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,
11 // (ins NZCV, Condition, Dest)
12 def SDT_A64br_cc : SDTypeProfile<0, 3, [SDTCisVT<0, i32>]>;
13 def A64br_cc : SDNode<"AArch64ISD::BR_CC", SDT_A64br_cc, [SDNPHasChain]>;
15 // (outs Result), (ins NZCV, IfTrue, IfFalse, Condition)
16 def SDT_A64select_cc : SDTypeProfile<1, 4, [SDTCisVT<1, i32>,
19 def A64select_cc : SDNode<"AArch64ISD::SELECT_CC", SDT_A64select_cc>;
21 // (outs NZCV), (ins LHS, RHS, Condition)
22 def SDT_A64setcc : SDTypeProfile<1, 3, [SDTCisVT<0, i32>,
24 def A64setcc : SDNode<"AArch64ISD::SETCC", SDT_A64setcc>;
27 // (outs GPR64), (ins)
28 def A64threadpointer : SDNode<"AArch64ISD::THREAD_POINTER", SDTPtrLeaf>;
30 // A64 compares don't care about the cond really (they set all flags) so a
31 // simple binary operator is useful.
32 def A64cmp : PatFrag<(ops node:$lhs, node:$rhs),
33 (A64setcc node:$lhs, node:$rhs, cond)>;
36 // When matching a notional (CMP op1, (sub 0, op2)), we'd like to use a CMN
37 // instruction on the grounds that "op1 - (-op2) == op1 + op2". However, the C
38 // and V flags can be set differently by this operation. It comes down to
39 // whether "SInt(~op2)+1 == SInt(~op2+1)" (and the same for UInt). If they are
40 // then everything is fine. If not then the optimization is wrong. Thus general
41 // comparisons are only valid if op2 != 0.
43 // So, finally, the only LLVM-native comparisons that don't mention C and V are
44 // SETEQ and SETNE. They're the only ones we can safely use CMN for in the
45 // absence of information about op2.
46 def equality_cond : PatLeaf<(cond), [{
47 return N->get() == ISD::SETEQ || N->get() == ISD::SETNE;
50 def A64cmn : PatFrag<(ops node:$lhs, node:$rhs),
51 (A64setcc node:$lhs, (sub 0, node:$rhs), equality_cond)>;
53 // There are two layers of indirection here, driven by the following
55 // + TableGen does not know CodeModel or Reloc so that decision should be
56 // made for a variable/address at ISelLowering.
57 // + The output of ISelLowering should be selectable (hence the Wrapper,
58 // rather than a bare target opcode)
59 def SDTAArch64Wrapper : SDTypeProfile<1, 3, [SDTCisSameAs<0, 1>,
64 def A64WrapperSmall : SDNode<"AArch64ISD::WrapperSmall", SDTAArch64Wrapper>;
67 def SDTAArch64GOTLoad : SDTypeProfile<1, 1, [SDTCisPtrTy<0>, SDTCisPtrTy<1>]>;
68 def A64GOTLoad : SDNode<"AArch64ISD::GOTLoad", SDTAArch64GOTLoad,
72 // (A64BFI LHS, RHS, LSB, Width)
73 def SDTA64BFI : SDTypeProfile<1, 4, [SDTCisSameAs<0, 1>,
78 def A64Bfi : SDNode<"AArch64ISD::BFI", SDTA64BFI>;
80 // (A64EXTR HiReg, LoReg, LSB)
81 def SDTA64EXTR : SDTypeProfile<1, 3, [SDTCisSameAs<0, 1>, SDTCisSameAs<1, 2>,
83 def A64Extr : SDNode<"AArch64ISD::EXTR", SDTA64EXTR>;
85 // (A64[SU]BFX Field, ImmR, ImmS).
87 // Note that ImmR and ImmS are already encoded for the actual instructions. The
88 // more natural LSB and Width mix together to form ImmR and ImmS, something
89 // which TableGen can't handle.
90 def SDTA64BFX : SDTypeProfile<1, 3, [SDTCisVT<2, i64>, SDTCisVT<3, i64>]>;
91 def A64Sbfx : SDNode<"AArch64ISD::SBFX", SDTA64BFX>;
93 def A64Ubfx : SDNode<"AArch64ISD::UBFX", SDTA64BFX>;
95 //===----------------------------------------------------------------------===//
96 // Call sequence pseudo-instructions
97 //===----------------------------------------------------------------------===//
100 def SDT_AArch64Call : SDTypeProfile<0, -1, [SDTCisPtrTy<0>]>;
101 def AArch64Call : SDNode<"AArch64ISD::Call", SDT_AArch64Call,
102 [SDNPHasChain, SDNPOptInGlue, SDNPOutGlue, SDNPVariadic]>;
104 def AArch64tcret : SDNode<"AArch64ISD::TC_RETURN", SDT_AArch64Call,
105 [SDNPHasChain, SDNPOptInGlue, SDNPVariadic]>;
107 // The TLSDESCCALL node is a variant call which goes to an indirectly calculated
108 // destination but needs a relocation against a fixed symbol. As such it has two
109 // certain operands: the callee and the relocated variable.
111 // The TLS ABI only allows it to be selected to a BLR instructin (with
112 // appropriate relocation).
113 def SDTTLSDescCall : SDTypeProfile<0, -2, [SDTCisPtrTy<0>, SDTCisPtrTy<1>]>;
115 def A64tlsdesc_blr : SDNode<"AArch64ISD::TLSDESCCALL", SDTTLSDescCall,
116 [SDNPInGlue, SDNPOutGlue, SDNPHasChain,
120 def SDT_AArch64CallSeqStart : SDCallSeqStart<[ SDTCisPtrTy<0> ]>;
121 def AArch64callseq_start : SDNode<"ISD::CALLSEQ_START", SDT_AArch64CallSeqStart,
122 [SDNPHasChain, SDNPOutGlue]>;
124 def SDT_AArch64CallSeqEnd : SDCallSeqEnd<[ SDTCisPtrTy<0>, SDTCisPtrTy<1> ]>;
125 def AArch64callseq_end : SDNode<"ISD::CALLSEQ_END", SDT_AArch64CallSeqEnd,
126 [SDNPHasChain, SDNPOptInGlue, SDNPOutGlue]>;
130 // These pseudo-instructions have special semantics by virtue of being passed to
131 // the InstrInfo constructor. CALLSEQ_START/CALLSEQ_END are produced by
132 // LowerCall to (in our case) tell the back-end about stack adjustments for
133 // arguments passed on the stack. Here we select those markers to
134 // pseudo-instructions which explicitly set the stack, and finally in the
135 // RegisterInfo we convert them to a true stack adjustment.
136 let Defs = [XSP], Uses = [XSP] in {
137 def ADJCALLSTACKDOWN : PseudoInst<(outs), (ins i64imm:$amt),
138 [(AArch64callseq_start timm:$amt)]>;
140 def ADJCALLSTACKUP : PseudoInst<(outs), (ins i64imm:$amt1, i64imm:$amt2),
141 [(AArch64callseq_end timm:$amt1, timm:$amt2)]>;
144 //===----------------------------------------------------------------------===//
145 // Atomic operation pseudo-instructions
146 //===----------------------------------------------------------------------===//
148 let usesCustomInserter = 1, Defs = [NZCV] in {
149 multiclass AtomicSizes<string opname> {
150 def _I8 : PseudoInst<(outs GPR32:$dst), (ins GPR64:$ptr, GPR32:$incr),
151 [(set GPR32:$dst, (!cast<SDNode>(opname # "_8") GPR64:$ptr, GPR32:$incr))]>;
152 def _I16 : PseudoInst<(outs GPR32:$dst), (ins GPR64:$ptr, GPR32:$incr),
153 [(set GPR32:$dst, (!cast<SDNode>(opname # "_16") GPR64:$ptr, GPR32:$incr))]>;
154 def _I32 : PseudoInst<(outs GPR32:$dst), (ins GPR64:$ptr, GPR32:$incr),
155 [(set GPR32:$dst, (!cast<SDNode>(opname # "_32") GPR64:$ptr, GPR32:$incr))]>;
156 def _I64 : PseudoInst<(outs GPR64:$dst), (ins GPR64:$ptr, GPR64:$incr),
157 [(set GPR64:$dst, (!cast<SDNode>(opname # "_64") GPR64:$ptr, GPR64:$incr))]>;
161 defm ATOMIC_LOAD_ADD : AtomicSizes<"atomic_load_add">;
162 defm ATOMIC_LOAD_SUB : AtomicSizes<"atomic_load_sub">;
163 defm ATOMIC_LOAD_AND : AtomicSizes<"atomic_load_and">;
164 defm ATOMIC_LOAD_OR : AtomicSizes<"atomic_load_or">;
165 defm ATOMIC_LOAD_XOR : AtomicSizes<"atomic_load_xor">;
166 defm ATOMIC_LOAD_NAND : AtomicSizes<"atomic_load_nand">;
167 defm ATOMIC_LOAD_MIN : AtomicSizes<"atomic_load_min">;
168 defm ATOMIC_LOAD_MAX : AtomicSizes<"atomic_load_max">;
169 defm ATOMIC_LOAD_UMIN : AtomicSizes<"atomic_load_umin">;
170 defm ATOMIC_LOAD_UMAX : AtomicSizes<"atomic_load_umax">;
171 defm ATOMIC_SWAP : AtomicSizes<"atomic_swap">;
173 let usesCustomInserter = 1, Defs = [NZCV] in {
174 def ATOMIC_CMP_SWAP_I8
175 : PseudoInst<(outs GPR32:$dst), (ins GPR64:$ptr, GPR32:$old, GPR32:$new),
177 (atomic_cmp_swap_8 GPR64:$ptr, GPR32:$old, GPR32:$new))]>;
178 def ATOMIC_CMP_SWAP_I16
179 : PseudoInst<(outs GPR32:$dst), (ins GPR64:$ptr, GPR32:$old, GPR32:$new),
181 (atomic_cmp_swap_16 GPR64:$ptr, GPR32:$old, GPR32:$new))]>;
182 def ATOMIC_CMP_SWAP_I32
183 : PseudoInst<(outs GPR32:$dst), (ins GPR64:$ptr, GPR32:$old, GPR32:$new),
185 (atomic_cmp_swap_32 GPR64:$ptr, GPR32:$old, GPR32:$new))]>;
186 def ATOMIC_CMP_SWAP_I64
187 : PseudoInst<(outs GPR64:$dst), (ins GPR64:$ptr, GPR64:$old, GPR64:$new),
189 (atomic_cmp_swap_64 GPR64:$ptr, GPR64:$old, GPR64:$new))]>;
192 //===----------------------------------------------------------------------===//
193 // Add-subtract (extended register) instructions
194 //===----------------------------------------------------------------------===//
195 // Contains: ADD, ADDS, SUB, SUBS + aliases CMN, CMP
197 // The RHS of these operations is conceptually a sign/zero-extended
198 // register, optionally shifted left by 1-4. The extension can be a
199 // NOP (e.g. "sxtx" sign-extending a 64-bit register to 64-bits) but
200 // must be specified with one exception:
202 // If one of the registers is sp/wsp then LSL is an alias for UXTW in
203 // 32-bit instructions and UXTX in 64-bit versions, the shift amount
204 // is not optional in that case (but can explicitly be 0), and the
205 // entire suffix can be skipped (e.g. "add sp, x3, x2").
207 multiclass extend_operands<string PREFIX> {
208 def _asmoperand : AsmOperandClass {
210 let RenderMethod = "addRegExtendOperands";
211 let PredicateMethod = "isRegExtend<A64SE::" # PREFIX # ">";
214 def _operand : Operand<i64>,
215 ImmLeaf<i64, [{ return Imm >= 0 && Imm <= 4; }]> {
216 let PrintMethod = "printRegExtendOperand<A64SE::" # PREFIX # ">";
217 let DecoderMethod = "DecodeRegExtendOperand";
218 let ParserMatchClass = !cast<AsmOperandClass>(PREFIX # "_asmoperand");
222 defm UXTB : extend_operands<"UXTB">;
223 defm UXTH : extend_operands<"UXTH">;
224 defm UXTW : extend_operands<"UXTW">;
225 defm UXTX : extend_operands<"UXTX">;
226 defm SXTB : extend_operands<"SXTB">;
227 defm SXTH : extend_operands<"SXTH">;
228 defm SXTW : extend_operands<"SXTW">;
229 defm SXTX : extend_operands<"SXTX">;
231 def LSL_extasmoperand : AsmOperandClass {
232 let Name = "RegExtendLSL";
233 let RenderMethod = "addRegExtendOperands";
236 def LSL_extoperand : Operand<i64> {
237 let ParserMatchClass = LSL_extasmoperand;
241 // The patterns for various sign-extensions are a little ugly and
242 // non-uniform because everything has already been promoted to the
243 // legal i64 and i32 types. We'll wrap the various variants up in a
244 // class for use later.
246 dag uxtb; dag uxth; dag uxtw; dag uxtx;
247 dag sxtb; dag sxth; dag sxtw; dag sxtx;
250 def extends_to_i64 : extend_types {
251 let uxtb = (and (anyext GPR32:$Rm), 255);
252 let uxth = (and (anyext GPR32:$Rm), 65535);
253 let uxtw = (zext GPR32:$Rm);
254 let uxtx = (i64 GPR64:$Rm);
256 let sxtb = (sext_inreg (anyext GPR32:$Rm), i8);
257 let sxth = (sext_inreg (anyext GPR32:$Rm), i16);
258 let sxtw = (sext GPR32:$Rm);
259 let sxtx = (i64 GPR64:$Rm);
263 def extends_to_i32 : extend_types {
264 let uxtb = (and GPR32:$Rm, 255);
265 let uxth = (and GPR32:$Rm, 65535);
266 let uxtw = (i32 GPR32:$Rm);
267 let uxtx = (i32 GPR32:$Rm);
269 let sxtb = (sext_inreg GPR32:$Rm, i8);
270 let sxth = (sext_inreg GPR32:$Rm, i16);
271 let sxtw = (i32 GPR32:$Rm);
272 let sxtx = (i32 GPR32:$Rm);
275 // Now, six of the extensions supported are easy and uniform: if the source size
276 // is 32-bits or less, then Rm is always a 32-bit register. We'll instantiate
277 // those instructions in one block.
279 // The uxtx/sxtx could potentially be merged in, but three facts dissuaded me:
280 // + It would break the naming scheme: either ADDxx_uxtx or ADDww_uxtx would
282 // + Patterns are very different as well.
283 // + Passing different registers would be ugly (more fields in extend_types
284 // would probably be the best option).
285 multiclass addsub_exts<bit sf, bit op, bit S, string asmop,
286 SDPatternOperator opfrag,
287 dag outs, extend_types exts, RegisterClass GPRsp> {
288 def w_uxtb : A64I_addsubext<sf, op, S, 0b00, 0b000,
290 (ins GPRsp:$Rn, GPR32:$Rm, UXTB_operand:$Imm3),
291 !strconcat(asmop, "$Rn, $Rm, $Imm3"),
292 [(opfrag GPRsp:$Rn, (shl exts.uxtb, UXTB_operand:$Imm3))],
294 def w_uxth : A64I_addsubext<sf, op, S, 0b00, 0b001,
296 (ins GPRsp:$Rn, GPR32:$Rm, UXTH_operand:$Imm3),
297 !strconcat(asmop, "$Rn, $Rm, $Imm3"),
298 [(opfrag GPRsp:$Rn, (shl exts.uxth, UXTH_operand:$Imm3))],
300 def w_uxtw : A64I_addsubext<sf, op, S, 0b00, 0b010,
302 (ins GPRsp:$Rn, GPR32:$Rm, UXTW_operand:$Imm3),
303 !strconcat(asmop, "$Rn, $Rm, $Imm3"),
304 [(opfrag GPRsp:$Rn, (shl exts.uxtw, UXTW_operand:$Imm3))],
307 def w_sxtb : A64I_addsubext<sf, op, S, 0b00, 0b100,
309 (ins GPRsp:$Rn, GPR32:$Rm, SXTB_operand:$Imm3),
310 !strconcat(asmop, "$Rn, $Rm, $Imm3"),
311 [(opfrag GPRsp:$Rn, (shl exts.sxtb, SXTB_operand:$Imm3))],
313 def w_sxth : A64I_addsubext<sf, op, S, 0b00, 0b101,
315 (ins GPRsp:$Rn, GPR32:$Rm, SXTH_operand:$Imm3),
316 !strconcat(asmop, "$Rn, $Rm, $Imm3"),
317 [(opfrag GPRsp:$Rn, (shl exts.sxth, SXTH_operand:$Imm3))],
319 def w_sxtw : A64I_addsubext<sf, op, S, 0b00, 0b110,
321 (ins GPRsp:$Rn, GPR32:$Rm, SXTW_operand:$Imm3),
322 !strconcat(asmop, "$Rn, $Rm, $Imm3"),
323 [(opfrag GPRsp:$Rn, (shl exts.sxtw, SXTW_operand:$Imm3))],
327 // These two could be merge in with the above, but their patterns aren't really
328 // necessary and the naming-scheme would necessarily break:
329 multiclass addsub_xxtx<bit op, bit S, string asmop, SDPatternOperator opfrag,
331 def x_uxtx : A64I_addsubext<0b1, op, S, 0b00, 0b011,
333 (ins GPR64xsp:$Rn, GPR64:$Rm, UXTX_operand:$Imm3),
334 !strconcat(asmop, "$Rn, $Rm, $Imm3"),
335 [(opfrag GPR64xsp:$Rn, (shl GPR64:$Rm, UXTX_operand:$Imm3))],
338 def x_sxtx : A64I_addsubext<0b1, op, S, 0b00, 0b111,
340 (ins GPR64xsp:$Rn, GPR64:$Rm, SXTX_operand:$Imm3),
341 !strconcat(asmop, "$Rn, $Rm, $Imm3"),
342 [/* No Pattern: same as uxtx */],
346 multiclass addsub_wxtx<bit op, bit S, string asmop, dag outs> {
347 def w_uxtx : A64I_addsubext<0b0, op, S, 0b00, 0b011,
349 (ins GPR32wsp:$Rn, GPR32:$Rm, UXTX_operand:$Imm3),
350 !strconcat(asmop, "$Rn, $Rm, $Imm3"),
351 [/* No pattern: probably same as uxtw */],
354 def w_sxtx : A64I_addsubext<0b0, op, S, 0b00, 0b111,
356 (ins GPR32wsp:$Rn, GPR32:$Rm, SXTX_operand:$Imm3),
357 !strconcat(asmop, "$Rn, $Rm, $Imm3"),
358 [/* No Pattern: probably same as uxtw */],
362 class SetRD<RegisterClass RC, SDPatternOperator op>
363 : PatFrag<(ops node:$lhs, node:$rhs), (set RC:$Rd, (op node:$lhs, node:$rhs))>;
364 class SetNZCV<SDPatternOperator op>
365 : PatFrag<(ops node:$lhs, node:$rhs), (set NZCV, (op node:$lhs, node:$rhs))>;
367 defm ADDxx :addsub_exts<0b1, 0b0, 0b0, "add\t$Rd, ", SetRD<GPR64xsp, add>,
368 (outs GPR64xsp:$Rd), extends_to_i64, GPR64xsp>,
369 addsub_xxtx< 0b0, 0b0, "add\t$Rd, ", SetRD<GPR64xsp, add>,
370 (outs GPR64xsp:$Rd)>;
371 defm ADDww :addsub_exts<0b0, 0b0, 0b0, "add\t$Rd, ", SetRD<GPR32wsp, add>,
372 (outs GPR32wsp:$Rd), extends_to_i32, GPR32wsp>,
373 addsub_wxtx< 0b0, 0b0, "add\t$Rd, ",
374 (outs GPR32wsp:$Rd)>;
375 defm SUBxx :addsub_exts<0b1, 0b1, 0b0, "sub\t$Rd, ", SetRD<GPR64xsp, sub>,
376 (outs GPR64xsp:$Rd), extends_to_i64, GPR64xsp>,
377 addsub_xxtx< 0b1, 0b0, "sub\t$Rd, ", SetRD<GPR64xsp, sub>,
378 (outs GPR64xsp:$Rd)>;
379 defm SUBww :addsub_exts<0b0, 0b1, 0b0, "sub\t$Rd, ", SetRD<GPR32wsp, sub>,
380 (outs GPR32wsp:$Rd), extends_to_i32, GPR32wsp>,
381 addsub_wxtx< 0b1, 0b0, "sub\t$Rd, ",
382 (outs GPR32wsp:$Rd)>;
384 let Defs = [NZCV] in {
385 defm ADDSxx :addsub_exts<0b1, 0b0, 0b1, "adds\t$Rd, ", SetRD<GPR64, addc>,
386 (outs GPR64:$Rd), extends_to_i64, GPR64xsp>,
387 addsub_xxtx< 0b0, 0b1, "adds\t$Rd, ", SetRD<GPR64, addc>,
389 defm ADDSww :addsub_exts<0b0, 0b0, 0b1, "adds\t$Rd, ", SetRD<GPR32, addc>,
390 (outs GPR32:$Rd), extends_to_i32, GPR32wsp>,
391 addsub_wxtx< 0b0, 0b1, "adds\t$Rd, ",
393 defm SUBSxx :addsub_exts<0b1, 0b1, 0b1, "subs\t$Rd, ", SetRD<GPR64, subc>,
394 (outs GPR64:$Rd), extends_to_i64, GPR64xsp>,
395 addsub_xxtx< 0b1, 0b1, "subs\t$Rd, ", SetRD<GPR64, subc>,
397 defm SUBSww :addsub_exts<0b0, 0b1, 0b1, "subs\t$Rd, ", SetRD<GPR32, subc>,
398 (outs GPR32:$Rd), extends_to_i32, GPR32wsp>,
399 addsub_wxtx< 0b1, 0b1, "subs\t$Rd, ",
403 let Rd = 0b11111, isCompare = 1 in {
404 defm CMNx : addsub_exts<0b1, 0b0, 0b1, "cmn\t", SetNZCV<A64cmn>,
405 (outs), extends_to_i64, GPR64xsp>,
406 addsub_xxtx< 0b0, 0b1, "cmn\t", SetNZCV<A64cmn>, (outs)>;
407 defm CMNw : addsub_exts<0b0, 0b0, 0b1, "cmn\t", SetNZCV<A64cmn>,
408 (outs), extends_to_i32, GPR32wsp>,
409 addsub_wxtx< 0b0, 0b1, "cmn\t", (outs)>;
410 defm CMPx : addsub_exts<0b1, 0b1, 0b1, "cmp\t", SetNZCV<A64cmp>,
411 (outs), extends_to_i64, GPR64xsp>,
412 addsub_xxtx< 0b1, 0b1, "cmp\t", SetNZCV<A64cmp>, (outs)>;
413 defm CMPw : addsub_exts<0b0, 0b1, 0b1, "cmp\t", SetNZCV<A64cmp>,
414 (outs), extends_to_i32, GPR32wsp>,
415 addsub_wxtx< 0b1, 0b1, "cmp\t", (outs)>;
419 // Now patterns for the operation without a shift being needed. No patterns are
420 // created for uxtx/sxtx since they're non-uniform and it's expected that
421 // add/sub (shifted register) will handle those cases anyway.
422 multiclass addsubext_noshift_patterns<string prefix, SDPatternOperator nodeop,
423 RegisterClass GPRsp, extend_types exts> {
424 def : Pat<(nodeop GPRsp:$Rn, exts.uxtb),
425 (!cast<Instruction>(prefix # "w_uxtb") GPRsp:$Rn, GPR32:$Rm, 0)>;
426 def : Pat<(nodeop GPRsp:$Rn, exts.uxth),
427 (!cast<Instruction>(prefix # "w_uxth") GPRsp:$Rn, GPR32:$Rm, 0)>;
428 def : Pat<(nodeop GPRsp:$Rn, exts.uxtw),
429 (!cast<Instruction>(prefix # "w_uxtw") GPRsp:$Rn, GPR32:$Rm, 0)>;
431 def : Pat<(nodeop GPRsp:$Rn, exts.sxtb),
432 (!cast<Instruction>(prefix # "w_sxtb") GPRsp:$Rn, GPR32:$Rm, 0)>;
433 def : Pat<(nodeop GPRsp:$Rn, exts.sxth),
434 (!cast<Instruction>(prefix # "w_sxth") GPRsp:$Rn, GPR32:$Rm, 0)>;
435 def : Pat<(nodeop GPRsp:$Rn, exts.sxtw),
436 (!cast<Instruction>(prefix # "w_sxtw") GPRsp:$Rn, GPR32:$Rm, 0)>;
439 defm : addsubext_noshift_patterns<"ADDxx", add, GPR64xsp, extends_to_i64>;
440 defm : addsubext_noshift_patterns<"ADDww", add, GPR32wsp, extends_to_i32>;
441 defm : addsubext_noshift_patterns<"SUBxx", sub, GPR64xsp, extends_to_i64>;
442 defm : addsubext_noshift_patterns<"SUBww", sub, GPR32wsp, extends_to_i32>;
444 defm : addsubext_noshift_patterns<"CMNx", A64cmn, GPR64xsp, extends_to_i64>;
445 defm : addsubext_noshift_patterns<"CMNw", A64cmn, GPR32wsp, extends_to_i32>;
446 defm : addsubext_noshift_patterns<"CMPx", A64cmp, GPR64xsp, extends_to_i64>;
447 defm : addsubext_noshift_patterns<"CMPw", A64cmp, GPR32wsp, extends_to_i32>;
449 // An extend of "lsl #imm" is valid if and only if one of Rn and Rd is
450 // sp/wsp. It is synonymous with uxtx/uxtw depending on the size of the
451 // operation. Also permitted in this case is complete omission of the argument,
452 // which implies "lsl #0".
453 multiclass lsl_aliases<string asmop, Instruction inst, RegisterClass GPR_Rd,
454 RegisterClass GPR_Rn, RegisterClass GPR_Rm> {
455 def : InstAlias<!strconcat(asmop, " $Rd, $Rn, $Rm"),
456 (inst GPR_Rd:$Rd, GPR_Rn:$Rn, GPR_Rm:$Rm, 0)>;
458 def : InstAlias<!strconcat(asmop, " $Rd, $Rn, $Rm, $LSL"),
459 (inst GPR_Rd:$Rd, GPR_Rn:$Rn, GPR_Rm:$Rm, LSL_extoperand:$LSL)>;
463 defm : lsl_aliases<"add", ADDxxx_uxtx, Rxsp, GPR64xsp, GPR64>;
464 defm : lsl_aliases<"add", ADDxxx_uxtx, GPR64xsp, Rxsp, GPR64>;
465 defm : lsl_aliases<"add", ADDwww_uxtw, Rwsp, GPR32wsp, GPR32>;
466 defm : lsl_aliases<"add", ADDwww_uxtw, GPR32wsp, Rwsp, GPR32>;
467 defm : lsl_aliases<"sub", SUBxxx_uxtx, Rxsp, GPR64xsp, GPR64>;
468 defm : lsl_aliases<"sub", SUBxxx_uxtx, GPR64xsp, Rxsp, GPR64>;
469 defm : lsl_aliases<"sub", SUBwww_uxtw, Rwsp, GPR32wsp, GPR32>;
470 defm : lsl_aliases<"sub", SUBwww_uxtw, GPR32wsp, Rwsp, GPR32>;
472 // Rd cannot be sp for flag-setting variants so only half of the aliases are
474 defm : lsl_aliases<"adds", ADDSxxx_uxtx, GPR64, Rxsp, GPR64>;
475 defm : lsl_aliases<"adds", ADDSwww_uxtw, GPR32, Rwsp, GPR32>;
476 defm : lsl_aliases<"subs", SUBSxxx_uxtx, GPR64, Rxsp, GPR64>;
477 defm : lsl_aliases<"subs", SUBSwww_uxtw, GPR32, Rwsp, GPR32>;
479 // CMP unfortunately has to be different because the instruction doesn't have a
481 multiclass cmp_lsl_aliases<string asmop, Instruction inst,
482 RegisterClass GPR_Rn, RegisterClass GPR_Rm> {
483 def : InstAlias<!strconcat(asmop, " $Rn, $Rm"),
484 (inst GPR_Rn:$Rn, GPR_Rm:$Rm, 0)>;
486 def : InstAlias<!strconcat(asmop, " $Rn, $Rm, $LSL"),
487 (inst GPR_Rn:$Rn, GPR_Rm:$Rm, LSL_extoperand:$LSL)>;
490 defm : cmp_lsl_aliases<"cmp", CMPxx_uxtx, Rxsp, GPR64>;
491 defm : cmp_lsl_aliases<"cmp", CMPww_uxtw, Rwsp, GPR32>;
492 defm : cmp_lsl_aliases<"cmn", CMNxx_uxtx, Rxsp, GPR64>;
493 defm : cmp_lsl_aliases<"cmn", CMNww_uxtw, Rwsp, GPR32>;
495 //===----------------------------------------------------------------------===//
496 // Add-subtract (immediate) instructions
497 //===----------------------------------------------------------------------===//
498 // Contains: ADD, ADDS, SUB, SUBS + aliases CMN, CMP, MOV
500 // These instructions accept a 12-bit unsigned immediate, optionally shifted
501 // left by 12 bits. Official assembly format specifies a 12 bit immediate with
502 // one of "", "LSL #0", "LSL #12" supplementary operands.
504 // There are surprisingly few ways to make this work with TableGen, so this
505 // implementation has separate instructions for the "LSL #0" and "LSL #12"
508 // If the MCInst retained a single combined immediate (which could be 0x123000,
509 // for example) then both components (imm & shift) would have to be delegated to
510 // a single assembly operand. This would entail a separate operand parser
511 // (because the LSL would have to live in the same AArch64Operand as the
512 // immediate to be accessible); assembly parsing is rather complex and
513 // error-prone C++ code.
515 // By splitting the immediate, we can delegate handling this optional operand to
516 // an InstAlias. Supporting functions to generate the correct MCInst are still
517 // required, but these are essentially trivial and parsing can remain generic.
519 // Rejected plans with rationale:
520 // ------------------------------
522 // In an ideal world you'de have two first class immediate operands (in
523 // InOperandList, specifying imm12 and shift). Unfortunately this is not
524 // selectable by any means I could discover.
526 // An Instruction with two MCOperands hidden behind a single entry in
527 // InOperandList (expanded by ComplexPatterns and MIOperandInfo) was functional,
528 // but required more C++ code to handle encoding/decoding. Parsing (the intended
529 // main beneficiary) ended up equally complex because of the optional nature of
532 // Attempting to circumvent the need for a custom OperandParser above by giving
533 // InstAliases without the "lsl #0" failed. add/sub could be accommodated but
534 // the cmp/cmn aliases didn't use the MIOperandInfo to determine how operands
535 // should be parsed: there was no way to accommodate an "lsl #12".
537 let ParserMethod = "ParseImmWithLSLOperand",
538 RenderMethod = "addImmWithLSLOperands" in {
539 // Derived PredicateMethod fields are different for each
540 def addsubimm_lsl0_asmoperand : AsmOperandClass {
541 let Name = "AddSubImmLSL0";
544 def addsubimm_lsl12_asmoperand : AsmOperandClass {
545 let Name = "AddSubImmLSL12";
549 def shr_12_XFORM : SDNodeXForm<imm, [{
550 return CurDAG->getTargetConstant(N->getSExtValue() >> 12, MVT::i32);
553 def shr_12_neg_XFORM : SDNodeXForm<imm, [{
554 return CurDAG->getTargetConstant((-N->getSExtValue()) >> 12, MVT::i32);
557 def neg_XFORM : SDNodeXForm<imm, [{
558 return CurDAG->getTargetConstant(-N->getSExtValue(), MVT::i32);
562 multiclass addsub_imm_operands<ValueType ty> {
563 let PrintMethod = "printAddSubImmLSL0Operand",
564 EncoderMethod = "getAddSubImmOpValue",
565 ParserMatchClass = addsubimm_lsl0_asmoperand in {
566 def _posimm_lsl0 : Operand<ty>,
567 ImmLeaf<ty, [{ return Imm >= 0 && (Imm & ~0xfff) == 0; }]>;
568 def _negimm_lsl0 : Operand<ty>,
569 ImmLeaf<ty, [{ return Imm < 0 && (-Imm & ~0xfff) == 0; }],
573 let PrintMethod = "printAddSubImmLSL12Operand",
574 EncoderMethod = "getAddSubImmOpValue",
575 ParserMatchClass = addsubimm_lsl12_asmoperand in {
576 def _posimm_lsl12 : Operand<ty>,
577 ImmLeaf<ty, [{ return Imm >= 0 && (Imm & ~0xfff000) == 0; }],
580 def _negimm_lsl12 : Operand<ty>,
581 ImmLeaf<ty, [{ return Imm < 0 && (-Imm & ~0xfff000) == 0; }],
586 // The add operands don't need any transformation
587 defm addsubimm_operand_i32 : addsub_imm_operands<i32>;
588 defm addsubimm_operand_i64 : addsub_imm_operands<i64>;
590 multiclass addsubimm_varieties<string prefix, bit sf, bit op, bits<2> shift,
591 string asmop, string cmpasmop,
592 Operand imm_operand, Operand cmp_imm_operand,
593 RegisterClass GPR, RegisterClass GPRsp,
595 // All registers for non-S variants allow SP
596 def _s : A64I_addsubimm<sf, op, 0b0, shift,
598 (ins GPRsp:$Rn, imm_operand:$Imm12),
599 !strconcat(asmop, "\t$Rd, $Rn, $Imm12"),
601 (add GPRsp:$Rn, imm_operand:$Imm12))],
605 // S variants can read SP but would write to ZR
606 def _S : A64I_addsubimm<sf, op, 0b1, shift,
608 (ins GPRsp:$Rn, imm_operand:$Imm12),
609 !strconcat(asmop, "s\t$Rd, $Rn, $Imm12"),
610 [(set GPR:$Rd, (addc GPRsp:$Rn, imm_operand:$Imm12))],
615 // Note that the pattern here for ADDS is subtle. Canonically CMP
616 // a, b becomes SUBS a, b. If b < 0 then this is equivalent to
617 // ADDS a, (-b). This is not true in general.
618 def _cmp : A64I_addsubimm<sf, op, 0b1, shift,
619 (outs), (ins GPRsp:$Rn, imm_operand:$Imm12),
620 !strconcat(cmpasmop, " $Rn, $Imm12"),
622 (A64cmp GPRsp:$Rn, cmp_imm_operand:$Imm12))],
631 multiclass addsubimm_shifts<string prefix, bit sf, bit op,
632 string asmop, string cmpasmop, string operand, string cmpoperand,
633 RegisterClass GPR, RegisterClass GPRsp, AArch64Reg ZR> {
634 defm _lsl0 : addsubimm_varieties<prefix # "_lsl0", sf, op, 0b00,
636 !cast<Operand>(operand # "_lsl0"),
637 !cast<Operand>(cmpoperand # "_lsl0"),
640 defm _lsl12 : addsubimm_varieties<prefix # "_lsl12", sf, op, 0b01,
642 !cast<Operand>(operand # "_lsl12"),
643 !cast<Operand>(cmpoperand # "_lsl12"),
647 defm ADDwwi : addsubimm_shifts<"ADDwi", 0b0, 0b0, "add", "cmn",
648 "addsubimm_operand_i32_posimm",
649 "addsubimm_operand_i32_negimm",
650 GPR32, GPR32wsp, WZR>;
651 defm ADDxxi : addsubimm_shifts<"ADDxi", 0b1, 0b0, "add", "cmn",
652 "addsubimm_operand_i64_posimm",
653 "addsubimm_operand_i64_negimm",
654 GPR64, GPR64xsp, XZR>;
655 defm SUBwwi : addsubimm_shifts<"SUBwi", 0b0, 0b1, "sub", "cmp",
656 "addsubimm_operand_i32_negimm",
657 "addsubimm_operand_i32_posimm",
658 GPR32, GPR32wsp, WZR>;
659 defm SUBxxi : addsubimm_shifts<"SUBxi", 0b1, 0b1, "sub", "cmp",
660 "addsubimm_operand_i64_negimm",
661 "addsubimm_operand_i64_posimm",
662 GPR64, GPR64xsp, XZR>;
664 multiclass MOVsp<RegisterClass GPRsp, RegisterClass SP, Instruction addop> {
665 def _fromsp : InstAlias<"mov $Rd, $Rn",
666 (addop GPRsp:$Rd, SP:$Rn, 0),
669 def _tosp : InstAlias<"mov $Rd, $Rn",
670 (addop SP:$Rd, GPRsp:$Rn, 0),
674 // Recall Rxsp is a RegisterClass containing *just* xsp.
675 defm MOVxx : MOVsp<GPR64xsp, Rxsp, ADDxxi_lsl0_s>;
676 defm MOVww : MOVsp<GPR32wsp, Rwsp, ADDwwi_lsl0_s>;
678 //===----------------------------------------------------------------------===//
679 // Add-subtract (shifted register) instructions
680 //===----------------------------------------------------------------------===//
681 // Contains: ADD, ADDS, SUB, SUBS + aliases CMN, CMP, NEG, NEGS
683 //===-------------------------------
684 // 1. The "shifed register" operands. Shared with logical insts.
685 //===-------------------------------
687 multiclass shift_operands<string prefix, string form> {
688 def _asmoperand_i32 : AsmOperandClass {
689 let Name = "Shift" # form # "i32";
690 let RenderMethod = "addShiftOperands";
692 = "isShift<A64SE::" # form # ", false>";
695 // Note that the operand type is intentionally i64 because the DAGCombiner
696 // puts these into a canonical form.
697 def _i32 : Operand<i64>, ImmLeaf<i64, [{ return Imm >= 0 && Imm <= 31; }]> {
699 = !cast<AsmOperandClass>(prefix # "_asmoperand_i32");
700 let PrintMethod = "printShiftOperand<A64SE::" # form # ">";
701 let DecoderMethod = "Decode32BitShiftOperand";
704 def _asmoperand_i64 : AsmOperandClass {
705 let Name = "Shift" # form # "i64";
706 let RenderMethod = "addShiftOperands";
708 = "isShift<A64SE::" # form # ", true>";
711 def _i64 : Operand<i64>, ImmLeaf<i64, [{ return Imm >= 0 && Imm <= 63; }]> {
713 = !cast<AsmOperandClass>(prefix # "_asmoperand_i64");
714 let PrintMethod = "printShiftOperand<A64SE::" # form # ">";
718 defm lsl_operand : shift_operands<"lsl_operand", "LSL">;
719 defm lsr_operand : shift_operands<"lsr_operand", "LSR">;
720 defm asr_operand : shift_operands<"asr_operand", "ASR">;
722 // Not used for add/sub, but defined here for completeness. The "logical
723 // (shifted register)" instructions *do* have an ROR variant.
724 defm ror_operand : shift_operands<"ror_operand", "ROR">;
726 //===-------------------------------
727 // 2. The basic 3.5-operand ADD/SUB/ADDS/SUBS instructions.
728 //===-------------------------------
730 // N.b. the commutable parameter is just !N. It will be first against the wall
731 // when the revolution comes.
732 multiclass addsub_shifts<string prefix, bit sf, bit op, bit s, bit commutable,
733 string asmop, SDPatternOperator opfrag, string sty,
734 RegisterClass GPR, list<Register> defs> {
735 let isCommutable = commutable, Defs = defs in {
736 def _lsl : A64I_addsubshift<sf, op, s, 0b00,
738 (ins GPR:$Rn, GPR:$Rm,
739 !cast<Operand>("lsl_operand_" # sty):$Imm6),
740 !strconcat(asmop, "\t$Rd, $Rn, $Rm, $Imm6"),
741 [(set GPR:$Rd, (opfrag GPR:$Rn, (shl GPR:$Rm,
742 !cast<Operand>("lsl_operand_" # sty):$Imm6))
746 def _lsr : A64I_addsubshift<sf, op, s, 0b01,
748 (ins GPR:$Rn, GPR:$Rm,
749 !cast<Operand>("lsr_operand_" # sty):$Imm6),
750 !strconcat(asmop, "\t$Rd, $Rn, $Rm, $Imm6"),
751 [(set GPR:$Rd, (opfrag GPR:$Rn, (srl GPR:$Rm,
752 !cast<Operand>("lsr_operand_" # sty):$Imm6))
756 def _asr : A64I_addsubshift<sf, op, s, 0b10,
758 (ins GPR:$Rn, GPR:$Rm,
759 !cast<Operand>("asr_operand_" # sty):$Imm6),
760 !strconcat(asmop, "\t$Rd, $Rn, $Rm, $Imm6"),
761 [(set GPR:$Rd, (opfrag GPR:$Rn, (sra GPR:$Rm,
762 !cast<Operand>("asr_operand_" # sty):$Imm6))
768 : InstAlias<!strconcat(asmop, " $Rd, $Rn, $Rm"),
769 (!cast<Instruction>(prefix # "_lsl") GPR:$Rd, GPR:$Rn,
772 def : Pat<(opfrag GPR:$Rn, GPR:$Rm),
773 (!cast<Instruction>(prefix # "_lsl") GPR:$Rn, GPR:$Rm, 0)>;
776 multiclass addsub_sizes<string prefix, bit op, bit s, bit commutable,
777 string asmop, SDPatternOperator opfrag,
778 list<Register> defs> {
779 defm xxx : addsub_shifts<prefix # "xxx", 0b1, op, s,
780 commutable, asmop, opfrag, "i64", GPR64, defs>;
781 defm www : addsub_shifts<prefix # "www", 0b0, op, s,
782 commutable, asmop, opfrag, "i32", GPR32, defs>;
786 defm ADD : addsub_sizes<"ADD", 0b0, 0b0, 0b1, "add", add, []>;
787 defm SUB : addsub_sizes<"SUB", 0b1, 0b0, 0b0, "sub", sub, []>;
789 defm ADDS : addsub_sizes<"ADDS", 0b0, 0b1, 0b1, "adds", addc, [NZCV]>;
790 defm SUBS : addsub_sizes<"SUBS", 0b1, 0b1, 0b0, "subs", subc, [NZCV]>;
792 //===-------------------------------
793 // 1. The NEG/NEGS aliases
794 //===-------------------------------
796 multiclass neg_alias<Instruction INST, RegisterClass GPR,
797 Register ZR, Operand shift_operand, SDNode shiftop> {
798 def : InstAlias<"neg $Rd, $Rm, $Imm6",
799 (INST GPR:$Rd, ZR, GPR:$Rm, shift_operand:$Imm6)>;
801 def : Pat<(sub 0, (shiftop GPR:$Rm, shift_operand:$Imm6)),
802 (INST ZR, GPR:$Rm, shift_operand:$Imm6)>;
805 defm : neg_alias<SUBwww_lsl, GPR32, WZR, lsl_operand_i32, shl>;
806 defm : neg_alias<SUBwww_lsr, GPR32, WZR, lsr_operand_i32, srl>;
807 defm : neg_alias<SUBwww_asr, GPR32, WZR, asr_operand_i32, sra>;
808 def : InstAlias<"neg $Rd, $Rm", (SUBwww_lsl GPR32:$Rd, WZR, GPR32:$Rm, 0)>;
809 def : Pat<(sub 0, GPR32:$Rm), (SUBwww_lsl WZR, GPR32:$Rm, 0)>;
811 defm : neg_alias<SUBxxx_lsl, GPR64, XZR, lsl_operand_i64, shl>;
812 defm : neg_alias<SUBxxx_lsr, GPR64, XZR, lsr_operand_i64, srl>;
813 defm : neg_alias<SUBxxx_asr, GPR64, XZR, asr_operand_i64, sra>;
814 def : InstAlias<"neg $Rd, $Rm", (SUBxxx_lsl GPR64:$Rd, XZR, GPR64:$Rm, 0)>;
815 def : Pat<(sub 0, GPR64:$Rm), (SUBxxx_lsl XZR, GPR64:$Rm, 0)>;
817 // NEGS doesn't get any patterns yet: defining multiple outputs means C++ has to
819 class negs_alias<Instruction INST, RegisterClass GPR,
820 Register ZR, Operand shift_operand, SDNode shiftop>
821 : InstAlias<"negs $Rd, $Rm, $Imm6",
822 (INST GPR:$Rd, ZR, GPR:$Rm, shift_operand:$Imm6)>;
824 def : negs_alias<SUBSwww_lsl, GPR32, WZR, lsl_operand_i32, shl>;
825 def : negs_alias<SUBSwww_lsr, GPR32, WZR, lsr_operand_i32, srl>;
826 def : negs_alias<SUBSwww_asr, GPR32, WZR, asr_operand_i32, sra>;
827 def : InstAlias<"negs $Rd, $Rm", (SUBSwww_lsl GPR32:$Rd, WZR, GPR32:$Rm, 0)>;
829 def : negs_alias<SUBSxxx_lsl, GPR64, XZR, lsl_operand_i64, shl>;
830 def : negs_alias<SUBSxxx_lsr, GPR64, XZR, lsr_operand_i64, srl>;
831 def : negs_alias<SUBSxxx_asr, GPR64, XZR, asr_operand_i64, sra>;
832 def : InstAlias<"negs $Rd, $Rm", (SUBSxxx_lsl GPR64:$Rd, XZR, GPR64:$Rm, 0)>;
834 //===-------------------------------
835 // 1. The CMP/CMN aliases
836 //===-------------------------------
838 multiclass cmp_shifts<string prefix, bit sf, bit op, bit commutable,
839 string asmop, SDPatternOperator opfrag, string sty,
841 let isCommutable = commutable, Rd = 0b11111, Defs = [NZCV] in {
842 def _lsl : A64I_addsubshift<sf, op, 0b1, 0b00,
844 (ins GPR:$Rn, GPR:$Rm,
845 !cast<Operand>("lsl_operand_" # sty):$Imm6),
846 !strconcat(asmop, "\t$Rn, $Rm, $Imm6"),
847 [(set NZCV, (opfrag GPR:$Rn, (shl GPR:$Rm,
848 !cast<Operand>("lsl_operand_" # sty):$Imm6))
852 def _lsr : A64I_addsubshift<sf, op, 0b1, 0b01,
854 (ins GPR:$Rn, GPR:$Rm,
855 !cast<Operand>("lsr_operand_" # sty):$Imm6),
856 !strconcat(asmop, "\t$Rn, $Rm, $Imm6"),
857 [(set NZCV, (opfrag GPR:$Rn, (srl GPR:$Rm,
858 !cast<Operand>("lsr_operand_" # sty):$Imm6))
862 def _asr : A64I_addsubshift<sf, op, 0b1, 0b10,
864 (ins GPR:$Rn, GPR:$Rm,
865 !cast<Operand>("asr_operand_" # sty):$Imm6),
866 !strconcat(asmop, "\t$Rn, $Rm, $Imm6"),
867 [(set NZCV, (opfrag GPR:$Rn, (sra GPR:$Rm,
868 !cast<Operand>("asr_operand_" # sty):$Imm6))
874 : InstAlias<!strconcat(asmop, " $Rn, $Rm"),
875 (!cast<Instruction>(prefix # "_lsl") GPR:$Rn, GPR:$Rm, 0)>;
877 def : Pat<(opfrag GPR:$Rn, GPR:$Rm),
878 (!cast<Instruction>(prefix # "_lsl") GPR:$Rn, GPR:$Rm, 0)>;
881 defm CMPww : cmp_shifts<"CMPww", 0b0, 0b1, 0b0, "cmp", A64cmp, "i32", GPR32>;
882 defm CMPxx : cmp_shifts<"CMPxx", 0b1, 0b1, 0b0, "cmp", A64cmp, "i64", GPR64>;
884 defm CMNww : cmp_shifts<"CMNww", 0b0, 0b0, 0b1, "cmn", A64cmn, "i32", GPR32>;
885 defm CMNxx : cmp_shifts<"CMNxx", 0b1, 0b0, 0b1, "cmn", A64cmn, "i64", GPR64>;
887 //===----------------------------------------------------------------------===//
888 // Add-subtract (with carry) instructions
889 //===----------------------------------------------------------------------===//
890 // Contains: ADC, ADCS, SBC, SBCS + aliases NGC, NGCS
892 multiclass A64I_addsubcarrySizes<bit op, bit s, string asmop> {
893 let Uses = [NZCV] in {
894 def www : A64I_addsubcarry<0b0, op, s, 0b000000,
895 (outs GPR32:$Rd), (ins GPR32:$Rn, GPR32:$Rm),
896 !strconcat(asmop, "\t$Rd, $Rn, $Rm"),
899 def xxx : A64I_addsubcarry<0b1, op, s, 0b000000,
900 (outs GPR64:$Rd), (ins GPR64:$Rn, GPR64:$Rm),
901 !strconcat(asmop, "\t$Rd, $Rn, $Rm"),
906 let isCommutable = 1 in {
907 defm ADC : A64I_addsubcarrySizes<0b0, 0b0, "adc">;
910 defm SBC : A64I_addsubcarrySizes<0b1, 0b0, "sbc">;
912 let Defs = [NZCV] in {
913 let isCommutable = 1 in {
914 defm ADCS : A64I_addsubcarrySizes<0b0, 0b1, "adcs">;
917 defm SBCS : A64I_addsubcarrySizes<0b1, 0b1, "sbcs">;
920 def : InstAlias<"ngc $Rd, $Rm", (SBCwww GPR32:$Rd, WZR, GPR32:$Rm)>;
921 def : InstAlias<"ngc $Rd, $Rm", (SBCxxx GPR64:$Rd, XZR, GPR64:$Rm)>;
922 def : InstAlias<"ngcs $Rd, $Rm", (SBCSwww GPR32:$Rd, WZR, GPR32:$Rm)>;
923 def : InstAlias<"ngcs $Rd, $Rm", (SBCSxxx GPR64:$Rd, XZR, GPR64:$Rm)>;
925 // Note that adde and sube can form a chain longer than two (e.g. for 256-bit
926 // addition). So the flag-setting instructions are appropriate.
927 def : Pat<(adde GPR32:$Rn, GPR32:$Rm), (ADCSwww GPR32:$Rn, GPR32:$Rm)>;
928 def : Pat<(adde GPR64:$Rn, GPR64:$Rm), (ADCSxxx GPR64:$Rn, GPR64:$Rm)>;
929 def : Pat<(sube GPR32:$Rn, GPR32:$Rm), (SBCSwww GPR32:$Rn, GPR32:$Rm)>;
930 def : Pat<(sube GPR64:$Rn, GPR64:$Rm), (SBCSxxx GPR64:$Rn, GPR64:$Rm)>;
932 //===----------------------------------------------------------------------===//
934 //===----------------------------------------------------------------------===//
935 // Contains: SBFM, BFM, UBFM, [SU]XT[BHW], ASR, LSR, LSL, SBFI[ZX], BFI, BFXIL,
938 // Because of the rather complicated nearly-overlapping aliases, the decoding of
939 // this range of instructions is handled manually. The architectural
940 // instructions are BFM, SBFM and UBFM but a disassembler should never produce
943 // In the end, the best option was to use BFM instructions for decoding under
944 // almost all circumstances, but to create aliasing *Instructions* for each of
945 // the canonical forms and specify a completely custom decoder which would
946 // substitute the correct MCInst as needed.
948 // This also simplifies instruction selection, parsing etc because the MCInsts
949 // have a shape that's closer to their use in code.
951 //===-------------------------------
952 // 1. The architectural BFM instructions
953 //===-------------------------------
955 def uimm5_asmoperand : AsmOperandClass {
957 let PredicateMethod = "isUImm<5>";
958 let RenderMethod = "addImmOperands";
961 def uimm6_asmoperand : AsmOperandClass {
963 let PredicateMethod = "isUImm<6>";
964 let RenderMethod = "addImmOperands";
967 def bitfield32_imm : Operand<i64>,
968 ImmLeaf<i64, [{ return Imm >= 0 && Imm < 32; }]> {
969 let ParserMatchClass = uimm5_asmoperand;
971 let DecoderMethod = "DecodeBitfield32ImmOperand";
975 def bitfield64_imm : Operand<i64>,
976 ImmLeaf<i64, [{ return Imm >= 0 && Imm < 64; }]> {
977 let ParserMatchClass = uimm6_asmoperand;
979 // Default decoder works in 64-bit case: the 6-bit field can take any value.
982 multiclass A64I_bitfieldSizes<bits<2> opc, string asmop> {
983 def wwii : A64I_bitfield<0b0, opc, 0b0, (outs GPR32:$Rd),
984 (ins GPR32:$Rn, bitfield32_imm:$ImmR, bitfield32_imm:$ImmS),
985 !strconcat(asmop, "\t$Rd, $Rn, $ImmR, $ImmS"),
987 let DecoderMethod = "DecodeBitfieldInstruction";
990 def xxii : A64I_bitfield<0b1, opc, 0b1, (outs GPR64:$Rd),
991 (ins GPR64:$Rn, bitfield64_imm:$ImmR, bitfield64_imm:$ImmS),
992 !strconcat(asmop, "\t$Rd, $Rn, $ImmR, $ImmS"),
994 let DecoderMethod = "DecodeBitfieldInstruction";
998 defm SBFM : A64I_bitfieldSizes<0b00, "sbfm">;
999 defm UBFM : A64I_bitfieldSizes<0b10, "ubfm">;
1001 // BFM instructions modify the destination register rather than defining it
1004 A64I_bitfield<0b0, 0b01, 0b0, (outs GPR32:$Rd),
1005 (ins GPR32:$src, GPR32:$Rn, bitfield32_imm:$ImmR, bitfield32_imm:$ImmS),
1006 "bfm\t$Rd, $Rn, $ImmR, $ImmS", [], NoItinerary> {
1007 let DecoderMethod = "DecodeBitfieldInstruction";
1008 let Constraints = "$src = $Rd";
1012 A64I_bitfield<0b1, 0b01, 0b1, (outs GPR64:$Rd),
1013 (ins GPR64:$src, GPR64:$Rn, bitfield64_imm:$ImmR, bitfield64_imm:$ImmS),
1014 "bfm\t$Rd, $Rn, $ImmR, $ImmS", [], NoItinerary> {
1015 let DecoderMethod = "DecodeBitfieldInstruction";
1016 let Constraints = "$src = $Rd";
1020 //===-------------------------------
1021 // 2. Extend aliases to 64-bit dest
1022 //===-------------------------------
1024 // Unfortunately the extensions that end up as 64-bits cannot be handled by an
1025 // instruction alias: their syntax is (for example) "SXTB x0, w0", which needs
1026 // to be mapped to "SBFM x0, x0, #0, 7" (changing the class of Rn). InstAlias is
1027 // not capable of such a map as far as I'm aware
1029 // Note that these instructions are strictly more specific than the
1030 // BFM ones (in ImmR) so they can handle their own decoding.
1031 class A64I_bf_ext<bit sf, bits<2> opc, RegisterClass GPRDest, string asmop,
1032 bits<6> imms, dag pattern>
1033 : A64I_bitfield<sf, opc, sf,
1034 (outs GPRDest:$Rd), (ins GPR32:$Rn),
1035 !strconcat(asmop, "\t$Rd, $Rn"),
1036 [(set GPRDest:$Rd, pattern)], NoItinerary> {
1037 let ImmR = 0b000000;
1041 // Signed extensions
1042 def SXTBxw : A64I_bf_ext<0b1, 0b00, GPR64, "sxtb", 7,
1043 (sext_inreg (anyext GPR32:$Rn), i8)>;
1044 def SXTBww : A64I_bf_ext<0b0, 0b00, GPR32, "sxtb", 7,
1045 (sext_inreg GPR32:$Rn, i8)>;
1046 def SXTHxw : A64I_bf_ext<0b1, 0b00, GPR64, "sxth", 15,
1047 (sext_inreg (anyext GPR32:$Rn), i16)>;
1048 def SXTHww : A64I_bf_ext<0b0, 0b00, GPR32, "sxth", 15,
1049 (sext_inreg GPR32:$Rn, i16)>;
1050 def SXTWxw : A64I_bf_ext<0b1, 0b00, GPR64, "sxtw", 31, (sext GPR32:$Rn)>;
1052 // Unsigned extensions
1053 def UXTBww : A64I_bf_ext<0b0, 0b10, GPR32, "uxtb", 7,
1054 (and GPR32:$Rn, 255)>;
1055 def UXTHww : A64I_bf_ext<0b0, 0b10, GPR32, "uxth", 15,
1056 (and GPR32:$Rn, 65535)>;
1058 // The 64-bit unsigned variants are not strictly architectural but recommended
1060 let isAsmParserOnly = 1 in {
1061 def UXTBxw : A64I_bf_ext<0b0, 0b10, GPR64, "uxtb", 7,
1062 (and (anyext GPR32:$Rn), 255)>;
1063 def UXTHxw : A64I_bf_ext<0b0, 0b10, GPR64, "uxth", 15,
1064 (and (anyext GPR32:$Rn), 65535)>;
1067 // Extra patterns for when the source register is actually 64-bits
1068 // too. There's no architectural difference here, it's just LLVM
1069 // shinanigans. There's no need for equivalent zero-extension patterns
1070 // because they'll already be caught by logical (immediate) matching.
1071 def : Pat<(sext_inreg GPR64:$Rn, i8),
1072 (SXTBxw (EXTRACT_SUBREG GPR64:$Rn, sub_32))>;
1073 def : Pat<(sext_inreg GPR64:$Rn, i16),
1074 (SXTHxw (EXTRACT_SUBREG GPR64:$Rn, sub_32))>;
1075 def : Pat<(sext_inreg GPR64:$Rn, i32),
1076 (SXTWxw (EXTRACT_SUBREG GPR64:$Rn, sub_32))>;
1079 //===-------------------------------
1080 // 3. Aliases for ASR and LSR (the simple shifts)
1081 //===-------------------------------
1083 // These also handle their own decoding because ImmS being set makes
1084 // them take precedence over BFM.
1085 multiclass A64I_shift<bits<2> opc, string asmop, SDNode opnode> {
1086 def wwi : A64I_bitfield<0b0, opc, 0b0,
1087 (outs GPR32:$Rd), (ins GPR32:$Rn, bitfield32_imm:$ImmR),
1088 !strconcat(asmop, "\t$Rd, $Rn, $ImmR"),
1089 [(set GPR32:$Rd, (opnode GPR32:$Rn, bitfield32_imm:$ImmR))],
1094 def xxi : A64I_bitfield<0b1, opc, 0b1,
1095 (outs GPR64:$Rd), (ins GPR64:$Rn, bitfield64_imm:$ImmR),
1096 !strconcat(asmop, "\t$Rd, $Rn, $ImmR"),
1097 [(set GPR64:$Rd, (opnode GPR64:$Rn, bitfield64_imm:$ImmR))],
1104 defm ASR : A64I_shift<0b00, "asr", sra>;
1105 defm LSR : A64I_shift<0b10, "lsr", srl>;
1107 //===-------------------------------
1108 // 4. Aliases for LSL
1109 //===-------------------------------
1111 // Unfortunately LSL and subsequent aliases are much more complicated. We need
1112 // to be able to say certain output instruction fields depend in a complex
1113 // manner on combinations of input assembly fields).
1115 // MIOperandInfo *might* have been able to do it, but at the cost of
1116 // significantly more C++ code.
1118 // N.b. contrary to usual practice these operands store the shift rather than
1119 // the machine bits in an MCInst. The complexity overhead of consistency
1120 // outweighed the benefits in this case (custom asmparser, printer and selection
1121 // vs custom encoder).
1122 def bitfield32_lsl_imm : Operand<i64>,
1123 ImmLeaf<i64, [{ return Imm >= 0 && Imm <= 31; }]> {
1124 let ParserMatchClass = uimm5_asmoperand;
1125 let EncoderMethod = "getBitfield32LSLOpValue";
1128 def bitfield64_lsl_imm : Operand<i64>,
1129 ImmLeaf<i64, [{ return Imm >= 0 && Imm <= 63; }]> {
1130 let ParserMatchClass = uimm6_asmoperand;
1131 let EncoderMethod = "getBitfield64LSLOpValue";
1134 class A64I_bitfield_lsl<bit sf, RegisterClass GPR, Operand operand>
1135 : A64I_bitfield<sf, 0b10, sf, (outs GPR:$Rd), (ins GPR:$Rn, operand:$FullImm),
1136 "lsl\t$Rd, $Rn, $FullImm",
1137 [(set GPR:$Rd, (shl GPR:$Rn, operand:$FullImm))],
1140 let ImmR = FullImm{5-0};
1141 let ImmS = FullImm{11-6};
1143 // No disassembler allowed because it would overlap with BFM which does the
1145 let isAsmParserOnly = 1;
1148 def LSLwwi : A64I_bitfield_lsl<0b0, GPR32, bitfield32_lsl_imm>;
1149 def LSLxxi : A64I_bitfield_lsl<0b1, GPR64, bitfield64_lsl_imm>;
1151 //===-------------------------------
1152 // 5. Aliases for bitfield extract instructions
1153 //===-------------------------------
1155 def bfx32_width_asmoperand : AsmOperandClass {
1156 let Name = "BFX32Width";
1157 let PredicateMethod = "isBitfieldWidth<32>";
1158 let RenderMethod = "addBFXWidthOperands";
1161 def bfx32_width : Operand<i64>, ImmLeaf<i64, [{ return true; }]> {
1162 let PrintMethod = "printBFXWidthOperand";
1163 let ParserMatchClass = bfx32_width_asmoperand;
1166 def bfx64_width_asmoperand : AsmOperandClass {
1167 let Name = "BFX64Width";
1168 let PredicateMethod = "isBitfieldWidth<64>";
1169 let RenderMethod = "addBFXWidthOperands";
1172 def bfx64_width : Operand<i64> {
1173 let PrintMethod = "printBFXWidthOperand";
1174 let ParserMatchClass = bfx64_width_asmoperand;
1178 multiclass A64I_bitfield_extract<bits<2> opc, string asmop, SDNode op> {
1179 def wwii : A64I_bitfield<0b0, opc, 0b0, (outs GPR32:$Rd),
1180 (ins GPR32:$Rn, bitfield32_imm:$ImmR, bfx32_width:$ImmS),
1181 !strconcat(asmop, "\t$Rd, $Rn, $ImmR, $ImmS"),
1182 [(set GPR32:$Rd, (op GPR32:$Rn, imm:$ImmR, imm:$ImmS))],
1184 // As above, no disassembler allowed.
1185 let isAsmParserOnly = 1;
1188 def xxii : A64I_bitfield<0b1, opc, 0b1, (outs GPR64:$Rd),
1189 (ins GPR64:$Rn, bitfield64_imm:$ImmR, bfx64_width:$ImmS),
1190 !strconcat(asmop, "\t$Rd, $Rn, $ImmR, $ImmS"),
1191 [(set GPR64:$Rd, (op GPR64:$Rn, imm:$ImmR, imm:$ImmS))],
1193 // As above, no disassembler allowed.
1194 let isAsmParserOnly = 1;
1198 defm SBFX : A64I_bitfield_extract<0b00, "sbfx", A64Sbfx>;
1199 defm UBFX : A64I_bitfield_extract<0b10, "ubfx", A64Ubfx>;
1201 // Again, variants based on BFM modify Rd so need it as an input too.
1202 def BFXILwwii : A64I_bitfield<0b0, 0b01, 0b0, (outs GPR32:$Rd),
1203 (ins GPR32:$src, GPR32:$Rn, bitfield32_imm:$ImmR, bfx32_width:$ImmS),
1204 "bfxil\t$Rd, $Rn, $ImmR, $ImmS", [], NoItinerary> {
1205 // As above, no disassembler allowed.
1206 let isAsmParserOnly = 1;
1207 let Constraints = "$src = $Rd";
1210 def BFXILxxii : A64I_bitfield<0b1, 0b01, 0b1, (outs GPR64:$Rd),
1211 (ins GPR64:$src, GPR64:$Rn, bitfield64_imm:$ImmR, bfx64_width:$ImmS),
1212 "bfxil\t$Rd, $Rn, $ImmR, $ImmS", [], NoItinerary> {
1213 // As above, no disassembler allowed.
1214 let isAsmParserOnly = 1;
1215 let Constraints = "$src = $Rd";
1218 // SBFX instructions can do a 1-instruction sign-extension of boolean values.
1219 def : Pat<(sext_inreg GPR64:$Rn, i1), (SBFXxxii GPR64:$Rn, 0, 0)>;
1220 def : Pat<(sext_inreg GPR32:$Rn, i1), (SBFXwwii GPR32:$Rn, 0, 0)>;
1221 def : Pat<(i64 (sext_inreg (anyext GPR32:$Rn), i1)),
1222 (SBFXxxii (SUBREG_TO_REG (i64 0), GPR32:$Rn, sub_32), 0, 0)>;
1224 // UBFX makes sense as an implementation of a 64-bit zero-extension too. Could
1225 // use either 64-bit or 32-bit variant, but 32-bit might be more efficient.
1226 def : Pat<(zext GPR32:$Rn), (SUBREG_TO_REG (i64 0), (UBFXwwii GPR32:$Rn, 0, 31),
1229 //===-------------------------------
1230 // 6. Aliases for bitfield insert instructions
1231 //===-------------------------------
1233 def bfi32_lsb_asmoperand : AsmOperandClass {
1234 let Name = "BFI32LSB";
1235 let PredicateMethod = "isUImm<5>";
1236 let RenderMethod = "addBFILSBOperands<32>";
1239 def bfi32_lsb : Operand<i64>,
1240 ImmLeaf<i64, [{ return Imm >= 0 && Imm <= 31; }]> {
1241 let PrintMethod = "printBFILSBOperand<32>";
1242 let ParserMatchClass = bfi32_lsb_asmoperand;
1245 def bfi64_lsb_asmoperand : AsmOperandClass {
1246 let Name = "BFI64LSB";
1247 let PredicateMethod = "isUImm<6>";
1248 let RenderMethod = "addBFILSBOperands<64>";
1251 def bfi64_lsb : Operand<i64>,
1252 ImmLeaf<i64, [{ return Imm >= 0 && Imm <= 63; }]> {
1253 let PrintMethod = "printBFILSBOperand<64>";
1254 let ParserMatchClass = bfi64_lsb_asmoperand;
1257 // Width verification is performed during conversion so width operand can be
1258 // shared between 32/64-bit cases. Still needed for the print method though
1259 // because ImmR encodes "width - 1".
1260 def bfi32_width_asmoperand : AsmOperandClass {
1261 let Name = "BFI32Width";
1262 let PredicateMethod = "isBitfieldWidth<32>";
1263 let RenderMethod = "addBFIWidthOperands";
1266 def bfi32_width : Operand<i64>,
1267 ImmLeaf<i64, [{ return Imm >= 1 && Imm <= 32; }]> {
1268 let PrintMethod = "printBFIWidthOperand";
1269 let ParserMatchClass = bfi32_width_asmoperand;
1272 def bfi64_width_asmoperand : AsmOperandClass {
1273 let Name = "BFI64Width";
1274 let PredicateMethod = "isBitfieldWidth<64>";
1275 let RenderMethod = "addBFIWidthOperands";
1278 def bfi64_width : Operand<i64>,
1279 ImmLeaf<i64, [{ return Imm >= 1 && Imm <= 64; }]> {
1280 let PrintMethod = "printBFIWidthOperand";
1281 let ParserMatchClass = bfi64_width_asmoperand;
1284 multiclass A64I_bitfield_insert<bits<2> opc, string asmop> {
1285 def wwii : A64I_bitfield<0b0, opc, 0b0, (outs GPR32:$Rd),
1286 (ins GPR32:$Rn, bfi32_lsb:$ImmR, bfi32_width:$ImmS),
1287 !strconcat(asmop, "\t$Rd, $Rn, $ImmR, $ImmS"),
1289 // As above, no disassembler allowed.
1290 let isAsmParserOnly = 1;
1293 def xxii : A64I_bitfield<0b1, opc, 0b1, (outs GPR64:$Rd),
1294 (ins GPR64:$Rn, bfi64_lsb:$ImmR, bfi64_width:$ImmS),
1295 !strconcat(asmop, "\t$Rd, $Rn, $ImmR, $ImmS"),
1297 // As above, no disassembler allowed.
1298 let isAsmParserOnly = 1;
1302 defm SBFIZ : A64I_bitfield_insert<0b00, "sbfiz">;
1303 defm UBFIZ : A64I_bitfield_insert<0b10, "ubfiz">;
1306 def BFIwwii : A64I_bitfield<0b0, 0b01, 0b0, (outs GPR32:$Rd),
1307 (ins GPR32:$src, GPR32:$Rn, bfi32_lsb:$ImmR, bfi32_width:$ImmS),
1308 "bfi\t$Rd, $Rn, $ImmR, $ImmS", [], NoItinerary> {
1309 // As above, no disassembler allowed.
1310 let isAsmParserOnly = 1;
1311 let Constraints = "$src = $Rd";
1314 def BFIxxii : A64I_bitfield<0b1, 0b01, 0b1, (outs GPR64:$Rd),
1315 (ins GPR64:$src, GPR64:$Rn, bfi64_lsb:$ImmR, bfi64_width:$ImmS),
1316 "bfi\t$Rd, $Rn, $ImmR, $ImmS", [], NoItinerary> {
1317 // As above, no disassembler allowed.
1318 let isAsmParserOnly = 1;
1319 let Constraints = "$src = $Rd";
1322 //===----------------------------------------------------------------------===//
1323 // Compare and branch (immediate)
1324 //===----------------------------------------------------------------------===//
1325 // Contains: CBZ, CBNZ
1327 class label_asmoperand<int width, int scale> : AsmOperandClass {
1328 let Name = "Label" # width # "_" # scale;
1329 let PredicateMethod = "isLabel<" # width # "," # scale # ">";
1330 let RenderMethod = "addLabelOperands<" # width # ", " # scale # ">";
1333 def label_wid19_scal4_asmoperand : label_asmoperand<19, 4>;
1335 // All conditional immediate branches are the same really: 19 signed bits scaled
1336 // by the instruction-size (4).
1337 def bcc_target : Operand<OtherVT> {
1338 // This label is a 19-bit offset from PC, scaled by the instruction-width: 4.
1339 let ParserMatchClass = label_wid19_scal4_asmoperand;
1340 let PrintMethod = "printLabelOperand<19, 4>";
1341 let EncoderMethod = "getLabelOpValue<AArch64::fixup_a64_condbr>";
1342 let OperandType = "OPERAND_PCREL";
1345 multiclass cmpbr_sizes<bit op, string asmop, ImmLeaf SETOP> {
1346 let isBranch = 1, isTerminator = 1 in {
1347 def x : A64I_cmpbr<0b1, op,
1349 (ins GPR64:$Rt, bcc_target:$Label),
1350 !strconcat(asmop,"\t$Rt, $Label"),
1351 [(A64br_cc (A64cmp GPR64:$Rt, 0), SETOP, bb:$Label)],
1354 def w : A64I_cmpbr<0b0, op,
1356 (ins GPR32:$Rt, bcc_target:$Label),
1357 !strconcat(asmop,"\t$Rt, $Label"),
1358 [(A64br_cc (A64cmp GPR32:$Rt, 0), SETOP, bb:$Label)],
1363 defm CBZ : cmpbr_sizes<0b0, "cbz", ImmLeaf<i32, [{
1364 return Imm == A64CC::EQ;
1366 defm CBNZ : cmpbr_sizes<0b1, "cbnz", ImmLeaf<i32, [{
1367 return Imm == A64CC::NE;
1370 //===----------------------------------------------------------------------===//
1371 // Conditional branch (immediate) instructions
1372 //===----------------------------------------------------------------------===//
1375 def cond_code_asmoperand : AsmOperandClass {
1376 let Name = "CondCode";
1379 def cond_code : Operand<i32>, ImmLeaf<i32, [{
1380 return Imm >= 0 && Imm <= 15;
1382 let PrintMethod = "printCondCodeOperand";
1383 let ParserMatchClass = cond_code_asmoperand;
1386 def Bcc : A64I_condbr<0b0, 0b0, (outs),
1387 (ins cond_code:$Cond, bcc_target:$Label),
1388 "b.$Cond $Label", [(A64br_cc NZCV, (i32 imm:$Cond), bb:$Label)],
1392 let isTerminator = 1;
1395 //===----------------------------------------------------------------------===//
1396 // Conditional compare (immediate) instructions
1397 //===----------------------------------------------------------------------===//
1398 // Contains: CCMN, CCMP
1400 def uimm4_asmoperand : AsmOperandClass {
1402 let PredicateMethod = "isUImm<4>";
1403 let RenderMethod = "addImmOperands";
1406 def uimm4 : Operand<i32> {
1407 let ParserMatchClass = uimm4_asmoperand;
1410 def uimm5 : Operand<i32> {
1411 let ParserMatchClass = uimm5_asmoperand;
1414 // The only difference between this operand and the one for instructions like
1415 // B.cc is that it's parsed manually. The other get parsed implicitly as part of
1416 // the mnemonic handling.
1417 def cond_code_op_asmoperand : AsmOperandClass {
1418 let Name = "CondCodeOp";
1419 let RenderMethod = "addCondCodeOperands";
1420 let PredicateMethod = "isCondCode";
1421 let ParserMethod = "ParseCondCodeOperand";
1424 def cond_code_op : Operand<i32> {
1425 let PrintMethod = "printCondCodeOperand";
1426 let ParserMatchClass = cond_code_op_asmoperand;
1429 class A64I_condcmpimmImpl<bit sf, bit op, RegisterClass GPR, string asmop>
1430 : A64I_condcmpimm<sf, op, 0b0, 0b0, 0b1, (outs),
1431 (ins GPR:$Rn, uimm5:$UImm5, uimm4:$NZCVImm, cond_code_op:$Cond),
1432 !strconcat(asmop, "\t$Rn, $UImm5, $NZCVImm, $Cond"),
1437 def CCMNwi : A64I_condcmpimmImpl<0b0, 0b0, GPR32, "ccmn">;
1438 def CCMNxi : A64I_condcmpimmImpl<0b1, 0b0, GPR64, "ccmn">;
1439 def CCMPwi : A64I_condcmpimmImpl<0b0, 0b1, GPR32, "ccmp">;
1440 def CCMPxi : A64I_condcmpimmImpl<0b1, 0b1, GPR64, "ccmp">;
1442 //===----------------------------------------------------------------------===//
1443 // Conditional compare (register) instructions
1444 //===----------------------------------------------------------------------===//
1445 // Contains: CCMN, CCMP
1447 class A64I_condcmpregImpl<bit sf, bit op, RegisterClass GPR, string asmop>
1448 : A64I_condcmpreg<sf, op, 0b0, 0b0, 0b1,
1450 (ins GPR:$Rn, GPR:$Rm, uimm4:$NZCVImm, cond_code_op:$Cond),
1451 !strconcat(asmop, "\t$Rn, $Rm, $NZCVImm, $Cond"),
1456 def CCMNww : A64I_condcmpregImpl<0b0, 0b0, GPR32, "ccmn">;
1457 def CCMNxx : A64I_condcmpregImpl<0b1, 0b0, GPR64, "ccmn">;
1458 def CCMPww : A64I_condcmpregImpl<0b0, 0b1, GPR32, "ccmp">;
1459 def CCMPxx : A64I_condcmpregImpl<0b1, 0b1, GPR64, "ccmp">;
1461 //===----------------------------------------------------------------------===//
1462 // Conditional select instructions
1463 //===----------------------------------------------------------------------===//
1464 // Contains: CSEL, CSINC, CSINV, CSNEG + aliases CSET, CSETM, CINC, CINV, CNEG
1466 // Condition code which is encoded as the inversion (semantically rather than
1467 // bitwise) in the instruction.
1468 def inv_cond_code_op_asmoperand : AsmOperandClass {
1469 let Name = "InvCondCodeOp";
1470 let RenderMethod = "addInvCondCodeOperands";
1471 let PredicateMethod = "isCondCode";
1472 let ParserMethod = "ParseCondCodeOperand";
1475 def inv_cond_code_op : Operand<i32> {
1476 let ParserMatchClass = inv_cond_code_op_asmoperand;
1479 // Having a separate operand for the selectable use-case is debatable, but gives
1480 // consistency with cond_code.
1481 def inv_cond_XFORM : SDNodeXForm<imm, [{
1482 A64CC::CondCodes CC = static_cast<A64CC::CondCodes>(N->getZExtValue());
1483 return CurDAG->getTargetConstant(A64InvertCondCode(CC), MVT::i32);
1487 : ImmLeaf<i32, [{ return Imm >= 0 && Imm <= 15; }], inv_cond_XFORM>;
1490 multiclass A64I_condselSizes<bit op, bits<2> op2, string asmop,
1491 SDPatternOperator select> {
1492 let Uses = [NZCV] in {
1493 def wwwc : A64I_condsel<0b0, op, 0b0, op2,
1495 (ins GPR32:$Rn, GPR32:$Rm, cond_code_op:$Cond),
1496 !strconcat(asmop, "\t$Rd, $Rn, $Rm, $Cond"),
1497 [(set GPR32:$Rd, (select GPR32:$Rn, GPR32:$Rm))],
1501 def xxxc : A64I_condsel<0b1, op, 0b0, op2,
1503 (ins GPR64:$Rn, GPR64:$Rm, cond_code_op:$Cond),
1504 !strconcat(asmop, "\t$Rd, $Rn, $Rm, $Cond"),
1505 [(set GPR64:$Rd, (select GPR64:$Rn, GPR64:$Rm))],
1511 : PatFrag<(ops node:$lhs, node:$rhs),
1512 (A64select_cc NZCV, node:$lhs, node:$rhs, (i32 imm:$Cond))>;
1514 class complex_select<SDPatternOperator opnode>
1515 : PatFrag<(ops node:$lhs, node:$rhs),
1516 (A64select_cc NZCV, node:$lhs, (opnode node:$rhs), (i32 imm:$Cond))>;
1519 defm CSEL : A64I_condselSizes<0b0, 0b00, "csel", simple_select>;
1520 defm CSINC : A64I_condselSizes<0b0, 0b01, "csinc",
1521 complex_select<PatFrag<(ops node:$val),
1522 (add node:$val, 1)>>>;
1523 defm CSINV : A64I_condselSizes<0b1, 0b00, "csinv", complex_select<not>>;
1524 defm CSNEG : A64I_condselSizes<0b1, 0b01, "csneg", complex_select<ineg>>;
1526 // Now the instruction aliases, which fit nicely into LLVM's model:
1528 def : InstAlias<"cset $Rd, $Cond",
1529 (CSINCwwwc GPR32:$Rd, WZR, WZR, inv_cond_code_op:$Cond)>;
1530 def : InstAlias<"cset $Rd, $Cond",
1531 (CSINCxxxc GPR64:$Rd, XZR, XZR, inv_cond_code_op:$Cond)>;
1532 def : InstAlias<"csetm $Rd, $Cond",
1533 (CSINVwwwc GPR32:$Rd, WZR, WZR, inv_cond_code_op:$Cond)>;
1534 def : InstAlias<"csetm $Rd, $Cond",
1535 (CSINVxxxc GPR64:$Rd, XZR, XZR, inv_cond_code_op:$Cond)>;
1536 def : InstAlias<"cinc $Rd, $Rn, $Cond",
1537 (CSINCwwwc GPR32:$Rd, GPR32:$Rn, GPR32:$Rn, inv_cond_code_op:$Cond)>;
1538 def : InstAlias<"cinc $Rd, $Rn, $Cond",
1539 (CSINCxxxc GPR64:$Rd, GPR64:$Rn, GPR64:$Rn, inv_cond_code_op:$Cond)>;
1540 def : InstAlias<"cinv $Rd, $Rn, $Cond",
1541 (CSINVwwwc GPR32:$Rd, GPR32:$Rn, GPR32:$Rn, inv_cond_code_op:$Cond)>;
1542 def : InstAlias<"cinv $Rd, $Rn, $Cond",
1543 (CSINVxxxc GPR64:$Rd, GPR64:$Rn, GPR64:$Rn, inv_cond_code_op:$Cond)>;
1544 def : InstAlias<"cneg $Rd, $Rn, $Cond",
1545 (CSNEGwwwc GPR32:$Rd, GPR32:$Rn, GPR32:$Rn, inv_cond_code_op:$Cond)>;
1546 def : InstAlias<"cneg $Rd, $Rn, $Cond",
1547 (CSNEGxxxc GPR64:$Rd, GPR64:$Rn, GPR64:$Rn, inv_cond_code_op:$Cond)>;
1549 // Finally some helper patterns.
1551 // For CSET (a.k.a. zero-extension of icmp)
1552 def : Pat<(A64select_cc NZCV, 0, 1, cond_code:$Cond),
1553 (CSINCwwwc WZR, WZR, cond_code:$Cond)>;
1554 def : Pat<(A64select_cc NZCV, 1, 0, inv_cond_code:$Cond),
1555 (CSINCwwwc WZR, WZR, inv_cond_code:$Cond)>;
1557 def : Pat<(A64select_cc NZCV, 0, 1, cond_code:$Cond),
1558 (CSINCxxxc XZR, XZR, cond_code:$Cond)>;
1559 def : Pat<(A64select_cc NZCV, 1, 0, inv_cond_code:$Cond),
1560 (CSINCxxxc XZR, XZR, inv_cond_code:$Cond)>;
1562 // For CSETM (a.k.a. sign-extension of icmp)
1563 def : Pat<(A64select_cc NZCV, 0, -1, cond_code:$Cond),
1564 (CSINVwwwc WZR, WZR, cond_code:$Cond)>;
1565 def : Pat<(A64select_cc NZCV, -1, 0, inv_cond_code:$Cond),
1566 (CSINVwwwc WZR, WZR, inv_cond_code:$Cond)>;
1568 def : Pat<(A64select_cc NZCV, 0, -1, cond_code:$Cond),
1569 (CSINVxxxc XZR, XZR, cond_code:$Cond)>;
1570 def : Pat<(A64select_cc NZCV, -1, 0, inv_cond_code:$Cond),
1571 (CSINVxxxc XZR, XZR, inv_cond_code:$Cond)>;
1573 // CINC, CINV and CNEG get dealt with automatically, which leaves the issue of
1574 // commutativity. The instructions are to complex for isCommutable to be used,
1575 // so we have to create the patterns manually:
1577 // No commutable pattern for CSEL since the commuted version is isomorphic.
1580 def :Pat<(A64select_cc NZCV, (add GPR32:$Rm, 1), GPR32:$Rn,
1581 inv_cond_code:$Cond),
1582 (CSINCwwwc GPR32:$Rn, GPR32:$Rm, inv_cond_code:$Cond)>;
1583 def :Pat<(A64select_cc NZCV, (add GPR64:$Rm, 1), GPR64:$Rn,
1584 inv_cond_code:$Cond),
1585 (CSINCxxxc GPR64:$Rn, GPR64:$Rm, inv_cond_code:$Cond)>;
1588 def :Pat<(A64select_cc NZCV, (not GPR32:$Rm), GPR32:$Rn, inv_cond_code:$Cond),
1589 (CSINVwwwc GPR32:$Rn, GPR32:$Rm, inv_cond_code:$Cond)>;
1590 def :Pat<(A64select_cc NZCV, (not GPR64:$Rm), GPR64:$Rn, inv_cond_code:$Cond),
1591 (CSINVxxxc GPR64:$Rn, GPR64:$Rm, inv_cond_code:$Cond)>;
1594 def :Pat<(A64select_cc NZCV, (ineg GPR32:$Rm), GPR32:$Rn, inv_cond_code:$Cond),
1595 (CSNEGwwwc GPR32:$Rn, GPR32:$Rm, inv_cond_code:$Cond)>;
1596 def :Pat<(A64select_cc NZCV, (ineg GPR64:$Rm), GPR64:$Rn, inv_cond_code:$Cond),
1597 (CSNEGxxxc GPR64:$Rn, GPR64:$Rm, inv_cond_code:$Cond)>;
1599 //===----------------------------------------------------------------------===//
1600 // Data Processing (1 source) instructions
1601 //===----------------------------------------------------------------------===//
1602 // Contains: RBIT, REV16, REV, REV32, CLZ, CLS.
1604 // We define an unary operator which always fails. We will use this to
1605 // define unary operators that cannot be matched.
1607 class A64I_dp_1src_impl<bit sf, bits<6> opcode, string asmop,
1608 list<dag> patterns, RegisterClass GPRrc,
1609 InstrItinClass itin>:
1614 !strconcat(asmop, "\t$Rd, $Rn"),
1620 multiclass A64I_dp_1src <bits<6> opcode, string asmop> {
1621 let neverHasSideEffects = 1 in {
1622 def ww : A64I_dp_1src_impl<0b0, opcode, asmop, [], GPR32, NoItinerary>;
1623 def xx : A64I_dp_1src_impl<0b1, opcode, asmop, [], GPR64, NoItinerary>;
1627 defm RBIT : A64I_dp_1src<0b000000, "rbit">;
1628 defm CLS : A64I_dp_1src<0b000101, "cls">;
1629 defm CLZ : A64I_dp_1src<0b000100, "clz">;
1631 def : Pat<(ctlz GPR32:$Rn), (CLZww GPR32:$Rn)>;
1632 def : Pat<(ctlz GPR64:$Rn), (CLZxx GPR64:$Rn)>;
1633 def : Pat<(ctlz_zero_undef GPR32:$Rn), (CLZww GPR32:$Rn)>;
1634 def : Pat<(ctlz_zero_undef GPR64:$Rn), (CLZxx GPR64:$Rn)>;
1636 def : Pat<(cttz GPR32:$Rn), (CLZww (RBITww GPR32:$Rn))>;
1637 def : Pat<(cttz GPR64:$Rn), (CLZxx (RBITxx GPR64:$Rn))>;
1638 def : Pat<(cttz_zero_undef GPR32:$Rn), (CLZww (RBITww GPR32:$Rn))>;
1639 def : Pat<(cttz_zero_undef GPR64:$Rn), (CLZxx (RBITxx GPR64:$Rn))>;
1642 def REVww : A64I_dp_1src_impl<0b0, 0b000010, "rev",
1643 [(set GPR32:$Rd, (bswap GPR32:$Rn))],
1644 GPR32, NoItinerary>;
1645 def REVxx : A64I_dp_1src_impl<0b1, 0b000011, "rev",
1646 [(set GPR64:$Rd, (bswap GPR64:$Rn))],
1647 GPR64, NoItinerary>;
1648 def REV32xx : A64I_dp_1src_impl<0b1, 0b000010, "rev32",
1649 [(set GPR64:$Rd, (bswap (rotr GPR64:$Rn, (i64 32))))],
1650 GPR64, NoItinerary>;
1651 def REV16ww : A64I_dp_1src_impl<0b0, 0b000001, "rev16",
1652 [(set GPR32:$Rd, (bswap (rotr GPR32:$Rn, (i64 16))))],
1655 def REV16xx : A64I_dp_1src_impl<0b1, 0b000001, "rev16", [], GPR64, NoItinerary>;
1657 //===----------------------------------------------------------------------===//
1658 // Data Processing (2 sources) instructions
1659 //===----------------------------------------------------------------------===//
1660 // Contains: UDIV, SDIV, LSLV, LSRV, ASRV, RORV + aliases LSL, LSR, ASR, ROR
1662 class dp_2src_impl<bit sf, bits<6> opcode, string asmop, list<dag> patterns,
1663 RegisterClass GPRsp,
1664 InstrItinClass itin>:
1668 !strconcat(asmop, "\t$Rd, $Rn, $Rm"),
1670 (ins GPRsp:$Rn, GPRsp:$Rm),
1674 multiclass dp_2src_zext <bits<6> opcode, string asmop, SDPatternOperator op> {
1675 def www : dp_2src_impl<0b0,
1679 (op GPR32:$Rn, (i64 (zext GPR32:$Rm))))],
1682 def xxx : dp_2src_impl<0b1,
1685 [(set GPR64:$Rd, (op GPR64:$Rn, GPR64:$Rm))],
1691 multiclass dp_2src <bits<6> opcode, string asmop, SDPatternOperator op> {
1692 def www : dp_2src_impl<0b0,
1695 [(set GPR32:$Rd, (op GPR32:$Rn, GPR32:$Rm))],
1698 def xxx : dp_2src_impl<0b1,
1701 [(set GPR64:$Rd, (op GPR64:$Rn, GPR64:$Rm))],
1706 // Here we define the data processing 2 source instructions.
1707 defm UDIV : dp_2src<0b000010, "udiv", udiv>;
1708 defm SDIV : dp_2src<0b000011, "sdiv", sdiv>;
1710 defm LSLV : dp_2src_zext<0b001000, "lsl", shl>;
1711 defm LSRV : dp_2src_zext<0b001001, "lsr", srl>;
1712 defm ASRV : dp_2src_zext<0b001010, "asr", sra>;
1713 defm RORV : dp_2src_zext<0b001011, "ror", rotr>;
1715 // Extra patterns for an incoming 64-bit value for a 32-bit
1716 // operation. Since the LLVM operations are undefined (as in C) if the
1717 // RHS is out of range, it's perfectly permissible to discard the high
1718 // bits of the GPR64.
1719 def : Pat<(shl GPR32:$Rn, GPR64:$Rm),
1720 (LSLVwww GPR32:$Rn, (EXTRACT_SUBREG GPR64:$Rm, sub_32))>;
1721 def : Pat<(srl GPR32:$Rn, GPR64:$Rm),
1722 (LSRVwww GPR32:$Rn, (EXTRACT_SUBREG GPR64:$Rm, sub_32))>;
1723 def : Pat<(sra GPR32:$Rn, GPR64:$Rm),
1724 (ASRVwww GPR32:$Rn, (EXTRACT_SUBREG GPR64:$Rm, sub_32))>;
1725 def : Pat<(rotr GPR32:$Rn, GPR64:$Rm),
1726 (RORVwww GPR32:$Rn, (EXTRACT_SUBREG GPR64:$Rm, sub_32))>;
1728 // Here we define the aliases for the data processing 2 source instructions.
1729 def LSL_mnemonic : MnemonicAlias<"lslv", "lsl">;
1730 def LSR_mnemonic : MnemonicAlias<"lsrv", "lsr">;
1731 def ASR_menmonic : MnemonicAlias<"asrv", "asr">;
1732 def ROR_menmonic : MnemonicAlias<"rorv", "ror">;
1734 //===----------------------------------------------------------------------===//
1735 // Data Processing (3 sources) instructions
1736 //===----------------------------------------------------------------------===//
1737 // Contains: MADD, MSUB, SMADDL, SMSUBL, SMULH, UMADDL, UMSUBL, UMULH
1738 // + aliases MUL, MNEG, SMULL, SMNEGL, UMULL, UMNEGL
1740 class A64I_dp3_4operand<bit sf, bits<6> opcode, RegisterClass AccReg,
1741 RegisterClass SrcReg, string asmop, dag pattern>
1742 : A64I_dp3<sf, opcode,
1743 (outs AccReg:$Rd), (ins SrcReg:$Rn, SrcReg:$Rm, AccReg:$Ra),
1744 !strconcat(asmop, "\t$Rd, $Rn, $Rm, $Ra"),
1745 [(set AccReg:$Rd, pattern)], NoItinerary> {
1746 RegisterClass AccGPR = AccReg;
1747 RegisterClass SrcGPR = SrcReg;
1750 def MADDwwww : A64I_dp3_4operand<0b0, 0b000000, GPR32, GPR32, "madd",
1751 (add GPR32:$Ra, (mul GPR32:$Rn, GPR32:$Rm))>;
1752 def MADDxxxx : A64I_dp3_4operand<0b1, 0b000000, GPR64, GPR64, "madd",
1753 (add GPR64:$Ra, (mul GPR64:$Rn, GPR64:$Rm))>;
1755 def MSUBwwww : A64I_dp3_4operand<0b0, 0b000001, GPR32, GPR32, "msub",
1756 (sub GPR32:$Ra, (mul GPR32:$Rn, GPR32:$Rm))>;
1757 def MSUBxxxx : A64I_dp3_4operand<0b1, 0b000001, GPR64, GPR64, "msub",
1758 (sub GPR64:$Ra, (mul GPR64:$Rn, GPR64:$Rm))>;
1760 def SMADDLxwwx : A64I_dp3_4operand<0b1, 0b000010, GPR64, GPR32, "smaddl",
1761 (add GPR64:$Ra, (mul (i64 (sext GPR32:$Rn)), (sext GPR32:$Rm)))>;
1762 def SMSUBLxwwx : A64I_dp3_4operand<0b1, 0b000011, GPR64, GPR32, "smsubl",
1763 (sub GPR64:$Ra, (mul (i64 (sext GPR32:$Rn)), (sext GPR32:$Rm)))>;
1765 def UMADDLxwwx : A64I_dp3_4operand<0b1, 0b001010, GPR64, GPR32, "umaddl",
1766 (add GPR64:$Ra, (mul (i64 (zext GPR32:$Rn)), (zext GPR32:$Rm)))>;
1767 def UMSUBLxwwx : A64I_dp3_4operand<0b1, 0b001011, GPR64, GPR32, "umsubl",
1768 (sub GPR64:$Ra, (mul (i64 (zext GPR32:$Rn)), (zext GPR32:$Rm)))>;
1770 let isCommutable = 1, PostEncoderMethod = "fixMulHigh" in {
1771 def UMULHxxx : A64I_dp3<0b1, 0b001100, (outs GPR64:$Rd),
1772 (ins GPR64:$Rn, GPR64:$Rm),
1773 "umulh\t$Rd, $Rn, $Rm",
1774 [(set GPR64:$Rd, (mulhu GPR64:$Rn, GPR64:$Rm))],
1777 def SMULHxxx : A64I_dp3<0b1, 0b000100, (outs GPR64:$Rd),
1778 (ins GPR64:$Rn, GPR64:$Rm),
1779 "smulh\t$Rd, $Rn, $Rm",
1780 [(set GPR64:$Rd, (mulhs GPR64:$Rn, GPR64:$Rm))],
1784 multiclass A64I_dp3_3operand<string asmop, A64I_dp3_4operand INST,
1785 Register ZR, dag pattern> {
1786 def : InstAlias<asmop # " $Rd, $Rn, $Rm",
1787 (INST INST.AccGPR:$Rd, INST.SrcGPR:$Rn, INST.SrcGPR:$Rm, ZR)>;
1789 def : Pat<pattern, (INST INST.SrcGPR:$Rn, INST.SrcGPR:$Rm, ZR)>;
1792 defm : A64I_dp3_3operand<"mul", MADDwwww, WZR, (mul GPR32:$Rn, GPR32:$Rm)>;
1793 defm : A64I_dp3_3operand<"mul", MADDxxxx, XZR, (mul GPR64:$Rn, GPR64:$Rm)>;
1795 defm : A64I_dp3_3operand<"mneg", MSUBwwww, WZR,
1796 (sub 0, (mul GPR32:$Rn, GPR32:$Rm))>;
1797 defm : A64I_dp3_3operand<"mneg", MSUBxxxx, XZR,
1798 (sub 0, (mul GPR64:$Rn, GPR64:$Rm))>;
1800 defm : A64I_dp3_3operand<"smull", SMADDLxwwx, XZR,
1801 (mul (i64 (sext GPR32:$Rn)), (sext GPR32:$Rm))>;
1802 defm : A64I_dp3_3operand<"smnegl", SMSUBLxwwx, XZR,
1803 (sub 0, (mul (i64 (sext GPR32:$Rn)), (sext GPR32:$Rm)))>;
1805 defm : A64I_dp3_3operand<"umull", UMADDLxwwx, XZR,
1806 (mul (i64 (zext GPR32:$Rn)), (zext GPR32:$Rm))>;
1807 defm : A64I_dp3_3operand<"umnegl", UMSUBLxwwx, XZR,
1808 (sub 0, (mul (i64 (zext GPR32:$Rn)), (zext GPR32:$Rm)))>;
1811 //===----------------------------------------------------------------------===//
1812 // Exception generation
1813 //===----------------------------------------------------------------------===//
1814 // Contains: SVC, HVC, SMC, BRK, HLT, DCPS1, DCPS2, DCPS3
1816 def uimm16_asmoperand : AsmOperandClass {
1817 let Name = "UImm16";
1818 let PredicateMethod = "isUImm<16>";
1819 let RenderMethod = "addImmOperands";
1822 def uimm16 : Operand<i32> {
1823 let ParserMatchClass = uimm16_asmoperand;
1826 class A64I_exceptImpl<bits<3> opc, bits<2> ll, string asmop>
1827 : A64I_exception<opc, 0b000, ll, (outs), (ins uimm16:$UImm16),
1828 !strconcat(asmop, "\t$UImm16"), [], NoItinerary> {
1830 let isTerminator = 1;
1833 def SVCi : A64I_exceptImpl<0b000, 0b01, "svc">;
1834 def HVCi : A64I_exceptImpl<0b000, 0b10, "hvc">;
1835 def SMCi : A64I_exceptImpl<0b000, 0b11, "smc">;
1836 def BRKi : A64I_exceptImpl<0b001, 0b00, "brk">;
1837 def HLTi : A64I_exceptImpl<0b010, 0b00, "hlt">;
1839 def DCPS1i : A64I_exceptImpl<0b101, 0b01, "dcps1">;
1840 def DCPS2i : A64I_exceptImpl<0b101, 0b10, "dcps2">;
1841 def DCPS3i : A64I_exceptImpl<0b101, 0b11, "dcps3">;
1843 // The immediate is optional for the DCPS instructions, defaulting to 0.
1844 def : InstAlias<"dcps1", (DCPS1i 0)>;
1845 def : InstAlias<"dcps2", (DCPS2i 0)>;
1846 def : InstAlias<"dcps3", (DCPS3i 0)>;
1848 //===----------------------------------------------------------------------===//
1849 // Extract (immediate)
1850 //===----------------------------------------------------------------------===//
1851 // Contains: EXTR + alias ROR
1853 def EXTRwwwi : A64I_extract<0b0, 0b000, 0b0,
1855 (ins GPR32:$Rn, GPR32:$Rm, bitfield32_imm:$LSB),
1856 "extr\t$Rd, $Rn, $Rm, $LSB",
1858 (A64Extr GPR32:$Rn, GPR32:$Rm, imm:$LSB))],
1860 def EXTRxxxi : A64I_extract<0b1, 0b000, 0b1,
1862 (ins GPR64:$Rn, GPR64:$Rm, bitfield64_imm:$LSB),
1863 "extr\t$Rd, $Rn, $Rm, $LSB",
1865 (A64Extr GPR64:$Rn, GPR64:$Rm, imm:$LSB))],
1868 def : InstAlias<"ror $Rd, $Rs, $LSB",
1869 (EXTRwwwi GPR32:$Rd, GPR32:$Rs, GPR32:$Rs, bitfield32_imm:$LSB)>;
1870 def : InstAlias<"ror $Rd, $Rs, $LSB",
1871 (EXTRxxxi GPR64:$Rd, GPR64:$Rs, GPR64:$Rs, bitfield64_imm:$LSB)>;
1873 def : Pat<(rotr GPR32:$Rn, bitfield32_imm:$LSB),
1874 (EXTRwwwi GPR32:$Rn, GPR32:$Rn, bitfield32_imm:$LSB)>;
1875 def : Pat<(rotr GPR64:$Rn, bitfield64_imm:$LSB),
1876 (EXTRxxxi GPR64:$Rn, GPR64:$Rn, bitfield64_imm:$LSB)>;
1878 //===----------------------------------------------------------------------===//
1879 // Floating-point compare instructions
1880 //===----------------------------------------------------------------------===//
1881 // Contains: FCMP, FCMPE
1883 def fpzero_asmoperand : AsmOperandClass {
1884 let Name = "FPZero";
1885 let ParserMethod = "ParseFPImmOperand";
1888 def fpz32 : Operand<f32>,
1889 ComplexPattern<f32, 1, "SelectFPZeroOperand", [fpimm]> {
1890 let ParserMatchClass = fpzero_asmoperand;
1891 let PrintMethod = "printFPZeroOperand";
1894 def fpz64 : Operand<f64>,
1895 ComplexPattern<f64, 1, "SelectFPZeroOperand", [fpimm]> {
1896 let ParserMatchClass = fpzero_asmoperand;
1897 let PrintMethod = "printFPZeroOperand";
1900 multiclass A64I_fpcmpSignal<bits<2> type, bit imm, dag ins, string asmop2,
1902 def _quiet : A64I_fpcmp<0b0, 0b0, type, 0b00, {0b0, imm, 0b0, 0b0, 0b0},
1903 (outs), ins, !strconcat("fcmp\t$Rn, ", asmop2),
1904 [pattern], NoItinerary> {
1908 def _sig : A64I_fpcmp<0b0, 0b0, type, 0b00, {0b1, imm, 0b0, 0b0, 0b0},
1909 (outs), ins, !strconcat("fcmpe\t$Rn, ", asmop2),
1915 defm FCMPss : A64I_fpcmpSignal<0b00, 0b0, (ins FPR32:$Rn, FPR32:$Rm), "$Rm",
1916 (set NZCV, (A64cmp (f32 FPR32:$Rn), FPR32:$Rm))>;
1917 defm FCMPdd : A64I_fpcmpSignal<0b01, 0b0, (ins FPR64:$Rn, FPR64:$Rm), "$Rm",
1918 (set NZCV, (A64cmp (f64 FPR64:$Rn), FPR64:$Rm))>;
1920 // What would be Rm should be written as 0, but anything is valid for
1921 // disassembly so we can't set the bits
1922 let PostEncoderMethod = "fixFCMPImm" in {
1923 defm FCMPsi : A64I_fpcmpSignal<0b00, 0b1, (ins FPR32:$Rn, fpz32:$Imm), "$Imm",
1924 (set NZCV, (A64cmp (f32 FPR32:$Rn), fpz32:$Imm))>;
1926 defm FCMPdi : A64I_fpcmpSignal<0b01, 0b1, (ins FPR64:$Rn, fpz64:$Imm), "$Imm",
1927 (set NZCV, (A64cmp (f64 FPR64:$Rn), fpz64:$Imm))>;
1931 //===----------------------------------------------------------------------===//
1932 // Floating-point conditional compare instructions
1933 //===----------------------------------------------------------------------===//
1934 // Contains: FCCMP, FCCMPE
1936 class A64I_fpccmpImpl<bits<2> type, bit op, RegisterClass FPR, string asmop>
1937 : A64I_fpccmp<0b0, 0b0, type, op,
1939 (ins FPR:$Rn, FPR:$Rm, uimm4:$NZCVImm, cond_code_op:$Cond),
1940 !strconcat(asmop, "\t$Rn, $Rm, $NZCVImm, $Cond"),
1945 def FCCMPss : A64I_fpccmpImpl<0b00, 0b0, FPR32, "fccmp">;
1946 def FCCMPEss : A64I_fpccmpImpl<0b00, 0b1, FPR32, "fccmpe">;
1947 def FCCMPdd : A64I_fpccmpImpl<0b01, 0b0, FPR64, "fccmp">;
1948 def FCCMPEdd : A64I_fpccmpImpl<0b01, 0b1, FPR64, "fccmpe">;
1950 //===----------------------------------------------------------------------===//
1951 // Floating-point conditional select instructions
1952 //===----------------------------------------------------------------------===//
1955 let Uses = [NZCV] in {
1956 def FCSELsssc : A64I_fpcondsel<0b0, 0b0, 0b00, (outs FPR32:$Rd),
1957 (ins FPR32:$Rn, FPR32:$Rm, cond_code_op:$Cond),
1958 "fcsel\t$Rd, $Rn, $Rm, $Cond",
1960 (simple_select (f32 FPR32:$Rn),
1965 def FCSELdddc : A64I_fpcondsel<0b0, 0b0, 0b01, (outs FPR64:$Rd),
1966 (ins FPR64:$Rn, FPR64:$Rm, cond_code_op:$Cond),
1967 "fcsel\t$Rd, $Rn, $Rm, $Cond",
1969 (simple_select (f64 FPR64:$Rn),
1974 //===----------------------------------------------------------------------===//
1975 // Floating-point data-processing (1 source)
1976 //===----------------------------------------------------------------------===//
1977 // Contains: FMOV, FABS, FNEG, FSQRT, FCVT, FRINT[NPMZAXI].
1979 def FPNoUnop : PatFrag<(ops node:$val), (fneg node:$val),
1980 [{ (void)N; return false; }]>;
1982 // First we do the fairly trivial bunch with uniform "OP s, s" and "OP d, d"
1983 // syntax. Default to no pattern because most are odd enough not to have one.
1984 multiclass A64I_fpdp1sizes<bits<6> opcode, string asmstr,
1985 SDPatternOperator opnode = FPNoUnop> {
1986 def ss : A64I_fpdp1<0b0, 0b0, 0b00, opcode, (outs FPR32:$Rd), (ins FPR32:$Rn),
1987 !strconcat(asmstr, "\t$Rd, $Rn"),
1988 [(set (f32 FPR32:$Rd), (opnode FPR32:$Rn))],
1991 def dd : A64I_fpdp1<0b0, 0b0, 0b01, opcode, (outs FPR64:$Rd), (ins FPR64:$Rn),
1992 !strconcat(asmstr, "\t$Rd, $Rn"),
1993 [(set (f64 FPR64:$Rd), (opnode FPR64:$Rn))],
1997 defm FMOV : A64I_fpdp1sizes<0b000000, "fmov">;
1998 defm FABS : A64I_fpdp1sizes<0b000001, "fabs", fabs>;
1999 defm FNEG : A64I_fpdp1sizes<0b000010, "fneg", fneg>;
2000 defm FSQRT : A64I_fpdp1sizes<0b000011, "fsqrt", fsqrt>;
2002 defm FRINTN : A64I_fpdp1sizes<0b001000, "frintn">;
2003 defm FRINTP : A64I_fpdp1sizes<0b001001, "frintp", fceil>;
2004 defm FRINTM : A64I_fpdp1sizes<0b001010, "frintm", ffloor>;
2005 defm FRINTZ : A64I_fpdp1sizes<0b001011, "frintz", ftrunc>;
2006 defm FRINTA : A64I_fpdp1sizes<0b001100, "frinta">;
2007 defm FRINTX : A64I_fpdp1sizes<0b001110, "frintx", frint>;
2008 defm FRINTI : A64I_fpdp1sizes<0b001111, "frinti", fnearbyint>;
2010 // The FCVT instrucitons have different source and destination register-types,
2011 // but the fields are uniform everywhere a D-register (say) crops up. Package
2012 // this information in a Record.
2013 class FCVTRegType<RegisterClass rc, bits<2> fld, ValueType vt> {
2014 RegisterClass Class = rc;
2020 def FCVT16 : FCVTRegType<FPR16, 0b11, f16>;
2021 def FCVT32 : FCVTRegType<FPR32, 0b00, f32>;
2022 def FCVT64 : FCVTRegType<FPR64, 0b01, f64>;
2024 class A64I_fpdp1_fcvt<FCVTRegType DestReg, FCVTRegType SrcReg, SDNode opnode>
2025 : A64I_fpdp1<0b0, 0b0, {SrcReg.t1, SrcReg.t0},
2026 {0,0,0,1, DestReg.t1, DestReg.t0},
2027 (outs DestReg.Class:$Rd), (ins SrcReg.Class:$Rn),
2029 [(set (DestReg.VT DestReg.Class:$Rd),
2030 (opnode (SrcReg.VT SrcReg.Class:$Rn)))], NoItinerary>;
2032 def FCVTds : A64I_fpdp1_fcvt<FCVT64, FCVT32, fextend>;
2033 def FCVThs : A64I_fpdp1_fcvt<FCVT16, FCVT32, fround>;
2034 def FCVTsd : A64I_fpdp1_fcvt<FCVT32, FCVT64, fround>;
2035 def FCVThd : A64I_fpdp1_fcvt<FCVT16, FCVT64, fround>;
2036 def FCVTsh : A64I_fpdp1_fcvt<FCVT32, FCVT16, fextend>;
2037 def FCVTdh : A64I_fpdp1_fcvt<FCVT64, FCVT16, fextend>;
2040 //===----------------------------------------------------------------------===//
2041 // Floating-point data-processing (2 sources) instructions
2042 //===----------------------------------------------------------------------===//
2043 // Contains: FMUL, FDIV, FADD, FSUB, FMAX, FMIN, FMAXNM, FMINNM, FNMUL
2045 def FPNoBinop : PatFrag<(ops node:$lhs, node:$rhs), (fadd node:$lhs, node:$rhs),
2046 [{ (void)N; return false; }]>;
2048 multiclass A64I_fpdp2sizes<bits<4> opcode, string asmstr,
2049 SDPatternOperator opnode> {
2050 def sss : A64I_fpdp2<0b0, 0b0, 0b00, opcode,
2052 (ins FPR32:$Rn, FPR32:$Rm),
2053 !strconcat(asmstr, "\t$Rd, $Rn, $Rm"),
2054 [(set (f32 FPR32:$Rd), (opnode FPR32:$Rn, FPR32:$Rm))],
2057 def ddd : A64I_fpdp2<0b0, 0b0, 0b01, opcode,
2059 (ins FPR64:$Rn, FPR64:$Rm),
2060 !strconcat(asmstr, "\t$Rd, $Rn, $Rm"),
2061 [(set (f64 FPR64:$Rd), (opnode FPR64:$Rn, FPR64:$Rm))],
2065 let isCommutable = 1 in {
2066 defm FMUL : A64I_fpdp2sizes<0b0000, "fmul", fmul>;
2067 defm FADD : A64I_fpdp2sizes<0b0010, "fadd", fadd>;
2069 // No patterns for these.
2070 defm FMAX : A64I_fpdp2sizes<0b0100, "fmax", FPNoBinop>;
2071 defm FMIN : A64I_fpdp2sizes<0b0101, "fmin", FPNoBinop>;
2072 defm FMAXNM : A64I_fpdp2sizes<0b0110, "fmaxnm", FPNoBinop>;
2073 defm FMINNM : A64I_fpdp2sizes<0b0111, "fminnm", FPNoBinop>;
2075 defm FNMUL : A64I_fpdp2sizes<0b1000, "fnmul",
2076 PatFrag<(ops node:$lhs, node:$rhs),
2077 (fneg (fmul node:$lhs, node:$rhs))> >;
2080 defm FDIV : A64I_fpdp2sizes<0b0001, "fdiv", fdiv>;
2081 defm FSUB : A64I_fpdp2sizes<0b0011, "fsub", fsub>;
2083 //===----------------------------------------------------------------------===//
2084 // Floating-point data-processing (3 sources) instructions
2085 //===----------------------------------------------------------------------===//
2086 // Contains: FMADD, FMSUB, FNMADD, FNMSUB
2088 def fmsub : PatFrag<(ops node:$Rn, node:$Rm, node:$Ra),
2089 (fma (fneg node:$Rn), node:$Rm, node:$Ra)>;
2090 def fnmadd : PatFrag<(ops node:$Rn, node:$Rm, node:$Ra),
2091 (fma node:$Rn, node:$Rm, (fneg node:$Ra))>;
2092 def fnmsub : PatFrag<(ops node:$Rn, node:$Rm, node:$Ra),
2093 (fma (fneg node:$Rn), node:$Rm, (fneg node:$Ra))>;
2095 class A64I_fpdp3Impl<string asmop, RegisterClass FPR, ValueType VT,
2096 bits<2> type, bit o1, bit o0, SDPatternOperator fmakind>
2097 : A64I_fpdp3<0b0, 0b0, type, o1, o0, (outs FPR:$Rd),
2098 (ins FPR:$Rn, FPR:$Rm, FPR:$Ra),
2099 !strconcat(asmop,"\t$Rd, $Rn, $Rm, $Ra"),
2100 [(set FPR:$Rd, (fmakind (VT FPR:$Rn), FPR:$Rm, FPR:$Ra))],
2103 def FMADDssss : A64I_fpdp3Impl<"fmadd", FPR32, f32, 0b00, 0b0, 0b0, fma>;
2104 def FMSUBssss : A64I_fpdp3Impl<"fmsub", FPR32, f32, 0b00, 0b0, 0b1, fmsub>;
2105 def FNMADDssss : A64I_fpdp3Impl<"fnmadd", FPR32, f32, 0b00, 0b1, 0b0, fnmadd>;
2106 def FNMSUBssss : A64I_fpdp3Impl<"fnmsub", FPR32, f32, 0b00, 0b1, 0b1, fnmsub>;
2108 def FMADDdddd : A64I_fpdp3Impl<"fmadd", FPR64, f64, 0b01, 0b0, 0b0, fma>;
2109 def FMSUBdddd : A64I_fpdp3Impl<"fmsub", FPR64, f64, 0b01, 0b0, 0b1, fmsub>;
2110 def FNMADDdddd : A64I_fpdp3Impl<"fnmadd", FPR64, f64, 0b01, 0b1, 0b0, fnmadd>;
2111 def FNMSUBdddd : A64I_fpdp3Impl<"fnmsub", FPR64, f64, 0b01, 0b1, 0b1, fnmsub>;
2113 //===----------------------------------------------------------------------===//
2114 // Floating-point <-> fixed-point conversion instructions
2115 //===----------------------------------------------------------------------===//
2116 // Contains: FCVTZS, FCVTZU, SCVTF, UCVTF
2118 // #1-#32 allowed, encoded as "64 - <specified imm>
2119 def fixedpos_asmoperand_i32 : AsmOperandClass {
2120 let Name = "CVTFixedPos32";
2121 let RenderMethod = "addCVTFixedPosOperands";
2122 let PredicateMethod = "isCVTFixedPos<32>";
2125 // Also encoded as "64 - <specified imm>" but #1-#64 allowed.
2126 def fixedpos_asmoperand_i64 : AsmOperandClass {
2127 let Name = "CVTFixedPos64";
2128 let RenderMethod = "addCVTFixedPosOperands";
2129 let PredicateMethod = "isCVTFixedPos<64>";
2132 // We need the cartesian product of f32/f64 i32/i64 operands for
2134 // + Selection needs to use operands of correct floating type
2135 // + Assembly parsing and decoding depend on integer width
2136 class cvtfix_i32_op<ValueType FloatVT>
2138 ComplexPattern<FloatVT, 1, "SelectCVTFixedPosOperand<32>", [fpimm]> {
2139 let ParserMatchClass = fixedpos_asmoperand_i32;
2140 let DecoderMethod = "DecodeCVT32FixedPosOperand";
2141 let PrintMethod = "printCVTFixedPosOperand";
2144 class cvtfix_i64_op<ValueType FloatVT>
2146 ComplexPattern<FloatVT, 1, "SelectCVTFixedPosOperand<64>", [fpimm]> {
2147 let ParserMatchClass = fixedpos_asmoperand_i64;
2148 let PrintMethod = "printCVTFixedPosOperand";
2151 // Because of the proliferation of weird operands, it's not really
2152 // worth going for a multiclass here. Oh well.
2154 class A64I_fptofix<bit sf, bits<2> type, bits<3> opcode,
2155 RegisterClass GPR, RegisterClass FPR, Operand scale_op,
2156 string asmop, SDNode cvtop>
2157 : A64I_fpfixed<sf, 0b0, type, 0b11, opcode,
2158 (outs GPR:$Rd), (ins FPR:$Rn, scale_op:$Scale),
2159 !strconcat(asmop, "\t$Rd, $Rn, $Scale"),
2160 [(set GPR:$Rd, (cvtop (fmul FPR:$Rn, scale_op:$Scale)))],
2163 def FCVTZSwsi : A64I_fptofix<0b0, 0b00, 0b000, GPR32, FPR32,
2164 cvtfix_i32_op<f32>, "fcvtzs", fp_to_sint>;
2165 def FCVTZSxsi : A64I_fptofix<0b1, 0b00, 0b000, GPR64, FPR32,
2166 cvtfix_i64_op<f32>, "fcvtzs", fp_to_sint>;
2167 def FCVTZUwsi : A64I_fptofix<0b0, 0b00, 0b001, GPR32, FPR32,
2168 cvtfix_i32_op<f32>, "fcvtzu", fp_to_uint>;
2169 def FCVTZUxsi : A64I_fptofix<0b1, 0b00, 0b001, GPR64, FPR32,
2170 cvtfix_i64_op<f32>, "fcvtzu", fp_to_uint>;
2172 def FCVTZSwdi : A64I_fptofix<0b0, 0b01, 0b000, GPR32, FPR64,
2173 cvtfix_i32_op<f64>, "fcvtzs", fp_to_sint>;
2174 def FCVTZSxdi : A64I_fptofix<0b1, 0b01, 0b000, GPR64, FPR64,
2175 cvtfix_i64_op<f64>, "fcvtzs", fp_to_sint>;
2176 def FCVTZUwdi : A64I_fptofix<0b0, 0b01, 0b001, GPR32, FPR64,
2177 cvtfix_i32_op<f64>, "fcvtzu", fp_to_uint>;
2178 def FCVTZUxdi : A64I_fptofix<0b1, 0b01, 0b001, GPR64, FPR64,
2179 cvtfix_i64_op<f64>, "fcvtzu", fp_to_uint>;
2182 class A64I_fixtofp<bit sf, bits<2> type, bits<3> opcode,
2183 RegisterClass FPR, RegisterClass GPR, Operand scale_op,
2184 string asmop, SDNode cvtop>
2185 : A64I_fpfixed<sf, 0b0, type, 0b00, opcode,
2186 (outs FPR:$Rd), (ins GPR:$Rn, scale_op:$Scale),
2187 !strconcat(asmop, "\t$Rd, $Rn, $Scale"),
2188 [(set FPR:$Rd, (fdiv (cvtop GPR:$Rn), scale_op:$Scale))],
2191 def SCVTFswi : A64I_fixtofp<0b0, 0b00, 0b010, FPR32, GPR32,
2192 cvtfix_i32_op<f32>, "scvtf", sint_to_fp>;
2193 def SCVTFsxi : A64I_fixtofp<0b1, 0b00, 0b010, FPR32, GPR64,
2194 cvtfix_i64_op<f32>, "scvtf", sint_to_fp>;
2195 def UCVTFswi : A64I_fixtofp<0b0, 0b00, 0b011, FPR32, GPR32,
2196 cvtfix_i32_op<f32>, "ucvtf", uint_to_fp>;
2197 def UCVTFsxi : A64I_fixtofp<0b1, 0b00, 0b011, FPR32, GPR64,
2198 cvtfix_i64_op<f32>, "ucvtf", uint_to_fp>;
2199 def SCVTFdwi : A64I_fixtofp<0b0, 0b01, 0b010, FPR64, GPR32,
2200 cvtfix_i32_op<f64>, "scvtf", sint_to_fp>;
2201 def SCVTFdxi : A64I_fixtofp<0b1, 0b01, 0b010, FPR64, GPR64,
2202 cvtfix_i64_op<f64>, "scvtf", sint_to_fp>;
2203 def UCVTFdwi : A64I_fixtofp<0b0, 0b01, 0b011, FPR64, GPR32,
2204 cvtfix_i32_op<f64>, "ucvtf", uint_to_fp>;
2205 def UCVTFdxi : A64I_fixtofp<0b1, 0b01, 0b011, FPR64, GPR64,
2206 cvtfix_i64_op<f64>, "ucvtf", uint_to_fp>;
2208 //===----------------------------------------------------------------------===//
2209 // Floating-point <-> integer conversion instructions
2210 //===----------------------------------------------------------------------===//
2211 // Contains: FCVTZS, FCVTZU, SCVTF, UCVTF
2213 class A64I_fpintI<bit sf, bits<2> type, bits<2> rmode, bits<3> opcode,
2214 RegisterClass DestPR, RegisterClass SrcPR, string asmop>
2215 : A64I_fpint<sf, 0b0, type, rmode, opcode, (outs DestPR:$Rd), (ins SrcPR:$Rn),
2216 !strconcat(asmop, "\t$Rd, $Rn"), [], NoItinerary>;
2218 multiclass A64I_fptointRM<bits<2> rmode, bit o2, string asmop> {
2219 def Sws : A64I_fpintI<0b0, 0b00, rmode, {o2, 0, 0},
2220 GPR32, FPR32, asmop # "s">;
2221 def Sxs : A64I_fpintI<0b1, 0b00, rmode, {o2, 0, 0},
2222 GPR64, FPR32, asmop # "s">;
2223 def Uws : A64I_fpintI<0b0, 0b00, rmode, {o2, 0, 1},
2224 GPR32, FPR32, asmop # "u">;
2225 def Uxs : A64I_fpintI<0b1, 0b00, rmode, {o2, 0, 1},
2226 GPR64, FPR32, asmop # "u">;
2228 def Swd : A64I_fpintI<0b0, 0b01, rmode, {o2, 0, 0},
2229 GPR32, FPR64, asmop # "s">;
2230 def Sxd : A64I_fpintI<0b1, 0b01, rmode, {o2, 0, 0},
2231 GPR64, FPR64, asmop # "s">;
2232 def Uwd : A64I_fpintI<0b0, 0b01, rmode, {o2, 0, 1},
2233 GPR32, FPR64, asmop # "u">;
2234 def Uxd : A64I_fpintI<0b1, 0b01, rmode, {o2, 0, 1},
2235 GPR64, FPR64, asmop # "u">;
2238 defm FCVTN : A64I_fptointRM<0b00, 0b0, "fcvtn">;
2239 defm FCVTP : A64I_fptointRM<0b01, 0b0, "fcvtp">;
2240 defm FCVTM : A64I_fptointRM<0b10, 0b0, "fcvtm">;
2241 defm FCVTZ : A64I_fptointRM<0b11, 0b0, "fcvtz">;
2242 defm FCVTA : A64I_fptointRM<0b00, 0b1, "fcvta">;
2244 def : Pat<(i32 (fp_to_sint FPR32:$Rn)), (FCVTZSws FPR32:$Rn)>;
2245 def : Pat<(i64 (fp_to_sint FPR32:$Rn)), (FCVTZSxs FPR32:$Rn)>;
2246 def : Pat<(i32 (fp_to_uint FPR32:$Rn)), (FCVTZUws FPR32:$Rn)>;
2247 def : Pat<(i64 (fp_to_uint FPR32:$Rn)), (FCVTZUxs FPR32:$Rn)>;
2248 def : Pat<(i32 (fp_to_sint (f64 FPR64:$Rn))), (FCVTZSwd FPR64:$Rn)>;
2249 def : Pat<(i64 (fp_to_sint (f64 FPR64:$Rn))), (FCVTZSxd FPR64:$Rn)>;
2250 def : Pat<(i32 (fp_to_uint (f64 FPR64:$Rn))), (FCVTZUwd FPR64:$Rn)>;
2251 def : Pat<(i64 (fp_to_uint (f64 FPR64:$Rn))), (FCVTZUxd FPR64:$Rn)>;
2253 multiclass A64I_inttofp<bit o0, string asmop> {
2254 def CVTFsw : A64I_fpintI<0b0, 0b00, 0b00, {0, 1, o0}, FPR32, GPR32, asmop>;
2255 def CVTFsx : A64I_fpintI<0b1, 0b00, 0b00, {0, 1, o0}, FPR32, GPR64, asmop>;
2256 def CVTFdw : A64I_fpintI<0b0, 0b01, 0b00, {0, 1, o0}, FPR64, GPR32, asmop>;
2257 def CVTFdx : A64I_fpintI<0b1, 0b01, 0b00, {0, 1, o0}, FPR64, GPR64, asmop>;
2260 defm S : A64I_inttofp<0b0, "scvtf">;
2261 defm U : A64I_inttofp<0b1, "ucvtf">;
2263 def : Pat<(f32 (sint_to_fp GPR32:$Rn)), (SCVTFsw GPR32:$Rn)>;
2264 def : Pat<(f32 (sint_to_fp GPR64:$Rn)), (SCVTFsx GPR64:$Rn)>;
2265 def : Pat<(f64 (sint_to_fp GPR32:$Rn)), (SCVTFdw GPR32:$Rn)>;
2266 def : Pat<(f64 (sint_to_fp GPR64:$Rn)), (SCVTFdx GPR64:$Rn)>;
2267 def : Pat<(f32 (uint_to_fp GPR32:$Rn)), (UCVTFsw GPR32:$Rn)>;
2268 def : Pat<(f32 (uint_to_fp GPR64:$Rn)), (UCVTFsx GPR64:$Rn)>;
2269 def : Pat<(f64 (uint_to_fp GPR32:$Rn)), (UCVTFdw GPR32:$Rn)>;
2270 def : Pat<(f64 (uint_to_fp GPR64:$Rn)), (UCVTFdx GPR64:$Rn)>;
2272 def FMOVws : A64I_fpintI<0b0, 0b00, 0b00, 0b110, GPR32, FPR32, "fmov">;
2273 def FMOVsw : A64I_fpintI<0b0, 0b00, 0b00, 0b111, FPR32, GPR32, "fmov">;
2274 def FMOVxd : A64I_fpintI<0b1, 0b01, 0b00, 0b110, GPR64, FPR64, "fmov">;
2275 def FMOVdx : A64I_fpintI<0b1, 0b01, 0b00, 0b111, FPR64, GPR64, "fmov">;
2277 def : Pat<(i32 (bitconvert (f32 FPR32:$Rn))), (FMOVws FPR32:$Rn)>;
2278 def : Pat<(f32 (bitconvert (i32 GPR32:$Rn))), (FMOVsw GPR32:$Rn)>;
2279 def : Pat<(i64 (bitconvert (f64 FPR64:$Rn))), (FMOVxd FPR64:$Rn)>;
2280 def : Pat<(f64 (bitconvert (i64 GPR64:$Rn))), (FMOVdx GPR64:$Rn)>;
2282 def lane1_asmoperand : AsmOperandClass {
2284 let RenderMethod = "addImmOperands";
2287 def lane1 : Operand<i32> {
2288 let ParserMatchClass = lane1_asmoperand;
2289 let PrintMethod = "printBareImmOperand";
2292 let DecoderMethod = "DecodeFMOVLaneInstruction" in {
2293 def FMOVxv : A64I_fpint<0b1, 0b0, 0b10, 0b01, 0b110,
2294 (outs GPR64:$Rd), (ins VPR128:$Rn, lane1:$Lane),
2295 "fmov\t$Rd, $Rn.d[$Lane]", [], NoItinerary>;
2297 def FMOVvx : A64I_fpint<0b1, 0b0, 0b10, 0b01, 0b111,
2298 (outs VPR128:$Rd), (ins GPR64:$Rn, lane1:$Lane),
2299 "fmov\t$Rd.d[$Lane], $Rn", [], NoItinerary>;
2302 def : InstAlias<"fmov $Rd, $Rn.2d[$Lane]",
2303 (FMOVxv GPR64:$Rd, VPR128:$Rn, lane1:$Lane), 0b0>;
2305 def : InstAlias<"fmov $Rd.2d[$Lane], $Rn",
2306 (FMOVvx VPR128:$Rd, GPR64:$Rn, lane1:$Lane), 0b0>;
2308 //===----------------------------------------------------------------------===//
2309 // Floating-point immediate instructions
2310 //===----------------------------------------------------------------------===//
2313 def fpimm_asmoperand : AsmOperandClass {
2314 let Name = "FMOVImm";
2315 let ParserMethod = "ParseFPImmOperand";
2318 // The MCOperand for these instructions are the encoded 8-bit values.
2319 def SDXF_fpimm : SDNodeXForm<fpimm, [{
2321 A64Imms::isFPImm(N->getValueAPF(), Imm8);
2322 return CurDAG->getTargetConstant(Imm8, MVT::i32);
2325 class fmov_operand<ValueType FT>
2327 PatLeaf<(FT fpimm), [{ return A64Imms::isFPImm(N->getValueAPF()); }],
2329 let PrintMethod = "printFPImmOperand";
2330 let ParserMatchClass = fpimm_asmoperand;
2333 def fmov32_operand : fmov_operand<f32>;
2334 def fmov64_operand : fmov_operand<f64>;
2336 class A64I_fpimm_impl<bits<2> type, RegisterClass Reg, ValueType VT,
2337 Operand fmov_operand>
2338 : A64I_fpimm<0b0, 0b0, type, 0b00000,
2340 (ins fmov_operand:$Imm8),
2342 [(set (VT Reg:$Rd), fmov_operand:$Imm8)],
2345 def FMOVsi : A64I_fpimm_impl<0b00, FPR32, f32, fmov32_operand>;
2346 def FMOVdi : A64I_fpimm_impl<0b01, FPR64, f64, fmov64_operand>;
2348 //===----------------------------------------------------------------------===//
2349 // Load-register (literal) instructions
2350 //===----------------------------------------------------------------------===//
2351 // Contains: LDR, LDRSW, PRFM
2353 def ldrlit_label_asmoperand : AsmOperandClass {
2354 let Name = "LoadLitLabel";
2355 let RenderMethod = "addLabelOperands<19, 4>";
2358 def ldrlit_label : Operand<i64> {
2359 let EncoderMethod = "getLoadLitLabelOpValue";
2361 // This label is a 19-bit offset from PC, scaled by the instruction-width: 4.
2362 let PrintMethod = "printLabelOperand<19, 4>";
2363 let ParserMatchClass = ldrlit_label_asmoperand;
2364 let OperandType = "OPERAND_PCREL";
2367 // Various instructions take an immediate value (which can always be used),
2368 // where some numbers have a symbolic name to make things easier. These operands
2369 // and the associated functions abstract away the differences.
2370 multiclass namedimm<string prefix, string mapper> {
2371 def _asmoperand : AsmOperandClass {
2372 let Name = "NamedImm" # prefix;
2373 let PredicateMethod = "isUImm";
2374 let RenderMethod = "addImmOperands";
2375 let ParserMethod = "ParseNamedImmOperand<" # mapper # ">";
2378 def _op : Operand<i32> {
2379 let ParserMatchClass = !cast<AsmOperandClass>(prefix # "_asmoperand");
2380 let PrintMethod = "printNamedImmOperand<" # mapper # ">";
2381 let DecoderMethod = "DecodeNamedImmOperand<" # mapper # ">";
2385 defm prefetch : namedimm<"prefetch", "A64PRFM::PRFMMapper">;
2387 class A64I_LDRlitSimple<bits<2> opc, bit v, RegisterClass OutReg,
2388 list<dag> patterns = []>
2389 : A64I_LDRlit<opc, v, (outs OutReg:$Rt), (ins ldrlit_label:$Imm19),
2390 "ldr\t$Rt, $Imm19", patterns, NoItinerary>;
2392 let mayLoad = 1 in {
2393 def LDRw_lit : A64I_LDRlitSimple<0b00, 0b0, GPR32>;
2394 def LDRx_lit : A64I_LDRlitSimple<0b01, 0b0, GPR64>;
2397 def LDRs_lit : A64I_LDRlitSimple<0b00, 0b1, FPR32,
2398 [(set (f32 FPR32:$Rt), (load constpool:$Imm19))]>;
2399 def LDRd_lit : A64I_LDRlitSimple<0b01, 0b1, FPR64,
2400 [(set (f64 FPR64:$Rt), (load constpool:$Imm19))]>;
2402 let mayLoad = 1 in {
2403 def LDRq_lit : A64I_LDRlitSimple<0b10, 0b1, FPR128>;
2406 def LDRSWx_lit : A64I_LDRlit<0b10, 0b0,
2408 (ins ldrlit_label:$Imm19),
2409 "ldrsw\t$Rt, $Imm19",
2412 def PRFM_lit : A64I_LDRlit<0b11, 0b0,
2413 (outs), (ins prefetch_op:$Rt, ldrlit_label:$Imm19),
2414 "prfm\t$Rt, $Imm19",
2418 //===----------------------------------------------------------------------===//
2419 // Load-store exclusive instructions
2420 //===----------------------------------------------------------------------===//
2421 // Contains: STXRB, STXRH, STXR, LDXRB, LDXRH, LDXR. STXP, LDXP, STLXRB,
2422 // STLXRH, STLXR, LDAXRB, LDAXRH, LDAXR, STLXP, LDAXP, STLRB,
2423 // STLRH, STLR, LDARB, LDARH, LDAR
2425 // Since these instructions have the undefined register bits set to 1 in
2426 // their canonical form, we need a post encoder method to set those bits
2427 // to 1 when encoding these instructions. We do this using the
2428 // fixLoadStoreExclusive function. This function has template parameters:
2430 // fixLoadStoreExclusive<int hasRs, int hasRt2>
2432 // hasRs indicates that the instruction uses the Rs field, so we won't set
2433 // it to 1 (and the same for Rt2). We don't need template parameters for
2434 // the other register fiels since Rt and Rn are always used.
2436 // This operand parses a GPR64xsp register, followed by an optional immediate
2438 def GPR64xsp0_asmoperand : AsmOperandClass {
2439 let Name = "GPR64xsp0";
2440 let PredicateMethod = "isWrappedReg";
2441 let RenderMethod = "addRegOperands";
2442 let ParserMethod = "ParseLSXAddressOperand";
2445 def GPR64xsp0 : RegisterOperand<GPR64xsp> {
2446 let ParserMatchClass = GPR64xsp0_asmoperand;
2449 //===----------------------------------
2450 // Store-exclusive (releasing & normal)
2451 //===----------------------------------
2453 class A64I_SRexs_impl<bits<2> size, bits<3> opcode, string asm, dag outs,
2454 dag ins, list<dag> pat,
2455 InstrItinClass itin> :
2456 A64I_LDSTex_stn <size,
2457 opcode{2}, 0, opcode{1}, opcode{0},
2459 !strconcat(asm, "\t$Rs, $Rt, [$Rn]"),
2462 let PostEncoderMethod = "fixLoadStoreExclusive<1,0>";
2465 multiclass A64I_SRex<string asmstr, bits<3> opcode, string prefix> {
2466 def _byte: A64I_SRexs_impl<0b00, opcode, !strconcat(asmstr, "b"),
2467 (outs GPR32:$Rs), (ins GPR32:$Rt, GPR64xsp0:$Rn),
2470 def _hword: A64I_SRexs_impl<0b01, opcode, !strconcat(asmstr, "h"),
2471 (outs GPR32:$Rs), (ins GPR32:$Rt, GPR64xsp0:$Rn),
2474 def _word: A64I_SRexs_impl<0b10, opcode, asmstr,
2475 (outs GPR32:$Rs), (ins GPR32:$Rt, GPR64xsp0:$Rn),
2478 def _dword: A64I_SRexs_impl<0b11, opcode, asmstr,
2479 (outs GPR32:$Rs), (ins GPR64:$Rt, GPR64xsp0:$Rn),
2483 defm STXR : A64I_SRex<"stxr", 0b000, "STXR">;
2484 defm STLXR : A64I_SRex<"stlxr", 0b001, "STLXR">;
2486 //===----------------------------------
2488 //===----------------------------------
2490 class A64I_LRexs_impl<bits<2> size, bits<3> opcode, string asm, dag outs,
2491 dag ins, list<dag> pat,
2492 InstrItinClass itin> :
2493 A64I_LDSTex_tn <size,
2494 opcode{2}, 1, opcode{1}, opcode{0},
2496 !strconcat(asm, "\t$Rt, [$Rn]"),
2499 let PostEncoderMethod = "fixLoadStoreExclusive<0,0>";
2502 multiclass A64I_LRex<string asmstr, bits<3> opcode> {
2503 def _byte: A64I_LRexs_impl<0b00, opcode, !strconcat(asmstr, "b"),
2504 (outs GPR32:$Rt), (ins GPR64xsp0:$Rn),
2507 def _hword: A64I_LRexs_impl<0b01, opcode, !strconcat(asmstr, "h"),
2508 (outs GPR32:$Rt), (ins GPR64xsp0:$Rn),
2511 def _word: A64I_LRexs_impl<0b10, opcode, asmstr,
2512 (outs GPR32:$Rt), (ins GPR64xsp0:$Rn),
2515 def _dword: A64I_LRexs_impl<0b11, opcode, asmstr,
2516 (outs GPR64:$Rt), (ins GPR64xsp0:$Rn),
2520 defm LDXR : A64I_LRex<"ldxr", 0b000>;
2521 defm LDAXR : A64I_LRex<"ldaxr", 0b001>;
2522 defm LDAR : A64I_LRex<"ldar", 0b101>;
2524 class acquiring_load<PatFrag base>
2525 : PatFrag<(ops node:$ptr), (base node:$ptr), [{
2526 return cast<AtomicSDNode>(N)->getOrdering() == Acquire;
2529 def atomic_load_acquire_8 : acquiring_load<atomic_load_8>;
2530 def atomic_load_acquire_16 : acquiring_load<atomic_load_16>;
2531 def atomic_load_acquire_32 : acquiring_load<atomic_load_32>;
2532 def atomic_load_acquire_64 : acquiring_load<atomic_load_64>;
2534 def : Pat<(atomic_load_acquire_8 GPR64xsp:$Rn), (LDAR_byte GPR64xsp0:$Rn)>;
2535 def : Pat<(atomic_load_acquire_16 GPR64xsp:$Rn), (LDAR_hword GPR64xsp0:$Rn)>;
2536 def : Pat<(atomic_load_acquire_32 GPR64xsp:$Rn), (LDAR_word GPR64xsp0:$Rn)>;
2537 def : Pat<(atomic_load_acquire_64 GPR64xsp:$Rn), (LDAR_dword GPR64xsp0:$Rn)>;
2539 //===----------------------------------
2540 // Store-release (no exclusivity)
2541 //===----------------------------------
2543 class A64I_SLexs_impl<bits<2> size, bits<3> opcode, string asm, dag outs,
2544 dag ins, list<dag> pat,
2545 InstrItinClass itin> :
2546 A64I_LDSTex_tn <size,
2547 opcode{2}, 0, opcode{1}, opcode{0},
2549 !strconcat(asm, "\t$Rt, [$Rn]"),
2552 let PostEncoderMethod = "fixLoadStoreExclusive<0,0>";
2555 class releasing_store<PatFrag base>
2556 : PatFrag<(ops node:$ptr, node:$val), (base node:$ptr, node:$val), [{
2557 return cast<AtomicSDNode>(N)->getOrdering() == Release;
2560 def atomic_store_release_8 : releasing_store<atomic_store_8>;
2561 def atomic_store_release_16 : releasing_store<atomic_store_16>;
2562 def atomic_store_release_32 : releasing_store<atomic_store_32>;
2563 def atomic_store_release_64 : releasing_store<atomic_store_64>;
2565 multiclass A64I_SLex<string asmstr, bits<3> opcode, string prefix> {
2566 def _byte: A64I_SLexs_impl<0b00, opcode, !strconcat(asmstr, "b"),
2567 (outs), (ins GPR32:$Rt, GPR64xsp0:$Rn),
2568 [(atomic_store_release_8 GPR64xsp0:$Rn, GPR32:$Rt)],
2571 def _hword: A64I_SLexs_impl<0b01, opcode, !strconcat(asmstr, "h"),
2572 (outs), (ins GPR32:$Rt, GPR64xsp0:$Rn),
2573 [(atomic_store_release_16 GPR64xsp0:$Rn, GPR32:$Rt)],
2576 def _word: A64I_SLexs_impl<0b10, opcode, asmstr,
2577 (outs), (ins GPR32:$Rt, GPR64xsp0:$Rn),
2578 [(atomic_store_release_32 GPR64xsp0:$Rn, GPR32:$Rt)],
2581 def _dword: A64I_SLexs_impl<0b11, opcode, asmstr,
2582 (outs), (ins GPR64:$Rt, GPR64xsp0:$Rn),
2583 [(atomic_store_release_64 GPR64xsp0:$Rn, GPR64:$Rt)],
2587 defm STLR : A64I_SLex<"stlr", 0b101, "STLR">;
2589 //===----------------------------------
2590 // Store-exclusive pair (releasing & normal)
2591 //===----------------------------------
2593 class A64I_SPexs_impl<bits<2> size, bits<3> opcode, string asm, dag outs,
2594 dag ins, list<dag> pat,
2595 InstrItinClass itin> :
2596 A64I_LDSTex_stt2n <size,
2597 opcode{2}, 0, opcode{1}, opcode{0},
2599 !strconcat(asm, "\t$Rs, $Rt, $Rt2, [$Rn]"),
2605 multiclass A64I_SPex<string asmstr, bits<3> opcode> {
2606 def _word: A64I_SPexs_impl<0b10, opcode, asmstr, (outs),
2607 (ins GPR32:$Rs, GPR32:$Rt, GPR32:$Rt2,
2611 def _dword: A64I_SPexs_impl<0b11, opcode, asmstr, (outs),
2612 (ins GPR32:$Rs, GPR64:$Rt, GPR64:$Rt2,
2617 defm STXP : A64I_SPex<"stxp", 0b010>;
2618 defm STLXP : A64I_SPex<"stlxp", 0b011>;
2620 //===----------------------------------
2621 // Load-exclusive pair (acquiring & normal)
2622 //===----------------------------------
2624 class A64I_LPexs_impl<bits<2> size, bits<3> opcode, string asm, dag outs,
2625 dag ins, list<dag> pat,
2626 InstrItinClass itin> :
2627 A64I_LDSTex_tt2n <size,
2628 opcode{2}, 1, opcode{1}, opcode{0},
2630 !strconcat(asm, "\t$Rt, $Rt2, [$Rn]"),
2633 let DecoderMethod = "DecodeLoadPairExclusiveInstruction";
2634 let PostEncoderMethod = "fixLoadStoreExclusive<0,1>";
2637 multiclass A64I_LPex<string asmstr, bits<3> opcode> {
2638 def _word: A64I_LPexs_impl<0b10, opcode, asmstr,
2639 (outs GPR32:$Rt, GPR32:$Rt2),
2640 (ins GPR64xsp0:$Rn),
2643 def _dword: A64I_LPexs_impl<0b11, opcode, asmstr,
2644 (outs GPR64:$Rt, GPR64:$Rt2),
2645 (ins GPR64xsp0:$Rn),
2649 defm LDXP : A64I_LPex<"ldxp", 0b010>;
2650 defm LDAXP : A64I_LPex<"ldaxp", 0b011>;
2652 //===----------------------------------------------------------------------===//
2653 // Load-store register (unscaled immediate) instructions
2654 //===----------------------------------------------------------------------===//
2655 // Contains: LDURB, LDURH, LDRUSB, LDRUSH, LDRUSW, STUR, STURB, STURH and PRFUM
2659 //===----------------------------------------------------------------------===//
2660 // Load-store register (register offset) instructions
2661 //===----------------------------------------------------------------------===//
2662 // Contains: LDRB, LDRH, LDRSB, LDRSH, LDRSW, STR, STRB, STRH and PRFM
2666 //===----------------------------------------------------------------------===//
2667 // Load-store register (unsigned immediate) instructions
2668 //===----------------------------------------------------------------------===//
2669 // Contains: LDRB, LDRH, LDRSB, LDRSH, LDRSW, STR, STRB, STRH and PRFM
2673 //===----------------------------------------------------------------------===//
2674 // Load-store register (immediate post-indexed) instructions
2675 //===----------------------------------------------------------------------===//
2676 // Contains: STRB, STRH, STR, LDRB, LDRH, LDR, LDRSB, LDRSH, LDRSW
2680 //===----------------------------------------------------------------------===//
2681 // Load-store register (immediate pre-indexed) instructions
2682 //===----------------------------------------------------------------------===//
2683 // Contains: STRB, STRH, STR, LDRB, LDRH, LDR, LDRSB, LDRSH, LDRSW
2685 // Note that patterns are much later on in a completely separate section (they
2686 // need ADRPxi to be defined).
2688 //===-------------------------------
2689 // 1. Various operands needed
2690 //===-------------------------------
2692 //===-------------------------------
2693 // 1.1 Unsigned 12-bit immediate operands
2694 //===-------------------------------
2695 // The addressing mode for these instructions consists of an unsigned 12-bit
2696 // immediate which is scaled by the size of the memory access.
2698 // We represent this in the MC layer by two operands:
2699 // 1. A base register.
2700 // 2. A 12-bit immediate: not multiplied by access size, so "LDR x0,[x0,#8]"
2701 // would have '1' in this field.
2702 // This means that separate functions are needed for converting representations
2703 // which *are* aware of the intended access size.
2705 // Anything that creates an MCInst (Decoding, selection and AsmParsing) has to
2706 // know the access size via some means. An isolated operand does not have this
2707 // information unless told from here, which means we need separate tablegen
2708 // Operands for each access size. This multiclass takes care of instantiating
2709 // the correct template functions in the rest of the backend.
2711 //===-------------------------------
2712 // 1.1 Unsigned 12-bit immediate operands
2713 //===-------------------------------
2715 multiclass offsets_uimm12<int MemSize, string prefix> {
2716 def uimm12_asmoperand : AsmOperandClass {
2717 let Name = "OffsetUImm12_" # MemSize;
2718 let PredicateMethod = "isOffsetUImm12<" # MemSize # ">";
2719 let RenderMethod = "addOffsetUImm12Operands<" # MemSize # ">";
2722 // Pattern is really no more than an ImmLeaf, but predicated on MemSize which
2723 // complicates things beyond TableGen's ken.
2724 def uimm12 : Operand<i64>,
2725 ComplexPattern<i64, 1, "SelectOffsetUImm12<" # MemSize # ">"> {
2726 let ParserMatchClass
2727 = !cast<AsmOperandClass>(prefix # uimm12_asmoperand);
2729 let PrintMethod = "printOffsetUImm12Operand<" # MemSize # ">";
2730 let EncoderMethod = "getOffsetUImm12OpValue<" # MemSize # ">";
2734 defm byte_ : offsets_uimm12<1, "byte_">;
2735 defm hword_ : offsets_uimm12<2, "hword_">;
2736 defm word_ : offsets_uimm12<4, "word_">;
2737 defm dword_ : offsets_uimm12<8, "dword_">;
2738 defm qword_ : offsets_uimm12<16, "qword_">;
2740 //===-------------------------------
2741 // 1.1 Signed 9-bit immediate operands
2742 //===-------------------------------
2744 // The MCInst is expected to store the bit-wise encoding of the value,
2745 // which amounts to lopping off the extended sign bits.
2746 def SDXF_simm9 : SDNodeXForm<imm, [{
2747 return CurDAG->getTargetConstant(N->getZExtValue() & 0x1ff, MVT::i32);
2750 def simm9_asmoperand : AsmOperandClass {
2752 let PredicateMethod = "isSImm<9>";
2753 let RenderMethod = "addSImmOperands<9>";
2756 def simm9 : Operand<i64>,
2757 ImmLeaf<i64, [{ return Imm >= -0x100 && Imm <= 0xff; }],
2759 let PrintMethod = "printOffsetSImm9Operand";
2760 let ParserMatchClass = simm9_asmoperand;
2764 //===-------------------------------
2765 // 1.3 Register offset extensions
2766 //===-------------------------------
2768 // The assembly-syntax for these addressing-modes is:
2769 // [<Xn|SP>, <R><m> {, <extend> {<amount>}}]
2771 // The essential semantics are:
2772 // + <amount> is a shift: #<log(transfer size)> or #0
2773 // + <R> can be W or X.
2774 // + If <R> is W, <extend> can be UXTW or SXTW
2775 // + If <R> is X, <extend> can be LSL or SXTX
2777 // The trickiest of those constraints is that Rm can be either GPR32 or GPR64,
2778 // which will need separate instructions for LLVM type-consistency. We'll also
2779 // need separate operands, of course.
2780 multiclass regexts<int MemSize, int RmSize, RegisterClass GPR,
2781 string Rm, string prefix> {
2782 def regext_asmoperand : AsmOperandClass {
2783 let Name = "AddrRegExtend_" # MemSize # "_" # Rm;
2784 let PredicateMethod = "isAddrRegExtend<" # MemSize # "," # RmSize # ">";
2785 let RenderMethod = "addAddrRegExtendOperands<" # MemSize # ">";
2788 def regext : Operand<i64> {
2790 = "printAddrRegExtendOperand<" # MemSize # ", " # RmSize # ">";
2792 let DecoderMethod = "DecodeAddrRegExtendOperand";
2793 let ParserMatchClass
2794 = !cast<AsmOperandClass>(prefix # regext_asmoperand);
2798 multiclass regexts_wx<int MemSize, string prefix> {
2799 // Rm is an X-register if LSL or SXTX are specified as the shift.
2800 defm Xm_ : regexts<MemSize, 64, GPR64, "Xm", prefix # "Xm_">;
2802 // Rm is a W-register if UXTW or SXTW are specified as the shift.
2803 defm Wm_ : regexts<MemSize, 32, GPR32, "Wm", prefix # "Wm_">;
2806 defm byte_ : regexts_wx<1, "byte_">;
2807 defm hword_ : regexts_wx<2, "hword_">;
2808 defm word_ : regexts_wx<4, "word_">;
2809 defm dword_ : regexts_wx<8, "dword_">;
2810 defm qword_ : regexts_wx<16, "qword_">;
2813 //===------------------------------
2814 // 2. The instructions themselves.
2815 //===------------------------------
2817 // We have the following instructions to implement:
2818 // | | B | H | W | X |
2819 // |-----------------+-------+-------+-------+--------|
2820 // | unsigned str | STRB | STRH | STR | STR |
2821 // | unsigned ldr | LDRB | LDRH | LDR | LDR |
2822 // | signed ldr to W | LDRSB | LDRSH | - | - |
2823 // | signed ldr to X | LDRSB | LDRSH | LDRSW | (PRFM) |
2825 // This will instantiate the LDR/STR instructions you'd expect to use for an
2826 // unsigned datatype (first two rows above) or floating-point register, which is
2827 // reasonably uniform across all access sizes.
2830 //===------------------------------
2831 // 2.1 Regular instructions
2832 //===------------------------------
2834 // This class covers the basic unsigned or irrelevantly-signed loads and stores,
2835 // to general-purpose and floating-point registers.
2837 class AddrParams<string prefix> {
2838 Operand uimm12 = !cast<Operand>(prefix # "_uimm12");
2840 Operand regextWm = !cast<Operand>(prefix # "_Wm_regext");
2841 Operand regextXm = !cast<Operand>(prefix # "_Xm_regext");
2844 def byte_addrparams : AddrParams<"byte">;
2845 def hword_addrparams : AddrParams<"hword">;
2846 def word_addrparams : AddrParams<"word">;
2847 def dword_addrparams : AddrParams<"dword">;
2848 def qword_addrparams : AddrParams<"qword">;
2850 multiclass A64I_LDRSTR_unsigned<string prefix, bits<2> size, bit v,
2851 bit high_opc, string asmsuffix,
2852 RegisterClass GPR, AddrParams params> {
2853 // Unsigned immediate
2854 def _STR : A64I_LSunsigimm<size, v, {high_opc, 0b0},
2855 (outs), (ins GPR:$Rt, GPR64xsp:$Rn, params.uimm12:$UImm12),
2856 "str" # asmsuffix # "\t$Rt, [$Rn, $UImm12]",
2860 def : InstAlias<"str" # asmsuffix # " $Rt, [$Rn]",
2861 (!cast<Instruction>(prefix # "_STR") GPR:$Rt, GPR64xsp:$Rn, 0)>;
2863 def _LDR : A64I_LSunsigimm<size, v, {high_opc, 0b1},
2864 (outs GPR:$Rt), (ins GPR64xsp:$Rn, params.uimm12:$UImm12),
2865 "ldr" # asmsuffix # "\t$Rt, [$Rn, $UImm12]",
2869 def : InstAlias<"ldr" # asmsuffix # " $Rt, [$Rn]",
2870 (!cast<Instruction>(prefix # "_LDR") GPR:$Rt, GPR64xsp:$Rn, 0)>;
2872 // Register offset (four of these: load/store and Wm/Xm).
2873 let mayLoad = 1 in {
2874 def _Wm_RegOffset_LDR : A64I_LSregoff<size, v, {high_opc, 0b1}, 0b0,
2876 (ins GPR64xsp:$Rn, GPR32:$Rm, params.regextWm:$Ext),
2877 "ldr" # asmsuffix # "\t$Rt, [$Rn, $Rm, $Ext]",
2880 def _Xm_RegOffset_LDR : A64I_LSregoff<size, v, {high_opc, 0b1}, 0b1,
2882 (ins GPR64xsp:$Rn, GPR64:$Rm, params.regextXm:$Ext),
2883 "ldr" # asmsuffix # "\t$Rt, [$Rn, $Rm, $Ext]",
2886 def : InstAlias<"ldr" # asmsuffix # " $Rt, [$Rn, $Rm]",
2887 (!cast<Instruction>(prefix # "_Xm_RegOffset_LDR") GPR:$Rt, GPR64xsp:$Rn,
2890 let mayStore = 1 in {
2891 def _Wm_RegOffset_STR : A64I_LSregoff<size, v, {high_opc, 0b0}, 0b0,
2892 (outs), (ins GPR:$Rt, GPR64xsp:$Rn, GPR32:$Rm,
2893 params.regextWm:$Ext),
2894 "str" # asmsuffix # "\t$Rt, [$Rn, $Rm, $Ext]",
2897 def _Xm_RegOffset_STR : A64I_LSregoff<size, v, {high_opc, 0b0}, 0b1,
2898 (outs), (ins GPR:$Rt, GPR64xsp:$Rn, GPR64:$Rm,
2899 params.regextXm:$Ext),
2900 "str" # asmsuffix # "\t$Rt, [$Rn, $Rm, $Ext]",
2903 def : InstAlias<"str" # asmsuffix # " $Rt, [$Rn, $Rm]",
2904 (!cast<Instruction>(prefix # "_Xm_RegOffset_STR") GPR:$Rt, GPR64xsp:$Rn,
2907 // Unaligned immediate
2908 def _STUR : A64I_LSunalimm<size, v, {high_opc, 0b0},
2909 (outs), (ins GPR:$Rt, GPR64xsp:$Rn, simm9:$SImm9),
2910 "stur" # asmsuffix # "\t$Rt, [$Rn, $SImm9]",
2914 def : InstAlias<"stur" # asmsuffix # " $Rt, [$Rn]",
2915 (!cast<Instruction>(prefix # "_STUR") GPR:$Rt, GPR64xsp:$Rn, 0)>;
2917 def _LDUR : A64I_LSunalimm<size, v, {high_opc, 0b1},
2918 (outs GPR:$Rt), (ins GPR64xsp:$Rn, simm9:$SImm9),
2919 "ldur" # asmsuffix # "\t$Rt, [$Rn, $SImm9]",
2923 def : InstAlias<"ldur" # asmsuffix # " $Rt, [$Rn]",
2924 (!cast<Instruction>(prefix # "_LDUR") GPR:$Rt, GPR64xsp:$Rn, 0)>;
2927 def _PostInd_STR : A64I_LSpostind<size, v, {high_opc, 0b0},
2928 (outs GPR64xsp:$Rn_wb),
2929 (ins GPR:$Rt, GPR64xsp:$Rn, simm9:$SImm9),
2930 "str" # asmsuffix # "\t$Rt, [$Rn], $SImm9",
2932 let Constraints = "$Rn = $Rn_wb";
2935 // Decoder only needed for unpredictability checking (FIXME).
2936 let DecoderMethod = "DecodeSingleIndexedInstruction";
2939 def _PostInd_LDR : A64I_LSpostind<size, v, {high_opc, 0b1},
2940 (outs GPR:$Rt, GPR64xsp:$Rn_wb),
2941 (ins GPR64xsp:$Rn, simm9:$SImm9),
2942 "ldr" # asmsuffix # "\t$Rt, [$Rn], $SImm9",
2945 let Constraints = "$Rn = $Rn_wb";
2946 let DecoderMethod = "DecodeSingleIndexedInstruction";
2950 def _PreInd_STR : A64I_LSpreind<size, v, {high_opc, 0b0},
2951 (outs GPR64xsp:$Rn_wb),
2952 (ins GPR:$Rt, GPR64xsp:$Rn, simm9:$SImm9),
2953 "str" # asmsuffix # "\t$Rt, [$Rn, $SImm9]!",
2955 let Constraints = "$Rn = $Rn_wb";
2958 // Decoder only needed for unpredictability checking (FIXME).
2959 let DecoderMethod = "DecodeSingleIndexedInstruction";
2962 def _PreInd_LDR : A64I_LSpreind<size, v, {high_opc, 0b1},
2963 (outs GPR:$Rt, GPR64xsp:$Rn_wb),
2964 (ins GPR64xsp:$Rn, simm9:$SImm9),
2965 "ldr" # asmsuffix # "\t$Rt, [$Rn, $SImm9]!",
2968 let Constraints = "$Rn = $Rn_wb";
2969 let DecoderMethod = "DecodeSingleIndexedInstruction";
2974 // STRB/LDRB: First define the instructions
2976 : A64I_LDRSTR_unsigned<"LS8", 0b00, 0b0, 0b0, "b", GPR32, byte_addrparams>;
2980 : A64I_LDRSTR_unsigned<"LS16", 0b01, 0b0, 0b0, "h", GPR32, hword_addrparams>;
2983 // STR/LDR to/from a W register
2985 : A64I_LDRSTR_unsigned<"LS32", 0b10, 0b0, 0b0, "", GPR32, word_addrparams>;
2987 // STR/LDR to/from an X register
2989 : A64I_LDRSTR_unsigned<"LS64", 0b11, 0b0, 0b0, "", GPR64, dword_addrparams>;
2991 // STR/LDR to/from a B register
2993 : A64I_LDRSTR_unsigned<"LSFP8", 0b00, 0b1, 0b0, "", FPR8, byte_addrparams>;
2995 // STR/LDR to/from an H register
2997 : A64I_LDRSTR_unsigned<"LSFP16", 0b01, 0b1, 0b0, "", FPR16, hword_addrparams>;
2999 // STR/LDR to/from an S register
3001 : A64I_LDRSTR_unsigned<"LSFP32", 0b10, 0b1, 0b0, "", FPR32, word_addrparams>;
3002 // STR/LDR to/from a D register
3004 : A64I_LDRSTR_unsigned<"LSFP64", 0b11, 0b1, 0b0, "", FPR64, dword_addrparams>;
3005 // STR/LDR to/from a Q register
3007 : A64I_LDRSTR_unsigned<"LSFP128", 0b00, 0b1, 0b1, "", FPR128,
3010 //===------------------------------
3012 //===------------------------------
3014 // Byte and half-word signed loads can both go into either an X or a W register,
3015 // so it's worth factoring out. Signed word loads don't fit because there is no
3017 multiclass A64I_LDR_signed<bits<2> size, string asmopcode, AddrParams params,
3020 def w : A64I_LSunsigimm<size, 0b0, 0b11,
3022 (ins GPR64xsp:$Rn, params.uimm12:$UImm12),
3023 "ldrs" # asmopcode # "\t$Rt, [$Rn, $UImm12]",
3027 def : InstAlias<"ldrs" # asmopcode # " $Rt, [$Rn]",
3028 (!cast<Instruction>(prefix # w) GPR32:$Rt, GPR64xsp:$Rn, 0)>;
3030 def x : A64I_LSunsigimm<size, 0b0, 0b10,
3032 (ins GPR64xsp:$Rn, params.uimm12:$UImm12),
3033 "ldrs" # asmopcode # "\t$Rt, [$Rn, $UImm12]",
3037 def : InstAlias<"ldrs" # asmopcode # " $Rt, [$Rn]",
3038 (!cast<Instruction>(prefix # x) GPR64:$Rt, GPR64xsp:$Rn, 0)>;
3041 let mayLoad = 1 in {
3042 def w_Wm_RegOffset : A64I_LSregoff<size, 0b0, 0b11, 0b0,
3044 (ins GPR64xsp:$Rn, GPR32:$Rm, params.regextWm:$Ext),
3045 "ldrs" # asmopcode # "\t$Rt, [$Rn, $Rm, $Ext]",
3048 def w_Xm_RegOffset : A64I_LSregoff<size, 0b0, 0b11, 0b1,
3050 (ins GPR64xsp:$Rn, GPR64:$Rm, params.regextXm:$Ext),
3051 "ldrs" # asmopcode # "\t$Rt, [$Rn, $Rm, $Ext]",
3054 def x_Wm_RegOffset : A64I_LSregoff<size, 0b0, 0b10, 0b0,
3056 (ins GPR64xsp:$Rn, GPR32:$Rm, params.regextWm:$Ext),
3057 "ldrs" # asmopcode # "\t$Rt, [$Rn, $Rm, $Ext]",
3060 def x_Xm_RegOffset : A64I_LSregoff<size, 0b0, 0b10, 0b1,
3062 (ins GPR64xsp:$Rn, GPR64:$Rm, params.regextXm:$Ext),
3063 "ldrs" # asmopcode # "\t$Rt, [$Rn, $Rm, $Ext]",
3066 def : InstAlias<"ldrs" # asmopcode # " $Rt, [$Rn, $Rm]",
3067 (!cast<Instruction>(prefix # "w_Xm_RegOffset") GPR32:$Rt, GPR64xsp:$Rn,
3070 def : InstAlias<"ldrs" # asmopcode # " $Rt, [$Rn, $Rm]",
3071 (!cast<Instruction>(prefix # "x_Xm_RegOffset") GPR64:$Rt, GPR64xsp:$Rn,
3075 let mayLoad = 1 in {
3077 def w_U : A64I_LSunalimm<size, 0b0, 0b11,
3079 (ins GPR64xsp:$Rn, simm9:$SImm9),
3080 "ldurs" # asmopcode # "\t$Rt, [$Rn, $SImm9]",
3083 def x_U : A64I_LSunalimm<size, 0b0, 0b10,
3085 (ins GPR64xsp:$Rn, simm9:$SImm9),
3086 "ldurs" # asmopcode # "\t$Rt, [$Rn, $SImm9]",
3091 def w_PostInd : A64I_LSpostind<size, 0b0, 0b11,
3092 (outs GPR32:$Rt, GPR64xsp:$Rn_wb),
3093 (ins GPR64xsp:$Rn, simm9:$SImm9),
3094 "ldrs" # asmopcode # "\t$Rt, [$Rn], $SImm9",
3096 let Constraints = "$Rn = $Rn_wb";
3097 let DecoderMethod = "DecodeSingleIndexedInstruction";
3100 def x_PostInd : A64I_LSpostind<size, 0b0, 0b10,
3101 (outs GPR64:$Rt, GPR64xsp:$Rn_wb),
3102 (ins GPR64xsp:$Rn, simm9:$SImm9),
3103 "ldrs" # asmopcode # "\t$Rt, [$Rn], $SImm9",
3105 let Constraints = "$Rn = $Rn_wb";
3106 let DecoderMethod = "DecodeSingleIndexedInstruction";
3110 def w_PreInd : A64I_LSpreind<size, 0b0, 0b11,
3111 (outs GPR32:$Rt, GPR64xsp:$Rn_wb),
3112 (ins GPR64xsp:$Rn, simm9:$SImm9),
3113 "ldrs" # asmopcode # "\t$Rt, [$Rn, $SImm9]!",
3115 let Constraints = "$Rn = $Rn_wb";
3116 let DecoderMethod = "DecodeSingleIndexedInstruction";
3119 def x_PreInd : A64I_LSpreind<size, 0b0, 0b10,
3120 (outs GPR64:$Rt, GPR64xsp:$Rn_wb),
3121 (ins GPR64xsp:$Rn, simm9:$SImm9),
3122 "ldrs" # asmopcode # "\t$Rt, [$Rn, $SImm9]!",
3124 let Constraints = "$Rn = $Rn_wb";
3125 let DecoderMethod = "DecodeSingleIndexedInstruction";
3127 } // let mayLoad = 1
3131 defm LDRSB : A64I_LDR_signed<0b00, "b", byte_addrparams, "LDRSB">;
3133 defm LDRSH : A64I_LDR_signed<0b01, "h", hword_addrparams, "LDRSH">;
3135 // LDRSW: load a 32-bit register, sign-extending to 64-bits.
3137 : A64I_LSunsigimm<0b10, 0b0, 0b10,
3139 (ins GPR64xsp:$Rn, word_uimm12:$UImm12),
3140 "ldrsw\t$Rt, [$Rn, $UImm12]",
3144 def : InstAlias<"ldrsw $Rt, [$Rn]", (LDRSWx GPR64:$Rt, GPR64xsp:$Rn, 0)>;
3146 let mayLoad = 1 in {
3147 def LDRSWx_Wm_RegOffset : A64I_LSregoff<0b10, 0b0, 0b10, 0b0,
3149 (ins GPR64xsp:$Rn, GPR32:$Rm, word_Wm_regext:$Ext),
3150 "ldrsw\t$Rt, [$Rn, $Rm, $Ext]",
3153 def LDRSWx_Xm_RegOffset : A64I_LSregoff<0b10, 0b0, 0b10, 0b1,
3155 (ins GPR64xsp:$Rn, GPR64:$Rm, word_Xm_regext:$Ext),
3156 "ldrsw\t$Rt, [$Rn, $Rm, $Ext]",
3159 def : InstAlias<"ldrsw $Rt, [$Rn, $Rm]",
3160 (LDRSWx_Xm_RegOffset GPR64:$Rt, GPR64xsp:$Rn, GPR64:$Rm, 2)>;
3164 : A64I_LSunalimm<0b10, 0b0, 0b10,
3166 (ins GPR64xsp:$Rn, simm9:$SImm9),
3167 "ldursw\t$Rt, [$Rn, $SImm9]",
3171 def : InstAlias<"ldursw $Rt, [$Rn]", (LDURSWx GPR64:$Rt, GPR64xsp:$Rn, 0)>;
3174 : A64I_LSpostind<0b10, 0b0, 0b10,
3175 (outs GPR64:$Rt, GPR64xsp:$Rn_wb),
3176 (ins GPR64xsp:$Rn, simm9:$SImm9),
3177 "ldrsw\t$Rt, [$Rn], $SImm9",
3180 let Constraints = "$Rn = $Rn_wb";
3181 let DecoderMethod = "DecodeSingleIndexedInstruction";
3184 def LDRSWx_PreInd : A64I_LSpreind<0b10, 0b0, 0b10,
3185 (outs GPR64:$Rt, GPR64xsp:$Rn_wb),
3186 (ins GPR64xsp:$Rn, simm9:$SImm9),
3187 "ldrsw\t$Rt, [$Rn, $SImm9]!",
3190 let Constraints = "$Rn = $Rn_wb";
3191 let DecoderMethod = "DecodeSingleIndexedInstruction";
3194 //===------------------------------
3195 // 2.4 Prefetch operations
3196 //===------------------------------
3198 def PRFM : A64I_LSunsigimm<0b11, 0b0, 0b10, (outs),
3199 (ins prefetch_op:$Rt, GPR64xsp:$Rn, dword_uimm12:$UImm12),
3200 "prfm\t$Rt, [$Rn, $UImm12]",
3204 def : InstAlias<"prfm $Rt, [$Rn]",
3205 (PRFM prefetch_op:$Rt, GPR64xsp:$Rn, 0)>;
3207 let mayLoad = 1 in {
3208 def PRFM_Wm_RegOffset : A64I_LSregoff<0b11, 0b0, 0b10, 0b0, (outs),
3209 (ins prefetch_op:$Rt, GPR64xsp:$Rn,
3210 GPR32:$Rm, dword_Wm_regext:$Ext),
3211 "prfm\t$Rt, [$Rn, $Rm, $Ext]",
3213 def PRFM_Xm_RegOffset : A64I_LSregoff<0b11, 0b0, 0b10, 0b1, (outs),
3214 (ins prefetch_op:$Rt, GPR64xsp:$Rn,
3215 GPR64:$Rm, dword_Xm_regext:$Ext),
3216 "prfm\t$Rt, [$Rn, $Rm, $Ext]",
3220 def : InstAlias<"prfm $Rt, [$Rn, $Rm]",
3221 (PRFM_Xm_RegOffset prefetch_op:$Rt, GPR64xsp:$Rn,
3225 def PRFUM : A64I_LSunalimm<0b11, 0b0, 0b10, (outs),
3226 (ins prefetch_op:$Rt, GPR64xsp:$Rn, simm9:$SImm9),
3227 "prfum\t$Rt, [$Rn, $SImm9]",
3231 def : InstAlias<"prfum $Rt, [$Rn]",
3232 (PRFUM prefetch_op:$Rt, GPR64xsp:$Rn, 0)>;
3234 //===----------------------------------------------------------------------===//
3235 // Load-store register (unprivileged) instructions
3236 //===----------------------------------------------------------------------===//
3237 // Contains: LDTRB, LDTRH, LDTRSB, LDTRSH, LDTRSW, STTR, STTRB and STTRH
3239 // These instructions very much mirror the "unscaled immediate" loads, but since
3240 // there are no floating-point variants we need to split them out into their own
3241 // section to avoid instantiation of "ldtr d0, [sp]" etc.
3243 multiclass A64I_LDTRSTTR<bits<2> size, string asmsuffix, RegisterClass GPR,
3245 def _UnPriv_STR : A64I_LSunpriv<size, 0b0, 0b00,
3246 (outs), (ins GPR:$Rt, GPR64xsp:$Rn, simm9:$SImm9),
3247 "sttr" # asmsuffix # "\t$Rt, [$Rn, $SImm9]",
3252 def : InstAlias<"sttr" # asmsuffix # " $Rt, [$Rn]",
3253 (!cast<Instruction>(prefix # "_UnPriv_STR") GPR:$Rt, GPR64xsp:$Rn, 0)>;
3255 def _UnPriv_LDR : A64I_LSunpriv<size, 0b0, 0b01,
3256 (outs GPR:$Rt), (ins GPR64xsp:$Rn, simm9:$SImm9),
3257 "ldtr" # asmsuffix # "\t$Rt, [$Rn, $SImm9]",
3262 def : InstAlias<"ldtr" # asmsuffix # " $Rt, [$Rn]",
3263 (!cast<Instruction>(prefix # "_UnPriv_LDR") GPR:$Rt, GPR64xsp:$Rn, 0)>;
3267 // STTRB/LDTRB: First define the instructions
3268 defm LS8 : A64I_LDTRSTTR<0b00, "b", GPR32, "LS8">;
3271 defm LS16 : A64I_LDTRSTTR<0b01, "h", GPR32, "LS16">;
3273 // STTR/LDTR to/from a W register
3274 defm LS32 : A64I_LDTRSTTR<0b10, "", GPR32, "LS32">;
3276 // STTR/LDTR to/from an X register
3277 defm LS64 : A64I_LDTRSTTR<0b11, "", GPR64, "LS64">;
3279 // Now a class for the signed instructions that can go to either 32 or 64
3281 multiclass A64I_LDTR_signed<bits<2> size, string asmopcode, string prefix> {
3282 let mayLoad = 1 in {
3283 def w : A64I_LSunpriv<size, 0b0, 0b11,
3285 (ins GPR64xsp:$Rn, simm9:$SImm9),
3286 "ldtrs" # asmopcode # "\t$Rt, [$Rn, $SImm9]",
3289 def x : A64I_LSunpriv<size, 0b0, 0b10,
3291 (ins GPR64xsp:$Rn, simm9:$SImm9),
3292 "ldtrs" # asmopcode # "\t$Rt, [$Rn, $SImm9]",
3296 def : InstAlias<"ldtrs" # asmopcode # " $Rt, [$Rn]",
3297 (!cast<Instruction>(prefix # "w") GPR32:$Rt, GPR64xsp:$Rn, 0)>;
3299 def : InstAlias<"ldtrs" # asmopcode # " $Rt, [$Rn]",
3300 (!cast<Instruction>(prefix # "x") GPR64:$Rt, GPR64xsp:$Rn, 0)>;
3305 defm LDTRSB : A64I_LDTR_signed<0b00, "b", "LDTRSB">;
3307 defm LDTRSH : A64I_LDTR_signed<0b01, "h", "LDTRSH">;
3309 // And finally LDTRSW which only goes to 64 bits.
3310 def LDTRSWx : A64I_LSunpriv<0b10, 0b0, 0b10,
3312 (ins GPR64xsp:$Rn, simm9:$SImm9),
3313 "ldtrsw\t$Rt, [$Rn, $SImm9]",
3317 def : InstAlias<"ldtrsw $Rt, [$Rn]", (LDTRSWx GPR64:$Rt, GPR64xsp:$Rn, 0)>;
3319 //===----------------------------------------------------------------------===//
3320 // Load-store register pair (offset) instructions
3321 //===----------------------------------------------------------------------===//
3325 //===----------------------------------------------------------------------===//
3326 // Load-store register pair (post-indexed) instructions
3327 //===----------------------------------------------------------------------===//
3328 // Contains: STP, LDP, LDPSW
3332 //===----------------------------------------------------------------------===//
3333 // Load-store register pair (pre-indexed) instructions
3334 //===----------------------------------------------------------------------===//
3335 // Contains: STP, LDP, LDPSW
3339 //===----------------------------------------------------------------------===//
3340 // Load-store non-temporal register pair (offset) instructions
3341 //===----------------------------------------------------------------------===//
3342 // Contains: STNP, LDNP
3345 // Anything that creates an MCInst (Decoding, selection and AsmParsing) has to
3346 // know the access size via some means. An isolated operand does not have this
3347 // information unless told from here, which means we need separate tablegen
3348 // Operands for each access size. This multiclass takes care of instantiating
3349 // the correct template functions in the rest of the backend.
3351 multiclass offsets_simm7<string MemSize, string prefix> {
3352 // The bare signed 7-bit immediate is used in post-indexed instructions, but
3353 // because of the scaling performed a generic "simm7" operand isn't
3354 // appropriate here either.
3355 def simm7_asmoperand : AsmOperandClass {
3356 let Name = "SImm7_Scaled" # MemSize;
3357 let PredicateMethod = "isSImm7Scaled<" # MemSize # ">";
3358 let RenderMethod = "addSImm7ScaledOperands<" # MemSize # ">";
3361 def simm7 : Operand<i64> {
3362 let PrintMethod = "printSImm7ScaledOperand<" # MemSize # ">";
3363 let ParserMatchClass = !cast<AsmOperandClass>(prefix # "simm7_asmoperand");
3367 defm word_ : offsets_simm7<"4", "word_">;
3368 defm dword_ : offsets_simm7<"8", "dword_">;
3369 defm qword_ : offsets_simm7<"16", "qword_">;
3371 multiclass A64I_LSPsimple<bits<2> opc, bit v, RegisterClass SomeReg,
3372 Operand simm7, string prefix> {
3373 def _STR : A64I_LSPoffset<opc, v, 0b0, (outs),
3374 (ins SomeReg:$Rt, SomeReg:$Rt2, GPR64xsp:$Rn, simm7:$SImm7),
3375 "stp\t$Rt, $Rt2, [$Rn, $SImm7]", [], NoItinerary> {
3377 let DecoderMethod = "DecodeLDSTPairInstruction";
3379 def : InstAlias<"stp $Rt, $Rt2, [$Rn]",
3380 (!cast<Instruction>(prefix # "_STR") SomeReg:$Rt,
3381 SomeReg:$Rt2, GPR64xsp:$Rn, 0)>;
3383 def _LDR : A64I_LSPoffset<opc, v, 0b1,
3384 (outs SomeReg:$Rt, SomeReg:$Rt2),
3385 (ins GPR64xsp:$Rn, simm7:$SImm7),
3386 "ldp\t$Rt, $Rt2, [$Rn, $SImm7]", [], NoItinerary> {
3388 let DecoderMethod = "DecodeLDSTPairInstruction";
3390 def : InstAlias<"ldp $Rt, $Rt2, [$Rn]",
3391 (!cast<Instruction>(prefix # "_LDR") SomeReg:$Rt,
3392 SomeReg:$Rt2, GPR64xsp:$Rn, 0)>;
3394 def _PostInd_STR : A64I_LSPpostind<opc, v, 0b0,
3395 (outs GPR64xsp:$Rn_wb),
3396 (ins SomeReg:$Rt, SomeReg:$Rt2,
3399 "stp\t$Rt, $Rt2, [$Rn], $SImm7",
3402 let Constraints = "$Rn = $Rn_wb";
3404 // Decoder only needed for unpredictability checking (FIXME).
3405 let DecoderMethod = "DecodeLDSTPairInstruction";
3408 def _PostInd_LDR : A64I_LSPpostind<opc, v, 0b1,
3409 (outs SomeReg:$Rt, SomeReg:$Rt2, GPR64xsp:$Rn_wb),
3410 (ins GPR64xsp:$Rn, simm7:$SImm7),
3411 "ldp\t$Rt, $Rt2, [$Rn], $SImm7",
3414 let Constraints = "$Rn = $Rn_wb";
3415 let DecoderMethod = "DecodeLDSTPairInstruction";
3418 def _PreInd_STR : A64I_LSPpreind<opc, v, 0b0, (outs GPR64xsp:$Rn_wb),
3419 (ins SomeReg:$Rt, SomeReg:$Rt2, GPR64xsp:$Rn, simm7:$SImm7),
3420 "stp\t$Rt, $Rt2, [$Rn, $SImm7]!",
3423 let Constraints = "$Rn = $Rn_wb";
3424 let DecoderMethod = "DecodeLDSTPairInstruction";
3427 def _PreInd_LDR : A64I_LSPpreind<opc, v, 0b1,
3428 (outs SomeReg:$Rt, SomeReg:$Rt2, GPR64xsp:$Rn_wb),
3429 (ins GPR64xsp:$Rn, simm7:$SImm7),
3430 "ldp\t$Rt, $Rt2, [$Rn, $SImm7]!",
3433 let Constraints = "$Rn = $Rn_wb";
3434 let DecoderMethod = "DecodeLDSTPairInstruction";
3437 def _NonTemp_STR : A64I_LSPnontemp<opc, v, 0b0, (outs),
3438 (ins SomeReg:$Rt, SomeReg:$Rt2, GPR64xsp:$Rn, simm7:$SImm7),
3439 "stnp\t$Rt, $Rt2, [$Rn, $SImm7]", [], NoItinerary> {
3441 let DecoderMethod = "DecodeLDSTPairInstruction";
3443 def : InstAlias<"stnp $Rt, $Rt2, [$Rn]",
3444 (!cast<Instruction>(prefix # "_NonTemp_STR") SomeReg:$Rt,
3445 SomeReg:$Rt2, GPR64xsp:$Rn, 0)>;
3447 def _NonTemp_LDR : A64I_LSPnontemp<opc, v, 0b1,
3448 (outs SomeReg:$Rt, SomeReg:$Rt2),
3449 (ins GPR64xsp:$Rn, simm7:$SImm7),
3450 "ldnp\t$Rt, $Rt2, [$Rn, $SImm7]", [], NoItinerary> {
3452 let DecoderMethod = "DecodeLDSTPairInstruction";
3454 def : InstAlias<"ldnp $Rt, $Rt2, [$Rn]",
3455 (!cast<Instruction>(prefix # "_NonTemp_LDR") SomeReg:$Rt,
3456 SomeReg:$Rt2, GPR64xsp:$Rn, 0)>;
3461 defm LSPair32 : A64I_LSPsimple<0b00, 0b0, GPR32, word_simm7, "LSPair32">;
3462 defm LSPair64 : A64I_LSPsimple<0b10, 0b0, GPR64, dword_simm7, "LSPair64">;
3463 defm LSFPPair32 : A64I_LSPsimple<0b00, 0b1, FPR32, word_simm7, "LSFPPair32">;
3464 defm LSFPPair64 : A64I_LSPsimple<0b01, 0b1, FPR64, dword_simm7, "LSFPPair64">;
3465 defm LSFPPair128 : A64I_LSPsimple<0b10, 0b1, FPR128, qword_simm7,
3469 def LDPSWx : A64I_LSPoffset<0b01, 0b0, 0b1,
3470 (outs GPR64:$Rt, GPR64:$Rt2),
3471 (ins GPR64xsp:$Rn, word_simm7:$SImm7),
3472 "ldpsw\t$Rt, $Rt2, [$Rn, $SImm7]", [], NoItinerary> {
3474 let DecoderMethod = "DecodeLDSTPairInstruction";
3476 def : InstAlias<"ldpsw $Rt, $Rt2, [$Rn]",
3477 (LDPSWx GPR64:$Rt, GPR64:$Rt2, GPR64xsp:$Rn, 0)>;
3479 def LDPSWx_PostInd : A64I_LSPpostind<0b01, 0b0, 0b1,
3480 (outs GPR64:$Rt, GPR64:$Rt2, GPR64:$Rn_wb),
3481 (ins GPR64xsp:$Rn, word_simm7:$SImm7),
3482 "ldpsw\t$Rt, $Rt2, [$Rn], $SImm7",
3485 let Constraints = "$Rn = $Rn_wb";
3486 let DecoderMethod = "DecodeLDSTPairInstruction";
3489 def LDPSWx_PreInd : A64I_LSPpreind<0b01, 0b0, 0b1,
3490 (outs GPR64:$Rt, GPR64:$Rt2, GPR64:$Rn_wb),
3491 (ins GPR64xsp:$Rn, word_simm7:$SImm7),
3492 "ldpsw\t$Rt, $Rt2, [$Rn, $SImm7]!",
3495 let Constraints = "$Rn = $Rn_wb";
3496 let DecoderMethod = "DecodeLDSTPairInstruction";
3499 //===----------------------------------------------------------------------===//
3500 // Logical (immediate) instructions
3501 //===----------------------------------------------------------------------===//
3502 // Contains: AND, ORR, EOR, ANDS, + aliases TST, MOV
3504 multiclass logical_imm_operands<string prefix, string note,
3505 int size, ValueType VT> {
3506 def _asmoperand : AsmOperandClass {
3507 let Name = "LogicalImm" # note # size;
3508 let PredicateMethod = "isLogicalImm" # note # "<" # size # ">";
3509 let RenderMethod = "addLogicalImmOperands<" # size # ">";
3513 : Operand<VT>, ComplexPattern<VT, 1, "SelectLogicalImm", [imm]> {
3514 let ParserMatchClass = !cast<AsmOperandClass>(prefix # "_asmoperand");
3515 let PrintMethod = "printLogicalImmOperand<" # size # ">";
3516 let DecoderMethod = "DecodeLogicalImmOperand<" # size # ">";
3520 defm logical_imm32 : logical_imm_operands<"logical_imm32", "", 32, i32>;
3521 defm logical_imm64 : logical_imm_operands<"logical_imm64", "", 64, i64>;
3523 // The mov versions only differ in assembly parsing, where they
3524 // exclude values representable with either MOVZ or MOVN.
3525 defm logical_imm32_mov
3526 : logical_imm_operands<"logical_imm32_mov", "MOV", 32, i32>;
3527 defm logical_imm64_mov
3528 : logical_imm_operands<"logical_imm64_mov", "MOV", 64, i64>;
3531 multiclass A64I_logimmSizes<bits<2> opc, string asmop, SDNode opnode> {
3532 def wwi : A64I_logicalimm<0b0, opc, (outs GPR32wsp:$Rd),
3533 (ins GPR32:$Rn, logical_imm32_operand:$Imm),
3534 !strconcat(asmop, "\t$Rd, $Rn, $Imm"),
3536 (opnode GPR32:$Rn, logical_imm32_operand:$Imm))],
3539 def xxi : A64I_logicalimm<0b1, opc, (outs GPR64xsp:$Rd),
3540 (ins GPR64:$Rn, logical_imm64_operand:$Imm),
3541 !strconcat(asmop, "\t$Rd, $Rn, $Imm"),
3543 (opnode GPR64:$Rn, logical_imm64_operand:$Imm))],
3547 defm AND : A64I_logimmSizes<0b00, "and", and>;
3548 defm ORR : A64I_logimmSizes<0b01, "orr", or>;
3549 defm EOR : A64I_logimmSizes<0b10, "eor", xor>;
3551 let Defs = [NZCV] in {
3552 def ANDSwwi : A64I_logicalimm<0b0, 0b11, (outs GPR32:$Rd),
3553 (ins GPR32:$Rn, logical_imm32_operand:$Imm),
3554 "ands\t$Rd, $Rn, $Imm",
3557 def ANDSxxi : A64I_logicalimm<0b1, 0b11, (outs GPR64:$Rd),
3558 (ins GPR64:$Rn, logical_imm64_operand:$Imm),
3559 "ands\t$Rd, $Rn, $Imm",
3564 def : InstAlias<"tst $Rn, $Imm",
3565 (ANDSwwi WZR, GPR32:$Rn, logical_imm32_operand:$Imm)>;
3566 def : InstAlias<"tst $Rn, $Imm",
3567 (ANDSxxi XZR, GPR64:$Rn, logical_imm64_operand:$Imm)>;
3568 def : InstAlias<"mov $Rd, $Imm",
3569 (ORRwwi GPR32wsp:$Rd, WZR, logical_imm32_mov_operand:$Imm)>;
3570 def : InstAlias<"mov $Rd, $Imm",
3571 (ORRxxi GPR64xsp:$Rd, XZR, logical_imm64_mov_operand:$Imm)>;
3573 //===----------------------------------------------------------------------===//
3574 // Logical (shifted register) instructions
3575 //===----------------------------------------------------------------------===//
3576 // Contains: AND, BIC, ORR, ORN, EOR, EON, ANDS, BICS + aliases TST, MVN, MOV
3578 // Operand for optimizing (icmp (and LHS, RHS), 0, SomeCode). In theory "ANDS"
3579 // behaves differently for unsigned comparisons, so we defensively only allow
3580 // signed or n/a as the operand. In practice "unsigned greater than 0" is "not
3581 // equal to 0" and LLVM gives us this.
3582 def signed_cond : PatLeaf<(cond), [{
3583 return !isUnsignedIntSetCC(N->get());
3587 // These instructions share their "shift" operands with add/sub (shifted
3588 // register instructions). They are defined there.
3590 // N.b. the commutable parameter is just !N. It will be first against the wall
3591 // when the revolution comes.
3592 multiclass logical_shifts<string prefix, bit sf, bits<2> opc,
3593 bit N, bit commutable,
3594 string asmop, SDPatternOperator opfrag, string sty,
3595 RegisterClass GPR, list<Register> defs> {
3596 let isCommutable = commutable, Defs = defs in {
3597 def _lsl : A64I_logicalshift<sf, opc, 0b00, N,
3599 (ins GPR:$Rn, GPR:$Rm,
3600 !cast<Operand>("lsl_operand_" # sty):$Imm6),
3601 !strconcat(asmop, "\t$Rd, $Rn, $Rm, $Imm6"),
3602 [(set GPR:$Rd, (opfrag GPR:$Rn, (shl GPR:$Rm,
3603 !cast<Operand>("lsl_operand_" # sty):$Imm6))
3607 def _lsr : A64I_logicalshift<sf, opc, 0b01, N,
3609 (ins GPR:$Rn, GPR:$Rm,
3610 !cast<Operand>("lsr_operand_" # sty):$Imm6),
3611 !strconcat(asmop, "\t$Rd, $Rn, $Rm, $Imm6"),
3612 [(set GPR:$Rd, (opfrag GPR:$Rn, (srl GPR:$Rm,
3613 !cast<Operand>("lsr_operand_" # sty):$Imm6))
3617 def _asr : A64I_logicalshift<sf, opc, 0b10, N,
3619 (ins GPR:$Rn, GPR:$Rm,
3620 !cast<Operand>("asr_operand_" # sty):$Imm6),
3621 !strconcat(asmop, "\t$Rd, $Rn, $Rm, $Imm6"),
3622 [(set GPR:$Rd, (opfrag GPR:$Rn, (sra GPR:$Rm,
3623 !cast<Operand>("asr_operand_" # sty):$Imm6))
3627 def _ror : A64I_logicalshift<sf, opc, 0b11, N,
3629 (ins GPR:$Rn, GPR:$Rm,
3630 !cast<Operand>("ror_operand_" # sty):$Imm6),
3631 !strconcat(asmop, "\t$Rd, $Rn, $Rm, $Imm6"),
3632 [(set GPR:$Rd, (opfrag GPR:$Rn, (rotr GPR:$Rm,
3633 !cast<Operand>("ror_operand_" # sty):$Imm6))
3639 : InstAlias<!strconcat(asmop, " $Rd, $Rn, $Rm"),
3640 (!cast<Instruction>(prefix # "_lsl") GPR:$Rd, GPR:$Rn,
3643 def : Pat<(opfrag GPR:$Rn, GPR:$Rm),
3644 (!cast<Instruction>(prefix # "_lsl") GPR:$Rn, GPR:$Rm, 0)>;
3647 multiclass logical_sizes<string prefix, bits<2> opc, bit N, bit commutable,
3648 string asmop, SDPatternOperator opfrag,
3649 list<Register> defs> {
3650 defm xxx : logical_shifts<prefix # "xxx", 0b1, opc, N,
3651 commutable, asmop, opfrag, "i64", GPR64, defs>;
3652 defm www : logical_shifts<prefix # "www", 0b0, opc, N,
3653 commutable, asmop, opfrag, "i32", GPR32, defs>;
3657 defm AND : logical_sizes<"AND", 0b00, 0b0, 0b1, "and", and, []>;
3658 defm ORR : logical_sizes<"ORR", 0b01, 0b0, 0b1, "orr", or, []>;
3659 defm EOR : logical_sizes<"EOR", 0b10, 0b0, 0b1, "eor", xor, []>;
3660 defm ANDS : logical_sizes<"ANDS", 0b11, 0b0, 0b1, "ands",
3661 PatFrag<(ops node:$lhs, node:$rhs), (and node:$lhs, node:$rhs),
3662 [{ (void)N; return false; }]>,
3665 defm BIC : logical_sizes<"BIC", 0b00, 0b1, 0b0, "bic",
3666 PatFrag<(ops node:$lhs, node:$rhs),
3667 (and node:$lhs, (not node:$rhs))>, []>;
3668 defm ORN : logical_sizes<"ORN", 0b01, 0b1, 0b0, "orn",
3669 PatFrag<(ops node:$lhs, node:$rhs),
3670 (or node:$lhs, (not node:$rhs))>, []>;
3671 defm EON : logical_sizes<"EON", 0b10, 0b1, 0b0, "eon",
3672 PatFrag<(ops node:$lhs, node:$rhs),
3673 (xor node:$lhs, (not node:$rhs))>, []>;
3674 defm BICS : logical_sizes<"BICS", 0b11, 0b1, 0b0, "bics",
3675 PatFrag<(ops node:$lhs, node:$rhs),
3676 (and node:$lhs, (not node:$rhs)),
3677 [{ (void)N; return false; }]>,
3680 multiclass tst_shifts<string prefix, bit sf, string sty, RegisterClass GPR> {
3681 let isCommutable = 1, Rd = 0b11111, Defs = [NZCV] in {
3682 def _lsl : A64I_logicalshift<sf, 0b11, 0b00, 0b0,
3684 (ins GPR:$Rn, GPR:$Rm,
3685 !cast<Operand>("lsl_operand_" # sty):$Imm6),
3686 "tst\t$Rn, $Rm, $Imm6",
3687 [(set NZCV, (A64setcc (and GPR:$Rn, (shl GPR:$Rm,
3688 !cast<Operand>("lsl_operand_" # sty):$Imm6)),
3693 def _lsr : A64I_logicalshift<sf, 0b11, 0b01, 0b0,
3695 (ins GPR:$Rn, GPR:$Rm,
3696 !cast<Operand>("lsr_operand_" # sty):$Imm6),
3697 "tst\t$Rn, $Rm, $Imm6",
3698 [(set NZCV, (A64setcc (and GPR:$Rn, (srl GPR:$Rm,
3699 !cast<Operand>("lsr_operand_" # sty):$Imm6)),
3703 def _asr : A64I_logicalshift<sf, 0b11, 0b10, 0b0,
3705 (ins GPR:$Rn, GPR:$Rm,
3706 !cast<Operand>("asr_operand_" # sty):$Imm6),
3707 "tst\t$Rn, $Rm, $Imm6",
3708 [(set NZCV, (A64setcc (and GPR:$Rn, (sra GPR:$Rm,
3709 !cast<Operand>("asr_operand_" # sty):$Imm6)),
3713 def _ror : A64I_logicalshift<sf, 0b11, 0b11, 0b0,
3715 (ins GPR:$Rn, GPR:$Rm,
3716 !cast<Operand>("ror_operand_" # sty):$Imm6),
3717 "tst\t$Rn, $Rm, $Imm6",
3718 [(set NZCV, (A64setcc (and GPR:$Rn, (rotr GPR:$Rm,
3719 !cast<Operand>("ror_operand_" # sty):$Imm6)),
3724 def _noshift : InstAlias<"tst $Rn, $Rm",
3725 (!cast<Instruction>(prefix # "_lsl") GPR:$Rn, GPR:$Rm, 0)>;
3727 def : Pat<(A64setcc (and GPR:$Rn, GPR:$Rm), 0, signed_cond),
3728 (!cast<Instruction>(prefix # "_lsl") GPR:$Rn, GPR:$Rm, 0)>;
3731 defm TSTxx : tst_shifts<"TSTxx", 0b1, "i64", GPR64>;
3732 defm TSTww : tst_shifts<"TSTww", 0b0, "i32", GPR32>;
3735 multiclass mvn_shifts<string prefix, bit sf, string sty, RegisterClass GPR> {
3736 let isCommutable = 0, Rn = 0b11111 in {
3737 def _lsl : A64I_logicalshift<sf, 0b01, 0b00, 0b1,
3740 !cast<Operand>("lsl_operand_" # sty):$Imm6),
3741 "mvn\t$Rd, $Rm, $Imm6",
3742 [(set GPR:$Rd, (not (shl GPR:$Rm,
3743 !cast<Operand>("lsl_operand_" # sty):$Imm6)))],
3747 def _lsr : A64I_logicalshift<sf, 0b01, 0b01, 0b1,
3750 !cast<Operand>("lsr_operand_" # sty):$Imm6),
3751 "mvn\t$Rd, $Rm, $Imm6",
3752 [(set GPR:$Rd, (not (srl GPR:$Rm,
3753 !cast<Operand>("lsr_operand_" # sty):$Imm6)))],
3756 def _asr : A64I_logicalshift<sf, 0b01, 0b10, 0b1,
3759 !cast<Operand>("asr_operand_" # sty):$Imm6),
3760 "mvn\t$Rd, $Rm, $Imm6",
3761 [(set GPR:$Rd, (not (sra GPR:$Rm,
3762 !cast<Operand>("asr_operand_" # sty):$Imm6)))],
3765 def _ror : A64I_logicalshift<sf, 0b01, 0b11, 0b1,
3768 !cast<Operand>("ror_operand_" # sty):$Imm6),
3769 "mvn\t$Rd, $Rm, $Imm6",
3770 [(set GPR:$Rd, (not (rotr GPR:$Rm,
3771 !cast<Operand>("lsl_operand_" # sty):$Imm6)))],
3775 def _noshift : InstAlias<"mvn $Rn, $Rm",
3776 (!cast<Instruction>(prefix # "_lsl") GPR:$Rn, GPR:$Rm, 0)>;
3778 def : Pat<(not GPR:$Rm),
3779 (!cast<Instruction>(prefix # "_lsl") GPR:$Rm, 0)>;
3782 defm MVNxx : mvn_shifts<"MVNxx", 0b1, "i64", GPR64>;
3783 defm MVNww : mvn_shifts<"MVNww", 0b0, "i32", GPR32>;
3785 def MOVxx :InstAlias<"mov $Rd, $Rm", (ORRxxx_lsl GPR64:$Rd, XZR, GPR64:$Rm, 0)>;
3786 def MOVww :InstAlias<"mov $Rd, $Rm", (ORRwww_lsl GPR32:$Rd, WZR, GPR32:$Rm, 0)>;
3788 //===----------------------------------------------------------------------===//
3789 // Move wide (immediate) instructions
3790 //===----------------------------------------------------------------------===//
3791 // Contains: MOVN, MOVZ, MOVK + MOV aliases
3793 // A wide variety of different relocations are needed for variants of these
3794 // instructions, so it turns out that we need a different operand for all of
3796 multiclass movw_operands<string prefix, string instname, int width> {
3797 def _imm_asmoperand : AsmOperandClass {
3798 let Name = instname # width # "Shifted" # shift;
3799 let PredicateMethod = "is" # instname # width # "Imm";
3800 let RenderMethod = "addMoveWideImmOperands";
3802 let ParserMethod = "ParseImmWithLSLOperand";
3805 def _imm : Operand<i32> {
3806 let ParserMatchClass = !cast<AsmOperandClass>(prefix # "_imm_asmoperand");
3807 let PrintMethod = "printMoveWideImmOperand";
3808 let EncoderMethod = "getMoveWideImmOpValue";
3809 let DecoderMethod = "DecodeMoveWideImmOperand<" # width # ">";
3811 let MIOperandInfo = (ops uimm16:$UImm16, imm:$Shift);
3815 defm movn32 : movw_operands<"movn32", "MOVN", 32>;
3816 defm movn64 : movw_operands<"movn64", "MOVN", 64>;
3817 defm movz32 : movw_operands<"movz32", "MOVZ", 32>;
3818 defm movz64 : movw_operands<"movz64", "MOVZ", 64>;
3819 defm movk32 : movw_operands<"movk32", "MOVK", 32>;
3820 defm movk64 : movw_operands<"movk64", "MOVK", 64>;
3822 multiclass A64I_movwSizes<bits<2> opc, string asmop, dag ins32bit,
3825 def wii : A64I_movw<0b0, opc, (outs GPR32:$Rd), ins32bit,
3826 !strconcat(asmop, "\t$Rd, $FullImm"),
3829 let UImm16 = FullImm{15-0};
3830 let Shift = FullImm{17-16};
3833 def xii : A64I_movw<0b1, opc, (outs GPR64:$Rd), ins64bit,
3834 !strconcat(asmop, "\t$Rd, $FullImm"),
3837 let UImm16 = FullImm{15-0};
3838 let Shift = FullImm{17-16};
3842 let isMoveImm = 1, isReMaterializable = 1,
3843 isAsCheapAsAMove = 1, neverHasSideEffects = 1 in {
3844 defm MOVN : A64I_movwSizes<0b00, "movn",
3845 (ins movn32_imm:$FullImm),
3846 (ins movn64_imm:$FullImm)>;
3848 // Some relocations are able to convert between a MOVZ and a MOVN. If these
3849 // are applied the instruction must be emitted with the corresponding bits as
3850 // 0, which means a MOVZ needs to override that bit from the default.
3851 let PostEncoderMethod = "fixMOVZ" in
3852 defm MOVZ : A64I_movwSizes<0b10, "movz",
3853 (ins movz32_imm:$FullImm),
3854 (ins movz64_imm:$FullImm)>;
3857 let Constraints = "$src = $Rd" in
3858 defm MOVK : A64I_movwSizes<0b11, "movk",
3859 (ins GPR32:$src, movk32_imm:$FullImm),
3860 (ins GPR64:$src, movk64_imm:$FullImm)>;
3863 // And now the "MOV" aliases. These also need their own operands because what
3864 // they accept is completely different to what the base instructions accept.
3865 multiclass movalias_operand<string prefix, string basename,
3866 string immpredicate, int width> {
3867 def _asmoperand : AsmOperandClass {
3868 let Name = basename # width # "MovAlias";
3870 = "isMoveWideMovAlias<" # width # ", A64Imms::" # immpredicate # ">";
3872 = "addMoveWideMovAliasOperands<" # width # ", "
3873 # "A64Imms::" # immpredicate # ">";
3876 def _movimm : Operand<i32> {
3877 let ParserMatchClass = !cast<AsmOperandClass>(prefix # "_asmoperand");
3879 let MIOperandInfo = (ops uimm16:$UImm16, imm:$Shift);
3883 defm movz32 : movalias_operand<"movz32", "MOVZ", "isMOVZImm", 32>;
3884 defm movz64 : movalias_operand<"movz64", "MOVZ", "isMOVZImm", 64>;
3885 defm movn32 : movalias_operand<"movn32", "MOVN", "isOnlyMOVNImm", 32>;
3886 defm movn64 : movalias_operand<"movn64", "MOVN", "isOnlyMOVNImm", 64>;
3888 // FIXME: these are officially canonical aliases, but TableGen is too limited to
3889 // print them at the moment. I believe in this case an "AliasPredicate" method
3890 // will need to be implemented. to allow it, as well as the more generally
3891 // useful handling of non-register, non-constant operands.
3892 class movalias<Instruction INST, RegisterClass GPR, Operand operand>
3893 : InstAlias<"mov $Rd, $FullImm", (INST GPR:$Rd, operand:$FullImm)>;
3895 def : movalias<MOVZwii, GPR32, movz32_movimm>;
3896 def : movalias<MOVZxii, GPR64, movz64_movimm>;
3897 def : movalias<MOVNwii, GPR32, movn32_movimm>;
3898 def : movalias<MOVNxii, GPR64, movn64_movimm>;
3900 //===----------------------------------------------------------------------===//
3901 // PC-relative addressing instructions
3902 //===----------------------------------------------------------------------===//
3903 // Contains: ADR, ADRP
3905 def adr_label : Operand<i64> {
3906 let EncoderMethod = "getLabelOpValue<AArch64::fixup_a64_adr_prel>";
3908 // This label is a 21-bit offset from PC, unscaled
3909 let PrintMethod = "printLabelOperand<21, 1>";
3910 let ParserMatchClass = label_asmoperand<21, 1>;
3911 let OperandType = "OPERAND_PCREL";
3914 def adrp_label_asmoperand : AsmOperandClass {
3915 let Name = "AdrpLabel";
3916 let RenderMethod = "addLabelOperands<21, 4096>";
3919 def adrp_label : Operand<i64> {
3920 let EncoderMethod = "getAdrpLabelOpValue";
3922 // This label is a 21-bit offset from PC, scaled by the page-size: 4096.
3923 let PrintMethod = "printLabelOperand<21, 4096>";
3924 let ParserMatchClass = adrp_label_asmoperand;
3925 let OperandType = "OPERAND_PCREL";
3928 let neverHasSideEffects = 1 in {
3929 def ADRxi : A64I_PCADR<0b0, (outs GPR64:$Rd), (ins adr_label:$Label),
3930 "adr\t$Rd, $Label", [], NoItinerary>;
3932 def ADRPxi : A64I_PCADR<0b1, (outs GPR64:$Rd), (ins adrp_label:$Label),
3933 "adrp\t$Rd, $Label", [], NoItinerary>;
3936 //===----------------------------------------------------------------------===//
3937 // System instructions
3938 //===----------------------------------------------------------------------===//
3939 // Contains: HINT, CLREX, DSB, DMB, ISB, MSR, SYS, SYSL, MRS
3940 // + aliases IC, DC, AT, TLBI, NOP, YIELD, WFE, WFI, SEV, SEVL
3942 // Op1 and Op2 fields are sometimes simple 3-bit unsigned immediate values.
3943 def uimm3_asmoperand : AsmOperandClass {
3945 let PredicateMethod = "isUImm<3>";
3946 let RenderMethod = "addImmOperands";
3949 def uimm3 : Operand<i32> {
3950 let ParserMatchClass = uimm3_asmoperand;
3953 // The HINT alias can accept a simple unsigned 7-bit immediate.
3954 def uimm7_asmoperand : AsmOperandClass {
3956 let PredicateMethod = "isUImm<7>";
3957 let RenderMethod = "addImmOperands";
3960 def uimm7 : Operand<i32> {
3961 let ParserMatchClass = uimm7_asmoperand;
3964 // Multiclass namedimm is defined with the prefetch operands. Most of these fit
3965 // into the NamedImmMapper scheme well: they either accept a named operand or
3966 // any immediate under a particular value (which may be 0, implying no immediate
3968 defm dbarrier : namedimm<"dbarrier", "A64DB::DBarrierMapper">;
3969 defm isb : namedimm<"isb", "A64ISB::ISBMapper">;
3970 defm ic : namedimm<"ic", "A64IC::ICMapper">;
3971 defm dc : namedimm<"dc", "A64DC::DCMapper">;
3972 defm at : namedimm<"at", "A64AT::ATMapper">;
3973 defm tlbi : namedimm<"tlbi", "A64TLBI::TLBIMapper">;
3975 // However, MRS and MSR are more complicated for a few reasons:
3976 // * There are ~1000 generic names S3_<op1>_<CRn>_<CRm>_<Op2> which have an
3977 // implementation-defined effect
3978 // * Most registers are shared, but some are read-only or write-only.
3979 // * There is a variant of MSR which accepts the same register name (SPSel),
3980 // but which would have a different encoding.
3982 // In principle these could be resolved in with more complicated subclasses of
3983 // NamedImmMapper, however that imposes an overhead on other "named
3984 // immediates". Both in concrete terms with virtual tables and in unnecessary
3987 // The solution adopted here is to take the MRS/MSR Mappers out of the usual
3988 // hierarchy (they're not derived from NamedImmMapper) and to add logic for
3989 // their special situation.
3990 def mrs_asmoperand : AsmOperandClass {
3992 let ParserMethod = "ParseSysRegOperand";
3995 def mrs_op : Operand<i32> {
3996 let ParserMatchClass = mrs_asmoperand;
3997 let PrintMethod = "printMRSOperand";
3998 let DecoderMethod = "DecodeMRSOperand";
4001 def msr_asmoperand : AsmOperandClass {
4002 let Name = "MSRWithReg";
4004 // Note that SPSel is valid for both this and the pstate operands, but with
4005 // different immediate encodings. This is why these operands provide a string
4006 // AArch64Operand rather than an immediate. The overlap is small enough that
4007 // it could be resolved with hackery now, but who can say in future?
4008 let ParserMethod = "ParseSysRegOperand";
4011 def msr_op : Operand<i32> {
4012 let ParserMatchClass = msr_asmoperand;
4013 let PrintMethod = "printMSROperand";
4014 let DecoderMethod = "DecodeMSROperand";
4017 def pstate_asmoperand : AsmOperandClass {
4018 let Name = "MSRPState";
4019 // See comment above about parser.
4020 let ParserMethod = "ParseSysRegOperand";
4023 def pstate_op : Operand<i32> {
4024 let ParserMatchClass = pstate_asmoperand;
4025 let PrintMethod = "printNamedImmOperand<A64PState::PStateMapper>";
4026 let DecoderMethod = "DecodeNamedImmOperand<A64PState::PStateMapper>";
4029 // When <CRn> is specified, an assembler should accept something like "C4", not
4030 // the usual "#4" immediate.
4031 def CRx_asmoperand : AsmOperandClass {
4033 let PredicateMethod = "isUImm<4>";
4034 let RenderMethod = "addImmOperands";
4035 let ParserMethod = "ParseCRxOperand";
4038 def CRx : Operand<i32> {
4039 let ParserMatchClass = CRx_asmoperand;
4040 let PrintMethod = "printCRxOperand";
4044 // Finally, we can start defining the instructions.
4046 // HINT is straightforward, with a few aliases.
4047 def HINTi : A64I_system<0b0, (outs), (ins uimm7:$UImm7), "hint\t$UImm7",
4050 let CRm = UImm7{6-3};
4051 let Op2 = UImm7{2-0};
4059 def : InstAlias<"nop", (HINTi 0)>;
4060 def : InstAlias<"yield", (HINTi 1)>;
4061 def : InstAlias<"wfe", (HINTi 2)>;
4062 def : InstAlias<"wfi", (HINTi 3)>;
4063 def : InstAlias<"sev", (HINTi 4)>;
4064 def : InstAlias<"sevl", (HINTi 5)>;
4066 // Quite a few instructions then follow a similar pattern of fixing common
4067 // fields in the bitpattern, we'll define a helper-class for them.
4068 class simple_sys<bits<2> op0, bits<3> op1, bits<4> crn, bits<3> op2,
4069 Operand operand, string asmop>
4070 : A64I_system<0b0, (outs), (ins operand:$CRm), !strconcat(asmop, "\t$CRm"),
4080 def CLREXi : simple_sys<0b00, 0b011, 0b0011, 0b010, uimm4, "clrex">;
4081 def DSBi : simple_sys<0b00, 0b011, 0b0011, 0b100, dbarrier_op, "dsb">;
4082 def DMBi : simple_sys<0b00, 0b011, 0b0011, 0b101, dbarrier_op, "dmb">;
4083 def ISBi : simple_sys<0b00, 0b011, 0b0011, 0b110, isb_op, "isb">;
4085 def : InstAlias<"clrex", (CLREXi 0b1111)>;
4086 def : InstAlias<"isb", (ISBi 0b1111)>;
4088 // (DMBi 0xb) is a "DMB ISH" instruciton, appropriate for Linux SMP
4089 // configurations at least.
4090 def : Pat<(atomic_fence imm, imm), (DMBi 0xb)>;
4092 // Any SYS bitpattern can be represented with a complex and opaque "SYS"
4094 def SYSiccix : A64I_system<0b0, (outs),
4095 (ins uimm3:$Op1, CRx:$CRn, CRx:$CRm,
4096 uimm3:$Op2, GPR64:$Rt),
4097 "sys\t$Op1, $CRn, $CRm, $Op2, $Rt",
4102 // You can skip the Xt argument whether it makes sense or not for the generic
4104 def : InstAlias<"sys $Op1, $CRn, $CRm, $Op2",
4105 (SYSiccix uimm3:$Op1, CRx:$CRn, CRx:$CRm, uimm3:$Op2, XZR)>;
4108 // But many have aliases, which obviously don't fit into
4109 class SYSalias<dag ins, string asmstring>
4110 : A64I_system<0b0, (outs), ins, asmstring, [], NoItinerary> {
4111 let isAsmParserOnly = 1;
4115 let Op1 = SysOp{13-11};
4116 let CRn = SysOp{10-7};
4117 let CRm = SysOp{6-3};
4118 let Op2 = SysOp{2-0};
4121 def ICix : SYSalias<(ins ic_op:$SysOp, GPR64:$Rt), "ic\t$SysOp, $Rt">;
4123 def ICi : SYSalias<(ins ic_op:$SysOp), "ic\t$SysOp"> {
4127 def DCix : SYSalias<(ins dc_op:$SysOp, GPR64:$Rt), "dc\t$SysOp, $Rt">;
4128 def ATix : SYSalias<(ins at_op:$SysOp, GPR64:$Rt), "at\t$SysOp, $Rt">;
4130 def TLBIix : SYSalias<(ins tlbi_op:$SysOp, GPR64:$Rt), "tlbi\t$SysOp, $Rt">;
4132 def TLBIi : SYSalias<(ins tlbi_op:$SysOp), "tlbi\t$SysOp"> {
4137 def SYSLxicci : A64I_system<0b1, (outs GPR64:$Rt),
4138 (ins uimm3:$Op1, CRx:$CRn, CRx:$CRm, uimm3:$Op2),
4139 "sysl\t$Rt, $Op1, $CRn, $CRm, $Op2",
4144 // The instructions themselves are rather simple for MSR and MRS.
4145 def MSRix : A64I_system<0b0, (outs), (ins msr_op:$SysReg, GPR64:$Rt),
4146 "msr\t$SysReg, $Rt", [], NoItinerary> {
4148 let Op0 = SysReg{15-14};
4149 let Op1 = SysReg{13-11};
4150 let CRn = SysReg{10-7};
4151 let CRm = SysReg{6-3};
4152 let Op2 = SysReg{2-0};
4155 def MRSxi : A64I_system<0b1, (outs GPR64:$Rt), (ins mrs_op:$SysReg),
4156 "mrs\t$Rt, $SysReg", [], NoItinerary> {
4158 let Op0 = SysReg{15-14};
4159 let Op1 = SysReg{13-11};
4160 let CRn = SysReg{10-7};
4161 let CRm = SysReg{6-3};
4162 let Op2 = SysReg{2-0};
4165 def MSRii : A64I_system<0b0, (outs), (ins pstate_op:$PState, uimm4:$CRm),
4166 "msr\t$PState, $CRm", [], NoItinerary> {
4170 let Op1 = PState{5-3};
4172 let Op2 = PState{2-0};
4176 //===----------------------------------------------------------------------===//
4177 // Test & branch (immediate) instructions
4178 //===----------------------------------------------------------------------===//
4179 // Contains: TBZ, TBNZ
4181 // The bit to test is a simple unsigned 6-bit immediate in the X-register
4183 def uimm6 : Operand<i64> {
4184 let ParserMatchClass = uimm6_asmoperand;
4187 def label_wid14_scal4_asmoperand : label_asmoperand<14, 4>;
4189 def tbimm_target : Operand<OtherVT> {
4190 let EncoderMethod = "getLabelOpValue<AArch64::fixup_a64_tstbr>";
4192 // This label is a 14-bit offset from PC, scaled by the instruction-width: 4.
4193 let PrintMethod = "printLabelOperand<14, 4>";
4194 let ParserMatchClass = label_wid14_scal4_asmoperand;
4196 let OperandType = "OPERAND_PCREL";
4199 def A64eq : ImmLeaf<i32, [{ return Imm == A64CC::EQ; }]>;
4200 def A64ne : ImmLeaf<i32, [{ return Imm == A64CC::NE; }]>;
4202 // These instructions correspond to patterns involving "and" with a power of
4203 // two, which we need to be able to select.
4204 def tstb64_pat : ComplexPattern<i64, 1, "SelectTSTBOperand<64>">;
4205 def tstb32_pat : ComplexPattern<i32, 1, "SelectTSTBOperand<32>">;
4207 let isBranch = 1, isTerminator = 1 in {
4208 def TBZxii : A64I_TBimm<0b0, (outs),
4209 (ins GPR64:$Rt, uimm6:$Imm, tbimm_target:$Label),
4210 "tbz\t$Rt, $Imm, $Label",
4211 [(A64br_cc (A64cmp (and GPR64:$Rt, tstb64_pat:$Imm), 0),
4215 def TBNZxii : A64I_TBimm<0b1, (outs),
4216 (ins GPR64:$Rt, uimm6:$Imm, tbimm_target:$Label),
4217 "tbnz\t$Rt, $Imm, $Label",
4218 [(A64br_cc (A64cmp (and GPR64:$Rt, tstb64_pat:$Imm), 0),
4223 // Note, these instructions overlap with the above 64-bit patterns. This is
4224 // intentional, "tbz x3, #1, somewhere" and "tbz w3, #1, somewhere" would both
4225 // do the same thing and are both permitted assembly. They also both have
4226 // sensible DAG patterns.
4227 def TBZwii : A64I_TBimm<0b0, (outs),
4228 (ins GPR32:$Rt, uimm5:$Imm, tbimm_target:$Label),
4229 "tbz\t$Rt, $Imm, $Label",
4230 [(A64br_cc (A64cmp (and GPR32:$Rt, tstb32_pat:$Imm), 0),
4236 def TBNZwii : A64I_TBimm<0b1, (outs),
4237 (ins GPR32:$Rt, uimm5:$Imm, tbimm_target:$Label),
4238 "tbnz\t$Rt, $Imm, $Label",
4239 [(A64br_cc (A64cmp (and GPR32:$Rt, tstb32_pat:$Imm), 0),
4246 //===----------------------------------------------------------------------===//
4247 // Unconditional branch (immediate) instructions
4248 //===----------------------------------------------------------------------===//
4251 def label_wid26_scal4_asmoperand : label_asmoperand<26, 4>;
4253 def bimm_target : Operand<OtherVT> {
4254 let EncoderMethod = "getLabelOpValue<AArch64::fixup_a64_uncondbr>";
4256 // This label is a 26-bit offset from PC, scaled by the instruction-width: 4.
4257 let PrintMethod = "printLabelOperand<26, 4>";
4258 let ParserMatchClass = label_wid26_scal4_asmoperand;
4260 let OperandType = "OPERAND_PCREL";
4263 def blimm_target : Operand<i64> {
4264 let EncoderMethod = "getLabelOpValue<AArch64::fixup_a64_call>";
4266 // This label is a 26-bit offset from PC, scaled by the instruction-width: 4.
4267 let PrintMethod = "printLabelOperand<26, 4>";
4268 let ParserMatchClass = label_wid26_scal4_asmoperand;
4270 let OperandType = "OPERAND_PCREL";
4273 class A64I_BimmImpl<bit op, string asmop, list<dag> patterns, Operand lbl_type>
4274 : A64I_Bimm<op, (outs), (ins lbl_type:$Label),
4275 !strconcat(asmop, "\t$Label"), patterns,
4278 let isBranch = 1 in {
4279 def Bimm : A64I_BimmImpl<0b0, "b", [(br bb:$Label)], bimm_target> {
4280 let isTerminator = 1;
4284 def BLimm : A64I_BimmImpl<0b1, "bl",
4285 [(AArch64Call tglobaladdr:$Label)], blimm_target> {
4291 def : Pat<(AArch64Call texternalsym:$Label), (BLimm texternalsym:$Label)>;
4293 //===----------------------------------------------------------------------===//
4294 // Unconditional branch (register) instructions
4295 //===----------------------------------------------------------------------===//
4296 // Contains: BR, BLR, RET, ERET, DRP.
4298 // Most of the notional opcode fields in the A64I_Breg format are fixed in A64
4300 class A64I_BregImpl<bits<4> opc,
4301 dag outs, dag ins, string asmstr, list<dag> patterns,
4302 InstrItinClass itin = NoItinerary>
4303 : A64I_Breg<opc, 0b11111, 0b000000, 0b00000,
4304 outs, ins, asmstr, patterns, itin> {
4306 let isIndirectBranch = 1;
4309 // Note that these are not marked isCall or isReturn because as far as LLVM is
4310 // concerned they're not. "ret" is just another jump unless it has been selected
4311 // by LLVM as the function's return.
4313 let isBranch = 1 in {
4314 def BRx : A64I_BregImpl<0b0000,(outs), (ins GPR64:$Rn),
4315 "br\t$Rn", [(brind GPR64:$Rn)]> {
4317 let isTerminator = 1;
4320 def BLRx : A64I_BregImpl<0b0001, (outs), (ins GPR64:$Rn),
4321 "blr\t$Rn", [(AArch64Call GPR64:$Rn)]> {
4327 def RETx : A64I_BregImpl<0b0010, (outs), (ins GPR64:$Rn),
4330 let isTerminator = 1;
4334 // Create a separate pseudo-instruction for codegen to use so that we don't
4335 // flag x30 as used in every function. It'll be restored before the RET by the
4336 // epilogue if it's legitimately used.
4337 def RET : A64PseudoExpand<(outs), (ins), [(A64ret)], (RETx (ops X30))> {
4338 let isTerminator = 1;
4343 def ERET : A64I_BregImpl<0b0100, (outs), (ins), "eret", []> {
4346 let isTerminator = 1;
4350 def DRPS : A64I_BregImpl<0b0101, (outs), (ins), "drps", []> {
4356 def RETAlias : InstAlias<"ret", (RETx X30)>;
4359 //===----------------------------------------------------------------------===//
4360 // Address generation patterns
4361 //===----------------------------------------------------------------------===//
4363 // Primary method of address generation for the small/absolute memory model is
4364 // an ADRP/ADR pair:
4365 // ADRP x0, some_variable
4366 // ADD x0, x0, #:lo12:some_variable
4368 // The load/store elision of the ADD is accomplished when selecting
4369 // addressing-modes. This just mops up the cases where that doesn't work and we
4370 // really need an address in some register.
4372 // This wrapper applies a LO12 modifier to the address. Otherwise we could just
4373 // use the same address.
4375 class ADRP_ADD<SDNode Wrapper, SDNode addrop>
4376 : Pat<(Wrapper addrop:$Hi, addrop:$Lo12, (i32 imm)),
4377 (ADDxxi_lsl0_s (ADRPxi addrop:$Hi), addrop:$Lo12)>;
4379 def : ADRP_ADD<A64WrapperSmall, tblockaddress>;
4380 def : ADRP_ADD<A64WrapperSmall, texternalsym>;
4381 def : ADRP_ADD<A64WrapperSmall, tglobaladdr>;
4382 def : ADRP_ADD<A64WrapperSmall, tglobaltlsaddr>;
4383 def : ADRP_ADD<A64WrapperSmall, tjumptable>;
4385 //===----------------------------------------------------------------------===//
4386 // GOT access patterns
4387 //===----------------------------------------------------------------------===//
4391 class GOTLoadSmall<SDNode addrfrag>
4392 : Pat<(A64GOTLoad (A64WrapperSmall addrfrag:$Hi, addrfrag:$Lo12, 8)),
4393 (LS64_LDR (ADRPxi addrfrag:$Hi), addrfrag:$Lo12)>;
4395 def : GOTLoadSmall<texternalsym>;
4396 def : GOTLoadSmall<tglobaladdr>;
4397 def : GOTLoadSmall<tglobaltlsaddr>;
4399 //===----------------------------------------------------------------------===//
4400 // Tail call handling
4401 //===----------------------------------------------------------------------===//
4403 let isCall = 1, isTerminator = 1, isReturn = 1, isBarrier = 1, Uses = [XSP] in {
4405 : PseudoInst<(outs), (ins i64imm:$dst, i32imm:$FPDiff),
4406 [(AArch64tcret tglobaladdr:$dst, (i32 timm:$FPDiff))]>;
4409 : PseudoInst<(outs), (ins tcGPR64:$dst, i32imm:$FPDiff),
4410 [(AArch64tcret tcGPR64:$dst, (i32 timm:$FPDiff))]>;
4413 let isCall = 1, isTerminator = 1, isReturn = 1, isBarrier = 1,
4415 def TAIL_Bimm : A64PseudoExpand<(outs), (ins bimm_target:$Label), [],
4416 (Bimm bimm_target:$Label)>;
4418 def TAIL_BRx : A64PseudoExpand<(outs), (ins tcGPR64:$Rd), [],
4423 def : Pat<(AArch64tcret texternalsym:$dst, (i32 timm:$FPDiff)),
4424 (TC_RETURNdi texternalsym:$dst, imm:$FPDiff)>;
4426 //===----------------------------------------------------------------------===//
4427 // Thread local storage
4428 //===----------------------------------------------------------------------===//
4430 // This is a pseudo-instruction representing the ".tlsdesccall" directive in
4431 // assembly. Its effect is to insert an R_AARCH64_TLSDESC_CALL relocation at the
4432 // current location. It should always be immediately followed by a BLR
4433 // instruction, and is intended solely for relaxation by the linker.
4435 def : Pat<(A64threadpointer), (MRSxi 0xde82)>;
4437 def TLSDESCCALL : PseudoInst<(outs), (ins i64imm:$Lbl), []> {
4438 let hasSideEffects = 1;
4441 def TLSDESC_BLRx : PseudoInst<(outs), (ins GPR64:$Rn, i64imm:$Var),
4442 [(A64tlsdesc_blr GPR64:$Rn, tglobaltlsaddr:$Var)]> {
4447 def : Pat<(A64tlsdesc_blr GPR64:$Rn, texternalsym:$Var),
4448 (TLSDESC_BLRx GPR64:$Rn, texternalsym:$Var)>;
4450 //===----------------------------------------------------------------------===//
4451 // Bitfield patterns
4452 //===----------------------------------------------------------------------===//
4454 def bfi32_lsb_to_immr : SDNodeXForm<imm, [{
4455 return CurDAG->getTargetConstant((32 - N->getZExtValue()) % 32, MVT::i64);
4458 def bfi64_lsb_to_immr : SDNodeXForm<imm, [{
4459 return CurDAG->getTargetConstant((64 - N->getZExtValue()) % 64, MVT::i64);
4462 def bfi_width_to_imms : SDNodeXForm<imm, [{
4463 return CurDAG->getTargetConstant(N->getZExtValue() - 1, MVT::i64);
4467 // The simpler patterns deal with cases where no AND mask is actually needed
4468 // (either all bits are used or the low 32 bits are used).
4469 let AddedComplexity = 10 in {
4471 def : Pat<(A64Bfi GPR64:$src, GPR64:$Rn, imm:$ImmR, imm:$ImmS),
4472 (BFIxxii GPR64:$src, GPR64:$Rn,
4473 (bfi64_lsb_to_immr (i64 imm:$ImmR)),
4474 (bfi_width_to_imms (i64 imm:$ImmS)))>;
4476 def : Pat<(A64Bfi GPR32:$src, GPR32:$Rn, imm:$ImmR, imm:$ImmS),
4477 (BFIwwii GPR32:$src, GPR32:$Rn,
4478 (bfi32_lsb_to_immr (i64 imm:$ImmR)),
4479 (bfi_width_to_imms (i64 imm:$ImmS)))>;
4482 def : Pat<(and (A64Bfi GPR64:$src, GPR64:$Rn, imm:$ImmR, imm:$ImmS),
4484 (SUBREG_TO_REG (i64 0),
4485 (BFIwwii (EXTRACT_SUBREG GPR64:$src, sub_32),
4486 (EXTRACT_SUBREG GPR64:$Rn, sub_32),
4487 (bfi32_lsb_to_immr (i64 imm:$ImmR)),
4488 (bfi_width_to_imms (i64 imm:$ImmS))),
4493 //===----------------------------------------------------------------------===//
4494 // Constant island entries
4495 //===----------------------------------------------------------------------===//
4497 // The constant island pass needs to create "instructions" in the middle of the
4498 // instruction stream to reresent its constants.
4500 def cpinst_operand : Operand<i32>;
4502 def CONSTPOOL_ENTRY : PseudoInst<(outs), (ins cpinst_operand:$instid,
4503 cpinst_operand:$cpidx,
4504 i32imm:$size), []> {
4505 let neverHasSideEffects = 1;
4506 let isNotDuplicable = 1;
4509 //===----------------------------------------------------------------------===//
4510 // Miscellaneous patterns
4511 //===----------------------------------------------------------------------===//
4513 // Truncation from 64 to 32-bits just involves renaming your register.
4514 def : Pat<(i32 (trunc (i64 GPR64:$val))), (EXTRACT_SUBREG GPR64:$val, sub_32)>;
4516 // Similarly, extension where we don't care about the high bits is
4518 def : Pat<(i64 (anyext (i32 GPR32:$val))),
4519 (INSERT_SUBREG (IMPLICIT_DEF), GPR32:$val, sub_32)>;
4521 // SELECT instructions providing f128 types need to be handled by a
4522 // pseudo-instruction since the eventual code will need to introduce basic
4523 // blocks and control flow.
4524 def F128CSEL : PseudoInst<(outs FPR128:$Rd),
4525 (ins FPR128:$Rn, FPR128:$Rm, cond_code_op:$Cond),
4526 [(set FPR128:$Rd, (simple_select (f128 FPR128:$Rn),
4529 let usesCustomInserter = 1;
4532 //===----------------------------------------------------------------------===//
4533 // Load/store patterns
4534 //===----------------------------------------------------------------------===//
4536 // There are lots of patterns here, because we need to allow at least three
4537 // parameters to vary independently.
4538 // 1. Instruction: "ldrb w9, [sp]", "ldrh w9, [sp]", ...
4539 // 2. LLVM source: zextloadi8, anyextloadi8, ...
4540 // 3. Address-generation: A64Wrapper, (add BASE, OFFSET), ...
4542 // The biggest problem turns out to be the address-generation variable. At the
4543 // point of instantiation we need to produce two DAGs, one for the pattern and
4544 // one for the instruction. Doing this at the lowest level of classes doesn't
4547 // Consider the simple uimm12 addressing mode, and the desire to match both (add
4548 // GPR64xsp:$Rn, uimm12:$Offset) and GPR64xsp:$Rn, particularly on the
4549 // instruction side. We'd need to insert either "GPR64xsp" and "uimm12" or
4550 // "GPR64xsp" and "0" into an unknown dag. !subst is not capable of this
4551 // operation, and PatFrags are for selection not output.
4553 // As a result, the address-generation patterns are the final
4554 // instantiations. However, we do still need to vary the operand for the address
4555 // further down (At the point we're deciding A64WrapperSmall, we don't know
4556 // the memory width of the operation).
4558 //===------------------------------
4559 // 1. Basic infrastructural defs
4560 //===------------------------------
4562 // First, some simple classes for !foreach and !subst to use:
4573 // You can't use !subst on an actual immediate, but you *can* use it on an
4574 // operand record that happens to match a single immediate. So we do.
4575 def imm_eq0 : ImmLeaf<i64, [{ return Imm == 0; }]>;
4576 def imm_eq1 : ImmLeaf<i64, [{ return Imm == 1; }]>;
4577 def imm_eq2 : ImmLeaf<i64, [{ return Imm == 2; }]>;
4578 def imm_eq3 : ImmLeaf<i64, [{ return Imm == 3; }]>;
4579 def imm_eq4 : ImmLeaf<i64, [{ return Imm == 4; }]>;
4581 // If the low bits of a pointer are known to be 0 then an "or" is just as good
4582 // as addition for computing an offset. This fragment forwards that check for
4584 def add_like_or : PatFrag<(ops node:$lhs, node:$rhs), (or node:$lhs, node:$rhs),
4586 return CurDAG->isBaseWithConstantOffset(SDValue(N, 0));
4589 // Load/store (unsigned immediate) operations with relocations against global
4590 // symbols (for lo12) are only valid if those symbols have correct alignment
4591 // (since the immediate offset is divided by the access scale, it can't have a
4594 // The guaranteed alignment is provided as part of the WrapperSmall
4595 // operation, and checked against one of these.
4596 def any_align : ImmLeaf<i32, [{ (void)Imm; return true; }]>;
4597 def min_align2 : ImmLeaf<i32, [{ return Imm >= 2; }]>;
4598 def min_align4 : ImmLeaf<i32, [{ return Imm >= 4; }]>;
4599 def min_align8 : ImmLeaf<i32, [{ return Imm >= 8; }]>;
4600 def min_align16 : ImmLeaf<i32, [{ return Imm >= 16; }]>;
4602 // "Normal" load/store instructions can be used on atomic operations, provided
4603 // the ordering parameter is at most "monotonic". Anything above that needs
4604 // special handling with acquire/release instructions.
4605 class simple_load<PatFrag base>
4606 : PatFrag<(ops node:$ptr), (base node:$ptr), [{
4607 return cast<AtomicSDNode>(N)->getOrdering() <= Monotonic;
4610 def atomic_load_simple_i8 : simple_load<atomic_load_8>;
4611 def atomic_load_simple_i16 : simple_load<atomic_load_16>;
4612 def atomic_load_simple_i32 : simple_load<atomic_load_32>;
4613 def atomic_load_simple_i64 : simple_load<atomic_load_64>;
4615 class simple_store<PatFrag base>
4616 : PatFrag<(ops node:$ptr, node:$val), (base node:$ptr, node:$val), [{
4617 return cast<AtomicSDNode>(N)->getOrdering() <= Monotonic;
4620 def atomic_store_simple_i8 : simple_store<atomic_store_8>;
4621 def atomic_store_simple_i16 : simple_store<atomic_store_16>;
4622 def atomic_store_simple_i32 : simple_store<atomic_store_32>;
4623 def atomic_store_simple_i64 : simple_store<atomic_store_64>;
4625 //===------------------------------
4626 // 2. UImm12 and SImm9
4627 //===------------------------------
4629 // These instructions have two operands providing the address so they can be
4630 // treated similarly for most purposes.
4632 //===------------------------------
4633 // 2.1 Base patterns covering extend/truncate semantics
4634 //===------------------------------
4636 // Atomic patterns can be shared between integer operations of all sizes, a
4637 // quick multiclass here allows reuse.
4638 multiclass ls_atomic_pats<Instruction LOAD, Instruction STORE, dag Base,
4639 dag Offset, dag address, RegisterClass TPR,
4641 def : Pat<(!cast<PatFrag>("atomic_load_simple_" # sty) address),
4642 (LOAD Base, Offset)>;
4644 def : Pat<(!cast<PatFrag>("atomic_store_simple_" # sty) address, TPR:$Rt),
4645 (STORE TPR:$Rt, Base, Offset)>;
4648 // Instructions accessing a memory chunk smaller than a register (or, in a
4649 // pinch, the same size) have a characteristic set of patterns they want to
4650 // match: extending loads and truncating stores. This class deals with the
4651 // sign-neutral version of those patterns.
4653 // It will be instantiated across multiple addressing-modes.
4654 multiclass ls_small_pats<Instruction LOAD, Instruction STORE,
4655 dag Base, dag Offset,
4656 dag address, ValueType sty>
4657 : ls_atomic_pats<LOAD, STORE, Base, Offset, address, GPR32, sty> {
4658 def : Pat<(!cast<SDNode>(zextload # sty) address), (LOAD Base, Offset)>;
4660 def : Pat<(!cast<SDNode>(extload # sty) address), (LOAD Base, Offset)>;
4662 // For zero-extension to 64-bits we have to tell LLVM that the whole 64-bit
4663 // register was actually set.
4664 def : Pat<(i64 (!cast<SDNode>(zextload # sty) address)),
4665 (SUBREG_TO_REG (i64 0), (LOAD Base, Offset), sub_32)>;
4667 def : Pat<(i64 (!cast<SDNode>(extload # sty) address)),
4668 (SUBREG_TO_REG (i64 0), (LOAD Base, Offset), sub_32)>;
4670 def : Pat<(!cast<SDNode>(truncstore # sty) GPR32:$Rt, address),
4671 (STORE GPR32:$Rt, Base, Offset)>;
4673 // For truncating store from 64-bits, we have to manually tell LLVM to
4674 // ignore the high bits of the x register.
4675 def : Pat<(!cast<SDNode>(truncstore # sty) GPR64:$Rt, address),
4676 (STORE (EXTRACT_SUBREG GPR64:$Rt, sub_32), Base, Offset)>;
4679 // Next come patterns for sign-extending loads.
4680 multiclass load_signed_pats<string T, string U, dag Base, dag Offset,
4681 dag address, ValueType sty> {
4682 def : Pat<(i32 (!cast<SDNode>("sextload" # sty) address)),
4683 (!cast<Instruction>("LDRS" # T # "w" # U) Base, Offset)>;
4685 def : Pat<(i64 (!cast<SDNode>("sextload" # sty) address)),
4686 (!cast<Instruction>("LDRS" # T # "x" # U) Base, Offset)>;
4690 // and finally "natural-width" loads and stores come next.
4691 multiclass ls_neutral_pats<Instruction LOAD, Instruction STORE, dag Base,
4692 dag Offset, dag address, RegisterClass TPR,
4694 def : Pat<(sty (load address)), (LOAD Base, Offset)>;
4695 def : Pat<(store (sty TPR:$Rt), address), (STORE TPR:$Rt, Base, Offset)>;
4698 // Integer operations also get atomic instructions to select for.
4699 multiclass ls_int_neutral_pats<Instruction LOAD, Instruction STORE, dag Base,
4700 dag Offset, dag address, RegisterClass TPR,
4702 : ls_neutral_pats<LOAD, STORE, Base, Offset, address, TPR, sty>,
4703 ls_atomic_pats<LOAD, STORE, Base, Offset, address, TPR, sty>;
4705 //===------------------------------
4706 // 2.2. Addressing-mode instantiations
4707 //===------------------------------
4709 multiclass uimm12_pats<dag address, dag Base, dag Offset> {
4710 defm : ls_small_pats<LS8_LDR, LS8_STR, Base,
4711 !foreach(decls.pattern, Offset,
4712 !subst(OFFSET, byte_uimm12, decls.pattern)),
4713 !foreach(decls.pattern, address,
4714 !subst(OFFSET, byte_uimm12,
4715 !subst(ALIGN, any_align, decls.pattern))),
4717 defm : ls_small_pats<LS16_LDR, LS16_STR, Base,
4718 !foreach(decls.pattern, Offset,
4719 !subst(OFFSET, hword_uimm12, decls.pattern)),
4720 !foreach(decls.pattern, address,
4721 !subst(OFFSET, hword_uimm12,
4722 !subst(ALIGN, min_align2, decls.pattern))),
4724 defm : ls_small_pats<LS32_LDR, LS32_STR, Base,
4725 !foreach(decls.pattern, Offset,
4726 !subst(OFFSET, word_uimm12, decls.pattern)),
4727 !foreach(decls.pattern, address,
4728 !subst(OFFSET, word_uimm12,
4729 !subst(ALIGN, min_align4, decls.pattern))),
4732 defm : ls_int_neutral_pats<LS32_LDR, LS32_STR, Base,
4733 !foreach(decls.pattern, Offset,
4734 !subst(OFFSET, word_uimm12, decls.pattern)),
4735 !foreach(decls.pattern, address,
4736 !subst(OFFSET, word_uimm12,
4737 !subst(ALIGN, min_align4, decls.pattern))),
4740 defm : ls_int_neutral_pats<LS64_LDR, LS64_STR, Base,
4741 !foreach(decls.pattern, Offset,
4742 !subst(OFFSET, dword_uimm12, decls.pattern)),
4743 !foreach(decls.pattern, address,
4744 !subst(OFFSET, dword_uimm12,
4745 !subst(ALIGN, min_align8, decls.pattern))),
4748 defm : ls_neutral_pats<LSFP16_LDR, LSFP16_STR, Base,
4749 !foreach(decls.pattern, Offset,
4750 !subst(OFFSET, hword_uimm12, decls.pattern)),
4751 !foreach(decls.pattern, address,
4752 !subst(OFFSET, hword_uimm12,
4753 !subst(ALIGN, min_align2, decls.pattern))),
4756 defm : ls_neutral_pats<LSFP32_LDR, LSFP32_STR, Base,
4757 !foreach(decls.pattern, Offset,
4758 !subst(OFFSET, word_uimm12, decls.pattern)),
4759 !foreach(decls.pattern, address,
4760 !subst(OFFSET, word_uimm12,
4761 !subst(ALIGN, min_align4, decls.pattern))),
4764 defm : ls_neutral_pats<LSFP64_LDR, LSFP64_STR, Base,
4765 !foreach(decls.pattern, Offset,
4766 !subst(OFFSET, dword_uimm12, decls.pattern)),
4767 !foreach(decls.pattern, address,
4768 !subst(OFFSET, dword_uimm12,
4769 !subst(ALIGN, min_align8, decls.pattern))),
4772 defm : ls_neutral_pats<LSFP128_LDR, LSFP128_STR, Base,
4773 !foreach(decls.pattern, Offset,
4774 !subst(OFFSET, qword_uimm12, decls.pattern)),
4775 !foreach(decls.pattern, address,
4776 !subst(OFFSET, qword_uimm12,
4777 !subst(ALIGN, min_align16, decls.pattern))),
4780 defm : load_signed_pats<"B", "", Base,
4781 !foreach(decls.pattern, Offset,
4782 !subst(OFFSET, byte_uimm12, decls.pattern)),
4783 !foreach(decls.pattern, address,
4784 !subst(OFFSET, byte_uimm12,
4785 !subst(ALIGN, any_align, decls.pattern))),
4788 defm : load_signed_pats<"H", "", Base,
4789 !foreach(decls.pattern, Offset,
4790 !subst(OFFSET, hword_uimm12, decls.pattern)),
4791 !foreach(decls.pattern, address,
4792 !subst(OFFSET, hword_uimm12,
4793 !subst(ALIGN, min_align2, decls.pattern))),
4796 def : Pat<(sextloadi32 !foreach(decls.pattern, address,
4797 !subst(OFFSET, word_uimm12,
4798 !subst(ALIGN, min_align4, decls.pattern)))),
4799 (LDRSWx Base, !foreach(decls.pattern, Offset,
4800 !subst(OFFSET, word_uimm12, decls.pattern)))>;
4803 // Straightforward patterns of last resort: a pointer with or without an
4804 // appropriate offset.
4805 defm : uimm12_pats<(i64 GPR64xsp:$Rn), (i64 GPR64xsp:$Rn), (i64 0)>;
4806 defm : uimm12_pats<(add GPR64xsp:$Rn, OFFSET:$UImm12),
4807 (i64 GPR64xsp:$Rn), (i64 OFFSET:$UImm12)>;
4809 // The offset could be hidden behind an "or", of course:
4810 defm : uimm12_pats<(add_like_or GPR64xsp:$Rn, OFFSET:$UImm12),
4811 (i64 GPR64xsp:$Rn), (i64 OFFSET:$UImm12)>;
4813 // Global addresses under the small-absolute model should use these
4814 // instructions. There are ELF relocations specifically for it.
4815 defm : uimm12_pats<(A64WrapperSmall tglobaladdr:$Hi, tglobaladdr:$Lo12, ALIGN),
4816 (ADRPxi tglobaladdr:$Hi), (i64 tglobaladdr:$Lo12)>;
4818 defm : uimm12_pats<(A64WrapperSmall tglobaltlsaddr:$Hi, tglobaltlsaddr:$Lo12,
4820 (ADRPxi tglobaltlsaddr:$Hi), (i64 tglobaltlsaddr:$Lo12)>;
4822 // External symbols that make it this far should also get standard relocations.
4823 defm : uimm12_pats<(A64WrapperSmall texternalsym:$Hi, texternalsym:$Lo12,
4825 (ADRPxi texternalsym:$Hi), (i64 texternalsym:$Lo12)>;
4828 // We also want to use uimm12 instructions for local variables at the moment.
4829 def tframeindex_XFORM : SDNodeXForm<frameindex, [{
4830 int FI = cast<FrameIndexSDNode>(N)->getIndex();
4831 return CurDAG->getTargetFrameIndex(FI, MVT::i64);
4834 defm : uimm12_pats<(i64 frameindex:$Rn),
4835 (tframeindex_XFORM tframeindex:$Rn), (i64 0)>;
4837 // These can be much simpler than uimm12 because we don't to change the operand
4838 // type (e.g. LDURB and LDURH take the same operands).
4839 multiclass simm9_pats<dag address, dag Base, dag Offset> {
4840 defm : ls_small_pats<LS8_LDUR, LS8_STUR, Base, Offset, address, i8>;
4841 defm : ls_small_pats<LS16_LDUR, LS16_STUR, Base, Offset, address, i16>;
4843 defm : ls_int_neutral_pats<LS32_LDUR, LS32_STUR, Base, Offset, address,
4845 defm : ls_int_neutral_pats<LS64_LDUR, LS64_STUR, Base, Offset, address,
4848 defm : ls_neutral_pats<LSFP16_LDUR, LSFP16_STUR, Base, Offset, address,
4850 defm : ls_neutral_pats<LSFP32_LDUR, LSFP32_STUR, Base, Offset, address,
4852 defm : ls_neutral_pats<LSFP64_LDUR, LSFP64_STUR, Base, Offset, address,
4854 defm : ls_neutral_pats<LSFP128_LDUR, LSFP128_STUR, Base, Offset, address,
4857 def : Pat<(i64 (zextloadi32 address)),
4858 (SUBREG_TO_REG (i64 0), (LS32_LDUR Base, Offset), sub_32)>;
4860 def : Pat<(truncstorei32 GPR64:$Rt, address),
4861 (LS32_STUR (EXTRACT_SUBREG GPR64:$Rt, sub_32), Base, Offset)>;
4863 defm : load_signed_pats<"B", "_U", Base, Offset, address, i8>;
4864 defm : load_signed_pats<"H", "_U", Base, Offset, address, i16>;
4865 def : Pat<(sextloadi32 address), (LDURSWx Base, Offset)>;
4868 defm : simm9_pats<(add GPR64xsp:$Rn, simm9:$SImm9),
4869 (i64 GPR64xsp:$Rn), (SDXF_simm9 simm9:$SImm9)>;
4871 defm : simm9_pats<(add_like_or GPR64xsp:$Rn, simm9:$SImm9),
4872 (i64 GPR64xsp:$Rn), (SDXF_simm9 simm9:$SImm9)>;
4875 //===------------------------------
4876 // 3. Register offset patterns
4877 //===------------------------------
4879 // Atomic patterns can be shared between integer operations of all sizes, a
4880 // quick multiclass here allows reuse.
4881 multiclass ro_atomic_pats<Instruction LOAD, Instruction STORE, dag Base,
4882 dag Offset, dag Extend, dag address,
4883 RegisterClass TPR, ValueType sty> {
4884 def : Pat<(!cast<PatFrag>("atomic_load_simple_" # sty) address),
4885 (LOAD Base, Offset, Extend)>;
4887 def : Pat<(!cast<PatFrag>("atomic_store_simple_" # sty) address, TPR:$Rt),
4888 (STORE TPR:$Rt, Base, Offset, Extend)>;
4891 // The register offset instructions take three operands giving the instruction,
4892 // and have an annoying split between instructions where Rm is 32-bit and
4893 // 64-bit. So we need a special hierarchy to describe them. Other than that the
4894 // same operations should be supported as for simm9 and uimm12 addressing.
4896 multiclass ro_small_pats<Instruction LOAD, Instruction STORE,
4897 dag Base, dag Offset, dag Extend,
4898 dag address, ValueType sty>
4899 : ro_atomic_pats<LOAD, STORE, Base, Offset, Extend, address, GPR32, sty> {
4900 def : Pat<(!cast<SDNode>(zextload # sty) address),
4901 (LOAD Base, Offset, Extend)>;
4903 def : Pat<(!cast<SDNode>(extload # sty) address),
4904 (LOAD Base, Offset, Extend)>;
4906 // For zero-extension to 64-bits we have to tell LLVM that the whole 64-bit
4907 // register was actually set.
4908 def : Pat<(i64 (!cast<SDNode>(zextload # sty) address)),
4909 (SUBREG_TO_REG (i64 0), (LOAD Base, Offset, Extend), sub_32)>;
4911 def : Pat<(i64 (!cast<SDNode>(extload # sty) address)),
4912 (SUBREG_TO_REG (i64 0), (LOAD Base, Offset, Extend), sub_32)>;
4914 def : Pat<(!cast<SDNode>(truncstore # sty) GPR32:$Rt, address),
4915 (STORE GPR32:$Rt, Base, Offset, Extend)>;
4917 // For truncating store from 64-bits, we have to manually tell LLVM to
4918 // ignore the high bits of the x register.
4919 def : Pat<(!cast<SDNode>(truncstore # sty) GPR64:$Rt, address),
4920 (STORE (EXTRACT_SUBREG GPR64:$Rt, sub_32), Base, Offset, Extend)>;
4924 // Next come patterns for sign-extending loads.
4925 multiclass ro_signed_pats<string T, string Rm, dag Base, dag Offset, dag Extend,
4926 dag address, ValueType sty> {
4927 def : Pat<(i32 (!cast<SDNode>("sextload" # sty) address)),
4928 (!cast<Instruction>("LDRS" # T # "w_" # Rm # "_RegOffset")
4929 Base, Offset, Extend)>;
4931 def : Pat<(i64 (!cast<SDNode>("sextload" # sty) address)),
4932 (!cast<Instruction>("LDRS" # T # "x_" # Rm # "_RegOffset")
4933 Base, Offset, Extend)>;
4936 // and finally "natural-width" loads and stores come next.
4937 multiclass ro_neutral_pats<Instruction LOAD, Instruction STORE,
4938 dag Base, dag Offset, dag Extend, dag address,
4939 RegisterClass TPR, ValueType sty> {
4940 def : Pat<(sty (load address)), (LOAD Base, Offset, Extend)>;
4941 def : Pat<(store (sty TPR:$Rt), address),
4942 (STORE TPR:$Rt, Base, Offset, Extend)>;
4945 multiclass ro_int_neutral_pats<Instruction LOAD, Instruction STORE,
4946 dag Base, dag Offset, dag Extend, dag address,
4947 RegisterClass TPR, ValueType sty>
4948 : ro_neutral_pats<LOAD, STORE, Base, Offset, Extend, address, TPR, sty>,
4949 ro_atomic_pats<LOAD, STORE, Base, Offset, Extend, address, TPR, sty>;
4951 multiclass regoff_pats<string Rm, dag address, dag Base, dag Offset,
4953 defm : ro_small_pats<!cast<Instruction>("LS8_" # Rm # "_RegOffset_LDR"),
4954 !cast<Instruction>("LS8_" # Rm # "_RegOffset_STR"),
4955 Base, Offset, Extend,
4956 !foreach(decls.pattern, address,
4957 !subst(SHIFT, imm_eq0, decls.pattern)),
4959 defm : ro_small_pats<!cast<Instruction>("LS16_" # Rm # "_RegOffset_LDR"),
4960 !cast<Instruction>("LS16_" # Rm # "_RegOffset_STR"),
4961 Base, Offset, Extend,
4962 !foreach(decls.pattern, address,
4963 !subst(SHIFT, imm_eq1, decls.pattern)),
4965 defm : ro_small_pats<!cast<Instruction>("LS32_" # Rm # "_RegOffset_LDR"),
4966 !cast<Instruction>("LS32_" # Rm # "_RegOffset_STR"),
4967 Base, Offset, Extend,
4968 !foreach(decls.pattern, address,
4969 !subst(SHIFT, imm_eq2, decls.pattern)),
4972 defm : ro_int_neutral_pats<
4973 !cast<Instruction>("LS32_" # Rm # "_RegOffset_LDR"),
4974 !cast<Instruction>("LS32_" # Rm # "_RegOffset_STR"),
4975 Base, Offset, Extend,
4976 !foreach(decls.pattern, address,
4977 !subst(SHIFT, imm_eq2, decls.pattern)),
4980 defm : ro_int_neutral_pats<
4981 !cast<Instruction>("LS64_" # Rm # "_RegOffset_LDR"),
4982 !cast<Instruction>("LS64_" # Rm # "_RegOffset_STR"),
4983 Base, Offset, Extend,
4984 !foreach(decls.pattern, address,
4985 !subst(SHIFT, imm_eq3, decls.pattern)),
4988 defm : ro_neutral_pats<!cast<Instruction>("LSFP16_" # Rm # "_RegOffset_LDR"),
4989 !cast<Instruction>("LSFP16_" # Rm # "_RegOffset_STR"),
4990 Base, Offset, Extend,
4991 !foreach(decls.pattern, address,
4992 !subst(SHIFT, imm_eq1, decls.pattern)),
4995 defm : ro_neutral_pats<!cast<Instruction>("LSFP32_" # Rm # "_RegOffset_LDR"),
4996 !cast<Instruction>("LSFP32_" # Rm # "_RegOffset_STR"),
4997 Base, Offset, Extend,
4998 !foreach(decls.pattern, address,
4999 !subst(SHIFT, imm_eq2, decls.pattern)),
5002 defm : ro_neutral_pats<!cast<Instruction>("LSFP64_" # Rm # "_RegOffset_LDR"),
5003 !cast<Instruction>("LSFP64_" # Rm # "_RegOffset_STR"),
5004 Base, Offset, Extend,
5005 !foreach(decls.pattern, address,
5006 !subst(SHIFT, imm_eq3, decls.pattern)),
5009 defm : ro_neutral_pats<!cast<Instruction>("LSFP128_" # Rm # "_RegOffset_LDR"),
5010 !cast<Instruction>("LSFP128_" # Rm # "_RegOffset_STR"),
5011 Base, Offset, Extend,
5012 !foreach(decls.pattern, address,
5013 !subst(SHIFT, imm_eq4, decls.pattern)),
5016 defm : ro_signed_pats<"B", Rm, Base, Offset, Extend,
5017 !foreach(decls.pattern, address,
5018 !subst(SHIFT, imm_eq0, decls.pattern)),
5021 defm : ro_signed_pats<"H", Rm, Base, Offset, Extend,
5022 !foreach(decls.pattern, address,
5023 !subst(SHIFT, imm_eq1, decls.pattern)),
5026 def : Pat<(sextloadi32 !foreach(decls.pattern, address,
5027 !subst(SHIFT, imm_eq2, decls.pattern))),
5028 (!cast<Instruction>("LDRSWx_" # Rm # "_RegOffset")
5029 Base, Offset, Extend)>;
5033 // Finally we're in a position to tell LLVM exactly what addresses are reachable
5034 // using register-offset instructions. Essentially a base plus a possibly
5035 // extended, possibly shifted (by access size) offset.
5037 defm : regoff_pats<"Wm", (add GPR64xsp:$Rn, (sext GPR32:$Rm)),
5038 (i64 GPR64xsp:$Rn), (i32 GPR32:$Rm), (i64 6)>;
5040 defm : regoff_pats<"Wm", (add GPR64xsp:$Rn, (shl (sext GPR32:$Rm), SHIFT)),
5041 (i64 GPR64xsp:$Rn), (i32 GPR32:$Rm), (i64 7)>;
5043 defm : regoff_pats<"Wm", (add GPR64xsp:$Rn, (zext GPR32:$Rm)),
5044 (i64 GPR64xsp:$Rn), (i32 GPR32:$Rm), (i64 2)>;
5046 defm : regoff_pats<"Wm", (add GPR64xsp:$Rn, (shl (zext GPR32:$Rm), SHIFT)),
5047 (i64 GPR64xsp:$Rn), (i32 GPR32:$Rm), (i64 3)>;
5049 defm : regoff_pats<"Xm", (add GPR64xsp:$Rn, GPR64:$Rm),
5050 (i64 GPR64xsp:$Rn), (i64 GPR64:$Rm), (i64 2)>;
5052 defm : regoff_pats<"Xm", (add GPR64xsp:$Rn, (shl GPR64:$Rm, SHIFT)),
5053 (i64 GPR64xsp:$Rn), (i64 GPR64:$Rm), (i64 3)>;