethod = "DecodeRegExtendOperand";
270 let ParserMatchClass = !cast<AsmOperandClass>(PREFIX # "_asmoperand");
274 defm UXTB : extend_operands<"UXTB", "Small">;
275 defm UXTH : extend_operands<"UXTH", "Small">;
276 defm UXTW : extend_operands<"UXTW", "Small">;
277 defm UXTX : extend_operands<"UXTX", "Large">;
278 defm SXTB : extend_operands<"SXTB", "Small">;
279 defm SXTH : extend_operands<"SXTH", "Small">;
280 defm SXTW : extend_operands<"SXTW", "Small">;
281 defm SXTX : extend_operands<"SXTX", "Large">;
283 def LSL_extasmoperand : AsmOperandClass {
284 let Name = "RegExtendLSL";
285 let RenderMethod = "addRegExtendOperands";
286 let DiagnosticType = "AddSubRegExtendLarge";
289 def LSL_extoperand : Operand<i64> {
290 let ParserMatchClass = LSL_extasmoperand;
294 // The patterns for various sign-extensions are a little ugly and
295 // non-uniform because everything has already been promoted to the
296 // legal i64 and i32 types. We'll wrap the various variants up in a
297 // class for use later.
299 dag uxtb; dag uxth; dag uxtw; dag uxtx;
300 dag sxtb; dag sxth; dag sxtw; dag sxtx;
305 def extends_to_i64 : extend_types {
306 let uxtb = (and (anyext i32:$Rm), 255);
307 let uxth = (and (anyext i32:$Rm), 65535);
308 let uxtw = (zext i32:$Rm);
309 let uxtx = (i64 $Rm);
311 let sxtb = (sext_inreg (anyext i32:$Rm), i8);
312 let sxth = (sext_inreg (anyext i32:$Rm), i16);
313 let sxtw = (sext i32:$Rm);
314 let sxtx = (i64 $Rm);
321 def extends_to_i32 : extend_types {
322 let uxtb = (and i32:$Rm, 255);
323 let uxth = (and i32:$Rm, 65535);
324 let uxtw = (i32 i32:$Rm);
325 let uxtx = (i32 i32:$Rm);
327 let sxtb = (sext_inreg i32:$Rm, i8);
328 let sxth = (sext_inreg i32:$Rm, i16);
329 let sxtw = (i32 i32:$Rm);
330 let sxtx = (i32 i32:$Rm);
336 // Now, six of the extensions supported are easy and uniform: if the source size
337 // is 32-bits or less, then Rm is always a 32-bit register. We'll instantiate
338 // those instructions in one block.
340 // The uxtx/sxtx could potentially be merged in, but three facts dissuaded me:
341 // + It would break the naming scheme: either ADDxx_uxtx or ADDww_uxtx would
343 // + Patterns are very different as well.
344 // + Passing different registers would be ugly (more fields in extend_types
345 // would probably be the best option).
346 multiclass addsub_exts<bit sf, bit op, bit S, string asmop,
347 SDPatternOperator opfrag,
348 dag outs, extend_types exts> {
349 def w_uxtb : A64I_addsubext<sf, op, S, 0b00, 0b000,
350 outs, (ins exts.GPR:$Rn, GPR32:$Rm, UXTB_operand:$Imm3),
351 !strconcat(asmop, "$Rn, $Rm, $Imm3"),
352 [(opfrag exts.ty:$Rn, (shl exts.uxtb, UXTB_operand:$Imm3))],
354 Sched<[WriteALU, ReadALU, ReadALU]>;
355 def w_uxth : A64I_addsubext<sf, op, S, 0b00, 0b001,
356 outs, (ins exts.GPR:$Rn, GPR32:$Rm, UXTH_operand:$Imm3),
357 !strconcat(asmop, "$Rn, $Rm, $Imm3"),
358 [(opfrag exts.ty:$Rn, (shl exts.uxth, UXTH_operand:$Imm3))],
360 Sched<[WriteALU, ReadALU, ReadALU]>;
361 def w_uxtw : A64I_addsubext<sf, op, S, 0b00, 0b010,
362 outs, (ins exts.GPR:$Rn, GPR32:$Rm, UXTW_operand:$Imm3),
363 !strconcat(asmop, "$Rn, $Rm, $Imm3"),
364 [(opfrag exts.ty:$Rn, (shl exts.uxtw, UXTW_operand:$Imm3))],
366 Sched<[WriteALU, ReadALU, ReadALU]>;
368 def w_sxtb : A64I_addsubext<sf, op, S, 0b00, 0b100,
369 outs, (ins exts.GPR:$Rn, GPR32:$Rm, SXTB_operand:$Imm3),
370 !strconcat(asmop, "$Rn, $Rm, $Imm3"),
371 [(opfrag exts.ty:$Rn, (shl exts.sxtb, SXTB_operand:$Imm3))],
373 Sched<[WriteALU, ReadALU, ReadALU]>;
374 def w_sxth : A64I_addsubext<sf, op, S, 0b00, 0b101,
375 outs, (ins exts.GPR:$Rn, GPR32:$Rm, SXTH_operand:$Imm3),
376 !strconcat(asmop, "$Rn, $Rm, $Imm3"),
377 [(opfrag exts.ty:$Rn, (shl exts.sxth, SXTH_operand:$Imm3))],
379 Sched<[WriteALU, ReadALU, ReadALU]>;
380 def w_sxtw : A64I_addsubext<sf, op, S, 0b00, 0b110,
381 outs, (ins exts.GPR:$Rn, GPR32:$Rm, SXTW_operand:$Imm3),
382 !strconcat(asmop, "$Rn, $Rm, $Imm3"),
383 [(opfrag exts.ty:$Rn, (shl exts.sxtw, SXTW_operand:$Imm3))],
385 Sched<[WriteALU, ReadALU, ReadALU]>;
388 // These two could be merge in with the above, but their patterns aren't really
389 // necessary and the naming-scheme would necessarily break:
390 multiclass addsub_xxtx<bit op, bit S, string asmop, SDPatternOperator opfrag,
392 def x_uxtx : A64I_addsubext<0b1, op, S, 0b00, 0b011,
394 (ins GPR64xsp:$Rn, GPR64:$Rm, UXTX_operand:$Imm3),
395 !strconcat(asmop, "$Rn, $Rm, $Imm3"),
396 [(opfrag i64:$Rn, (shl i64:$Rm, UXTX_operand:$Imm3))],
398 Sched<[WriteALU, ReadALU, ReadALU]>;
400 def x_sxtx : A64I_addsubext<0b1, op, S, 0b00, 0b111,
402 (ins GPR64xsp:$Rn, GPR64:$Rm, SXTX_operand:$Imm3),
403 !strconcat(asmop, "$Rn, $Rm, $Imm3"),
404 [/* No Pattern: same as uxtx */],
406 Sched<[WriteALU, ReadALU, ReadALU]>;
409 multiclass addsub_wxtx<bit op, bit S, string asmop, dag outs> {
410 def w_uxtx : A64I_addsubext<0b0, op, S, 0b00, 0b011,
411 outs, (ins GPR32wsp:$Rn, GPR32:$Rm, UXTX_operand:$Imm3),
412 !strconcat(asmop, "$Rn, $Rm, $Imm3"),
413 [/* No pattern: probably same as uxtw */],
415 Sched<[WriteALU, ReadALU, ReadALU]>;
417 def w_sxtx : A64I_addsubext<0b0, op, S, 0b00, 0b111,
418 outs, (ins GPR32wsp:$Rn, GPR32:$Rm, SXTX_operand:$Imm3),
419 !strconcat(asmop, "$Rn, $Rm, $Imm3"),
420 [/* No Pattern: probably same as uxtw */],
422 Sched<[WriteALU, ReadALU, ReadALU]>;
425 class SetRD<RegisterClass RC, SDPatternOperator op>
426 : PatFrag<(ops node:$lhs, node:$rhs), (set RC:$Rd, (op node:$lhs, node:$rhs))>;
427 class SetNZCV<SDPatternOperator op>
428 : PatFrag<(ops node:$lhs, node:$rhs), (set NZCV, (op node:$lhs, node:$rhs))>;
430 defm ADDxx :addsub_exts<0b1, 0b0, 0b0, "add\t$Rd, ", SetRD<GPR64xsp, add>,
431 (outs GPR64xsp:$Rd), extends_to_i64>,
432 addsub_xxtx< 0b0, 0b0, "add\t$Rd, ", SetRD<GPR64xsp, add>,
433 (outs GPR64xsp:$Rd)>;
434 defm ADDww :addsub_exts<0b0, 0b0, 0b0, "add\t$Rd, ", SetRD<GPR32wsp, add>,
435 (outs GPR32wsp:$Rd), extends_to_i32>,
436 addsub_wxtx< 0b0, 0b0, "add\t$Rd, ",
437 (outs GPR32wsp:$Rd)>;
438 defm SUBxx :addsub_exts<0b1, 0b1, 0b0, "sub\t$Rd, ", SetRD<GPR64xsp, sub>,
439 (outs GPR64xsp:$Rd), extends_to_i64>,
440 addsub_xxtx< 0b1, 0b0, "sub\t$Rd, ", SetRD<GPR64xsp, sub>,
441 (outs GPR64xsp:$Rd)>;
442 defm SUBww :addsub_exts<0b0, 0b1, 0b0, "sub\t$Rd, ", SetRD<GPR32wsp, sub>,
443 (outs GPR32wsp:$Rd), extends_to_i32>,
444 addsub_wxtx< 0b1, 0b0, "sub\t$Rd, ",
445 (outs GPR32wsp:$Rd)>;
447 let Defs = [NZCV] in {
448 defm ADDSxx :addsub_exts<0b1, 0b0, 0b1, "adds\t$Rd, ", SetRD<GPR64, addc>,
449 (outs GPR64:$Rd), extends_to_i64>,
450 addsub_xxtx< 0b0, 0b1, "adds\t$Rd, ", SetRD<GPR64, addc>,
452 defm ADDSww :addsub_exts<0b0, 0b0, 0b1, "adds\t$Rd, ", SetRD<GPR32, addc>,
453 (outs GPR32:$Rd), extends_to_i32>,
454 addsub_wxtx< 0b0, 0b1, "adds\t$Rd, ",
456 defm SUBSxx :addsub_exts<0b1, 0b1, 0b1, "subs\t$Rd, ", SetRD<GPR64, subc>,
457 (outs GPR64:$Rd), extends_to_i64>,
458 addsub_xxtx< 0b1, 0b1, "subs\t$Rd, ", SetRD<GPR64, subc>,
460 defm SUBSww :addsub_exts<0b0, 0b1, 0b1, "subs\t$Rd, ", SetRD<GPR32, subc>,
461 (outs GPR32:$Rd), extends_to_i32>,
462 addsub_wxtx< 0b1, 0b1, "subs\t$Rd, ",
466 let SchedRW = [WriteCMP, ReadCMP, ReadCMP], Rd = 0b11111, isCompare = 1 in {
467 defm CMNx : addsub_exts<0b1, 0b0, 0b1, "cmn\t", SetNZCV<A64cmn>,
468 (outs), extends_to_i64>,
469 addsub_xxtx< 0b0, 0b1, "cmn\t", SetNZCV<A64cmn>, (outs)>;
470 defm CMNw : addsub_exts<0b0, 0b0, 0b1, "cmn\t", SetNZCV<A64cmn>,
471 (outs), extends_to_i32>,
472 addsub_wxtx< 0b0, 0b1, "cmn\t", (outs)>;
473 defm CMPx : addsub_exts<0b1, 0b1, 0b1, "cmp\t", SetNZCV<A64cmp>,
474 (outs), extends_to_i64>,
475 addsub_xxtx< 0b1, 0b1, "cmp\t", SetNZCV<A64cmp>, (outs)>;
476 defm CMPw : addsub_exts<0b0, 0b1, 0b1, "cmp\t", SetNZCV<A64cmp>,
477 (outs), extends_to_i32>,
478 addsub_wxtx< 0b1, 0b1, "cmp\t", (outs)>;
482 // Now patterns for the operation without a shift being needed. No patterns are
483 // created for uxtx/sxtx since they're non-uniform and it's expected that
484 // add/sub (shifted register) will handle those cases anyway.
485 multiclass addsubext_noshift_patterns<string prefix, SDPatternOperator nodeop,
487 def : Pat<(nodeop exts.ty:$Rn, exts.uxtb),
488 (!cast<Instruction>(prefix # "w_uxtb") $Rn, $Rm, 0)>;
489 def : Pat<(nodeop exts.ty:$Rn, exts.uxth),
490 (!cast<Instruction>(prefix # "w_uxth") $Rn, $Rm, 0)>;
491 def : Pat<(nodeop exts.ty:$Rn, exts.uxtw),
492 (!cast<Instruction>(prefix # "w_uxtw") $Rn, $Rm, 0)>;
494 def : Pat<(nodeop exts.ty:$Rn, exts.sxtb),
495 (!cast<Instruction>(prefix # "w_sxtb") $Rn, $Rm, 0)>;
496 def : Pat<(nodeop exts.ty:$Rn, exts.sxth),
497 (!cast<Instruction>(prefix # "w_sxth") $Rn, $Rm, 0)>;
498 def : Pat<(nodeop exts.ty:$Rn, exts.sxtw),
499 (!cast<Instruction>(prefix # "w_sxtw") $Rn, $Rm, 0)>;
502 defm : addsubext_noshift_patterns<"ADDxx", add, extends_to_i64>;
503 defm : addsubext_noshift_patterns<"ADDww", add, extends_to_i32>;
504 defm : addsubext_noshift_patterns<"SUBxx", sub, extends_to_i64>;
505 defm : addsubext_noshift_patterns<"SUBww", sub, extends_to_i32>;
507 defm : addsubext_noshift_patterns<"CMNx", A64cmn, extends_to_i64>;
508 defm : addsubext_noshift_patterns<"CMNw", A64cmn, extends_to_i32>;
509 defm : addsubext_noshift_patterns<"CMPx", A64cmp, extends_to_i64>;
510 defm : addsubext_noshift_patterns<"CMPw", A64cmp, extends_to_i32>;
512 // An extend of "lsl #imm" is valid if and only if one of Rn and Rd is
513 // sp/wsp. It is synonymous with uxtx/uxtw depending on the size of the
514 // operation. Also permitted in this case is complete omission of the argument,
515 // which implies "lsl #0".
516 multiclass lsl_aliases<string asmop, Instruction inst, RegisterClass GPR_Rd,
517 RegisterClass GPR_Rn, RegisterClass GPR_Rm> {
518 def : InstAlias<!strconcat(asmop, " $Rd, $Rn, $Rm"),
519 (inst GPR_Rd:$Rd, GPR_Rn:$Rn, GPR_Rm:$Rm, 0)>;
521 def : InstAlias<!strconcat(asmop, " $Rd, $Rn, $Rm, $LSL"),
522 (inst GPR_Rd:$Rd, GPR_Rn:$Rn, GPR_Rm:$Rm, LSL_extoperand:$LSL),
527 defm : lsl_aliases<"add", ADDxxx_uxtx, Rxsp, GPR64xsp, GPR64>;
528 defm : lsl_aliases<"add", ADDxxx_uxtx, GPR64xsp, Rxsp, GPR64>;
529 defm : lsl_aliases<"add", ADDwww_uxtw, Rwsp, GPR32wsp, GPR32>;
530 defm : lsl_aliases<"add", ADDwww_uxtw, GPR32wsp, Rwsp, GPR32>;
531 defm : lsl_aliases<"sub", SUBxxx_uxtx, Rxsp, GPR64xsp, GPR64>;
532 defm : lsl_aliases<"sub", SUBxxx_uxtx, GPR64xsp, Rxsp, GPR64>;
533 defm : lsl_aliases<"sub", SUBwww_uxtw, Rwsp, GPR32wsp, GPR32>;
534 defm : lsl_aliases<"sub", SUBwww_uxtw, GPR32wsp, Rwsp, GPR32>;
536 // Rd cannot be sp for flag-setting variants so only half of the aliases are
538 defm : lsl_aliases<"adds", ADDSxxx_uxtx, GPR64, Rxsp, GPR64>;
539 defm : lsl_aliases<"adds", ADDSwww_uxtw, GPR32, Rwsp, GPR32>;
540 defm : lsl_aliases<"subs", SUBSxxx_uxtx, GPR64, Rxsp, GPR64>;
541 defm : lsl_aliases<"subs", SUBSwww_uxtw, GPR32, Rwsp, GPR32>;
543 // CMP unfortunately has to be different because the instruction doesn't have a
545 multiclass cmp_lsl_aliases<string asmop, Instruction inst,
546 RegisterClass GPR_Rn, RegisterClass GPR_Rm> {
547 def : InstAlias<!strconcat(asmop, " $Rn, $Rm"),
548 (inst GPR_Rn:$Rn, GPR_Rm:$Rm, 0)>;
550 def : InstAlias<!strconcat(asmop, " $Rn, $Rm, $LSL"),
551 (inst GPR_Rn:$Rn, GPR_Rm:$Rm, LSL_extoperand:$LSL)>;
554 defm : cmp_lsl_aliases<"cmp", CMPxx_uxtx, Rxsp, GPR64>;
555 defm : cmp_lsl_aliases<"cmp", CMPww_uxtw, Rwsp, GPR32>;
556 defm : cmp_lsl_aliases<"cmn", CMNxx_uxtx, Rxsp, GPR64>;
557 defm : cmp_lsl_aliases<"cmn", CMNww_uxtw, Rwsp, GPR32>;
559 //===----------------------------------------------------------------------===//
560 // Add-subtract (immediate) instructions
561 //===----------------------------------------------------------------------===//
562 // Contains: ADD, ADDS, SUB, SUBS + aliases CMN, CMP, MOV
564 // These instructions accept a 12-bit unsigned immediate, optionally shifted
565 // left by 12 bits. Official assembly format specifies a 12 bit immediate with
566 // one of "", "LSL #0", "LSL #12" supplementary operands.
568 // There are surprisingly few ways to make this work with TableGen, so this
569 // implementation has separate instructions for the "LSL #0" and "LSL #12"
572 // If the MCInst retained a single combined immediate (which could be 0x123000,
573 // for example) then both components (imm & shift) would have to be delegated to
574 // a single assembly operand. This would entail a separate operand parser
575 // (because the LSL would have to live in the same AArch64Operand as the
576 // immediate to be accessible); assembly parsing is rather complex and
577 // error-prone C++ code.
579 // By splitting the immediate, we can delegate handling this optional operand to
580 // an InstAlias. Supporting functions to generate the correct MCInst are still
581 // required, but these are essentially trivial and parsing can remain generic.
583 // Rejected plans with rationale:
584 // ------------------------------
586 // In an ideal world you'de have two first class immediate operands (in
587 // InOperandList, specifying imm12 and shift). Unfortunately this is not
588 // selectable by any means I could discover.
590 // An Instruction with two MCOperands hidden behind a single entry in
591 // InOperandList (expanded by ComplexPatterns and MIOperandInfo) was functional,
592 // but required more C++ code to handle encoding/decoding. Parsing (the intended
593 // main beneficiary) ended up equally complex because of the optional nature of
596 // Attempting to circumvent the need for a custom OperandParser above by giving
597 // InstAliases without the "lsl #0" failed. add/sub could be accommodated but
598 // the cmp/cmn aliases didn't use the MIOperandInfo to determine how operands
599 // should be parsed: there was no way to accommodate an "lsl #12".
601 let ParserMethod = "ParseImmWithLSLOperand",
602 RenderMethod = "addImmWithLSLOperands" in {
603 // Derived PredicateMethod fields are different for each
604 def addsubimm_lsl0_asmoperand : AsmOperandClass {
605 let Name = "AddSubImmLSL0";
606 // If an error is reported against this operand, instruction could also be a
608 let DiagnosticType = "AddSubSecondSource";
611 def addsubimm_lsl12_asmoperand : AsmOperandClass {
612 let Name = "AddSubImmLSL12";
613 let DiagnosticType = "AddSubSecondSource";
617 def shr_12_XFORM : SDNodeXForm<imm, [{
618 return CurDAG->getTargetConstant(N->getSExtValue() >> 12, MVT::i32);
621 def shr_12_neg_XFORM : SDNodeXForm<imm, [{
622 return CurDAG->getTargetConstant((-N->getSExtValue()) >> 12, MVT::i32);
625 def neg_XFORM : SDNodeXForm<imm, [{
626 return CurDAG->getTargetConstant(-N->getSExtValue(), MVT::i32);
630 multiclass addsub_imm_operands<ValueType ty> {
631 let PrintMethod = "printAddSubImmLSL0Operand",
632 EncoderMethod = "getAddSubImmOpValue",
633 ParserMatchClass = addsubimm_lsl0_asmoperand in {
634 def _posimm_lsl0 : Operand<ty>,
635 ImmLeaf<ty, [{ return Imm >= 0 && (Imm & ~0xfff) == 0; }]>;
636 def _negimm_lsl0 : Operand<ty>,
637 ImmLeaf<ty, [{ return Imm < 0 && (-Imm & ~0xfff) == 0; }],
641 let PrintMethod = "printAddSubImmLSL12Operand",
642 EncoderMethod = "getAddSubImmOpValue",
643 ParserMatchClass = addsubimm_lsl12_asmoperand in {
644 def _posimm_lsl12 : Operand<ty>,
645 ImmLeaf<ty, [{ return Imm >= 0 && (Imm & ~0xfff000) == 0; }],
648 def _negimm_lsl12 : Operand<ty>,
649 ImmLeaf<ty, [{ return Imm < 0 && (-Imm & ~0xfff000) == 0; }],
654 // The add operands don't need any transformation
655 defm addsubimm_operand_i32 : addsub_imm_operands<i32>;
656 defm addsubimm_operand_i64 : addsub_imm_operands<i64>;
658 multiclass addsubimm_varieties<string prefix, bit sf, bit op, bits<2> shift,
659 string asmop, string cmpasmop,
660 Operand imm_operand, Operand cmp_imm_operand,
661 RegisterClass GPR, RegisterClass GPRsp,
662 AArch64Reg ZR, ValueType Ty> {
663 // All registers for non-S variants allow SP
664 def _s : A64I_addsubimm<sf, op, 0b0, shift,
666 (ins GPRsp:$Rn, imm_operand:$Imm12),
667 !strconcat(asmop, "\t$Rd, $Rn, $Imm12"),
668 [(set Ty:$Rd, (add Ty:$Rn, imm_operand:$Imm12))],
670 Sched<[WriteALU, ReadALU]>;
673 // S variants can read SP but would write to ZR
674 def _S : A64I_addsubimm<sf, op, 0b1, shift,
676 (ins GPRsp:$Rn, imm_operand:$Imm12),
677 !strconcat(asmop, "s\t$Rd, $Rn, $Imm12"),
678 [(set Ty:$Rd, (addc Ty:$Rn, imm_operand:$Imm12))],
680 Sched<[WriteALU, ReadALU]> {
684 // Note that the pattern here for ADDS is subtle. Canonically CMP
685 // a, b becomes SUBS a, b. If b < 0 then this is equivalent to
686 // ADDS a, (-b). This is not true in general.
687 def _cmp : A64I_addsubimm<sf, op, 0b1, shift,
688 (outs), (ins GPRsp:$Rn, imm_operand:$Imm12),
689 !strconcat(cmpasmop, " $Rn, $Imm12"),
691 (A64cmp Ty:$Rn, cmp_imm_operand:$Imm12))],
693 Sched<[WriteCMP, ReadCMP]> {
701 multiclass addsubimm_shifts<string prefix, bit sf, bit op,
702 string asmop, string cmpasmop, string operand, string cmpoperand,
703 RegisterClass GPR, RegisterClass GPRsp, AArch64Reg ZR,
705 defm _lsl0 : addsubimm_varieties<prefix # "_lsl0", sf, op, 0b00,
707 !cast<Operand>(operand # "_lsl0"),
708 !cast<Operand>(cmpoperand # "_lsl0"),
711 defm _lsl12 : addsubimm_varieties<prefix # "_lsl12", sf, op, 0b01,
713 !cast<Operand>(operand # "_lsl12"),
714 !cast<Operand>(cmpoperand # "_lsl12"),
718 defm ADDwwi : addsubimm_shifts<"ADDwi", 0b0, 0b0, "add", "cmn",
719 "addsubimm_operand_i32_posimm",
720 "addsubimm_operand_i32_negimm",
721 GPR32, GPR32wsp, WZR, i32>;
722 defm ADDxxi : addsubimm_shifts<"ADDxi", 0b1, 0b0, "add", "cmn",
723 "addsubimm_operand_i64_posimm",
724 "addsubimm_operand_i64_negimm",
725 GPR64, GPR64xsp, XZR, i64>;
726 defm SUBwwi : addsubimm_shifts<"SUBwi", 0b0, 0b1, "sub", "cmp",
727 "addsubimm_operand_i32_negimm",
728 "addsubimm_operand_i32_posimm",
729 GPR32, GPR32wsp, WZR, i32>;
730 defm SUBxxi : addsubimm_shifts<"SUBxi", 0b1, 0b1, "sub", "cmp",
731 "addsubimm_operand_i64_negimm",
732 "addsubimm_operand_i64_posimm",
733 GPR64, GPR64xsp, XZR, i64>;
735 multiclass MOVsp<RegisterClass GPRsp, RegisterClass SP, Instruction addop> {
736 def _fromsp : InstAlias<"mov $Rd, $Rn",
737 (addop GPRsp:$Rd, SP:$Rn, 0),
740 def _tosp : InstAlias<"mov $Rd, $Rn",
741 (addop SP:$Rd, GPRsp:$Rn, 0),
745 // Recall Rxsp is a RegisterClass containing *just* xsp.
746 defm MOVxx : MOVsp<GPR64xsp, Rxsp, ADDxxi_lsl0_s>;
747 defm MOVww : MOVsp<GPR32wsp, Rwsp, ADDwwi_lsl0_s>;
749 //===----------------------------------------------------------------------===//
750 // Add-subtract (shifted register) instructions
751 //===----------------------------------------------------------------------===//
752 // Contains: ADD, ADDS, SUB, SUBS + aliases CMN, CMP, NEG, NEGS
754 //===-------------------------------
755 // 1. The "shifted register" operands. Shared with logical insts.
756 //===-------------------------------
758 multiclass shift_operands<string prefix, string form> {
759 def _asmoperand_i32 : AsmOperandClass {
760 let Name = "Shift" # form # "i32";
761 let RenderMethod = "addShiftOperands";
762 let PredicateMethod = "isShift<A64SE::" # form # ", false>";
763 let DiagnosticType = "AddSubRegShift32";
766 // Note that the operand type is intentionally i64 because the DAGCombiner
767 // puts these into a canonical form.
768 def _i32 : Operand<i64>, ImmLeaf<i64, [{ return Imm >= 0 && Imm <= 31; }]> {
770 = !cast<AsmOperandClass>(prefix # "_asmoperand_i32");
771 let PrintMethod = "printShiftOperand<A64SE::" # form # ">";
772 let DecoderMethod = "Decode32BitShiftOperand";
775 def _asmoperand_i64 : AsmOperandClass {
776 let Name = "Shift" # form # "i64";
777 let RenderMethod = "addShiftOperands";
778 let PredicateMethod = "isShift<A64SE::" # form # ", true>";
779 let DiagnosticType = "AddSubRegShift64";
782 def _i64 : Operand<i64>, ImmLeaf<i64, [{ return Imm >= 0 && Imm <= 63; }]> {
784 = !cast<AsmOperandClass>(prefix # "_asmoperand_i64");
785 let PrintMethod = "printShiftOperand<A64SE::" # form # ">";
789 defm lsl_operand : shift_operands<"lsl_operand", "LSL">;
790 defm lsr_operand : shift_operands<"lsr_operand", "LSR">;
791 defm asr_operand : shift_operands<"asr_operand", "ASR">;
793 // Not used for add/sub, but defined here for completeness. The "logical
794 // (shifted register)" instructions *do* have an ROR variant.
795 defm ror_operand : shift_operands<"ror_operand", "ROR">;
797 //===-------------------------------
798 // 2. The basic 3.5-operand ADD/SUB/ADDS/SUBS instructions.
799 //===-------------------------------
801 // N.b. the commutable parameter is just !N. It will be first against the wall
802 // when the revolution comes.
803 multiclass addsub_shifts<string prefix, bit sf, bit op, bit s, bit commutable,
804 string asmop, SDPatternOperator opfrag, ValueType ty,
805 RegisterClass GPR, list<Register> defs> {
806 let isCommutable = commutable, Defs = defs in {
807 def _lsl : A64I_addsubshift<sf, op, s, 0b00,
809 (ins GPR:$Rn, GPR:$Rm,
810 !cast<Operand>("lsl_operand_" # ty):$Imm6),
811 !strconcat(asmop, "\t$Rd, $Rn, $Rm, $Imm6"),
812 [(set GPR:$Rd, (opfrag ty:$Rn, (shl ty:$Rm,
813 !cast<Operand>("lsl_operand_" # ty):$Imm6))
816 Sched<[WriteALU, ReadALU]>;
818 def _lsr : A64I_addsubshift<sf, op, s, 0b01,
820 (ins GPR:$Rn, GPR:$Rm,
821 !cast<Operand>("lsr_operand_" # ty):$Imm6),
822 !strconcat(asmop, "\t$Rd, $Rn, $Rm, $Imm6"),
823 [(set ty:$Rd, (opfrag ty:$Rn, (srl ty:$Rm,
824 !cast<Operand>("lsr_operand_" # ty):$Imm6))
827 Sched<[WriteALU, ReadALU]>;
829 def _asr : A64I_addsubshift<sf, op, s, 0b10,
831 (ins GPR:$Rn, GPR:$Rm,
832 !cast<Operand>("asr_operand_" # ty):$Imm6),
833 !strconcat(asmop, "\t$Rd, $Rn, $Rm, $Imm6"),
834 [(set ty:$Rd, (opfrag ty:$Rn, (sra ty:$Rm,
835 !cast<Operand>("asr_operand_" # ty):$Imm6))
838 Sched<[WriteALU, ReadALU]>;
842 : InstAlias<!strconcat(asmop, " $Rd, $Rn, $Rm"),
843 (!cast<Instruction>(prefix # "_lsl") GPR:$Rd, GPR:$Rn,
846 def : Pat<(opfrag ty:$Rn, ty:$Rm),
847 (!cast<Instruction>(prefix # "_lsl") $Rn, $Rm, 0)>;
850 multiclass addsub_sizes<string prefix, bit op, bit s, bit commutable,
851 string asmop, SDPatternOperator opfrag,
852 list<Register> defs> {
853 defm xxx : addsub_shifts<prefix # "xxx", 0b1, op, s,
854 commutable, asmop, opfrag, i64, GPR64, defs>;
855 defm www : addsub_shifts<prefix # "www", 0b0, op, s,
856 commutable, asmop, opfrag, i32, GPR32, defs>;
860 defm ADD : addsub_sizes<"ADD", 0b0, 0b0, 0b1, "add", add, []>;
861 defm SUB : addsub_sizes<"SUB", 0b1, 0b0, 0b0, "sub", sub, []>;
863 defm ADDS : addsub_sizes<"ADDS", 0b0, 0b1, 0b1, "adds", addc, [NZCV]>;
864 defm SUBS : addsub_sizes<"SUBS", 0b1, 0b1, 0b0, "subs", subc, [NZCV]>;
866 //===-------------------------------
867 // 1. The NEG/NEGS aliases
868 //===-------------------------------
870 multiclass neg_alias<Instruction INST, RegisterClass GPR, Register ZR,
871 ValueType ty, Operand shift_operand, SDNode shiftop> {
872 def : InstAlias<"neg $Rd, $Rm, $Imm6",
873 (INST GPR:$Rd, ZR, GPR:$Rm, shift_operand:$Imm6)>;
875 def : Pat<(sub 0, (shiftop ty:$Rm, shift_operand:$Imm6)),
876 (INST ZR, $Rm, shift_operand:$Imm6)>;
879 defm : neg_alias<SUBwww_lsl, GPR32, WZR, i32, lsl_operand_i32, shl>;
880 defm : neg_alias<SUBwww_lsr, GPR32, WZR, i32, lsr_operand_i32, srl>;
881 defm : neg_alias<SUBwww_asr, GPR32, WZR, i32, asr_operand_i32, sra>;
882 def : InstAlias<"neg $Rd, $Rm", (SUBwww_lsl GPR32:$Rd, WZR, GPR32:$Rm, 0)>;
883 def : Pat<(sub 0, i32:$Rm), (SUBwww_lsl WZR, $Rm, 0)>;
885 defm : neg_alias<SUBxxx_lsl, GPR64, XZR, i64, lsl_operand_i64, shl>;
886 defm : neg_alias<SUBxxx_lsr, GPR64, XZR, i64, lsr_operand_i64, srl>;
887 defm : neg_alias<SUBxxx_asr, GPR64, XZR, i64, asr_operand_i64, sra>;
888 def : InstAlias<"neg $Rd, $Rm", (SUBxxx_lsl GPR64:$Rd, XZR, GPR64:$Rm, 0)>;
889 def : Pat<(sub 0, i64:$Rm), (SUBxxx_lsl XZR, $Rm, 0)>;
891 // NEGS doesn't get any patterns yet: defining multiple outputs means C++ has to
893 class negs_alias<Instruction INST, RegisterClass GPR,
894 Register ZR, Operand shift_operand, SDNode shiftop>
895 : InstAlias<"negs $Rd, $Rm, $Imm6",
896 (INST GPR:$Rd, ZR, GPR:$Rm, shift_operand:$Imm6)>;
898 def : negs_alias<SUBSwww_lsl, GPR32, WZR, lsl_operand_i32, shl>;
899 def : negs_alias<SUBSwww_lsr, GPR32, WZR, lsr_operand_i32, srl>;
900 def : negs_alias<SUBSwww_asr, GPR32, WZR, asr_operand_i32, sra>;
901 def : InstAlias<"negs $Rd, $Rm", (SUBSwww_lsl GPR32:$Rd, WZR, GPR32:$Rm, 0)>;
903 def : negs_alias<SUBSxxx_lsl, GPR64, XZR, lsl_operand_i64, shl>;
904 def : negs_alias<SUBSxxx_lsr, GPR64, XZR, lsr_operand_i64, srl>;
905 def : negs_alias<SUBSxxx_asr, GPR64, XZR, asr_operand_i64, sra>;
906 def : InstAlias<"negs $Rd, $Rm", (SUBSxxx_lsl GPR64:$Rd, XZR, GPR64:$Rm, 0)>;
908 //===-------------------------------
909 // 1. The CMP/CMN aliases
910 //===-------------------------------
912 multiclass cmp_shifts<string prefix, bit sf, bit op, bit commutable,
913 string asmop, SDPatternOperator opfrag, ValueType ty,
915 let isCommutable = commutable, Rd = 0b11111, Defs = [NZCV] in {
916 def _lsl : A64I_addsubshift<sf, op, 0b1, 0b00,
918 (ins GPR:$Rn, GPR:$Rm,
919 !cast<Operand>("lsl_operand_" # ty):$Imm6),
920 !strconcat(asmop, "\t$Rn, $Rm, $Imm6"),
921 [(set NZCV, (opfrag ty:$Rn, (shl ty:$Rm,
922 !cast<Operand>("lsl_operand_" # ty):$Imm6))
925 Sched<[WriteCMP, ReadCMP, ReadCMP]>;
927 def _lsr : A64I_addsubshift<sf, op, 0b1, 0b01,
929 (ins GPR:$Rn, GPR:$Rm,
930 !cast<Operand>("lsr_operand_" # ty):$Imm6),
931 !strconcat(asmop, "\t$Rn, $Rm, $Imm6"),
932 [(set NZCV, (opfrag ty:$Rn, (srl ty:$Rm,
933 !cast<Operand>("lsr_operand_" # ty):$Imm6))
936 Sched<[WriteCMP, ReadCMP, ReadCMP]>;
938 def _asr : A64I_addsubshift<sf, op, 0b1, 0b10,
940 (ins GPR:$Rn, GPR:$Rm,
941 !cast<Operand>("asr_operand_" # ty):$Imm6),
942 !strconcat(asmop, "\t$Rn, $Rm, $Imm6"),
943 [(set NZCV, (opfrag ty:$Rn, (sra ty:$Rm,
944 !cast<Operand>("asr_operand_" # ty):$Imm6))
947 Sched<[WriteCMP, ReadCMP, ReadCMP]>;
951 : InstAlias<!strconcat(asmop, " $Rn, $Rm"),
952 (!cast<Instruction>(prefix # "_lsl") GPR:$Rn, GPR:$Rm, 0)>;
954 def : Pat<(opfrag ty:$Rn, ty:$Rm),
955 (!cast<Instruction>(prefix # "_lsl") $Rn, $Rm, 0)>;
958 defm CMPww : cmp_shifts<"CMPww", 0b0, 0b1, 0b0, "cmp", A64cmp, i32, GPR32>;
959 defm CMPxx : cmp_shifts<"CMPxx", 0b1, 0b1, 0b0, "cmp", A64cmp, i64, GPR64>;
961 defm CMNww : cmp_shifts<"CMNww", 0b0, 0b0, 0b1, "cmn", A64cmn, i32, GPR32>;
962 defm CMNxx : cmp_shifts<"CMNxx", 0b1, 0b0, 0b1, "cmn", A64cmn, i64, GPR64>;
964 //===----------------------------------------------------------------------===//
965 // Add-subtract (with carry) instructions
966 //===----------------------------------------------------------------------===//
967 // Contains: ADC, ADCS, SBC, SBCS + aliases NGC, NGCS
969 multiclass A64I_addsubcarrySizes<bit op, bit s, string asmop> {
970 let Uses = [NZCV] in {
971 def www : A64I_addsubcarry<0b0, op, s, 0b000000,
972 (outs GPR32:$Rd), (ins GPR32:$Rn, GPR32:$Rm),
973 !strconcat(asmop, "\t$Rd, $Rn, $Rm"),
975 Sched<[WriteALU, ReadALU, ReadALU]>;
977 def xxx : A64I_addsubcarry<0b1, op, s, 0b000000,
978 (outs GPR64:$Rd), (ins GPR64:$Rn, GPR64:$Rm),
979 !strconcat(asmop, "\t$Rd, $Rn, $Rm"),
981 Sched<[WriteALU, ReadALU, ReadALU]>;
985 let isCommutable = 1 in {
986 defm ADC : A64I_addsubcarrySizes<0b0, 0b0, "adc">;
989 defm SBC : A64I_addsubcarrySizes<0b1, 0b0, "sbc">;
991 let Defs = [NZCV] in {
992 let isCommutable = 1 in {
993 defm ADCS : A64I_addsubcarrySizes<0b0, 0b1, "adcs">;
996 defm SBCS : A64I_addsubcarrySizes<0b1, 0b1, "sbcs">;
999 def : InstAlias<"ngc $Rd, $Rm", (SBCwww GPR32:$Rd, WZR, GPR32:$Rm)>;
1000 def : InstAlias<"ngc $Rd, $Rm", (SBCxxx GPR64:$Rd, XZR, GPR64:$Rm)>;
1001 def : InstAlias<"ngcs $Rd, $Rm", (SBCSwww GPR32:$Rd, WZR, GPR32:$Rm)>;
1002 def : InstAlias<"ngcs $Rd, $Rm", (SBCSxxx GPR64:$Rd, XZR, GPR64:$Rm)>;
1004 // Note that adde and sube can form a chain longer than two (e.g. for 256-bit
1005 // addition). So the flag-setting instructions are appropriate.
1006 def : Pat<(adde i32:$Rn, i32:$Rm), (ADCSwww $Rn, $Rm)>;
1007 def : Pat<(adde i64:$Rn, i64:$Rm), (ADCSxxx $Rn, $Rm)>;
1008 def : Pat<(sube i32:$Rn, i32:$Rm), (SBCSwww $Rn, $Rm)>;
1009 def : Pat<(sube i64:$Rn, i64:$Rm), (SBCSxxx $Rn, $Rm)>;
1011 //===----------------------------------------------------------------------===//
1013 //===----------------------------------------------------------------------===//
1014 // Contains: SBFM, BFM, UBFM, [SU]XT[BHW], ASR, LSR, LSL, SBFI[ZX], BFI, BFXIL,
1017 // Because of the rather complicated nearly-overlapping aliases, the decoding of
1018 // this range of instructions is handled manually. The architectural
1019 // instructions are BFM, SBFM and UBFM but a disassembler should never produce
1022 // In the end, the best option was to use BFM instructions for decoding under
1023 // almost all circumstances, but to create aliasing *Instructions* for each of
1024 // the canonical forms and specify a completely custom decoder which would
1025 // substitute the correct MCInst as needed.
1027 // This also simplifies instruction selection, parsing etc because the MCInsts
1028 // have a shape that's closer to their use in code.
1030 //===-------------------------------
1031 // 1. The architectural BFM instructions
1032 //===-------------------------------
1034 def uimm5_asmoperand : AsmOperandClass {
1036 let PredicateMethod = "isUImm<5>";
1037 let RenderMethod = "addImmOperands";
1038 let DiagnosticType = "UImm5";
1041 def uimm6_asmoperand : AsmOperandClass {
1043 let PredicateMethod = "isUImm<6>";
1044 let RenderMethod = "addImmOperands";
1045 let DiagnosticType = "UImm6";
1048 def bitfield32_imm : Operand<i64>,
1049 ImmLeaf<i64, [{ return Imm >= 0 && Imm < 32; }]> {
1050 let ParserMatchClass = uimm5_asmoperand;
1052 let DecoderMethod = "DecodeBitfield32ImmOperand";
1056 def bitfield64_imm : Operand<i64>,
1057 ImmLeaf<i64, [{ return Imm >= 0 && Imm < 64; }]> {
1058 let ParserMatchClass = uimm6_asmoperand;
1060 // Default decoder works in 64-bit case: the 6-bit field can take any value.
1063 multiclass A64I_bitfieldSizes<bits<2> opc, string asmop> {
1064 def wwii : A64I_bitfield<0b0, opc, 0b0, (outs GPR32:$Rd),
1065 (ins GPR32:$Rn, bitfield32_imm:$ImmR, bitfield32_imm:$ImmS),
1066 !strconcat(asmop, "\t$Rd, $Rn, $ImmR, $ImmS"),
1068 Sched<[WriteALU, ReadALU]> {
1069 let DecoderMethod = "DecodeBitfieldInstruction";
1072 def xxii : A64I_bitfield<0b1, opc, 0b1, (outs GPR64:$Rd),
1073 (ins GPR64:$Rn, bitfield64_imm:$ImmR, bitfield64_imm:$ImmS),
1074 !strconcat(asmop, "\t$Rd, $Rn, $ImmR, $ImmS"),
1076 Sched<[WriteALU, ReadALU]> {
1077 let DecoderMethod = "DecodeBitfieldInstruction";
1081 defm SBFM : A64I_bitfieldSizes<0b00, "sbfm">;
1082 defm UBFM : A64I_bitfieldSizes<0b10, "ubfm">;
1084 // BFM instructions modify the destination register rather than defining it
1087 A64I_bitfield<0b0, 0b01, 0b0, (outs GPR32:$Rd),
1088 (ins GPR32:$src, GPR32:$Rn, bitfield32_imm:$ImmR, bitfield32_imm:$ImmS),
1089 "bfm\t$Rd, $Rn, $ImmR, $ImmS", [], NoItinerary>,
1090 Sched<[WriteALU, ReadALU, ReadALU]> {
1091 let DecoderMethod = "DecodeBitfieldInstruction";
1092 let Constraints = "$src = $Rd";
1096 A64I_bitfield<0b1, 0b01, 0b1, (outs GPR64:$Rd),
1097 (ins GPR64:$src, GPR64:$Rn, bitfield64_imm:$ImmR, bitfield64_imm:$ImmS),
1098 "bfm\t$Rd, $Rn, $ImmR, $ImmS", [], NoItinerary>,
1099 Sched<[WriteALU, ReadALU, ReadALU]> {
1100 let DecoderMethod = "DecodeBitfieldInstruction";
1101 let Constraints = "$src = $Rd";
1105 //===-------------------------------
1106 // 2. Extend aliases to 64-bit dest
1107 //===-------------------------------
1109 // Unfortunately the extensions that end up as 64-bits cannot be handled by an
1110 // instruction alias: their syntax is (for example) "SXTB x0, w0", which needs
1111 // to be mapped to "SBFM x0, x0, #0, 7" (changing the class of Rn). InstAlias is
1112 // not capable of such a map as far as I'm aware
1114 // Note that these instructions are strictly more specific than the
1115 // BFM ones (in ImmR) so they can handle their own decoding.
1116 class A64I_bf_ext<bit sf, bits<2> opc, RegisterClass GPRDest, ValueType dty,
1117 string asmop, bits<6> imms, dag pattern>
1118 : A64I_bitfield<sf, opc, sf,
1119 (outs GPRDest:$Rd), (ins GPR32:$Rn),
1120 !strconcat(asmop, "\t$Rd, $Rn"),
1121 [(set dty:$Rd, pattern)], NoItinerary>,
1122 Sched<[WriteALU, ReadALU]> {
1123 let ImmR = 0b000000;
1127 // Signed extensions
1128 def SXTBxw : A64I_bf_ext<0b1, 0b00, GPR64, i64, "sxtb", 7,
1129 (sext_inreg (anyext i32:$Rn), i8)>;
1130 def SXTBww : A64I_bf_ext<0b0, 0b00, GPR32, i32, "sxtb", 7,
1131 (sext_inreg i32:$Rn, i8)>;
1132 def SXTHxw : A64I_bf_ext<0b1, 0b00, GPR64, i64, "sxth", 15,
1133 (sext_inreg (anyext i32:$Rn), i16)>;
1134 def SXTHww : A64I_bf_ext<0b0, 0b00, GPR32, i32, "sxth", 15,
1135 (sext_inreg i32:$Rn, i16)>;
1136 def SXTWxw : A64I_bf_ext<0b1, 0b00, GPR64, i64, "sxtw", 31, (sext i32:$Rn)>;
1138 // Unsigned extensions
1139 def UXTBww : A64I_bf_ext<0b0, 0b10, GPR32, i32, "uxtb", 7,
1140 (and i32:$Rn, 255)>;
1141 def UXTHww : A64I_bf_ext<0b0, 0b10, GPR32, i32, "uxth", 15,
1142 (and i32:$Rn, 65535)>;
1144 // The 64-bit unsigned variants are not strictly architectural but recommended
1146 let isAsmParserOnly = 1 in {
1147 def UXTBxw : A64I_bf_ext<0b0, 0b10, GPR64, i64, "uxtb", 7,
1148 (and (anyext i32:$Rn), 255)>;
1149 def UXTHxw : A64I_bf_ext<0b0, 0b10, GPR64, i64, "uxth", 15,
1150 (and (anyext i32:$Rn), 65535)>;
1153 // Extra patterns for when the source register is actually 64-bits
1154 // too. There's no architectural difference here, it's just LLVM
1155 // shinanigans. There's no need for equivalent zero-extension patterns
1156 // because they'll already be caught by logical (immediate) matching.
1157 def : Pat<(sext_inreg i64:$Rn, i8),
1158 (SXTBxw (EXTRACT_SUBREG $Rn, sub_32))>;
1159 def : Pat<(sext_inreg i64:$Rn, i16),
1160 (SXTHxw (EXTRACT_SUBREG $Rn, sub_32))>;
1161 def : Pat<(sext_inreg i64:$Rn, i32),
1162 (SXTWxw (EXTRACT_SUBREG $Rn, sub_32))>;
1165 //===-------------------------------
1166 // 3. Aliases for ASR and LSR (the simple shifts)
1167 //===-------------------------------
1169 // These also handle their own decoding because ImmS being set makes
1170 // them take precedence over BFM.
1171 multiclass A64I_shift<bits<2> opc, string asmop, SDNode opnode> {
1172 def wwi : A64I_bitfield<0b0, opc, 0b0,
1173 (outs GPR32:$Rd), (ins GPR32:$Rn, bitfield32_imm:$ImmR),
1174 !strconcat(asmop, "\t$Rd, $Rn, $ImmR"),
1175 [(set i32:$Rd, (opnode i32:$Rn, bitfield32_imm:$ImmR))],
1177 Sched<[WriteALU, ReadALU]> {
1181 def xxi : A64I_bitfield<0b1, opc, 0b1,
1182 (outs GPR64:$Rd), (ins GPR64:$Rn, bitfield64_imm:$ImmR),
1183 !strconcat(asmop, "\t$Rd, $Rn, $ImmR"),
1184 [(set i64:$Rd, (opnode i64:$Rn, bitfield64_imm:$ImmR))],
1186 Sched<[WriteALU, ReadALU]> {
1192 defm ASR : A64I_shift<0b00, "asr", sra>;
1193 defm LSR : A64I_shift<0b10, "lsr", srl>;
1195 //===-------------------------------
1196 // 4. Aliases for LSL
1197 //===-------------------------------
1199 // Unfortunately LSL and subsequent aliases are much more complicated. We need
1200 // to be able to say certain output instruction fields depend in a complex
1201 // manner on combinations of input assembly fields).
1203 // MIOperandInfo *might* have been able to do it, but at the cost of
1204 // significantly more C++ code.
1206 // N.b. contrary to usual practice these operands store the shift rather than
1207 // the machine bits in an MCInst. The complexity overhead of consistency
1208 // outweighed the benefits in this case (custom asmparser, printer and selection
1209 // vs custom encoder).
1210 def bitfield32_lsl_imm : Operand<i64>,
1211 ImmLeaf<i64, [{ return Imm >= 0 && Imm <= 31; }]> {
1212 let ParserMatchClass = uimm5_asmoperand;
1213 let EncoderMethod = "getBitfield32LSLOpValue";
1216 def bitfield64_lsl_imm : Operand<i64>,
1217 ImmLeaf<i64, [{ return Imm >= 0 && Imm <= 63; }]> {
1218 let ParserMatchClass = uimm6_asmoperand;
1219 let EncoderMethod = "getBitfield64LSLOpValue";
1222 class A64I_bitfield_lsl<bit sf, RegisterClass GPR, ValueType ty,
1224 : A64I_bitfield<sf, 0b10, sf, (outs GPR:$Rd), (ins GPR:$Rn, operand:$FullImm),
1225 "lsl\t$Rd, $Rn, $FullImm",
1226 [(set ty:$Rd, (shl ty:$Rn, operand:$FullImm))],
1228 Sched<[WriteALU, ReadALU]> {
1230 let ImmR = FullImm{5-0};
1231 let ImmS = FullImm{11-6};
1233 // No disassembler allowed because it would overlap with BFM which does the
1235 let isAsmParserOnly = 1;
1238 def LSLwwi : A64I_bitfield_lsl<0b0, GPR32, i32, bitfield32_lsl_imm>;
1239 def LSLxxi : A64I_bitfield_lsl<0b1, GPR64, i64, bitfield64_lsl_imm>;
1241 //===-------------------------------
1242 // 5. Aliases for bitfield extract instructions
1243 //===-------------------------------
1245 def bfx32_width_asmoperand : AsmOperandClass {
1246 let Name = "BFX32Width";
1247 let PredicateMethod = "isBitfieldWidth<32>";
1248 let RenderMethod = "addBFXWidthOperands";
1249 let DiagnosticType = "Width32";
1252 def bfx32_width : Operand<i64>, ImmLeaf<i64, [{ return true; }]> {
1253 let PrintMethod = "printBFXWidthOperand";
1254 let ParserMatchClass = bfx32_width_asmoperand;
1257 def bfx64_width_asmoperand : AsmOperandClass {
1258 let Name = "BFX64Width";
1259 let PredicateMethod = "isBitfieldWidth<64>";
1260 let RenderMethod = "addBFXWidthOperands";
1261 let DiagnosticType = "Width64";
1264 def bfx64_width : Operand<i64> {
1265 let PrintMethod = "printBFXWidthOperand";
1266 let ParserMatchClass = bfx64_width_asmoperand;
1270 multiclass A64I_bitfield_extract<bits<2> opc, string asmop, SDNode op> {
1271 def wwii : A64I_bitfield<0b0, opc, 0b0, (outs GPR32:$Rd),
1272 (ins GPR32:$Rn, bitfield32_imm:$ImmR, bfx32_width:$ImmS),
1273 !strconcat(asmop, "\t$Rd, $Rn, $ImmR, $ImmS"),
1274 [(set i32:$Rd, (op i32:$Rn, imm:$ImmR, imm:$ImmS))],
1276 Sched<[WriteALU, ReadALU]> {
1277 // As above, no disassembler allowed.
1278 let isAsmParserOnly = 1;
1281 def xxii : A64I_bitfield<0b1, opc, 0b1, (outs GPR64:$Rd),
1282 (ins GPR64:$Rn, bitfield64_imm:$ImmR, bfx64_width:$ImmS),
1283 !strconcat(asmop, "\t$Rd, $Rn, $ImmR, $ImmS"),
1284 [(set i64:$Rd, (op i64:$Rn, imm:$ImmR, imm:$ImmS))],
1286 Sched<[WriteALU, ReadALU]> {
1287 // As above, no disassembler allowed.
1288 let isAsmParserOnly = 1;
1292 defm SBFX : A64I_bitfield_extract<0b00, "sbfx", A64Sbfx>;
1293 defm UBFX : A64I_bitfield_extract<0b10, "ubfx", A64Ubfx>;
1295 // Again, variants based on BFM modify Rd so need it as an input too.
1296 def BFXILwwii : A64I_bitfield<0b0, 0b01, 0b0, (outs GPR32:$Rd),
1297 (ins GPR32:$src, GPR32:$Rn, bitfield32_imm:$ImmR, bfx32_width:$ImmS),
1298 "bfxil\t$Rd, $Rn, $ImmR, $ImmS", [], NoItinerary>,
1299 Sched<[WriteALU, ReadALU, ReadALU]> {
1300 // As above, no disassembler allowed.
1301 let isAsmParserOnly = 1;
1302 let Constraints = "$src = $Rd";
1305 def BFXILxxii : A64I_bitfield<0b1, 0b01, 0b1, (outs GPR64:$Rd),
1306 (ins GPR64:$src, GPR64:$Rn, bitfield64_imm:$ImmR, bfx64_width:$ImmS),
1307 "bfxil\t$Rd, $Rn, $ImmR, $ImmS", [], NoItinerary>,
1308 Sched<[WriteALU, ReadALU, ReadALU]> {
1309 // As above, no disassembler allowed.
1310 let isAsmParserOnly = 1;
1311 let Constraints = "$src = $Rd";
1314 // SBFX instructions can do a 1-instruction sign-extension of boolean values.
1315 def : Pat<(sext_inreg i64:$Rn, i1), (SBFXxxii $Rn, 0, 0)>;
1316 def : Pat<(sext_inreg i32:$Rn, i1), (SBFXwwii $Rn, 0, 0)>;
1317 def : Pat<(i64 (sext_inreg (anyext i32:$Rn), i1)),
1318 (SBFXxxii (SUBREG_TO_REG (i64 0), $Rn, sub_32), 0, 0)>;
1320 // UBFX makes sense as an implementation of a 64-bit zero-extension too. Could
1321 // use either 64-bit or 32-bit variant, but 32-bit might be more efficient.
1322 def : Pat<(i64 (zext i32:$Rn)), (SUBREG_TO_REG (i64 0), (UBFXwwii $Rn, 0, 31),
1325 //===-------------------------------
1326 // 6. Aliases for bitfield insert instructions
1327 //===-------------------------------
1329 def bfi32_lsb_asmoperand : AsmOperandClass {
1330 let Name = "BFI32LSB";
1331 let PredicateMethod = "isUImm<5>";
1332 let RenderMethod = "addBFILSBOperands<32>";
1333 let DiagnosticType = "UImm5";
1336 def bfi32_lsb : Operand<i64>,
1337 ImmLeaf<i64, [{ return Imm >= 0 && Imm <= 31; }]> {
1338 let PrintMethod = "printBFILSBOperand<32>";
1339 let ParserMatchClass = bfi32_lsb_asmoperand;
1342 def bfi64_lsb_asmoperand : AsmOperandClass {
1343 let Name = "BFI64LSB";
1344 let PredicateMethod = "isUImm<6>";
1345 let RenderMethod = "addBFILSBOperands<64>";
1346 let DiagnosticType = "UImm6";
1349 def bfi64_lsb : Operand<i64>,
1350 ImmLeaf<i64, [{ return Imm >= 0 && Imm <= 63; }]> {
1351 let PrintMethod = "printBFILSBOperand<64>";
1352 let ParserMatchClass = bfi64_lsb_asmoperand;
1355 // Width verification is performed during conversion so width operand can be
1356 // shared between 32/64-bit cases. Still needed for the print method though
1357 // because ImmR encodes "width - 1".
1358 def bfi32_width_asmoperand : AsmOperandClass {
1359 let Name = "BFI32Width";
1360 let PredicateMethod = "isBitfieldWidth<32>";
1361 let RenderMethod = "addBFIWidthOperands";
1362 let DiagnosticType = "Width32";
1365 def bfi32_width : Operand<i64>,
1366 ImmLeaf<i64, [{ return Imm >= 1 && Imm <= 32; }]> {
1367 let PrintMethod = "printBFIWidthOperand";
1368 let ParserMatchClass = bfi32_width_asmoperand;
1371 def bfi64_width_asmoperand : AsmOperandClass {
1372 let Name = "BFI64Width";
1373 let PredicateMethod = "isBitfieldWidth<64>";
1374 let RenderMethod = "addBFIWidthOperands";
1375 let DiagnosticType = "Width64";
1378 def bfi64_width : Operand<i64>,
1379 ImmLeaf<i64, [{ return Imm >= 1 && Imm <= 64; }]> {
1380 let PrintMethod = "printBFIWidthOperand";
1381 let ParserMatchClass = bfi64_width_asmoperand;
1384 multiclass A64I_bitfield_insert<bits<2> opc, string asmop> {
1385 def wwii : A64I_bitfield<0b0, opc, 0b0, (outs GPR32:$Rd),
1386 (ins GPR32:$Rn, bfi32_lsb:$ImmR, bfi32_width:$ImmS),
1387 !strconcat(asmop, "\t$Rd, $Rn, $ImmR, $ImmS"),
1389 Sched<[WriteALU, ReadALU]> {
1390 // As above, no disassembler allowed.
1391 let isAsmParserOnly = 1;
1394 def xxii : A64I_bitfield<0b1, opc, 0b1, (outs GPR64:$Rd),
1395 (ins GPR64:$Rn, bfi64_lsb:$ImmR, bfi64_width:$ImmS),
1396 !strconcat(asmop, "\t$Rd, $Rn, $ImmR, $ImmS"),
1398 Sched<[WriteALU, ReadALU]> {
1399 // As above, no disassembler allowed.
1400 let isAsmParserOnly = 1;
1404 defm SBFIZ : A64I_bitfield_insert<0b00, "sbfiz">;
1405 defm UBFIZ : A64I_bitfield_insert<0b10, "ubfiz">;
1408 def BFIwwii : A64I_bitfield<0b0, 0b01, 0b0, (outs GPR32:$Rd),
1409 (ins GPR32:$src, GPR32:$Rn, bfi32_lsb:$ImmR, bfi32_width:$ImmS),
1410 "bfi\t$Rd, $Rn, $ImmR, $ImmS", [], NoItinerary>,
1411 Sched<[WriteALU, ReadALU, ReadALU]> {
1412 // As above, no disassembler allowed.
1413 let isAsmParserOnly = 1;
1414 let Constraints = "$src = $Rd";
1417 def BFIxxii : A64I_bitfield<0b1, 0b01, 0b1, (outs GPR64:$Rd),
1418 (ins GPR64:$src, GPR64:$Rn, bfi64_lsb:$ImmR, bfi64_width:$ImmS),
1419 "bfi\t$Rd, $Rn, $ImmR, $ImmS", [], NoItinerary>,
1420 Sched<[WriteALU, ReadALU, ReadALU]> {
1421 // As above, no disassembler allowed.
1422 let isAsmParserOnly = 1;
1423 let Constraints = "$src = $Rd";
1426 //===----------------------------------------------------------------------===//
1427 // Compare and branch (immediate)
1428 //===----------------------------------------------------------------------===//
1429 // Contains: CBZ, CBNZ
1431 class label_asmoperand<int width, int scale> : AsmOperandClass {
1432 let Name = "Label" # width # "_" # scale;
1433 let PredicateMethod = "isLabel<" # width # "," # scale # ">";
1434 let RenderMethod = "addLabelOperands<" # width # ", " # scale # ">";
1435 let DiagnosticType = "Label";
1438 def label_wid19_scal4_asmoperand : label_asmoperand<19, 4>;
1440 // All conditional immediate branches are the same really: 19 signed bits scaled
1441 // by the instruction-size (4).
1442 def bcc_target : Operand<OtherVT> {
1443 // This label is a 19-bit offset from PC, scaled by the instruction-width: 4.
1444 let ParserMatchClass = label_wid19_scal4_asmoperand;
1445 let PrintMethod = "printLabelOperand<19, 4>";
1446 let EncoderMethod = "getLabelOpValue<AArch64::fixup_a64_condbr>";
1447 let OperandType = "OPERAND_PCREL";
1450 multiclass cmpbr_sizes<bit op, string asmop, ImmLeaf SETOP> {
1451 let isBranch = 1, isTerminator = 1 in {
1452 def x : A64I_cmpbr<0b1, op,
1454 (ins GPR64:$Rt, bcc_target:$Label),
1455 !strconcat(asmop,"\t$Rt, $Label"),
1456 [(A64br_cc (A64cmp i64:$Rt, 0), SETOP, bb:$Label)],
1458 Sched<[WriteBr, ReadBr]>;
1460 def w : A64I_cmpbr<0b0, op,
1462 (ins GPR32:$Rt, bcc_target:$Label),
1463 !strconcat(asmop,"\t$Rt, $Label"),
1464 [(A64br_cc (A64cmp i32:$Rt, 0), SETOP, bb:$Label)],
1466 Sched<[WriteBr, ReadBr]>;
1470 defm CBZ : cmpbr_sizes<0b0, "cbz", ImmLeaf<i32, [{
1471 return Imm == A64CC::EQ;
1473 defm CBNZ : cmpbr_sizes<0b1, "cbnz", ImmLeaf<i32, [{
1474 return Imm == A64CC::NE;
1477 //===----------------------------------------------------------------------===//
1478 // Conditional branch (immediate) instructions
1479 //===----------------------------------------------------------------------===//
1482 def cond_code_asmoperand : AsmOperandClass {
1483 let Name = "CondCode";
1484 let DiagnosticType = "CondCode";
1487 def cond_code : Operand<i32>, ImmLeaf<i32, [{
1488 return Imm >= 0 && Imm <= 15;
1490 let PrintMethod = "printCondCodeOperand";
1491 let ParserMatchClass = cond_code_asmoperand;
1494 def Bcc : A64I_condbr<0b0, 0b0, (outs),
1495 (ins cond_code:$Cond, bcc_target:$Label),
1496 "b.$Cond $Label", [(A64br_cc NZCV, (i32 imm:$Cond), bb:$Label)],
1501 let isTerminator = 1;
1504 //===----------------------------------------------------------------------===//
1505 // Conditional compare (immediate) instructions
1506 //===----------------------------------------------------------------------===//
1507 // Contains: CCMN, CCMP
1509 def uimm4_asmoperand : AsmOperandClass {
1511 let PredicateMethod = "isUImm<4>";
1512 let RenderMethod = "addImmOperands";
1513 let DiagnosticType = "UImm4";
1516 def uimm4 : Operand<i32> {
1517 let ParserMatchClass = uimm4_asmoperand;
1520 def uimm5 : Operand<i32> {
1521 let ParserMatchClass = uimm5_asmoperand;
1524 // The only difference between this operand and the one for instructions like
1525 // B.cc is that it's parsed manually. The other get parsed implicitly as part of
1526 // the mnemonic handling.
1527 def cond_code_op_asmoperand : AsmOperandClass {
1528 let Name = "CondCodeOp";
1529 let RenderMethod = "addCondCodeOperands";
1530 let PredicateMethod = "isCondCode";
1531 let ParserMethod = "ParseCondCodeOperand";
1532 let DiagnosticType = "CondCode";
1535 def cond_code_op : Operand<i32> {
1536 let PrintMethod = "printCondCodeOperand";
1537 let ParserMatchClass = cond_code_op_asmoperand;
1540 class A64I_condcmpimmImpl<bit sf, bit op, RegisterClass GPR, string asmop>
1541 : A64I_condcmpimm<sf, op, 0b0, 0b0, 0b1, (outs),
1542 (ins GPR:$Rn, uimm5:$UImm5, uimm4:$NZCVImm, cond_code_op:$Cond),
1543 !strconcat(asmop, "\t$Rn, $UImm5, $NZCVImm, $Cond"),
1545 Sched<[WriteCMP, ReadCMP]> {
1549 def CCMNwi : A64I_condcmpimmImpl<0b0, 0b0, GPR32, "ccmn">;
1550 def CCMNxi : A64I_condcmpimmImpl<0b1, 0b0, GPR64, "ccmn">;
1551 def CCMPwi : A64I_condcmpimmImpl<0b0, 0b1, GPR32, "ccmp">;
1552 def CCMPxi : A64I_condcmpimmImpl<0b1, 0b1, GPR64, "ccmp">;
1554 //===----------------------------------------------------------------------===//
1555 // Conditional compare (register) instructions
1556 //===----------------------------------------------------------------------===//
1557 // Contains: CCMN, CCMP
1559 class A64I_condcmpregImpl<bit sf, bit op, RegisterClass GPR, string asmop>
1560 : A64I_condcmpreg<sf, op, 0b0, 0b0, 0b1,
1562 (ins GPR:$Rn, GPR:$Rm, uimm4:$NZCVImm, cond_code_op:$Cond),
1563 !strconcat(asmop, "\t$Rn, $Rm, $NZCVImm, $Cond"),
1565 Sched<[WriteCMP, ReadCMP, ReadCMP]> {
1569 def CCMNww : A64I_condcmpregImpl<0b0, 0b0, GPR32, "ccmn">;
1570 def CCMNxx : A64I_condcmpregImpl<0b1, 0b0, GPR64, "ccmn">;
1571 def CCMPww : A64I_condcmpregImpl<0b0, 0b1, GPR32, "ccmp">;
1572 def CCMPxx : A64I_condcmpregImpl<0b1, 0b1, GPR64, "ccmp">;
1574 //===----------------------------------------------------------------------===//
1575 // Conditional select instructions
1576 //===----------------------------------------------------------------------===//
1577 // Contains: CSEL, CSINC, CSINV, CSNEG + aliases CSET, CSETM, CINC, CINV, CNEG
1579 // Condition code which is encoded as the inversion (semantically rather than
1580 // bitwise) in the instruction.
1581 def inv_cond_code_op_asmoperand : AsmOperandClass {
1582 let Name = "InvCondCodeOp";
1583 let RenderMethod = "addInvCondCodeOperands";
1584 let PredicateMethod = "isCondCode";
1585 let ParserMethod = "ParseCondCodeOperand";
1586 let DiagnosticType = "CondCode";
1589 def inv_cond_code_op : Operand<i32> {
1590 let ParserMatchClass = inv_cond_code_op_asmoperand;
1591 let PrintMethod = "printInverseCondCodeOperand";
1594 // Having a separate operand for the selectable use-case is debatable, but gives
1595 // consistency with cond_code.
1596 def inv_cond_XFORM : SDNodeXForm<imm, [{
1597 A64CC::CondCodes CC = static_cast<A64CC::CondCodes>(N->getZExtValue());
1598 return CurDAG->getTargetConstant(A64InvertCondCode(CC), MVT::i32);
1602 : ImmLeaf<i32, [{ return Imm >= 0 && Imm <= 15; }], inv_cond_XFORM>;
1605 multiclass A64I_condselSizes<bit op, bits<2> op2, string asmop,
1606 SDPatternOperator select> {
1607 let Uses = [NZCV] in {
1608 def wwwc : A64I_condsel<0b0, op, 0b0, op2,
1610 (ins GPR32:$Rn, GPR32:$Rm, cond_code_op:$Cond),
1611 !strconcat(asmop, "\t$Rd, $Rn, $Rm, $Cond"),
1612 [(set i32:$Rd, (select i32:$Rn, i32:$Rm))],
1614 Sched<[WriteCMP, ReadCMP, ReadCMP]>;
1617 def xxxc : A64I_condsel<0b1, op, 0b0, op2,
1619 (ins GPR64:$Rn, GPR64:$Rm, cond_code_op:$Cond),
1620 !strconcat(asmop, "\t$Rd, $Rn, $Rm, $Cond"),
1621 [(set i64:$Rd, (select i64:$Rn, i64:$Rm))],
1623 Sched<[WriteCMP, ReadCMP, ReadCMP]>;
1628 : PatFrag<(ops node:$lhs, node:$rhs),
1629 (A64select_cc NZCV, node:$lhs, node:$rhs, (i32 imm:$Cond))>;
1631 class complex_select<SDPatternOperator opnode>
1632 : PatFrag<(ops node:$lhs, node:$rhs),
1633 (A64select_cc NZCV, node:$lhs, (opnode node:$rhs), (i32 imm:$Cond))>;
1636 defm CSEL : A64I_condselSizes<0b0, 0b00, "csel", simple_select>;
1637 defm CSINC : A64I_condselSizes<0b0, 0b01, "csinc",
1638 complex_select<PatFrag<(ops node:$val),
1639 (add node:$val, 1)>>>;
1640 defm CSINV : A64I_condselSizes<0b1, 0b00, "csinv", complex_select<not>>;
1641 defm CSNEG : A64I_condselSizes<0b1, 0b01, "csneg", complex_select<ineg>>;
1643 // Now the instruction aliases, which fit nicely into LLVM's model:
1645 def : InstAlias<"cset $Rd, $Cond",
1646 (CSINCwwwc GPR32:$Rd, WZR, WZR, inv_cond_code_op:$Cond)>;
1647 def : InstAlias<"cset $Rd, $Cond",
1648 (CSINCxxxc GPR64:$Rd, XZR, XZR, inv_cond_code_op:$Cond)>;
1649 def : InstAlias<"csetm $Rd, $Cond",
1650 (CSINVwwwc GPR32:$Rd, WZR, WZR, inv_cond_code_op:$Cond)>;
1651 def : InstAlias<"csetm $Rd, $Cond",
1652 (CSINVxxxc GPR64:$Rd, XZR, XZR, inv_cond_code_op:$Cond)>;
1653 def : InstAlias<"cinc $Rd, $Rn, $Cond",
1654 (CSINCwwwc GPR32:$Rd, GPR32:$Rn, GPR32:$Rn, inv_cond_code_op:$Cond)>;
1655 def : InstAlias<"cinc $Rd, $Rn, $Cond",
1656 (CSINCxxxc GPR64:$Rd, GPR64:$Rn, GPR64:$Rn, inv_cond_code_op:$Cond)>;
1657 def : InstAlias<"cinv $Rd, $Rn, $Cond",
1658 (CSINVwwwc GPR32:$Rd, GPR32:$Rn, GPR32:$Rn, inv_cond_code_op:$Cond)>;
1659 def : InstAlias<"cinv $Rd, $Rn, $Cond",
1660 (CSINVxxxc GPR64:$Rd, GPR64:$Rn, GPR64:$Rn, inv_cond_code_op:$Cond)>;
1661 def : InstAlias<"cneg $Rd, $Rn, $Cond",
1662 (CSNEGwwwc GPR32:$Rd, GPR32:$Rn, GPR32:$Rn, inv_cond_code_op:$Cond)>;
1663 def : InstAlias<"cneg $Rd, $Rn, $Cond",
1664 (CSNEGxxxc GPR64:$Rd, GPR64:$Rn, GPR64:$Rn, inv_cond_code_op:$Cond)>;
1666 // Finally some helper patterns.
1668 // For CSET (a.k.a. zero-extension of icmp)
1669 def : Pat<(A64select_cc NZCV, 0, 1, cond_code:$Cond),
1670 (CSINCwwwc WZR, WZR, cond_code:$Cond)>;
1671 def : Pat<(A64select_cc NZCV, 1, 0, inv_cond_code:$Cond),
1672 (CSINCwwwc WZR, WZR, inv_cond_code:$Cond)>;
1674 def : Pat<(A64select_cc NZCV, 0, 1, cond_code:$Cond),
1675 (CSINCxxxc XZR, XZR, cond_code:$Cond)>;
1676 def : Pat<(A64select_cc NZCV, 1, 0, inv_cond_code:$Cond),
1677 (CSINCxxxc XZR, XZR, inv_cond_code:$Cond)>;
1679 // For CSETM (a.k.a. sign-extension of icmp)
1680 def : Pat<(A64select_cc NZCV, 0, -1, cond_code:$Cond),
1681 (CSINVwwwc WZR, WZR, cond_code:$Cond)>;
1682 def : Pat<(A64select_cc NZCV, -1, 0, inv_cond_code:$Cond),
1683 (CSINVwwwc WZR, WZR, inv_cond_code:$Cond)>;
1685 def : Pat<(A64select_cc NZCV, 0, -1, cond_code:$Cond),
1686 (CSINVxxxc XZR, XZR, cond_code:$Cond)>;
1687 def : Pat<(A64select_cc NZCV, -1, 0, inv_cond_code:$Cond),
1688 (CSINVxxxc XZR, XZR, inv_cond_code:$Cond)>;
1690 // CINC, CINV and CNEG get dealt with automatically, which leaves the issue of
1691 // commutativity. The instructions are to complex for isCommutable to be used,
1692 // so we have to create the patterns manually:
1694 // No commutable pattern for CSEL since the commuted version is isomorphic.
1697 def :Pat<(A64select_cc NZCV, (add i32:$Rm, 1), i32:$Rn, inv_cond_code:$Cond),
1698 (CSINCwwwc $Rn, $Rm, inv_cond_code:$Cond)>;
1699 def :Pat<(A64select_cc NZCV, (add i64:$Rm, 1), i64:$Rn, inv_cond_code:$Cond),
1700 (CSINCxxxc $Rn, $Rm, inv_cond_code:$Cond)>;
1703 def :Pat<(A64select_cc NZCV, (not i32:$Rm), i32:$Rn, inv_cond_code:$Cond),
1704 (CSINVwwwc $Rn, $Rm, inv_cond_code:$Cond)>;
1705 def :Pat<(A64select_cc NZCV, (not i64:$Rm), i64:$Rn, inv_cond_code:$Cond),
1706 (CSINVxxxc $Rn, $Rm, inv_cond_code:$Cond)>;
1709 def :Pat<(A64select_cc NZCV, (ineg i32:$Rm), i32:$Rn, inv_cond_code:$Cond),
1710 (CSNEGwwwc $Rn, $Rm, inv_cond_code:$Cond)>;
1711 def :Pat<(A64select_cc NZCV, (ineg i64:$Rm), i64:$Rn, inv_cond_code:$Cond),
1712 (CSNEGxxxc $Rn, $Rm, inv_cond_code:$Cond)>;
1714 //===----------------------------------------------------------------------===//
1715 // Data Processing (1 source) instructions
1716 //===----------------------------------------------------------------------===//
1717 // Contains: RBIT, REV16, REV, REV32, CLZ, CLS.
1719 // We define an unary operator which always fails. We will use this to
1720 // define unary operators that cannot be matched.
1722 class A64I_dp_1src_impl<bit sf, bits<6> opcode, string asmop,
1723 list<dag> patterns, RegisterClass GPRrc,
1724 InstrItinClass itin>:
1729 !strconcat(asmop, "\t$Rd, $Rn"),
1734 Sched<[WriteALU, ReadALU]>;
1736 multiclass A64I_dp_1src <bits<6> opcode, string asmop> {
1737 let hasSideEffects = 0 in {
1738 def ww : A64I_dp_1src_impl<0b0, opcode, asmop, [], GPR32, NoItinerary>;
1739 def xx : A64I_dp_1src_impl<0b1, opcode, asmop, [], GPR64, NoItinerary>;
1743 defm RBIT : A64I_dp_1src<0b000000, "rbit">;
1744 defm CLS : A64I_dp_1src<0b000101, "cls">;
1745 defm CLZ : A64I_dp_1src<0b000100, "clz">;
1747 def : Pat<(ctlz i32:$Rn), (CLZww $Rn)>;
1748 def : Pat<(ctlz i64:$Rn), (CLZxx $Rn)>;
1749 def : Pat<(ctlz_zero_undef i32:$Rn), (CLZww $Rn)>;
1750 def : Pat<(ctlz_zero_undef i64:$Rn), (CLZxx $Rn)>;
1752 def : Pat<(cttz i32:$Rn), (CLZww (RBITww $Rn))>;
1753 def : Pat<(cttz i64:$Rn), (CLZxx (RBITxx $Rn))>;
1754 def : Pat<(cttz_zero_undef i32:$Rn), (CLZww (RBITww $Rn))>;
1755 def : Pat<(cttz_zero_undef i64:$Rn), (CLZxx (RBITxx $Rn))>;
1758 def REVww : A64I_dp_1src_impl<0b0, 0b000010, "rev",
1759 [(set i32:$Rd, (bswap i32:$Rn))],
1760 GPR32, NoItinerary>;
1761 def REVxx : A64I_dp_1src_impl<0b1, 0b000011, "rev",
1762 [(set i64:$Rd, (bswap i64:$Rn))],
1763 GPR64, NoItinerary>;
1764 def REV32xx : A64I_dp_1src_impl<0b1, 0b000010, "rev32",
1765 [(set i64:$Rd, (bswap (rotr i64:$Rn, (i64 32))))],
1766 GPR64, NoItinerary>;
1767 def REV16ww : A64I_dp_1src_impl<0b0, 0b000001, "rev16",
1768 [(set i32:$Rd, (bswap (rotr i32:$Rn, (i64 16))))],
1771 def REV16xx : A64I_dp_1src_impl<0b1, 0b000001, "rev16", [], GPR64, NoItinerary>;
1773 //===----------------------------------------------------------------------===//
1774 // Data Processing (2 sources) instructions
1775 //===----------------------------------------------------------------------===//
1776 // Contains: CRC32C?[BHWX], UDIV, SDIV, LSLV, LSRV, ASRV, RORV + aliases LSL,
1780 class dp_2src_impl<bit sf, bits<6> opcode, string asmop, list<dag> patterns,
1781 RegisterClass GPRsp,
1782 InstrItinClass itin>:
1786 !strconcat(asmop, "\t$Rd, $Rn, $Rm"),
1788 (ins GPRsp:$Rn, GPRsp:$Rm),
1791 Sched<[WriteALU, ReadALU, ReadALU]>;
1793 multiclass dp_2src_crc<bit c, string asmop> {
1794 def B_www : dp_2src_impl<0b0, {0, 1, 0, c, 0, 0},
1795 !strconcat(asmop, "b"), [], GPR32, NoItinerary>;
1796 def H_www : dp_2src_impl<0b0, {0, 1, 0, c, 0, 1},
1797 !strconcat(asmop, "h"), [], GPR32, NoItinerary>;
1798 def W_www : dp_2src_impl<0b0, {0, 1, 0, c, 1, 0},
1799 !strconcat(asmop, "w"), [], GPR32, NoItinerary>;
1800 def X_wwx : A64I_dp_2src<0b1, {0, 1, 0, c, 1, 1}, 0b0,
1801 !strconcat(asmop, "x\t$Rd, $Rn, $Rm"),
1802 (outs GPR32:$Rd), (ins GPR32:$Rn, GPR64:$Rm), [],
1804 Sched<[WriteALU, ReadALU, ReadALU]>;
1807 multiclass dp_2src_zext <bits<6> opcode, string asmop, SDPatternOperator op> {
1808 def www : dp_2src_impl<0b0,
1812 (op i32:$Rn, (i64 (zext i32:$Rm))))],
1815 def xxx : dp_2src_impl<0b1,
1818 [(set i64:$Rd, (op i64:$Rn, i64:$Rm))],
1824 multiclass dp_2src <bits<6> opcode, string asmop, SDPatternOperator op> {
1825 def www : dp_2src_impl<0b0,
1828 [(set i32:$Rd, (op i32:$Rn, i32:$Rm))],
1831 def xxx : dp_2src_impl<0b1,
1834 [(set i64:$Rd, (op i64:$Rn, i64:$Rm))],
1839 // Here we define the data processing 2 source instructions.
1840 defm CRC32 : dp_2src_crc<0b0, "crc32">;
1841 defm CRC32C : dp_2src_crc<0b1, "crc32c">;
1843 let SchedRW = [WriteDiv, ReadDiv, ReadDiv] in {
1844 defm UDIV : dp_2src<0b000010, "udiv", udiv>;
1845 defm SDIV : dp_2src<0b000011, "sdiv", sdiv>;
1848 let SchedRW = [WriteALUs, ReadALU, ReadALU] in {
1849 defm LSLV : dp_2src_zext<0b001000, "lsl", shl>;
1850 defm LSRV : dp_2src_zext<0b001001, "lsr", srl>;
1851 defm ASRV : dp_2src_zext<0b001010, "asr", sra>;
1852 defm RORV : dp_2src_zext<0b001011, "ror", rotr>;
1855 // Extra patterns for an incoming 64-bit value for a 32-bit
1856 // operation. Since the LLVM operations are undefined (as in C) if the
1857 // RHS is out of range, it's perfectly permissible to discard the high
1858 // bits of the GPR64.
1859 def : Pat<(shl i32:$Rn, i64:$Rm),
1860 (LSLVwww $Rn, (EXTRACT_SUBREG $Rm, sub_32))>;
1861 def : Pat<(srl i32:$Rn, i64:$Rm),
1862 (LSRVwww $Rn, (EXTRACT_SUBREG $Rm, sub_32))>;
1863 def : Pat<(sra i32:$Rn, i64:$Rm),
1864 (ASRVwww $Rn, (EXTRACT_SUBREG $Rm, sub_32))>;
1865 def : Pat<(rotr i32:$Rn, i64:$Rm),
1866 (RORVwww $Rn, (EXTRACT_SUBREG $Rm, sub_32))>;
1868 // Here we define the aliases for the data processing 2 source instructions.
1869 def LSL_mnemonic : MnemonicAlias<"lslv", "lsl">;
1870 def LSR_mnemonic : MnemonicAlias<"lsrv", "lsr">;
1871 def ASR_menmonic : MnemonicAlias<"asrv", "asr">;
1872 def ROR_menmonic : MnemonicAlias<"rorv", "ror">;
1874 //===----------------------------------------------------------------------===//
1875 // Data Processing (3 sources) instructions
1876 //===----------------------------------------------------------------------===//
1877 // Contains: MADD, MSUB, SMADDL, SMSUBL, SMULH, UMADDL, UMSUBL, UMULH
1878 // + aliases MUL, MNEG, SMULL, SMNEGL, UMULL, UMNEGL
1880 class A64I_dp3_4operand<bit sf, bits<6> opcode, RegisterClass AccReg,
1881 ValueType AccTy, RegisterClass SrcReg,
1882 string asmop, dag pattern>
1883 : A64I_dp3<sf, opcode,
1884 (outs AccReg:$Rd), (ins SrcReg:$Rn, SrcReg:$Rm, AccReg:$Ra),
1885 !strconcat(asmop, "\t$Rd, $Rn, $Rm, $Ra"),
1886 [(set AccTy:$Rd, pattern)], NoItinerary>,
1887 Sched<[WriteMAC, ReadMAC, ReadMAC, ReadMAC]> {
1889 let Inst{14-10} = Ra;
1891 RegisterClass AccGPR = AccReg;
1892 RegisterClass SrcGPR = SrcReg;
1895 def MADDwwww : A64I_dp3_4operand<0b0, 0b000000, GPR32, i32, GPR32, "madd",
1896 (add i32:$Ra, (mul i32:$Rn, i32:$Rm))>;
1897 def MADDxxxx : A64I_dp3_4operand<0b1, 0b000000, GPR64, i64, GPR64, "madd",
1898 (add i64:$Ra, (mul i64:$Rn, i64:$Rm))>;
1900 def MSUBwwww : A64I_dp3_4operand<0b0, 0b000001, GPR32, i32, GPR32, "msub",
1901 (sub i32:$Ra, (mul i32:$Rn, i32:$Rm))>;
1902 def MSUBxxxx : A64I_dp3_4operand<0b1, 0b000001, GPR64, i64, GPR64, "msub",
1903 (sub i64:$Ra, (mul i64:$Rn, i64:$Rm))>;
1905 def SMADDLxwwx : A64I_dp3_4operand<0b1, 0b000010, GPR64, i64, GPR32, "smaddl",
1906 (add i64:$Ra, (mul (i64 (sext i32:$Rn)), (sext i32:$Rm)))>;
1907 def SMSUBLxwwx : A64I_dp3_4operand<0b1, 0b000011, GPR64, i64, GPR32, "smsubl",
1908 (sub i64:$Ra, (mul (i64 (sext i32:$Rn)), (sext i32:$Rm)))>;
1910 def UMADDLxwwx : A64I_dp3_4operand<0b1, 0b001010, GPR64, i64, GPR32, "umaddl",
1911 (add i64:$Ra, (mul (i64 (zext i32:$Rn)), (zext i32:$Rm)))>;
1912 def UMSUBLxwwx : A64I_dp3_4operand<0b1, 0b001011, GPR64, i64, GPR32, "umsubl",
1913 (sub i64:$Ra, (mul (i64 (zext i32:$Rn)), (zext i32:$Rm)))>;
1915 let isCommutable = 1, PostEncoderMethod = "fixMulHigh" in {
1916 def UMULHxxx : A64I_dp3<0b1, 0b001100, (outs GPR64:$Rd),
1917 (ins GPR64:$Rn, GPR64:$Rm),
1918 "umulh\t$Rd, $Rn, $Rm",
1919 [(set i64:$Rd, (mulhu i64:$Rn, i64:$Rm))],
1921 Sched<[WriteMAC, ReadMAC, ReadMAC]>;
1923 def SMULHxxx : A64I_dp3<0b1, 0b000100, (outs GPR64:$Rd),
1924 (ins GPR64:$Rn, GPR64:$Rm),
1925 "smulh\t$Rd, $Rn, $Rm",
1926 [(set i64:$Rd, (mulhs i64:$Rn, i64:$Rm))],
1928 Sched<[WriteMAC, ReadMAC, ReadMAC]>;
1931 multiclass A64I_dp3_3operand<string asmop, A64I_dp3_4operand INST,
1932 Register ZR, dag pattern> {
1933 def : InstAlias<asmop # " $Rd, $Rn, $Rm",
1934 (INST INST.AccGPR:$Rd, INST.SrcGPR:$Rn, INST.SrcGPR:$Rm, ZR)>;
1936 def : Pat<pattern, (INST $Rn, $Rm, ZR)>;
1939 defm : A64I_dp3_3operand<"mul", MADDwwww, WZR, (mul i32:$Rn, i32:$Rm)>;
1940 defm : A64I_dp3_3operand<"mul", MADDxxxx, XZR, (mul i64:$Rn, i64:$Rm)>;
1942 defm : A64I_dp3_3operand<"mneg", MSUBwwww, WZR,
1943 (sub 0, (mul i32:$Rn, i32:$Rm))>;
1944 defm : A64I_dp3_3operand<"mneg", MSUBxxxx, XZR,
1945 (sub 0, (mul i64:$Rn, i64:$Rm))>;
1947 defm : A64I_dp3_3operand<"smull", SMADDLxwwx, XZR,
1948 (mul (i64 (sext i32:$Rn)), (sext i32:$Rm))>;
1949 defm : A64I_dp3_3operand<"smnegl", SMSUBLxwwx, XZR,
1950 (sub 0, (mul (i64 (sext i32:$Rn)), (sext i32:$Rm)))>;
1952 defm : A64I_dp3_3operand<"umull", UMADDLxwwx, XZR,
1953 (mul (i64 (zext i32:$Rn)), (zext i32:$Rm))>;
1954 defm : A64I_dp3_3operand<"umnegl", UMSUBLxwwx, XZR,
1955 (sub 0, (mul (i64 (zext i32:$Rn)), (zext i32:$Rm)))>;
1958 //===----------------------------------------------------------------------===//
1959 // Exception generation
1960 //===----------------------------------------------------------------------===//
1961 // Contains: SVC, HVC, SMC, BRK, HLT, DCPS1, DCPS2, DCPS3
1963 def uimm16_asmoperand : AsmOperandClass {
1964 let Name = "UImm16";
1965 let PredicateMethod = "isUImm<16>";
1966 let RenderMethod = "addImmOperands";
1967 let DiagnosticType = "UImm16";
1970 def uimm16 : Operand<i32> {
1971 let ParserMatchClass = uimm16_asmoperand;
1974 class A64I_exceptImpl<bits<3> opc, bits<2> ll, string asmop>
1975 : A64I_exception<opc, 0b000, ll, (outs), (ins uimm16:$UImm16),
1976 !strconcat(asmop, "\t$UImm16"), [], NoItinerary>,
1979 let isTerminator = 1;
1982 def SVCi : A64I_exceptImpl<0b000, 0b01, "svc">;
1983 def HVCi : A64I_exceptImpl<0b000, 0b10, "hvc">;
1984 def SMCi : A64I_exceptImpl<0b000, 0b11, "smc">;
1985 def BRKi : A64I_exceptImpl<0b001, 0b00, "brk">;
1986 def HLTi : A64I_exceptImpl<0b010, 0b00, "hlt">;
1988 def DCPS1i : A64I_exceptImpl<0b101, 0b01, "dcps1">;
1989 def DCPS2i : A64I_exceptImpl<0b101, 0b10, "dcps2">;
1990 def DCPS3i : A64I_exceptImpl<0b101, 0b11, "dcps3">;
1992 // The immediate is optional for the DCPS instructions, defaulting to 0.
1993 def : InstAlias<"dcps1", (DCPS1i 0)>;
1994 def : InstAlias<"dcps2", (DCPS2i 0)>;
1995 def : InstAlias<"dcps3", (DCPS3i 0)>;
1997 //===----------------------------------------------------------------------===//
1998 // Extract (immediate)
1999 //===----------------------------------------------------------------------===//
2000 // Contains: EXTR + alias ROR
2002 def EXTRwwwi : A64I_extract<0b0, 0b000, 0b0,
2004 (ins GPR32:$Rn, GPR32:$Rm, bitfield32_imm:$LSB),
2005 "extr\t$Rd, $Rn, $Rm, $LSB",
2007 (A64Extr i32:$Rn, i32:$Rm, imm:$LSB))],
2009 Sched<[WriteALU, ReadALU, ReadALU]>;
2010 def EXTRxxxi : A64I_extract<0b1, 0b000, 0b1,
2012 (ins GPR64:$Rn, GPR64:$Rm, bitfield64_imm:$LSB),
2013 "extr\t$Rd, $Rn, $Rm, $LSB",
2015 (A64Extr i64:$Rn, i64:$Rm, imm:$LSB))],
2017 Sched<[WriteALU, ReadALU, ReadALU]>;
2019 def : InstAlias<"ror $Rd, $Rs, $LSB",
2020 (EXTRwwwi GPR32:$Rd, GPR32:$Rs, GPR32:$Rs, bitfield32_imm:$LSB)>;
2021 def : InstAlias<"ror $Rd, $Rs, $LSB",
2022 (EXTRxxxi GPR64:$Rd, GPR64:$Rs, GPR64:$Rs, bitfield64_imm:$LSB)>;
2024 def : Pat<(rotr i32:$Rn, bitfield32_imm:$LSB),
2025 (EXTRwwwi $Rn, $Rn, bitfield32_imm:$LSB)>;
2026 def : Pat<(rotr i64:$Rn, bitfield64_imm:$LSB),
2027 (EXTRxxxi $Rn, $Rn, bitfield64_imm:$LSB)>;
2029 //===----------------------------------------------------------------------===//
2030 // Floating-point compare instructions
2031 //===----------------------------------------------------------------------===//
2032 // Contains: FCMP, FCMPE
2034 def fpzero_asmoperand : AsmOperandClass {
2035 let Name = "FPZero";
2036 let ParserMethod = "ParseFPImmOperand";
2037 let DiagnosticType = "FPZero";
2040 def fpz32 : Operand<f32>,
2041 ComplexPattern<f32, 1, "SelectFPZeroOperand", [fpimm]> {
2042 let ParserMatchClass = fpzero_asmoperand;
2043 let PrintMethod = "printFPZeroOperand";
2044 let DecoderMethod = "DecodeFPZeroOperand";
2047 def fpz64 : Operand<f64>,
2048 ComplexPattern<f64, 1, "SelectFPZeroOperand", [fpimm]> {
2049 let ParserMatchClass = fpzero_asmoperand;
2050 let PrintMethod = "printFPZeroOperand";
2051 let DecoderMethod = "DecodeFPZeroOperand";
2054 def fpz64movi : Operand<i64>,
2055 ComplexPattern<f64, 1, "SelectFPZeroOperand", [fpimm]> {
2056 let ParserMatchClass = fpzero_asmoperand;
2057 let PrintMethod = "printFPZeroOperand";
2058 let DecoderMethod = "DecodeFPZeroOperand";
2061 multiclass A64I_fpcmpSignal<bits<2> type, bit imm, dag ins, dag pattern> {
2062 def _quiet : A64I_fpcmp<0b0, 0b0, type, 0b00, {0b0, imm, 0b0, 0b0, 0b0},
2063 (outs), ins, "fcmp\t$Rn, $Rm", [pattern],
2065 Sched<[WriteFPALU, ReadFPALU, ReadFPALU]> {
2069 def _sig : A64I_fpcmp<0b0, 0b0, type, 0b00, {0b1, imm, 0b0, 0b0, 0b0},
2070 (outs), ins, "fcmpe\t$Rn, $Rm", [], NoItinerary>,
2071 Sched<[WriteFPALU, ReadFPALU, ReadFPALU]> {
2076 defm FCMPss : A64I_fpcmpSignal<0b00, 0b0, (ins FPR32:$Rn, FPR32:$Rm),
2077 (set NZCV, (A64cmp f32:$Rn, f32:$Rm))>;
2078 defm FCMPdd : A64I_fpcmpSignal<0b01, 0b0, (ins FPR64:$Rn, FPR64:$Rm),
2079 (set NZCV, (A64cmp f64:$Rn, f64:$Rm))>;
2081 // What would be Rm should be written as 0; note that even though it's called
2082 // "$Rm" here to fit in with the InstrFormats, it's actually an immediate.
2083 defm FCMPsi : A64I_fpcmpSignal<0b00, 0b1, (ins FPR32:$Rn, fpz32:$Rm),
2084 (set NZCV, (A64cmp f32:$Rn, fpz32:$Rm))>;
2086 defm FCMPdi : A64I_fpcmpSignal<0b01, 0b1, (ins FPR64:$Rn, fpz64:$Rm),
2087 (set NZCV, (A64cmp f64:$Rn, fpz64:$Rm))>;
2090 //===----------------------------------------------------------------------===//
2091 // Floating-point conditional compare instructions
2092 //===----------------------------------------------------------------------===//
2093 // Contains: FCCMP, FCCMPE
2095 class A64I_fpccmpImpl<bits<2> type, bit op, RegisterClass FPR, string asmop>
2096 : A64I_fpccmp<0b0, 0b0, type, op,
2098 (ins FPR:$Rn, FPR:$Rm, uimm4:$NZCVImm, cond_code_op:$Cond),
2099 !strconcat(asmop, "\t$Rn, $Rm, $NZCVImm, $Cond"),
2101 Sched<[WriteFPALU, ReadFPALU, ReadFPALU]> {
2105 def FCCMPss : A64I_fpccmpImpl<0b00, 0b0, FPR32, "fccmp">;
2106 def FCCMPEss : A64I_fpccmpImpl<0b00, 0b1, FPR32, "fccmpe">;
2107 def FCCMPdd : A64I_fpccmpImpl<0b01, 0b0, FPR64, "fccmp">;
2108 def FCCMPEdd : A64I_fpccmpImpl<0b01, 0b1, FPR64, "fccmpe">;
2110 //===----------------------------------------------------------------------===//
2111 // Floating-point conditional select instructions
2112 //===----------------------------------------------------------------------===//
2115 let Uses = [NZCV] in {
2116 def FCSELsssc : A64I_fpcondsel<0b0, 0b0, 0b00, (outs FPR32:$Rd),
2117 (ins FPR32:$Rn, FPR32:$Rm, cond_code_op:$Cond),
2118 "fcsel\t$Rd, $Rn, $Rm, $Cond",
2120 (simple_select f32:$Rn, f32:$Rm))],
2122 Sched<[WriteFPALU, ReadFPALU, ReadFPALU]>;
2125 def FCSELdddc : A64I_fpcondsel<0b0, 0b0, 0b01, (outs FPR64:$Rd),
2126 (ins FPR64:$Rn, FPR64:$Rm, cond_code_op:$Cond),
2127 "fcsel\t$Rd, $Rn, $Rm, $Cond",
2129 (simple_select f64:$Rn, f64:$Rm))],
2131 Sched<[WriteFPALU, ReadFPALU, ReadFPALU]>;
2134 //===----------------------------------------------------------------------===//
2135 // Floating-point data-processing (1 source)
2136 //===----------------------------------------------------------------------===//
2137 // Contains: FMOV, FABS, FNEG, FSQRT, FCVT, FRINT[NPMZAXI].
2139 def FPNoUnop : PatFrag<(ops node:$val), (fneg node:$val),
2140 [{ (void)N; return false; }]>;
2142 // First we do the fairly trivial bunch with uniform "OP s, s" and "OP d, d"
2143 // syntax. Default to no pattern because most are odd enough not to have one.
2144 multiclass A64I_fpdp1sizes<bits<6> opcode, string asmstr,
2145 SDPatternOperator opnode = FPNoUnop> {
2146 def ss : A64I_fpdp1<0b0, 0b0, 0b00, opcode, (outs FPR32:$Rd), (ins FPR32:$Rn),
2147 !strconcat(asmstr, "\t$Rd, $Rn"),
2148 [(set f32:$Rd, (opnode f32:$Rn))],
2150 Sched<[WriteFPALU, ReadFPALU]>;
2152 def dd : A64I_fpdp1<0b0, 0b0, 0b01, opcode, (outs FPR64:$Rd), (ins FPR64:$Rn),
2153 !strconcat(asmstr, "\t$Rd, $Rn"),
2154 [(set f64:$Rd, (opnode f64:$Rn))],
2156 Sched<[WriteFPALU, ReadFPALU]>;
2159 defm FMOV : A64I_fpdp1sizes<0b000000, "fmov">;
2160 defm FABS : A64I_fpdp1sizes<0b000001, "fabs", fabs>;
2161 defm FNEG : A64I_fpdp1sizes<0b000010, "fneg", fneg>;
2162 let SchedRW = [WriteFPSqrt, ReadFPSqrt] in {
2163 defm FSQRT : A64I_fpdp1sizes<0b000011, "fsqrt", fsqrt>;
2166 defm FRINTN : A64I_fpdp1sizes<0b001000, "frintn">;
2167 defm FRINTP : A64I_fpdp1sizes<0b001001, "frintp", fceil>;
2168 defm FRINTM : A64I_fpdp1sizes<0b001010, "frintm", ffloor>;
2169 defm FRINTZ : A64I_fpdp1sizes<0b001011, "frintz", ftrunc>;
2170 defm FRINTA : A64I_fpdp1sizes<0b001100, "frinta">;
2171 defm FRINTX : A64I_fpdp1sizes<0b001110, "frintx", frint>;
2172 defm FRINTI : A64I_fpdp1sizes<0b001111, "frinti", fnearbyint>;
2174 // The FCVT instrucitons have different source and destination register-types,
2175 // but the fields are uniform everywhere a D-register (say) crops up. Package
2176 // this information in a Record.
2177 class FCVTRegType<RegisterClass rc, bits<2> fld, ValueType vt> {
2178 RegisterClass Class = rc;
2184 def FCVT16 : FCVTRegType<FPR16, 0b11, f16>;
2185 def FCVT32 : FCVTRegType<FPR32, 0b00, f32>;
2186 def FCVT64 : FCVTRegType<FPR64, 0b01, f64>;
2188 class A64I_fpdp1_fcvt<FCVTRegType DestReg, FCVTRegType SrcReg, SDNode opnode>
2189 : A64I_fpdp1<0b0, 0b0, {SrcReg.t1, SrcReg.t0},
2190 {0,0,0,1, DestReg.t1, DestReg.t0},
2191 (outs DestReg.Class:$Rd), (ins SrcReg.Class:$Rn),
2193 [(set DestReg.VT:$Rd, (opnode SrcReg.VT:$Rn))], NoItinerary>,
2194 Sched<[WriteFPALU, ReadFPALU]>;
2196 def FCVTds : A64I_fpdp1_fcvt<FCVT64, FCVT32, fextend>;
2197 def FCVThs : A64I_fpdp1_fcvt<FCVT16, FCVT32, fround>;
2198 def FCVTsd : A64I_fpdp1_fcvt<FCVT32, FCVT64, fround>;
2199 def FCVThd : A64I_fpdp1_fcvt<FCVT16, FCVT64, fround>;
2200 def FCVTsh : A64I_fpdp1_fcvt<FCVT32, FCVT16, fextend>;
2201 def FCVTdh : A64I_fpdp1_fcvt<FCVT64, FCVT16, fextend>;
2204 //===----------------------------------------------------------------------===//
2205 // Floating-point data-processing (2 sources) instructions
2206 //===----------------------------------------------------------------------===//
2207 // Contains: FMUL, FDIV, FADD, FSUB, FMAX, FMIN, FMAXNM, FMINNM, FNMUL
2209 def FPNoBinop : PatFrag<(ops node:$lhs, node:$rhs), (fadd node:$lhs, node:$rhs),
2210 [{ (void)N; return false; }]>;
2212 multiclass A64I_fpdp2sizes<bits<4> opcode, string asmstr,
2213 SDPatternOperator opnode> {
2214 def sss : A64I_fpdp2<0b0, 0b0, 0b00, opcode,
2216 (ins FPR32:$Rn, FPR32:$Rm),
2217 !strconcat(asmstr, "\t$Rd, $Rn, $Rm"),
2218 [(set f32:$Rd, (opnode f32:$Rn, f32:$Rm))],
2220 Sched<[WriteFPALU, ReadFPALU, ReadFPALU]>;
2222 def ddd : A64I_fpdp2<0b0, 0b0, 0b01, opcode,
2224 (ins FPR64:$Rn, FPR64:$Rm),
2225 !strconcat(asmstr, "\t$Rd, $Rn, $Rm"),
2226 [(set f64:$Rd, (opnode f64:$Rn, f64:$Rm))],
2228 Sched<[WriteFPALU, ReadFPALU, ReadFPALU]>;
2231 let isCommutable = 1 in {
2232 let SchedRW = [WriteFPMul, ReadFPMul, ReadFPMul] in {
2233 defm FMUL : A64I_fpdp2sizes<0b0000, "fmul", fmul>;
2235 defm FADD : A64I_fpdp2sizes<0b0010, "fadd", fadd>;
2237 // No patterns for these.
2238 defm FMAX : A64I_fpdp2sizes<0b0100, "fmax", FPNoBinop>;
2239 defm FMIN : A64I_fpdp2sizes<0b0101, "fmin", FPNoBinop>;
2240 defm FMAXNM : A64I_fpdp2sizes<0b0110, "fmaxnm", FPNoBinop>;
2241 defm FMINNM : A64I_fpdp2sizes<0b0111, "fminnm", FPNoBinop>;
2243 let SchedRW = [WriteFPMul, ReadFPMul, ReadFPMul] in {
2244 defm FNMUL : A64I_fpdp2sizes<0b1000, "fnmul",
2245 PatFrag<(ops node:$lhs, node:$rhs),
2246 (fneg (fmul node:$lhs, node:$rhs))> >;
2250 let SchedRW = [WriteFPDiv, ReadFPDiv, ReadFPDiv] in {
2251 defm FDIV : A64I_fpdp2sizes<0b0001, "fdiv", fdiv>;
2253 defm FSUB : A64I_fpdp2sizes<0b0011, "fsub", fsub>;
2255 //===----------------------------------------------------------------------===//
2256 // Floating-point data-processing (3 sources) instructions
2257 //===----------------------------------------------------------------------===//
2258 // Contains: FMADD, FMSUB, FNMADD, FNMSUB
2260 def fmsub : PatFrag<(ops node:$Rn, node:$Rm, node:$Ra),
2261 (fma (fneg node:$Rn), node:$Rm, node:$Ra)>;
2262 def fnmsub : PatFrag<(ops node:$Rn, node:$Rm, node:$Ra),
2263 (fma node:$Rn, node:$Rm, (fneg node:$Ra))>;
2264 def fnmadd : PatFrag<(ops node:$Rn, node:$Rm, node:$Ra),
2265 (fma (fneg node:$Rn), node:$Rm, (fneg node:$Ra))>;
2267 class A64I_fpdp3Impl<string asmop, RegisterClass FPR, ValueType VT,
2268 bits<2> type, bit o1, bit o0, SDPatternOperator fmakind>
2269 : A64I_fpdp3<0b0, 0b0, type, o1, o0, (outs FPR:$Rd),
2270 (ins FPR:$Rn, FPR:$Rm, FPR:$Ra),
2271 !strconcat(asmop,"\t$Rd, $Rn, $Rm, $Ra"),
2272 [(set VT:$Rd, (fmakind VT:$Rn, VT:$Rm, VT:$Ra))],
2274 Sched<[WriteFPMAC, ReadFPMAC, ReadFPMAC, ReadFPMAC]>;
2276 def FMADDssss : A64I_fpdp3Impl<"fmadd", FPR32, f32, 0b00, 0b0, 0b0, fma>;
2277 def FMSUBssss : A64I_fpdp3Impl<"fmsub", FPR32, f32, 0b00, 0b0, 0b1, fmsub>;
2278 def FNMADDssss : A64I_fpdp3Impl<"fnmadd", FPR32, f32, 0b00, 0b1, 0b0, fnmadd>;
2279 def FNMSUBssss : A64I_fpdp3Impl<"fnmsub", FPR32, f32, 0b00, 0b1, 0b1, fnmsub>;
2281 def FMADDdddd : A64I_fpdp3Impl<"fmadd", FPR64, f64, 0b01, 0b0, 0b0, fma>;
2282 def FMSUBdddd : A64I_fpdp3Impl<"fmsub", FPR64, f64, 0b01, 0b0, 0b1, fmsub>;
2283 def FNMADDdddd : A64I_fpdp3Impl<"fnmadd", FPR64, f64, 0b01, 0b1, 0b0, fnmadd>;
2284 def FNMSUBdddd : A64I_fpdp3Impl<"fnmsub", FPR64, f64, 0b01, 0b1, 0b1, fnmsub>;
2286 // Extra patterns for when we're allowed to optimise separate multiplication and
2288 let Predicates = [HasFPARMv8, UseFusedMAC] in {
2289 def : Pat<(f32 (fadd FPR32:$Ra, (f32 (fmul_su FPR32:$Rn, FPR32:$Rm)))),
2290 (FMADDssss FPR32:$Rn, FPR32:$Rm, FPR32:$Ra)>;
2291 def : Pat<(f32 (fsub FPR32:$Ra, (f32 (fmul_su FPR32:$Rn, FPR32:$Rm)))),
2292 (FMSUBssss FPR32:$Rn, FPR32:$Rm, FPR32:$Ra)>;
2293 def : Pat<(f32 (fsub (f32 (fneg FPR32:$Ra)), (f32 (fmul_su FPR32:$Rn, FPR32:$Rm)))),
2294 (FNMADDssss FPR32:$Rn, FPR32:$Rm, FPR32:$Ra)>;
2295 def : Pat<(f32 (fsub (f32 (fmul_su FPR32:$Rn, FPR32:$Rm)), FPR32:$Ra)),
2296 (FNMSUBssss FPR32:$Rn, FPR32:$Rm, FPR32:$Ra)>;
2298 def : Pat<(f64 (fadd FPR64:$Ra, (f64 (fmul_su FPR64:$Rn, FPR64:$Rm)))),
2299 (FMADDdddd FPR64:$Rn, FPR64:$Rm, FPR64:$Ra)>;
2300 def : Pat<(f64 (fsub FPR64:$Ra, (f64 (fmul_su FPR64:$Rn, FPR64:$Rm)))),
2301 (FMSUBdddd FPR64:$Rn, FPR64:$Rm, FPR64:$Ra)>;
2302 def : Pat<(f64 (fsub (f64 (fneg FPR64:$Ra)), (f64 (fmul_su FPR64:$Rn, FPR64:$Rm)))),
2303 (FNMADDdddd FPR64:$Rn, FPR64:$Rm, FPR64:$Ra)>;
2304 def : Pat<(f64 (fsub (f64 (fmul_su FPR64:$Rn, FPR64:$Rm)), FPR64:$Ra)),
2305 (FNMSUBdddd FPR64:$Rn, FPR64:$Rm, FPR64:$Ra)>;
2309 //===----------------------------------------------------------------------===//
2310 // Floating-point <-> fixed-point conversion instructions
2311 //===----------------------------------------------------------------------===//
2312 // Contains: FCVTZS, FCVTZU, SCVTF, UCVTF
2314 // #1-#32 allowed, encoded as "64 - <specified imm>
2315 def fixedpos_asmoperand_i32 : AsmOperandClass {
2316 let Name = "CVTFixedPos32";
2317 let RenderMethod = "addCVTFixedPosOperands";
2318 let PredicateMethod = "isCVTFixedPos<32>";
2319 let DiagnosticType = "CVTFixedPos32";
2322 // Also encoded as "64 - <specified imm>" but #1-#64 allowed.
2323 def fixedpos_asmoperand_i64 : AsmOperandClass {
2324 let Name = "CVTFixedPos64";
2325 let RenderMethod = "addCVTFixedPosOperands";
2326 let PredicateMethod = "isCVTFixedPos<64>";
2327 let DiagnosticType = "CVTFixedPos64";
2330 // We need the cartesian product of f32/f64 i32/i64 operands for
2332 // + Selection needs to use operands of correct floating type
2333 // + Assembly parsing and decoding depend on integer width
2334 class cvtfix_i32_op<ValueType FloatVT>
2336 ComplexPattern<FloatVT, 1, "SelectCVTFixedPosOperand<32>", [fpimm]> {
2337 let ParserMatchClass = fixedpos_asmoperand_i32;
2338 let DecoderMethod = "DecodeCVT32FixedPosOperand";
2339 let PrintMethod = "printCVTFixedPosOperand";
2342 class cvtfix_i64_op<ValueType FloatVT>
2344 ComplexPattern<FloatVT, 1, "SelectCVTFixedPosOperand<64>", [fpimm]> {
2345 let ParserMatchClass = fixedpos_asmoperand_i64;
2346 let PrintMethod = "printCVTFixedPosOperand";
2349 // Because of the proliferation of weird operands, it's not really
2350 // worth going for a multiclass here. Oh well.
2352 class A64I_fptofix<bit sf, bits<2> type, bits<3> opcode,
2353 RegisterClass GPR, RegisterClass FPR,
2354 ValueType DstTy, ValueType SrcTy,
2355 Operand scale_op, string asmop, SDNode cvtop>
2356 : A64I_fpfixed<sf, 0b0, type, 0b11, opcode,
2357 (outs GPR:$Rd), (ins FPR:$Rn, scale_op:$Scale),
2358 !strconcat(asmop, "\t$Rd, $Rn, $Scale"),
2359 [(set DstTy:$Rd, (cvtop (fmul SrcTy:$Rn, scale_op:$Scale)))],
2361 Sched<[WriteFPALU, ReadFPALU]>;
2363 def FCVTZSwsi : A64I_fptofix<0b0, 0b00, 0b000, GPR32, FPR32, i32, f32,
2364 cvtfix_i32_op<f32>, "fcvtzs", fp_to_sint>;
2365 def FCVTZSxsi : A64I_fptofix<0b1, 0b00, 0b000, GPR64, FPR32, i64, f32,
2366 cvtfix_i64_op<f32>, "fcvtzs", fp_to_sint>;
2367 def FCVTZUwsi : A64I_fptofix<0b0, 0b00, 0b001, GPR32, FPR32, i32, f32,
2368 cvtfix_i32_op<f32>, "fcvtzu", fp_to_uint>;
2369 def FCVTZUxsi : A64I_fptofix<0b1, 0b00, 0b001, GPR64, FPR32, i64, f32,
2370 cvtfix_i64_op<f32>, "fcvtzu", fp_to_uint>;
2372 def FCVTZSwdi : A64I_fptofix<0b0, 0b01, 0b000, GPR32, FPR64, i32, f64,
2373 cvtfix_i32_op<f64>, "fcvtzs", fp_to_sint>;
2374 def FCVTZSxdi : A64I_fptofix<0b1, 0b01, 0b000, GPR64, FPR64, i64, f64,
2375 cvtfix_i64_op<f64>, "fcvtzs", fp_to_sint>;
2376 def FCVTZUwdi : A64I_fptofix<0b0, 0b01, 0b001, GPR32, FPR64, i32, f64,
2377 cvtfix_i32_op<f64>, "fcvtzu", fp_to_uint>;
2378 def FCVTZUxdi : A64I_fptofix<0b1, 0b01, 0b001, GPR64, FPR64, i64, f64,
2379 cvtfix_i64_op<f64>, "fcvtzu", fp_to_uint>;
2382 class A64I_fixtofp<bit sf, bits<2> type, bits<3> opcode,
2383 RegisterClass FPR, RegisterClass GPR,
2384 ValueType DstTy, ValueType SrcTy,
2385 Operand scale_op, string asmop, SDNode cvtop>
2386 : A64I_fpfixed<sf, 0b0, type, 0b00, opcode,
2387 (outs FPR:$Rd), (ins GPR:$Rn, scale_op:$Scale),
2388 !strconcat(asmop, "\t$Rd, $Rn, $Scale"),
2389 [(set DstTy:$Rd, (fdiv (cvtop SrcTy:$Rn), scale_op:$Scale))],
2391 Sched<[WriteFPALU, ReadFPALU]>;
2393 def SCVTFswi : A64I_fixtofp<0b0, 0b00, 0b010, FPR32, GPR32, f32, i32,
2394 cvtfix_i32_op<f32>, "scvtf", sint_to_fp>;
2395 def SCVTFsxi : A64I_fixtofp<0b1, 0b00, 0b010, FPR32, GPR64, f32, i64,
2396 cvtfix_i64_op<f32>, "scvtf", sint_to_fp>;
2397 def UCVTFswi : A64I_fixtofp<0b0, 0b00, 0b011, FPR32, GPR32, f32, i32,
2398 cvtfix_i32_op<f32>, "ucvtf", uint_to_fp>;
2399 def UCVTFsxi : A64I_fixtofp<0b1, 0b00, 0b011, FPR32, GPR64, f32, i64,
2400 cvtfix_i64_op<f32>, "ucvtf", uint_to_fp>;
2401 def SCVTFdwi : A64I_fixtofp<0b0, 0b01, 0b010, FPR64, GPR32, f64, i32,
2402 cvtfix_i32_op<f64>, "scvtf", sint_to_fp>;
2403 def SCVTFdxi : A64I_fixtofp<0b1, 0b01, 0b010, FPR64, GPR64, f64, i64,
2404 cvtfix_i64_op<f64>, "scvtf", sint_to_fp>;
2405 def UCVTFdwi : A64I_fixtofp<0b0, 0b01, 0b011, FPR64, GPR32, f64, i32,
2406 cvtfix_i32_op<f64>, "ucvtf", uint_to_fp>;
2407 def UCVTFdxi : A64I_fixtofp<0b1, 0b01, 0b011, FPR64, GPR64, f64, i64,
2408 cvtfix_i64_op<f64>, "ucvtf", uint_to_fp>;
2410 //===----------------------------------------------------------------------===//
2411 // Floating-point <-> integer conversion instructions
2412 //===----------------------------------------------------------------------===//
2413 // Contains: FCVTZS, FCVTZU, SCVTF, UCVTF
2415 class A64I_fpintI<bit sf, bits<2> type, bits<2> rmode, bits<3> opcode,
2416 RegisterClass DestPR, RegisterClass SrcPR, string asmop>
2417 : A64I_fpint<sf, 0b0, type, rmode, opcode, (outs DestPR:$Rd), (ins SrcPR:$Rn),
2418 !strconcat(asmop, "\t$Rd, $Rn"), [], NoItinerary>,
2419 Sched<[WriteFPALU, ReadFPALU]>;
2421 multiclass A64I_fptointRM<bits<2> rmode, bit o2, string asmop> {
2422 def Sws : A64I_fpintI<0b0, 0b00, rmode, {o2, 0, 0},
2423 GPR32, FPR32, asmop # "s">;
2424 def Sxs : A64I_fpintI<0b1, 0b00, rmode, {o2, 0, 0},
2425 GPR64, FPR32, asmop # "s">;
2426 def Uws : A64I_fpintI<0b0, 0b00, rmode, {o2, 0, 1},
2427 GPR32, FPR32, asmop # "u">;
2428 def Uxs : A64I_fpintI<0b1, 0b00, rmode, {o2, 0, 1},
2429 GPR64, FPR32, asmop # "u">;
2431 def Swd : A64I_fpintI<0b0, 0b01, rmode, {o2, 0, 0},
2432 GPR32, FPR64, asmop # "s">;
2433 def Sxd : A64I_fpintI<0b1, 0b01, rmode, {o2, 0, 0},
2434 GPR64, FPR64, asmop # "s">;
2435 def Uwd : A64I_fpintI<0b0, 0b01, rmode, {o2, 0, 1},
2436 GPR32, FPR64, asmop # "u">;
2437 def Uxd : A64I_fpintI<0b1, 0b01, rmode, {o2, 0, 1},
2438 GPR64, FPR64, asmop # "u">;
2441 defm FCVTN : A64I_fptointRM<0b00, 0b0, "fcvtn">;
2442 defm FCVTP : A64I_fptointRM<0b01, 0b0, "fcvtp">;
2443 defm FCVTM : A64I_fptointRM<0b10, 0b0, "fcvtm">;
2444 defm FCVTZ : A64I_fptointRM<0b11, 0b0, "fcvtz">;
2445 defm FCVTA : A64I_fptointRM<0b00, 0b1, "fcvta">;
2447 let Predicates = [HasFPARMv8] in {
2448 def : Pat<(i32 (fp_to_sint f32:$Rn)), (FCVTZSws $Rn)>;
2449 def : Pat<(i64 (fp_to_sint f32:$Rn)), (FCVTZSxs $Rn)>;
2450 def : Pat<(i32 (fp_to_uint f32:$Rn)), (FCVTZUws $Rn)>;
2451 def : Pat<(i64 (fp_to_uint f32:$Rn)), (FCVTZUxs $Rn)>;
2452 def : Pat<(i32 (fp_to_sint f64:$Rn)), (FCVTZSwd $Rn)>;
2453 def : Pat<(i64 (fp_to_sint f64:$Rn)), (FCVTZSxd $Rn)>;
2454 def : Pat<(i32 (fp_to_uint f64:$Rn)), (FCVTZUwd $Rn)>;
2455 def : Pat<(i64 (fp_to_uint f64:$Rn)), (FCVTZUxd $Rn)>;
2458 multiclass A64I_inttofp<bit o0, string asmop> {
2459 def CVTFsw : A64I_fpintI<0b0, 0b00, 0b00, {0, 1, o0}, FPR32, GPR32, asmop>;
2460 def CVTFsx : A64I_fpintI<0b1, 0b00, 0b00, {0, 1, o0}, FPR32, GPR64, asmop>;
2461 def CVTFdw : A64I_fpintI<0b0, 0b01, 0b00, {0, 1, o0}, FPR64, GPR32, asmop>;
2462 def CVTFdx : A64I_fpintI<0b1, 0b01, 0b00, {0, 1, o0}, FPR64, GPR64, asmop>;
2465 defm S : A64I_inttofp<0b0, "scvtf">;
2466 defm U : A64I_inttofp<0b1, "ucvtf">;
2468 let Predicates = [HasFPARMv8] in {
2469 def : Pat<(f32 (sint_to_fp i32:$Rn)), (SCVTFsw $Rn)>;
2470 def : Pat<(f32 (sint_to_fp i64:$Rn)), (SCVTFsx $Rn)>;
2471 def : Pat<(f64 (sint_to_fp i32:$Rn)), (SCVTFdw $Rn)>;
2472 def : Pat<(f64 (sint_to_fp i64:$Rn)), (SCVTFdx $Rn)>;
2473 def : Pat<(f32 (uint_to_fp i32:$Rn)), (UCVTFsw $Rn)>;
2474 def : Pat<(f32 (uint_to_fp i64:$Rn)), (UCVTFsx $Rn)>;
2475 def : Pat<(f64 (uint_to_fp i32:$Rn)), (UCVTFdw $Rn)>;
2476 def : Pat<(f64 (uint_to_fp i64:$Rn)), (UCVTFdx $Rn)>;
2479 def FMOVws : A64I_fpintI<0b0, 0b00, 0b00, 0b110, GPR32, FPR32, "fmov">;
2480 def FMOVsw : A64I_fpintI<0b0, 0b00, 0b00, 0b111, FPR32, GPR32, "fmov">;
2481 def FMOVxd : A64I_fpintI<0b1, 0b01, 0b00, 0b110, GPR64, FPR64, "fmov">;
2482 def FMOVdx : A64I_fpintI<0b1, 0b01, 0b00, 0b111, FPR64, GPR64, "fmov">;
2484 let Predicates = [HasFPARMv8] in {
2485 def : Pat<(i32 (bitconvert f32:$Rn)), (FMOVws $Rn)>;
2486 def : Pat<(f32 (bitconvert i32:$Rn)), (FMOVsw $Rn)>;
2487 def : Pat<(i64 (bitconvert f64:$Rn)), (FMOVxd $Rn)>;
2488 def : Pat<(f64 (bitconvert i64:$Rn)), (FMOVdx $Rn)>;
2491 def lane1_asmoperand : AsmOperandClass {
2493 let RenderMethod = "addImmOperands";
2494 let DiagnosticType = "Lane1";
2497 def lane1 : Operand<i32> {
2498 let ParserMatchClass = lane1_asmoperand;
2499 let PrintMethod = "printBareImmOperand";
2502 let DecoderMethod = "DecodeFMOVLaneInstruction" in {
2503 def FMOVxv : A64I_fpint<0b1, 0b0, 0b10, 0b01, 0b110,
2504 (outs GPR64:$Rd), (ins VPR128:$Rn, lane1:$Lane),
2505 "fmov\t$Rd, $Rn.d[$Lane]", [], NoItinerary>,
2506 Sched<[WriteFPALU, ReadFPALU]>;
2508 def FMOVvx : A64I_fpint<0b1, 0b0, 0b10, 0b01, 0b111,
2509 (outs VPR128:$Rd), (ins GPR64:$Rn, lane1:$Lane),
2510 "fmov\t$Rd.d[$Lane], $Rn", [], NoItinerary>,
2511 Sched<[WriteFPALU, ReadFPALU]>;
2514 let Predicates = [HasFPARMv8] in {
2515 def : InstAlias<"fmov $Rd, $Rn.2d[$Lane]",
2516 (FMOVxv GPR64:$Rd, VPR128:$Rn, lane1:$Lane), 0b0>;
2518 def : InstAlias<"fmov $Rd.2d[$Lane], $Rn",
2519 (FMOVvx VPR128:$Rd, GPR64:$Rn, lane1:$Lane), 0b0>;
2522 //===----------------------------------------------------------------------===//
2523 // Floating-point immediate instructions
2524 //===----------------------------------------------------------------------===//
2527 def fpimm_asmoperand : AsmOperandClass {
2528 let Name = "FMOVImm";
2529 let ParserMethod = "ParseFPImmOperand";
2530 let DiagnosticType = "FPImm";
2533 // The MCOperand for these instructions are the encoded 8-bit values.
2534 def SDXF_fpimm : SDNodeXForm<fpimm, [{
2536 A64Imms::isFPImm(N->getValueAPF(), Imm8);
2537 return CurDAG->getTargetConstant(Imm8, MVT::i32);
2540 class fmov_operand<ValueType FT>
2542 PatLeaf<(FT fpimm), [{ return A64Imms::isFPImm(N->getValueAPF()); }],
2544 let PrintMethod = "printFPImmOperand";
2545 let ParserMatchClass = fpimm_asmoperand;
2548 def fmov32_operand : fmov_operand<f32>;
2549 def fmov64_operand : fmov_operand<f64>;
2551 class A64I_fpimm_impl<bits<2> type, RegisterClass Reg, ValueType VT,
2552 Operand fmov_operand>
2553 : A64I_fpimm<0b0, 0b0, type, 0b00000,
2555 (ins fmov_operand:$Imm8),
2557 [(set VT:$Rd, fmov_operand:$Imm8)],
2559 Sched<[WriteFPALU]>;
2561 def FMOVsi : A64I_fpimm_impl<0b00, FPR32, f32, fmov32_operand>;
2562 def FMOVdi : A64I_fpimm_impl<0b01, FPR64, f64, fmov64_operand>;
2564 //===----------------------------------------------------------------------===//
2565 // Load-register (literal) instructions
2566 //===----------------------------------------------------------------------===//
2567 // Contains: LDR, LDRSW, PRFM
2569 def ldrlit_label_asmoperand : AsmOperandClass {
2570 let Name = "LoadLitLabel";
2571 let RenderMethod = "addLabelOperands<19, 4>";
2572 let DiagnosticType = "Label";
2575 def ldrlit_label : Operand<i64> {
2576 let EncoderMethod = "getLoadLitLabelOpValue";
2578 // This label is a 19-bit offset from PC, scaled by the instruction-width: 4.
2579 let PrintMethod = "printLabelOperand<19, 4>";
2580 let ParserMatchClass = ldrlit_label_asmoperand;
2581 let OperandType = "OPERAND_PCREL";
2584 // Various instructions take an immediate value (which can always be used),
2585 // where some numbers have a symbolic name to make things easier. These operands
2586 // and the associated functions abstract away the differences.
2587 multiclass namedimm<string prefix, string mapper> {
2588 def _asmoperand : AsmOperandClass {
2589 let Name = "NamedImm" # prefix;
2590 let PredicateMethod = "isUImm";
2591 let RenderMethod = "addImmOperands";
2592 let ParserMethod = "ParseNamedImmOperand<" # mapper # ">";
2593 let DiagnosticType = "NamedImm_" # prefix;
2596 def _op : Operand<i32> {
2597 let ParserMatchClass = !cast<AsmOperandClass>(prefix # "_asmoperand");
2598 let PrintMethod = "printNamedImmOperand<" # mapper # ">";
2599 let DecoderMethod = "DecodeNamedImmOperand<" # mapper # ">";
2603 defm prefetch : namedimm<"prefetch", "A64PRFM::PRFMMapper">;
2605 class A64I_LDRlitSimple<bits<2> opc, bit v, RegisterClass OutReg,
2606 list<dag> patterns = []>
2607 : A64I_LDRlit<opc, v, (outs OutReg:$Rt), (ins ldrlit_label:$Imm19),
2608 "ldr\t$Rt, $Imm19", patterns, NoItinerary>,
2611 let mayLoad = 1 in {
2612 def LDRw_lit : A64I_LDRlitSimple<0b00, 0b0, GPR32>;
2613 def LDRx_lit : A64I_LDRlitSimple<0b01, 0b0, GPR64>;
2616 let Predicates = [HasFPARMv8] in {
2617 def LDRs_lit : A64I_LDRlitSimple<0b00, 0b1, FPR32>;
2618 def LDRd_lit : A64I_LDRlitSimple<0b01, 0b1, FPR64>;
2621 let mayLoad = 1 in {
2622 let Predicates = [HasFPARMv8] in {
2623 def LDRq_lit : A64I_LDRlitSimple<0b10, 0b1, FPR128>;
2626 def LDRSWx_lit : A64I_LDRlit<0b10, 0b0,
2628 (ins ldrlit_label:$Imm19),
2629 "ldrsw\t$Rt, $Imm19",
2633 def PRFM_lit : A64I_LDRlit<0b11, 0b0,
2634 (outs), (ins prefetch_op:$Rt, ldrlit_label:$Imm19),
2635 "prfm\t$Rt, $Imm19",
2637 Sched<[WriteLd, ReadLd]>;
2640 //===----------------------------------------------------------------------===//
2641 // Load-store exclusive instructions
2642 //===----------------------------------------------------------------------===//
2643 // Contains: STXRB, STXRH, STXR, LDXRB, LDXRH, LDXR. STXP, LDXP, STLXRB,
2644 // STLXRH, STLXR, LDAXRB, LDAXRH, LDAXR, STLXP, LDAXP, STLRB,
2645 // STLRH, STLR, LDARB, LDARH, LDAR
2647 // Since these instructions have the undefined register bits set to 1 in
2648 // their canonical form, we need a post encoder method to set those bits
2649 // to 1 when encoding these instructions. We do this using the
2650 // fixLoadStoreExclusive function. This function has template parameters:
2652 // fixLoadStoreExclusive<int hasRs, int hasRt2>
2654 // hasRs indicates that the instruction uses the Rs field, so we won't set
2655 // it to 1 (and the same for Rt2). We don't need template parameters for
2656 // the other register fiels since Rt and Rn are always used.
2658 // This operand parses a GPR64xsp register, followed by an optional immediate
2660 def GPR64xsp0_asmoperand : AsmOperandClass {
2661 let Name = "GPR64xsp0";
2662 let PredicateMethod = "isWrappedReg";
2663 let RenderMethod = "addRegOperands";
2664 let ParserMethod = "ParseLSXAddressOperand";
2665 // Diagnostics are provided by ParserMethod
2668 def GPR64xsp0 : RegisterOperand<GPR64xsp> {
2669 let ParserMatchClass = GPR64xsp0_asmoperand;
2672 //===----------------------------------
2673 // Store-exclusive (releasing & normal)
2674 //===----------------------------------
2676 class A64I_SRexs_impl<bits<2> size, bits<3> opcode, string asm, dag outs,
2677 dag ins, list<dag> pat,
2678 InstrItinClass itin> :
2679 A64I_LDSTex_stn <size,
2680 opcode{2}, 0, opcode{1}, opcode{0},
2682 !strconcat(asm, "\t$Rs, $Rt, [$Rn]"),
2685 let PostEncoderMethod = "fixLoadStoreExclusive<1,0>";
2686 let Constraints = "@earlyclobber $Rs";
2689 multiclass A64I_SRex<string asmstr, bits<3> opcode, string prefix> {
2690 def _byte: A64I_SRexs_impl<0b00, opcode, !strconcat(asmstr, "b"),
2691 (outs GPR32:$Rs), (ins GPR32:$Rt, GPR64xsp0:$Rn),
2693 Sched<[WriteSt, ReadSt, ReadSt]>;
2695 def _hword: A64I_SRexs_impl<0b01, opcode, !strconcat(asmstr, "h"),
2696 (outs GPR32:$Rs), (ins GPR32:$Rt, GPR64xsp0:$Rn),
2698 Sched<[WriteSt, ReadSt, ReadSt]>;
2700 def _word: A64I_SRexs_impl<0b10, opcode, asmstr,
2701 (outs GPR32:$Rs), (ins GPR32:$Rt, GPR64xsp0:$Rn),
2703 Sched<[WriteSt, ReadSt, ReadSt]>;
2705 def _dword: A64I_SRexs_impl<0b11, opcode, asmstr,
2706 (outs GPR32:$Rs), (ins GPR64:$Rt, GPR64xsp0:$Rn),
2708 Sched<[WriteSt, ReadSt, ReadSt]>;
2711 defm STXR : A64I_SRex<"stxr", 0b000, "STXR">;
2712 defm STLXR : A64I_SRex<"stlxr", 0b001, "STLXR">;
2714 //===----------------------------------
2716 //===----------------------------------
2718 class A64I_LRexs_impl<bits<2> size, bits<3> opcode, string asm, dag outs,
2719 dag ins, list<dag> pat,
2720 InstrItinClass itin> :
2721 A64I_LDSTex_tn <size,
2722 opcode{2}, 1, opcode{1}, opcode{0},
2724 !strconcat(asm, "\t$Rt, [$Rn]"),
2727 let PostEncoderMethod = "fixLoadStoreExclusive<0,0>";
2730 multiclass A64I_LRex<string asmstr, bits<3> opcode> {
2731 def _byte: A64I_LRexs_impl<0b00, opcode, !strconcat(asmstr, "b"),
2732 (outs GPR32:$Rt), (ins GPR64xsp0:$Rn),
2736 def _hword: A64I_LRexs_impl<0b01, opcode, !strconcat(asmstr, "h"),
2737 (outs GPR32:$Rt), (ins GPR64xsp0:$Rn),
2741 def _word: A64I_LRexs_impl<0b10, opcode, asmstr,
2742 (outs GPR32:$Rt), (ins GPR64xsp0:$Rn),
2746 def _dword: A64I_LRexs_impl<0b11, opcode, asmstr,
2747 (outs GPR64:$Rt), (ins GPR64xsp0:$Rn),
2752 defm LDXR : A64I_LRex<"ldxr", 0b000>;
2753 defm LDAXR : A64I_LRex<"ldaxr", 0b001>;
2754 defm LDAR : A64I_LRex<"ldar", 0b101>;
2756 class acquiring_load<PatFrag base>
2757 : PatFrag<(ops node:$ptr), (base node:$ptr), [{
2758 AtomicOrdering Ordering = cast<AtomicSDNode>(N)->getOrdering();
2759 return Ordering == Acquire || Ordering == SequentiallyConsistent;
2762 def atomic_load_acquire_8 : acquiring_load<atomic_load_8>;
2763 def atomic_load_acquire_16 : acquiring_load<atomic_load_16>;
2764 def atomic_load_acquire_32 : acquiring_load<atomic_load_32>;
2765 def atomic_load_acquire_64 : acquiring_load<atomic_load_64>;
2767 def : Pat<(atomic_load_acquire_8 i64:$Rn), (LDAR_byte $Rn)>;
2768 def : Pat<(atomic_load_acquire_16 i64:$Rn), (LDAR_hword $Rn)>;
2769 def : Pat<(atomic_load_acquire_32 i64:$Rn), (LDAR_word $Rn)>;
2770 def : Pat<(atomic_load_acquire_64 i64:$Rn), (LDAR_dword $Rn)>;
2772 //===----------------------------------
2773 // Store-release (no exclusivity)
2774 //===----------------------------------
2776 class A64I_SLexs_impl<bits<2> size, bits<3> opcode, string asm, dag outs,
2777 dag ins, list<dag> pat,
2778 InstrItinClass itin> :
2779 A64I_LDSTex_tn <size,
2780 opcode{2}, 0, opcode{1}, opcode{0},
2782 !strconcat(asm, "\t$Rt, [$Rn]"),
2785 let PostEncoderMethod = "fixLoadStoreExclusive<0,0>";
2788 class releasing_store<PatFrag base>
2789 : PatFrag<(ops node:$ptr, node:$val), (base node:$ptr, node:$val), [{
2790 AtomicOrdering Ordering = cast<AtomicSDNode>(N)->getOrdering();
2791 return Ordering == Release || Ordering == SequentiallyConsistent;
2794 def atomic_store_release_8 : releasing_store<atomic_store_8>;
2795 def atomic_store_release_16 : releasing_store<atomic_store_16>;
2796 def atomic_store_release_32 : releasing_store<atomic_store_32>;
2797 def atomic_store_release_64 : releasing_store<atomic_store_64>;
2799 multiclass A64I_SLex<string asmstr, bits<3> opcode, string prefix> {
2800 def _byte: A64I_SLexs_impl<0b00, opcode, !strconcat(asmstr, "b"),
2801 (outs), (ins GPR32:$Rt, GPR64xsp0:$Rn),
2802 [(atomic_store_release_8 i64:$Rn, i32:$Rt)],
2804 Sched<[WriteSt, ReadSt, ReadSt]>;
2806 def _hword: A64I_SLexs_impl<0b01, opcode, !strconcat(asmstr, "h"),
2807 (outs), (ins GPR32:$Rt, GPR64xsp0:$Rn),
2808 [(atomic_store_release_16 i64:$Rn, i32:$Rt)],
2810 Sched<[WriteSt, ReadSt, ReadSt]>;
2812 def _word: A64I_SLexs_impl<0b10, opcode, asmstr,
2813 (outs), (ins GPR32:$Rt, GPR64xsp0:$Rn),
2814 [(atomic_store_release_32 i64:$Rn, i32:$Rt)],
2816 Sched<[WriteSt, ReadSt, ReadSt]>;
2818 def _dword: A64I_SLexs_impl<0b11, opcode, asmstr,
2819 (outs), (ins GPR64:$Rt, GPR64xsp0:$Rn),
2820 [(atomic_store_release_64 i64:$Rn, i64:$Rt)],
2822 Sched<[WriteSt, ReadSt, ReadSt]>;
2825 defm STLR : A64I_SLex<"stlr", 0b101, "STLR">;
2827 //===----------------------------------
2828 // Store-exclusive pair (releasing & normal)
2829 //===----------------------------------
2831 class A64I_SPexs_impl<bits<2> size, bits<3> opcode, string asm, dag outs,
2832 dag ins, list<dag> pat,
2833 InstrItinClass itin> :
2834 A64I_LDSTex_stt2n <size,
2835 opcode{2}, 0, opcode{1}, opcode{0},
2837 !strconcat(asm, "\t$Rs, $Rt, $Rt2, [$Rn]"),
2843 multiclass A64I_SPex<string asmstr, bits<3> opcode> {
2844 def _word: A64I_SPexs_impl<0b10, opcode, asmstr, (outs),
2845 (ins GPR32:$Rs, GPR32:$Rt, GPR32:$Rt2,
2848 Sched<[WriteSt, ReadSt, ReadSt, ReadSt]>;
2850 def _dword: A64I_SPexs_impl<0b11, opcode, asmstr, (outs),
2851 (ins GPR32:$Rs, GPR64:$Rt, GPR64:$Rt2,
2854 Sched<[WriteSt, ReadSt, ReadSt, ReadSt]>;
2857 defm STXP : A64I_SPex<"stxp", 0b010>;
2858 defm STLXP : A64I_SPex<"stlxp", 0b011>;
2860 //===----------------------------------
2861 // Load-exclusive pair (acquiring & normal)
2862 //===----------------------------------
2864 class A64I_LPexs_impl<bits<2> size, bits<3> opcode, string asm, dag outs,
2865 dag ins, list<dag> pat,
2866 InstrItinClass itin> :
2867 A64I_LDSTex_tt2n <size,
2868 opcode{2}, 1, opcode{1}, opcode{0},
2870 !strconcat(asm, "\t$Rt, $Rt2, [$Rn]"),
2873 let DecoderMethod = "DecodeLoadPairExclusiveInstruction";
2874 let PostEncoderMethod = "fixLoadStoreExclusive<0,1>";
2877 multiclass A64I_LPex<string asmstr, bits<3> opcode> {
2878 def _word: A64I_LPexs_impl<0b10, opcode, asmstr,
2879 (outs GPR32:$Rt, GPR32:$Rt2),
2880 (ins GPR64xsp0:$Rn),
2882 Sched<[WriteLd, WriteLd, ReadLd]>;
2884 def _dword: A64I_LPexs_impl<0b11, opcode, asmstr,
2885 (outs GPR64:$Rt, GPR64:$Rt2),
2886 (ins GPR64xsp0:$Rn),
2888 Sched<[WriteLd, WriteLd, ReadLd]>;
2891 defm LDXP : A64I_LPex<"ldxp", 0b010>;
2892 defm LDAXP : A64I_LPex<"ldaxp", 0b011>;
2894 //===----------------------------------------------------------------------===//
2895 // Load-store register (unscaled immediate) instructions
2896 //===----------------------------------------------------------------------===//
2897 // Contains: LDURB, LDURH, LDRUSB, LDRUSH, LDRUSW, STUR, STURB, STURH and PRFUM
2901 //===----------------------------------------------------------------------===//
2902 // Load-store register (register offset) instructions
2903 //===----------------------------------------------------------------------===//
2904 // Contains: LDRB, LDRH, LDRSB, LDRSH, LDRSW, STR, STRB, STRH and PRFM
2908 //===----------------------------------------------------------------------===//
2909 // Load-store register (unsigned immediate) instructions
2910 //===----------------------------------------------------------------------===//
2911 // Contains: LDRB, LDRH, LDRSB, LDRSH, LDRSW, STR, STRB, STRH and PRFM
2915 //===----------------------------------------------------------------------===//
2916 // Load-store register (immediate post-indexed) instructions
2917 //===----------------------------------------------------------------------===//
2918 // Contains: STRB, STRH, STR, LDRB, LDRH, LDR, LDRSB, LDRSH, LDRSW
2922 //===----------------------------------------------------------------------===//
2923 // Load-store register (immediate pre-indexed) instructions
2924 //===----------------------------------------------------------------------===//
2925 // Contains: STRB, STRH, STR, LDRB, LDRH, LDR, LDRSB, LDRSH, LDRSW
2927 // Note that patterns are much later on in a completely separate section (they
2928 // need ADRPxi to be defined).
2930 //===-------------------------------
2931 // 1. Various operands needed
2932 //===-------------------------------
2934 //===-------------------------------
2935 // 1.1 Unsigned 12-bit immediate operands
2936 //===-------------------------------
2937 // The addressing mode for these instructions consists of an unsigned 12-bit
2938 // immediate which is scaled by the size of the memory access.
2940 // We represent this in the MC layer by two operands:
2941 // 1. A base register.
2942 // 2. A 12-bit immediate: not multiplied by access size, so "LDR x0,[x0,#8]"
2943 // would have '1' in this field.
2944 // This means that separate functions are needed for converting representations
2945 // which *are* aware of the intended access size.
2947 // Anything that creates an MCInst (Decoding, selection and AsmParsing) has to
2948 // know the access size via some means. An isolated operand does not have this
2949 // information unless told from here, which means we need separate tablegen
2950 // Operands for each access size. This multiclass takes care of instantiating
2951 // the correct template functions in the rest of the backend.
2953 //===-------------------------------
2954 // 1.1 Unsigned 12-bit immediate operands
2955 //===-------------------------------
2957 multiclass offsets_uimm12<int MemSize, string prefix> {
2958 def uimm12_asmoperand : AsmOperandClass {
2959 let Name = "OffsetUImm12_" # MemSize;
2960 let PredicateMethod = "isOffsetUImm12<" # MemSize # ">";
2961 let RenderMethod = "addOffsetUImm12Operands<" # MemSize # ">";
2962 let DiagnosticType = "LoadStoreUImm12_" # MemSize;
2965 // Pattern is really no more than an ImmLeaf, but predicated on MemSize which
2966 // complicates things beyond TableGen's ken.
2967 def uimm12 : Operand<i64>,
2968 ComplexPattern<i64, 1, "SelectOffsetUImm12<" # MemSize # ">"> {
2969 let ParserMatchClass
2970 = !cast<AsmOperandClass>(prefix # uimm12_asmoperand);
2972 let PrintMethod = "printOffsetUImm12Operand<" # MemSize # ">";
2973 let EncoderMethod = "getOffsetUImm12OpValue<" # MemSize # ">";
2977 defm byte_ : offsets_uimm12<1, "byte_">;
2978 defm hword_ : offsets_uimm12<2, "hword_">;
2979 defm word_ : offsets_uimm12<4, "word_">;
2980 defm dword_ : offsets_uimm12<8, "dword_">;
2981 defm qword_ : offsets_uimm12<16, "qword_">;
2983 //===-------------------------------
2984 // 1.1 Signed 9-bit immediate operands
2985 //===-------------------------------
2987 // The MCInst is expected to store the bit-wise encoding of the value,
2988 // which amounts to lopping off the extended sign bits.
2989 def SDXF_simm9 : SDNodeXForm<imm, [{
2990 return CurDAG->getTargetConstant(N->getZExtValue() & 0x1ff, MVT::i32);
2993 def simm9_asmoperand : AsmOperandClass {
2995 let PredicateMethod = "isSImm<9>";
2996 let RenderMethod = "addSImmOperands<9>";
2997 let DiagnosticType = "LoadStoreSImm9";
3000 def simm9 : Operand<i64>,
3001 ImmLeaf<i64, [{ return Imm >= -0x100 && Imm <= 0xff; }],
3003 let PrintMethod = "printOffsetSImm9Operand";
3004 let ParserMatchClass = simm9_asmoperand;
3008 //===-------------------------------
3009 // 1.3 Register offset extensions
3010 //===-------------------------------
3012 // The assembly-syntax for these addressing-modes is:
3013 // [<Xn|SP>, <R><m> {, <extend> {<amount>}}]
3015 // The essential semantics are:
3016 // + <amount> is a shift: #<log(transfer size)> or #0
3017 // + <R> can be W or X.
3018 // + If <R> is W, <extend> can be UXTW or SXTW
3019 // + If <R> is X, <extend> can be LSL or SXTX
3021 // The trickiest of those constraints is that Rm can be either GPR32 or GPR64,
3022 // which will need separate instructions for LLVM type-consistency. We'll also
3023 // need separate operands, of course.
3024 multiclass regexts<int MemSize, int RmSize, RegisterClass GPR,
3025 string Rm, string prefix> {
3026 def regext_asmoperand : AsmOperandClass {
3027 let Name = "AddrRegExtend_" # MemSize # "_" # Rm;
3028 let PredicateMethod = "isAddrRegExtend<" # MemSize # "," # RmSize # ">";
3029 let RenderMethod = "addAddrRegExtendOperands<" # MemSize # ">";
3030 let DiagnosticType = "LoadStoreExtend" # RmSize # "_" # MemSize;
3033 def regext : Operand<i64> {
3035 = "printAddrRegExtendOperand<" # MemSize # ", " # RmSize # ">";
3037 let DecoderMethod = "DecodeAddrRegExtendOperand";
3038 let ParserMatchClass
3039 = !cast<AsmOperandClass>(prefix # regext_asmoperand);
3043 multiclass regexts_wx<int MemSize, string prefix> {
3044 // Rm is an X-register if LSL or SXTX are specified as the shift.
3045 defm Xm_ : regexts<MemSize, 64, GPR64, "Xm", prefix # "Xm_">;
3047 // Rm is a W-register if UXTW or SXTW are specified as the shift.
3048 defm Wm_ : regexts<MemSize, 32, GPR32, "Wm", prefix # "Wm_">;
3051 defm byte_ : regexts_wx<1, "byte_">;
3052 defm hword_ : regexts_wx<2, "hword_">;
3053 defm word_ : regexts_wx<4, "word_">;
3054 defm dword_ : regexts_wx<8, "dword_">;
3055 defm qword_ : regexts_wx<16, "qword_">;
3058 //===------------------------------
3059 // 2. The instructions themselves.
3060 //===------------------------------
3062 // We have the following instructions to implement:
3063 // | | B | H | W | X |
3064 // |-----------------+-------+-------+-------+--------|
3065 // | unsigned str | STRB | STRH | STR | STR |
3066 // | unsigned ldr | LDRB | LDRH | LDR | LDR |
3067 // | signed ldr to W | LDRSB | LDRSH | - | - |
3068 // | signed ldr to X | LDRSB | LDRSH | LDRSW | (PRFM) |
3070 // This will instantiate the LDR/STR instructions you'd expect to use for an
3071 // unsigned datatype (first two rows above) or floating-point register, which is
3072 // reasonably uniform across all access sizes.
3075 //===------------------------------
3076 // 2.1 Regular instructions
3077 //===------------------------------
3079 // This class covers the basic unsigned or irrelevantly-signed loads and stores,
3080 // to general-purpose and floating-point registers.
3082 class AddrParams<string prefix> {
3083 Operand uimm12 = !cast<Operand>(prefix # "_uimm12");
3085 Operand regextWm = !cast<Operand>(prefix # "_Wm_regext");
3086 Operand regextXm = !cast<Operand>(prefix # "_Xm_regext");
3089 def byte_addrparams : AddrParams<"byte">;
3090 def hword_addrparams : AddrParams<"hword">;
3091 def word_addrparams : AddrParams<"word">;
3092 def dword_addrparams : AddrParams<"dword">;
3093 def qword_addrparams : AddrParams<"qword">;
3095 multiclass A64I_LDRSTR_unsigned<string prefix, bits<2> size, bit v,
3096 bit high_opc, string asmsuffix,
3097 RegisterClass GPR, AddrParams params> {
3098 // Unsigned immediate
3099 def _STR : A64I_LSunsigimm<size, v, {high_opc, 0b0},
3100 (outs), (ins GPR:$Rt, GPR64xsp:$Rn, params.uimm12:$UImm12),
3101 "str" # asmsuffix # "\t$Rt, [$Rn, $UImm12]",
3103 Sched<[WriteSt, ReadSt, ReadSt]> {
3106 def : InstAlias<"str" # asmsuffix # " $Rt, [$Rn]",
3107 (!cast<Instruction>(prefix # "_STR") GPR:$Rt, GPR64xsp:$Rn, 0)>;
3109 def _LDR : A64I_LSunsigimm<size, v, {high_opc, 0b1},
3110 (outs GPR:$Rt), (ins GPR64xsp:$Rn, params.uimm12:$UImm12),
3111 "ldr" # asmsuffix # "\t$Rt, [$Rn, $UImm12]",
3113 Sched<[WriteLd, ReadLd]> {
3116 def : InstAlias<"ldr" # asmsuffix # " $Rt, [$Rn]",
3117 (!cast<Instruction>(prefix # "_LDR") GPR:$Rt, GPR64xsp:$Rn, 0)>;
3119 // Register offset (four of these: load/store and Wm/Xm).
3120 let mayLoad = 1 in {
3121 def _Wm_RegOffset_LDR : A64I_LSregoff<size, v, {high_opc, 0b1}, 0b0,
3123 (ins GPR64xsp:$Rn, GPR32:$Rm, params.regextWm:$Ext),
3124 "ldr" # asmsuffix # "\t$Rt, [$Rn, $Rm, $Ext]",
3126 Sched<[WriteLd, ReadLd, ReadLd]>;
3128 def _Xm_RegOffset_LDR : A64I_LSregoff<size, v, {high_opc, 0b1}, 0b1,
3130 (ins GPR64xsp:$Rn, GPR64:$Rm, params.regextXm:$Ext),
3131 "ldr" # asmsuffix # "\t$Rt, [$Rn, $Rm, $Ext]",
3133 Sched<[WriteLd, ReadLd, ReadLd]>;
3135 def : InstAlias<"ldr" # asmsuffix # " $Rt, [$Rn, $Rm]",
3136 (!cast<Instruction>(prefix # "_Xm_RegOffset_LDR") GPR:$Rt, GPR64xsp:$Rn,
3139 let mayStore = 1 in {
3140 def _Wm_RegOffset_STR : A64I_LSregoff<size, v, {high_opc, 0b0}, 0b0,
3141 (outs), (ins GPR:$Rt, GPR64xsp:$Rn, GPR32:$Rm,
3142 params.regextWm:$Ext),
3143 "str" # asmsuffix # "\t$Rt, [$Rn, $Rm, $Ext]",
3145 Sched<[WriteSt, ReadSt, ReadSt, ReadSt]>;
3147 def _Xm_RegOffset_STR : A64I_LSregoff<size, v, {high_opc, 0b0}, 0b1,
3148 (outs), (ins GPR:$Rt, GPR64xsp:$Rn, GPR64:$Rm,
3149 params.regextXm:$Ext),
3150 "str" # asmsuffix # "\t$Rt, [$Rn, $Rm, $Ext]",
3152 Sched<[WriteSt, ReadSt, ReadSt, ReadSt]>;
3154 def : InstAlias<"str" # asmsuffix # " $Rt, [$Rn, $Rm]",
3155 (!cast<Instruction>(prefix # "_Xm_RegOffset_STR") GPR:$Rt, GPR64xsp:$Rn,
3158 // Unaligned immediate
3159 def _STUR : A64I_LSunalimm<size, v, {high_opc, 0b0},
3160 (outs), (ins GPR:$Rt, GPR64xsp:$Rn, simm9:$SImm9),
3161 "stur" # asmsuffix # "\t$Rt, [$Rn, $SImm9]",
3163 Sched<[WriteSt, ReadSt, ReadSt]> {
3166 def : InstAlias<"stur" # asmsuffix # " $Rt, [$Rn]",
3167 (!cast<Instruction>(prefix # "_STUR") GPR:$Rt, GPR64xsp:$Rn, 0)>;
3169 def _LDUR : A64I_LSunalimm<size, v, {high_opc, 0b1},
3170 (outs GPR:$Rt), (ins GPR64xsp:$Rn, simm9:$SImm9),
3171 "ldur" # asmsuffix # "\t$Rt, [$Rn, $SImm9]",
3173 Sched<[WriteLd, ReadLd]> {
3176 def : InstAlias<"ldur" # asmsuffix # " $Rt, [$Rn]",
3177 (!cast<Instruction>(prefix # "_LDUR") GPR:$Rt, GPR64xsp:$Rn, 0)>;
3180 def _PostInd_STR : A64I_LSpostind<size, v, {high_opc, 0b0},
3181 (outs GPR64xsp:$Rn_wb),
3182 (ins GPR:$Rt, GPR64xsp:$Rn, simm9:$SImm9),
3183 "str" # asmsuffix # "\t$Rt, [$Rn], $SImm9",
3185 Sched<[WriteSt, ReadSt, ReadSt]> {
3186 let Constraints = "$Rn = $Rn_wb";
3189 // Decoder only needed for unpredictability checking (FIXME).
3190 let DecoderMethod = "DecodeSingleIndexedInstruction";
3193 def _PostInd_LDR : A64I_LSpostind<size, v, {high_opc, 0b1},
3194 (outs GPR:$Rt, GPR64xsp:$Rn_wb),
3195 (ins GPR64xsp:$Rn, simm9:$SImm9),
3196 "ldr" # asmsuffix # "\t$Rt, [$Rn], $SImm9",
3198 Sched<[WriteLd, WriteLd, ReadLd]> {
3200 let Constraints = "$Rn = $Rn_wb";
3201 let DecoderMethod = "DecodeSingleIndexedInstruction";
3205 def _PreInd_STR : A64I_LSpreind<size, v, {high_opc, 0b0},
3206 (outs GPR64xsp:$Rn_wb),
3207 (ins GPR:$Rt, GPR64xsp:$Rn, simm9:$SImm9),
3208 "str" # asmsuffix # "\t$Rt, [$Rn, $SImm9]!",
3210 Sched<[WriteSt, ReadSt, ReadSt]> {
3211 let Constraints = "$Rn = $Rn_wb";
3214 // Decoder only needed for unpredictability checking (FIXME).
3215 let DecoderMethod = "DecodeSingleIndexedInstruction";
3218 def _PreInd_LDR : A64I_LSpreind<size, v, {high_opc, 0b1},
3219 (outs GPR:$Rt, GPR64xsp:$Rn_wb),
3220 (ins GPR64xsp:$Rn, simm9:$SImm9),
3221 "ldr" # asmsuffix # "\t$Rt, [$Rn, $SImm9]!",
3223 Sched<[WriteLd, WriteLd, ReadLd]> {
3225 let Constraints = "$Rn = $Rn_wb";
3226 let DecoderMethod = "DecodeSingleIndexedInstruction";
3231 // STRB/LDRB: First define the instructions
3233 : A64I_LDRSTR_unsigned<"LS8", 0b00, 0b0, 0b0, "b", GPR32, byte_addrparams>;
3237 : A64I_LDRSTR_unsigned<"LS16", 0b01, 0b0, 0b0, "h", GPR32, hword_addrparams>;
3240 // STR/LDR to/from a W register
3242 : A64I_LDRSTR_unsigned<"LS32", 0b10, 0b0, 0b0, "", GPR32, word_addrparams>;
3244 // STR/LDR to/from an X register
3246 : A64I_LDRSTR_unsigned<"LS64", 0b11, 0b0, 0b0, "", GPR64, dword_addrparams>;
3248 let Predicates = [HasFPARMv8] in {
3249 // STR/LDR to/from a B register
3251 : A64I_LDRSTR_unsigned<"LSFP8", 0b00, 0b1, 0b0, "", FPR8, byte_addrparams>;
3253 // STR/LDR to/from an H register
3255 : A64I_LDRSTR_unsigned<"LSFP16", 0b01, 0b1, 0b0, "", FPR16, hword_addrparams>;
3257 // STR/LDR to/from an S register
3259 : A64I_LDRSTR_unsigned<"LSFP32", 0b10, 0b1, 0b0, "", FPR32, word_addrparams>;
3260 // STR/LDR to/from a D register
3262 : A64I_LDRSTR_unsigned<"LSFP64", 0b11, 0b1, 0b0, "", FPR64, dword_addrparams>;
3263 // STR/LDR to/from a Q register
3265 : A64I_LDRSTR_unsigned<"LSFP128", 0b00, 0b1, 0b1, "", FPR128,
3269 //===------------------------------
3271 //===------------------------------
3273 // Byte and half-word signed loads can both go into either an X or a W register,
3274 // so it's worth factoring out. Signed word loads don't fit because there is no
3276 multiclass A64I_LDR_signed<bits<2> size, string asmopcode, AddrParams params,
3279 def w : A64I_LSunsigimm<size, 0b0, 0b11,
3281 (ins GPR64xsp:$Rn, params.uimm12:$UImm12),
3282 "ldrs" # asmopcode # "\t$Rt, [$Rn, $UImm12]",
3284 Sched<[WriteLd, ReadLd]> {
3287 def : InstAlias<"ldrs" # asmopcode # " $Rt, [$Rn]",
3288 (!cast<Instruction>(prefix # w) GPR32:$Rt, GPR64xsp:$Rn, 0)>;
3290 def x : A64I_LSunsigimm<size, 0b0, 0b10,
3292 (ins GPR64xsp:$Rn, params.uimm12:$UImm12),
3293 "ldrs" # asmopcode # "\t$Rt, [$Rn, $UImm12]",
3295 Sched<[WriteLd, ReadLd]> {
3298 def : InstAlias<"ldrs" # asmopcode # " $Rt, [$Rn]",
3299 (!cast<Instruction>(prefix # x) GPR64:$Rt, GPR64xsp:$Rn, 0)>;
3302 let mayLoad = 1 in {
3303 def w_Wm_RegOffset : A64I_LSregoff<size, 0b0, 0b11, 0b0,
3305 (ins GPR64xsp:$Rn, GPR32:$Rm, params.regextWm:$Ext),
3306 "ldrs" # asmopcode # "\t$Rt, [$Rn, $Rm, $Ext]",
3308 Sched<[WriteLd, ReadLd, ReadLd]>;
3310 def w_Xm_RegOffset : A64I_LSregoff<size, 0b0, 0b11, 0b1,
3312 (ins GPR64xsp:$Rn, GPR64:$Rm, params.regextXm:$Ext),
3313 "ldrs" # asmopcode # "\t$Rt, [$Rn, $Rm, $Ext]",
3315 Sched<[WriteLd, ReadLd, ReadLd]>;
3317 def x_Wm_RegOffset : A64I_LSregoff<size, 0b0, 0b10, 0b0,
3319 (ins GPR64xsp:$Rn, GPR32:$Rm, params.regextWm:$Ext),
3320 "ldrs" # asmopcode # "\t$Rt, [$Rn, $Rm, $Ext]",
3322 Sched<[WriteLd, ReadLd, ReadLd]>;
3324 def x_Xm_RegOffset : A64I_LSregoff<size, 0b0, 0b10, 0b1,
3326 (ins GPR64xsp:$Rn, GPR64:$Rm, params.regextXm:$Ext),
3327 "ldrs" # asmopcode # "\t$Rt, [$Rn, $Rm, $Ext]",
3329 Sched<[WriteLd, ReadLd, ReadLd]>;
3331 def : InstAlias<"ldrs" # asmopcode # " $Rt, [$Rn, $Rm]",
3332 (!cast<Instruction>(prefix # "w_Xm_RegOffset") GPR32:$Rt, GPR64xsp:$Rn,
3335 def : InstAlias<"ldrs" # asmopcode # " $Rt, [$Rn, $Rm]",
3336 (!cast<Instruction>(prefix # "x_Xm_RegOffset") GPR64:$Rt, GPR64xsp:$Rn,
3340 let mayLoad = 1 in {
3342 def w_U : A64I_LSunalimm<size, 0b0, 0b11,
3344 (ins GPR64xsp:$Rn, simm9:$SImm9),
3345 "ldurs" # asmopcode # "\t$Rt, [$Rn, $SImm9]",
3347 Sched<[WriteLd, ReadLd]>;
3349 def x_U : A64I_LSunalimm<size, 0b0, 0b10,
3351 (ins GPR64xsp:$Rn, simm9:$SImm9),
3352 "ldurs" # asmopcode # "\t$Rt, [$Rn, $SImm9]",
3354 Sched<[WriteLd, ReadLd]>;
3358 def w_PostInd : A64I_LSpostind<size, 0b0, 0b11,
3359 (outs GPR32:$Rt, GPR64xsp:$Rn_wb),
3360 (ins GPR64xsp:$Rn, simm9:$SImm9),
3361 "ldrs" # asmopcode # "\t$Rt, [$Rn], $SImm9",
3363 Sched<[WriteLd, WriteLd, ReadLd]> {
3364 let Constraints = "$Rn = $Rn_wb";
3365 let DecoderMethod = "DecodeSingleIndexedInstruction";
3368 def x_PostInd : A64I_LSpostind<size, 0b0, 0b10,
3369 (outs GPR64:$Rt, GPR64xsp:$Rn_wb),
3370 (ins GPR64xsp:$Rn, simm9:$SImm9),
3371 "ldrs" # asmopcode # "\t$Rt, [$Rn], $SImm9",
3373 Sched<[WriteLd, WriteLd, ReadLd]> {
3374 let Constraints = "$Rn = $Rn_wb";
3375 let DecoderMethod = "DecodeSingleIndexedInstruction";
3379 def w_PreInd : A64I_LSpreind<size, 0b0, 0b11,
3380 (outs GPR32:$Rt, GPR64xsp:$Rn_wb),
3381 (ins GPR64xsp:$Rn, simm9:$SImm9),
3382 "ldrs" # asmopcode # "\t$Rt, [$Rn, $SImm9]!",
3384 Sched<[WriteLd, WriteLd, ReadLd]> {
3385 let Constraints = "$Rn = $Rn_wb";
3386 let DecoderMethod = "DecodeSingleIndexedInstruction";
3389 def x_PreInd : A64I_LSpreind<size, 0b0, 0b10,
3390 (outs GPR64:$Rt, GPR64xsp:$Rn_wb),
3391 (ins GPR64xsp:$Rn, simm9:$SImm9),
3392 "ldrs" # asmopcode # "\t$Rt, [$Rn, $SImm9]!",
3394 Sched<[WriteLd, WriteLd, ReadLd]> {
3395 let Constraints = "$Rn = $Rn_wb";
3396 let DecoderMethod = "DecodeSingleIndexedInstruction";
3398 } // let mayLoad = 1
3402 defm LDRSB : A64I_LDR_signed<0b00, "b", byte_addrparams, "LDRSB">;
3404 defm LDRSH : A64I_LDR_signed<0b01, "h", hword_addrparams, "LDRSH">;
3406 // LDRSW: load a 32-bit register, sign-extending to 64-bits.
3408 : A64I_LSunsigimm<0b10, 0b0, 0b10,
3410 (ins GPR64xsp:$Rn, word_uimm12:$UImm12),
3411 "ldrsw\t$Rt, [$Rn, $UImm12]",
3413 Sched<[WriteLd, ReadLd]> {
3416 def : InstAlias<"ldrsw $Rt, [$Rn]", (LDRSWx GPR64:$Rt, GPR64xsp:$Rn, 0)>;
3418 let mayLoad = 1 in {
3419 def LDRSWx_Wm_RegOffset : A64I_LSregoff<0b10, 0b0, 0b10, 0b0,
3421 (ins GPR64xsp:$Rn, GPR32:$Rm, word_Wm_regext:$Ext),
3422 "ldrsw\t$Rt, [$Rn, $Rm, $Ext]",
3424 Sched<[WriteLd, ReadLd, ReadLd]>;
3426 def LDRSWx_Xm_RegOffset : A64I_LSregoff<0b10, 0b0, 0b10, 0b1,
3428 (ins GPR64xsp:$Rn, GPR64:$Rm, word_Xm_regext:$Ext),
3429 "ldrsw\t$Rt, [$Rn, $Rm, $Ext]",
3431 Sched<[WriteLd, ReadLd, ReadLd]>;
3433 def : InstAlias<"ldrsw $Rt, [$Rn, $Rm]",
3434 (LDRSWx_Xm_RegOffset GPR64:$Rt, GPR64xsp:$Rn, GPR64:$Rm, 2)>;
3438 : A64I_LSunalimm<0b10, 0b0, 0b10,
3440 (ins GPR64xsp:$Rn, simm9:$SImm9),
3441 "ldursw\t$Rt, [$Rn, $SImm9]",
3443 Sched<[WriteLd, ReadLd]> {
3446 def : InstAlias<"ldursw $Rt, [$Rn]", (LDURSWx GPR64:$Rt, GPR64xsp:$Rn, 0)>;
3449 : A64I_LSpostind<0b10, 0b0, 0b10,
3450 (outs GPR64:$Rt, GPR64xsp:$Rn_wb),
3451 (ins GPR64xsp:$Rn, simm9:$SImm9),
3452 "ldrsw\t$Rt, [$Rn], $SImm9",
3454 Sched<[WriteLd, WriteLd, ReadLd]> {
3456 let Constraints = "$Rn = $Rn_wb";
3457 let DecoderMethod = "DecodeSingleIndexedInstruction";
3460 def LDRSWx_PreInd : A64I_LSpreind<0b10, 0b0, 0b10,
3461 (outs GPR64:$Rt, GPR64xsp:$Rn_wb),
3462 (ins GPR64xsp:$Rn, simm9:$SImm9),
3463 "ldrsw\t$Rt, [$Rn, $SImm9]!",
3465 Sched<[WriteLd, WriteLd, ReadLd]> {
3467 let Constraints = "$Rn = $Rn_wb";
3468 let DecoderMethod = "DecodeSingleIndexedInstruction";
3471 //===------------------------------
3472 // 2.4 Prefetch operations
3473 //===------------------------------
3475 def PRFM : A64I_LSunsigimm<0b11, 0b0, 0b10, (outs),
3476 (ins prefetch_op:$Rt, GPR64xsp:$Rn, dword_uimm12:$UImm12),
3477 "prfm\t$Rt, [$Rn, $UImm12]",
3479 Sched<[WritePreLd, ReadPreLd]> {
3482 def : InstAlias<"prfm $Rt, [$Rn]",
3483 (PRFM prefetch_op:$Rt, GPR64xsp:$Rn, 0)>;
3485 let mayLoad = 1 in {
3486 def PRFM_Wm_RegOffset : A64I_LSregoff<0b11, 0b0, 0b10, 0b0, (outs),
3487 (ins prefetch_op:$Rt, GPR64xsp:$Rn,
3488 GPR32:$Rm, dword_Wm_regext:$Ext),
3489 "prfm\t$Rt, [$Rn, $Rm, $Ext]",
3491 Sched<[WritePreLd, ReadPreLd]>;
3492 def PRFM_Xm_RegOffset : A64I_LSregoff<0b11, 0b0, 0b10, 0b1, (outs),
3493 (ins prefetch_op:$Rt, GPR64xsp:$Rn,
3494 GPR64:$Rm, dword_Xm_regext:$Ext),
3495 "prfm\t$Rt, [$Rn, $Rm, $Ext]",
3497 Sched<[WritePreLd, ReadPreLd]>;
3500 def : InstAlias<"prfm $Rt, [$Rn, $Rm]",
3501 (PRFM_Xm_RegOffset prefetch_op:$Rt, GPR64xsp:$Rn,
3505 def PRFUM : A64I_LSunalimm<0b11, 0b0, 0b10, (outs),
3506 (ins prefetch_op:$Rt, GPR64xsp:$Rn, simm9:$SImm9),
3507 "prfum\t$Rt, [$Rn, $SImm9]",
3509 Sched<[WritePreLd, ReadPreLd]> {
3512 def : InstAlias<"prfum $Rt, [$Rn]",
3513 (PRFUM prefetch_op:$Rt, GPR64xsp:$Rn, 0)>;
3515 //===----------------------------------------------------------------------===//
3516 // Load-store register (unprivileged) instructions
3517 //===----------------------------------------------------------------------===//
3518 // Contains: LDTRB, LDTRH, LDTRSB, LDTRSH, LDTRSW, STTR, STTRB and STTRH
3520 // These instructions very much mirror the "unscaled immediate" loads, but since
3521 // there are no floating-point variants we need to split them out into their own
3522 // section to avoid instantiation of "ldtr d0, [sp]" etc.
3524 multiclass A64I_LDTRSTTR<bits<2> size, string asmsuffix, RegisterClass GPR,
3526 def _UnPriv_STR : A64I_LSunpriv<size, 0b0, 0b00,
3527 (outs), (ins GPR:$Rt, GPR64xsp:$Rn, simm9:$SImm9),
3528 "sttr" # asmsuffix # "\t$Rt, [$Rn, $SImm9]",
3530 Sched<[WriteLd, ReadLd]> {
3534 def : InstAlias<"sttr" # asmsuffix # " $Rt, [$Rn]",
3535 (!cast<Instruction>(prefix # "_UnPriv_STR") GPR:$Rt, GPR64xsp:$Rn, 0)>;
3537 def _UnPriv_LDR : A64I_LSunpriv<size, 0b0, 0b01,
3538 (outs GPR:$Rt), (ins GPR64xsp:$Rn, simm9:$SImm9),
3539 "ldtr" # asmsuffix # "\t$Rt, [$Rn, $SImm9]",
3541 Sched<[WriteLd, ReadLd]> {
3545 def : InstAlias<"ldtr" # asmsuffix # " $Rt, [$Rn]",
3546 (!cast<Instruction>(prefix # "_UnPriv_LDR") GPR:$Rt, GPR64xsp:$Rn, 0)>;
3550 // STTRB/LDTRB: First define the instructions
3551 defm LS8 : A64I_LDTRSTTR<0b00, "b", GPR32, "LS8">;
3554 defm LS16 : A64I_LDTRSTTR<0b01, "h", GPR32, "LS16">;
3556 // STTR/LDTR to/from a W register
3557 defm LS32 : A64I_LDTRSTTR<0b10, "", GPR32, "LS32">;
3559 // STTR/LDTR to/from an X register
3560 defm LS64 : A64I_LDTRSTTR<0b11, "", GPR64, "LS64">;
3562 // Now a class for the signed instructions that can go to either 32 or 64
3564 multiclass A64I_LDTR_signed<bits<2> size, string asmopcode, string prefix> {
3565 let mayLoad = 1 in {
3566 def w : A64I_LSunpriv<size, 0b0, 0b11,
3568 (ins GPR64xsp:$Rn, simm9:$SImm9),
3569 "ldtrs" # asmopcode # "\t$Rt, [$Rn, $SImm9]",
3571 Sched<[WriteLd, ReadLd]>;
3573 def x : A64I_LSunpriv<size, 0b0, 0b10,
3575 (ins GPR64xsp:$Rn, simm9:$SImm9),
3576 "ldtrs" # asmopcode # "\t$Rt, [$Rn, $SImm9]",
3578 Sched<[WriteLd, ReadLd]>;
3581 def : InstAlias<"ldtrs" # asmopcode # " $Rt, [$Rn]",
3582 (!cast<Instruction>(prefix # "w") GPR32:$Rt, GPR64xsp:$Rn, 0)>;
3584 def : InstAlias<"ldtrs" # asmopcode # " $Rt, [$Rn]",
3585 (!cast<Instruction>(prefix # "x") GPR64:$Rt, GPR64xsp:$Rn, 0)>;
3590 defm LDTRSB : A64I_LDTR_signed<0b00, "b", "LDTRSB">;
3592 defm LDTRSH : A64I_LDTR_signed<0b01, "h", "LDTRSH">;
3594 // And finally LDTRSW which only goes to 64 bits.
3595 def LDTRSWx : A64I_LSunpriv<0b10, 0b0, 0b10,
3597 (ins GPR64xsp:$Rn, simm9:$SImm9),
3598 "ldtrsw\t$Rt, [$Rn, $SImm9]",
3600 Sched<[WriteLd, ReadLd]> {
3603 def : InstAlias<"ldtrsw $Rt, [$Rn]", (LDTRSWx GPR64:$Rt, GPR64xsp:$Rn, 0)>;
3605 //===----------------------------------------------------------------------===//
3606 // Load-store register pair (offset) instructions
3607 //===----------------------------------------------------------------------===//
3611 //===----------------------------------------------------------------------===//
3612 // Load-store register pair (post-indexed) instructions
3613 //===----------------------------------------------------------------------===//
3614 // Contains: STP, LDP, LDPSW
3618 //===----------------------------------------------------------------------===//
3619 // Load-store register pair (pre-indexed) instructions
3620 //===----------------------------------------------------------------------===//
3621 // Contains: STP, LDP, LDPSW
3625 //===----------------------------------------------------------------------===//
3626 // Load-store non-temporal register pair (offset) instructions
3627 //===----------------------------------------------------------------------===//
3628 // Contains: STNP, LDNP
3631 // Anything that creates an MCInst (Decoding, selection and AsmParsing) has to
3632 // know the access size via some means. An isolated operand does not have this
3633 // information unless told from here, which means we need separate tablegen
3634 // Operands for each access size. This multiclass takes care of instantiating
3635 // the correct template functions in the rest of the backend.
3637 multiclass offsets_simm7<string MemSize, string prefix> {
3638 // The bare signed 7-bit immediate is used in post-indexed instructions, but
3639 // because of the scaling performed a generic "simm7" operand isn't
3640 // appropriate here either.
3641 def simm7_asmoperand : AsmOperandClass {
3642 let Name = "SImm7_Scaled" # MemSize;
3643 let PredicateMethod = "isSImm7Scaled<" # MemSize # ">";
3644 let RenderMethod = "addSImm7ScaledOperands<" # MemSize # ">";
3645 let DiagnosticType = "LoadStoreSImm7_" # MemSize;
3648 def simm7 : Operand<i64> {
3649 let PrintMethod = "printSImm7ScaledOperand<" # MemSize # ">";
3650 let ParserMatchClass = !cast<AsmOperandClass>(prefix # "simm7_asmoperand");
3654 defm word_ : offsets_simm7<"4", "word_">;
3655 defm dword_ : offsets_simm7<"8", "dword_">;
3656 defm qword_ : offsets_simm7<"16", "qword_">;
3658 multiclass A64I_LSPsimple<bits<2> opc, bit v, RegisterClass SomeReg,
3659 Operand simm7, string prefix> {
3660 def _STR : A64I_LSPoffset<opc, v, 0b0, (outs),
3661 (ins SomeReg:$Rt, SomeReg:$Rt2, GPR64xsp:$Rn, simm7:$SImm7),
3662 "stp\t$Rt, $Rt2, [$Rn, $SImm7]", [], NoItinerary>,
3663 Sched<[WriteLd, ReadLd]> {
3665 let DecoderMethod = "DecodeLDSTPairInstruction";
3667 def : InstAlias<"stp $Rt, $Rt2, [$Rn]",
3668 (!cast<Instruction>(prefix # "_STR") SomeReg:$Rt,
3669 SomeReg:$Rt2, GPR64xsp:$Rn, 0)>;
3671 def _LDR : A64I_LSPoffset<opc, v, 0b1,
3672 (outs SomeReg:$Rt, SomeReg:$Rt2),
3673 (ins GPR64xsp:$Rn, simm7:$SImm7),
3674 "ldp\t$Rt, $Rt2, [$Rn, $SImm7]", [], NoItinerary>,
3675 Sched<[WriteLd, WriteLd, ReadLd]> {
3677 let DecoderMethod = "DecodeLDSTPairInstruction";
3679 def : InstAlias<"ldp $Rt, $Rt2, [$Rn]",
3680 (!cast<Instruction>(prefix # "_LDR") SomeReg:$Rt,
3681 SomeReg:$Rt2, GPR64xsp:$Rn, 0)>;
3683 def _PostInd_STR : A64I_LSPpostind<opc, v, 0b0,
3684 (outs GPR64xsp:$Rn_wb),
3685 (ins SomeReg:$Rt, SomeReg:$Rt2,
3688 "stp\t$Rt, $Rt2, [$Rn], $SImm7",
3690 Sched<[WriteSt, ReadSt, ReadSt, ReadSt]> {
3692 let Constraints = "$Rn = $Rn_wb";
3694 // Decoder only needed for unpredictability checking (FIXME).
3695 let DecoderMethod = "DecodeLDSTPairInstruction";
3698 def _PostInd_LDR : A64I_LSPpostind<opc, v, 0b1,
3699 (outs SomeReg:$Rt, SomeReg:$Rt2, GPR64xsp:$Rn_wb),
3700 (ins GPR64xsp:$Rn, simm7:$SImm7),
3701 "ldp\t$Rt, $Rt2, [$Rn], $SImm7",
3703 Sched<[WriteLd, WriteLd, WriteLd, ReadLd]> {
3705 let Constraints = "$Rn = $Rn_wb";
3706 let DecoderMethod = "DecodeLDSTPairInstruction";
3709 def _PreInd_STR : A64I_LSPpreind<opc, v, 0b0, (outs GPR64xsp:$Rn_wb),
3710 (ins SomeReg:$Rt, SomeReg:$Rt2, GPR64xsp:$Rn, simm7:$SImm7),
3711 "stp\t$Rt, $Rt2, [$Rn, $SImm7]!",
3713 Sched<[WriteSt, ReadSt, ReadSt, ReadSt]> {
3715 let Constraints = "$Rn = $Rn_wb";
3716 let DecoderMethod = "DecodeLDSTPairInstruction";
3719 def _PreInd_LDR : A64I_LSPpreind<opc, v, 0b1,
3720 (outs SomeReg:$Rt, SomeReg:$Rt2, GPR64xsp:$Rn_wb),
3721 (ins GPR64xsp:$Rn, simm7:$SImm7),
3722 "ldp\t$Rt, $Rt2, [$Rn, $SImm7]!",
3724 Sched<[WriteLd, WriteLd, WriteLd, ReadLd]> {
3726 let Constraints = "$Rn = $Rn_wb";
3727 let DecoderMethod = "DecodeLDSTPairInstruction";
3730 def _NonTemp_STR : A64I_LSPnontemp<opc, v, 0b0, (outs),
3731 (ins SomeReg:$Rt, SomeReg:$Rt2, GPR64xsp:$Rn, simm7:$SImm7),
3732 "stnp\t$Rt, $Rt2, [$Rn, $SImm7]", [], NoItinerary>,
3733 Sched<[WriteSt, ReadSt, ReadSt, ReadSt]> {
3735 let DecoderMethod = "DecodeLDSTPairInstruction";
3737 def : InstAlias<"stnp $Rt, $Rt2, [$Rn]",
3738 (!cast<Instruction>(prefix # "_NonTemp_STR") SomeReg:$Rt,
3739 SomeReg:$Rt2, GPR64xsp:$Rn, 0)>;
3741 def _NonTemp_LDR : A64I_LSPnontemp<opc, v, 0b1,
3742 (outs SomeReg:$Rt, SomeReg:$Rt2),
3743 (ins GPR64xsp:$Rn, simm7:$SImm7),
3744 "ldnp\t$Rt, $Rt2, [$Rn, $SImm7]", [], NoItinerary>,
3745 Sched<[WriteLd, WriteLd, ReadLd]> {
3747 let DecoderMethod = "DecodeLDSTPairInstruction";
3749 def : InstAlias<"ldnp $Rt, $Rt2, [$Rn]",
3750 (!cast<Instruction>(prefix # "_NonTemp_LDR") SomeReg:$Rt,
3751 SomeReg:$Rt2, GPR64xsp:$Rn, 0)>;
3756 defm LSPair32 : A64I_LSPsimple<0b00, 0b0, GPR32, word_simm7, "LSPair32">;
3757 defm LSPair64 : A64I_LSPsimple<0b10, 0b0, GPR64, dword_simm7, "LSPair64">;
3759 let Predicates = [HasFPARMv8] in {
3760 defm LSFPPair32 : A64I_LSPsimple<0b00, 0b1, FPR32, word_simm7, "LSFPPair32">;
3761 defm LSFPPair64 : A64I_LSPsimple<0b01, 0b1, FPR64, dword_simm7, "LSFPPair64">;
3762 defm LSFPPair128 : A64I_LSPsimple<0b10, 0b1, FPR128, qword_simm7,
3767 def LDPSWx : A64I_LSPoffset<0b01, 0b0, 0b1,
3768 (outs GPR64:$Rt, GPR64:$Rt2),
3769 (ins GPR64xsp:$Rn, word_simm7:$SImm7),
3770 "ldpsw\t$Rt, $Rt2, [$Rn, $SImm7]", [], NoItinerary>,
3771 Sched<[WriteLd, WriteLd, ReadLd]> {
3773 let DecoderMethod = "DecodeLDSTPairInstruction";
3775 def : InstAlias<"ldpsw $Rt, $Rt2, [$Rn]",
3776 (LDPSWx GPR64:$Rt, GPR64:$Rt2, GPR64xsp:$Rn, 0)>;
3778 def LDPSWx_PostInd : A64I_LSPpostind<0b01, 0b0, 0b1,
3779 (outs GPR64:$Rt, GPR64:$Rt2, GPR64:$Rn_wb),
3780 (ins GPR64xsp:$Rn, word_simm7:$SImm7),
3781 "ldpsw\t$Rt, $Rt2, [$Rn], $SImm7",
3783 Sched<[WriteLd, WriteLd, WriteLd, ReadLd]> {
3785 let Constraints = "$Rn = $Rn_wb";
3786 let DecoderMethod = "DecodeLDSTPairInstruction";
3789 def LDPSWx_PreInd : A64I_LSPpreind<0b01, 0b0, 0b1,
3790 (outs GPR64:$Rt, GPR64:$Rt2, GPR64:$Rn_wb),
3791 (ins GPR64xsp:$Rn, word_simm7:$SImm7),
3792 "ldpsw\t$Rt, $Rt2, [$Rn, $SImm7]!",
3794 Sched<[WriteLd, WriteLd, WriteLd, ReadLd]> {
3796 let Constraints = "$Rn = $Rn_wb";
3797 let DecoderMethod = "DecodeLDSTPairInstruction";
3800 //===----------------------------------------------------------------------===//
3801 // Logical (immediate) instructions
3802 //===----------------------------------------------------------------------===//
3803 // Contains: AND, ORR, EOR, ANDS, + aliases TST, MOV
3805 multiclass logical_imm_operands<string prefix, string note,
3806 int size, ValueType VT> {
3807 def _asmoperand : AsmOperandClass {
3808 let Name = "LogicalImm" # note # size;
3809 let PredicateMethod = "isLogicalImm" # note # "<" # size # ">";
3810 let RenderMethod = "addLogicalImmOperands<" # size # ">";
3811 let DiagnosticType = "LogicalSecondSource";
3815 : Operand<VT>, ComplexPattern<VT, 1, "SelectLogicalImm", [imm]> {
3816 let ParserMatchClass = !cast<AsmOperandClass>(prefix # "_asmoperand");
3817 let PrintMethod = "printLogicalImmOperand<" # size # ">";
3818 let DecoderMethod = "DecodeLogicalImmOperand<" # size # ">";
3822 defm logical_imm32 : logical_imm_operands<"logical_imm32", "", 32, i32>;
3823 defm logical_imm64 : logical_imm_operands<"logical_imm64", "", 64, i64>;
3825 // The mov versions only differ in assembly parsing, where they
3826 // exclude values representable with either MOVZ or MOVN.
3827 defm logical_imm32_mov
3828 : logical_imm_operands<"logical_imm32_mov", "MOV", 32, i32>;
3829 defm logical_imm64_mov
3830 : logical_imm_operands<"logical_imm64_mov", "MOV", 64, i64>;
3833 multiclass A64I_logimmSizes<bits<2> opc, string asmop, SDNode opnode> {
3834 def wwi : A64I_logicalimm<0b0, opc, (outs GPR32wsp:$Rd),
3835 (ins GPR32:$Rn, logical_imm32_operand:$Imm),
3836 !strconcat(asmop, "\t$Rd, $Rn, $Imm"),
3838 (opnode i32:$Rn, logical_imm32_operand:$Imm))],
3840 Sched<[WriteALU, ReadALU]>;
3842 def xxi : A64I_logicalimm<0b1, opc, (outs GPR64xsp:$Rd),
3843 (ins GPR64:$Rn, logical_imm64_operand:$Imm),
3844 !strconcat(asmop, "\t$Rd, $Rn, $Imm"),
3846 (opnode i64:$Rn, logical_imm64_operand:$Imm))],
3848 Sched<[WriteALU, ReadALU]>;
3851 defm AND : A64I_logimmSizes<0b00, "and", and>;
3852 defm ORR : A64I_logimmSizes<0b01, "orr", or>;
3853 defm EOR : A64I_logimmSizes<0b10, "eor", xor>;
3855 let Defs = [NZCV] in {
3856 def ANDSwwi : A64I_logicalimm<0b0, 0b11, (outs GPR32:$Rd),
3857 (ins GPR32:$Rn, logical_imm32_operand:$Imm),
3858 "ands\t$Rd, $Rn, $Imm",
3860 Sched<[WriteALU, ReadALU]>;
3862 def ANDSxxi : A64I_logicalimm<0b1, 0b11, (outs GPR64:$Rd),
3863 (ins GPR64:$Rn, logical_imm64_operand:$Imm),
3864 "ands\t$Rd, $Rn, $Imm",
3866 Sched<[WriteALU, ReadALU]>;
3869 def : InstAlias<"tst $Rn, $Imm",
3870 (ANDSwwi WZR, GPR32:$Rn, logical_imm32_operand:$Imm)>;
3871 def : InstAlias<"tst $Rn, $Imm",
3872 (ANDSxxi XZR, GPR64:$Rn, logical_imm64_operand:$Imm)>;
3873 // FIXME: these sometimes are canonical.
3874 def : InstAlias<"mov $Rd, $Imm",
3875 (ORRwwi GPR32wsp:$Rd, WZR, logical_imm32_mov_operand:$Imm), 0>;
3876 def : InstAlias<"mov $Rd, $Imm",
3877 (ORRxxi GPR64xsp:$Rd, XZR, logical_imm64_mov_operand:$Imm), 0>;
3879 //===----------------------------------------------------------------------===//
3880 // Logical (shifted register) instructions
3881 //===----------------------------------------------------------------------===//
3882 // Contains: AND, BIC, ORR, ORN, EOR, EON, ANDS, BICS + aliases TST, MVN, MOV
3884 // Operand for optimizing (icmp (and LHS, RHS), 0, SomeCode). In theory "ANDS"
3885 // behaves differently for unsigned comparisons, so we defensively only allow
3886 // signed or n/a as the operand. In practice "unsigned greater than 0" is "not
3887 // equal to 0" and LLVM gives us this.
3888 def signed_cond : PatLeaf<(cond), [{
3889 return !isUnsignedIntSetCC(N->get());
3893 // These instructions share their "shift" operands with add/sub (shifted
3894 // register instructions). They are defined there.
3896 // N.b. the commutable parameter is just !N. It will be first against the wall
3897 // when the revolution comes.
3898 multiclass logical_shifts<string prefix, bit sf, bits<2> opc,
3899 bit N, bit commutable,
3900 string asmop, SDPatternOperator opfrag, ValueType ty,
3901 RegisterClass GPR, list<Register> defs> {
3902 let isCommutable = commutable, Defs = defs in {
3903 def _lsl : A64I_logicalshift<sf, opc, 0b00, N,
3905 (ins GPR:$Rn, GPR:$Rm,
3906 !cast<Operand>("lsl_operand_" # ty):$Imm6),
3907 !strconcat(asmop, "\t$Rd, $Rn, $Rm, $Imm6"),
3908 [(set ty:$Rd, (opfrag ty:$Rn, (shl ty:$Rm,
3909 !cast<Operand>("lsl_operand_" # ty):$Imm6))
3912 Sched<[WriteALU, ReadALU, ReadALU]>;
3914 def _lsr : A64I_logicalshift<sf, opc, 0b01, N,
3916 (ins GPR:$Rn, GPR:$Rm,
3917 !cast<Operand>("lsr_operand_" # ty):$Imm6),
3918 !strconcat(asmop, "\t$Rd, $Rn, $Rm, $Imm6"),
3919 [(set ty:$Rd, (opfrag ty:$Rn, (srl ty:$Rm,
3920 !cast<Operand>("lsr_operand_" # ty):$Imm6))
3923 Sched<[WriteALU, ReadALU, ReadALU]>;
3925 def _asr : A64I_logicalshift<sf, opc, 0b10, N,
3927 (ins GPR:$Rn, GPR:$Rm,
3928 !cast<Operand>("asr_operand_" # ty):$Imm6),
3929 !strconcat(asmop, "\t$Rd, $Rn, $Rm, $Imm6"),
3930 [(set ty:$Rd, (opfrag ty:$Rn, (sra ty:$Rm,
3931 !cast<Operand>("asr_operand_" # ty):$Imm6))
3934 Sched<[WriteALU, ReadALU, ReadALU]>;
3936 def _ror : A64I_logicalshift<sf, opc, 0b11, N,
3938 (ins GPR:$Rn, GPR:$Rm,
3939 !cast<Operand>("ror_operand_" # ty):$Imm6),
3940 !strconcat(asmop, "\t$Rd, $Rn, $Rm, $Imm6"),
3941 [(set ty:$Rd, (opfrag ty:$Rn, (rotr ty:$Rm,
3942 !cast<Operand>("ror_operand_" # ty):$Imm6))
3945 Sched<[WriteALU, ReadALU, ReadALU]>;
3949 : InstAlias<!strconcat(asmop, " $Rd, $Rn, $Rm"),
3950 (!cast<Instruction>(prefix # "_lsl") GPR:$Rd, GPR:$Rn,
3953 def : Pat<(opfrag ty:$Rn, ty:$Rm),
3954 (!cast<Instruction>(prefix # "_lsl") $Rn, $Rm, 0)>;
3957 multiclass logical_sizes<string prefix, bits<2> opc, bit N, bit commutable,
3958 string asmop, SDPatternOperator opfrag,
3959 list<Register> defs> {
3960 defm xxx : logical_shifts<prefix # "xxx", 0b1, opc, N,
3961 commutable, asmop, opfrag, i64, GPR64, defs>;
3962 defm www : logical_shifts<prefix # "www", 0b0, opc, N,
3963 commutable, asmop, opfrag, i32, GPR32, defs>;
3967 defm AND : logical_sizes<"AND", 0b00, 0b0, 0b1, "and", and, []>;
3968 defm ORR : logical_sizes<"ORR", 0b01, 0b0, 0b1, "orr", or, []>;
3969 defm EOR : logical_sizes<"EOR", 0b10, 0b0, 0b1, "eor", xor, []>;
3970 defm ANDS : logical_sizes<"ANDS", 0b11, 0b0, 0b1, "ands",
3971 PatFrag<(ops node:$lhs, node:$rhs), (and node:$lhs, node:$rhs),
3972 [{ (void)N; return false; }]>,
3975 defm BIC : logical_sizes<"BIC", 0b00, 0b1, 0b0, "bic",
3976 PatFrag<(ops node:$lhs, node:$rhs),
3977 (and node:$lhs, (not node:$rhs))>, []>;
3978 defm ORN : logical_sizes<"ORN", 0b01, 0b1, 0b0, "orn",
3979 PatFrag<(ops node:$lhs, node:$rhs),
3980 (or node:$lhs, (not node:$rhs))>, []>;
3981 defm EON : logical_sizes<"EON", 0b10, 0b1, 0b0, "eon",
3982 PatFrag<(ops node:$lhs, node:$rhs),
3983 (xor node:$lhs, (not node:$rhs))>, []>;
3984 defm BICS : logical_sizes<"BICS", 0b11, 0b1, 0b0, "bics",
3985 PatFrag<(ops node:$lhs, node:$rhs),
3986 (and node:$lhs, (not node:$rhs)),
3987 [{ (void)N; return false; }]>,
3990 multiclass tst_shifts<string prefix, bit sf, ValueType ty, RegisterClass GPR> {
3991 let isCommutable = 1, Rd = 0b11111, Defs = [NZCV] in {
3992 def _lsl : A64I_logicalshift<sf, 0b11, 0b00, 0b0,
3994 (ins GPR:$Rn, GPR:$Rm,
3995 !cast<Operand>("lsl_operand_" # ty):$Imm6),
3996 "tst\t$Rn, $Rm, $Imm6",
3997 [(set NZCV, (A64setcc (and ty:$Rn, (shl ty:$Rm,
3998 !cast<Operand>("lsl_operand_" # ty):$Imm6)),
4001 Sched<[WriteALU, ReadALU, ReadALU]>;
4004 def _lsr : A64I_logicalshift<sf, 0b11, 0b01, 0b0,
4006 (ins GPR:$Rn, GPR:$Rm,
4007 !cast<Operand>("lsr_operand_" # ty):$Imm6),
4008 "tst\t$Rn, $Rm, $Imm6",
4009 [(set NZCV, (A64setcc (and ty:$Rn, (srl ty:$Rm,
4010 !cast<Operand>("lsr_operand_" # ty):$Imm6)),
4013 Sched<[WriteALU, ReadALU, ReadALU]>;
4015 def _asr : A64I_logicalshift<sf, 0b11, 0b10, 0b0,
4017 (ins GPR:$Rn, GPR:$Rm,
4018 !cast<Operand>("asr_operand_" # ty):$Imm6),
4019 "tst\t$Rn, $Rm, $Imm6",
4020 [(set NZCV, (A64setcc (and ty:$Rn, (sra ty:$Rm,
4021 !cast<Operand>("asr_operand_" # ty):$Imm6)),
4024 Sched<[WriteALU, ReadALU, ReadALU]>;
4026 def _ror : A64I_logicalshift<sf, 0b11, 0b11, 0b0,
4028 (ins GPR:$Rn, GPR:$Rm,
4029 !cast<Operand>("ror_operand_" # ty):$Imm6),
4030 "tst\t$Rn, $Rm, $Imm6",
4031 [(set NZCV, (A64setcc (and ty:$Rn, (rotr ty:$Rm,
4032 !cast<Operand>("ror_operand_" # ty):$Imm6)),
4035 Sched<[WriteALU, ReadALU, ReadALU]>;
4038 def _noshift : InstAlias<"tst $Rn, $Rm",
4039 (!cast<Instruction>(prefix # "_lsl") GPR:$Rn, GPR:$Rm, 0)>;
4041 def : Pat<(A64setcc (and ty:$Rn, ty:$Rm), 0, signed_cond),
4042 (!cast<Instruction>(prefix # "_lsl") $Rn, $Rm, 0)>;
4045 defm TSTxx : tst_shifts<"TSTxx", 0b1, i64, GPR64>;
4046 defm TSTww : tst_shifts<"TSTww", 0b0, i32, GPR32>;
4049 multiclass mvn_shifts<string prefix, bit sf, ValueType ty, RegisterClass GPR> {
4050 let isCommutable = 0, Rn = 0b11111 in {
4051 def _lsl : A64I_logicalshift<sf, 0b01, 0b00, 0b1,
4054 !cast<Operand>("lsl_operand_" # ty):$Imm6),
4055 "mvn\t$Rd, $Rm, $Imm6",
4056 [(set ty:$Rd, (not (shl ty:$Rm,
4057 !cast<Operand>("lsl_operand_" # ty):$Imm6)))],
4059 Sched<[WriteALU, ReadALU, ReadALU]>;
4062 def _lsr : A64I_logicalshift<sf, 0b01, 0b01, 0b1,
4065 !cast<Operand>("lsr_operand_" # ty):$Imm6),
4066 "mvn\t$Rd, $Rm, $Imm6",
4067 [(set ty:$Rd, (not (srl ty:$Rm,
4068 !cast<Operand>("lsr_operand_" # ty):$Imm6)))],
4070 Sched<[WriteALU, ReadALU, ReadALU]>;
4072 def _asr : A64I_logicalshift<sf, 0b01, 0b10, 0b1,
4075 !cast<Operand>("asr_operand_" # ty):$Imm6),
4076 "mvn\t$Rd, $Rm, $Imm6",
4077 [(set ty:$Rd, (not (sra ty:$Rm,
4078 !cast<Operand>("asr_operand_" # ty):$Imm6)))],
4080 Sched<[WriteALU, ReadALU, ReadALU]>;
4082 def _ror : A64I_logicalshift<sf, 0b01, 0b11, 0b1,
4085 !cast<Operand>("ror_operand_" # ty):$Imm6),
4086 "mvn\t$Rd, $Rm, $Imm6",
4087 [(set ty:$Rd, (not (rotr ty:$Rm,
4088 !cast<Operand>("lsl_operand_" # ty):$Imm6)))],
4090 Sched<[WriteALU, ReadALU, ReadALU]>;
4093 def _noshift : InstAlias<"mvn $Rn, $Rm",
4094 (!cast<Instruction>(prefix # "_lsl") GPR:$Rn, GPR:$Rm, 0)>;
4096 def : Pat<(not ty:$Rm),
4097 (!cast<Instruction>(prefix # "_lsl") $Rm, 0)>;
4100 defm MVNxx : mvn_shifts<"MVNxx", 0b1, i64, GPR64>;
4101 defm MVNww : mvn_shifts<"MVNww", 0b0, i32, GPR32>;
4103 def MOVxx :InstAlias<"mov $Rd, $Rm", (ORRxxx_lsl GPR64:$Rd, XZR, GPR64:$Rm, 0)>;
4104 def MOVww :InstAlias<"mov $Rd, $Rm", (ORRwww_lsl GPR32:$Rd, WZR, GPR32:$Rm, 0)>;
4106 //===----------------------------------------------------------------------===//
4107 // Move wide (immediate) instructions
4108 //===----------------------------------------------------------------------===//
4109 // Contains: MOVN, MOVZ, MOVK + MOV aliases
4111 // A wide variety of different relocations are needed for variants of these
4112 // instructions, so it turns out that we need a different operand for all of
4114 multiclass movw_operands<string prefix, string instname, int width> {
4115 def _imm_asmoperand : AsmOperandClass {
4116 let Name = instname # width # "Shifted" # shift;
4117 let PredicateMethod = "is" # instname # width # "Imm";
4118 let RenderMethod = "addMoveWideImmOperands";
4119 let ParserMethod = "ParseImmWithLSLOperand";
4120 let DiagnosticType = "MOVWUImm16";
4123 def _imm : Operand<i64> {
4124 let ParserMatchClass = !cast<AsmOperandClass>(prefix # "_imm_asmoperand");
4125 let PrintMethod = "printMoveWideImmOperand";
4126 let EncoderMethod = "getMoveWideImmOpValue";
4127 let DecoderMethod = "DecodeMoveWideImmOperand<" # width # ">";
4129 let MIOperandInfo = (ops uimm16:$UImm16, imm:$Shift);
4133 defm movn32 : movw_operands<"movn32", "MOVN", 32>;
4134 defm movn64 : movw_operands<"movn64", "MOVN", 64>;
4135 defm movz32 : movw_operands<"movz32", "MOVZ", 32>;
4136 defm movz64 : movw_operands<"movz64", "MOVZ", 64>;
4137 defm movk32 : movw_operands<"movk32", "MOVK", 32>;
4138 defm movk64 : movw_operands<"movk64", "MOVK", 64>;
4140 multiclass A64I_movwSizes<bits<2> opc, string asmop, dag ins32bit,
4143 def wii : A64I_movw<0b0, opc, (outs GPR32:$Rd), ins32bit,
4144 !strconcat(asmop, "\t$Rd, $FullImm"),
4148 let UImm16 = FullImm{15-0};
4149 let Shift = FullImm{17-16};
4152 def xii : A64I_movw<0b1, opc, (outs GPR64:$Rd), ins64bit,
4153 !strconcat(asmop, "\t$Rd, $FullImm"),
4157 let UImm16 = FullImm{15-0};
4158 let Shift = FullImm{17-16};
4162 let isMoveImm = 1, isReMaterializable = 1,
4163 isAsCheapAsAMove = 1, hasSideEffects = 0 in {
4164 defm MOVN : A64I_movwSizes<0b00, "movn",
4165 (ins movn32_imm:$FullImm),
4166 (ins movn64_imm:$FullImm)>;
4168 // Some relocations are able to convert between a MOVZ and a MOVN. If these
4169 // are applied the instruction must be emitted with the corresponding bits as
4170 // 0, which means a MOVZ needs to override that bit from the default.
4171 let PostEncoderMethod = "fixMOVZ" in
4172 defm MOVZ : A64I_movwSizes<0b10, "movz",
4173 (ins movz32_imm:$FullImm),
4174 (ins movz64_imm:$FullImm)>;
4177 let Constraints = "$src = $Rd",
4178 SchedRW = [WriteALU, ReadALU] in
4179 defm MOVK : A64I_movwSizes<0b11, "movk",
4180 (ins GPR32:$src, movk32_imm:$FullImm),
4181 (ins GPR64:$src, movk64_imm:$FullImm)>;
4184 // And now the "MOV" aliases. These also need their own operands because what
4185 // they accept is completely different to what the base instructions accept.
4186 multiclass movalias_operand<string prefix, string basename,
4187 string immpredicate, int width> {
4188 def _asmoperand : AsmOperandClass {
4189 let Name = basename # width # "MovAlias";
4191 = "isMoveWideMovAlias<" # width # ", A64Imms::" # immpredicate # ">";
4193 = "addMoveWideMovAliasOperands<" # width # ", "
4194 # "A64Imms::" # immpredicate # ">";
4197 def _movimm : Operand<i64> {
4198 let ParserMatchClass = !cast<AsmOperandClass>(prefix # "_asmoperand");
4200 let MIOperandInfo = (ops uimm16:$UImm16, imm:$Shift);
4204 defm movz32 : movalias_operand<"movz32", "MOVZ", "isMOVZImm", 32>;
4205 defm movz64 : movalias_operand<"movz64", "MOVZ", "isMOVZImm", 64>;
4206 defm movn32 : movalias_operand<"movn32", "MOVN", "isOnlyMOVNImm", 32>;
4207 defm movn64 : movalias_operand<"movn64", "MOVN", "isOnlyMOVNImm", 64>;
4209 // FIXME: these are officially canonical aliases, but TableGen is too limited to
4210 // print them at the moment. I believe in this case an "AliasPredicate" method
4211 // will need to be implemented. to allow it, as well as the more generally
4212 // useful handling of non-register, non-constant operands.
4213 class movalias<Instruction INST, RegisterClass GPR, Operand operand>
4214 : InstAlias<"mov $Rd, $FullImm", (INST GPR:$Rd, operand:$FullImm), 0>;
4216 def : movalias<MOVZwii, GPR32, movz32_movimm>;
4217 def : movalias<MOVZxii, GPR64, movz64_movimm>;
4218 def : movalias<MOVNwii, GPR32, movn32_movimm>;
4219 def : movalias<MOVNxii, GPR64, movn64_movimm>;
4221 def movw_addressref_g0 : ComplexPattern<i64, 2, "SelectMOVWAddressRef<0>">;
4222 def movw_addressref_g1 : ComplexPattern<i64, 2, "SelectMOVWAddressRef<1>">;
4223 def movw_addressref_g2 : ComplexPattern<i64, 2, "SelectMOVWAddressRef<2>">;
4224 def movw_addressref_g3 : ComplexPattern<i64, 2, "SelectMOVWAddressRef<3>">;
4226 def : Pat<(A64WrapperLarge movw_addressref_g3:$G3, movw_addressref_g2:$G2,
4227 movw_addressref_g1:$G1, movw_addressref_g0:$G0),
4228 (MOVKxii (MOVKxii (MOVKxii (MOVZxii movw_addressref_g3:$G3),
4229 movw_addressref_g2:$G2),
4230 movw_addressref_g1:$G1),
4231 movw_addressref_g0:$G0)>;
4233 //===----------------------------------------------------------------------===//
4234 // PC-relative addressing instructions
4235 //===----------------------------------------------------------------------===//
4236 // Contains: ADR, ADRP
4238 def adr_label : Operand<i64> {
4239 let EncoderMethod = "getLabelOpValue<AArch64::fixup_a64_adr_prel>";
4241 // This label is a 21-bit offset from PC, unscaled
4242 let PrintMethod = "printLabelOperand<21, 1>";
4243 let ParserMatchClass = label_asmoperand<21, 1>;
4244 let OperandType = "OPERAND_PCREL";
4247 def adrp_label_asmoperand : AsmOperandClass {
4248 let Name = "AdrpLabel";
4249 let RenderMethod = "addLabelOperands<21, 4096>";
4250 let DiagnosticType = "Label";
4253 def adrp_label : Operand<i64> {
4254 let EncoderMethod = "getAdrpLabelOpValue";
4256 // This label is a 21-bit offset from PC, scaled by the page-size: 4096.
4257 let PrintMethod = "printLabelOperand<21, 4096>";
4258 let ParserMatchClass = adrp_label_asmoperand;
4259 let OperandType = "OPERAND_PCREL";
4262 let hasSideEffects = 0 in {
4263 def ADRxi : A64I_PCADR<0b0, (outs GPR64:$Rd), (ins adr_label:$Label),
4264 "adr\t$Rd, $Label", [], NoItinerary>,
4267 def ADRPxi : A64I_PCADR<0b1, (outs GPR64:$Rd), (ins adrp_label:$Label),
4268 "adrp\t$Rd, $Label", [], NoItinerary>,
4272 //===----------------------------------------------------------------------===//
4273 // System instructions
4274 //===----------------------------------------------------------------------===//
4275 // Contains: HINT, CLREX, DSB, DMB, ISB, MSR, SYS, SYSL, MRS
4276 // + aliases IC, DC, AT, TLBI, NOP, YIELD, WFE, WFI, SEV, SEVL
4278 // Op1 and Op2 fields are sometimes simple 3-bit unsigned immediate values.
4279 def uimm3_asmoperand : AsmOperandClass {
4281 let PredicateMethod = "isUImm<3>";
4282 let RenderMethod = "addImmOperands";
4283 let DiagnosticType = "UImm3";
4286 def uimm3 : Operand<i32> {
4287 let ParserMatchClass = uimm3_asmoperand;
4290 // The HINT alias can accept a simple unsigned 7-bit immediate.
4291 def uimm7_asmoperand : AsmOperandClass {
4293 let PredicateMethod = "isUImm<7>";
4294 let RenderMethod = "addImmOperands";
4295 let DiagnosticType = "UImm7";
4298 def uimm7 : Operand<i32> {
4299 let ParserMatchClass = uimm7_asmoperand;
4302 // Multiclass namedimm is defined with the prefetch operands. Most of these fit
4303 // into the NamedImmMapper scheme well: they either accept a named operand or
4304 // any immediate under a particular value (which may be 0, implying no immediate
4306 defm dbarrier : namedimm<"dbarrier", "A64DB::DBarrierMapper">;
4307 defm isb : namedimm<"isb", "A64ISB::ISBMapper">;
4308 defm ic : namedimm<"ic", "A64IC::ICMapper">;
4309 defm dc : namedimm<"dc", "A64DC::DCMapper">;
4310 defm at : namedimm<"at", "A64AT::ATMapper">;
4311 defm tlbi : namedimm<"tlbi", "A64TLBI::TLBIMapper">;
4313 // However, MRS and MSR are more complicated for a few reasons:
4314 // * There are ~1000 generic names S3_<op1>_<CRn>_<CRm>_<Op2> which have an
4315 // implementation-defined effect
4316 // * Most registers are shared, but some are read-only or write-only.
4317 // * There is a variant of MSR which accepts the same register name (SPSel),
4318 // but which would have a different encoding.
4320 // In principle these could be resolved in with more complicated subclasses of
4321 // NamedImmMapper, however that imposes an overhead on other "named
4322 // immediates". Both in concrete terms with virtual tables and in unnecessary
4325 // The solution adopted here is to take the MRS/MSR Mappers out of the usual
4326 // hierarchy (they're not derived from NamedImmMapper) and to add logic for
4327 // their special situation.
4328 def mrs_asmoperand : AsmOperandClass {
4330 let ParserMethod = "ParseSysRegOperand";
4331 let DiagnosticType = "MRS";
4334 def mrs_op : Operand<i32> {
4335 let ParserMatchClass = mrs_asmoperand;
4336 let PrintMethod = "printMRSOperand";
4337 let DecoderMethod = "DecodeMRSOperand";
4340 def msr_asmoperand : AsmOperandClass {
4341 let Name = "MSRWithReg";
4343 // Note that SPSel is valid for both this and the pstate operands, but with
4344 // different immediate encodings. This is why these operands provide a string
4345 // AArch64Operand rather than an immediate. The overlap is small enough that
4346 // it could be resolved with hackery now, but who can say in future?
4347 let ParserMethod = "ParseSysRegOperand";
4348 let DiagnosticType = "MSR";
4351 def msr_op : Operand<i32> {
4352 let ParserMatchClass = msr_asmoperand;
4353 let PrintMethod = "printMSROperand";
4354 let DecoderMethod = "DecodeMSROperand";
4357 def pstate_asmoperand : AsmOperandClass {
4358 let Name = "MSRPState";
4359 // See comment above about parser.
4360 let ParserMethod = "ParseSysRegOperand";
4361 let DiagnosticType = "MSR";
4364 def pstate_op : Operand<i32> {
4365 let ParserMatchClass = pstate_asmoperand;
4366 let PrintMethod = "printNamedImmOperand<A64PState::PStateMapper>";
4367 let DecoderMethod = "DecodeNamedImmOperand<A64PState::PStateMapper>";
4370 // When <CRn> is specified, an assembler should accept something like "C4", not
4371 // the usual "#4" immediate.
4372 def CRx_asmoperand : AsmOperandClass {
4374 let PredicateMethod = "isUImm<4>";
4375 let RenderMethod = "addImmOperands";
4376 let ParserMethod = "ParseCRxOperand";
4377 // Diagnostics are handled in all cases by ParseCRxOperand.
4380 def CRx : Operand<i32> {
4381 let ParserMatchClass = CRx_asmoperand;
4382 let PrintMethod = "printCRxOperand";
4386 // Finally, we can start defining the instructions.
4388 // HINT is straightforward, with a few aliases.
4389 def HINTi : A64I_system<0b0, (outs), (ins uimm7:$UImm7), "hint\t$UImm7",
4392 let CRm = UImm7{6-3};
4393 let Op2 = UImm7{2-0};
4401 def : InstAlias<"nop", (HINTi 0)>;
4402 def : InstAlias<"yield", (HINTi 1)>;
4403 def : InstAlias<"wfe", (HINTi 2)>;
4404 def : InstAlias<"wfi", (HINTi 3)>;
4405 def : InstAlias<"sev", (HINTi 4)>;
4406 def : InstAlias<"sevl", (HINTi 5)>;
4408 // Quite a few instructions then follow a similar pattern of fixing common
4409 // fields in the bitpattern, we'll define a helper-class for them.
4410 class simple_sys<bits<2> op0, bits<3> op1, bits<4> crn, bits<3> op2,
4411 Operand operand, string asmop>
4412 : A64I_system<0b0, (outs), (ins operand:$CRm), !strconcat(asmop, "\t$CRm"),
4422 def CLREXi : simple_sys<0b00, 0b011, 0b0011, 0b010, uimm4, "clrex">;
4423 def DSBi : simple_sys<0b00, 0b011, 0b0011, 0b100, dbarrier_op, "dsb">;
4424 def DMBi : simple_sys<0b00, 0b011, 0b0011, 0b101, dbarrier_op, "dmb">;
4425 def ISBi : simple_sys<0b00, 0b011, 0b0011, 0b110, isb_op, "isb">;
4427 def : InstAlias<"clrex", (CLREXi 0b1111)>;
4428 def : InstAlias<"isb", (ISBi 0b1111)>;
4430 // (DMBi 0xb) is a "DMB ISH" instruciton, appropriate for Linux SMP
4431 // configurations at least.
4432 def : Pat<(atomic_fence imm, imm), (DMBi 0xb)>;
4434 // Any SYS bitpattern can be represented with a complex and opaque "SYS"
4436 def SYSiccix : A64I_system<0b0, (outs),
4437 (ins uimm3:$Op1, CRx:$CRn, CRx:$CRm,
4438 uimm3:$Op2, GPR64:$Rt),
4439 "sys\t$Op1, $CRn, $CRm, $Op2, $Rt",
4444 // You can skip the Xt argument whether it makes sense or not for the generic
4446 def : InstAlias<"sys $Op1, $CRn, $CRm, $Op2",
4447 (SYSiccix uimm3:$Op1, CRx:$CRn, CRx:$CRm, uimm3:$Op2, XZR)>;
4450 // But many have aliases, which obviously don't fit into
4451 class SYSalias<dag ins, string asmstring>
4452 : A64I_system<0b0, (outs), ins, asmstring, [], NoItinerary> {
4453 let isAsmParserOnly = 1;
4457 let Op1 = SysOp{13-11};
4458 let CRn = SysOp{10-7};
4459 let CRm = SysOp{6-3};
4460 let Op2 = SysOp{2-0};
4463 def ICix : SYSalias<(ins ic_op:$SysOp, GPR64:$Rt), "ic\t$SysOp, $Rt">;
4465 def ICi : SYSalias<(ins ic_op:$SysOp), "ic\t$SysOp"> {
4469 def DCix : SYSalias<(ins dc_op:$SysOp, GPR64:$Rt), "dc\t$SysOp, $Rt">;
4470 def ATix : SYSalias<(ins at_op:$SysOp, GPR64:$Rt), "at\t$SysOp, $Rt">;
4472 def TLBIix : SYSalias<(ins tlbi_op:$SysOp, GPR64:$Rt), "tlbi\t$SysOp, $Rt">;
4474 def TLBIi : SYSalias<(ins tlbi_op:$SysOp), "tlbi\t$SysOp"> {
4479 def SYSLxicci : A64I_system<0b1, (outs GPR64:$Rt),
4480 (ins uimm3:$Op1, CRx:$CRn, CRx:$CRm, uimm3:$Op2),
4481 "sysl\t$Rt, $Op1, $CRn, $CRm, $Op2",
4486 // The instructions themselves are rather simple for MSR and MRS.
4487 def MSRix : A64I_system<0b0, (outs), (ins msr_op:$SysReg, GPR64:$Rt),
4488 "msr\t$SysReg, $Rt", [], NoItinerary> {
4490 let Op0 = SysReg{15-14};
4491 let Op1 = SysReg{13-11};
4492 let CRn = SysReg{10-7};
4493 let CRm = SysReg{6-3};
4494 let Op2 = SysReg{2-0};
4497 def MRSxi : A64I_system<0b1, (outs GPR64:$Rt), (ins mrs_op:$SysReg),
4498 "mrs\t$Rt, $SysReg", [], NoItinerary> {
4500 let Op0 = SysReg{15-14};
4501 let Op1 = SysReg{13-11};
4502 let CRn = SysReg{10-7};
4503 let CRm = SysReg{6-3};
4504 let Op2 = SysReg{2-0};
4507 def MSRii : A64I_system<0b0, (outs), (ins pstate_op:$PState, uimm4:$CRm),
4508 "msr\t$PState, $CRm", [], NoItinerary> {
4512 let Op1 = PState{5-3};
4514 let Op2 = PState{2-0};
4518 //===----------------------------------------------------------------------===//
4519 // Test & branch (immediate) instructions
4520 //===----------------------------------------------------------------------===//
4521 // Contains: TBZ, TBNZ
4523 // The bit to test is a simple unsigned 6-bit immediate in the X-register
4525 def uimm6 : Operand<i64> {
4526 let ParserMatchClass = uimm6_asmoperand;
4529 def label_wid14_scal4_asmoperand : label_asmoperand<14, 4>;
4531 def tbimm_target : Operand<OtherVT> {
4532 let EncoderMethod = "getLabelOpValue<AArch64::fixup_a64_tstbr>";
4534 // This label is a 14-bit offset from PC, scaled by the instruction-width: 4.
4535 let PrintMethod = "printLabelOperand<14, 4>";
4536 let ParserMatchClass = label_wid14_scal4_asmoperand;
4538 let OperandType = "OPERAND_PCREL";
4541 def A64eq : ImmLeaf<i32, [{ return Imm == A64CC::EQ; }]>;
4542 def A64ne : ImmLeaf<i32, [{ return Imm == A64CC::NE; }]>;
4544 // These instructions correspond to patterns involving "and" with a power of
4545 // two, which we need to be able to select.
4546 def tstb64_pat : ComplexPattern<i64, 1, "SelectTSTBOperand<64>">;
4547 def tstb32_pat : ComplexPattern<i32, 1, "SelectTSTBOperand<32>">;
4549 let isBranch = 1, isTerminator = 1 in {
4550 def TBZxii : A64I_TBimm<0b0, (outs),
4551 (ins GPR64:$Rt, uimm6:$Imm, tbimm_target:$Label),
4552 "tbz\t$Rt, $Imm, $Label",
4553 [(A64br_cc (A64cmp (and i64:$Rt, tstb64_pat:$Imm), 0),
4558 def TBNZxii : A64I_TBimm<0b1, (outs),
4559 (ins GPR64:$Rt, uimm6:$Imm, tbimm_target:$Label),
4560 "tbnz\t$Rt, $Imm, $Label",
4561 [(A64br_cc (A64cmp (and i64:$Rt, tstb64_pat:$Imm), 0),
4567 // Note, these instructions overlap with the above 64-bit patterns. This is
4568 // intentional, "tbz x3, #1, somewhere" and "tbz w3, #1, somewhere" would both
4569 // do the same thing and are both permitted assembly. They also both have
4570 // sensible DAG patterns.
4571 def TBZwii : A64I_TBimm<0b0, (outs),
4572 (ins GPR32:$Rt, uimm5:$Imm, tbimm_target:$Label),
4573 "tbz\t$Rt, $Imm, $Label",
4574 [(A64br_cc (A64cmp (and i32:$Rt, tstb32_pat:$Imm), 0),
4581 def TBNZwii : A64I_TBimm<0b1, (outs),
4582 (ins GPR32:$Rt, uimm5:$Imm, tbimm_target:$Label),
4583 "tbnz\t$Rt, $Imm, $Label",
4584 [(A64br_cc (A64cmp (and i32:$Rt, tstb32_pat:$Imm), 0),
4592 //===----------------------------------------------------------------------===//
4593 // Unconditional branch (immediate) instructions
4594 //===----------------------------------------------------------------------===//
4597 def label_wid26_scal4_asmoperand : label_asmoperand<26, 4>;
4599 def bimm_target : Operand<OtherVT> {
4600 let EncoderMethod = "getLabelOpValue<AArch64::fixup_a64_uncondbr>";
4602 // This label is a 26-bit offset from PC, scaled by the instruction-width: 4.
4603 let PrintMethod = "printLabelOperand<26, 4>";
4604 let ParserMatchClass = label_wid26_scal4_asmoperand;
4606 let OperandType = "OPERAND_PCREL";
4609 def blimm_target : Operand<i64> {
4610 let EncoderMethod = "getLabelOpValue<AArch64::fixup_a64_call>";
4612 // This label is a 26-bit offset from PC, scaled by the instruction-width: 4.
4613 let PrintMethod = "printLabelOperand<26, 4>";
4614 let ParserMatchClass = label_wid26_scal4_asmoperand;
4616 let OperandType = "OPERAND_PCREL";
4619 class A64I_BimmImpl<bit op, string asmop, list<dag> patterns, Operand lbl_type>
4620 : A64I_Bimm<op, (outs), (ins lbl_type:$Label),
4621 !strconcat(asmop, "\t$Label"), patterns,
4625 let isBranch = 1 in {
4626 def Bimm : A64I_BimmImpl<0b0, "b", [(br bb:$Label)], bimm_target> {
4627 let isTerminator = 1;
4631 let SchedRW = [WriteBrL] in {
4632 def BLimm : A64I_BimmImpl<0b1, "bl",
4633 [(AArch64Call tglobaladdr:$Label)], blimm_target> {
4640 def : Pat<(AArch64Call texternalsym:$Label), (BLimm texternalsym:$Label)>;
4642 //===----------------------------------------------------------------------===//
4643 // Unconditional branch (register) instructions
4644 //===----------------------------------------------------------------------===//
4645 // Contains: BR, BLR, RET, ERET, DRP.
4647 // Most of the notional opcode fields in the A64I_Breg format are fixed in A64
4649 class A64I_BregImpl<bits<4> opc,
4650 dag outs, dag ins, string asmstr, list<dag> patterns,
4651 InstrItinClass itin = NoItinerary>
4652 : A64I_Breg<opc, 0b11111, 0b000000, 0b00000,
4653 outs, ins, asmstr, patterns, itin>,
4656 let isIndirectBranch = 1;
4659 // Note that these are not marked isCall or isReturn because as far as LLVM is
4660 // concerned they're not. "ret" is just another jump unless it has been selected
4661 // by LLVM as the function's return.
4663 let isBranch = 1 in {
4664 def BRx : A64I_BregImpl<0b0000,(outs), (ins GPR64:$Rn),
4665 "br\t$Rn", [(brind i64:$Rn)]> {
4667 let isTerminator = 1;
4670 let SchedRW = [WriteBrL] in {
4671 def BLRx : A64I_BregImpl<0b0001, (outs), (ins GPR64:$Rn),
4672 "blr\t$Rn", [(AArch64Call i64:$Rn)]> {
4679 def RETx : A64I_BregImpl<0b0010, (outs), (ins GPR64:$Rn),
4682 let isTerminator = 1;
4686 // Create a separate pseudo-instruction for codegen to use so that we don't
4687 // flag x30 as used in every function. It'll be restored before the RET by the
4688 // epilogue if it's legitimately used.
4689 def RET : A64PseudoExpand<(outs), (ins), [(A64ret)], (RETx (ops X30))> {
4690 let isTerminator = 1;
4695 def ERET : A64I_BregImpl<0b0100, (outs), (ins), "eret", []> {
4698 let isTerminator = 1;
4702 def DRPS : A64I_BregImpl<0b0101, (outs), (ins), "drps", []> {
4708 def RETAlias : InstAlias<"ret", (RETx X30)>;
4711 //===----------------------------------------------------------------------===//
4712 // Address generation patterns
4713 //===----------------------------------------------------------------------===//
4715 // Primary method of address generation for the small/absolute memory model is
4716 // an ADRP/ADR pair:
4717 // ADRP x0, some_variable
4718 // ADD x0, x0, #:lo12:some_variable
4720 // The load/store elision of the ADD is accomplished when selecting
4721 // addressing-modes. This just mops up the cases where that doesn't work and we
4722 // really need an address in some register.
4724 // This wrapper applies a LO12 modifier to the address. Otherwise we could just
4725 // use the same address.
4727 class ADRP_ADD<SDNode Wrapper, SDNode addrop>
4728 : Pat<(Wrapper addrop:$Hi, addrop:$Lo12, (i32 imm)),
4729 (ADDxxi_lsl0_s (ADRPxi addrop:$Hi), addrop:$Lo12)>;
4731 def : ADRP_ADD<A64WrapperSmall, tblockaddress>;
4732 def : ADRP_ADD<A64WrapperSmall, texternalsym>;
4733 def : ADRP_ADD<A64WrapperSmall, tglobaladdr>;
4734 def : ADRP_ADD<A64WrapperSmall, tglobaltlsaddr>;
4735 def : ADRP_ADD<A64WrapperSmall, tjumptable>;
4736 def : ADRP_ADD<A64WrapperSmall, tconstpool>;
4738 //===----------------------------------------------------------------------===//
4739 // GOT access patterns
4740 //===----------------------------------------------------------------------===//
4742 class GOTLoadSmall<SDNode addrfrag>
4743 : Pat<(A64GOTLoad (A64WrapperSmall addrfrag:$Hi, addrfrag:$Lo12, 8)),
4744 (LS64_LDR (ADRPxi addrfrag:$Hi), addrfrag:$Lo12)>;
4746 def : GOTLoadSmall<texternalsym>;
4747 def : GOTLoadSmall<tglobaladdr>;
4748 def : GOTLoadSmall<tglobaltlsaddr>;
4750 //===----------------------------------------------------------------------===//
4751 // Tail call handling
4752 //===----------------------------------------------------------------------===//
4754 let isCall = 1, isTerminator = 1, isReturn = 1, isBarrier = 1, Uses = [XSP] in {
4756 : PseudoInst<(outs), (ins i64imm:$dst, i32imm:$FPDiff),
4757 [(AArch64tcret tglobaladdr:$dst, (i32 timm:$FPDiff))]>;
4760 : PseudoInst<(outs), (ins tcGPR64:$dst, i32imm:$FPDiff),
4761 [(AArch64tcret i64:$dst, (i32 timm:$FPDiff))]>;
4764 let isCall = 1, isTerminator = 1, isReturn = 1, isBarrier = 1,
4766 def TAIL_Bimm : A64PseudoExpand<(outs), (ins bimm_target:$Label), [],
4767 (Bimm bimm_target:$Label)>;
4769 def TAIL_BRx : A64PseudoExpand<(outs), (ins tcGPR64:$Rd), [],
4774 def : Pat<(AArch64tcret texternalsym:$dst, (i32 timm:$FPDiff)),
4775 (TC_RETURNdi texternalsym:$dst, imm:$FPDiff)>;
4777 //===----------------------------------------------------------------------===//
4778 // Thread local storage
4779 //===----------------------------------------------------------------------===//
4781 // This is a pseudo-instruction representing the ".tlsdesccall" directive in
4782 // assembly. Its effect is to insert an R_AARCH64_TLSDESC_CALL relocation at the
4783 // current location. It should always be immediately followed by a BLR
4784 // instruction, and is intended solely for relaxation by the linker.
4786 def : Pat<(A64threadpointer), (MRSxi 0xde82)>;
4788 def TLSDESCCALL : PseudoInst<(outs), (ins i64imm:$Lbl), []> {
4789 let hasSideEffects = 1;
4792 def TLSDESC_BLRx : PseudoInst<(outs), (ins GPR64:$Rn, i64imm:$Var),
4793 [(A64tlsdesc_blr i64:$Rn, tglobaltlsaddr:$Var)]> {
4798 def : Pat<(A64tlsdesc_blr i64:$Rn, texternalsym:$Var),
4799 (TLSDESC_BLRx $Rn, texternalsym:$Var)>;
4801 //===----------------------------------------------------------------------===//
4802 // Bitfield patterns
4803 //===----------------------------------------------------------------------===//
4805 def bfi32_lsb_to_immr : SDNodeXForm<imm, [{
4806 return CurDAG->getTargetConstant((32 - N->getZExtValue()) % 32, MVT::i64);
4809 def bfi64_lsb_to_immr : SDNodeXForm<imm, [{
4810 return CurDAG->getTargetConstant((64 - N->getZExtValue()) % 64, MVT::i64);
4813 def bfi_width_to_imms : SDNodeXForm<imm, [{
4814 return CurDAG->getTargetConstant(N->getZExtValue() - 1, MVT::i64);
4818 // The simpler patterns deal with cases where no AND mask is actually needed
4819 // (either all bits are used or the low 32 bits are used).
4820 let AddedComplexity = 10 in {
4822 def : Pat<(A64Bfi i64:$src, i64:$Rn, imm:$ImmR, imm:$ImmS),
4824 (bfi64_lsb_to_immr (i64 imm:$ImmR)),
4825 (bfi_width_to_imms (i64 imm:$ImmS)))>;
4827 def : Pat<(A64Bfi i32:$src, i32:$Rn, imm:$ImmR, imm:$ImmS),
4829 (bfi32_lsb_to_immr (i64 imm:$ImmR)),
4830 (bfi_width_to_imms (i64 imm:$ImmS)))>;
4833 def : Pat<(and (A64Bfi i64:$src, i64:$Rn, imm:$ImmR, imm:$ImmS),
4835 (SUBREG_TO_REG (i64 0),
4836 (BFIwwii (EXTRACT_SUBREG $src, sub_32),
4837 (EXTRACT_SUBREG $Rn, sub_32),
4838 (bfi32_lsb_to_immr (i64 imm:$ImmR)),
4839 (bfi_width_to_imms (i64 imm:$ImmS))),
4844 //===----------------------------------------------------------------------===//
4845 // Miscellaneous patterns
4846 //===----------------------------------------------------------------------===//
4848 // Truncation from 64 to 32-bits just involves renaming your register.
4849 def : Pat<(i32 (trunc i64:$val)), (EXTRACT_SUBREG $val, sub_32)>;
4851 // Similarly, extension where we don't care about the high bits is
4853 def : Pat<(i64 (anyext i32:$val)),
4854 (INSERT_SUBREG (IMPLICIT_DEF), $val, sub_32)>;
4856 // SELECT instructions providing f128 types need to be handled by a
4857 // pseudo-instruction since the eventual code will need to introduce basic
4858 // blocks and control flow.
4859 def F128CSEL : PseudoInst<(outs FPR128:$Rd),
4860 (ins FPR128:$Rn, FPR128:$Rm, cond_code_op:$Cond),
4861 [(set f128:$Rd, (simple_select f128:$Rn, f128:$Rm))]> {
4863 let usesCustomInserter = 1;
4866 //===----------------------------------------------------------------------===//
4867 // Load/store patterns
4868 //===----------------------------------------------------------------------===//
4870 // There are lots of patterns here, because we need to allow at least three
4871 // parameters to vary independently.
4872 // 1. Instruction: "ldrb w9, [sp]", "ldrh w9, [sp]", ...
4873 // 2. LLVM source: zextloadi8, anyextloadi8, ...
4874 // 3. Address-generation: A64Wrapper, (add BASE, OFFSET), ...
4876 // The biggest problem turns out to be the address-generation variable. At the
4877 // point of instantiation we need to produce two DAGs, one for the pattern and
4878 // one for the instruction. Doing this at the lowest level of classes doesn't
4881 // Consider the simple uimm12 addressing mode, and the desire to match both (add
4882 // GPR64xsp:$Rn, uimm12:$Offset) and GPR64xsp:$Rn, particularly on the
4883 // instruction side. We'd need to insert either "GPR64xsp" and "uimm12" or
4884 // "GPR64xsp" and "0" into an unknown dag. !subst is not capable of this
4885 // operation, and PatFrags are for selection not output.
4887 // As a result, the address-generation patterns are the final
4888 // instantiations. However, we do still need to vary the operand for the address
4889 // further down (At the point we're deciding A64WrapperSmall, we don't know
4890 // the memory width of the operation).
4892 //===------------------------------
4893 // 1. Basic infrastructural defs
4894 //===------------------------------
4896 // First, some simple classes for !foreach and !subst to use:
4907 // You can't use !subst on an actual immediate, but you *can* use it on an
4908 // operand record that happens to match a single immediate. So we do.
4909 def imm_eq0 : ImmLeaf<i64, [{ return Imm == 0; }]>;
4910 def imm_eq1 : ImmLeaf<i64, [{ return Imm == 1; }]>;
4911 def imm_eq2 : ImmLeaf<i64, [{ return Imm == 2; }]>;
4912 def imm_eq3 : ImmLeaf<i64, [{ return Imm == 3; }]>;
4913 def imm_eq4 : ImmLeaf<i64, [{ return Imm == 4; }]>;
4915 // If the low bits of a pointer are known to be 0 then an "or" is just as good
4916 // as addition for computing an offset. This fragment forwards that check for
4918 def add_like_or : PatFrag<(ops node:$lhs, node:$rhs), (or node:$lhs, node:$rhs),
4920 return CurDAG->isBaseWithConstantOffset(SDValue(N, 0));
4923 // Load/store (unsigned immediate) operations with relocations against global
4924 // symbols (for lo12) are only valid if those symbols have correct alignment
4925 // (since the immediate offset is divided by the access scale, it can't have a
4928 // The guaranteed alignment is provided as part of the WrapperSmall
4929 // operation, and checked against one of these.
4930 def any_align : ImmLeaf<i32, [{ (void)Imm; return true; }]>;
4931 def min_align2 : ImmLeaf<i32, [{ return Imm >= 2; }]>;
4932 def min_align4 : ImmLeaf<i32, [{ return Imm >= 4; }]>;
4933 def min_align8 : ImmLeaf<i32, [{ return Imm >= 8; }]>;
4934 def min_align16 : ImmLeaf<i32, [{ return Imm >= 16; }]>;
4936 // "Normal" load/store instructions can be used on atomic operations, provided
4937 // the ordering parameter is at most "monotonic". Anything above that needs
4938 // special handling with acquire/release instructions.
4939 class simple_load<PatFrag base>
4940 : PatFrag<(ops node:$ptr), (base node:$ptr), [{
4941 return cast<AtomicSDNode>(N)->getOrdering() <= Monotonic;
4944 def atomic_load_simple_i8 : simple_load<atomic_load_8>;
4945 def atomic_load_simple_i16 : simple_load<atomic_load_16>;
4946 def atomic_load_simple_i32 : simple_load<atomic_load_32>;
4947 def atomic_load_simple_i64 : simple_load<atomic_load_64>;
4949 class simple_store<PatFrag base>
4950 : PatFrag<(ops node:$ptr, node:$val), (base node:$ptr, node:$val), [{
4951 return cast<AtomicSDNode>(N)->getOrdering() <= Monotonic;
4954 def atomic_store_simple_i8 : simple_store<atomic_store_8>;
4955 def atomic_store_simple_i16 : simple_store<atomic_store_16>;
4956 def atomic_store_simple_i32 : simple_store<atomic_store_32>;
4957 def atomic_store_simple_i64 : simple_store<atomic_store_64>;
4959 //===------------------------------
4960 // 2. UImm12 and SImm9
4961 //===------------------------------
4963 // These instructions have two operands providing the address so they can be
4964 // treated similarly for most purposes.
4966 //===------------------------------
4967 // 2.1 Base patterns covering extend/truncate semantics
4968 //===------------------------------
4970 // Atomic patterns can be shared between integer operations of all sizes, a
4971 // quick multiclass here allows reuse.
4972 multiclass ls_atomic_pats<Instruction LOAD, Instruction STORE, dag Base,
4973 dag Offset, dag address, ValueType transty,
4975 def : Pat<(!cast<PatFrag>("atomic_load_simple_" # sty) address),
4976 (LOAD Base, Offset)>;
4978 def : Pat<(!cast<PatFrag>("atomic_store_simple_" # sty) address, transty:$Rt),
4979 (STORE $Rt, Base, Offset)>;
4982 // Instructions accessing a memory chunk smaller than a register (or, in a
4983 // pinch, the same size) have a characteristic set of patterns they want to
4984 // match: extending loads and truncating stores. This class deals with the
4985 // sign-neutral version of those patterns.
4987 // It will be instantiated across multiple addressing-modes.
4988 multiclass ls_small_pats<Instruction LOAD, Instruction STORE,
4989 dag Base, dag Offset,
4990 dag address, ValueType sty>
4991 : ls_atomic_pats<LOAD, STORE, Base, Offset, address, i32, sty> {
4992 def : Pat<(!cast<SDNode>(zextload # sty) address), (LOAD Base, Offset)>;
4994 def : Pat<(!cast<SDNode>(extload # sty) address), (LOAD Base, Offset)>;
4996 // For zero-extension to 64-bits we have to tell LLVM that the whole 64-bit
4997 // register was actually set.
4998 def : Pat<(i64 (!cast<SDNode>(zextload # sty) address)),
4999 (SUBREG_TO_REG (i64 0), (LOAD Base, Offset), sub_32)>;
5001 def : Pat<(i64 (!cast<SDNode>(extload # sty) address)),
5002 (SUBREG_TO_REG (i64 0), (LOAD Base, Offset), sub_32)>;
5004 def : Pat<(!cast<SDNode>(truncstore # sty) i32:$Rt, address),
5005 (STORE $Rt, Base, Offset)>;
5007 // For truncating store from 64-bits, we have to manually tell LLVM to
5008 // ignore the high bits of the x register.
5009 def : Pat<(!cast<SDNode>(truncstore # sty) i64:$Rt, address),
5010 (STORE (EXTRACT_SUBREG $Rt, sub_32), Base, Offset)>;
5013 // Next come patterns for sign-extending loads.
5014 multiclass load_signed_pats<string T, string U, dag Base, dag Offset,
5015 dag address, ValueType sty> {
5016 def : Pat<(i32 (!cast<SDNode>("sextload" # sty) address)),
5017 (!cast<Instruction>("LDRS" # T # "w" # U) Base, Offset)>;
5019 def : Pat<(i64 (!cast<SDNode>("sextload" # sty) address)),
5020 (!cast<Instruction>("LDRS" # T # "x" # U) Base, Offset)>;
5024 // and finally "natural-width" loads and stores come next.
5025 multiclass ls_neutral_pats<Instruction LOAD, Instruction STORE, dag Base,
5026 dag Offset, dag address, ValueType sty> {
5027 def : Pat<(sty (load address)), (LOAD Base, Offset)>;
5028 def : Pat<(store sty:$Rt, address), (STORE $Rt, Base, Offset)>;
5031 // Integer operations also get atomic instructions to select for.
5032 multiclass ls_int_neutral_pats<Instruction LOAD, Instruction STORE, dag Base,
5033 dag Offset, dag address, ValueType sty>
5034 : ls_neutral_pats<LOAD, STORE, Base, Offset, address, sty>,
5035 ls_atomic_pats<LOAD, STORE, Base, Offset, address, sty, sty>;
5037 //===------------------------------
5038 // 2.2. Addressing-mode instantiations
5039 //===------------------------------
5041 multiclass uimm12_pats<dag address, dag Base, dag Offset> {
5042 defm : ls_small_pats<LS8_LDR, LS8_STR, Base,
5043 !foreach(decls.pattern, Offset,
5044 !subst(OFFSET, byte_uimm12, decls.pattern)),
5045 !foreach(decls.pattern, address,
5046 !subst(OFFSET, byte_uimm12,
5047 !subst(ALIGN, any_align, decls.pattern))),
5049 defm : ls_small_pats<LS16_LDR, LS16_STR, Base,
5050 !foreach(decls.pattern, Offset,
5051 !subst(OFFSET, hword_uimm12, decls.pattern)),
5052 !foreach(decls.pattern, address,
5053 !subst(OFFSET, hword_uimm12,
5054 !subst(ALIGN, min_align2, decls.pattern))),
5056 defm : ls_small_pats<LS32_LDR, LS32_STR, Base,
5057 !foreach(decls.pattern, Offset,
5058 !subst(OFFSET, word_uimm12, decls.pattern)),
5059 !foreach(decls.pattern, address,
5060 !subst(OFFSET, word_uimm12,
5061 !subst(ALIGN, min_align4, decls.pattern))),
5064 defm : ls_int_neutral_pats<LS32_LDR, LS32_STR, Base,
5065 !foreach(decls.pattern, Offset,
5066 !subst(OFFSET, word_uimm12, decls.pattern)),
5067 !foreach(decls.pattern, address,
5068 !subst(OFFSET, word_uimm12,
5069 !subst(ALIGN, min_align4, decls.pattern))),
5072 defm : ls_int_neutral_pats<LS64_LDR, LS64_STR, Base,
5073 !foreach(decls.pattern, Offset,
5074 !subst(OFFSET, dword_uimm12, decls.pattern)),
5075 !foreach(decls.pattern, address,
5076 !subst(OFFSET, dword_uimm12,
5077 !subst(ALIGN, min_align8, decls.pattern))),
5080 defm : ls_neutral_pats<LSFP16_LDR, LSFP16_STR, Base,
5081 !foreach(decls.pattern, Offset,
5082 !subst(OFFSET, hword_uimm12, decls.pattern)),
5083 !foreach(decls.pattern, address,
5084 !subst(OFFSET, hword_uimm12,
5085 !subst(ALIGN, min_align2, decls.pattern))),
5088 defm : ls_neutral_pats<LSFP32_LDR, LSFP32_STR, Base,
5089 !foreach(decls.pattern, Offset,
5090 !subst(OFFSET, word_uimm12, decls.pattern)),
5091 !foreach(decls.pattern, address,
5092 !subst(OFFSET, word_uimm12,
5093 !subst(ALIGN, min_align4, decls.pattern))),
5096 defm : ls_neutral_pats<LSFP64_LDR, LSFP64_STR, Base,
5097 !foreach(decls.pattern, Offset,
5098 !subst(OFFSET, dword_uimm12, decls.pattern)),
5099 !foreach(decls.pattern, address,
5100 !subst(OFFSET, dword_uimm12,
5101 !subst(ALIGN, min_align8, decls.pattern))),
5104 defm : ls_neutral_pats<LSFP128_LDR, LSFP128_STR, Base,
5105 !foreach(decls.pattern, Offset,
5106 !subst(OFFSET, qword_uimm12, decls.pattern)),
5107 !foreach(decls.pattern, address,
5108 !subst(OFFSET, qword_uimm12,
5109 !subst(ALIGN, min_align16, decls.pattern))),
5112 defm : load_signed_pats<"B", "", Base,
5113 !foreach(decls.pattern, Offset,
5114 !subst(OFFSET, byte_uimm12, decls.pattern)),
5115 !foreach(decls.pattern, address,
5116 !subst(OFFSET, byte_uimm12,
5117 !subst(ALIGN, any_align, decls.pattern))),
5120 defm : load_signed_pats<"H", "", Base,
5121 !foreach(decls.pattern, Offset,
5122 !subst(OFFSET, hword_uimm12, decls.pattern)),
5123 !foreach(decls.pattern, address,
5124 !subst(OFFSET, hword_uimm12,
5125 !subst(ALIGN, min_align2, decls.pattern))),
5128 def : Pat<(sextloadi32 !foreach(decls.pattern, address,
5129 !subst(OFFSET, word_uimm12,
5130 !subst(ALIGN, min_align4, decls.pattern)))),
5131 (LDRSWx Base, !foreach(decls.pattern, Offset,
5132 !subst(OFFSET, word_uimm12, decls.pattern)))>;
5135 // Straightforward patterns of last resort: a pointer with or without an
5136 // appropriate offset.
5137 defm : uimm12_pats<(i64 i64:$Rn), (i64 i64:$Rn), (i64 0)>;
5138 defm : uimm12_pats<(add i64:$Rn, OFFSET:$UImm12),
5139 (i64 i64:$Rn), (i64 OFFSET:$UImm12)>;
5141 // The offset could be hidden behind an "or", of course:
5142 defm : uimm12_pats<(add_like_or i64:$Rn, OFFSET:$UImm12),
5143 (i64 i64:$Rn), (i64 OFFSET:$UImm12)>;
5145 // Global addresses under the small-absolute model should use these
5146 // instructions. There are ELF relocations specifically for it.
5147 defm : uimm12_pats<(A64WrapperSmall tglobaladdr:$Hi, tglobaladdr:$Lo12, ALIGN),
5148 (ADRPxi tglobaladdr:$Hi), (i64 tglobaladdr:$Lo12)>;
5150 defm : uimm12_pats<(A64WrapperSmall tglobaltlsaddr:$Hi, tglobaltlsaddr:$Lo12,
5152 (ADRPxi tglobaltlsaddr:$Hi), (i64 tglobaltlsaddr:$Lo12)>;
5154 // External symbols that make it this far should also get standard relocations.
5155 defm : uimm12_pats<(A64WrapperSmall texternalsym:$Hi, texternalsym:$Lo12,
5157 (ADRPxi texternalsym:$Hi), (i64 texternalsym:$Lo12)>;
5159 defm : uimm12_pats<(A64WrapperSmall tconstpool:$Hi, tconstpool:$Lo12, ALIGN),
5160 (ADRPxi tconstpool:$Hi), (i64 tconstpool:$Lo12)>;
5162 // We also want to use uimm12 instructions for local variables at the moment.
5163 def tframeindex_XFORM : SDNodeXForm<frameindex, [{
5164 int FI = cast<FrameIndexSDNode>(N)->getIndex();
5165 return CurDAG->getTargetFrameIndex(FI, MVT::i64);
5168 defm : uimm12_pats<(i64 frameindex:$Rn),
5169 (tframeindex_XFORM tframeindex:$Rn), (i64 0)>;
5171 // These can be much simpler than uimm12 because we don't to change the operand
5172 // type (e.g. LDURB and LDURH take the same operands).
5173 multiclass simm9_pats<dag address, dag Base, dag Offset> {
5174 defm : ls_small_pats<LS8_LDUR, LS8_STUR, Base, Offset, address, i8>;
5175 defm : ls_small_pats<LS16_LDUR, LS16_STUR, Base, Offset, address, i16>;
5177 defm : ls_int_neutral_pats<LS32_LDUR, LS32_STUR, Base, Offset, address, i32>;
5178 defm : ls_int_neutral_pats<LS64_LDUR, LS64_STUR, Base, Offset, address, i64>;
5180 defm : ls_neutral_pats<LSFP16_LDUR, LSFP16_STUR, Base, Offset, address, f16>;
5181 defm : ls_neutral_pats<LSFP32_LDUR, LSFP32_STUR, Base, Offset, address, f32>;
5182 defm : ls_neutral_pats<LSFP64_LDUR, LSFP64_STUR, Base, Offset, address, f64>;
5183 defm : ls_neutral_pats<LSFP128_LDUR, LSFP128_STUR, Base, Offset, address,
5186 def : Pat<(i64 (zextloadi32 address)),
5187 (SUBREG_TO_REG (i64 0), (LS32_LDUR Base, Offset), sub_32)>;
5189 def : Pat<(truncstorei32 i64:$Rt, address),
5190 (LS32_STUR (EXTRACT_SUBREG $Rt, sub_32), Base, Offset)>;
5192 defm : load_signed_pats<"B", "_U", Base, Offset, address, i8>;
5193 defm : load_signed_pats<"H", "_U", Base, Offset, address, i16>;
5194 def : Pat<(sextloadi32 address), (LDURSWx Base, Offset)>;
5197 defm : simm9_pats<(add i64:$Rn, simm9:$SImm9),
5198 (i64 $Rn), (SDXF_simm9 simm9:$SImm9)>;
5200 defm : simm9_pats<(add_like_or i64:$Rn, simm9:$SImm9),
5201 (i64 $Rn), (SDXF_simm9 simm9:$SImm9)>;
5204 //===------------------------------
5205 // 3. Register offset patterns
5206 //===------------------------------
5208 // Atomic patterns can be shared between integer operations of all sizes, a
5209 // quick multiclass here allows reuse.
5210 multiclass ro_atomic_pats<Instruction LOAD, Instruction STORE, dag Base,
5211 dag Offset, dag Extend, dag address,
5212 ValueType transty, ValueType sty> {
5213 def : Pat<(!cast<PatFrag>("atomic_load_simple_" # sty) address),
5214 (LOAD Base, Offset, Extend)>;
5216 def : Pat<(!cast<PatFrag>("atomic_store_simple_" # sty) address, transty:$Rt),
5217 (STORE $Rt, Base, Offset, Extend)>;
5220 // The register offset instructions take three operands giving the instruction,
5221 // and have an annoying split between instructions where Rm is 32-bit and
5222 // 64-bit. So we need a special hierarchy to describe them. Other than that the
5223 // same operations should be supported as for simm9 and uimm12 addressing.
5225 multiclass ro_small_pats<Instruction LOAD, Instruction STORE,
5226 dag Base, dag Offset, dag Extend,
5227 dag address, ValueType sty>
5228 : ro_atomic_pats<LOAD, STORE, Base, Offset, Extend, address, i32, sty> {
5229 def : Pat<(!cast<SDNode>(zextload # sty) address),
5230 (LOAD Base, Offset, Extend)>;
5232 def : Pat<(!cast<SDNode>(extload # sty) address),
5233 (LOAD Base, Offset, Extend)>;
5235 // For zero-extension to 64-bits we have to tell LLVM that the whole 64-bit
5236 // register was actually set.
5237 def : Pat<(i64 (!cast<SDNode>(zextload # sty) address)),
5238 (SUBREG_TO_REG (i64 0), (LOAD Base, Offset, Extend), sub_32)>;
5240 def : Pat<(i64 (!cast<SDNode>(extload # sty) address)),
5241 (SUBREG_TO_REG (i64 0), (LOAD Base, Offset, Extend), sub_32)>;
5243 def : Pat<(!cast<SDNode>(truncstore # sty) i32:$Rt, address),
5244 (STORE $Rt, Base, Offset, Extend)>;
5246 // For truncating store from 64-bits, we have to manually tell LLVM to
5247 // ignore the high bits of the x register.
5248 def : Pat<(!cast<SDNode>(truncstore # sty) i64:$Rt, address),
5249 (STORE (EXTRACT_SUBREG $Rt, sub_32), Base, Offset, Extend)>;
5253 // Next come patterns for sign-extending loads.
5254 multiclass ro_signed_pats<string T, string Rm, dag Base, dag Offset, dag Extend,
5255 dag address, ValueType sty> {
5256 def : Pat<(i32 (!cast<SDNode>("sextload" # sty) address)),
5257 (!cast<Instruction>("LDRS" # T # "w_" # Rm # "_RegOffset")
5258 Base, Offset, Extend)>;
5260 def : Pat<(i64 (!cast<SDNode>("sextload" # sty) address)),
5261 (!cast<Instruction>("LDRS" # T # "x_" # Rm # "_RegOffset")
5262 Base, Offset, Extend)>;
5265 // and finally "natural-width" loads and stores come next.
5266 multiclass ro_neutral_pats<Instruction LOAD, Instruction STORE,
5267 dag Base, dag Offset, dag Extend, dag address,
5269 def : Pat<(sty (load address)), (LOAD Base, Offset, Extend)>;
5270 def : Pat<(store sty:$Rt, address),
5271 (STORE $Rt, Base, Offset, Extend)>;
5274 multiclass ro_int_neutral_pats<Instruction LOAD, Instruction STORE,
5275 dag Base, dag Offset, dag Extend, dag address,
5277 : ro_neutral_pats<LOAD, STORE, Base, Offset, Extend, address, sty>,
5278 ro_atomic_pats<LOAD, STORE, Base, Offset, Extend, address, sty, sty>;
5280 multiclass regoff_pats<string Rm, dag address, dag Base, dag Offset,
5282 defm : ro_small_pats<!cast<Instruction>("LS8_" # Rm # "_RegOffset_LDR"),
5283 !cast<Instruction>("LS8_" # Rm # "_RegOffset_STR"),
5284 Base, Offset, Extend,
5285 !foreach(decls.pattern, address,
5286 !subst(SHIFT, imm_eq0, decls.pattern)),
5288 defm : ro_small_pats<!cast<Instruction>("LS16_" # Rm # "_RegOffset_LDR"),
5289 !cast<Instruction>("LS16_" # Rm # "_RegOffset_STR"),
5290 Base, Offset, Extend,
5291 !foreach(decls.pattern, address,
5292 !subst(SHIFT, imm_eq1, decls.pattern)),
5294 defm : ro_small_pats<!cast<Instruction>("LS32_" # Rm # "_RegOffset_LDR"),
5295 !cast<Instruction>("LS32_" # Rm # "_RegOffset_STR"),
5296 Base, Offset, Extend,
5297 !foreach(decls.pattern, address,
5298 !subst(SHIFT, imm_eq2, decls.pattern)),
5301 defm : ro_int_neutral_pats<
5302 !cast<Instruction>("LS32_" # Rm # "_RegOffset_LDR"),
5303 !cast<Instruction>("LS32_" # Rm # "_RegOffset_STR"),
5304 Base, Offset, Extend,
5305 !foreach(decls.pattern, address,
5306 !subst(SHIFT, imm_eq2, decls.pattern)),
5309 defm : ro_int_neutral_pats<
5310 !cast<Instruction>("LS64_" # Rm # "_RegOffset_LDR"),
5311 !cast<Instruction>("LS64_" # Rm # "_RegOffset_STR"),
5312 Base, Offset, Extend,
5313 !foreach(decls.pattern, address,
5314 !subst(SHIFT, imm_eq3, decls.pattern)),
5317 defm : ro_neutral_pats<!cast<Instruction>("LSFP16_" # Rm # "_RegOffset_LDR"),
5318 !cast<Instruction>("LSFP16_" # Rm # "_RegOffset_STR"),
5319 Base, Offset, Extend,
5320 !foreach(decls.pattern, address,
5321 !subst(SHIFT, imm_eq1, decls.pattern)),
5324 defm : ro_neutral_pats<!cast<Instruction>("LSFP32_" # Rm # "_RegOffset_LDR"),
5325 !cast<Instruction>("LSFP32_" # Rm # "_RegOffset_STR"),
5326 Base, Offset, Extend,
5327 !foreach(decls.pattern, address,
5328 !subst(SHIFT, imm_eq2, decls.pattern)),
5331 defm : ro_neutral_pats<!cast<Instruction>("LSFP64_" # Rm # "_RegOffset_LDR"),
5332 !cast<Instruction>("LSFP64_" # Rm # "_RegOffset_STR"),
5333 Base, Offset, Extend,
5334 !foreach(decls.pattern, address,
5335 !subst(SHIFT, imm_eq3, decls.pattern)),
5338 defm : ro_neutral_pats<!cast<Instruction>("LSFP128_" # Rm # "_RegOffset_LDR"),
5339 !cast<Instruction>("LSFP128_" # Rm # "_RegOffset_STR"),
5340 Base, Offset, Extend,
5341 !foreach(decls.pattern, address,
5342 !subst(SHIFT, imm_eq4, decls.pattern)),
5345 defm : ro_signed_pats<"B", Rm, Base, Offset, Extend,
5346 !foreach(decls.pattern, address,
5347 !subst(SHIFT, imm_eq0, decls.pattern)),
5350 defm : ro_signed_pats<"H", Rm, Base, Offset, Extend,
5351 !foreach(decls.pattern, address,
5352 !subst(SHIFT, imm_eq1, decls.pattern)),
5355 def : Pat<(sextloadi32 !foreach(decls.pattern, address,
5356 !subst(SHIFT, imm_eq2, decls.pattern))),
5357 (!cast<Instruction>("LDRSWx_" # Rm # "_RegOffset")
5358 Base, Offset, Extend)>;
5362 // Finally we're in a position to tell LLVM exactly what addresses are reachable
5363 // using register-offset instructions. Essentially a base plus a possibly
5364 // extended, possibly shifted (by access size) offset.
5366 defm : regoff_pats<"Wm", (add i64:$Rn, (sext i32:$Rm)),
5367 (i64 i64:$Rn), (i32 i32:$Rm), (i64 6)>;
5369 defm : regoff_pats<"Wm", (add i64:$Rn, (shl (sext i32:$Rm), SHIFT)),
5370 (i64 i64:$Rn), (i32 i32:$Rm), (i64 7)>;
5372 defm : regoff_pats<"Wm", (add i64:$Rn, (zext i32:$Rm)),
5373 (i64 i64:$Rn), (i32 i32:$Rm), (i64 2)>;
5375 defm : regoff_pats<"Wm", (add i64:$Rn, (shl (zext i32:$Rm), SHIFT)),
5376 (i64 i64:$Rn), (i32 i32:$Rm), (i64 3)>;
5378 defm : regoff_pats<"Xm", (add i64:$Rn, i64:$Rm),
5379 (i64 i64:$Rn), (i64 i64:$Rm), (i64 2)>;
5381 defm : regoff_pats<"Xm", (add i64:$Rn, (shl i64:$Rm, SHIFT)),
5382 (i64 i64:$Rn), (i64 i64:$Rm), (i64 3)>;
5384 //===----------------------------------------------------------------------===//
5385 // Advanced SIMD (NEON) Support
5388 include "AArch64InstrNEON.td"