1 include "AArch64InstrFormats.td"
3 //===----------------------------------------------------------------------===//
4 // Target-specific ISD nodes and profiles
5 //===----------------------------------------------------------------------===//
7 def SDT_A64ret : SDTypeProfile<0, 0, []>;
8 def A64ret : SDNode<"AArch64ISD::Ret", SDT_A64ret, [SDNPHasChain,
12 // (ins NZCV, Condition, Dest)
13 def SDT_A64br_cc : SDTypeProfile<0, 3, [SDTCisVT<0, i32>]>;
14 def A64br_cc : SDNode<"AArch64ISD::BR_CC", SDT_A64br_cc, [SDNPHasChain]>;
16 // (outs Result), (ins NZCV, IfTrue, IfFalse, Condition)
17 def SDT_A64select_cc : SDTypeProfile<1, 4, [SDTCisVT<1, i32>,
20 def A64select_cc : SDNode<"AArch64ISD::SELECT_CC", SDT_A64select_cc>;
22 // (outs NZCV), (ins LHS, RHS, Condition)
23 def SDT_A64setcc : SDTypeProfile<1, 3, [SDTCisVT<0, i32>,
25 def A64setcc : SDNode<"AArch64ISD::SETCC", SDT_A64setcc>;
28 // (outs GPR64), (ins)
29 def A64threadpointer : SDNode<"AArch64ISD::THREAD_POINTER", SDTPtrLeaf>;
31 // A64 compares don't care about the cond really (they set all flags) so a
32 // simple binary operator is useful.
33 def A64cmp : PatFrag<(ops node:$lhs, node:$rhs),
34 (A64setcc node:$lhs, node:$rhs, cond)>;
37 // When matching a notional (CMP op1, (sub 0, op2)), we'd like to use a CMN
38 // instruction on the grounds that "op1 - (-op2) == op1 + op2". However, the C
39 // and V flags can be set differently by this operation. It comes down to
40 // whether "SInt(~op2)+1 == SInt(~op2+1)" (and the same for UInt). If they are
41 // then everything is fine. If not then the optimization is wrong. Thus general
42 // comparisons are only valid if op2 != 0.
44 // So, finally, the only LLVM-native comparisons that don't mention C and V are
45 // SETEQ and SETNE. They're the only ones we can safely use CMN for in the
46 // absence of information about op2.
47 def equality_cond : PatLeaf<(cond), [{
48 return N->get() == ISD::SETEQ || N->get() == ISD::SETNE;
51 def A64cmn : PatFrag<(ops node:$lhs, node:$rhs),
52 (A64setcc node:$lhs, (sub 0, node:$rhs), equality_cond)>;
54 // There are two layers of indirection here, driven by the following
56 // + TableGen does not know CodeModel or Reloc so that decision should be
57 // made for a variable/address at ISelLowering.
58 // + The output of ISelLowering should be selectable (hence the Wrapper,
59 // rather than a bare target opcode)
60 def SDTAArch64Wrapper : SDTypeProfile<1, 3, [SDTCisSameAs<0, 1>,
65 def A64WrapperSmall : SDNode<"AArch64ISD::WrapperSmall", SDTAArch64Wrapper>;
68 def SDTAArch64GOTLoad : SDTypeProfile<1, 1, [SDTCisPtrTy<0>, SDTCisPtrTy<1>]>;
69 def A64GOTLoad : SDNode<"AArch64ISD::GOTLoad", SDTAArch64GOTLoad,
73 // (A64BFI LHS, RHS, LSB, Width)
74 def SDTA64BFI : SDTypeProfile<1, 4, [SDTCisSameAs<0, 1>,
79 def A64Bfi : SDNode<"AArch64ISD::BFI", SDTA64BFI>;
81 // (A64EXTR HiReg, LoReg, LSB)
82 def SDTA64EXTR : SDTypeProfile<1, 3, [SDTCisSameAs<0, 1>, SDTCisSameAs<1, 2>,
84 def A64Extr : SDNode<"AArch64ISD::EXTR", SDTA64EXTR>;
86 // (A64[SU]BFX Field, ImmR, ImmS).
88 // Note that ImmR and ImmS are already encoded for the actual instructions. The
89 // more natural LSB and Width mix together to form ImmR and ImmS, something
90 // which TableGen can't handle.
91 def SDTA64BFX : SDTypeProfile<1, 3, [SDTCisVT<2, i64>, SDTCisVT<3, i64>]>;
92 def A64Sbfx : SDNode<"AArch64ISD::SBFX", SDTA64BFX>;
94 def A64Ubfx : SDNode<"AArch64ISD::UBFX", SDTA64BFX>;
96 //===----------------------------------------------------------------------===//
97 // Call sequence pseudo-instructions
98 //===----------------------------------------------------------------------===//
101 def SDT_AArch64Call : SDTypeProfile<0, -1, [SDTCisPtrTy<0>]>;
102 def AArch64Call : SDNode<"AArch64ISD::Call", SDT_AArch64Call,
103 [SDNPHasChain, SDNPOptInGlue, SDNPOutGlue, SDNPVariadic]>;
105 def AArch64tcret : SDNode<"AArch64ISD::TC_RETURN", SDT_AArch64Call,
106 [SDNPHasChain, SDNPOptInGlue, SDNPVariadic]>;
108 // The TLSDESCCALL node is a variant call which goes to an indirectly calculated
109 // destination but needs a relocation against a fixed symbol. As such it has two
110 // certain operands: the callee and the relocated variable.
112 // The TLS ABI only allows it to be selected to a BLR instructin (with
113 // appropriate relocation).
114 def SDTTLSDescCall : SDTypeProfile<0, -2, [SDTCisPtrTy<0>, SDTCisPtrTy<1>]>;
116 def A64tlsdesc_blr : SDNode<"AArch64ISD::TLSDESCCALL", SDTTLSDescCall,
117 [SDNPInGlue, SDNPOutGlue, SDNPHasChain,
121 def SDT_AArch64CallSeqStart : SDCallSeqStart<[ SDTCisPtrTy<0> ]>;
122 def AArch64callseq_start : SDNode<"ISD::CALLSEQ_START", SDT_AArch64CallSeqStart,
123 [SDNPHasChain, SDNPOutGlue]>;
125 def SDT_AArch64CallSeqEnd : SDCallSeqEnd<[ SDTCisPtrTy<0>, SDTCisPtrTy<1> ]>;
126 def AArch64callseq_end : SDNode<"ISD::CALLSEQ_END", SDT_AArch64CallSeqEnd,
127 [SDNPHasChain, SDNPOptInGlue, SDNPOutGlue]>;
131 // These pseudo-instructions have special semantics by virtue of being passed to
132 // the InstrInfo constructor. CALLSEQ_START/CALLSEQ_END are produced by
133 // LowerCall to (in our case) tell the back-end about stack adjustments for
134 // arguments passed on the stack. Here we select those markers to
135 // pseudo-instructions which explicitly set the stack, and finally in the
136 // RegisterInfo we convert them to a true stack adjustment.
137 let Defs = [XSP], Uses = [XSP] in {
138 def ADJCALLSTACKDOWN : PseudoInst<(outs), (ins i64imm:$amt),
139 [(AArch64callseq_start timm:$amt)]>;
141 def ADJCALLSTACKUP : PseudoInst<(outs), (ins i64imm:$amt1, i64imm:$amt2),
142 [(AArch64callseq_end timm:$amt1, timm:$amt2)]>;
145 //===----------------------------------------------------------------------===//
146 // Atomic operation pseudo-instructions
147 //===----------------------------------------------------------------------===//
149 let usesCustomInserter = 1, Defs = [NZCV] in {
150 multiclass AtomicSizes<string opname> {
151 def _I8 : PseudoInst<(outs GPR32:$dst), (ins GPR64:$ptr, GPR32:$incr),
152 [(set GPR32:$dst, (!cast<SDNode>(opname # "_8") GPR64:$ptr, GPR32:$incr))]>;
153 def _I16 : PseudoInst<(outs GPR32:$dst), (ins GPR64:$ptr, GPR32:$incr),
154 [(set GPR32:$dst, (!cast<SDNode>(opname # "_16") GPR64:$ptr, GPR32:$incr))]>;
155 def _I32 : PseudoInst<(outs GPR32:$dst), (ins GPR64:$ptr, GPR32:$incr),
156 [(set GPR32:$dst, (!cast<SDNode>(opname # "_32") GPR64:$ptr, GPR32:$incr))]>;
157 def _I64 : PseudoInst<(outs GPR64:$dst), (ins GPR64:$ptr, GPR64:$incr),
158 [(set GPR64:$dst, (!cast<SDNode>(opname # "_64") GPR64:$ptr, GPR64:$incr))]>;
162 defm ATOMIC_LOAD_ADD : AtomicSizes<"atomic_load_add">;
163 defm ATOMIC_LOAD_SUB : AtomicSizes<"atomic_load_sub">;
164 defm ATOMIC_LOAD_AND : AtomicSizes<"atomic_load_and">;
165 defm ATOMIC_LOAD_OR : AtomicSizes<"atomic_load_or">;
166 defm ATOMIC_LOAD_XOR : AtomicSizes<"atomic_load_xor">;
167 defm ATOMIC_LOAD_NAND : AtomicSizes<"atomic_load_nand">;
168 defm ATOMIC_LOAD_MIN : AtomicSizes<"atomic_load_min">;
169 defm ATOMIC_LOAD_MAX : AtomicSizes<"atomic_load_max">;
170 defm ATOMIC_LOAD_UMIN : AtomicSizes<"atomic_load_umin">;
171 defm ATOMIC_LOAD_UMAX : AtomicSizes<"atomic_load_umax">;
172 defm ATOMIC_SWAP : AtomicSizes<"atomic_swap">;
174 let usesCustomInserter = 1, Defs = [NZCV] in {
175 def ATOMIC_CMP_SWAP_I8
176 : PseudoInst<(outs GPR32:$dst), (ins GPR64:$ptr, GPR32:$old, GPR32:$new),
178 (atomic_cmp_swap_8 GPR64:$ptr, GPR32:$old, GPR32:$new))]>;
179 def ATOMIC_CMP_SWAP_I16
180 : PseudoInst<(outs GPR32:$dst), (ins GPR64:$ptr, GPR32:$old, GPR32:$new),
182 (atomic_cmp_swap_16 GPR64:$ptr, GPR32:$old, GPR32:$new))]>;
183 def ATOMIC_CMP_SWAP_I32
184 : PseudoInst<(outs GPR32:$dst), (ins GPR64:$ptr, GPR32:$old, GPR32:$new),
186 (atomic_cmp_swap_32 GPR64:$ptr, GPR32:$old, GPR32:$new))]>;
187 def ATOMIC_CMP_SWAP_I64
188 : PseudoInst<(outs GPR64:$dst), (ins GPR64:$ptr, GPR64:$old, GPR64:$new),
190 (atomic_cmp_swap_64 GPR64:$ptr, GPR64:$old, GPR64:$new))]>;
193 //===----------------------------------------------------------------------===//
194 // Add-subtract (extended register) instructions
195 //===----------------------------------------------------------------------===//
196 // Contains: ADD, ADDS, SUB, SUBS + aliases CMN, CMP
198 // The RHS of these operations is conceptually a sign/zero-extended
199 // register, optionally shifted left by 1-4. The extension can be a
200 // NOP (e.g. "sxtx" sign-extending a 64-bit register to 64-bits) but
201 // must be specified with one exception:
203 // If one of the registers is sp/wsp then LSL is an alias for UXTW in
204 // 32-bit instructions and UXTX in 64-bit versions, the shift amount
205 // is not optional in that case (but can explicitly be 0), and the
206 // entire suffix can be skipped (e.g. "add sp, x3, x2").
208 multiclass extend_operands<string PREFIX> {
209 def _asmoperand : AsmOperandClass {
211 let RenderMethod = "addRegExtendOperands";
212 let PredicateMethod = "isRegExtend<A64SE::" # PREFIX # ">";
215 def _operand : Operand<i64>,
216 ImmLeaf<i64, [{ return Imm >= 0 && Imm <= 4; }]> {
217 let PrintMethod = "printRegExtendOperand<A64SE::" # PREFIX # ">";
218 let DecoderMethod = "DecodeRegExtendOperand";
219 let ParserMatchClass = !cast<AsmOperandClass>(PREFIX # "_asmoperand");
223 defm UXTB : extend_operands<"UXTB">;
224 defm UXTH : extend_operands<"UXTH">;
225 defm UXTW : extend_operands<"UXTW">;
226 defm UXTX : extend_operands<"UXTX">;
227 defm SXTB : extend_operands<"SXTB">;
228 defm SXTH : extend_operands<"SXTH">;
229 defm SXTW : extend_operands<"SXTW">;
230 defm SXTX : extend_operands<"SXTX">;
232 def LSL_extasmoperand : AsmOperandClass {
233 let Name = "RegExtendLSL";
234 let RenderMethod = "addRegExtendOperands";
237 def LSL_extoperand : Operand<i64> {
238 let ParserMatchClass = LSL_extasmoperand;
242 // The patterns for various sign-extensions are a little ugly and
243 // non-uniform because everything has already been promoted to the
244 // legal i64 and i32 types. We'll wrap the various variants up in a
245 // class for use later.
247 dag uxtb; dag uxth; dag uxtw; dag uxtx;
248 dag sxtb; dag sxth; dag sxtw; dag sxtx;
251 def extends_to_i64 : extend_types {
252 let uxtb = (and (anyext GPR32:$Rm), 255);
253 let uxth = (and (anyext GPR32:$Rm), 65535);
254 let uxtw = (zext GPR32:$Rm);
255 let uxtx = (i64 GPR64:$Rm);
257 let sxtb = (sext_inreg (anyext GPR32:$Rm), i8);
258 let sxth = (sext_inreg (anyext GPR32:$Rm), i16);
259 let sxtw = (sext GPR32:$Rm);
260 let sxtx = (i64 GPR64:$Rm);
264 def extends_to_i32 : extend_types {
265 let uxtb = (and GPR32:$Rm, 255);
266 let uxth = (and GPR32:$Rm, 65535);
267 let uxtw = (i32 GPR32:$Rm);
268 let uxtx = (i32 GPR32:$Rm);
270 let sxtb = (sext_inreg GPR32:$Rm, i8);
271 let sxth = (sext_inreg GPR32:$Rm, i16);
272 let sxtw = (i32 GPR32:$Rm);
273 let sxtx = (i32 GPR32:$Rm);
276 // Now, six of the extensions supported are easy and uniform: if the source size
277 // is 32-bits or less, then Rm is always a 32-bit register. We'll instantiate
278 // those instructions in one block.
280 // The uxtx/sxtx could potentially be merged in, but three facts dissuaded me:
281 // + It would break the naming scheme: either ADDxx_uxtx or ADDww_uxtx would
283 // + Patterns are very different as well.
284 // + Passing different registers would be ugly (more fields in extend_types
285 // would probably be the best option).
286 multiclass addsub_exts<bit sf, bit op, bit S, string asmop,
287 SDPatternOperator opfrag,
288 dag outs, extend_types exts, RegisterClass GPRsp> {
289 def w_uxtb : A64I_addsubext<sf, op, S, 0b00, 0b000,
291 (ins GPRsp:$Rn, GPR32:$Rm, UXTB_operand:$Imm3),
292 !strconcat(asmop, "$Rn, $Rm, $Imm3"),
293 [(opfrag GPRsp:$Rn, (shl exts.uxtb, UXTB_operand:$Imm3))],
295 def w_uxth : A64I_addsubext<sf, op, S, 0b00, 0b001,
297 (ins GPRsp:$Rn, GPR32:$Rm, UXTH_operand:$Imm3),
298 !strconcat(asmop, "$Rn, $Rm, $Imm3"),
299 [(opfrag GPRsp:$Rn, (shl exts.uxth, UXTH_operand:$Imm3))],
301 def w_uxtw : A64I_addsubext<sf, op, S, 0b00, 0b010,
303 (ins GPRsp:$Rn, GPR32:$Rm, UXTW_operand:$Imm3),
304 !strconcat(asmop, "$Rn, $Rm, $Imm3"),
305 [(opfrag GPRsp:$Rn, (shl exts.uxtw, UXTW_operand:$Imm3))],
308 def w_sxtb : A64I_addsubext<sf, op, S, 0b00, 0b100,
310 (ins GPRsp:$Rn, GPR32:$Rm, SXTB_operand:$Imm3),
311 !strconcat(asmop, "$Rn, $Rm, $Imm3"),
312 [(opfrag GPRsp:$Rn, (shl exts.sxtb, SXTB_operand:$Imm3))],
314 def w_sxth : A64I_addsubext<sf, op, S, 0b00, 0b101,
316 (ins GPRsp:$Rn, GPR32:$Rm, SXTH_operand:$Imm3),
317 !strconcat(asmop, "$Rn, $Rm, $Imm3"),
318 [(opfrag GPRsp:$Rn, (shl exts.sxth, SXTH_operand:$Imm3))],
320 def w_sxtw : A64I_addsubext<sf, op, S, 0b00, 0b110,
322 (ins GPRsp:$Rn, GPR32:$Rm, SXTW_operand:$Imm3),
323 !strconcat(asmop, "$Rn, $Rm, $Imm3"),
324 [(opfrag GPRsp:$Rn, (shl exts.sxtw, SXTW_operand:$Imm3))],
328 // These two could be merge in with the above, but their patterns aren't really
329 // necessary and the naming-scheme would necessarily break:
330 multiclass addsub_xxtx<bit op, bit S, string asmop, SDPatternOperator opfrag,
332 def x_uxtx : A64I_addsubext<0b1, op, S, 0b00, 0b011,
334 (ins GPR64xsp:$Rn, GPR64:$Rm, UXTX_operand:$Imm3),
335 !strconcat(asmop, "$Rn, $Rm, $Imm3"),
336 [(opfrag GPR64xsp:$Rn, (shl GPR64:$Rm, UXTX_operand:$Imm3))],
339 def x_sxtx : A64I_addsubext<0b1, op, S, 0b00, 0b111,
341 (ins GPR64xsp:$Rn, GPR64:$Rm, SXTX_operand:$Imm3),
342 !strconcat(asmop, "$Rn, $Rm, $Imm3"),
343 [/* No Pattern: same as uxtx */],
347 multiclass addsub_wxtx<bit op, bit S, string asmop, dag outs> {
348 def w_uxtx : A64I_addsubext<0b0, op, S, 0b00, 0b011,
350 (ins GPR32wsp:$Rn, GPR32:$Rm, UXTX_operand:$Imm3),
351 !strconcat(asmop, "$Rn, $Rm, $Imm3"),
352 [/* No pattern: probably same as uxtw */],
355 def w_sxtx : A64I_addsubext<0b0, op, S, 0b00, 0b111,
357 (ins GPR32wsp:$Rn, GPR32:$Rm, SXTX_operand:$Imm3),
358 !strconcat(asmop, "$Rn, $Rm, $Imm3"),
359 [/* No Pattern: probably same as uxtw */],
363 class SetRD<RegisterClass RC, SDPatternOperator op>
364 : PatFrag<(ops node:$lhs, node:$rhs), (set RC:$Rd, (op node:$lhs, node:$rhs))>;
365 class SetNZCV<SDPatternOperator op>
366 : PatFrag<(ops node:$lhs, node:$rhs), (set NZCV, (op node:$lhs, node:$rhs))>;
368 defm ADDxx :addsub_exts<0b1, 0b0, 0b0, "add\t$Rd, ", SetRD<GPR64xsp, add>,
369 (outs GPR64xsp:$Rd), extends_to_i64, GPR64xsp>,
370 addsub_xxtx< 0b0, 0b0, "add\t$Rd, ", SetRD<GPR64xsp, add>,
371 (outs GPR64xsp:$Rd)>;
372 defm ADDww :addsub_exts<0b0, 0b0, 0b0, "add\t$Rd, ", SetRD<GPR32wsp, add>,
373 (outs GPR32wsp:$Rd), extends_to_i32, GPR32wsp>,
374 addsub_wxtx< 0b0, 0b0, "add\t$Rd, ",
375 (outs GPR32wsp:$Rd)>;
376 defm SUBxx :addsub_exts<0b1, 0b1, 0b0, "sub\t$Rd, ", SetRD<GPR64xsp, sub>,
377 (outs GPR64xsp:$Rd), extends_to_i64, GPR64xsp>,
378 addsub_xxtx< 0b1, 0b0, "sub\t$Rd, ", SetRD<GPR64xsp, sub>,
379 (outs GPR64xsp:$Rd)>;
380 defm SUBww :addsub_exts<0b0, 0b1, 0b0, "sub\t$Rd, ", SetRD<GPR32wsp, sub>,
381 (outs GPR32wsp:$Rd), extends_to_i32, GPR32wsp>,
382 addsub_wxtx< 0b1, 0b0, "sub\t$Rd, ",
383 (outs GPR32wsp:$Rd)>;
385 let Defs = [NZCV] in {
386 defm ADDSxx :addsub_exts<0b1, 0b0, 0b1, "adds\t$Rd, ", SetRD<GPR64, addc>,
387 (outs GPR64:$Rd), extends_to_i64, GPR64xsp>,
388 addsub_xxtx< 0b0, 0b1, "adds\t$Rd, ", SetRD<GPR64, addc>,
390 defm ADDSww :addsub_exts<0b0, 0b0, 0b1, "adds\t$Rd, ", SetRD<GPR32, addc>,
391 (outs GPR32:$Rd), extends_to_i32, GPR32wsp>,
392 addsub_wxtx< 0b0, 0b1, "adds\t$Rd, ",
394 defm SUBSxx :addsub_exts<0b1, 0b1, 0b1, "subs\t$Rd, ", SetRD<GPR64, subc>,
395 (outs GPR64:$Rd), extends_to_i64, GPR64xsp>,
396 addsub_xxtx< 0b1, 0b1, "subs\t$Rd, ", SetRD<GPR64, subc>,
398 defm SUBSww :addsub_exts<0b0, 0b1, 0b1, "subs\t$Rd, ", SetRD<GPR32, subc>,
399 (outs GPR32:$Rd), extends_to_i32, GPR32wsp>,
400 addsub_wxtx< 0b1, 0b1, "subs\t$Rd, ",
404 let Rd = 0b11111, isCompare = 1 in {
405 defm CMNx : addsub_exts<0b1, 0b0, 0b1, "cmn\t", SetNZCV<A64cmn>,
406 (outs), extends_to_i64, GPR64xsp>,
407 addsub_xxtx< 0b0, 0b1, "cmn\t", SetNZCV<A64cmn>, (outs)>;
408 defm CMNw : addsub_exts<0b0, 0b0, 0b1, "cmn\t", SetNZCV<A64cmn>,
409 (outs), extends_to_i32, GPR32wsp>,
410 addsub_wxtx< 0b0, 0b1, "cmn\t", (outs)>;
411 defm CMPx : addsub_exts<0b1, 0b1, 0b1, "cmp\t", SetNZCV<A64cmp>,
412 (outs), extends_to_i64, GPR64xsp>,
413 addsub_xxtx< 0b1, 0b1, "cmp\t", SetNZCV<A64cmp>, (outs)>;
414 defm CMPw : addsub_exts<0b0, 0b1, 0b1, "cmp\t", SetNZCV<A64cmp>,
415 (outs), extends_to_i32, GPR32wsp>,
416 addsub_wxtx< 0b1, 0b1, "cmp\t", (outs)>;
420 // Now patterns for the operation without a shift being needed. No patterns are
421 // created for uxtx/sxtx since they're non-uniform and it's expected that
422 // add/sub (shifted register) will handle those cases anyway.
423 multiclass addsubext_noshift_patterns<string prefix, SDPatternOperator nodeop,
424 RegisterClass GPRsp, extend_types exts> {
425 def : Pat<(nodeop GPRsp:$Rn, exts.uxtb),
426 (!cast<Instruction>(prefix # "w_uxtb") GPRsp:$Rn, GPR32:$Rm, 0)>;
427 def : Pat<(nodeop GPRsp:$Rn, exts.uxth),
428 (!cast<Instruction>(prefix # "w_uxth") GPRsp:$Rn, GPR32:$Rm, 0)>;
429 def : Pat<(nodeop GPRsp:$Rn, exts.uxtw),
430 (!cast<Instruction>(prefix # "w_uxtw") GPRsp:$Rn, GPR32:$Rm, 0)>;
432 def : Pat<(nodeop GPRsp:$Rn, exts.sxtb),
433 (!cast<Instruction>(prefix # "w_sxtb") GPRsp:$Rn, GPR32:$Rm, 0)>;
434 def : Pat<(nodeop GPRsp:$Rn, exts.sxth),
435 (!cast<Instruction>(prefix # "w_sxth") GPRsp:$Rn, GPR32:$Rm, 0)>;
436 def : Pat<(nodeop GPRsp:$Rn, exts.sxtw),
437 (!cast<Instruction>(prefix # "w_sxtw") GPRsp:$Rn, GPR32:$Rm, 0)>;
440 defm : addsubext_noshift_patterns<"ADDxx", add, GPR64xsp, extends_to_i64>;
441 defm : addsubext_noshift_patterns<"ADDww", add, GPR32wsp, extends_to_i32>;
442 defm : addsubext_noshift_patterns<"SUBxx", sub, GPR64xsp, extends_to_i64>;
443 defm : addsubext_noshift_patterns<"SUBww", sub, GPR32wsp, extends_to_i32>;
445 defm : addsubext_noshift_patterns<"CMNx", A64cmn, GPR64xsp, extends_to_i64>;
446 defm : addsubext_noshift_patterns<"CMNw", A64cmn, GPR32wsp, extends_to_i32>;
447 defm : addsubext_noshift_patterns<"CMPx", A64cmp, GPR64xsp, extends_to_i64>;
448 defm : addsubext_noshift_patterns<"CMPw", A64cmp, GPR32wsp, extends_to_i32>;
450 // An extend of "lsl #imm" is valid if and only if one of Rn and Rd is
451 // sp/wsp. It is synonymous with uxtx/uxtw depending on the size of the
452 // operation. Also permitted in this case is complete omission of the argument,
453 // which implies "lsl #0".
454 multiclass lsl_aliases<string asmop, Instruction inst, RegisterClass GPR_Rd,
455 RegisterClass GPR_Rn, RegisterClass GPR_Rm> {
456 def : InstAlias<!strconcat(asmop, " $Rd, $Rn, $Rm"),
457 (inst GPR_Rd:$Rd, GPR_Rn:$Rn, GPR_Rm:$Rm, 0)>;
459 def : InstAlias<!strconcat(asmop, " $Rd, $Rn, $Rm, $LSL"),
460 (inst GPR_Rd:$Rd, GPR_Rn:$Rn, GPR_Rm:$Rm, LSL_extoperand:$LSL)>;
464 defm : lsl_aliases<"add", ADDxxx_uxtx, Rxsp, GPR64xsp, GPR64>;
465 defm : lsl_aliases<"add", ADDxxx_uxtx, GPR64xsp, Rxsp, GPR64>;
466 defm : lsl_aliases<"add", ADDwww_uxtw, Rwsp, GPR32wsp, GPR32>;
467 defm : lsl_aliases<"add", ADDwww_uxtw, GPR32wsp, Rwsp, GPR32>;
468 defm : lsl_aliases<"sub", SUBxxx_uxtx, Rxsp, GPR64xsp, GPR64>;
469 defm : lsl_aliases<"sub", SUBxxx_uxtx, GPR64xsp, Rxsp, GPR64>;
470 defm : lsl_aliases<"sub", SUBwww_uxtw, Rwsp, GPR32wsp, GPR32>;
471 defm : lsl_aliases<"sub", SUBwww_uxtw, GPR32wsp, Rwsp, GPR32>;
473 // Rd cannot be sp for flag-setting variants so only half of the aliases are
475 defm : lsl_aliases<"adds", ADDSxxx_uxtx, GPR64, Rxsp, GPR64>;
476 defm : lsl_aliases<"adds", ADDSwww_uxtw, GPR32, Rwsp, GPR32>;
477 defm : lsl_aliases<"subs", SUBSxxx_uxtx, GPR64, Rxsp, GPR64>;
478 defm : lsl_aliases<"subs", SUBSwww_uxtw, GPR32, Rwsp, GPR32>;
480 // CMP unfortunately has to be different because the instruction doesn't have a
482 multiclass cmp_lsl_aliases<string asmop, Instruction inst,
483 RegisterClass GPR_Rn, RegisterClass GPR_Rm> {
484 def : InstAlias<!strconcat(asmop, " $Rn, $Rm"),
485 (inst GPR_Rn:$Rn, GPR_Rm:$Rm, 0)>;
487 def : InstAlias<!strconcat(asmop, " $Rn, $Rm, $LSL"),
488 (inst GPR_Rn:$Rn, GPR_Rm:$Rm, LSL_extoperand:$LSL)>;
491 defm : cmp_lsl_aliases<"cmp", CMPxx_uxtx, Rxsp, GPR64>;
492 defm : cmp_lsl_aliases<"cmp", CMPww_uxtw, Rwsp, GPR32>;
493 defm : cmp_lsl_aliases<"cmn", CMNxx_uxtx, Rxsp, GPR64>;
494 defm : cmp_lsl_aliases<"cmn", CMNww_uxtw, Rwsp, GPR32>;
496 //===----------------------------------------------------------------------===//
497 // Add-subtract (immediate) instructions
498 //===----------------------------------------------------------------------===//
499 // Contains: ADD, ADDS, SUB, SUBS + aliases CMN, CMP, MOV
501 // These instructions accept a 12-bit unsigned immediate, optionally shifted
502 // left by 12 bits. Official assembly format specifies a 12 bit immediate with
503 // one of "", "LSL #0", "LSL #12" supplementary operands.
505 // There are surprisingly few ways to make this work with TableGen, so this
506 // implementation has separate instructions for the "LSL #0" and "LSL #12"
509 // If the MCInst retained a single combined immediate (which could be 0x123000,
510 // for example) then both components (imm & shift) would have to be delegated to
511 // a single assembly operand. This would entail a separate operand parser
512 // (because the LSL would have to live in the same AArch64Operand as the
513 // immediate to be accessible); assembly parsing is rather complex and
514 // error-prone C++ code.
516 // By splitting the immediate, we can delegate handling this optional operand to
517 // an InstAlias. Supporting functions to generate the correct MCInst are still
518 // required, but these are essentially trivial and parsing can remain generic.
520 // Rejected plans with rationale:
521 // ------------------------------
523 // In an ideal world you'de have two first class immediate operands (in
524 // InOperandList, specifying imm12 and shift). Unfortunately this is not
525 // selectable by any means I could discover.
527 // An Instruction with two MCOperands hidden behind a single entry in
528 // InOperandList (expanded by ComplexPatterns and MIOperandInfo) was functional,
529 // but required more C++ code to handle encoding/decoding. Parsing (the intended
530 // main beneficiary) ended up equally complex because of the optional nature of
533 // Attempting to circumvent the need for a custom OperandParser above by giving
534 // InstAliases without the "lsl #0" failed. add/sub could be accommodated but
535 // the cmp/cmn aliases didn't use the MIOperandInfo to determine how operands
536 // should be parsed: there was no way to accommodate an "lsl #12".
538 let ParserMethod = "ParseImmWithLSLOperand",
539 RenderMethod = "addImmWithLSLOperands" in {
540 // Derived PredicateMethod fields are different for each
541 def addsubimm_lsl0_asmoperand : AsmOperandClass {
542 let Name = "AddSubImmLSL0";
545 def addsubimm_lsl12_asmoperand : AsmOperandClass {
546 let Name = "AddSubImmLSL12";
550 def shr_12_XFORM : SDNodeXForm<imm, [{
551 return CurDAG->getTargetConstant(N->getSExtValue() >> 12, MVT::i32);
554 def shr_12_neg_XFORM : SDNodeXForm<imm, [{
555 return CurDAG->getTargetConstant((-N->getSExtValue()) >> 12, MVT::i32);
558 def neg_XFORM : SDNodeXForm<imm, [{
559 return CurDAG->getTargetConstant(-N->getSExtValue(), MVT::i32);
563 multiclass addsub_imm_operands<ValueType ty> {
564 let PrintMethod = "printAddSubImmLSL0Operand",
565 EncoderMethod = "getAddSubImmOpValue",
566 ParserMatchClass = addsubimm_lsl0_asmoperand in {
567 def _posimm_lsl0 : Operand<ty>,
568 ImmLeaf<ty, [{ return Imm >= 0 && (Imm & ~0xfff) == 0; }]>;
569 def _negimm_lsl0 : Operand<ty>,
570 ImmLeaf<ty, [{ return Imm < 0 && (-Imm & ~0xfff) == 0; }],
574 let PrintMethod = "printAddSubImmLSL12Operand",
575 EncoderMethod = "getAddSubImmOpValue",
576 ParserMatchClass = addsubimm_lsl12_asmoperand in {
577 def _posimm_lsl12 : Operand<ty>,
578 ImmLeaf<ty, [{ return Imm >= 0 && (Imm & ~0xfff000) == 0; }],
581 def _negimm_lsl12 : Operand<ty>,
582 ImmLeaf<ty, [{ return Imm < 0 && (-Imm & ~0xfff000) == 0; }],
587 // The add operands don't need any transformation
588 defm addsubimm_operand_i32 : addsub_imm_operands<i32>;
589 defm addsubimm_operand_i64 : addsub_imm_operands<i64>;
591 multiclass addsubimm_varieties<string prefix, bit sf, bit op, bits<2> shift,
592 string asmop, string cmpasmop,
593 Operand imm_operand, Operand cmp_imm_operand,
594 RegisterClass GPR, RegisterClass GPRsp,
596 // All registers for non-S variants allow SP
597 def _s : A64I_addsubimm<sf, op, 0b0, shift,
599 (ins GPRsp:$Rn, imm_operand:$Imm12),
600 !strconcat(asmop, "\t$Rd, $Rn, $Imm12"),
602 (add GPRsp:$Rn, imm_operand:$Imm12))],
606 // S variants can read SP but would write to ZR
607 def _S : A64I_addsubimm<sf, op, 0b1, shift,
609 (ins GPRsp:$Rn, imm_operand:$Imm12),
610 !strconcat(asmop, "s\t$Rd, $Rn, $Imm12"),
611 [(set GPR:$Rd, (addc GPRsp:$Rn, imm_operand:$Imm12))],
616 // Note that the pattern here for ADDS is subtle. Canonically CMP
617 // a, b becomes SUBS a, b. If b < 0 then this is equivalent to
618 // ADDS a, (-b). This is not true in general.
619 def _cmp : A64I_addsubimm<sf, op, 0b1, shift,
620 (outs), (ins GPRsp:$Rn, imm_operand:$Imm12),
621 !strconcat(cmpasmop, " $Rn, $Imm12"),
623 (A64cmp GPRsp:$Rn, cmp_imm_operand:$Imm12))],
632 multiclass addsubimm_shifts<string prefix, bit sf, bit op,
633 string asmop, string cmpasmop, string operand, string cmpoperand,
634 RegisterClass GPR, RegisterClass GPRsp, AArch64Reg ZR> {
635 defm _lsl0 : addsubimm_varieties<prefix # "_lsl0", sf, op, 0b00,
637 !cast<Operand>(operand # "_lsl0"),
638 !cast<Operand>(cmpoperand # "_lsl0"),
641 defm _lsl12 : addsubimm_varieties<prefix # "_lsl12", sf, op, 0b01,
643 !cast<Operand>(operand # "_lsl12"),
644 !cast<Operand>(cmpoperand # "_lsl12"),
648 defm ADDwwi : addsubimm_shifts<"ADDwi", 0b0, 0b0, "add", "cmn",
649 "addsubimm_operand_i32_posimm",
650 "addsubimm_operand_i32_negimm",
651 GPR32, GPR32wsp, WZR>;
652 defm ADDxxi : addsubimm_shifts<"ADDxi", 0b1, 0b0, "add", "cmn",
653 "addsubimm_operand_i64_posimm",
654 "addsubimm_operand_i64_negimm",
655 GPR64, GPR64xsp, XZR>;
656 defm SUBwwi : addsubimm_shifts<"SUBwi", 0b0, 0b1, "sub", "cmp",
657 "addsubimm_operand_i32_negimm",
658 "addsubimm_operand_i32_posimm",
659 GPR32, GPR32wsp, WZR>;
660 defm SUBxxi : addsubimm_shifts<"SUBxi", 0b1, 0b1, "sub", "cmp",
661 "addsubimm_operand_i64_negimm",
662 "addsubimm_operand_i64_posimm",
663 GPR64, GPR64xsp, XZR>;
665 multiclass MOVsp<RegisterClass GPRsp, RegisterClass SP, Instruction addop> {
666 def _fromsp : InstAlias<"mov $Rd, $Rn",
667 (addop GPRsp:$Rd, SP:$Rn, 0),
670 def _tosp : InstAlias<"mov $Rd, $Rn",
671 (addop SP:$Rd, GPRsp:$Rn, 0),
675 // Recall Rxsp is a RegisterClass containing *just* xsp.
676 defm MOVxx : MOVsp<GPR64xsp, Rxsp, ADDxxi_lsl0_s>;
677 defm MOVww : MOVsp<GPR32wsp, Rwsp, ADDwwi_lsl0_s>;
679 //===----------------------------------------------------------------------===//
680 // Add-subtract (shifted register) instructions
681 //===----------------------------------------------------------------------===//
682 // Contains: ADD, ADDS, SUB, SUBS + aliases CMN, CMP, NEG, NEGS
684 //===-------------------------------
685 // 1. The "shifed register" operands. Shared with logical insts.
686 //===-------------------------------
688 multiclass shift_operands<string prefix, string form> {
689 def _asmoperand_i32 : AsmOperandClass {
690 let Name = "Shift" # form # "i32";
691 let RenderMethod = "addShiftOperands";
693 = "isShift<A64SE::" # form # ", false>";
696 // Note that the operand type is intentionally i64 because the DAGCombiner
697 // puts these into a canonical form.
698 def _i32 : Operand<i64>, ImmLeaf<i64, [{ return Imm >= 0 && Imm <= 31; }]> {
700 = !cast<AsmOperandClass>(prefix # "_asmoperand_i32");
701 let PrintMethod = "printShiftOperand<A64SE::" # form # ">";
702 let DecoderMethod = "Decode32BitShiftOperand";
705 def _asmoperand_i64 : AsmOperandClass {
706 let Name = "Shift" # form # "i64";
707 let RenderMethod = "addShiftOperands";
709 = "isShift<A64SE::" # form # ", true>";
712 def _i64 : Operand<i64>, ImmLeaf<i64, [{ return Imm >= 0 && Imm <= 63; }]> {
714 = !cast<AsmOperandClass>(prefix # "_asmoperand_i64");
715 let PrintMethod = "printShiftOperand<A64SE::" # form # ">";
719 defm lsl_operand : shift_operands<"lsl_operand", "LSL">;
720 defm lsr_operand : shift_operands<"lsr_operand", "LSR">;
721 defm asr_operand : shift_operands<"asr_operand", "ASR">;
723 // Not used for add/sub, but defined here for completeness. The "logical
724 // (shifted register)" instructions *do* have an ROR variant.
725 defm ror_operand : shift_operands<"ror_operand", "ROR">;
727 //===-------------------------------
728 // 2. The basic 3.5-operand ADD/SUB/ADDS/SUBS instructions.
729 //===-------------------------------
731 // N.b. the commutable parameter is just !N. It will be first against the wall
732 // when the revolution comes.
733 multiclass addsub_shifts<string prefix, bit sf, bit op, bit s, bit commutable,
734 string asmop, SDPatternOperator opfrag, string sty,
735 RegisterClass GPR, list<Register> defs> {
736 let isCommutable = commutable, Defs = defs in {
737 def _lsl : A64I_addsubshift<sf, op, s, 0b00,
739 (ins GPR:$Rn, GPR:$Rm,
740 !cast<Operand>("lsl_operand_" # sty):$Imm6),
741 !strconcat(asmop, "\t$Rd, $Rn, $Rm, $Imm6"),
742 [(set GPR:$Rd, (opfrag GPR:$Rn, (shl GPR:$Rm,
743 !cast<Operand>("lsl_operand_" # sty):$Imm6))
747 def _lsr : A64I_addsubshift<sf, op, s, 0b01,
749 (ins GPR:$Rn, GPR:$Rm,
750 !cast<Operand>("lsr_operand_" # sty):$Imm6),
751 !strconcat(asmop, "\t$Rd, $Rn, $Rm, $Imm6"),
752 [(set GPR:$Rd, (opfrag GPR:$Rn, (srl GPR:$Rm,
753 !cast<Operand>("lsr_operand_" # sty):$Imm6))
757 def _asr : A64I_addsubshift<sf, op, s, 0b10,
759 (ins GPR:$Rn, GPR:$Rm,
760 !cast<Operand>("asr_operand_" # sty):$Imm6),
761 !strconcat(asmop, "\t$Rd, $Rn, $Rm, $Imm6"),
762 [(set GPR:$Rd, (opfrag GPR:$Rn, (sra GPR:$Rm,
763 !cast<Operand>("asr_operand_" # sty):$Imm6))
769 : InstAlias<!strconcat(asmop, " $Rd, $Rn, $Rm"),
770 (!cast<Instruction>(prefix # "_lsl") GPR:$Rd, GPR:$Rn,
773 def : Pat<(opfrag GPR:$Rn, GPR:$Rm),
774 (!cast<Instruction>(prefix # "_lsl") GPR:$Rn, GPR:$Rm, 0)>;
777 multiclass addsub_sizes<string prefix, bit op, bit s, bit commutable,
778 string asmop, SDPatternOperator opfrag,
779 list<Register> defs> {
780 defm xxx : addsub_shifts<prefix # "xxx", 0b1, op, s,
781 commutable, asmop, opfrag, "i64", GPR64, defs>;
782 defm www : addsub_shifts<prefix # "www", 0b0, op, s,
783 commutable, asmop, opfrag, "i32", GPR32, defs>;
787 defm ADD : addsub_sizes<"ADD", 0b0, 0b0, 0b1, "add", add, []>;
788 defm SUB : addsub_sizes<"SUB", 0b1, 0b0, 0b0, "sub", sub, []>;
790 defm ADDS : addsub_sizes<"ADDS", 0b0, 0b1, 0b1, "adds", addc, [NZCV]>;
791 defm SUBS : addsub_sizes<"SUBS", 0b1, 0b1, 0b0, "subs", subc, [NZCV]>;
793 //===-------------------------------
794 // 1. The NEG/NEGS aliases
795 //===-------------------------------
797 multiclass neg_alias<Instruction INST, RegisterClass GPR,
798 Register ZR, Operand shift_operand, SDNode shiftop> {
799 def : InstAlias<"neg $Rd, $Rm, $Imm6",
800 (INST GPR:$Rd, ZR, GPR:$Rm, shift_operand:$Imm6)>;
802 def : Pat<(sub 0, (shiftop GPR:$Rm, shift_operand:$Imm6)),
803 (INST ZR, GPR:$Rm, shift_operand:$Imm6)>;
806 defm : neg_alias<SUBwww_lsl, GPR32, WZR, lsl_operand_i32, shl>;
807 defm : neg_alias<SUBwww_lsr, GPR32, WZR, lsr_operand_i32, srl>;
808 defm : neg_alias<SUBwww_asr, GPR32, WZR, asr_operand_i32, sra>;
809 def : InstAlias<"neg $Rd, $Rm", (SUBwww_lsl GPR32:$Rd, WZR, GPR32:$Rm, 0)>;
810 def : Pat<(sub 0, GPR32:$Rm), (SUBwww_lsl WZR, GPR32:$Rm, 0)>;
812 defm : neg_alias<SUBxxx_lsl, GPR64, XZR, lsl_operand_i64, shl>;
813 defm : neg_alias<SUBxxx_lsr, GPR64, XZR, lsr_operand_i64, srl>;
814 defm : neg_alias<SUBxxx_asr, GPR64, XZR, asr_operand_i64, sra>;
815 def : InstAlias<"neg $Rd, $Rm", (SUBxxx_lsl GPR64:$Rd, XZR, GPR64:$Rm, 0)>;
816 def : Pat<(sub 0, GPR64:$Rm), (SUBxxx_lsl XZR, GPR64:$Rm, 0)>;
818 // NEGS doesn't get any patterns yet: defining multiple outputs means C++ has to
820 class negs_alias<Instruction INST, RegisterClass GPR,
821 Register ZR, Operand shift_operand, SDNode shiftop>
822 : InstAlias<"negs $Rd, $Rm, $Imm6",
823 (INST GPR:$Rd, ZR, GPR:$Rm, shift_operand:$Imm6)>;
825 def : negs_alias<SUBSwww_lsl, GPR32, WZR, lsl_operand_i32, shl>;
826 def : negs_alias<SUBSwww_lsr, GPR32, WZR, lsr_operand_i32, srl>;
827 def : negs_alias<SUBSwww_asr, GPR32, WZR, asr_operand_i32, sra>;
828 def : InstAlias<"negs $Rd, $Rm", (SUBSwww_lsl GPR32:$Rd, WZR, GPR32:$Rm, 0)>;
830 def : negs_alias<SUBSxxx_lsl, GPR64, XZR, lsl_operand_i64, shl>;
831 def : negs_alias<SUBSxxx_lsr, GPR64, XZR, lsr_operand_i64, srl>;
832 def : negs_alias<SUBSxxx_asr, GPR64, XZR, asr_operand_i64, sra>;
833 def : InstAlias<"negs $Rd, $Rm", (SUBSxxx_lsl GPR64:$Rd, XZR, GPR64:$Rm, 0)>;
835 //===-------------------------------
836 // 1. The CMP/CMN aliases
837 //===-------------------------------
839 multiclass cmp_shifts<string prefix, bit sf, bit op, bit commutable,
840 string asmop, SDPatternOperator opfrag, string sty,
842 let isCommutable = commutable, Rd = 0b11111, Defs = [NZCV] in {
843 def _lsl : A64I_addsubshift<sf, op, 0b1, 0b00,
845 (ins GPR:$Rn, GPR:$Rm,
846 !cast<Operand>("lsl_operand_" # sty):$Imm6),
847 !strconcat(asmop, "\t$Rn, $Rm, $Imm6"),
848 [(set NZCV, (opfrag GPR:$Rn, (shl GPR:$Rm,
849 !cast<Operand>("lsl_operand_" # sty):$Imm6))
853 def _lsr : A64I_addsubshift<sf, op, 0b1, 0b01,
855 (ins GPR:$Rn, GPR:$Rm,
856 !cast<Operand>("lsr_operand_" # sty):$Imm6),
857 !strconcat(asmop, "\t$Rn, $Rm, $Imm6"),
858 [(set NZCV, (opfrag GPR:$Rn, (srl GPR:$Rm,
859 !cast<Operand>("lsr_operand_" # sty):$Imm6))
863 def _asr : A64I_addsubshift<sf, op, 0b1, 0b10,
865 (ins GPR:$Rn, GPR:$Rm,
866 !cast<Operand>("asr_operand_" # sty):$Imm6),
867 !strconcat(asmop, "\t$Rn, $Rm, $Imm6"),
868 [(set NZCV, (opfrag GPR:$Rn, (sra GPR:$Rm,
869 !cast<Operand>("asr_operand_" # sty):$Imm6))
875 : InstAlias<!strconcat(asmop, " $Rn, $Rm"),
876 (!cast<Instruction>(prefix # "_lsl") GPR:$Rn, GPR:$Rm, 0)>;
878 def : Pat<(opfrag GPR:$Rn, GPR:$Rm),
879 (!cast<Instruction>(prefix # "_lsl") GPR:$Rn, GPR:$Rm, 0)>;
882 defm CMPww : cmp_shifts<"CMPww", 0b0, 0b1, 0b0, "cmp", A64cmp, "i32", GPR32>;
883 defm CMPxx : cmp_shifts<"CMPxx", 0b1, 0b1, 0b0, "cmp", A64cmp, "i64", GPR64>;
885 defm CMNww : cmp_shifts<"CMNww", 0b0, 0b0, 0b1, "cmn", A64cmn, "i32", GPR32>;
886 defm CMNxx : cmp_shifts<"CMNxx", 0b1, 0b0, 0b1, "cmn", A64cmn, "i64", GPR64>;
888 //===----------------------------------------------------------------------===//
889 // Add-subtract (with carry) instructions
890 //===----------------------------------------------------------------------===//
891 // Contains: ADC, ADCS, SBC, SBCS + aliases NGC, NGCS
893 multiclass A64I_addsubcarrySizes<bit op, bit s, string asmop> {
894 let Uses = [NZCV] in {
895 def www : A64I_addsubcarry<0b0, op, s, 0b000000,
896 (outs GPR32:$Rd), (ins GPR32:$Rn, GPR32:$Rm),
897 !strconcat(asmop, "\t$Rd, $Rn, $Rm"),
900 def xxx : A64I_addsubcarry<0b1, op, s, 0b000000,
901 (outs GPR64:$Rd), (ins GPR64:$Rn, GPR64:$Rm),
902 !strconcat(asmop, "\t$Rd, $Rn, $Rm"),
907 let isCommutable = 1 in {
908 defm ADC : A64I_addsubcarrySizes<0b0, 0b0, "adc">;
911 defm SBC : A64I_addsubcarrySizes<0b1, 0b0, "sbc">;
913 let Defs = [NZCV] in {
914 let isCommutable = 1 in {
915 defm ADCS : A64I_addsubcarrySizes<0b0, 0b1, "adcs">;
918 defm SBCS : A64I_addsubcarrySizes<0b1, 0b1, "sbcs">;
921 def : InstAlias<"ngc $Rd, $Rm", (SBCwww GPR32:$Rd, WZR, GPR32:$Rm)>;
922 def : InstAlias<"ngc $Rd, $Rm", (SBCxxx GPR64:$Rd, XZR, GPR64:$Rm)>;
923 def : InstAlias<"ngcs $Rd, $Rm", (SBCSwww GPR32:$Rd, WZR, GPR32:$Rm)>;
924 def : InstAlias<"ngcs $Rd, $Rm", (SBCSxxx GPR64:$Rd, XZR, GPR64:$Rm)>;
926 // Note that adde and sube can form a chain longer than two (e.g. for 256-bit
927 // addition). So the flag-setting instructions are appropriate.
928 def : Pat<(adde GPR32:$Rn, GPR32:$Rm), (ADCSwww GPR32:$Rn, GPR32:$Rm)>;
929 def : Pat<(adde GPR64:$Rn, GPR64:$Rm), (ADCSxxx GPR64:$Rn, GPR64:$Rm)>;
930 def : Pat<(sube GPR32:$Rn, GPR32:$Rm), (SBCSwww GPR32:$Rn, GPR32:$Rm)>;
931 def : Pat<(sube GPR64:$Rn, GPR64:$Rm), (SBCSxxx GPR64:$Rn, GPR64:$Rm)>;
933 //===----------------------------------------------------------------------===//
935 //===----------------------------------------------------------------------===//
936 // Contains: SBFM, BFM, UBFM, [SU]XT[BHW], ASR, LSR, LSL, SBFI[ZX], BFI, BFXIL,
939 // Because of the rather complicated nearly-overlapping aliases, the decoding of
940 // this range of instructions is handled manually. The architectural
941 // instructions are BFM, SBFM and UBFM but a disassembler should never produce
944 // In the end, the best option was to use BFM instructions for decoding under
945 // almost all circumstances, but to create aliasing *Instructions* for each of
946 // the canonical forms and specify a completely custom decoder which would
947 // substitute the correct MCInst as needed.
949 // This also simplifies instruction selection, parsing etc because the MCInsts
950 // have a shape that's closer to their use in code.
952 //===-------------------------------
953 // 1. The architectural BFM instructions
954 //===-------------------------------
956 def uimm5_asmoperand : AsmOperandClass {
958 let PredicateMethod = "isUImm<5>";
959 let RenderMethod = "addImmOperands";
962 def uimm6_asmoperand : AsmOperandClass {
964 let PredicateMethod = "isUImm<6>";
965 let RenderMethod = "addImmOperands";
968 def bitfield32_imm : Operand<i64>,
969 ImmLeaf<i64, [{ return Imm >= 0 && Imm < 32; }]> {
970 let ParserMatchClass = uimm5_asmoperand;
972 let DecoderMethod = "DecodeBitfield32ImmOperand";
976 def bitfield64_imm : Operand<i64>,
977 ImmLeaf<i64, [{ return Imm >= 0 && Imm < 64; }]> {
978 let ParserMatchClass = uimm6_asmoperand;
980 // Default decoder works in 64-bit case: the 6-bit field can take any value.
983 multiclass A64I_bitfieldSizes<bits<2> opc, string asmop> {
984 def wwii : A64I_bitfield<0b0, opc, 0b0, (outs GPR32:$Rd),
985 (ins GPR32:$Rn, bitfield32_imm:$ImmR, bitfield32_imm:$ImmS),
986 !strconcat(asmop, "\t$Rd, $Rn, $ImmR, $ImmS"),
988 let DecoderMethod = "DecodeBitfieldInstruction";
991 def xxii : A64I_bitfield<0b1, opc, 0b1, (outs GPR64:$Rd),
992 (ins GPR64:$Rn, bitfield64_imm:$ImmR, bitfield64_imm:$ImmS),
993 !strconcat(asmop, "\t$Rd, $Rn, $ImmR, $ImmS"),
995 let DecoderMethod = "DecodeBitfieldInstruction";
999 defm SBFM : A64I_bitfieldSizes<0b00, "sbfm">;
1000 defm UBFM : A64I_bitfieldSizes<0b10, "ubfm">;
1002 // BFM instructions modify the destination register rather than defining it
1005 A64I_bitfield<0b0, 0b01, 0b0, (outs GPR32:$Rd),
1006 (ins GPR32:$src, GPR32:$Rn, bitfield32_imm:$ImmR, bitfield32_imm:$ImmS),
1007 "bfm\t$Rd, $Rn, $ImmR, $ImmS", [], NoItinerary> {
1008 let DecoderMethod = "DecodeBitfieldInstruction";
1009 let Constraints = "$src = $Rd";
1013 A64I_bitfield<0b1, 0b01, 0b1, (outs GPR64:$Rd),
1014 (ins GPR64:$src, GPR64:$Rn, bitfield64_imm:$ImmR, bitfield64_imm:$ImmS),
1015 "bfm\t$Rd, $Rn, $ImmR, $ImmS", [], NoItinerary> {
1016 let DecoderMethod = "DecodeBitfieldInstruction";
1017 let Constraints = "$src = $Rd";
1021 //===-------------------------------
1022 // 2. Extend aliases to 64-bit dest
1023 //===-------------------------------
1025 // Unfortunately the extensions that end up as 64-bits cannot be handled by an
1026 // instruction alias: their syntax is (for example) "SXTB x0, w0", which needs
1027 // to be mapped to "SBFM x0, x0, #0, 7" (changing the class of Rn). InstAlias is
1028 // not capable of such a map as far as I'm aware
1030 // Note that these instructions are strictly more specific than the
1031 // BFM ones (in ImmR) so they can handle their own decoding.
1032 class A64I_bf_ext<bit sf, bits<2> opc, RegisterClass GPRDest, string asmop,
1033 bits<6> imms, dag pattern>
1034 : A64I_bitfield<sf, opc, sf,
1035 (outs GPRDest:$Rd), (ins GPR32:$Rn),
1036 !strconcat(asmop, "\t$Rd, $Rn"),
1037 [(set GPRDest:$Rd, pattern)], NoItinerary> {
1038 let ImmR = 0b000000;
1042 // Signed extensions
1043 def SXTBxw : A64I_bf_ext<0b1, 0b00, GPR64, "sxtb", 7,
1044 (sext_inreg (anyext GPR32:$Rn), i8)>;
1045 def SXTBww : A64I_bf_ext<0b0, 0b00, GPR32, "sxtb", 7,
1046 (sext_inreg GPR32:$Rn, i8)>;
1047 def SXTHxw : A64I_bf_ext<0b1, 0b00, GPR64, "sxth", 15,
1048 (sext_inreg (anyext GPR32:$Rn), i16)>;
1049 def SXTHww : A64I_bf_ext<0b0, 0b00, GPR32, "sxth", 15,
1050 (sext_inreg GPR32:$Rn, i16)>;
1051 def SXTWxw : A64I_bf_ext<0b1, 0b00, GPR64, "sxtw", 31, (sext GPR32:$Rn)>;
1053 // Unsigned extensions
1054 def UXTBww : A64I_bf_ext<0b0, 0b10, GPR32, "uxtb", 7,
1055 (and GPR32:$Rn, 255)>;
1056 def UXTHww : A64I_bf_ext<0b0, 0b10, GPR32, "uxth", 15,
1057 (and GPR32:$Rn, 65535)>;
1059 // The 64-bit unsigned variants are not strictly architectural but recommended
1061 let isAsmParserOnly = 1 in {
1062 def UXTBxw : A64I_bf_ext<0b0, 0b10, GPR64, "uxtb", 7,
1063 (and (anyext GPR32:$Rn), 255)>;
1064 def UXTHxw : A64I_bf_ext<0b0, 0b10, GPR64, "uxth", 15,
1065 (and (anyext GPR32:$Rn), 65535)>;
1068 // Extra patterns for when the source register is actually 64-bits
1069 // too. There's no architectural difference here, it's just LLVM
1070 // shinanigans. There's no need for equivalent zero-extension patterns
1071 // because they'll already be caught by logical (immediate) matching.
1072 def : Pat<(sext_inreg GPR64:$Rn, i8),
1073 (SXTBxw (EXTRACT_SUBREG GPR64:$Rn, sub_32))>;
1074 def : Pat<(sext_inreg GPR64:$Rn, i16),
1075 (SXTHxw (EXTRACT_SUBREG GPR64:$Rn, sub_32))>;
1076 def : Pat<(sext_inreg GPR64:$Rn, i32),
1077 (SXTWxw (EXTRACT_SUBREG GPR64:$Rn, sub_32))>;
1080 //===-------------------------------
1081 // 3. Aliases for ASR and LSR (the simple shifts)
1082 //===-------------------------------
1084 // These also handle their own decoding because ImmS being set makes
1085 // them take precedence over BFM.
1086 multiclass A64I_shift<bits<2> opc, string asmop, SDNode opnode> {
1087 def wwi : A64I_bitfield<0b0, opc, 0b0,
1088 (outs GPR32:$Rd), (ins GPR32:$Rn, bitfield32_imm:$ImmR),
1089 !strconcat(asmop, "\t$Rd, $Rn, $ImmR"),
1090 [(set GPR32:$Rd, (opnode GPR32:$Rn, bitfield32_imm:$ImmR))],
1095 def xxi : A64I_bitfield<0b1, opc, 0b1,
1096 (outs GPR64:$Rd), (ins GPR64:$Rn, bitfield64_imm:$ImmR),
1097 !strconcat(asmop, "\t$Rd, $Rn, $ImmR"),
1098 [(set GPR64:$Rd, (opnode GPR64:$Rn, bitfield64_imm:$ImmR))],
1105 defm ASR : A64I_shift<0b00, "asr", sra>;
1106 defm LSR : A64I_shift<0b10, "lsr", srl>;
1108 //===-------------------------------
1109 // 4. Aliases for LSL
1110 //===-------------------------------
1112 // Unfortunately LSL and subsequent aliases are much more complicated. We need
1113 // to be able to say certain output instruction fields depend in a complex
1114 // manner on combinations of input assembly fields).
1116 // MIOperandInfo *might* have been able to do it, but at the cost of
1117 // significantly more C++ code.
1119 // N.b. contrary to usual practice these operands store the shift rather than
1120 // the machine bits in an MCInst. The complexity overhead of consistency
1121 // outweighed the benefits in this case (custom asmparser, printer and selection
1122 // vs custom encoder).
1123 def bitfield32_lsl_imm : Operand<i64>,
1124 ImmLeaf<i64, [{ return Imm >= 0 && Imm <= 31; }]> {
1125 let ParserMatchClass = uimm5_asmoperand;
1126 let EncoderMethod = "getBitfield32LSLOpValue";
1129 def bitfield64_lsl_imm : Operand<i64>,
1130 ImmLeaf<i64, [{ return Imm >= 0 && Imm <= 63; }]> {
1131 let ParserMatchClass = uimm6_asmoperand;
1132 let EncoderMethod = "getBitfield64LSLOpValue";
1135 class A64I_bitfield_lsl<bit sf, RegisterClass GPR, Operand operand>
1136 : A64I_bitfield<sf, 0b10, sf, (outs GPR:$Rd), (ins GPR:$Rn, operand:$FullImm),
1137 "lsl\t$Rd, $Rn, $FullImm",
1138 [(set GPR:$Rd, (shl GPR:$Rn, operand:$FullImm))],
1141 let ImmR = FullImm{5-0};
1142 let ImmS = FullImm{11-6};
1144 // No disassembler allowed because it would overlap with BFM which does the
1146 let isAsmParserOnly = 1;
1149 def LSLwwi : A64I_bitfield_lsl<0b0, GPR32, bitfield32_lsl_imm>;
1150 def LSLxxi : A64I_bitfield_lsl<0b1, GPR64, bitfield64_lsl_imm>;
1152 //===-------------------------------
1153 // 5. Aliases for bitfield extract instructions
1154 //===-------------------------------
1156 def bfx32_width_asmoperand : AsmOperandClass {
1157 let Name = "BFX32Width";
1158 let PredicateMethod = "isBitfieldWidth<32>";
1159 let RenderMethod = "addBFXWidthOperands";
1162 def bfx32_width : Operand<i64>, ImmLeaf<i64, [{ return true; }]> {
1163 let PrintMethod = "printBFXWidthOperand";
1164 let ParserMatchClass = bfx32_width_asmoperand;
1167 def bfx64_width_asmoperand : AsmOperandClass {
1168 let Name = "BFX64Width";
1169 let PredicateMethod = "isBitfieldWidth<64>";
1170 let RenderMethod = "addBFXWidthOperands";
1173 def bfx64_width : Operand<i64> {
1174 let PrintMethod = "printBFXWidthOperand";
1175 let ParserMatchClass = bfx64_width_asmoperand;
1179 multiclass A64I_bitfield_extract<bits<2> opc, string asmop, SDNode op> {
1180 def wwii : A64I_bitfield<0b0, opc, 0b0, (outs GPR32:$Rd),
1181 (ins GPR32:$Rn, bitfield32_imm:$ImmR, bfx32_width:$ImmS),
1182 !strconcat(asmop, "\t$Rd, $Rn, $ImmR, $ImmS"),
1183 [(set GPR32:$Rd, (op GPR32:$Rn, imm:$ImmR, imm:$ImmS))],
1185 // As above, no disassembler allowed.
1186 let isAsmParserOnly = 1;
1189 def xxii : A64I_bitfield<0b1, opc, 0b1, (outs GPR64:$Rd),
1190 (ins GPR64:$Rn, bitfield64_imm:$ImmR, bfx64_width:$ImmS),
1191 !strconcat(asmop, "\t$Rd, $Rn, $ImmR, $ImmS"),
1192 [(set GPR64:$Rd, (op GPR64:$Rn, imm:$ImmR, imm:$ImmS))],
1194 // As above, no disassembler allowed.
1195 let isAsmParserOnly = 1;
1199 defm SBFX : A64I_bitfield_extract<0b00, "sbfx", A64Sbfx>;
1200 defm UBFX : A64I_bitfield_extract<0b10, "ubfx", A64Ubfx>;
1202 // Again, variants based on BFM modify Rd so need it as an input too.
1203 def BFXILwwii : A64I_bitfield<0b0, 0b01, 0b0, (outs GPR32:$Rd),
1204 (ins GPR32:$src, GPR32:$Rn, bitfield32_imm:$ImmR, bfx32_width:$ImmS),
1205 "bfxil\t$Rd, $Rn, $ImmR, $ImmS", [], NoItinerary> {
1206 // As above, no disassembler allowed.
1207 let isAsmParserOnly = 1;
1208 let Constraints = "$src = $Rd";
1211 def BFXILxxii : A64I_bitfield<0b1, 0b01, 0b1, (outs GPR64:$Rd),
1212 (ins GPR64:$src, GPR64:$Rn, bitfield64_imm:$ImmR, bfx64_width:$ImmS),
1213 "bfxil\t$Rd, $Rn, $ImmR, $ImmS", [], NoItinerary> {
1214 // As above, no disassembler allowed.
1215 let isAsmParserOnly = 1;
1216 let Constraints = "$src = $Rd";
1219 // SBFX instructions can do a 1-instruction sign-extension of boolean values.
1220 def : Pat<(sext_inreg GPR64:$Rn, i1), (SBFXxxii GPR64:$Rn, 0, 0)>;
1221 def : Pat<(sext_inreg GPR32:$Rn, i1), (SBFXwwii GPR32:$Rn, 0, 0)>;
1222 def : Pat<(i64 (sext_inreg (anyext GPR32:$Rn), i1)),
1223 (SBFXxxii (SUBREG_TO_REG (i64 0), GPR32:$Rn, sub_32), 0, 0)>;
1225 // UBFX makes sense as an implementation of a 64-bit zero-extension too. Could
1226 // use either 64-bit or 32-bit variant, but 32-bit might be more efficient.
1227 def : Pat<(zext GPR32:$Rn), (SUBREG_TO_REG (i64 0), (UBFXwwii GPR32:$Rn, 0, 31),
1230 //===-------------------------------
1231 // 6. Aliases for bitfield insert instructions
1232 //===-------------------------------
1234 def bfi32_lsb_asmoperand : AsmOperandClass {
1235 let Name = "BFI32LSB";
1236 let PredicateMethod = "isUImm<5>";
1237 let RenderMethod = "addBFILSBOperands<32>";
1240 def bfi32_lsb : Operand<i64>,
1241 ImmLeaf<i64, [{ return Imm >= 0 && Imm <= 31; }]> {
1242 let PrintMethod = "printBFILSBOperand<32>";
1243 let ParserMatchClass = bfi32_lsb_asmoperand;
1246 def bfi64_lsb_asmoperand : AsmOperandClass {
1247 let Name = "BFI64LSB";
1248 let PredicateMethod = "isUImm<6>";
1249 let RenderMethod = "addBFILSBOperands<64>";
1252 def bfi64_lsb : Operand<i64>,
1253 ImmLeaf<i64, [{ return Imm >= 0 && Imm <= 63; }]> {
1254 let PrintMethod = "printBFILSBOperand<64>";
1255 let ParserMatchClass = bfi64_lsb_asmoperand;
1258 // Width verification is performed during conversion so width operand can be
1259 // shared between 32/64-bit cases. Still needed for the print method though
1260 // because ImmR encodes "width - 1".
1261 def bfi32_width_asmoperand : AsmOperandClass {
1262 let Name = "BFI32Width";
1263 let PredicateMethod = "isBitfieldWidth<32>";
1264 let RenderMethod = "addBFIWidthOperands";
1267 def bfi32_width : Operand<i64>,
1268 ImmLeaf<i64, [{ return Imm >= 1 && Imm <= 32; }]> {
1269 let PrintMethod = "printBFIWidthOperand";
1270 let ParserMatchClass = bfi32_width_asmoperand;
1273 def bfi64_width_asmoperand : AsmOperandClass {
1274 let Name = "BFI64Width";
1275 let PredicateMethod = "isBitfieldWidth<64>";
1276 let RenderMethod = "addBFIWidthOperands";
1279 def bfi64_width : Operand<i64>,
1280 ImmLeaf<i64, [{ return Imm >= 1 && Imm <= 64; }]> {
1281 let PrintMethod = "printBFIWidthOperand";
1282 let ParserMatchClass = bfi64_width_asmoperand;
1285 multiclass A64I_bitfield_insert<bits<2> opc, string asmop> {
1286 def wwii : A64I_bitfield<0b0, opc, 0b0, (outs GPR32:$Rd),
1287 (ins GPR32:$Rn, bfi32_lsb:$ImmR, bfi32_width:$ImmS),
1288 !strconcat(asmop, "\t$Rd, $Rn, $ImmR, $ImmS"),
1290 // As above, no disassembler allowed.
1291 let isAsmParserOnly = 1;
1294 def xxii : A64I_bitfield<0b1, opc, 0b1, (outs GPR64:$Rd),
1295 (ins GPR64:$Rn, bfi64_lsb:$ImmR, bfi64_width:$ImmS),
1296 !strconcat(asmop, "\t$Rd, $Rn, $ImmR, $ImmS"),
1298 // As above, no disassembler allowed.
1299 let isAsmParserOnly = 1;
1303 defm SBFIZ : A64I_bitfield_insert<0b00, "sbfiz">;
1304 defm UBFIZ : A64I_bitfield_insert<0b10, "ubfiz">;
1307 def BFIwwii : A64I_bitfield<0b0, 0b01, 0b0, (outs GPR32:$Rd),
1308 (ins GPR32:$src, GPR32:$Rn, bfi32_lsb:$ImmR, bfi32_width:$ImmS),
1309 "bfi\t$Rd, $Rn, $ImmR, $ImmS", [], NoItinerary> {
1310 // As above, no disassembler allowed.
1311 let isAsmParserOnly = 1;
1312 let Constraints = "$src = $Rd";
1315 def BFIxxii : A64I_bitfield<0b1, 0b01, 0b1, (outs GPR64:$Rd),
1316 (ins GPR64:$src, GPR64:$Rn, bfi64_lsb:$ImmR, bfi64_width:$ImmS),
1317 "bfi\t$Rd, $Rn, $ImmR, $ImmS", [], NoItinerary> {
1318 // As above, no disassembler allowed.
1319 let isAsmParserOnly = 1;
1320 let Constraints = "$src = $Rd";
1323 //===----------------------------------------------------------------------===//
1324 // Compare and branch (immediate)
1325 //===----------------------------------------------------------------------===//
1326 // Contains: CBZ, CBNZ
1328 class label_asmoperand<int width, int scale> : AsmOperandClass {
1329 let Name = "Label" # width # "_" # scale;
1330 let PredicateMethod = "isLabel<" # width # "," # scale # ">";
1331 let RenderMethod = "addLabelOperands<" # width # ", " # scale # ">";
1334 def label_wid19_scal4_asmoperand : label_asmoperand<19, 4>;
1336 // All conditional immediate branches are the same really: 19 signed bits scaled
1337 // by the instruction-size (4).
1338 def bcc_target : Operand<OtherVT> {
1339 // This label is a 19-bit offset from PC, scaled by the instruction-width: 4.
1340 let ParserMatchClass = label_wid19_scal4_asmoperand;
1341 let PrintMethod = "printLabelOperand<19, 4>";
1342 let EncoderMethod = "getLabelOpValue<AArch64::fixup_a64_condbr>";
1343 let OperandType = "OPERAND_PCREL";
1346 multiclass cmpbr_sizes<bit op, string asmop, ImmLeaf SETOP> {
1347 let isBranch = 1, isTerminator = 1 in {
1348 def x : A64I_cmpbr<0b1, op,
1350 (ins GPR64:$Rt, bcc_target:$Label),
1351 !strconcat(asmop,"\t$Rt, $Label"),
1352 [(A64br_cc (A64cmp GPR64:$Rt, 0), SETOP, bb:$Label)],
1355 def w : A64I_cmpbr<0b0, op,
1357 (ins GPR32:$Rt, bcc_target:$Label),
1358 !strconcat(asmop,"\t$Rt, $Label"),
1359 [(A64br_cc (A64cmp GPR32:$Rt, 0), SETOP, bb:$Label)],
1364 defm CBZ : cmpbr_sizes<0b0, "cbz", ImmLeaf<i32, [{
1365 return Imm == A64CC::EQ;
1367 defm CBNZ : cmpbr_sizes<0b1, "cbnz", ImmLeaf<i32, [{
1368 return Imm == A64CC::NE;
1371 //===----------------------------------------------------------------------===//
1372 // Conditional branch (immediate) instructions
1373 //===----------------------------------------------------------------------===//
1376 def cond_code_asmoperand : AsmOperandClass {
1377 let Name = "CondCode";
1380 def cond_code : Operand<i32>, ImmLeaf<i32, [{
1381 return Imm >= 0 && Imm <= 15;
1383 let PrintMethod = "printCondCodeOperand";
1384 let ParserMatchClass = cond_code_asmoperand;
1387 def Bcc : A64I_condbr<0b0, 0b0, (outs),
1388 (ins cond_code:$Cond, bcc_target:$Label),
1389 "b.$Cond $Label", [(A64br_cc NZCV, (i32 imm:$Cond), bb:$Label)],
1393 let isTerminator = 1;
1396 //===----------------------------------------------------------------------===//
1397 // Conditional compare (immediate) instructions
1398 //===----------------------------------------------------------------------===//
1399 // Contains: CCMN, CCMP
1401 def uimm4_asmoperand : AsmOperandClass {
1403 let PredicateMethod = "isUImm<4>";
1404 let RenderMethod = "addImmOperands";
1407 def uimm4 : Operand<i32> {
1408 let ParserMatchClass = uimm4_asmoperand;
1411 def uimm5 : Operand<i32> {
1412 let ParserMatchClass = uimm5_asmoperand;
1415 // The only difference between this operand and the one for instructions like
1416 // B.cc is that it's parsed manually. The other get parsed implicitly as part of
1417 // the mnemonic handling.
1418 def cond_code_op_asmoperand : AsmOperandClass {
1419 let Name = "CondCodeOp";
1420 let RenderMethod = "addCondCodeOperands";
1421 let PredicateMethod = "isCondCode";
1422 let ParserMethod = "ParseCondCodeOperand";
1425 def cond_code_op : Operand<i32> {
1426 let PrintMethod = "printCondCodeOperand";
1427 let ParserMatchClass = cond_code_op_asmoperand;
1430 class A64I_condcmpimmImpl<bit sf, bit op, RegisterClass GPR, string asmop>
1431 : A64I_condcmpimm<sf, op, 0b0, 0b0, 0b1, (outs),
1432 (ins GPR:$Rn, uimm5:$UImm5, uimm4:$NZCVImm, cond_code_op:$Cond),
1433 !strconcat(asmop, "\t$Rn, $UImm5, $NZCVImm, $Cond"),
1438 def CCMNwi : A64I_condcmpimmImpl<0b0, 0b0, GPR32, "ccmn">;
1439 def CCMNxi : A64I_condcmpimmImpl<0b1, 0b0, GPR64, "ccmn">;
1440 def CCMPwi : A64I_condcmpimmImpl<0b0, 0b1, GPR32, "ccmp">;
1441 def CCMPxi : A64I_condcmpimmImpl<0b1, 0b1, GPR64, "ccmp">;
1443 //===----------------------------------------------------------------------===//
1444 // Conditional compare (register) instructions
1445 //===----------------------------------------------------------------------===//
1446 // Contains: CCMN, CCMP
1448 class A64I_condcmpregImpl<bit sf, bit op, RegisterClass GPR, string asmop>
1449 : A64I_condcmpreg<sf, op, 0b0, 0b0, 0b1,
1451 (ins GPR:$Rn, GPR:$Rm, uimm4:$NZCVImm, cond_code_op:$Cond),
1452 !strconcat(asmop, "\t$Rn, $Rm, $NZCVImm, $Cond"),
1457 def CCMNww : A64I_condcmpregImpl<0b0, 0b0, GPR32, "ccmn">;
1458 def CCMNxx : A64I_condcmpregImpl<0b1, 0b0, GPR64, "ccmn">;
1459 def CCMPww : A64I_condcmpregImpl<0b0, 0b1, GPR32, "ccmp">;
1460 def CCMPxx : A64I_condcmpregImpl<0b1, 0b1, GPR64, "ccmp">;
1462 //===----------------------------------------------------------------------===//
1463 // Conditional select instructions
1464 //===----------------------------------------------------------------------===//
1465 // Contains: CSEL, CSINC, CSINV, CSNEG + aliases CSET, CSETM, CINC, CINV, CNEG
1467 // Condition code which is encoded as the inversion (semantically rather than
1468 // bitwise) in the instruction.
1469 def inv_cond_code_op_asmoperand : AsmOperandClass {
1470 let Name = "InvCondCodeOp";
1471 let RenderMethod = "addInvCondCodeOperands";
1472 let PredicateMethod = "isCondCode";
1473 let ParserMethod = "ParseCondCodeOperand";
1476 def inv_cond_code_op : Operand<i32> {
1477 let ParserMatchClass = inv_cond_code_op_asmoperand;
1480 // Having a separate operand for the selectable use-case is debatable, but gives
1481 // consistency with cond_code.
1482 def inv_cond_XFORM : SDNodeXForm<imm, [{
1483 A64CC::CondCodes CC = static_cast<A64CC::CondCodes>(N->getZExtValue());
1484 return CurDAG->getTargetConstant(A64InvertCondCode(CC), MVT::i32);
1488 : ImmLeaf<i32, [{ return Imm >= 0 && Imm <= 15; }], inv_cond_XFORM>;
1491 multiclass A64I_condselSizes<bit op, bits<2> op2, string asmop,
1492 SDPatternOperator select> {
1493 let Uses = [NZCV] in {
1494 def wwwc : A64I_condsel<0b0, op, 0b0, op2,
1496 (ins GPR32:$Rn, GPR32:$Rm, cond_code_op:$Cond),
1497 !strconcat(asmop, "\t$Rd, $Rn, $Rm, $Cond"),
1498 [(set GPR32:$Rd, (select GPR32:$Rn, GPR32:$Rm))],
1502 def xxxc : A64I_condsel<0b1, op, 0b0, op2,
1504 (ins GPR64:$Rn, GPR64:$Rm, cond_code_op:$Cond),
1505 !strconcat(asmop, "\t$Rd, $Rn, $Rm, $Cond"),
1506 [(set GPR64:$Rd, (select GPR64:$Rn, GPR64:$Rm))],
1512 : PatFrag<(ops node:$lhs, node:$rhs),
1513 (A64select_cc NZCV, node:$lhs, node:$rhs, (i32 imm:$Cond))>;
1515 class complex_select<SDPatternOperator opnode>
1516 : PatFrag<(ops node:$lhs, node:$rhs),
1517 (A64select_cc NZCV, node:$lhs, (opnode node:$rhs), (i32 imm:$Cond))>;
1520 defm CSEL : A64I_condselSizes<0b0, 0b00, "csel", simple_select>;
1521 defm CSINC : A64I_condselSizes<0b0, 0b01, "csinc",
1522 complex_select<PatFrag<(ops node:$val),
1523 (add node:$val, 1)>>>;
1524 defm CSINV : A64I_condselSizes<0b1, 0b00, "csinv", complex_select<not>>;
1525 defm CSNEG : A64I_condselSizes<0b1, 0b01, "csneg", complex_select<ineg>>;
1527 // Now the instruction aliases, which fit nicely into LLVM's model:
1529 def : InstAlias<"cset $Rd, $Cond",
1530 (CSINCwwwc GPR32:$Rd, WZR, WZR, inv_cond_code_op:$Cond)>;
1531 def : InstAlias<"cset $Rd, $Cond",
1532 (CSINCxxxc GPR64:$Rd, XZR, XZR, inv_cond_code_op:$Cond)>;
1533 def : InstAlias<"csetm $Rd, $Cond",
1534 (CSINVwwwc GPR32:$Rd, WZR, WZR, inv_cond_code_op:$Cond)>;
1535 def : InstAlias<"csetm $Rd, $Cond",
1536 (CSINVxxxc GPR64:$Rd, XZR, XZR, inv_cond_code_op:$Cond)>;
1537 def : InstAlias<"cinc $Rd, $Rn, $Cond",
1538 (CSINCwwwc GPR32:$Rd, GPR32:$Rn, GPR32:$Rn, inv_cond_code_op:$Cond)>;
1539 def : InstAlias<"cinc $Rd, $Rn, $Cond",
1540 (CSINCxxxc GPR64:$Rd, GPR64:$Rn, GPR64:$Rn, inv_cond_code_op:$Cond)>;
1541 def : InstAlias<"cinv $Rd, $Rn, $Cond",
1542 (CSINVwwwc GPR32:$Rd, GPR32:$Rn, GPR32:$Rn, inv_cond_code_op:$Cond)>;
1543 def : InstAlias<"cinv $Rd, $Rn, $Cond",
1544 (CSINVxxxc GPR64:$Rd, GPR64:$Rn, GPR64:$Rn, inv_cond_code_op:$Cond)>;
1545 def : InstAlias<"cneg $Rd, $Rn, $Cond",
1546 (CSNEGwwwc GPR32:$Rd, GPR32:$Rn, GPR32:$Rn, inv_cond_code_op:$Cond)>;
1547 def : InstAlias<"cneg $Rd, $Rn, $Cond",
1548 (CSNEGxxxc GPR64:$Rd, GPR64:$Rn, GPR64:$Rn, inv_cond_code_op:$Cond)>;
1550 // Finally some helper patterns.
1552 // For CSET (a.k.a. zero-extension of icmp)
1553 def : Pat<(A64select_cc NZCV, 0, 1, cond_code:$Cond),
1554 (CSINCwwwc WZR, WZR, cond_code:$Cond)>;
1555 def : Pat<(A64select_cc NZCV, 1, 0, inv_cond_code:$Cond),
1556 (CSINCwwwc WZR, WZR, inv_cond_code:$Cond)>;
1558 def : Pat<(A64select_cc NZCV, 0, 1, cond_code:$Cond),
1559 (CSINCxxxc XZR, XZR, cond_code:$Cond)>;
1560 def : Pat<(A64select_cc NZCV, 1, 0, inv_cond_code:$Cond),
1561 (CSINCxxxc XZR, XZR, inv_cond_code:$Cond)>;
1563 // For CSETM (a.k.a. sign-extension of icmp)
1564 def : Pat<(A64select_cc NZCV, 0, -1, cond_code:$Cond),
1565 (CSINVwwwc WZR, WZR, cond_code:$Cond)>;
1566 def : Pat<(A64select_cc NZCV, -1, 0, inv_cond_code:$Cond),
1567 (CSINVwwwc WZR, WZR, inv_cond_code:$Cond)>;
1569 def : Pat<(A64select_cc NZCV, 0, -1, cond_code:$Cond),
1570 (CSINVxxxc XZR, XZR, cond_code:$Cond)>;
1571 def : Pat<(A64select_cc NZCV, -1, 0, inv_cond_code:$Cond),
1572 (CSINVxxxc XZR, XZR, inv_cond_code:$Cond)>;
1574 // CINC, CINV and CNEG get dealt with automatically, which leaves the issue of
1575 // commutativity. The instructions are to complex for isCommutable to be used,
1576 // so we have to create the patterns manually:
1578 // No commutable pattern for CSEL since the commuted version is isomorphic.
1581 def :Pat<(A64select_cc NZCV, (add GPR32:$Rm, 1), GPR32:$Rn,
1582 inv_cond_code:$Cond),
1583 (CSINCwwwc GPR32:$Rn, GPR32:$Rm, inv_cond_code:$Cond)>;
1584 def :Pat<(A64select_cc NZCV, (add GPR64:$Rm, 1), GPR64:$Rn,
1585 inv_cond_code:$Cond),
1586 (CSINCxxxc GPR64:$Rn, GPR64:$Rm, inv_cond_code:$Cond)>;
1589 def :Pat<(A64select_cc NZCV, (not GPR32:$Rm), GPR32:$Rn, inv_cond_code:$Cond),
1590 (CSINVwwwc GPR32:$Rn, GPR32:$Rm, inv_cond_code:$Cond)>;
1591 def :Pat<(A64select_cc NZCV, (not GPR64:$Rm), GPR64:$Rn, inv_cond_code:$Cond),
1592 (CSINVxxxc GPR64:$Rn, GPR64:$Rm, inv_cond_code:$Cond)>;
1595 def :Pat<(A64select_cc NZCV, (ineg GPR32:$Rm), GPR32:$Rn, inv_cond_code:$Cond),
1596 (CSNEGwwwc GPR32:$Rn, GPR32:$Rm, inv_cond_code:$Cond)>;
1597 def :Pat<(A64select_cc NZCV, (ineg GPR64:$Rm), GPR64:$Rn, inv_cond_code:$Cond),
1598 (CSNEGxxxc GPR64:$Rn, GPR64:$Rm, inv_cond_code:$Cond)>;
1600 //===----------------------------------------------------------------------===//
1601 // Data Processing (1 source) instructions
1602 //===----------------------------------------------------------------------===//
1603 // Contains: RBIT, REV16, REV, REV32, CLZ, CLS.
1605 // We define an unary operator which always fails. We will use this to
1606 // define unary operators that cannot be matched.
1608 class A64I_dp_1src_impl<bit sf, bits<6> opcode, string asmop,
1609 list<dag> patterns, RegisterClass GPRrc,
1610 InstrItinClass itin>:
1615 !strconcat(asmop, "\t$Rd, $Rn"),
1621 multiclass A64I_dp_1src <bits<6> opcode, string asmop> {
1622 let neverHasSideEffects = 1 in {
1623 def ww : A64I_dp_1src_impl<0b0, opcode, asmop, [], GPR32, NoItinerary>;
1624 def xx : A64I_dp_1src_impl<0b1, opcode, asmop, [], GPR64, NoItinerary>;
1628 defm RBIT : A64I_dp_1src<0b000000, "rbit">;
1629 defm CLS : A64I_dp_1src<0b000101, "cls">;
1630 defm CLZ : A64I_dp_1src<0b000100, "clz">;
1632 def : Pat<(ctlz GPR32:$Rn), (CLZww GPR32:$Rn)>;
1633 def : Pat<(ctlz GPR64:$Rn), (CLZxx GPR64:$Rn)>;
1634 def : Pat<(ctlz_zero_undef GPR32:$Rn), (CLZww GPR32:$Rn)>;
1635 def : Pat<(ctlz_zero_undef GPR64:$Rn), (CLZxx GPR64:$Rn)>;
1637 def : Pat<(cttz GPR32:$Rn), (CLZww (RBITww GPR32:$Rn))>;
1638 def : Pat<(cttz GPR64:$Rn), (CLZxx (RBITxx GPR64:$Rn))>;
1639 def : Pat<(cttz_zero_undef GPR32:$Rn), (CLZww (RBITww GPR32:$Rn))>;
1640 def : Pat<(cttz_zero_undef GPR64:$Rn), (CLZxx (RBITxx GPR64:$Rn))>;
1643 def REVww : A64I_dp_1src_impl<0b0, 0b000010, "rev",
1644 [(set GPR32:$Rd, (bswap GPR32:$Rn))],
1645 GPR32, NoItinerary>;
1646 def REVxx : A64I_dp_1src_impl<0b1, 0b000011, "rev",
1647 [(set GPR64:$Rd, (bswap GPR64:$Rn))],
1648 GPR64, NoItinerary>;
1649 def REV32xx : A64I_dp_1src_impl<0b1, 0b000010, "rev32",
1650 [(set GPR64:$Rd, (bswap (rotr GPR64:$Rn, (i64 32))))],
1651 GPR64, NoItinerary>;
1652 def REV16ww : A64I_dp_1src_impl<0b0, 0b000001, "rev16",
1653 [(set GPR32:$Rd, (bswap (rotr GPR32:$Rn, (i64 16))))],
1656 def REV16xx : A64I_dp_1src_impl<0b1, 0b000001, "rev16", [], GPR64, NoItinerary>;
1658 //===----------------------------------------------------------------------===//
1659 // Data Processing (2 sources) instructions
1660 //===----------------------------------------------------------------------===//
1661 // Contains: UDIV, SDIV, LSLV, LSRV, ASRV, RORV + aliases LSL, LSR, ASR, ROR
1663 class dp_2src_impl<bit sf, bits<6> opcode, string asmop, list<dag> patterns,
1664 RegisterClass GPRsp,
1665 InstrItinClass itin>:
1669 !strconcat(asmop, "\t$Rd, $Rn, $Rm"),
1671 (ins GPRsp:$Rn, GPRsp:$Rm),
1675 multiclass dp_2src_zext <bits<6> opcode, string asmop, SDPatternOperator op> {
1676 def www : dp_2src_impl<0b0,
1680 (op GPR32:$Rn, (i64 (zext GPR32:$Rm))))],
1683 def xxx : dp_2src_impl<0b1,
1686 [(set GPR64:$Rd, (op GPR64:$Rn, GPR64:$Rm))],
1692 multiclass dp_2src <bits<6> opcode, string asmop, SDPatternOperator op> {
1693 def www : dp_2src_impl<0b0,
1696 [(set GPR32:$Rd, (op GPR32:$Rn, GPR32:$Rm))],
1699 def xxx : dp_2src_impl<0b1,
1702 [(set GPR64:$Rd, (op GPR64:$Rn, GPR64:$Rm))],
1707 // Here we define the data processing 2 source instructions.
1708 defm UDIV : dp_2src<0b000010, "udiv", udiv>;
1709 defm SDIV : dp_2src<0b000011, "sdiv", sdiv>;
1711 defm LSLV : dp_2src_zext<0b001000, "lsl", shl>;
1712 defm LSRV : dp_2src_zext<0b001001, "lsr", srl>;
1713 defm ASRV : dp_2src_zext<0b001010, "asr", sra>;
1714 defm RORV : dp_2src_zext<0b001011, "ror", rotr>;
1716 // Extra patterns for an incoming 64-bit value for a 32-bit
1717 // operation. Since the LLVM operations are undefined (as in C) if the
1718 // RHS is out of range, it's perfectly permissible to discard the high
1719 // bits of the GPR64.
1720 def : Pat<(shl GPR32:$Rn, GPR64:$Rm),
1721 (LSLVwww GPR32:$Rn, (EXTRACT_SUBREG GPR64:$Rm, sub_32))>;
1722 def : Pat<(srl GPR32:$Rn, GPR64:$Rm),
1723 (LSRVwww GPR32:$Rn, (EXTRACT_SUBREG GPR64:$Rm, sub_32))>;
1724 def : Pat<(sra GPR32:$Rn, GPR64:$Rm),
1725 (ASRVwww GPR32:$Rn, (EXTRACT_SUBREG GPR64:$Rm, sub_32))>;
1726 def : Pat<(rotr GPR32:$Rn, GPR64:$Rm),
1727 (RORVwww GPR32:$Rn, (EXTRACT_SUBREG GPR64:$Rm, sub_32))>;
1729 // Here we define the aliases for the data processing 2 source instructions.
1730 def LSL_mnemonic : MnemonicAlias<"lslv", "lsl">;
1731 def LSR_mnemonic : MnemonicAlias<"lsrv", "lsr">;
1732 def ASR_menmonic : MnemonicAlias<"asrv", "asr">;
1733 def ROR_menmonic : MnemonicAlias<"rorv", "ror">;
1735 //===----------------------------------------------------------------------===//
1736 // Data Processing (3 sources) instructions
1737 //===----------------------------------------------------------------------===//
1738 // Contains: MADD, MSUB, SMADDL, SMSUBL, SMULH, UMADDL, UMSUBL, UMULH
1739 // + aliases MUL, MNEG, SMULL, SMNEGL, UMULL, UMNEGL
1741 class A64I_dp3_4operand<bit sf, bits<6> opcode, RegisterClass AccReg,
1742 RegisterClass SrcReg, string asmop, dag pattern>
1743 : A64I_dp3<sf, opcode,
1744 (outs AccReg:$Rd), (ins SrcReg:$Rn, SrcReg:$Rm, AccReg:$Ra),
1745 !strconcat(asmop, "\t$Rd, $Rn, $Rm, $Ra"),
1746 [(set AccReg:$Rd, pattern)], NoItinerary> {
1747 RegisterClass AccGPR = AccReg;
1748 RegisterClass SrcGPR = SrcReg;
1751 def MADDwwww : A64I_dp3_4operand<0b0, 0b000000, GPR32, GPR32, "madd",
1752 (add GPR32:$Ra, (mul GPR32:$Rn, GPR32:$Rm))>;
1753 def MADDxxxx : A64I_dp3_4operand<0b1, 0b000000, GPR64, GPR64, "madd",
1754 (add GPR64:$Ra, (mul GPR64:$Rn, GPR64:$Rm))>;
1756 def MSUBwwww : A64I_dp3_4operand<0b0, 0b000001, GPR32, GPR32, "msub",
1757 (sub GPR32:$Ra, (mul GPR32:$Rn, GPR32:$Rm))>;
1758 def MSUBxxxx : A64I_dp3_4operand<0b1, 0b000001, GPR64, GPR64, "msub",
1759 (sub GPR64:$Ra, (mul GPR64:$Rn, GPR64:$Rm))>;
1761 def SMADDLxwwx : A64I_dp3_4operand<0b1, 0b000010, GPR64, GPR32, "smaddl",
1762 (add GPR64:$Ra, (mul (i64 (sext GPR32:$Rn)), (sext GPR32:$Rm)))>;
1763 def SMSUBLxwwx : A64I_dp3_4operand<0b1, 0b000011, GPR64, GPR32, "smsubl",
1764 (sub GPR64:$Ra, (mul (i64 (sext GPR32:$Rn)), (sext GPR32:$Rm)))>;
1766 def UMADDLxwwx : A64I_dp3_4operand<0b1, 0b001010, GPR64, GPR32, "umaddl",
1767 (add GPR64:$Ra, (mul (i64 (zext GPR32:$Rn)), (zext GPR32:$Rm)))>;
1768 def UMSUBLxwwx : A64I_dp3_4operand<0b1, 0b001011, GPR64, GPR32, "umsubl",
1769 (sub GPR64:$Ra, (mul (i64 (zext GPR32:$Rn)), (zext GPR32:$Rm)))>;
1771 let isCommutable = 1, PostEncoderMethod = "fixMulHigh" in {
1772 def UMULHxxx : A64I_dp3<0b1, 0b001100, (outs GPR64:$Rd),
1773 (ins GPR64:$Rn, GPR64:$Rm),
1774 "umulh\t$Rd, $Rn, $Rm",
1775 [(set GPR64:$Rd, (mulhu GPR64:$Rn, GPR64:$Rm))],
1778 def SMULHxxx : A64I_dp3<0b1, 0b000100, (outs GPR64:$Rd),
1779 (ins GPR64:$Rn, GPR64:$Rm),
1780 "smulh\t$Rd, $Rn, $Rm",
1781 [(set GPR64:$Rd, (mulhs GPR64:$Rn, GPR64:$Rm))],
1785 multiclass A64I_dp3_3operand<string asmop, A64I_dp3_4operand INST,
1786 Register ZR, dag pattern> {
1787 def : InstAlias<asmop # " $Rd, $Rn, $Rm",
1788 (INST INST.AccGPR:$Rd, INST.SrcGPR:$Rn, INST.SrcGPR:$Rm, ZR)>;
1790 def : Pat<pattern, (INST INST.SrcGPR:$Rn, INST.SrcGPR:$Rm, ZR)>;
1793 defm : A64I_dp3_3operand<"mul", MADDwwww, WZR, (mul GPR32:$Rn, GPR32:$Rm)>;
1794 defm : A64I_dp3_3operand<"mul", MADDxxxx, XZR, (mul GPR64:$Rn, GPR64:$Rm)>;
1796 defm : A64I_dp3_3operand<"mneg", MSUBwwww, WZR,
1797 (sub 0, (mul GPR32:$Rn, GPR32:$Rm))>;
1798 defm : A64I_dp3_3operand<"mneg", MSUBxxxx, XZR,
1799 (sub 0, (mul GPR64:$Rn, GPR64:$Rm))>;
1801 defm : A64I_dp3_3operand<"smull", SMADDLxwwx, XZR,
1802 (mul (i64 (sext GPR32:$Rn)), (sext GPR32:$Rm))>;
1803 defm : A64I_dp3_3operand<"smnegl", SMSUBLxwwx, XZR,
1804 (sub 0, (mul (i64 (sext GPR32:$Rn)), (sext GPR32:$Rm)))>;
1806 defm : A64I_dp3_3operand<"umull", UMADDLxwwx, XZR,
1807 (mul (i64 (zext GPR32:$Rn)), (zext GPR32:$Rm))>;
1808 defm : A64I_dp3_3operand<"umnegl", UMSUBLxwwx, XZR,
1809 (sub 0, (mul (i64 (zext GPR32:$Rn)), (zext GPR32:$Rm)))>;
1812 //===----------------------------------------------------------------------===//
1813 // Exception generation
1814 //===----------------------------------------------------------------------===//
1815 // Contains: SVC, HVC, SMC, BRK, HLT, DCPS1, DCPS2, DCPS3
1817 def uimm16_asmoperand : AsmOperandClass {
1818 let Name = "UImm16";
1819 let PredicateMethod = "isUImm<16>";
1820 let RenderMethod = "addImmOperands";
1823 def uimm16 : Operand<i32> {
1824 let ParserMatchClass = uimm16_asmoperand;
1827 class A64I_exceptImpl<bits<3> opc, bits<2> ll, string asmop>
1828 : A64I_exception<opc, 0b000, ll, (outs), (ins uimm16:$UImm16),
1829 !strconcat(asmop, "\t$UImm16"), [], NoItinerary> {
1831 let isTerminator = 1;
1834 def SVCi : A64I_exceptImpl<0b000, 0b01, "svc">;
1835 def HVCi : A64I_exceptImpl<0b000, 0b10, "hvc">;
1836 def SMCi : A64I_exceptImpl<0b000, 0b11, "smc">;
1837 def BRKi : A64I_exceptImpl<0b001, 0b00, "brk">;
1838 def HLTi : A64I_exceptImpl<0b010, 0b00, "hlt">;
1840 def DCPS1i : A64I_exceptImpl<0b101, 0b01, "dcps1">;
1841 def DCPS2i : A64I_exceptImpl<0b101, 0b10, "dcps2">;
1842 def DCPS3i : A64I_exceptImpl<0b101, 0b11, "dcps3">;
1844 // The immediate is optional for the DCPS instructions, defaulting to 0.
1845 def : InstAlias<"dcps1", (DCPS1i 0)>;
1846 def : InstAlias<"dcps2", (DCPS2i 0)>;
1847 def : InstAlias<"dcps3", (DCPS3i 0)>;
1849 //===----------------------------------------------------------------------===//
1850 // Extract (immediate)
1851 //===----------------------------------------------------------------------===//
1852 // Contains: EXTR + alias ROR
1854 def EXTRwwwi : A64I_extract<0b0, 0b000, 0b0,
1856 (ins GPR32:$Rn, GPR32:$Rm, bitfield32_imm:$LSB),
1857 "extr\t$Rd, $Rn, $Rm, $LSB",
1859 (A64Extr GPR32:$Rn, GPR32:$Rm, imm:$LSB))],
1861 def EXTRxxxi : A64I_extract<0b1, 0b000, 0b1,
1863 (ins GPR64:$Rn, GPR64:$Rm, bitfield64_imm:$LSB),
1864 "extr\t$Rd, $Rn, $Rm, $LSB",
1866 (A64Extr GPR64:$Rn, GPR64:$Rm, imm:$LSB))],
1869 def : InstAlias<"ror $Rd, $Rs, $LSB",
1870 (EXTRwwwi GPR32:$Rd, GPR32:$Rs, GPR32:$Rs, bitfield32_imm:$LSB)>;
1871 def : InstAlias<"ror $Rd, $Rs, $LSB",
1872 (EXTRxxxi GPR64:$Rd, GPR64:$Rs, GPR64:$Rs, bitfield64_imm:$LSB)>;
1874 def : Pat<(rotr GPR32:$Rn, bitfield32_imm:$LSB),
1875 (EXTRwwwi GPR32:$Rn, GPR32:$Rn, bitfield32_imm:$LSB)>;
1876 def : Pat<(rotr GPR64:$Rn, bitfield64_imm:$LSB),
1877 (EXTRxxxi GPR64:$Rn, GPR64:$Rn, bitfield64_imm:$LSB)>;
1879 //===----------------------------------------------------------------------===//
1880 // Floating-point compare instructions
1881 //===----------------------------------------------------------------------===//
1882 // Contains: FCMP, FCMPE
1884 def fpzero_asmoperand : AsmOperandClass {
1885 let Name = "FPZero";
1886 let ParserMethod = "ParseFPImmOperand";
1889 def fpz32 : Operand<f32>,
1890 ComplexPattern<f32, 1, "SelectFPZeroOperand", [fpimm]> {
1891 let ParserMatchClass = fpzero_asmoperand;
1892 let PrintMethod = "printFPZeroOperand";
1895 def fpz64 : Operand<f64>,
1896 ComplexPattern<f64, 1, "SelectFPZeroOperand", [fpimm]> {
1897 let ParserMatchClass = fpzero_asmoperand;
1898 let PrintMethod = "printFPZeroOperand";
1901 multiclass A64I_fpcmpSignal<bits<2> type, bit imm, dag ins, string asmop2,
1903 def _quiet : A64I_fpcmp<0b0, 0b0, type, 0b00, {0b0, imm, 0b0, 0b0, 0b0},
1904 (outs), ins, !strconcat("fcmp\t$Rn, ", asmop2),
1905 [pattern], NoItinerary> {
1909 def _sig : A64I_fpcmp<0b0, 0b0, type, 0b00, {0b1, imm, 0b0, 0b0, 0b0},
1910 (outs), ins, !strconcat("fcmpe\t$Rn, ", asmop2),
1916 defm FCMPss : A64I_fpcmpSignal<0b00, 0b0, (ins FPR32:$Rn, FPR32:$Rm), "$Rm",
1917 (set NZCV, (A64cmp (f32 FPR32:$Rn), FPR32:$Rm))>;
1918 defm FCMPdd : A64I_fpcmpSignal<0b01, 0b0, (ins FPR64:$Rn, FPR64:$Rm), "$Rm",
1919 (set NZCV, (A64cmp (f64 FPR64:$Rn), FPR64:$Rm))>;
1921 // What would be Rm should be written as 0, but anything is valid for
1922 // disassembly so we can't set the bits
1923 let PostEncoderMethod = "fixFCMPImm" in {
1924 defm FCMPsi : A64I_fpcmpSignal<0b00, 0b1, (ins FPR32:$Rn, fpz32:$Imm), "$Imm",
1925 (set NZCV, (A64cmp (f32 FPR32:$Rn), fpz32:$Imm))>;
1927 defm FCMPdi : A64I_fpcmpSignal<0b01, 0b1, (ins FPR64:$Rn, fpz64:$Imm), "$Imm",
1928 (set NZCV, (A64cmp (f64 FPR64:$Rn), fpz64:$Imm))>;
1932 //===----------------------------------------------------------------------===//
1933 // Floating-point conditional compare instructions
1934 //===----------------------------------------------------------------------===//
1935 // Contains: FCCMP, FCCMPE
1937 class A64I_fpccmpImpl<bits<2> type, bit op, RegisterClass FPR, string asmop>
1938 : A64I_fpccmp<0b0, 0b0, type, op,
1940 (ins FPR:$Rn, FPR:$Rm, uimm4:$NZCVImm, cond_code_op:$Cond),
1941 !strconcat(asmop, "\t$Rn, $Rm, $NZCVImm, $Cond"),
1946 def FCCMPss : A64I_fpccmpImpl<0b00, 0b0, FPR32, "fccmp">;
1947 def FCCMPEss : A64I_fpccmpImpl<0b00, 0b1, FPR32, "fccmpe">;
1948 def FCCMPdd : A64I_fpccmpImpl<0b01, 0b0, FPR64, "fccmp">;
1949 def FCCMPEdd : A64I_fpccmpImpl<0b01, 0b1, FPR64, "fccmpe">;
1951 //===----------------------------------------------------------------------===//
1952 // Floating-point conditional select instructions
1953 //===----------------------------------------------------------------------===//
1956 let Uses = [NZCV] in {
1957 def FCSELsssc : A64I_fpcondsel<0b0, 0b0, 0b00, (outs FPR32:$Rd),
1958 (ins FPR32:$Rn, FPR32:$Rm, cond_code_op:$Cond),
1959 "fcsel\t$Rd, $Rn, $Rm, $Cond",
1961 (simple_select (f32 FPR32:$Rn),
1966 def FCSELdddc : A64I_fpcondsel<0b0, 0b0, 0b01, (outs FPR64:$Rd),
1967 (ins FPR64:$Rn, FPR64:$Rm, cond_code_op:$Cond),
1968 "fcsel\t$Rd, $Rn, $Rm, $Cond",
1970 (simple_select (f64 FPR64:$Rn),
1975 //===----------------------------------------------------------------------===//
1976 // Floating-point data-processing (1 source)
1977 //===----------------------------------------------------------------------===//
1978 // Contains: FMOV, FABS, FNEG, FSQRT, FCVT, FRINT[NPMZAXI].
1980 def FPNoUnop : PatFrag<(ops node:$val), (fneg node:$val),
1981 [{ (void)N; return false; }]>;
1983 // First we do the fairly trivial bunch with uniform "OP s, s" and "OP d, d"
1984 // syntax. Default to no pattern because most are odd enough not to have one.
1985 multiclass A64I_fpdp1sizes<bits<6> opcode, string asmstr,
1986 SDPatternOperator opnode = FPNoUnop> {
1987 def ss : A64I_fpdp1<0b0, 0b0, 0b00, opcode, (outs FPR32:$Rd), (ins FPR32:$Rn),
1988 !strconcat(asmstr, "\t$Rd, $Rn"),
1989 [(set (f32 FPR32:$Rd), (opnode FPR32:$Rn))],
1992 def dd : A64I_fpdp1<0b0, 0b0, 0b01, opcode, (outs FPR64:$Rd), (ins FPR64:$Rn),
1993 !strconcat(asmstr, "\t$Rd, $Rn"),
1994 [(set (f64 FPR64:$Rd), (opnode FPR64:$Rn))],
1998 defm FMOV : A64I_fpdp1sizes<0b000000, "fmov">;
1999 defm FABS : A64I_fpdp1sizes<0b000001, "fabs", fabs>;
2000 defm FNEG : A64I_fpdp1sizes<0b000010, "fneg", fneg>;
2001 defm FSQRT : A64I_fpdp1sizes<0b000011, "fsqrt", fsqrt>;
2003 defm FRINTN : A64I_fpdp1sizes<0b001000, "frintn">;
2004 defm FRINTP : A64I_fpdp1sizes<0b001001, "frintp", fceil>;
2005 defm FRINTM : A64I_fpdp1sizes<0b001010, "frintm", ffloor>;
2006 defm FRINTZ : A64I_fpdp1sizes<0b001011, "frintz", ftrunc>;
2007 defm FRINTA : A64I_fpdp1sizes<0b001100, "frinta">;
2008 defm FRINTX : A64I_fpdp1sizes<0b001110, "frintx", frint>;
2009 defm FRINTI : A64I_fpdp1sizes<0b001111, "frinti", fnearbyint>;
2011 // The FCVT instrucitons have different source and destination register-types,
2012 // but the fields are uniform everywhere a D-register (say) crops up. Package
2013 // this information in a Record.
2014 class FCVTRegType<RegisterClass rc, bits<2> fld, ValueType vt> {
2015 RegisterClass Class = rc;
2021 def FCVT16 : FCVTRegType<FPR16, 0b11, f16>;
2022 def FCVT32 : FCVTRegType<FPR32, 0b00, f32>;
2023 def FCVT64 : FCVTRegType<FPR64, 0b01, f64>;
2025 class A64I_fpdp1_fcvt<FCVTRegType DestReg, FCVTRegType SrcReg, SDNode opnode>
2026 : A64I_fpdp1<0b0, 0b0, {SrcReg.t1, SrcReg.t0},
2027 {0,0,0,1, DestReg.t1, DestReg.t0},
2028 (outs DestReg.Class:$Rd), (ins SrcReg.Class:$Rn),
2030 [(set (DestReg.VT DestReg.Class:$Rd),
2031 (opnode (SrcReg.VT SrcReg.Class:$Rn)))], NoItinerary>;
2033 def FCVTds : A64I_fpdp1_fcvt<FCVT64, FCVT32, fextend>;
2034 def FCVThs : A64I_fpdp1_fcvt<FCVT16, FCVT32, fround>;
2035 def FCVTsd : A64I_fpdp1_fcvt<FCVT32, FCVT64, fround>;
2036 def FCVThd : A64I_fpdp1_fcvt<FCVT16, FCVT64, fround>;
2037 def FCVTsh : A64I_fpdp1_fcvt<FCVT32, FCVT16, fextend>;
2038 def FCVTdh : A64I_fpdp1_fcvt<FCVT64, FCVT16, fextend>;
2041 //===----------------------------------------------------------------------===//
2042 // Floating-point data-processing (2 sources) instructions
2043 //===----------------------------------------------------------------------===//
2044 // Contains: FMUL, FDIV, FADD, FSUB, FMAX, FMIN, FMAXNM, FMINNM, FNMUL
2046 def FPNoBinop : PatFrag<(ops node:$lhs, node:$rhs), (fadd node:$lhs, node:$rhs),
2047 [{ (void)N; return false; }]>;
2049 multiclass A64I_fpdp2sizes<bits<4> opcode, string asmstr,
2050 SDPatternOperator opnode> {
2051 def sss : A64I_fpdp2<0b0, 0b0, 0b00, opcode,
2053 (ins FPR32:$Rn, FPR32:$Rm),
2054 !strconcat(asmstr, "\t$Rd, $Rn, $Rm"),
2055 [(set (f32 FPR32:$Rd), (opnode FPR32:$Rn, FPR32:$Rm))],
2058 def ddd : A64I_fpdp2<0b0, 0b0, 0b01, opcode,
2060 (ins FPR64:$Rn, FPR64:$Rm),
2061 !strconcat(asmstr, "\t$Rd, $Rn, $Rm"),
2062 [(set (f64 FPR64:$Rd), (opnode FPR64:$Rn, FPR64:$Rm))],
2066 let isCommutable = 1 in {
2067 defm FMUL : A64I_fpdp2sizes<0b0000, "fmul", fmul>;
2068 defm FADD : A64I_fpdp2sizes<0b0010, "fadd", fadd>;
2070 // No patterns for these.
2071 defm FMAX : A64I_fpdp2sizes<0b0100, "fmax", FPNoBinop>;
2072 defm FMIN : A64I_fpdp2sizes<0b0101, "fmin", FPNoBinop>;
2073 defm FMAXNM : A64I_fpdp2sizes<0b0110, "fmaxnm", FPNoBinop>;
2074 defm FMINNM : A64I_fpdp2sizes<0b0111, "fminnm", FPNoBinop>;
2076 defm FNMUL : A64I_fpdp2sizes<0b1000, "fnmul",
2077 PatFrag<(ops node:$lhs, node:$rhs),
2078 (fneg (fmul node:$lhs, node:$rhs))> >;
2081 defm FDIV : A64I_fpdp2sizes<0b0001, "fdiv", fdiv>;
2082 defm FSUB : A64I_fpdp2sizes<0b0011, "fsub", fsub>;
2084 //===----------------------------------------------------------------------===//
2085 // Floating-point data-processing (3 sources) instructions
2086 //===----------------------------------------------------------------------===//
2087 // Contains: FMADD, FMSUB, FNMADD, FNMSUB
2089 def fmsub : PatFrag<(ops node:$Rn, node:$Rm, node:$Ra),
2090 (fma (fneg node:$Rn), node:$Rm, node:$Ra)>;
2091 def fnmadd : PatFrag<(ops node:$Rn, node:$Rm, node:$Ra),
2092 (fma node:$Rn, node:$Rm, (fneg node:$Ra))>;
2093 def fnmsub : PatFrag<(ops node:$Rn, node:$Rm, node:$Ra),
2094 (fma (fneg node:$Rn), node:$Rm, (fneg node:$Ra))>;
2096 class A64I_fpdp3Impl<string asmop, RegisterClass FPR, ValueType VT,
2097 bits<2> type, bit o1, bit o0, SDPatternOperator fmakind>
2098 : A64I_fpdp3<0b0, 0b0, type, o1, o0, (outs FPR:$Rd),
2099 (ins FPR:$Rn, FPR:$Rm, FPR:$Ra),
2100 !strconcat(asmop,"\t$Rd, $Rn, $Rm, $Ra"),
2101 [(set FPR:$Rd, (fmakind (VT FPR:$Rn), FPR:$Rm, FPR:$Ra))],
2104 def FMADDssss : A64I_fpdp3Impl<"fmadd", FPR32, f32, 0b00, 0b0, 0b0, fma>;
2105 def FMSUBssss : A64I_fpdp3Impl<"fmsub", FPR32, f32, 0b00, 0b0, 0b1, fmsub>;
2106 def FNMADDssss : A64I_fpdp3Impl<"fnmadd", FPR32, f32, 0b00, 0b1, 0b0, fnmadd>;
2107 def FNMSUBssss : A64I_fpdp3Impl<"fnmsub", FPR32, f32, 0b00, 0b1, 0b1, fnmsub>;
2109 def FMADDdddd : A64I_fpdp3Impl<"fmadd", FPR64, f64, 0b01, 0b0, 0b0, fma>;
2110 def FMSUBdddd : A64I_fpdp3Impl<"fmsub", FPR64, f64, 0b01, 0b0, 0b1, fmsub>;
2111 def FNMADDdddd : A64I_fpdp3Impl<"fnmadd", FPR64, f64, 0b01, 0b1, 0b0, fnmadd>;
2112 def FNMSUBdddd : A64I_fpdp3Impl<"fnmsub", FPR64, f64, 0b01, 0b1, 0b1, fnmsub>;
2114 //===----------------------------------------------------------------------===//
2115 // Floating-point <-> fixed-point conversion instructions
2116 //===----------------------------------------------------------------------===//
2117 // Contains: FCVTZS, FCVTZU, SCVTF, UCVTF
2119 // #1-#32 allowed, encoded as "64 - <specified imm>
2120 def fixedpos_asmoperand_i32 : AsmOperandClass {
2121 let Name = "CVTFixedPos32";
2122 let RenderMethod = "addCVTFixedPosOperands";
2123 let PredicateMethod = "isCVTFixedPos<32>";
2126 // Also encoded as "64 - <specified imm>" but #1-#64 allowed.
2127 def fixedpos_asmoperand_i64 : AsmOperandClass {
2128 let Name = "CVTFixedPos64";
2129 let RenderMethod = "addCVTFixedPosOperands";
2130 let PredicateMethod = "isCVTFixedPos<64>";
2133 // We need the cartesian product of f32/f64 i32/i64 operands for
2135 // + Selection needs to use operands of correct floating type
2136 // + Assembly parsing and decoding depend on integer width
2137 class cvtfix_i32_op<ValueType FloatVT>
2139 ComplexPattern<FloatVT, 1, "SelectCVTFixedPosOperand<32>", [fpimm]> {
2140 let ParserMatchClass = fixedpos_asmoperand_i32;
2141 let DecoderMethod = "DecodeCVT32FixedPosOperand";
2142 let PrintMethod = "printCVTFixedPosOperand";
2145 class cvtfix_i64_op<ValueType FloatVT>
2147 ComplexPattern<FloatVT, 1, "SelectCVTFixedPosOperand<64>", [fpimm]> {
2148 let ParserMatchClass = fixedpos_asmoperand_i64;
2149 let PrintMethod = "printCVTFixedPosOperand";
2152 // Because of the proliferation of weird operands, it's not really
2153 // worth going for a multiclass here. Oh well.
2155 class A64I_fptofix<bit sf, bits<2> type, bits<3> opcode,
2156 RegisterClass GPR, RegisterClass FPR, Operand scale_op,
2157 string asmop, SDNode cvtop>
2158 : A64I_fpfixed<sf, 0b0, type, 0b11, opcode,
2159 (outs GPR:$Rd), (ins FPR:$Rn, scale_op:$Scale),
2160 !strconcat(asmop, "\t$Rd, $Rn, $Scale"),
2161 [(set GPR:$Rd, (cvtop (fmul FPR:$Rn, scale_op:$Scale)))],
2164 def FCVTZSwsi : A64I_fptofix<0b0, 0b00, 0b000, GPR32, FPR32,
2165 cvtfix_i32_op<f32>, "fcvtzs", fp_to_sint>;
2166 def FCVTZSxsi : A64I_fptofix<0b1, 0b00, 0b000, GPR64, FPR32,
2167 cvtfix_i64_op<f32>, "fcvtzs", fp_to_sint>;
2168 def FCVTZUwsi : A64I_fptofix<0b0, 0b00, 0b001, GPR32, FPR32,
2169 cvtfix_i32_op<f32>, "fcvtzu", fp_to_uint>;
2170 def FCVTZUxsi : A64I_fptofix<0b1, 0b00, 0b001, GPR64, FPR32,
2171 cvtfix_i64_op<f32>, "fcvtzu", fp_to_uint>;
2173 def FCVTZSwdi : A64I_fptofix<0b0, 0b01, 0b000, GPR32, FPR64,
2174 cvtfix_i32_op<f64>, "fcvtzs", fp_to_sint>;
2175 def FCVTZSxdi : A64I_fptofix<0b1, 0b01, 0b000, GPR64, FPR64,
2176 cvtfix_i64_op<f64>, "fcvtzs", fp_to_sint>;
2177 def FCVTZUwdi : A64I_fptofix<0b0, 0b01, 0b001, GPR32, FPR64,
2178 cvtfix_i32_op<f64>, "fcvtzu", fp_to_uint>;
2179 def FCVTZUxdi : A64I_fptofix<0b1, 0b01, 0b001, GPR64, FPR64,
2180 cvtfix_i64_op<f64>, "fcvtzu", fp_to_uint>;
2183 class A64I_fixtofp<bit sf, bits<2> type, bits<3> opcode,
2184 RegisterClass FPR, RegisterClass GPR, Operand scale_op,
2185 string asmop, SDNode cvtop>
2186 : A64I_fpfixed<sf, 0b0, type, 0b00, opcode,
2187 (outs FPR:$Rd), (ins GPR:$Rn, scale_op:$Scale),
2188 !strconcat(asmop, "\t$Rd, $Rn, $Scale"),
2189 [(set FPR:$Rd, (fdiv (cvtop GPR:$Rn), scale_op:$Scale))],
2192 def SCVTFswi : A64I_fixtofp<0b0, 0b00, 0b010, FPR32, GPR32,
2193 cvtfix_i32_op<f32>, "scvtf", sint_to_fp>;
2194 def SCVTFsxi : A64I_fixtofp<0b1, 0b00, 0b010, FPR32, GPR64,
2195 cvtfix_i64_op<f32>, "scvtf", sint_to_fp>;
2196 def UCVTFswi : A64I_fixtofp<0b0, 0b00, 0b011, FPR32, GPR32,
2197 cvtfix_i32_op<f32>, "ucvtf", uint_to_fp>;
2198 def UCVTFsxi : A64I_fixtofp<0b1, 0b00, 0b011, FPR32, GPR64,
2199 cvtfix_i64_op<f32>, "ucvtf", uint_to_fp>;
2200 def SCVTFdwi : A64I_fixtofp<0b0, 0b01, 0b010, FPR64, GPR32,
2201 cvtfix_i32_op<f64>, "scvtf", sint_to_fp>;
2202 def SCVTFdxi : A64I_fixtofp<0b1, 0b01, 0b010, FPR64, GPR64,
2203 cvtfix_i64_op<f64>, "scvtf", sint_to_fp>;
2204 def UCVTFdwi : A64I_fixtofp<0b0, 0b01, 0b011, FPR64, GPR32,
2205 cvtfix_i32_op<f64>, "ucvtf", uint_to_fp>;
2206 def UCVTFdxi : A64I_fixtofp<0b1, 0b01, 0b011, FPR64, GPR64,
2207 cvtfix_i64_op<f64>, "ucvtf", uint_to_fp>;
2209 //===----------------------------------------------------------------------===//
2210 // Floating-point <-> integer conversion instructions
2211 //===----------------------------------------------------------------------===//
2212 // Contains: FCVTZS, FCVTZU, SCVTF, UCVTF
2214 class A64I_fpintI<bit sf, bits<2> type, bits<2> rmode, bits<3> opcode,
2215 RegisterClass DestPR, RegisterClass SrcPR, string asmop>
2216 : A64I_fpint<sf, 0b0, type, rmode, opcode, (outs DestPR:$Rd), (ins SrcPR:$Rn),
2217 !strconcat(asmop, "\t$Rd, $Rn"), [], NoItinerary>;
2219 multiclass A64I_fptointRM<bits<2> rmode, bit o2, string asmop> {
2220 def Sws : A64I_fpintI<0b0, 0b00, rmode, {o2, 0, 0},
2221 GPR32, FPR32, asmop # "s">;
2222 def Sxs : A64I_fpintI<0b1, 0b00, rmode, {o2, 0, 0},
2223 GPR64, FPR32, asmop # "s">;
2224 def Uws : A64I_fpintI<0b0, 0b00, rmode, {o2, 0, 1},
2225 GPR32, FPR32, asmop # "u">;
2226 def Uxs : A64I_fpintI<0b1, 0b00, rmode, {o2, 0, 1},
2227 GPR64, FPR32, asmop # "u">;
2229 def Swd : A64I_fpintI<0b0, 0b01, rmode, {o2, 0, 0},
2230 GPR32, FPR64, asmop # "s">;
2231 def Sxd : A64I_fpintI<0b1, 0b01, rmode, {o2, 0, 0},
2232 GPR64, FPR64, asmop # "s">;
2233 def Uwd : A64I_fpintI<0b0, 0b01, rmode, {o2, 0, 1},
2234 GPR32, FPR64, asmop # "u">;
2235 def Uxd : A64I_fpintI<0b1, 0b01, rmode, {o2, 0, 1},
2236 GPR64, FPR64, asmop # "u">;
2239 defm FCVTN : A64I_fptointRM<0b00, 0b0, "fcvtn">;
2240 defm FCVTP : A64I_fptointRM<0b01, 0b0, "fcvtp">;
2241 defm FCVTM : A64I_fptointRM<0b10, 0b0, "fcvtm">;
2242 defm FCVTZ : A64I_fptointRM<0b11, 0b0, "fcvtz">;
2243 defm FCVTA : A64I_fptointRM<0b00, 0b1, "fcvta">;
2245 def : Pat<(i32 (fp_to_sint FPR32:$Rn)), (FCVTZSws FPR32:$Rn)>;
2246 def : Pat<(i64 (fp_to_sint FPR32:$Rn)), (FCVTZSxs FPR32:$Rn)>;
2247 def : Pat<(i32 (fp_to_uint FPR32:$Rn)), (FCVTZUws FPR32:$Rn)>;
2248 def : Pat<(i64 (fp_to_uint FPR32:$Rn)), (FCVTZUxs FPR32:$Rn)>;
2249 def : Pat<(i32 (fp_to_sint (f64 FPR64:$Rn))), (FCVTZSwd FPR64:$Rn)>;
2250 def : Pat<(i64 (fp_to_sint (f64 FPR64:$Rn))), (FCVTZSxd FPR64:$Rn)>;
2251 def : Pat<(i32 (fp_to_uint (f64 FPR64:$Rn))), (FCVTZUwd FPR64:$Rn)>;
2252 def : Pat<(i64 (fp_to_uint (f64 FPR64:$Rn))), (FCVTZUxd FPR64:$Rn)>;
2254 multiclass A64I_inttofp<bit o0, string asmop> {
2255 def CVTFsw : A64I_fpintI<0b0, 0b00, 0b00, {0, 1, o0}, FPR32, GPR32, asmop>;
2256 def CVTFsx : A64I_fpintI<0b1, 0b00, 0b00, {0, 1, o0}, FPR32, GPR64, asmop>;
2257 def CVTFdw : A64I_fpintI<0b0, 0b01, 0b00, {0, 1, o0}, FPR64, GPR32, asmop>;
2258 def CVTFdx : A64I_fpintI<0b1, 0b01, 0b00, {0, 1, o0}, FPR64, GPR64, asmop>;
2261 defm S : A64I_inttofp<0b0, "scvtf">;
2262 defm U : A64I_inttofp<0b1, "ucvtf">;
2264 def : Pat<(f32 (sint_to_fp GPR32:$Rn)), (SCVTFsw GPR32:$Rn)>;
2265 def : Pat<(f32 (sint_to_fp GPR64:$Rn)), (SCVTFsx GPR64:$Rn)>;
2266 def : Pat<(f64 (sint_to_fp GPR32:$Rn)), (SCVTFdw GPR32:$Rn)>;
2267 def : Pat<(f64 (sint_to_fp GPR64:$Rn)), (SCVTFdx GPR64:$Rn)>;
2268 def : Pat<(f32 (uint_to_fp GPR32:$Rn)), (UCVTFsw GPR32:$Rn)>;
2269 def : Pat<(f32 (uint_to_fp GPR64:$Rn)), (UCVTFsx GPR64:$Rn)>;
2270 def : Pat<(f64 (uint_to_fp GPR32:$Rn)), (UCVTFdw GPR32:$Rn)>;
2271 def : Pat<(f64 (uint_to_fp GPR64:$Rn)), (UCVTFdx GPR64:$Rn)>;
2273 def FMOVws : A64I_fpintI<0b0, 0b00, 0b00, 0b110, GPR32, FPR32, "fmov">;
2274 def FMOVsw : A64I_fpintI<0b0, 0b00, 0b00, 0b111, FPR32, GPR32, "fmov">;
2275 def FMOVxd : A64I_fpintI<0b1, 0b01, 0b00, 0b110, GPR64, FPR64, "fmov">;
2276 def FMOVdx : A64I_fpintI<0b1, 0b01, 0b00, 0b111, FPR64, GPR64, "fmov">;
2278 def : Pat<(i32 (bitconvert (f32 FPR32:$Rn))), (FMOVws FPR32:$Rn)>;
2279 def : Pat<(f32 (bitconvert (i32 GPR32:$Rn))), (FMOVsw GPR32:$Rn)>;
2280 def : Pat<(i64 (bitconvert (f64 FPR64:$Rn))), (FMOVxd FPR64:$Rn)>;
2281 def : Pat<(f64 (bitconvert (i64 GPR64:$Rn))), (FMOVdx GPR64:$Rn)>;
2283 def lane1_asmoperand : AsmOperandClass {
2285 let RenderMethod = "addImmOperands";
2288 def lane1 : Operand<i32> {
2289 let ParserMatchClass = lane1_asmoperand;
2290 let PrintMethod = "printBareImmOperand";
2293 let DecoderMethod = "DecodeFMOVLaneInstruction" in {
2294 def FMOVxv : A64I_fpint<0b1, 0b0, 0b10, 0b01, 0b110,
2295 (outs GPR64:$Rd), (ins VPR128:$Rn, lane1:$Lane),
2296 "fmov\t$Rd, $Rn.d[$Lane]", [], NoItinerary>;
2298 def FMOVvx : A64I_fpint<0b1, 0b0, 0b10, 0b01, 0b111,
2299 (outs VPR128:$Rd), (ins GPR64:$Rn, lane1:$Lane),
2300 "fmov\t$Rd.d[$Lane], $Rn", [], NoItinerary>;
2303 def : InstAlias<"fmov $Rd, $Rn.2d[$Lane]",
2304 (FMOVxv GPR64:$Rd, VPR128:$Rn, lane1:$Lane), 0b0>;
2306 def : InstAlias<"fmov $Rd.2d[$Lane], $Rn",
2307 (FMOVvx VPR128:$Rd, GPR64:$Rn, lane1:$Lane), 0b0>;
2309 //===----------------------------------------------------------------------===//
2310 // Floating-point immediate instructions
2311 //===----------------------------------------------------------------------===//
2314 def fpimm_asmoperand : AsmOperandClass {
2315 let Name = "FMOVImm";
2316 let ParserMethod = "ParseFPImmOperand";
2319 // The MCOperand for these instructions are the encoded 8-bit values.
2320 def SDXF_fpimm : SDNodeXForm<fpimm, [{
2322 A64Imms::isFPImm(N->getValueAPF(), Imm8);
2323 return CurDAG->getTargetConstant(Imm8, MVT::i32);
2326 class fmov_operand<ValueType FT>
2328 PatLeaf<(FT fpimm), [{ return A64Imms::isFPImm(N->getValueAPF()); }],
2330 let PrintMethod = "printFPImmOperand";
2331 let ParserMatchClass = fpimm_asmoperand;
2334 def fmov32_operand : fmov_operand<f32>;
2335 def fmov64_operand : fmov_operand<f64>;
2337 class A64I_fpimm_impl<bits<2> type, RegisterClass Reg, ValueType VT,
2338 Operand fmov_operand>
2339 : A64I_fpimm<0b0, 0b0, type, 0b00000,
2341 (ins fmov_operand:$Imm8),
2343 [(set (VT Reg:$Rd), fmov_operand:$Imm8)],
2346 def FMOVsi : A64I_fpimm_impl<0b00, FPR32, f32, fmov32_operand>;
2347 def FMOVdi : A64I_fpimm_impl<0b01, FPR64, f64, fmov64_operand>;
2349 //===----------------------------------------------------------------------===//
2350 // Load-register (literal) instructions
2351 //===----------------------------------------------------------------------===//
2352 // Contains: LDR, LDRSW, PRFM
2354 def ldrlit_label_asmoperand : AsmOperandClass {
2355 let Name = "LoadLitLabel";
2356 let RenderMethod = "addLabelOperands<19, 4>";
2359 def ldrlit_label : Operand<i64> {
2360 let EncoderMethod = "getLoadLitLabelOpValue";
2362 // This label is a 19-bit offset from PC, scaled by the instruction-width: 4.
2363 let PrintMethod = "printLabelOperand<19, 4>";
2364 let ParserMatchClass = ldrlit_label_asmoperand;
2365 let OperandType = "OPERAND_PCREL";
2368 // Various instructions take an immediate value (which can always be used),
2369 // where some numbers have a symbolic name to make things easier. These operands
2370 // and the associated functions abstract away the differences.
2371 multiclass namedimm<string prefix, string mapper> {
2372 def _asmoperand : AsmOperandClass {
2373 let Name = "NamedImm" # prefix;
2374 let PredicateMethod = "isUImm";
2375 let RenderMethod = "addImmOperands";
2376 let ParserMethod = "ParseNamedImmOperand<" # mapper # ">";
2379 def _op : Operand<i32> {
2380 let ParserMatchClass = !cast<AsmOperandClass>(prefix # "_asmoperand");
2381 let PrintMethod = "printNamedImmOperand<" # mapper # ">";
2382 let DecoderMethod = "DecodeNamedImmOperand<" # mapper # ">";
2386 defm prefetch : namedimm<"prefetch", "A64PRFM::PRFMMapper">;
2388 class A64I_LDRlitSimple<bits<2> opc, bit v, RegisterClass OutReg,
2389 list<dag> patterns = []>
2390 : A64I_LDRlit<opc, v, (outs OutReg:$Rt), (ins ldrlit_label:$Imm19),
2391 "ldr\t$Rt, $Imm19", patterns, NoItinerary>;
2393 let mayLoad = 1 in {
2394 def LDRw_lit : A64I_LDRlitSimple<0b00, 0b0, GPR32>;
2395 def LDRx_lit : A64I_LDRlitSimple<0b01, 0b0, GPR64>;
2398 def LDRs_lit : A64I_LDRlitSimple<0b00, 0b1, FPR32,
2399 [(set (f32 FPR32:$Rt), (load constpool:$Imm19))]>;
2400 def LDRd_lit : A64I_LDRlitSimple<0b01, 0b1, FPR64,
2401 [(set (f64 FPR64:$Rt), (load constpool:$Imm19))]>;
2403 let mayLoad = 1 in {
2404 def LDRq_lit : A64I_LDRlitSimple<0b10, 0b1, FPR128>;
2407 def LDRSWx_lit : A64I_LDRlit<0b10, 0b0,
2409 (ins ldrlit_label:$Imm19),
2410 "ldrsw\t$Rt, $Imm19",
2413 def PRFM_lit : A64I_LDRlit<0b11, 0b0,
2414 (outs), (ins prefetch_op:$Rt, ldrlit_label:$Imm19),
2415 "prfm\t$Rt, $Imm19",
2419 //===----------------------------------------------------------------------===//
2420 // Load-store exclusive instructions
2421 //===----------------------------------------------------------------------===//
2422 // Contains: STXRB, STXRH, STXR, LDXRB, LDXRH, LDXR. STXP, LDXP, STLXRB,
2423 // STLXRH, STLXR, LDAXRB, LDAXRH, LDAXR, STLXP, LDAXP, STLRB,
2424 // STLRH, STLR, LDARB, LDARH, LDAR
2426 // Since these instructions have the undefined register bits set to 1 in
2427 // their canonical form, we need a post encoder method to set those bits
2428 // to 1 when encoding these instructions. We do this using the
2429 // fixLoadStoreExclusive function. This function has template parameters:
2431 // fixLoadStoreExclusive<int hasRs, int hasRt2>
2433 // hasRs indicates that the instruction uses the Rs field, so we won't set
2434 // it to 1 (and the same for Rt2). We don't need template parameters for
2435 // the other register fiels since Rt and Rn are always used.
2437 // This operand parses a GPR64xsp register, followed by an optional immediate
2439 def GPR64xsp0_asmoperand : AsmOperandClass {
2440 let Name = "GPR64xsp0";
2441 let PredicateMethod = "isWrappedReg";
2442 let RenderMethod = "addRegOperands";
2443 let ParserMethod = "ParseLSXAddressOperand";
2446 def GPR64xsp0 : RegisterOperand<GPR64xsp> {
2447 let ParserMatchClass = GPR64xsp0_asmoperand;
2450 //===----------------------------------
2451 // Store-exclusive (releasing & normal)
2452 //===----------------------------------
2454 class A64I_SRexs_impl<bits<2> size, bits<3> opcode, string asm, dag outs,
2455 dag ins, list<dag> pat,
2456 InstrItinClass itin> :
2457 A64I_LDSTex_stn <size,
2458 opcode{2}, 0, opcode{1}, opcode{0},
2460 !strconcat(asm, "\t$Rs, $Rt, [$Rn]"),
2463 let PostEncoderMethod = "fixLoadStoreExclusive<1,0>";
2466 multiclass A64I_SRex<string asmstr, bits<3> opcode, string prefix> {
2467 def _byte: A64I_SRexs_impl<0b00, opcode, !strconcat(asmstr, "b"),
2468 (outs GPR32:$Rs), (ins GPR32:$Rt, GPR64xsp0:$Rn),
2471 def _hword: A64I_SRexs_impl<0b01, opcode, !strconcat(asmstr, "h"),
2472 (outs GPR32:$Rs), (ins GPR32:$Rt, GPR64xsp0:$Rn),
2475 def _word: A64I_SRexs_impl<0b10, opcode, asmstr,
2476 (outs GPR32:$Rs), (ins GPR32:$Rt, GPR64xsp0:$Rn),
2479 def _dword: A64I_SRexs_impl<0b11, opcode, asmstr,
2480 (outs GPR32:$Rs), (ins GPR64:$Rt, GPR64xsp0:$Rn),
2484 defm STXR : A64I_SRex<"stxr", 0b000, "STXR">;
2485 defm STLXR : A64I_SRex<"stlxr", 0b001, "STLXR">;
2487 //===----------------------------------
2489 //===----------------------------------
2491 class A64I_LRexs_impl<bits<2> size, bits<3> opcode, string asm, dag outs,
2492 dag ins, list<dag> pat,
2493 InstrItinClass itin> :
2494 A64I_LDSTex_tn <size,
2495 opcode{2}, 1, opcode{1}, opcode{0},
2497 !strconcat(asm, "\t$Rt, [$Rn]"),
2500 let PostEncoderMethod = "fixLoadStoreExclusive<0,0>";
2503 multiclass A64I_LRex<string asmstr, bits<3> opcode> {
2504 def _byte: A64I_LRexs_impl<0b00, opcode, !strconcat(asmstr, "b"),
2505 (outs GPR32:$Rt), (ins GPR64xsp0:$Rn),
2508 def _hword: A64I_LRexs_impl<0b01, opcode, !strconcat(asmstr, "h"),
2509 (outs GPR32:$Rt), (ins GPR64xsp0:$Rn),
2512 def _word: A64I_LRexs_impl<0b10, opcode, asmstr,
2513 (outs GPR32:$Rt), (ins GPR64xsp0:$Rn),
2516 def _dword: A64I_LRexs_impl<0b11, opcode, asmstr,
2517 (outs GPR64:$Rt), (ins GPR64xsp0:$Rn),
2521 defm LDXR : A64I_LRex<"ldxr", 0b000>;
2522 defm LDAXR : A64I_LRex<"ldaxr", 0b001>;
2523 defm LDAR : A64I_LRex<"ldar", 0b101>;
2525 class acquiring_load<PatFrag base>
2526 : PatFrag<(ops node:$ptr), (base node:$ptr), [{
2527 return cast<AtomicSDNode>(N)->getOrdering() == Acquire;
2530 def atomic_load_acquire_8 : acquiring_load<atomic_load_8>;
2531 def atomic_load_acquire_16 : acquiring_load<atomic_load_16>;
2532 def atomic_load_acquire_32 : acquiring_load<atomic_load_32>;
2533 def atomic_load_acquire_64 : acquiring_load<atomic_load_64>;
2535 def : Pat<(atomic_load_acquire_8 GPR64xsp:$Rn), (LDAR_byte GPR64xsp0:$Rn)>;
2536 def : Pat<(atomic_load_acquire_16 GPR64xsp:$Rn), (LDAR_hword GPR64xsp0:$Rn)>;
2537 def : Pat<(atomic_load_acquire_32 GPR64xsp:$Rn), (LDAR_word GPR64xsp0:$Rn)>;
2538 def : Pat<(atomic_load_acquire_64 GPR64xsp:$Rn), (LDAR_dword GPR64xsp0:$Rn)>;
2540 //===----------------------------------
2541 // Store-release (no exclusivity)
2542 //===----------------------------------
2544 class A64I_SLexs_impl<bits<2> size, bits<3> opcode, string asm, dag outs,
2545 dag ins, list<dag> pat,
2546 InstrItinClass itin> :
2547 A64I_LDSTex_tn <size,
2548 opcode{2}, 0, opcode{1}, opcode{0},
2550 !strconcat(asm, "\t$Rt, [$Rn]"),
2553 let PostEncoderMethod = "fixLoadStoreExclusive<0,0>";
2556 class releasing_store<PatFrag base>
2557 : PatFrag<(ops node:$ptr, node:$val), (base node:$ptr, node:$val), [{
2558 return cast<AtomicSDNode>(N)->getOrdering() == Release;
2561 def atomic_store_release_8 : releasing_store<atomic_store_8>;
2562 def atomic_store_release_16 : releasing_store<atomic_store_16>;
2563 def atomic_store_release_32 : releasing_store<atomic_store_32>;
2564 def atomic_store_release_64 : releasing_store<atomic_store_64>;
2566 multiclass A64I_SLex<string asmstr, bits<3> opcode, string prefix> {
2567 def _byte: A64I_SLexs_impl<0b00, opcode, !strconcat(asmstr, "b"),
2568 (outs), (ins GPR32:$Rt, GPR64xsp0:$Rn),
2569 [(atomic_store_release_8 GPR64xsp0:$Rn, GPR32:$Rt)],
2572 def _hword: A64I_SLexs_impl<0b01, opcode, !strconcat(asmstr, "h"),
2573 (outs), (ins GPR32:$Rt, GPR64xsp0:$Rn),
2574 [(atomic_store_release_16 GPR64xsp0:$Rn, GPR32:$Rt)],
2577 def _word: A64I_SLexs_impl<0b10, opcode, asmstr,
2578 (outs), (ins GPR32:$Rt, GPR64xsp0:$Rn),
2579 [(atomic_store_release_32 GPR64xsp0:$Rn, GPR32:$Rt)],
2582 def _dword: A64I_SLexs_impl<0b11, opcode, asmstr,
2583 (outs), (ins GPR64:$Rt, GPR64xsp0:$Rn),
2584 [(atomic_store_release_64 GPR64xsp0:$Rn, GPR64:$Rt)],
2588 defm STLR : A64I_SLex<"stlr", 0b101, "STLR">;
2590 //===----------------------------------
2591 // Store-exclusive pair (releasing & normal)
2592 //===----------------------------------
2594 class A64I_SPexs_impl<bits<2> size, bits<3> opcode, string asm, dag outs,
2595 dag ins, list<dag> pat,
2596 InstrItinClass itin> :
2597 A64I_LDSTex_stt2n <size,
2598 opcode{2}, 0, opcode{1}, opcode{0},
2600 !strconcat(asm, "\t$Rs, $Rt, $Rt2, [$Rn]"),
2606 multiclass A64I_SPex<string asmstr, bits<3> opcode> {
2607 def _word: A64I_SPexs_impl<0b10, opcode, asmstr, (outs),
2608 (ins GPR32:$Rs, GPR32:$Rt, GPR32:$Rt2,
2612 def _dword: A64I_SPexs_impl<0b11, opcode, asmstr, (outs),
2613 (ins GPR32:$Rs, GPR64:$Rt, GPR64:$Rt2,
2618 defm STXP : A64I_SPex<"stxp", 0b010>;
2619 defm STLXP : A64I_SPex<"stlxp", 0b011>;
2621 //===----------------------------------
2622 // Load-exclusive pair (acquiring & normal)
2623 //===----------------------------------
2625 class A64I_LPexs_impl<bits<2> size, bits<3> opcode, string asm, dag outs,
2626 dag ins, list<dag> pat,
2627 InstrItinClass itin> :
2628 A64I_LDSTex_tt2n <size,
2629 opcode{2}, 1, opcode{1}, opcode{0},
2631 !strconcat(asm, "\t$Rt, $Rt2, [$Rn]"),
2634 let DecoderMethod = "DecodeLoadPairExclusiveInstruction";
2635 let PostEncoderMethod = "fixLoadStoreExclusive<0,1>";
2638 multiclass A64I_LPex<string asmstr, bits<3> opcode> {
2639 def _word: A64I_LPexs_impl<0b10, opcode, asmstr,
2640 (outs GPR32:$Rt, GPR32:$Rt2),
2641 (ins GPR64xsp0:$Rn),
2644 def _dword: A64I_LPexs_impl<0b11, opcode, asmstr,
2645 (outs GPR64:$Rt, GPR64:$Rt2),
2646 (ins GPR64xsp0:$Rn),
2650 defm LDXP : A64I_LPex<"ldxp", 0b010>;
2651 defm LDAXP : A64I_LPex<"ldaxp", 0b011>;
2653 //===----------------------------------------------------------------------===//
2654 // Load-store register (unscaled immediate) instructions
2655 //===----------------------------------------------------------------------===//
2656 // Contains: LDURB, LDURH, LDRUSB, LDRUSH, LDRUSW, STUR, STURB, STURH and PRFUM
2660 //===----------------------------------------------------------------------===//
2661 // Load-store register (register offset) instructions
2662 //===----------------------------------------------------------------------===//
2663 // Contains: LDRB, LDRH, LDRSB, LDRSH, LDRSW, STR, STRB, STRH and PRFM
2667 //===----------------------------------------------------------------------===//
2668 // Load-store register (unsigned immediate) instructions
2669 //===----------------------------------------------------------------------===//
2670 // Contains: LDRB, LDRH, LDRSB, LDRSH, LDRSW, STR, STRB, STRH and PRFM
2674 //===----------------------------------------------------------------------===//
2675 // Load-store register (immediate post-indexed) instructions
2676 //===----------------------------------------------------------------------===//
2677 // Contains: STRB, STRH, STR, LDRB, LDRH, LDR, LDRSB, LDRSH, LDRSW
2681 //===----------------------------------------------------------------------===//
2682 // Load-store register (immediate pre-indexed) instructions
2683 //===----------------------------------------------------------------------===//
2684 // Contains: STRB, STRH, STR, LDRB, LDRH, LDR, LDRSB, LDRSH, LDRSW
2686 // Note that patterns are much later on in a completely separate section (they
2687 // need ADRPxi to be defined).
2689 //===-------------------------------
2690 // 1. Various operands needed
2691 //===-------------------------------
2693 //===-------------------------------
2694 // 1.1 Unsigned 12-bit immediate operands
2695 //===-------------------------------
2696 // The addressing mode for these instructions consists of an unsigned 12-bit
2697 // immediate which is scaled by the size of the memory access.
2699 // We represent this in the MC layer by two operands:
2700 // 1. A base register.
2701 // 2. A 12-bit immediate: not multiplied by access size, so "LDR x0,[x0,#8]"
2702 // would have '1' in this field.
2703 // This means that separate functions are needed for converting representations
2704 // which *are* aware of the intended access size.
2706 // Anything that creates an MCInst (Decoding, selection and AsmParsing) has to
2707 // know the access size via some means. An isolated operand does not have this
2708 // information unless told from here, which means we need separate tablegen
2709 // Operands for each access size. This multiclass takes care of instantiating
2710 // the correct template functions in the rest of the backend.
2712 //===-------------------------------
2713 // 1.1 Unsigned 12-bit immediate operands
2714 //===-------------------------------
2716 multiclass offsets_uimm12<int MemSize, string prefix> {
2717 def uimm12_asmoperand : AsmOperandClass {
2718 let Name = "OffsetUImm12_" # MemSize;
2719 let PredicateMethod = "isOffsetUImm12<" # MemSize # ">";
2720 let RenderMethod = "addOffsetUImm12Operands<" # MemSize # ">";
2723 // Pattern is really no more than an ImmLeaf, but predicated on MemSize which
2724 // complicates things beyond TableGen's ken.
2725 def uimm12 : Operand<i64>,
2726 ComplexPattern<i64, 1, "SelectOffsetUImm12<" # MemSize # ">"> {
2727 let ParserMatchClass
2728 = !cast<AsmOperandClass>(prefix # uimm12_asmoperand);
2730 let PrintMethod = "printOffsetUImm12Operand<" # MemSize # ">";
2731 let EncoderMethod = "getOffsetUImm12OpValue<" # MemSize # ">";
2735 defm byte_ : offsets_uimm12<1, "byte_">;
2736 defm hword_ : offsets_uimm12<2, "hword_">;
2737 defm word_ : offsets_uimm12<4, "word_">;
2738 defm dword_ : offsets_uimm12<8, "dword_">;
2739 defm qword_ : offsets_uimm12<16, "qword_">;
2741 //===-------------------------------
2742 // 1.1 Signed 9-bit immediate operands
2743 //===-------------------------------
2745 // The MCInst is expected to store the bit-wise encoding of the value,
2746 // which amounts to lopping off the extended sign bits.
2747 def SDXF_simm9 : SDNodeXForm<imm, [{
2748 return CurDAG->getTargetConstant(N->getZExtValue() & 0x1ff, MVT::i32);
2751 def simm9_asmoperand : AsmOperandClass {
2753 let PredicateMethod = "isSImm<9>";
2754 let RenderMethod = "addSImmOperands<9>";
2757 def simm9 : Operand<i64>,
2758 ImmLeaf<i64, [{ return Imm >= -0x100 && Imm <= 0xff; }],
2760 let PrintMethod = "printOffsetSImm9Operand";
2761 let ParserMatchClass = simm9_asmoperand;
2765 //===-------------------------------
2766 // 1.3 Register offset extensions
2767 //===-------------------------------
2769 // The assembly-syntax for these addressing-modes is:
2770 // [<Xn|SP>, <R><m> {, <extend> {<amount>}}]
2772 // The essential semantics are:
2773 // + <amount> is a shift: #<log(transfer size)> or #0
2774 // + <R> can be W or X.
2775 // + If <R> is W, <extend> can be UXTW or SXTW
2776 // + If <R> is X, <extend> can be LSL or SXTX
2778 // The trickiest of those constraints is that Rm can be either GPR32 or GPR64,
2779 // which will need separate instructions for LLVM type-consistency. We'll also
2780 // need separate operands, of course.
2781 multiclass regexts<int MemSize, int RmSize, RegisterClass GPR,
2782 string Rm, string prefix> {
2783 def regext_asmoperand : AsmOperandClass {
2784 let Name = "AddrRegExtend_" # MemSize # "_" # Rm;
2785 let PredicateMethod = "isAddrRegExtend<" # MemSize # "," # RmSize # ">";
2786 let RenderMethod = "addAddrRegExtendOperands<" # MemSize # ">";
2789 def regext : Operand<i64> {
2791 = "printAddrRegExtendOperand<" # MemSize # ", " # RmSize # ">";
2793 let DecoderMethod = "DecodeAddrRegExtendOperand";
2794 let ParserMatchClass
2795 = !cast<AsmOperandClass>(prefix # regext_asmoperand);
2799 multiclass regexts_wx<int MemSize, string prefix> {
2800 // Rm is an X-register if LSL or SXTX are specified as the shift.
2801 defm Xm_ : regexts<MemSize, 64, GPR64, "Xm", prefix # "Xm_">;
2803 // Rm is a W-register if UXTW or SXTW are specified as the shift.
2804 defm Wm_ : regexts<MemSize, 32, GPR32, "Wm", prefix # "Wm_">;
2807 defm byte_ : regexts_wx<1, "byte_">;
2808 defm hword_ : regexts_wx<2, "hword_">;
2809 defm word_ : regexts_wx<4, "word_">;
2810 defm dword_ : regexts_wx<8, "dword_">;
2811 defm qword_ : regexts_wx<16, "qword_">;
2814 //===------------------------------
2815 // 2. The instructions themselves.
2816 //===------------------------------
2818 // We have the following instructions to implement:
2819 // | | B | H | W | X |
2820 // |-----------------+-------+-------+-------+--------|
2821 // | unsigned str | STRB | STRH | STR | STR |
2822 // | unsigned ldr | LDRB | LDRH | LDR | LDR |
2823 // | signed ldr to W | LDRSB | LDRSH | - | - |
2824 // | signed ldr to X | LDRSB | LDRSH | LDRSW | (PRFM) |
2826 // This will instantiate the LDR/STR instructions you'd expect to use for an
2827 // unsigned datatype (first two rows above) or floating-point register, which is
2828 // reasonably uniform across all access sizes.
2831 //===------------------------------
2832 // 2.1 Regular instructions
2833 //===------------------------------
2835 // This class covers the basic unsigned or irrelevantly-signed loads and stores,
2836 // to general-purpose and floating-point registers.
2838 class AddrParams<string prefix> {
2839 Operand uimm12 = !cast<Operand>(prefix # "_uimm12");
2841 Operand regextWm = !cast<Operand>(prefix # "_Wm_regext");
2842 Operand regextXm = !cast<Operand>(prefix # "_Xm_regext");
2845 def byte_addrparams : AddrParams<"byte">;
2846 def hword_addrparams : AddrParams<"hword">;
2847 def word_addrparams : AddrParams<"word">;
2848 def dword_addrparams : AddrParams<"dword">;
2849 def qword_addrparams : AddrParams<"qword">;
2851 multiclass A64I_LDRSTR_unsigned<string prefix, bits<2> size, bit v,
2852 bit high_opc, string asmsuffix,
2853 RegisterClass GPR, AddrParams params> {
2854 // Unsigned immediate
2855 def _STR : A64I_LSunsigimm<size, v, {high_opc, 0b0},
2856 (outs), (ins GPR:$Rt, GPR64xsp:$Rn, params.uimm12:$UImm12),
2857 "str" # asmsuffix # "\t$Rt, [$Rn, $UImm12]",
2861 def : InstAlias<"str" # asmsuffix # " $Rt, [$Rn]",
2862 (!cast<Instruction>(prefix # "_STR") GPR:$Rt, GPR64xsp:$Rn, 0)>;
2864 def _LDR : A64I_LSunsigimm<size, v, {high_opc, 0b1},
2865 (outs GPR:$Rt), (ins GPR64xsp:$Rn, params.uimm12:$UImm12),
2866 "ldr" # asmsuffix # "\t$Rt, [$Rn, $UImm12]",
2870 def : InstAlias<"ldr" # asmsuffix # " $Rt, [$Rn]",
2871 (!cast<Instruction>(prefix # "_LDR") GPR:$Rt, GPR64xsp:$Rn, 0)>;
2873 // Register offset (four of these: load/store and Wm/Xm).
2874 let mayLoad = 1 in {
2875 def _Wm_RegOffset_LDR : A64I_LSregoff<size, v, {high_opc, 0b1}, 0b0,
2877 (ins GPR64xsp:$Rn, GPR32:$Rm, params.regextWm:$Ext),
2878 "ldr" # asmsuffix # "\t$Rt, [$Rn, $Rm, $Ext]",
2881 def _Xm_RegOffset_LDR : A64I_LSregoff<size, v, {high_opc, 0b1}, 0b1,
2883 (ins GPR64xsp:$Rn, GPR64:$Rm, params.regextXm:$Ext),
2884 "ldr" # asmsuffix # "\t$Rt, [$Rn, $Rm, $Ext]",
2887 def : InstAlias<"ldr" # asmsuffix # " $Rt, [$Rn, $Rm]",
2888 (!cast<Instruction>(prefix # "_Xm_RegOffset_LDR") GPR:$Rt, GPR64xsp:$Rn,
2891 let mayStore = 1 in {
2892 def _Wm_RegOffset_STR : A64I_LSregoff<size, v, {high_opc, 0b0}, 0b0,
2893 (outs), (ins GPR:$Rt, GPR64xsp:$Rn, GPR32:$Rm,
2894 params.regextWm:$Ext),
2895 "str" # asmsuffix # "\t$Rt, [$Rn, $Rm, $Ext]",
2898 def _Xm_RegOffset_STR : A64I_LSregoff<size, v, {high_opc, 0b0}, 0b1,
2899 (outs), (ins GPR:$Rt, GPR64xsp:$Rn, GPR64:$Rm,
2900 params.regextXm:$Ext),
2901 "str" # asmsuffix # "\t$Rt, [$Rn, $Rm, $Ext]",
2904 def : InstAlias<"str" # asmsuffix # " $Rt, [$Rn, $Rm]",
2905 (!cast<Instruction>(prefix # "_Xm_RegOffset_STR") GPR:$Rt, GPR64xsp:$Rn,
2908 // Unaligned immediate
2909 def _STUR : A64I_LSunalimm<size, v, {high_opc, 0b0},
2910 (outs), (ins GPR:$Rt, GPR64xsp:$Rn, simm9:$SImm9),
2911 "stur" # asmsuffix # "\t$Rt, [$Rn, $SImm9]",
2915 def : InstAlias<"stur" # asmsuffix # " $Rt, [$Rn]",
2916 (!cast<Instruction>(prefix # "_STUR") GPR:$Rt, GPR64xsp:$Rn, 0)>;
2918 def _LDUR : A64I_LSunalimm<size, v, {high_opc, 0b1},
2919 (outs GPR:$Rt), (ins GPR64xsp:$Rn, simm9:$SImm9),
2920 "ldur" # asmsuffix # "\t$Rt, [$Rn, $SImm9]",
2924 def : InstAlias<"ldur" # asmsuffix # " $Rt, [$Rn]",
2925 (!cast<Instruction>(prefix # "_LDUR") GPR:$Rt, GPR64xsp:$Rn, 0)>;
2928 def _PostInd_STR : A64I_LSpostind<size, v, {high_opc, 0b0},
2929 (outs GPR64xsp:$Rn_wb),
2930 (ins GPR:$Rt, GPR64xsp:$Rn, simm9:$SImm9),
2931 "str" # asmsuffix # "\t$Rt, [$Rn], $SImm9",
2933 let Constraints = "$Rn = $Rn_wb";
2936 // Decoder only needed for unpredictability checking (FIXME).
2937 let DecoderMethod = "DecodeSingleIndexedInstruction";
2940 def _PostInd_LDR : A64I_LSpostind<size, v, {high_opc, 0b1},
2941 (outs GPR:$Rt, GPR64xsp:$Rn_wb),
2942 (ins GPR64xsp:$Rn, simm9:$SImm9),
2943 "ldr" # asmsuffix # "\t$Rt, [$Rn], $SImm9",
2946 let Constraints = "$Rn = $Rn_wb";
2947 let DecoderMethod = "DecodeSingleIndexedInstruction";
2951 def _PreInd_STR : A64I_LSpreind<size, v, {high_opc, 0b0},
2952 (outs GPR64xsp:$Rn_wb),
2953 (ins GPR:$Rt, GPR64xsp:$Rn, simm9:$SImm9),
2954 "str" # asmsuffix # "\t$Rt, [$Rn, $SImm9]!",
2956 let Constraints = "$Rn = $Rn_wb";
2959 // Decoder only needed for unpredictability checking (FIXME).
2960 let DecoderMethod = "DecodeSingleIndexedInstruction";
2963 def _PreInd_LDR : A64I_LSpreind<size, v, {high_opc, 0b1},
2964 (outs GPR:$Rt, GPR64xsp:$Rn_wb),
2965 (ins GPR64xsp:$Rn, simm9:$SImm9),
2966 "ldr" # asmsuffix # "\t$Rt, [$Rn, $SImm9]!",
2969 let Constraints = "$Rn = $Rn_wb";
2970 let DecoderMethod = "DecodeSingleIndexedInstruction";
2975 // STRB/LDRB: First define the instructions
2977 : A64I_LDRSTR_unsigned<"LS8", 0b00, 0b0, 0b0, "b", GPR32, byte_addrparams>;
2981 : A64I_LDRSTR_unsigned<"LS16", 0b01, 0b0, 0b0, "h", GPR32, hword_addrparams>;
2984 // STR/LDR to/from a W register
2986 : A64I_LDRSTR_unsigned<"LS32", 0b10, 0b0, 0b0, "", GPR32, word_addrparams>;
2988 // STR/LDR to/from an X register
2990 : A64I_LDRSTR_unsigned<"LS64", 0b11, 0b0, 0b0, "", GPR64, dword_addrparams>;
2992 // STR/LDR to/from a B register
2994 : A64I_LDRSTR_unsigned<"LSFP8", 0b00, 0b1, 0b0, "", FPR8, byte_addrparams>;
2996 // STR/LDR to/from an H register
2998 : A64I_LDRSTR_unsigned<"LSFP16", 0b01, 0b1, 0b0, "", FPR16, hword_addrparams>;
3000 // STR/LDR to/from an S register
3002 : A64I_LDRSTR_unsigned<"LSFP32", 0b10, 0b1, 0b0, "", FPR32, word_addrparams>;
3003 // STR/LDR to/from a D register
3005 : A64I_LDRSTR_unsigned<"LSFP64", 0b11, 0b1, 0b0, "", FPR64, dword_addrparams>;
3006 // STR/LDR to/from a Q register
3008 : A64I_LDRSTR_unsigned<"LSFP128", 0b00, 0b1, 0b1, "", FPR128,
3011 //===------------------------------
3013 //===------------------------------
3015 // Byte and half-word signed loads can both go into either an X or a W register,
3016 // so it's worth factoring out. Signed word loads don't fit because there is no
3018 multiclass A64I_LDR_signed<bits<2> size, string asmopcode, AddrParams params,
3021 def w : A64I_LSunsigimm<size, 0b0, 0b11,
3023 (ins GPR64xsp:$Rn, params.uimm12:$UImm12),
3024 "ldrs" # asmopcode # "\t$Rt, [$Rn, $UImm12]",
3028 def : InstAlias<"ldrs" # asmopcode # " $Rt, [$Rn]",
3029 (!cast<Instruction>(prefix # w) GPR32:$Rt, GPR64xsp:$Rn, 0)>;
3031 def x : A64I_LSunsigimm<size, 0b0, 0b10,
3033 (ins GPR64xsp:$Rn, params.uimm12:$UImm12),
3034 "ldrs" # asmopcode # "\t$Rt, [$Rn, $UImm12]",
3038 def : InstAlias<"ldrs" # asmopcode # " $Rt, [$Rn]",
3039 (!cast<Instruction>(prefix # x) GPR64:$Rt, GPR64xsp:$Rn, 0)>;
3042 let mayLoad = 1 in {
3043 def w_Wm_RegOffset : A64I_LSregoff<size, 0b0, 0b11, 0b0,
3045 (ins GPR64xsp:$Rn, GPR32:$Rm, params.regextWm:$Ext),
3046 "ldrs" # asmopcode # "\t$Rt, [$Rn, $Rm, $Ext]",
3049 def w_Xm_RegOffset : A64I_LSregoff<size, 0b0, 0b11, 0b1,
3051 (ins GPR64xsp:$Rn, GPR64:$Rm, params.regextXm:$Ext),
3052 "ldrs" # asmopcode # "\t$Rt, [$Rn, $Rm, $Ext]",
3055 def x_Wm_RegOffset : A64I_LSregoff<size, 0b0, 0b10, 0b0,
3057 (ins GPR64xsp:$Rn, GPR32:$Rm, params.regextWm:$Ext),
3058 "ldrs" # asmopcode # "\t$Rt, [$Rn, $Rm, $Ext]",
3061 def x_Xm_RegOffset : A64I_LSregoff<size, 0b0, 0b10, 0b1,
3063 (ins GPR64xsp:$Rn, GPR64:$Rm, params.regextXm:$Ext),
3064 "ldrs" # asmopcode # "\t$Rt, [$Rn, $Rm, $Ext]",
3067 def : InstAlias<"ldrs" # asmopcode # " $Rt, [$Rn, $Rm]",
3068 (!cast<Instruction>(prefix # "w_Xm_RegOffset") GPR32:$Rt, GPR64xsp:$Rn,
3071 def : InstAlias<"ldrs" # asmopcode # " $Rt, [$Rn, $Rm]",
3072 (!cast<Instruction>(prefix # "x_Xm_RegOffset") GPR64:$Rt, GPR64xsp:$Rn,
3076 let mayLoad = 1 in {
3078 def w_U : A64I_LSunalimm<size, 0b0, 0b11,
3080 (ins GPR64xsp:$Rn, simm9:$SImm9),
3081 "ldurs" # asmopcode # "\t$Rt, [$Rn, $SImm9]",
3084 def x_U : A64I_LSunalimm<size, 0b0, 0b10,
3086 (ins GPR64xsp:$Rn, simm9:$SImm9),
3087 "ldurs" # asmopcode # "\t$Rt, [$Rn, $SImm9]",
3092 def w_PostInd : A64I_LSpostind<size, 0b0, 0b11,
3093 (outs GPR32:$Rt, GPR64xsp:$Rn_wb),
3094 (ins GPR64xsp:$Rn, simm9:$SImm9),
3095 "ldrs" # asmopcode # "\t$Rt, [$Rn], $SImm9",
3097 let Constraints = "$Rn = $Rn_wb";
3098 let DecoderMethod = "DecodeSingleIndexedInstruction";
3101 def x_PostInd : A64I_LSpostind<size, 0b0, 0b10,
3102 (outs GPR64:$Rt, GPR64xsp:$Rn_wb),
3103 (ins GPR64xsp:$Rn, simm9:$SImm9),
3104 "ldrs" # asmopcode # "\t$Rt, [$Rn], $SImm9",
3106 let Constraints = "$Rn = $Rn_wb";
3107 let DecoderMethod = "DecodeSingleIndexedInstruction";
3111 def w_PreInd : A64I_LSpreind<size, 0b0, 0b11,
3112 (outs GPR32:$Rt, GPR64xsp:$Rn_wb),
3113 (ins GPR64xsp:$Rn, simm9:$SImm9),
3114 "ldrs" # asmopcode # "\t$Rt, [$Rn, $SImm9]!",
3116 let Constraints = "$Rn = $Rn_wb";
3117 let DecoderMethod = "DecodeSingleIndexedInstruction";
3120 def x_PreInd : A64I_LSpreind<size, 0b0, 0b10,
3121 (outs GPR64:$Rt, GPR64xsp:$Rn_wb),
3122 (ins GPR64xsp:$Rn, simm9:$SImm9),
3123 "ldrs" # asmopcode # "\t$Rt, [$Rn, $SImm9]!",
3125 let Constraints = "$Rn = $Rn_wb";
3126 let DecoderMethod = "DecodeSingleIndexedInstruction";
3128 } // let mayLoad = 1
3132 defm LDRSB : A64I_LDR_signed<0b00, "b", byte_addrparams, "LDRSB">;
3134 defm LDRSH : A64I_LDR_signed<0b01, "h", hword_addrparams, "LDRSH">;
3136 // LDRSW: load a 32-bit register, sign-extending to 64-bits.
3138 : A64I_LSunsigimm<0b10, 0b0, 0b10,
3140 (ins GPR64xsp:$Rn, word_uimm12:$UImm12),
3141 "ldrsw\t$Rt, [$Rn, $UImm12]",
3145 def : InstAlias<"ldrsw $Rt, [$Rn]", (LDRSWx GPR64:$Rt, GPR64xsp:$Rn, 0)>;
3147 let mayLoad = 1 in {
3148 def LDRSWx_Wm_RegOffset : A64I_LSregoff<0b10, 0b0, 0b10, 0b0,
3150 (ins GPR64xsp:$Rn, GPR32:$Rm, word_Wm_regext:$Ext),
3151 "ldrsw\t$Rt, [$Rn, $Rm, $Ext]",
3154 def LDRSWx_Xm_RegOffset : A64I_LSregoff<0b10, 0b0, 0b10, 0b1,
3156 (ins GPR64xsp:$Rn, GPR64:$Rm, word_Xm_regext:$Ext),
3157 "ldrsw\t$Rt, [$Rn, $Rm, $Ext]",
3160 def : InstAlias<"ldrsw $Rt, [$Rn, $Rm]",
3161 (LDRSWx_Xm_RegOffset GPR64:$Rt, GPR64xsp:$Rn, GPR64:$Rm, 2)>;
3165 : A64I_LSunalimm<0b10, 0b0, 0b10,
3167 (ins GPR64xsp:$Rn, simm9:$SImm9),
3168 "ldursw\t$Rt, [$Rn, $SImm9]",
3172 def : InstAlias<"ldursw $Rt, [$Rn]", (LDURSWx GPR64:$Rt, GPR64xsp:$Rn, 0)>;
3175 : A64I_LSpostind<0b10, 0b0, 0b10,
3176 (outs GPR64:$Rt, GPR64xsp:$Rn_wb),
3177 (ins GPR64xsp:$Rn, simm9:$SImm9),
3178 "ldrsw\t$Rt, [$Rn], $SImm9",
3181 let Constraints = "$Rn = $Rn_wb";
3182 let DecoderMethod = "DecodeSingleIndexedInstruction";
3185 def LDRSWx_PreInd : A64I_LSpreind<0b10, 0b0, 0b10,
3186 (outs GPR64:$Rt, GPR64xsp:$Rn_wb),
3187 (ins GPR64xsp:$Rn, simm9:$SImm9),
3188 "ldrsw\t$Rt, [$Rn, $SImm9]!",
3191 let Constraints = "$Rn = $Rn_wb";
3192 let DecoderMethod = "DecodeSingleIndexedInstruction";
3195 //===------------------------------
3196 // 2.4 Prefetch operations
3197 //===------------------------------
3199 def PRFM : A64I_LSunsigimm<0b11, 0b0, 0b10, (outs),
3200 (ins prefetch_op:$Rt, GPR64xsp:$Rn, dword_uimm12:$UImm12),
3201 "prfm\t$Rt, [$Rn, $UImm12]",
3205 def : InstAlias<"prfm $Rt, [$Rn]",
3206 (PRFM prefetch_op:$Rt, GPR64xsp:$Rn, 0)>;
3208 let mayLoad = 1 in {
3209 def PRFM_Wm_RegOffset : A64I_LSregoff<0b11, 0b0, 0b10, 0b0, (outs),
3210 (ins prefetch_op:$Rt, GPR64xsp:$Rn,
3211 GPR32:$Rm, dword_Wm_regext:$Ext),
3212 "prfm\t$Rt, [$Rn, $Rm, $Ext]",
3214 def PRFM_Xm_RegOffset : A64I_LSregoff<0b11, 0b0, 0b10, 0b1, (outs),
3215 (ins prefetch_op:$Rt, GPR64xsp:$Rn,
3216 GPR64:$Rm, dword_Xm_regext:$Ext),
3217 "prfm\t$Rt, [$Rn, $Rm, $Ext]",
3221 def : InstAlias<"prfm $Rt, [$Rn, $Rm]",
3222 (PRFM_Xm_RegOffset prefetch_op:$Rt, GPR64xsp:$Rn,
3226 def PRFUM : A64I_LSunalimm<0b11, 0b0, 0b10, (outs),
3227 (ins prefetch_op:$Rt, GPR64xsp:$Rn, simm9:$SImm9),
3228 "prfum\t$Rt, [$Rn, $SImm9]",
3232 def : InstAlias<"prfum $Rt, [$Rn]",
3233 (PRFUM prefetch_op:$Rt, GPR64xsp:$Rn, 0)>;
3235 //===----------------------------------------------------------------------===//
3236 // Load-store register (unprivileged) instructions
3237 //===----------------------------------------------------------------------===//
3238 // Contains: LDTRB, LDTRH, LDTRSB, LDTRSH, LDTRSW, STTR, STTRB and STTRH
3240 // These instructions very much mirror the "unscaled immediate" loads, but since
3241 // there are no floating-point variants we need to split them out into their own
3242 // section to avoid instantiation of "ldtr d0, [sp]" etc.
3244 multiclass A64I_LDTRSTTR<bits<2> size, string asmsuffix, RegisterClass GPR,
3246 def _UnPriv_STR : A64I_LSunpriv<size, 0b0, 0b00,
3247 (outs), (ins GPR:$Rt, GPR64xsp:$Rn, simm9:$SImm9),
3248 "sttr" # asmsuffix # "\t$Rt, [$Rn, $SImm9]",
3253 def : InstAlias<"sttr" # asmsuffix # " $Rt, [$Rn]",
3254 (!cast<Instruction>(prefix # "_UnPriv_STR") GPR:$Rt, GPR64xsp:$Rn, 0)>;
3256 def _UnPriv_LDR : A64I_LSunpriv<size, 0b0, 0b01,
3257 (outs GPR:$Rt), (ins GPR64xsp:$Rn, simm9:$SImm9),
3258 "ldtr" # asmsuffix # "\t$Rt, [$Rn, $SImm9]",
3263 def : InstAlias<"ldtr" # asmsuffix # " $Rt, [$Rn]",
3264 (!cast<Instruction>(prefix # "_UnPriv_LDR") GPR:$Rt, GPR64xsp:$Rn, 0)>;
3268 // STTRB/LDTRB: First define the instructions
3269 defm LS8 : A64I_LDTRSTTR<0b00, "b", GPR32, "LS8">;
3272 defm LS16 : A64I_LDTRSTTR<0b01, "h", GPR32, "LS16">;
3274 // STTR/LDTR to/from a W register
3275 defm LS32 : A64I_LDTRSTTR<0b10, "", GPR32, "LS32">;
3277 // STTR/LDTR to/from an X register
3278 defm LS64 : A64I_LDTRSTTR<0b11, "", GPR64, "LS64">;
3280 // Now a class for the signed instructions that can go to either 32 or 64
3282 multiclass A64I_LDTR_signed<bits<2> size, string asmopcode, string prefix> {
3283 let mayLoad = 1 in {
3284 def w : A64I_LSunpriv<size, 0b0, 0b11,
3286 (ins GPR64xsp:$Rn, simm9:$SImm9),
3287 "ldtrs" # asmopcode # "\t$Rt, [$Rn, $SImm9]",
3290 def x : A64I_LSunpriv<size, 0b0, 0b10,
3292 (ins GPR64xsp:$Rn, simm9:$SImm9),
3293 "ldtrs" # asmopcode # "\t$Rt, [$Rn, $SImm9]",
3297 def : InstAlias<"ldtrs" # asmopcode # " $Rt, [$Rn]",
3298 (!cast<Instruction>(prefix # "w") GPR32:$Rt, GPR64xsp:$Rn, 0)>;
3300 def : InstAlias<"ldtrs" # asmopcode # " $Rt, [$Rn]",
3301 (!cast<Instruction>(prefix # "x") GPR64:$Rt, GPR64xsp:$Rn, 0)>;
3306 defm LDTRSB : A64I_LDTR_signed<0b00, "b", "LDTRSB">;
3308 defm LDTRSH : A64I_LDTR_signed<0b01, "h", "LDTRSH">;
3310 // And finally LDTRSW which only goes to 64 bits.
3311 def LDTRSWx : A64I_LSunpriv<0b10, 0b0, 0b10,
3313 (ins GPR64xsp:$Rn, simm9:$SImm9),
3314 "ldtrsw\t$Rt, [$Rn, $SImm9]",
3318 def : InstAlias<"ldtrsw $Rt, [$Rn]", (LDTRSWx GPR64:$Rt, GPR64xsp:$Rn, 0)>;
3320 //===----------------------------------------------------------------------===//
3321 // Load-store register pair (offset) instructions
3322 //===----------------------------------------------------------------------===//
3326 //===----------------------------------------------------------------------===//
3327 // Load-store register pair (post-indexed) instructions
3328 //===----------------------------------------------------------------------===//
3329 // Contains: STP, LDP, LDPSW
3333 //===----------------------------------------------------------------------===//
3334 // Load-store register pair (pre-indexed) instructions
3335 //===----------------------------------------------------------------------===//
3336 // Contains: STP, LDP, LDPSW
3340 //===----------------------------------------------------------------------===//
3341 // Load-store non-temporal register pair (offset) instructions
3342 //===----------------------------------------------------------------------===//
3343 // Contains: STNP, LDNP
3346 // Anything that creates an MCInst (Decoding, selection and AsmParsing) has to
3347 // know the access size via some means. An isolated operand does not have this
3348 // information unless told from here, which means we need separate tablegen
3349 // Operands for each access size. This multiclass takes care of instantiating
3350 // the correct template functions in the rest of the backend.
3352 multiclass offsets_simm7<string MemSize, string prefix> {
3353 // The bare signed 7-bit immediate is used in post-indexed instructions, but
3354 // because of the scaling performed a generic "simm7" operand isn't
3355 // appropriate here either.
3356 def simm7_asmoperand : AsmOperandClass {
3357 let Name = "SImm7_Scaled" # MemSize;
3358 let PredicateMethod = "isSImm7Scaled<" # MemSize # ">";
3359 let RenderMethod = "addSImm7ScaledOperands<" # MemSize # ">";
3362 def simm7 : Operand<i64> {
3363 let PrintMethod = "printSImm7ScaledOperand<" # MemSize # ">";
3364 let ParserMatchClass = !cast<AsmOperandClass>(prefix # "simm7_asmoperand");
3368 defm word_ : offsets_simm7<"4", "word_">;
3369 defm dword_ : offsets_simm7<"8", "dword_">;
3370 defm qword_ : offsets_simm7<"16", "qword_">;
3372 multiclass A64I_LSPsimple<bits<2> opc, bit v, RegisterClass SomeReg,
3373 Operand simm7, string prefix> {
3374 def _STR : A64I_LSPoffset<opc, v, 0b0, (outs),
3375 (ins SomeReg:$Rt, SomeReg:$Rt2, GPR64xsp:$Rn, simm7:$SImm7),
3376 "stp\t$Rt, $Rt2, [$Rn, $SImm7]", [], NoItinerary> {
3378 let DecoderMethod = "DecodeLDSTPairInstruction";
3380 def : InstAlias<"stp $Rt, $Rt2, [$Rn]",
3381 (!cast<Instruction>(prefix # "_STR") SomeReg:$Rt,
3382 SomeReg:$Rt2, GPR64xsp:$Rn, 0)>;
3384 def _LDR : A64I_LSPoffset<opc, v, 0b1,
3385 (outs SomeReg:$Rt, SomeReg:$Rt2),
3386 (ins GPR64xsp:$Rn, simm7:$SImm7),
3387 "ldp\t$Rt, $Rt2, [$Rn, $SImm7]", [], NoItinerary> {
3389 let DecoderMethod = "DecodeLDSTPairInstruction";
3391 def : InstAlias<"ldp $Rt, $Rt2, [$Rn]",
3392 (!cast<Instruction>(prefix # "_LDR") SomeReg:$Rt,
3393 SomeReg:$Rt2, GPR64xsp:$Rn, 0)>;
3395 def _PostInd_STR : A64I_LSPpostind<opc, v, 0b0,
3396 (outs GPR64xsp:$Rn_wb),
3397 (ins SomeReg:$Rt, SomeReg:$Rt2,
3400 "stp\t$Rt, $Rt2, [$Rn], $SImm7",
3403 let Constraints = "$Rn = $Rn_wb";
3405 // Decoder only needed for unpredictability checking (FIXME).
3406 let DecoderMethod = "DecodeLDSTPairInstruction";
3409 def _PostInd_LDR : A64I_LSPpostind<opc, v, 0b1,
3410 (outs SomeReg:$Rt, SomeReg:$Rt2, GPR64xsp:$Rn_wb),
3411 (ins GPR64xsp:$Rn, simm7:$SImm7),
3412 "ldp\t$Rt, $Rt2, [$Rn], $SImm7",
3415 let Constraints = "$Rn = $Rn_wb";
3416 let DecoderMethod = "DecodeLDSTPairInstruction";
3419 def _PreInd_STR : A64I_LSPpreind<opc, v, 0b0, (outs GPR64xsp:$Rn_wb),
3420 (ins SomeReg:$Rt, SomeReg:$Rt2, GPR64xsp:$Rn, simm7:$SImm7),
3421 "stp\t$Rt, $Rt2, [$Rn, $SImm7]!",
3424 let Constraints = "$Rn = $Rn_wb";
3425 let DecoderMethod = "DecodeLDSTPairInstruction";
3428 def _PreInd_LDR : A64I_LSPpreind<opc, v, 0b1,
3429 (outs SomeReg:$Rt, SomeReg:$Rt2, GPR64xsp:$Rn_wb),
3430 (ins GPR64xsp:$Rn, simm7:$SImm7),
3431 "ldp\t$Rt, $Rt2, [$Rn, $SImm7]!",
3434 let Constraints = "$Rn = $Rn_wb";
3435 let DecoderMethod = "DecodeLDSTPairInstruction";
3438 def _NonTemp_STR : A64I_LSPnontemp<opc, v, 0b0, (outs),
3439 (ins SomeReg:$Rt, SomeReg:$Rt2, GPR64xsp:$Rn, simm7:$SImm7),
3440 "stnp\t$Rt, $Rt2, [$Rn, $SImm7]", [], NoItinerary> {
3442 let DecoderMethod = "DecodeLDSTPairInstruction";
3444 def : InstAlias<"stnp $Rt, $Rt2, [$Rn]",
3445 (!cast<Instruction>(prefix # "_NonTemp_STR") SomeReg:$Rt,
3446 SomeReg:$Rt2, GPR64xsp:$Rn, 0)>;
3448 def _NonTemp_LDR : A64I_LSPnontemp<opc, v, 0b1,
3449 (outs SomeReg:$Rt, SomeReg:$Rt2),
3450 (ins GPR64xsp:$Rn, simm7:$SImm7),
3451 "ldnp\t$Rt, $Rt2, [$Rn, $SImm7]", [], NoItinerary> {
3453 let DecoderMethod = "DecodeLDSTPairInstruction";
3455 def : InstAlias<"ldnp $Rt, $Rt2, [$Rn]",
3456 (!cast<Instruction>(prefix # "_NonTemp_LDR") SomeReg:$Rt,
3457 SomeReg:$Rt2, GPR64xsp:$Rn, 0)>;
3462 defm LSPair32 : A64I_LSPsimple<0b00, 0b0, GPR32, word_simm7, "LSPair32">;
3463 defm LSPair64 : A64I_LSPsimple<0b10, 0b0, GPR64, dword_simm7, "LSPair64">;
3464 defm LSFPPair32 : A64I_LSPsimple<0b00, 0b1, FPR32, word_simm7, "LSFPPair32">;
3465 defm LSFPPair64 : A64I_LSPsimple<0b01, 0b1, FPR64, dword_simm7, "LSFPPair64">;
3466 defm LSFPPair128 : A64I_LSPsimple<0b10, 0b1, FPR128, qword_simm7,
3470 def LDPSWx : A64I_LSPoffset<0b01, 0b0, 0b1,
3471 (outs GPR64:$Rt, GPR64:$Rt2),
3472 (ins GPR64xsp:$Rn, word_simm7:$SImm7),
3473 "ldpsw\t$Rt, $Rt2, [$Rn, $SImm7]", [], NoItinerary> {
3475 let DecoderMethod = "DecodeLDSTPairInstruction";
3477 def : InstAlias<"ldpsw $Rt, $Rt2, [$Rn]",
3478 (LDPSWx GPR64:$Rt, GPR64:$Rt2, GPR64xsp:$Rn, 0)>;
3480 def LDPSWx_PostInd : A64I_LSPpostind<0b01, 0b0, 0b1,
3481 (outs GPR64:$Rt, GPR64:$Rt2, GPR64:$Rn_wb),
3482 (ins GPR64xsp:$Rn, word_simm7:$SImm7),
3483 "ldpsw\t$Rt, $Rt2, [$Rn], $SImm7",
3486 let Constraints = "$Rn = $Rn_wb";
3487 let DecoderMethod = "DecodeLDSTPairInstruction";
3490 def LDPSWx_PreInd : A64I_LSPpreind<0b01, 0b0, 0b1,
3491 (outs GPR64:$Rt, GPR64:$Rt2, GPR64:$Rn_wb),
3492 (ins GPR64xsp:$Rn, word_simm7:$SImm7),
3493 "ldpsw\t$Rt, $Rt2, [$Rn, $SImm7]!",
3496 let Constraints = "$Rn = $Rn_wb";
3497 let DecoderMethod = "DecodeLDSTPairInstruction";
3500 //===----------------------------------------------------------------------===//
3501 // Logical (immediate) instructions
3502 //===----------------------------------------------------------------------===//
3503 // Contains: AND, ORR, EOR, ANDS, + aliases TST, MOV
3505 multiclass logical_imm_operands<string prefix, string note,
3506 int size, ValueType VT> {
3507 def _asmoperand : AsmOperandClass {
3508 let Name = "LogicalImm" # note # size;
3509 let PredicateMethod = "isLogicalImm" # note # "<" # size # ">";
3510 let RenderMethod = "addLogicalImmOperands<" # size # ">";
3514 : Operand<VT>, ComplexPattern<VT, 1, "SelectLogicalImm", [imm]> {
3515 let ParserMatchClass = !cast<AsmOperandClass>(prefix # "_asmoperand");
3516 let PrintMethod = "printLogicalImmOperand<" # size # ">";
3517 let DecoderMethod = "DecodeLogicalImmOperand<" # size # ">";
3521 defm logical_imm32 : logical_imm_operands<"logical_imm32", "", 32, i32>;
3522 defm logical_imm64 : logical_imm_operands<"logical_imm64", "", 64, i64>;
3524 // The mov versions only differ in assembly parsing, where they
3525 // exclude values representable with either MOVZ or MOVN.
3526 defm logical_imm32_mov
3527 : logical_imm_operands<"logical_imm32_mov", "MOV", 32, i32>;
3528 defm logical_imm64_mov
3529 : logical_imm_operands<"logical_imm64_mov", "MOV", 64, i64>;
3532 multiclass A64I_logimmSizes<bits<2> opc, string asmop, SDNode opnode> {
3533 def wwi : A64I_logicalimm<0b0, opc, (outs GPR32wsp:$Rd),
3534 (ins GPR32:$Rn, logical_imm32_operand:$Imm),
3535 !strconcat(asmop, "\t$Rd, $Rn, $Imm"),
3537 (opnode GPR32:$Rn, logical_imm32_operand:$Imm))],
3540 def xxi : A64I_logicalimm<0b1, opc, (outs GPR64xsp:$Rd),
3541 (ins GPR64:$Rn, logical_imm64_operand:$Imm),
3542 !strconcat(asmop, "\t$Rd, $Rn, $Imm"),
3544 (opnode GPR64:$Rn, logical_imm64_operand:$Imm))],
3548 defm AND : A64I_logimmSizes<0b00, "and", and>;
3549 defm ORR : A64I_logimmSizes<0b01, "orr", or>;
3550 defm EOR : A64I_logimmSizes<0b10, "eor", xor>;
3552 let Defs = [NZCV] in {
3553 def ANDSwwi : A64I_logicalimm<0b0, 0b11, (outs GPR32:$Rd),
3554 (ins GPR32:$Rn, logical_imm32_operand:$Imm),
3555 "ands\t$Rd, $Rn, $Imm",
3558 def ANDSxxi : A64I_logicalimm<0b1, 0b11, (outs GPR64:$Rd),
3559 (ins GPR64:$Rn, logical_imm64_operand:$Imm),
3560 "ands\t$Rd, $Rn, $Imm",
3565 def : InstAlias<"tst $Rn, $Imm",
3566 (ANDSwwi WZR, GPR32:$Rn, logical_imm32_operand:$Imm)>;
3567 def : InstAlias<"tst $Rn, $Imm",
3568 (ANDSxxi XZR, GPR64:$Rn, logical_imm64_operand:$Imm)>;
3569 def : InstAlias<"mov $Rd, $Imm",
3570 (ORRwwi GPR32wsp:$Rd, WZR, logical_imm32_mov_operand:$Imm)>;
3571 def : InstAlias<"mov $Rd, $Imm",
3572 (ORRxxi GPR64xsp:$Rd, XZR, logical_imm64_mov_operand:$Imm)>;
3574 //===----------------------------------------------------------------------===//
3575 // Logical (shifted register) instructions
3576 //===----------------------------------------------------------------------===//
3577 // Contains: AND, BIC, ORR, ORN, EOR, EON, ANDS, BICS + aliases TST, MVN, MOV
3579 // Operand for optimizing (icmp (and LHS, RHS), 0, SomeCode). In theory "ANDS"
3580 // behaves differently for unsigned comparisons, so we defensively only allow
3581 // signed or n/a as the operand. In practice "unsigned greater than 0" is "not
3582 // equal to 0" and LLVM gives us this.
3583 def signed_cond : PatLeaf<(cond), [{
3584 return !isUnsignedIntSetCC(N->get());
3588 // These instructions share their "shift" operands with add/sub (shifted
3589 // register instructions). They are defined there.
3591 // N.b. the commutable parameter is just !N. It will be first against the wall
3592 // when the revolution comes.
3593 multiclass logical_shifts<string prefix, bit sf, bits<2> opc,
3594 bit N, bit commutable,
3595 string asmop, SDPatternOperator opfrag, string sty,
3596 RegisterClass GPR, list<Register> defs> {
3597 let isCommutable = commutable, Defs = defs in {
3598 def _lsl : A64I_logicalshift<sf, opc, 0b00, N,
3600 (ins GPR:$Rn, GPR:$Rm,
3601 !cast<Operand>("lsl_operand_" # sty):$Imm6),
3602 !strconcat(asmop, "\t$Rd, $Rn, $Rm, $Imm6"),
3603 [(set GPR:$Rd, (opfrag GPR:$Rn, (shl GPR:$Rm,
3604 !cast<Operand>("lsl_operand_" # sty):$Imm6))
3608 def _lsr : A64I_logicalshift<sf, opc, 0b01, N,
3610 (ins GPR:$Rn, GPR:$Rm,
3611 !cast<Operand>("lsr_operand_" # sty):$Imm6),
3612 !strconcat(asmop, "\t$Rd, $Rn, $Rm, $Imm6"),
3613 [(set GPR:$Rd, (opfrag GPR:$Rn, (srl GPR:$Rm,
3614 !cast<Operand>("lsr_operand_" # sty):$Imm6))
3618 def _asr : A64I_logicalshift<sf, opc, 0b10, N,
3620 (ins GPR:$Rn, GPR:$Rm,
3621 !cast<Operand>("asr_operand_" # sty):$Imm6),
3622 !strconcat(asmop, "\t$Rd, $Rn, $Rm, $Imm6"),
3623 [(set GPR:$Rd, (opfrag GPR:$Rn, (sra GPR:$Rm,
3624 !cast<Operand>("asr_operand_" # sty):$Imm6))
3628 def _ror : A64I_logicalshift<sf, opc, 0b11, N,
3630 (ins GPR:$Rn, GPR:$Rm,
3631 !cast<Operand>("ror_operand_" # sty):$Imm6),
3632 !strconcat(asmop, "\t$Rd, $Rn, $Rm, $Imm6"),
3633 [(set GPR:$Rd, (opfrag GPR:$Rn, (rotr GPR:$Rm,
3634 !cast<Operand>("ror_operand_" # sty):$Imm6))
3640 : InstAlias<!strconcat(asmop, " $Rd, $Rn, $Rm"),
3641 (!cast<Instruction>(prefix # "_lsl") GPR:$Rd, GPR:$Rn,
3644 def : Pat<(opfrag GPR:$Rn, GPR:$Rm),
3645 (!cast<Instruction>(prefix # "_lsl") GPR:$Rn, GPR:$Rm, 0)>;
3648 multiclass logical_sizes<string prefix, bits<2> opc, bit N, bit commutable,
3649 string asmop, SDPatternOperator opfrag,
3650 list<Register> defs> {
3651 defm xxx : logical_shifts<prefix # "xxx", 0b1, opc, N,
3652 commutable, asmop, opfrag, "i64", GPR64, defs>;
3653 defm www : logical_shifts<prefix # "www", 0b0, opc, N,
3654 commutable, asmop, opfrag, "i32", GPR32, defs>;
3658 defm AND : logical_sizes<"AND", 0b00, 0b0, 0b1, "and", and, []>;
3659 defm ORR : logical_sizes<"ORR", 0b01, 0b0, 0b1, "orr", or, []>;
3660 defm EOR : logical_sizes<"EOR", 0b10, 0b0, 0b1, "eor", xor, []>;
3661 defm ANDS : logical_sizes<"ANDS", 0b11, 0b0, 0b1, "ands",
3662 PatFrag<(ops node:$lhs, node:$rhs), (and node:$lhs, node:$rhs),
3663 [{ (void)N; return false; }]>,
3666 defm BIC : logical_sizes<"BIC", 0b00, 0b1, 0b0, "bic",
3667 PatFrag<(ops node:$lhs, node:$rhs),
3668 (and node:$lhs, (not node:$rhs))>, []>;
3669 defm ORN : logical_sizes<"ORN", 0b01, 0b1, 0b0, "orn",
3670 PatFrag<(ops node:$lhs, node:$rhs),
3671 (or node:$lhs, (not node:$rhs))>, []>;
3672 defm EON : logical_sizes<"EON", 0b10, 0b1, 0b0, "eon",
3673 PatFrag<(ops node:$lhs, node:$rhs),
3674 (xor node:$lhs, (not node:$rhs))>, []>;
3675 defm BICS : logical_sizes<"BICS", 0b11, 0b1, 0b0, "bics",
3676 PatFrag<(ops node:$lhs, node:$rhs),
3677 (and node:$lhs, (not node:$rhs)),
3678 [{ (void)N; return false; }]>,
3681 multiclass tst_shifts<string prefix, bit sf, string sty, RegisterClass GPR> {
3682 let isCommutable = 1, Rd = 0b11111, Defs = [NZCV] in {
3683 def _lsl : A64I_logicalshift<sf, 0b11, 0b00, 0b0,
3685 (ins GPR:$Rn, GPR:$Rm,
3686 !cast<Operand>("lsl_operand_" # sty):$Imm6),
3687 "tst\t$Rn, $Rm, $Imm6",
3688 [(set NZCV, (A64setcc (and GPR:$Rn, (shl GPR:$Rm,
3689 !cast<Operand>("lsl_operand_" # sty):$Imm6)),
3694 def _lsr : A64I_logicalshift<sf, 0b11, 0b01, 0b0,
3696 (ins GPR:$Rn, GPR:$Rm,
3697 !cast<Operand>("lsr_operand_" # sty):$Imm6),
3698 "tst\t$Rn, $Rm, $Imm6",
3699 [(set NZCV, (A64setcc (and GPR:$Rn, (srl GPR:$Rm,
3700 !cast<Operand>("lsr_operand_" # sty):$Imm6)),
3704 def _asr : A64I_logicalshift<sf, 0b11, 0b10, 0b0,
3706 (ins GPR:$Rn, GPR:$Rm,
3707 !cast<Operand>("asr_operand_" # sty):$Imm6),
3708 "tst\t$Rn, $Rm, $Imm6",
3709 [(set NZCV, (A64setcc (and GPR:$Rn, (sra GPR:$Rm,
3710 !cast<Operand>("asr_operand_" # sty):$Imm6)),
3714 def _ror : A64I_logicalshift<sf, 0b11, 0b11, 0b0,
3716 (ins GPR:$Rn, GPR:$Rm,
3717 !cast<Operand>("ror_operand_" # sty):$Imm6),
3718 "tst\t$Rn, $Rm, $Imm6",
3719 [(set NZCV, (A64setcc (and GPR:$Rn, (rotr GPR:$Rm,
3720 !cast<Operand>("ror_operand_" # sty):$Imm6)),
3725 def _noshift : InstAlias<"tst $Rn, $Rm",
3726 (!cast<Instruction>(prefix # "_lsl") GPR:$Rn, GPR:$Rm, 0)>;
3728 def : Pat<(A64setcc (and GPR:$Rn, GPR:$Rm), 0, signed_cond),
3729 (!cast<Instruction>(prefix # "_lsl") GPR:$Rn, GPR:$Rm, 0)>;
3732 defm TSTxx : tst_shifts<"TSTxx", 0b1, "i64", GPR64>;
3733 defm TSTww : tst_shifts<"TSTww", 0b0, "i32", GPR32>;
3736 multiclass mvn_shifts<string prefix, bit sf, string sty, RegisterClass GPR> {
3737 let isCommutable = 0, Rn = 0b11111 in {
3738 def _lsl : A64I_logicalshift<sf, 0b01, 0b00, 0b1,
3741 !cast<Operand>("lsl_operand_" # sty):$Imm6),
3742 "mvn\t$Rd, $Rm, $Imm6",
3743 [(set GPR:$Rd, (not (shl GPR:$Rm,
3744 !cast<Operand>("lsl_operand_" # sty):$Imm6)))],
3748 def _lsr : A64I_logicalshift<sf, 0b01, 0b01, 0b1,
3751 !cast<Operand>("lsr_operand_" # sty):$Imm6),
3752 "mvn\t$Rd, $Rm, $Imm6",
3753 [(set GPR:$Rd, (not (srl GPR:$Rm,
3754 !cast<Operand>("lsr_operand_" # sty):$Imm6)))],
3757 def _asr : A64I_logicalshift<sf, 0b01, 0b10, 0b1,
3760 !cast<Operand>("asr_operand_" # sty):$Imm6),
3761 "mvn\t$Rd, $Rm, $Imm6",
3762 [(set GPR:$Rd, (not (sra GPR:$Rm,
3763 !cast<Operand>("asr_operand_" # sty):$Imm6)))],
3766 def _ror : A64I_logicalshift<sf, 0b01, 0b11, 0b1,
3769 !cast<Operand>("ror_operand_" # sty):$Imm6),
3770 "mvn\t$Rd, $Rm, $Imm6",
3771 [(set GPR:$Rd, (not (rotr GPR:$Rm,
3772 !cast<Operand>("lsl_operand_" # sty):$Imm6)))],
3776 def _noshift : InstAlias<"mvn $Rn, $Rm",
3777 (!cast<Instruction>(prefix # "_lsl") GPR:$Rn, GPR:$Rm, 0)>;
3779 def : Pat<(not GPR:$Rm),
3780 (!cast<Instruction>(prefix # "_lsl") GPR:$Rm, 0)>;
3783 defm MVNxx : mvn_shifts<"MVNxx", 0b1, "i64", GPR64>;
3784 defm MVNww : mvn_shifts<"MVNww", 0b0, "i32", GPR32>;
3786 def MOVxx :InstAlias<"mov $Rd, $Rm", (ORRxxx_lsl GPR64:$Rd, XZR, GPR64:$Rm, 0)>;
3787 def MOVww :InstAlias<"mov $Rd, $Rm", (ORRwww_lsl GPR32:$Rd, WZR, GPR32:$Rm, 0)>;
3789 //===----------------------------------------------------------------------===//
3790 // Move wide (immediate) instructions
3791 //===----------------------------------------------------------------------===//
3792 // Contains: MOVN, MOVZ, MOVK + MOV aliases
3794 // A wide variety of different relocations are needed for variants of these
3795 // instructions, so it turns out that we need a different operand for all of
3797 multiclass movw_operands<string prefix, string instname, int width> {
3798 def _imm_asmoperand : AsmOperandClass {
3799 let Name = instname # width # "Shifted" # shift;
3800 let PredicateMethod = "is" # instname # width # "Imm";
3801 let RenderMethod = "addMoveWideImmOperands";
3803 let ParserMethod = "ParseImmWithLSLOperand";
3806 def _imm : Operand<i32> {
3807 let ParserMatchClass = !cast<AsmOperandClass>(prefix # "_imm_asmoperand");
3808 let PrintMethod = "printMoveWideImmOperand";
3809 let EncoderMethod = "getMoveWideImmOpValue";
3810 let DecoderMethod = "DecodeMoveWideImmOperand<" # width # ">";
3812 let MIOperandInfo = (ops uimm16:$UImm16, imm:$Shift);
3816 defm movn32 : movw_operands<"movn32", "MOVN", 32>;
3817 defm movn64 : movw_operands<"movn64", "MOVN", 64>;
3818 defm movz32 : movw_operands<"movz32", "MOVZ", 32>;
3819 defm movz64 : movw_operands<"movz64", "MOVZ", 64>;
3820 defm movk32 : movw_operands<"movk32", "MOVK", 32>;
3821 defm movk64 : movw_operands<"movk64", "MOVK", 64>;
3823 multiclass A64I_movwSizes<bits<2> opc, string asmop, dag ins32bit,
3826 def wii : A64I_movw<0b0, opc, (outs GPR32:$Rd), ins32bit,
3827 !strconcat(asmop, "\t$Rd, $FullImm"),
3830 let UImm16 = FullImm{15-0};
3831 let Shift = FullImm{17-16};
3834 def xii : A64I_movw<0b1, opc, (outs GPR64:$Rd), ins64bit,
3835 !strconcat(asmop, "\t$Rd, $FullImm"),
3838 let UImm16 = FullImm{15-0};
3839 let Shift = FullImm{17-16};
3843 let isMoveImm = 1, isReMaterializable = 1,
3844 isAsCheapAsAMove = 1, neverHasSideEffects = 1 in {
3845 defm MOVN : A64I_movwSizes<0b00, "movn",
3846 (ins movn32_imm:$FullImm),
3847 (ins movn64_imm:$FullImm)>;
3849 // Some relocations are able to convert between a MOVZ and a MOVN. If these
3850 // are applied the instruction must be emitted with the corresponding bits as
3851 // 0, which means a MOVZ needs to override that bit from the default.
3852 let PostEncoderMethod = "fixMOVZ" in
3853 defm MOVZ : A64I_movwSizes<0b10, "movz",
3854 (ins movz32_imm:$FullImm),
3855 (ins movz64_imm:$FullImm)>;
3858 let Constraints = "$src = $Rd" in
3859 defm MOVK : A64I_movwSizes<0b11, "movk",
3860 (ins GPR32:$src, movk32_imm:$FullImm),
3861 (ins GPR64:$src, movk64_imm:$FullImm)>;
3864 // And now the "MOV" aliases. These also need their own operands because what
3865 // they accept is completely different to what the base instructions accept.
3866 multiclass movalias_operand<string prefix, string basename,
3867 string immpredicate, int width> {
3868 def _asmoperand : AsmOperandClass {
3869 let Name = basename # width # "MovAlias";
3871 = "isMoveWideMovAlias<" # width # ", A64Imms::" # immpredicate # ">";
3873 = "addMoveWideMovAliasOperands<" # width # ", "
3874 # "A64Imms::" # immpredicate # ">";
3877 def _movimm : Operand<i32> {
3878 let ParserMatchClass = !cast<AsmOperandClass>(prefix # "_asmoperand");
3880 let MIOperandInfo = (ops uimm16:$UImm16, imm:$Shift);
3884 defm movz32 : movalias_operand<"movz32", "MOVZ", "isMOVZImm", 32>;
3885 defm movz64 : movalias_operand<"movz64", "MOVZ", "isMOVZImm", 64>;
3886 defm movn32 : movalias_operand<"movn32", "MOVN", "isOnlyMOVNImm", 32>;
3887 defm movn64 : movalias_operand<"movn64", "MOVN", "isOnlyMOVNImm", 64>;
3889 // FIXME: these are officially canonical aliases, but TableGen is too limited to
3890 // print them at the moment. I believe in this case an "AliasPredicate" method
3891 // will need to be implemented. to allow it, as well as the more generally
3892 // useful handling of non-register, non-constant operands.
3893 class movalias<Instruction INST, RegisterClass GPR, Operand operand>
3894 : InstAlias<"mov $Rd, $FullImm", (INST GPR:$Rd, operand:$FullImm)>;
3896 def : movalias<MOVZwii, GPR32, movz32_movimm>;
3897 def : movalias<MOVZxii, GPR64, movz64_movimm>;
3898 def : movalias<MOVNwii, GPR32, movn32_movimm>;
3899 def : movalias<MOVNxii, GPR64, movn64_movimm>;
3901 //===----------------------------------------------------------------------===//
3902 // PC-relative addressing instructions
3903 //===----------------------------------------------------------------------===//
3904 // Contains: ADR, ADRP
3906 def adr_label : Operand<i64> {
3907 let EncoderMethod = "getLabelOpValue<AArch64::fixup_a64_adr_prel>";
3909 // This label is a 21-bit offset from PC, unscaled
3910 let PrintMethod = "printLabelOperand<21, 1>";
3911 let ParserMatchClass = label_asmoperand<21, 1>;
3912 let OperandType = "OPERAND_PCREL";
3915 def adrp_label_asmoperand : AsmOperandClass {
3916 let Name = "AdrpLabel";
3917 let RenderMethod = "addLabelOperands<21, 4096>";
3920 def adrp_label : Operand<i64> {
3921 let EncoderMethod = "getAdrpLabelOpValue";
3923 // This label is a 21-bit offset from PC, scaled by the page-size: 4096.
3924 let PrintMethod = "printLabelOperand<21, 4096>";
3925 let ParserMatchClass = adrp_label_asmoperand;
3926 let OperandType = "OPERAND_PCREL";
3929 let neverHasSideEffects = 1 in {
3930 def ADRxi : A64I_PCADR<0b0, (outs GPR64:$Rd), (ins adr_label:$Label),
3931 "adr\t$Rd, $Label", [], NoItinerary>;
3933 def ADRPxi : A64I_PCADR<0b1, (outs GPR64:$Rd), (ins adrp_label:$Label),
3934 "adrp\t$Rd, $Label", [], NoItinerary>;
3937 //===----------------------------------------------------------------------===//
3938 // System instructions
3939 //===----------------------------------------------------------------------===//
3940 // Contains: HINT, CLREX, DSB, DMB, ISB, MSR, SYS, SYSL, MRS
3941 // + aliases IC, DC, AT, TLBI, NOP, YIELD, WFE, WFI, SEV, SEVL
3943 // Op1 and Op2 fields are sometimes simple 3-bit unsigned immediate values.
3944 def uimm3_asmoperand : AsmOperandClass {
3946 let PredicateMethod = "isUImm<3>";
3947 let RenderMethod = "addImmOperands";
3950 def uimm3 : Operand<i32> {
3951 let ParserMatchClass = uimm3_asmoperand;
3954 // The HINT alias can accept a simple unsigned 7-bit immediate.
3955 def uimm7_asmoperand : AsmOperandClass {
3957 let PredicateMethod = "isUImm<7>";
3958 let RenderMethod = "addImmOperands";
3961 def uimm7 : Operand<i32> {
3962 let ParserMatchClass = uimm7_asmoperand;
3965 // Multiclass namedimm is defined with the prefetch operands. Most of these fit
3966 // into the NamedImmMapper scheme well: they either accept a named operand or
3967 // any immediate under a particular value (which may be 0, implying no immediate
3969 defm dbarrier : namedimm<"dbarrier", "A64DB::DBarrierMapper">;
3970 defm isb : namedimm<"isb", "A64ISB::ISBMapper">;
3971 defm ic : namedimm<"ic", "A64IC::ICMapper">;
3972 defm dc : namedimm<"dc", "A64DC::DCMapper">;
3973 defm at : namedimm<"at", "A64AT::ATMapper">;
3974 defm tlbi : namedimm<"tlbi", "A64TLBI::TLBIMapper">;
3976 // However, MRS and MSR are more complicated for a few reasons:
3977 // * There are ~1000 generic names S3_<op1>_<CRn>_<CRm>_<Op2> which have an
3978 // implementation-defined effect
3979 // * Most registers are shared, but some are read-only or write-only.
3980 // * There is a variant of MSR which accepts the same register name (SPSel),
3981 // but which would have a different encoding.
3983 // In principle these could be resolved in with more complicated subclasses of
3984 // NamedImmMapper, however that imposes an overhead on other "named
3985 // immediates". Both in concrete terms with virtual tables and in unnecessary
3988 // The solution adopted here is to take the MRS/MSR Mappers out of the usual
3989 // hierarchy (they're not derived from NamedImmMapper) and to add logic for
3990 // their special situation.
3991 def mrs_asmoperand : AsmOperandClass {
3993 let ParserMethod = "ParseSysRegOperand";
3996 def mrs_op : Operand<i32> {
3997 let ParserMatchClass = mrs_asmoperand;
3998 let PrintMethod = "printMRSOperand";
3999 let DecoderMethod = "DecodeMRSOperand";
4002 def msr_asmoperand : AsmOperandClass {
4003 let Name = "MSRWithReg";
4005 // Note that SPSel is valid for both this and the pstate operands, but with
4006 // different immediate encodings. This is why these operands provide a string
4007 // AArch64Operand rather than an immediate. The overlap is small enough that
4008 // it could be resolved with hackery now, but who can say in future?
4009 let ParserMethod = "ParseSysRegOperand";
4012 def msr_op : Operand<i32> {
4013 let ParserMatchClass = msr_asmoperand;
4014 let PrintMethod = "printMSROperand";
4015 let DecoderMethod = "DecodeMSROperand";
4018 def pstate_asmoperand : AsmOperandClass {
4019 let Name = "MSRPState";
4020 // See comment above about parser.
4021 let ParserMethod = "ParseSysRegOperand";
4024 def pstate_op : Operand<i32> {
4025 let ParserMatchClass = pstate_asmoperand;
4026 let PrintMethod = "printNamedImmOperand<A64PState::PStateMapper>";
4027 let DecoderMethod = "DecodeNamedImmOperand<A64PState::PStateMapper>";
4030 // When <CRn> is specified, an assembler should accept something like "C4", not
4031 // the usual "#4" immediate.
4032 def CRx_asmoperand : AsmOperandClass {
4034 let PredicateMethod = "isUImm<4>";
4035 let RenderMethod = "addImmOperands";
4036 let ParserMethod = "ParseCRxOperand";
4039 def CRx : Operand<i32> {
4040 let ParserMatchClass = CRx_asmoperand;
4041 let PrintMethod = "printCRxOperand";
4045 // Finally, we can start defining the instructions.
4047 // HINT is straightforward, with a few aliases.
4048 def HINTi : A64I_system<0b0, (outs), (ins uimm7:$UImm7), "hint\t$UImm7",
4051 let CRm = UImm7{6-3};
4052 let Op2 = UImm7{2-0};
4060 def : InstAlias<"nop", (HINTi 0)>;
4061 def : InstAlias<"yield", (HINTi 1)>;
4062 def : InstAlias<"wfe", (HINTi 2)>;
4063 def : InstAlias<"wfi", (HINTi 3)>;
4064 def : InstAlias<"sev", (HINTi 4)>;
4065 def : InstAlias<"sevl", (HINTi 5)>;
4067 // Quite a few instructions then follow a similar pattern of fixing common
4068 // fields in the bitpattern, we'll define a helper-class for them.
4069 class simple_sys<bits<2> op0, bits<3> op1, bits<4> crn, bits<3> op2,
4070 Operand operand, string asmop>
4071 : A64I_system<0b0, (outs), (ins operand:$CRm), !strconcat(asmop, "\t$CRm"),
4081 def CLREXi : simple_sys<0b00, 0b011, 0b0011, 0b010, uimm4, "clrex">;
4082 def DSBi : simple_sys<0b00, 0b011, 0b0011, 0b100, dbarrier_op, "dsb">;
4083 def DMBi : simple_sys<0b00, 0b011, 0b0011, 0b101, dbarrier_op, "dmb">;
4084 def ISBi : simple_sys<0b00, 0b011, 0b0011, 0b110, isb_op, "isb">;
4086 def : InstAlias<"clrex", (CLREXi 0b1111)>;
4087 def : InstAlias<"isb", (ISBi 0b1111)>;
4089 // (DMBi 0xb) is a "DMB ISH" instruciton, appropriate for Linux SMP
4090 // configurations at least.
4091 def : Pat<(atomic_fence imm, imm), (DMBi 0xb)>;
4093 // Any SYS bitpattern can be represented with a complex and opaque "SYS"
4095 def SYSiccix : A64I_system<0b0, (outs),
4096 (ins uimm3:$Op1, CRx:$CRn, CRx:$CRm,
4097 uimm3:$Op2, GPR64:$Rt),
4098 "sys\t$Op1, $CRn, $CRm, $Op2, $Rt",
4103 // You can skip the Xt argument whether it makes sense or not for the generic
4105 def : InstAlias<"sys $Op1, $CRn, $CRm, $Op2",
4106 (SYSiccix uimm3:$Op1, CRx:$CRn, CRx:$CRm, uimm3:$Op2, XZR)>;
4109 // But many have aliases, which obviously don't fit into
4110 class SYSalias<dag ins, string asmstring>
4111 : A64I_system<0b0, (outs), ins, asmstring, [], NoItinerary> {
4112 let isAsmParserOnly = 1;
4116 let Op1 = SysOp{13-11};
4117 let CRn = SysOp{10-7};
4118 let CRm = SysOp{6-3};
4119 let Op2 = SysOp{2-0};
4122 def ICix : SYSalias<(ins ic_op:$SysOp, GPR64:$Rt), "ic\t$SysOp, $Rt">;
4124 def ICi : SYSalias<(ins ic_op:$SysOp), "ic\t$SysOp"> {
4128 def DCix : SYSalias<(ins dc_op:$SysOp, GPR64:$Rt), "dc\t$SysOp, $Rt">;
4129 def ATix : SYSalias<(ins at_op:$SysOp, GPR64:$Rt), "at\t$SysOp, $Rt">;
4131 def TLBIix : SYSalias<(ins tlbi_op:$SysOp, GPR64:$Rt), "tlbi\t$SysOp, $Rt">;
4133 def TLBIi : SYSalias<(ins tlbi_op:$SysOp), "tlbi\t$SysOp"> {
4138 def SYSLxicci : A64I_system<0b1, (outs GPR64:$Rt),
4139 (ins uimm3:$Op1, CRx:$CRn, CRx:$CRm, uimm3:$Op2),
4140 "sysl\t$Rt, $Op1, $CRn, $CRm, $Op2",
4145 // The instructions themselves are rather simple for MSR and MRS.
4146 def MSRix : A64I_system<0b0, (outs), (ins msr_op:$SysReg, GPR64:$Rt),
4147 "msr\t$SysReg, $Rt", [], NoItinerary> {
4149 let Op0 = SysReg{15-14};
4150 let Op1 = SysReg{13-11};
4151 let CRn = SysReg{10-7};
4152 let CRm = SysReg{6-3};
4153 let Op2 = SysReg{2-0};
4156 def MRSxi : A64I_system<0b1, (outs GPR64:$Rt), (ins mrs_op:$SysReg),
4157 "mrs\t$Rt, $SysReg", [], NoItinerary> {
4159 let Op0 = SysReg{15-14};
4160 let Op1 = SysReg{13-11};
4161 let CRn = SysReg{10-7};
4162 let CRm = SysReg{6-3};
4163 let Op2 = SysReg{2-0};
4166 def MSRii : A64I_system<0b0, (outs), (ins pstate_op:$PState, uimm4:$CRm),
4167 "msr\t$PState, $CRm", [], NoItinerary> {
4171 let Op1 = PState{5-3};
4173 let Op2 = PState{2-0};
4177 //===----------------------------------------------------------------------===//
4178 // Test & branch (immediate) instructions
4179 //===----------------------------------------------------------------------===//
4180 // Contains: TBZ, TBNZ
4182 // The bit to test is a simple unsigned 6-bit immediate in the X-register
4184 def uimm6 : Operand<i64> {
4185 let ParserMatchClass = uimm6_asmoperand;
4188 def label_wid14_scal4_asmoperand : label_asmoperand<14, 4>;
4190 def tbimm_target : Operand<OtherVT> {
4191 let EncoderMethod = "getLabelOpValue<AArch64::fixup_a64_tstbr>";
4193 // This label is a 14-bit offset from PC, scaled by the instruction-width: 4.
4194 let PrintMethod = "printLabelOperand<14, 4>";
4195 let ParserMatchClass = label_wid14_scal4_asmoperand;
4197 let OperandType = "OPERAND_PCREL";
4200 def A64eq : ImmLeaf<i32, [{ return Imm == A64CC::EQ; }]>;
4201 def A64ne : ImmLeaf<i32, [{ return Imm == A64CC::NE; }]>;
4203 // These instructions correspond to patterns involving "and" with a power of
4204 // two, which we need to be able to select.
4205 def tstb64_pat : ComplexPattern<i64, 1, "SelectTSTBOperand<64>">;
4206 def tstb32_pat : ComplexPattern<i32, 1, "SelectTSTBOperand<32>">;
4208 let isBranch = 1, isTerminator = 1 in {
4209 def TBZxii : A64I_TBimm<0b0, (outs),
4210 (ins GPR64:$Rt, uimm6:$Imm, tbimm_target:$Label),
4211 "tbz\t$Rt, $Imm, $Label",
4212 [(A64br_cc (A64cmp (and GPR64:$Rt, tstb64_pat:$Imm), 0),
4216 def TBNZxii : A64I_TBimm<0b1, (outs),
4217 (ins GPR64:$Rt, uimm6:$Imm, tbimm_target:$Label),
4218 "tbnz\t$Rt, $Imm, $Label",
4219 [(A64br_cc (A64cmp (and GPR64:$Rt, tstb64_pat:$Imm), 0),
4224 // Note, these instructions overlap with the above 64-bit patterns. This is
4225 // intentional, "tbz x3, #1, somewhere" and "tbz w3, #1, somewhere" would both
4226 // do the same thing and are both permitted assembly. They also both have
4227 // sensible DAG patterns.
4228 def TBZwii : A64I_TBimm<0b0, (outs),
4229 (ins GPR32:$Rt, uimm5:$Imm, tbimm_target:$Label),
4230 "tbz\t$Rt, $Imm, $Label",
4231 [(A64br_cc (A64cmp (and GPR32:$Rt, tstb32_pat:$Imm), 0),
4237 def TBNZwii : A64I_TBimm<0b1, (outs),
4238 (ins GPR32:$Rt, uimm5:$Imm, tbimm_target:$Label),
4239 "tbnz\t$Rt, $Imm, $Label",
4240 [(A64br_cc (A64cmp (and GPR32:$Rt, tstb32_pat:$Imm), 0),
4247 //===----------------------------------------------------------------------===//
4248 // Unconditional branch (immediate) instructions
4249 //===----------------------------------------------------------------------===//
4252 def label_wid26_scal4_asmoperand : label_asmoperand<26, 4>;
4254 def bimm_target : Operand<OtherVT> {
4255 let EncoderMethod = "getLabelOpValue<AArch64::fixup_a64_uncondbr>";
4257 // This label is a 26-bit offset from PC, scaled by the instruction-width: 4.
4258 let PrintMethod = "printLabelOperand<26, 4>";
4259 let ParserMatchClass = label_wid26_scal4_asmoperand;
4261 let OperandType = "OPERAND_PCREL";
4264 def blimm_target : Operand<i64> {
4265 let EncoderMethod = "getLabelOpValue<AArch64::fixup_a64_call>";
4267 // This label is a 26-bit offset from PC, scaled by the instruction-width: 4.
4268 let PrintMethod = "printLabelOperand<26, 4>";
4269 let ParserMatchClass = label_wid26_scal4_asmoperand;
4271 let OperandType = "OPERAND_PCREL";
4274 class A64I_BimmImpl<bit op, string asmop, list<dag> patterns, Operand lbl_type>
4275 : A64I_Bimm<op, (outs), (ins lbl_type:$Label),
4276 !strconcat(asmop, "\t$Label"), patterns,
4279 let isBranch = 1 in {
4280 def Bimm : A64I_BimmImpl<0b0, "b", [(br bb:$Label)], bimm_target> {
4281 let isTerminator = 1;
4285 def BLimm : A64I_BimmImpl<0b1, "bl",
4286 [(AArch64Call tglobaladdr:$Label)], blimm_target> {
4292 def : Pat<(AArch64Call texternalsym:$Label), (BLimm texternalsym:$Label)>;
4294 //===----------------------------------------------------------------------===//
4295 // Unconditional branch (register) instructions
4296 //===----------------------------------------------------------------------===//
4297 // Contains: BR, BLR, RET, ERET, DRP.
4299 // Most of the notional opcode fields in the A64I_Breg format are fixed in A64
4301 class A64I_BregImpl<bits<4> opc,
4302 dag outs, dag ins, string asmstr, list<dag> patterns,
4303 InstrItinClass itin = NoItinerary>
4304 : A64I_Breg<opc, 0b11111, 0b000000, 0b00000,
4305 outs, ins, asmstr, patterns, itin> {
4307 let isIndirectBranch = 1;
4310 // Note that these are not marked isCall or isReturn because as far as LLVM is
4311 // concerned they're not. "ret" is just another jump unless it has been selected
4312 // by LLVM as the function's return.
4314 let isBranch = 1 in {
4315 def BRx : A64I_BregImpl<0b0000,(outs), (ins GPR64:$Rn),
4316 "br\t$Rn", [(brind GPR64:$Rn)]> {
4318 let isTerminator = 1;
4321 def BLRx : A64I_BregImpl<0b0001, (outs), (ins GPR64:$Rn),
4322 "blr\t$Rn", [(AArch64Call GPR64:$Rn)]> {
4328 def RETx : A64I_BregImpl<0b0010, (outs), (ins GPR64:$Rn),
4331 let isTerminator = 1;
4335 // Create a separate pseudo-instruction for codegen to use so that we don't
4336 // flag x30 as used in every function. It'll be restored before the RET by the
4337 // epilogue if it's legitimately used.
4338 def RET : A64PseudoExpand<(outs), (ins), [(A64ret)], (RETx (ops X30))> {
4339 let isTerminator = 1;
4344 def ERET : A64I_BregImpl<0b0100, (outs), (ins), "eret", []> {
4347 let isTerminator = 1;
4351 def DRPS : A64I_BregImpl<0b0101, (outs), (ins), "drps", []> {
4357 def RETAlias : InstAlias<"ret", (RETx X30)>;
4360 //===----------------------------------------------------------------------===//
4361 // Address generation patterns
4362 //===----------------------------------------------------------------------===//
4364 // Primary method of address generation for the small/absolute memory model is
4365 // an ADRP/ADR pair:
4366 // ADRP x0, some_variable
4367 // ADD x0, x0, #:lo12:some_variable
4369 // The load/store elision of the ADD is accomplished when selecting
4370 // addressing-modes. This just mops up the cases where that doesn't work and we
4371 // really need an address in some register.
4373 // This wrapper applies a LO12 modifier to the address. Otherwise we could just
4374 // use the same address.
4376 class ADRP_ADD<SDNode Wrapper, SDNode addrop>
4377 : Pat<(Wrapper addrop:$Hi, addrop:$Lo12, (i32 imm)),
4378 (ADDxxi_lsl0_s (ADRPxi addrop:$Hi), addrop:$Lo12)>;
4380 def : ADRP_ADD<A64WrapperSmall, tblockaddress>;
4381 def : ADRP_ADD<A64WrapperSmall, texternalsym>;
4382 def : ADRP_ADD<A64WrapperSmall, tglobaladdr>;
4383 def : ADRP_ADD<A64WrapperSmall, tglobaltlsaddr>;
4384 def : ADRP_ADD<A64WrapperSmall, tjumptable>;
4386 //===----------------------------------------------------------------------===//
4387 // GOT access patterns
4388 //===----------------------------------------------------------------------===//
4392 class GOTLoadSmall<SDNode addrfrag>
4393 : Pat<(A64GOTLoad (A64WrapperSmall addrfrag:$Hi, addrfrag:$Lo12, 8)),
4394 (LS64_LDR (ADRPxi addrfrag:$Hi), addrfrag:$Lo12)>;
4396 def : GOTLoadSmall<texternalsym>;
4397 def : GOTLoadSmall<tglobaladdr>;
4398 def : GOTLoadSmall<tglobaltlsaddr>;
4400 //===----------------------------------------------------------------------===//
4401 // Tail call handling
4402 //===----------------------------------------------------------------------===//
4404 let isCall = 1, isTerminator = 1, isReturn = 1, isBarrier = 1, Uses = [XSP] in {
4406 : PseudoInst<(outs), (ins i64imm:$dst, i32imm:$FPDiff),
4407 [(AArch64tcret tglobaladdr:$dst, (i32 timm:$FPDiff))]>;
4410 : PseudoInst<(outs), (ins tcGPR64:$dst, i32imm:$FPDiff),
4411 [(AArch64tcret tcGPR64:$dst, (i32 timm:$FPDiff))]>;
4414 let isCall = 1, isTerminator = 1, isReturn = 1, isBarrier = 1,
4416 def TAIL_Bimm : A64PseudoExpand<(outs), (ins bimm_target:$Label), [],
4417 (Bimm bimm_target:$Label)>;
4419 def TAIL_BRx : A64PseudoExpand<(outs), (ins tcGPR64:$Rd), [],
4424 def : Pat<(AArch64tcret texternalsym:$dst, (i32 timm:$FPDiff)),
4425 (TC_RETURNdi texternalsym:$dst, imm:$FPDiff)>;
4427 //===----------------------------------------------------------------------===//
4428 // Thread local storage
4429 //===----------------------------------------------------------------------===//
4431 // This is a pseudo-instruction representing the ".tlsdesccall" directive in
4432 // assembly. Its effect is to insert an R_AARCH64_TLSDESC_CALL relocation at the
4433 // current location. It should always be immediately followed by a BLR
4434 // instruction, and is intended solely for relaxation by the linker.
4436 def : Pat<(A64threadpointer), (MRSxi 0xde82)>;
4438 def TLSDESCCALL : PseudoInst<(outs), (ins i64imm:$Lbl), []> {
4439 let hasSideEffects = 1;
4442 def TLSDESC_BLRx : PseudoInst<(outs), (ins GPR64:$Rn, i64imm:$Var),
4443 [(A64tlsdesc_blr GPR64:$Rn, tglobaltlsaddr:$Var)]> {
4448 def : Pat<(A64tlsdesc_blr GPR64:$Rn, texternalsym:$Var),
4449 (TLSDESC_BLRx GPR64:$Rn, texternalsym:$Var)>;
4451 //===----------------------------------------------------------------------===//
4452 // Bitfield patterns
4453 //===----------------------------------------------------------------------===//
4455 def bfi32_lsb_to_immr : SDNodeXForm<imm, [{
4456 return CurDAG->getTargetConstant((32 - N->getZExtValue()) % 32, MVT::i64);
4459 def bfi64_lsb_to_immr : SDNodeXForm<imm, [{
4460 return CurDAG->getTargetConstant((64 - N->getZExtValue()) % 64, MVT::i64);
4463 def bfi_width_to_imms : SDNodeXForm<imm, [{
4464 return CurDAG->getTargetConstant(N->getZExtValue() - 1, MVT::i64);
4468 // The simpler patterns deal with cases where no AND mask is actually needed
4469 // (either all bits are used or the low 32 bits are used).
4470 let AddedComplexity = 10 in {
4472 def : Pat<(A64Bfi GPR64:$src, GPR64:$Rn, imm:$ImmR, imm:$ImmS),
4473 (BFIxxii GPR64:$src, GPR64:$Rn,
4474 (bfi64_lsb_to_immr (i64 imm:$ImmR)),
4475 (bfi_width_to_imms (i64 imm:$ImmS)))>;
4477 def : Pat<(A64Bfi GPR32:$src, GPR32:$Rn, imm:$ImmR, imm:$ImmS),
4478 (BFIwwii GPR32:$src, GPR32:$Rn,
4479 (bfi32_lsb_to_immr (i64 imm:$ImmR)),
4480 (bfi_width_to_imms (i64 imm:$ImmS)))>;
4483 def : Pat<(and (A64Bfi GPR64:$src, GPR64:$Rn, imm:$ImmR, imm:$ImmS),
4485 (SUBREG_TO_REG (i64 0),
4486 (BFIwwii (EXTRACT_SUBREG GPR64:$src, sub_32),
4487 (EXTRACT_SUBREG GPR64:$Rn, sub_32),
4488 (bfi32_lsb_to_immr (i64 imm:$ImmR)),
4489 (bfi_width_to_imms (i64 imm:$ImmS))),
4494 //===----------------------------------------------------------------------===//
4495 // Constant island entries
4496 //===----------------------------------------------------------------------===//
4498 // The constant island pass needs to create "instructions" in the middle of the
4499 // instruction stream to reresent its constants.
4501 def cpinst_operand : Operand<i32>;
4503 def CONSTPOOL_ENTRY : PseudoInst<(outs), (ins cpinst_operand:$instid,
4504 cpinst_operand:$cpidx,
4505 i32imm:$size), []> {
4506 let neverHasSideEffects = 1;
4507 let isNotDuplicable = 1;
4510 //===----------------------------------------------------------------------===//
4511 // Miscellaneous patterns
4512 //===----------------------------------------------------------------------===//
4514 // Truncation from 64 to 32-bits just involves renaming your register.
4515 def : Pat<(i32 (trunc (i64 GPR64:$val))), (EXTRACT_SUBREG GPR64:$val, sub_32)>;
4517 // Similarly, extension where we don't care about the high bits is
4519 def : Pat<(i64 (anyext (i32 GPR32:$val))),
4520 (INSERT_SUBREG (IMPLICIT_DEF), GPR32:$val, sub_32)>;
4522 // SELECT instructions providing f128 types need to be handled by a
4523 // pseudo-instruction since the eventual code will need to introduce basic
4524 // blocks and control flow.
4525 def F128CSEL : PseudoInst<(outs FPR128:$Rd),
4526 (ins FPR128:$Rn, FPR128:$Rm, cond_code_op:$Cond),
4527 [(set FPR128:$Rd, (simple_select (f128 FPR128:$Rn),
4530 let usesCustomInserter = 1;
4533 //===----------------------------------------------------------------------===//
4534 // Load/store patterns
4535 //===----------------------------------------------------------------------===//
4537 // There are lots of patterns here, because we need to allow at least three
4538 // parameters to vary independently.
4539 // 1. Instruction: "ldrb w9, [sp]", "ldrh w9, [sp]", ...
4540 // 2. LLVM source: zextloadi8, anyextloadi8, ...
4541 // 3. Address-generation: A64Wrapper, (add BASE, OFFSET), ...
4543 // The biggest problem turns out to be the address-generation variable. At the
4544 // point of instantiation we need to produce two DAGs, one for the pattern and
4545 // one for the instruction. Doing this at the lowest level of classes doesn't
4548 // Consider the simple uimm12 addressing mode, and the desire to match both (add
4549 // GPR64xsp:$Rn, uimm12:$Offset) and GPR64xsp:$Rn, particularly on the
4550 // instruction side. We'd need to insert either "GPR64xsp" and "uimm12" or
4551 // "GPR64xsp" and "0" into an unknown dag. !subst is not capable of this
4552 // operation, and PatFrags are for selection not output.
4554 // As a result, the address-generation patterns are the final
4555 // instantiations. However, we do still need to vary the operand for the address
4556 // further down (At the point we're deciding A64WrapperSmall, we don't know
4557 // the memory width of the operation).
4559 //===------------------------------
4560 // 1. Basic infrastructural defs
4561 //===------------------------------
4563 // First, some simple classes for !foreach and !subst to use:
4574 // You can't use !subst on an actual immediate, but you *can* use it on an
4575 // operand record that happens to match a single immediate. So we do.
4576 def imm_eq0 : ImmLeaf<i64, [{ return Imm == 0; }]>;
4577 def imm_eq1 : ImmLeaf<i64, [{ return Imm == 1; }]>;
4578 def imm_eq2 : ImmLeaf<i64, [{ return Imm == 2; }]>;
4579 def imm_eq3 : ImmLeaf<i64, [{ return Imm == 3; }]>;
4580 def imm_eq4 : ImmLeaf<i64, [{ return Imm == 4; }]>;
4582 // If the low bits of a pointer are known to be 0 then an "or" is just as good
4583 // as addition for computing an offset. This fragment forwards that check for
4585 def add_like_or : PatFrag<(ops node:$lhs, node:$rhs), (or node:$lhs, node:$rhs),
4587 return CurDAG->isBaseWithConstantOffset(SDValue(N, 0));
4590 // Load/store (unsigned immediate) operations with relocations against global
4591 // symbols (for lo12) are only valid if those symbols have correct alignment
4592 // (since the immediate offset is divided by the access scale, it can't have a
4595 // The guaranteed alignment is provided as part of the WrapperSmall
4596 // operation, and checked against one of these.
4597 def any_align : ImmLeaf<i32, [{ (void)Imm; return true; }]>;
4598 def min_align2 : ImmLeaf<i32, [{ return Imm >= 2; }]>;
4599 def min_align4 : ImmLeaf<i32, [{ return Imm >= 4; }]>;
4600 def min_align8 : ImmLeaf<i32, [{ return Imm >= 8; }]>;
4601 def min_align16 : ImmLeaf<i32, [{ return Imm >= 16; }]>;
4603 // "Normal" load/store instructions can be used on atomic operations, provided
4604 // the ordering parameter is at most "monotonic". Anything above that needs
4605 // special handling with acquire/release instructions.
4606 class simple_load<PatFrag base>
4607 : PatFrag<(ops node:$ptr), (base node:$ptr), [{
4608 return cast<AtomicSDNode>(N)->getOrdering() <= Monotonic;
4611 def atomic_load_simple_i8 : simple_load<atomic_load_8>;
4612 def atomic_load_simple_i16 : simple_load<atomic_load_16>;
4613 def atomic_load_simple_i32 : simple_load<atomic_load_32>;
4614 def atomic_load_simple_i64 : simple_load<atomic_load_64>;
4616 class simple_store<PatFrag base>
4617 : PatFrag<(ops node:$ptr, node:$val), (base node:$ptr, node:$val), [{
4618 return cast<AtomicSDNode>(N)->getOrdering() <= Monotonic;
4621 def atomic_store_simple_i8 : simple_store<atomic_store_8>;
4622 def atomic_store_simple_i16 : simple_store<atomic_store_16>;
4623 def atomic_store_simple_i32 : simple_store<atomic_store_32>;
4624 def atomic_store_simple_i64 : simple_store<atomic_store_64>;
4626 //===------------------------------
4627 // 2. UImm12 and SImm9
4628 //===------------------------------
4630 // These instructions have two operands providing the address so they can be
4631 // treated similarly for most purposes.
4633 //===------------------------------
4634 // 2.1 Base patterns covering extend/truncate semantics
4635 //===------------------------------
4637 // Atomic patterns can be shared between integer operations of all sizes, a
4638 // quick multiclass here allows reuse.
4639 multiclass ls_atomic_pats<Instruction LOAD, Instruction STORE, dag Base,
4640 dag Offset, dag address, RegisterClass TPR,
4642 def : Pat<(!cast<PatFrag>("atomic_load_simple_" # sty) address),
4643 (LOAD Base, Offset)>;
4645 def : Pat<(!cast<PatFrag>("atomic_store_simple_" # sty) address, TPR:$Rt),
4646 (STORE TPR:$Rt, Base, Offset)>;
4649 // Instructions accessing a memory chunk smaller than a register (or, in a
4650 // pinch, the same size) have a characteristic set of patterns they want to
4651 // match: extending loads and truncating stores. This class deals with the
4652 // sign-neutral version of those patterns.
4654 // It will be instantiated across multiple addressing-modes.
4655 multiclass ls_small_pats<Instruction LOAD, Instruction STORE,
4656 dag Base, dag Offset,
4657 dag address, ValueType sty>
4658 : ls_atomic_pats<LOAD, STORE, Base, Offset, address, GPR32, sty> {
4659 def : Pat<(!cast<SDNode>(zextload # sty) address), (LOAD Base, Offset)>;
4661 def : Pat<(!cast<SDNode>(extload # sty) address), (LOAD Base, Offset)>;
4663 // For zero-extension to 64-bits we have to tell LLVM that the whole 64-bit
4664 // register was actually set.
4665 def : Pat<(i64 (!cast<SDNode>(zextload # sty) address)),
4666 (SUBREG_TO_REG (i64 0), (LOAD Base, Offset), sub_32)>;
4668 def : Pat<(i64 (!cast<SDNode>(extload # sty) address)),
4669 (SUBREG_TO_REG (i64 0), (LOAD Base, Offset), sub_32)>;
4671 def : Pat<(!cast<SDNode>(truncstore # sty) GPR32:$Rt, address),
4672 (STORE GPR32:$Rt, Base, Offset)>;
4674 // For truncating store from 64-bits, we have to manually tell LLVM to
4675 // ignore the high bits of the x register.
4676 def : Pat<(!cast<SDNode>(truncstore # sty) GPR64:$Rt, address),
4677 (STORE (EXTRACT_SUBREG GPR64:$Rt, sub_32), Base, Offset)>;
4680 // Next come patterns for sign-extending loads.
4681 multiclass load_signed_pats<string T, string U, dag Base, dag Offset,
4682 dag address, ValueType sty> {
4683 def : Pat<(i32 (!cast<SDNode>("sextload" # sty) address)),
4684 (!cast<Instruction>("LDRS" # T # "w" # U) Base, Offset)>;
4686 def : Pat<(i64 (!cast<SDNode>("sextload" # sty) address)),
4687 (!cast<Instruction>("LDRS" # T # "x" # U) Base, Offset)>;
4691 // and finally "natural-width" loads and stores come next.
4692 multiclass ls_neutral_pats<Instruction LOAD, Instruction STORE, dag Base,
4693 dag Offset, dag address, RegisterClass TPR,
4695 def : Pat<(sty (load address)), (LOAD Base, Offset)>;
4696 def : Pat<(store (sty TPR:$Rt), address), (STORE TPR:$Rt, Base, Offset)>;
4699 // Integer operations also get atomic instructions to select for.
4700 multiclass ls_int_neutral_pats<Instruction LOAD, Instruction STORE, dag Base,
4701 dag Offset, dag address, RegisterClass TPR,
4703 : ls_neutral_pats<LOAD, STORE, Base, Offset, address, TPR, sty>,
4704 ls_atomic_pats<LOAD, STORE, Base, Offset, address, TPR, sty>;
4706 //===------------------------------
4707 // 2.2. Addressing-mode instantiations
4708 //===------------------------------
4710 multiclass uimm12_pats<dag address, dag Base, dag Offset> {
4711 defm : ls_small_pats<LS8_LDR, LS8_STR, Base,
4712 !foreach(decls.pattern, Offset,
4713 !subst(OFFSET, byte_uimm12, decls.pattern)),
4714 !foreach(decls.pattern, address,
4715 !subst(OFFSET, byte_uimm12,
4716 !subst(ALIGN, any_align, decls.pattern))),
4718 defm : ls_small_pats<LS16_LDR, LS16_STR, Base,
4719 !foreach(decls.pattern, Offset,
4720 !subst(OFFSET, hword_uimm12, decls.pattern)),
4721 !foreach(decls.pattern, address,
4722 !subst(OFFSET, hword_uimm12,
4723 !subst(ALIGN, min_align2, decls.pattern))),
4725 defm : ls_small_pats<LS32_LDR, LS32_STR, Base,
4726 !foreach(decls.pattern, Offset,
4727 !subst(OFFSET, word_uimm12, decls.pattern)),
4728 !foreach(decls.pattern, address,
4729 !subst(OFFSET, word_uimm12,
4730 !subst(ALIGN, min_align4, decls.pattern))),
4733 defm : ls_int_neutral_pats<LS32_LDR, LS32_STR, Base,
4734 !foreach(decls.pattern, Offset,
4735 !subst(OFFSET, word_uimm12, decls.pattern)),
4736 !foreach(decls.pattern, address,
4737 !subst(OFFSET, word_uimm12,
4738 !subst(ALIGN, min_align4, decls.pattern))),
4741 defm : ls_int_neutral_pats<LS64_LDR, LS64_STR, Base,
4742 !foreach(decls.pattern, Offset,
4743 !subst(OFFSET, dword_uimm12, decls.pattern)),
4744 !foreach(decls.pattern, address,
4745 !subst(OFFSET, dword_uimm12,
4746 !subst(ALIGN, min_align8, decls.pattern))),
4749 defm : ls_neutral_pats<LSFP16_LDR, LSFP16_STR, Base,
4750 !foreach(decls.pattern, Offset,
4751 !subst(OFFSET, hword_uimm12, decls.pattern)),
4752 !foreach(decls.pattern, address,
4753 !subst(OFFSET, hword_uimm12,
4754 !subst(ALIGN, min_align2, decls.pattern))),
4757 defm : ls_neutral_pats<LSFP32_LDR, LSFP32_STR, Base,
4758 !foreach(decls.pattern, Offset,
4759 !subst(OFFSET, word_uimm12, decls.pattern)),
4760 !foreach(decls.pattern, address,
4761 !subst(OFFSET, word_uimm12,
4762 !subst(ALIGN, min_align4, decls.pattern))),
4765 defm : ls_neutral_pats<LSFP64_LDR, LSFP64_STR, Base,
4766 !foreach(decls.pattern, Offset,
4767 !subst(OFFSET, dword_uimm12, decls.pattern)),
4768 !foreach(decls.pattern, address,
4769 !subst(OFFSET, dword_uimm12,
4770 !subst(ALIGN, min_align8, decls.pattern))),
4773 defm : ls_neutral_pats<LSFP128_LDR, LSFP128_STR, Base,
4774 !foreach(decls.pattern, Offset,
4775 !subst(OFFSET, qword_uimm12, decls.pattern)),
4776 !foreach(decls.pattern, address,
4777 !subst(OFFSET, qword_uimm12,
4778 !subst(ALIGN, min_align16, decls.pattern))),
4781 defm : load_signed_pats<"B", "", Base,
4782 !foreach(decls.pattern, Offset,
4783 !subst(OFFSET, byte_uimm12, decls.pattern)),
4784 !foreach(decls.pattern, address,
4785 !subst(OFFSET, byte_uimm12,
4786 !subst(ALIGN, any_align, decls.pattern))),
4789 defm : load_signed_pats<"H", "", Base,
4790 !foreach(decls.pattern, Offset,
4791 !subst(OFFSET, hword_uimm12, decls.pattern)),
4792 !foreach(decls.pattern, address,
4793 !subst(OFFSET, hword_uimm12,
4794 !subst(ALIGN, min_align2, decls.pattern))),
4797 def : Pat<(sextloadi32 !foreach(decls.pattern, address,
4798 !subst(OFFSET, word_uimm12,
4799 !subst(ALIGN, min_align4, decls.pattern)))),
4800 (LDRSWx Base, !foreach(decls.pattern, Offset,
4801 !subst(OFFSET, word_uimm12, decls.pattern)))>;
4804 // Straightforward patterns of last resort: a pointer with or without an
4805 // appropriate offset.
4806 defm : uimm12_pats<(i64 GPR64xsp:$Rn), (i64 GPR64xsp:$Rn), (i64 0)>;
4807 defm : uimm12_pats<(add GPR64xsp:$Rn, OFFSET:$UImm12),
4808 (i64 GPR64xsp:$Rn), (i64 OFFSET:$UImm12)>;
4810 // The offset could be hidden behind an "or", of course:
4811 defm : uimm12_pats<(add_like_or GPR64xsp:$Rn, OFFSET:$UImm12),
4812 (i64 GPR64xsp:$Rn), (i64 OFFSET:$UImm12)>;
4814 // Global addresses under the small-absolute model should use these
4815 // instructions. There are ELF relocations specifically for it.
4816 defm : uimm12_pats<(A64WrapperSmall tglobaladdr:$Hi, tglobaladdr:$Lo12, ALIGN),
4817 (ADRPxi tglobaladdr:$Hi), (i64 tglobaladdr:$Lo12)>;
4819 defm : uimm12_pats<(A64WrapperSmall tglobaltlsaddr:$Hi, tglobaltlsaddr:$Lo12,
4821 (ADRPxi tglobaltlsaddr:$Hi), (i64 tglobaltlsaddr:$Lo12)>;
4823 // External symbols that make it this far should also get standard relocations.
4824 defm : uimm12_pats<(A64WrapperSmall texternalsym:$Hi, texternalsym:$Lo12,
4826 (ADRPxi texternalsym:$Hi), (i64 texternalsym:$Lo12)>;
4829 // We also want to use uimm12 instructions for local variables at the moment.
4830 def tframeindex_XFORM : SDNodeXForm<frameindex, [{
4831 int FI = cast<FrameIndexSDNode>(N)->getIndex();
4832 return CurDAG->getTargetFrameIndex(FI, MVT::i64);
4835 defm : uimm12_pats<(i64 frameindex:$Rn),
4836 (tframeindex_XFORM tframeindex:$Rn), (i64 0)>;
4838 // These can be much simpler than uimm12 because we don't to change the operand
4839 // type (e.g. LDURB and LDURH take the same operands).
4840 multiclass simm9_pats<dag address, dag Base, dag Offset> {
4841 defm : ls_small_pats<LS8_LDUR, LS8_STUR, Base, Offset, address, i8>;
4842 defm : ls_small_pats<LS16_LDUR, LS16_STUR, Base, Offset, address, i16>;
4844 defm : ls_int_neutral_pats<LS32_LDUR, LS32_STUR, Base, Offset, address,
4846 defm : ls_int_neutral_pats<LS64_LDUR, LS64_STUR, Base, Offset, address,
4849 defm : ls_neutral_pats<LSFP16_LDUR, LSFP16_STUR, Base, Offset, address,
4851 defm : ls_neutral_pats<LSFP32_LDUR, LSFP32_STUR, Base, Offset, address,
4853 defm : ls_neutral_pats<LSFP64_LDUR, LSFP64_STUR, Base, Offset, address,
4855 defm : ls_neutral_pats<LSFP128_LDUR, LSFP128_STUR, Base, Offset, address,
4858 def : Pat<(i64 (zextloadi32 address)),
4859 (SUBREG_TO_REG (i64 0), (LS32_LDUR Base, Offset), sub_32)>;
4861 def : Pat<(truncstorei32 GPR64:$Rt, address),
4862 (LS32_STUR (EXTRACT_SUBREG GPR64:$Rt, sub_32), Base, Offset)>;
4864 defm : load_signed_pats<"B", "_U", Base, Offset, address, i8>;
4865 defm : load_signed_pats<"H", "_U", Base, Offset, address, i16>;
4866 def : Pat<(sextloadi32 address), (LDURSWx Base, Offset)>;
4869 defm : simm9_pats<(add GPR64xsp:$Rn, simm9:$SImm9),
4870 (i64 GPR64xsp:$Rn), (SDXF_simm9 simm9:$SImm9)>;
4872 defm : simm9_pats<(add_like_or GPR64xsp:$Rn, simm9:$SImm9),
4873 (i64 GPR64xsp:$Rn), (SDXF_simm9 simm9:$SImm9)>;
4876 //===------------------------------
4877 // 3. Register offset patterns
4878 //===------------------------------
4880 // Atomic patterns can be shared between integer operations of all sizes, a
4881 // quick multiclass here allows reuse.
4882 multiclass ro_atomic_pats<Instruction LOAD, Instruction STORE, dag Base,
4883 dag Offset, dag Extend, dag address,
4884 RegisterClass TPR, ValueType sty> {
4885 def : Pat<(!cast<PatFrag>("atomic_load_simple_" # sty) address),
4886 (LOAD Base, Offset, Extend)>;
4888 def : Pat<(!cast<PatFrag>("atomic_store_simple_" # sty) address, TPR:$Rt),
4889 (STORE TPR:$Rt, Base, Offset, Extend)>;
4892 // The register offset instructions take three operands giving the instruction,
4893 // and have an annoying split between instructions where Rm is 32-bit and
4894 // 64-bit. So we need a special hierarchy to describe them. Other than that the
4895 // same operations should be supported as for simm9 and uimm12 addressing.
4897 multiclass ro_small_pats<Instruction LOAD, Instruction STORE,
4898 dag Base, dag Offset, dag Extend,
4899 dag address, ValueType sty>
4900 : ro_atomic_pats<LOAD, STORE, Base, Offset, Extend, address, GPR32, sty> {
4901 def : Pat<(!cast<SDNode>(zextload # sty) address),
4902 (LOAD Base, Offset, Extend)>;
4904 def : Pat<(!cast<SDNode>(extload # sty) address),
4905 (LOAD Base, Offset, Extend)>;
4907 // For zero-extension to 64-bits we have to tell LLVM that the whole 64-bit
4908 // register was actually set.
4909 def : Pat<(i64 (!cast<SDNode>(zextload # sty) address)),
4910 (SUBREG_TO_REG (i64 0), (LOAD Base, Offset, Extend), sub_32)>;
4912 def : Pat<(i64 (!cast<SDNode>(extload # sty) address)),
4913 (SUBREG_TO_REG (i64 0), (LOAD Base, Offset, Extend), sub_32)>;
4915 def : Pat<(!cast<SDNode>(truncstore # sty) GPR32:$Rt, address),
4916 (STORE GPR32:$Rt, Base, Offset, Extend)>;
4918 // For truncating store from 64-bits, we have to manually tell LLVM to
4919 // ignore the high bits of the x register.
4920 def : Pat<(!cast<SDNode>(truncstore # sty) GPR64:$Rt, address),
4921 (STORE (EXTRACT_SUBREG GPR64:$Rt, sub_32), Base, Offset, Extend)>;
4925 // Next come patterns for sign-extending loads.
4926 multiclass ro_signed_pats<string T, string Rm, dag Base, dag Offset, dag Extend,
4927 dag address, ValueType sty> {
4928 def : Pat<(i32 (!cast<SDNode>("sextload" # sty) address)),
4929 (!cast<Instruction>("LDRS" # T # "w_" # Rm # "_RegOffset")
4930 Base, Offset, Extend)>;
4932 def : Pat<(i64 (!cast<SDNode>("sextload" # sty) address)),
4933 (!cast<Instruction>("LDRS" # T # "x_" # Rm # "_RegOffset")
4934 Base, Offset, Extend)>;
4937 // and finally "natural-width" loads and stores come next.
4938 multiclass ro_neutral_pats<Instruction LOAD, Instruction STORE,
4939 dag Base, dag Offset, dag Extend, dag address,
4940 RegisterClass TPR, ValueType sty> {
4941 def : Pat<(sty (load address)), (LOAD Base, Offset, Extend)>;
4942 def : Pat<(store (sty TPR:$Rt), address),
4943 (STORE TPR:$Rt, Base, Offset, Extend)>;
4946 multiclass ro_int_neutral_pats<Instruction LOAD, Instruction STORE,
4947 dag Base, dag Offset, dag Extend, dag address,
4948 RegisterClass TPR, ValueType sty>
4949 : ro_neutral_pats<LOAD, STORE, Base, Offset, Extend, address, TPR, sty>,
4950 ro_atomic_pats<LOAD, STORE, Base, Offset, Extend, address, TPR, sty>;
4952 multiclass regoff_pats<string Rm, dag address, dag Base, dag Offset,
4954 defm : ro_small_pats<!cast<Instruction>("LS8_" # Rm # "_RegOffset_LDR"),
4955 !cast<Instruction>("LS8_" # Rm # "_RegOffset_STR"),
4956 Base, Offset, Extend,
4957 !foreach(decls.pattern, address,
4958 !subst(SHIFT, imm_eq0, decls.pattern)),
4960 defm : ro_small_pats<!cast<Instruction>("LS16_" # Rm # "_RegOffset_LDR"),
4961 !cast<Instruction>("LS16_" # Rm # "_RegOffset_STR"),
4962 Base, Offset, Extend,
4963 !foreach(decls.pattern, address,
4964 !subst(SHIFT, imm_eq1, decls.pattern)),
4966 defm : ro_small_pats<!cast<Instruction>("LS32_" # Rm # "_RegOffset_LDR"),
4967 !cast<Instruction>("LS32_" # Rm # "_RegOffset_STR"),
4968 Base, Offset, Extend,
4969 !foreach(decls.pattern, address,
4970 !subst(SHIFT, imm_eq2, decls.pattern)),
4973 defm : ro_int_neutral_pats<
4974 !cast<Instruction>("LS32_" # Rm # "_RegOffset_LDR"),
4975 !cast<Instruction>("LS32_" # Rm # "_RegOffset_STR"),
4976 Base, Offset, Extend,
4977 !foreach(decls.pattern, address,
4978 !subst(SHIFT, imm_eq2, decls.pattern)),
4981 defm : ro_int_neutral_pats<
4982 !cast<Instruction>("LS64_" # Rm # "_RegOffset_LDR"),
4983 !cast<Instruction>("LS64_" # Rm # "_RegOffset_STR"),
4984 Base, Offset, Extend,
4985 !foreach(decls.pattern, address,
4986 !subst(SHIFT, imm_eq3, decls.pattern)),
4989 defm : ro_neutral_pats<!cast<Instruction>("LSFP16_" # Rm # "_RegOffset_LDR"),
4990 !cast<Instruction>("LSFP16_" # Rm # "_RegOffset_STR"),
4991 Base, Offset, Extend,
4992 !foreach(decls.pattern, address,
4993 !subst(SHIFT, imm_eq1, decls.pattern)),
4996 defm : ro_neutral_pats<!cast<Instruction>("LSFP32_" # Rm # "_RegOffset_LDR"),
4997 !cast<Instruction>("LSFP32_" # Rm # "_RegOffset_STR"),
4998 Base, Offset, Extend,
4999 !foreach(decls.pattern, address,
5000 !subst(SHIFT, imm_eq2, decls.pattern)),
5003 defm : ro_neutral_pats<!cast<Instruction>("LSFP64_" # Rm # "_RegOffset_LDR"),
5004 !cast<Instruction>("LSFP64_" # Rm # "_RegOffset_STR"),
5005 Base, Offset, Extend,
5006 !foreach(decls.pattern, address,
5007 !subst(SHIFT, imm_eq3, decls.pattern)),
5010 defm : ro_neutral_pats<!cast<Instruction>("LSFP128_" # Rm # "_RegOffset_LDR"),
5011 !cast<Instruction>("LSFP128_" # Rm # "_RegOffset_STR"),
5012 Base, Offset, Extend,
5013 !foreach(decls.pattern, address,
5014 !subst(SHIFT, imm_eq4, decls.pattern)),
5017 defm : ro_signed_pats<"B", Rm, Base, Offset, Extend,
5018 !foreach(decls.pattern, address,
5019 !subst(SHIFT, imm_eq0, decls.pattern)),
5022 defm : ro_signed_pats<"H", Rm, Base, Offset, Extend,
5023 !foreach(decls.pattern, address,
5024 !subst(SHIFT, imm_eq1, decls.pattern)),
5027 def : Pat<(sextloadi32 !foreach(decls.pattern, address,
5028 !subst(SHIFT, imm_eq2, decls.pattern))),
5029 (!cast<Instruction>("LDRSWx_" # Rm # "_RegOffset")
5030 Base, Offset, Extend)>;
5034 // Finally we're in a position to tell LLVM exactly what addresses are reachable
5035 // using register-offset instructions. Essentially a base plus a possibly
5036 // extended, possibly shifted (by access size) offset.
5038 defm : regoff_pats<"Wm", (add GPR64xsp:$Rn, (sext GPR32:$Rm)),
5039 (i64 GPR64xsp:$Rn), (i32 GPR32:$Rm), (i64 6)>;
5041 defm : regoff_pats<"Wm", (add GPR64xsp:$Rn, (shl (sext GPR32:$Rm), SHIFT)),
5042 (i64 GPR64xsp:$Rn), (i32 GPR32:$Rm), (i64 7)>;
5044 defm : regoff_pats<"Wm", (add GPR64xsp:$Rn, (zext GPR32:$Rm)),
5045 (i64 GPR64xsp:$Rn), (i32 GPR32:$Rm), (i64 2)>;
5047 defm : regoff_pats<"Wm", (add GPR64xsp:$Rn, (shl (zext GPR32:$Rm), SHIFT)),
5048 (i64 GPR64xsp:$Rn), (i32 GPR32:$Rm), (i64 3)>;
5050 defm : regoff_pats<"Xm", (add GPR64xsp:$Rn, GPR64:$Rm),
5051 (i64 GPR64xsp:$Rn), (i64 GPR64:$Rm), (i64 2)>;
5053 defm : regoff_pats<"Xm", (add GPR64xsp:$Rn, (shl GPR64:$Rm, SHIFT)),
5054 (i64 GPR64xsp:$Rn), (i64 GPR64:$Rm), (i64 3)>;