xref: /freebsd/contrib/llvm-project/llvm/lib/Target/ARM/ARMScheduleM55.td (revision 5ca8e32633c4ffbbcd6762e5888b6a4ba0708c6c)
1//==- ARMScheduleM55.td - Arm Cortex-M55 Scheduling Definitions -*- tablegen -*-=//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This file defines the scheduling model for the Arm Cortex-M55 processors.
10//
11//===----------------------------------------------------------------------===//
12
13// ===---------------------------------------------------------------------===//
14// Cortex-M55 is a lot like the M4/M33 in terms of scheduling. It technically
15// has an extra pipeline stage but that is unimportant for scheduling, just
16// starting our model a stage later. The main points of interest over an
17// Cortex-M4 are MVE instructions and the ability to dual issue thumb1
18// instructions.
19//
20//
21// MVE
22//
23// The EPU pipelines now include both MVE and FP instructions. It has four
24// pipelines across 4 stages (E1-E4). These pipelines are "control",
25// "load/store", "integer" and "float/mul". We start the schedule at E2 to line
26// up with the rest of the pipeline we model, and take the latency as the time
27// between reading registers (almost always in E2) and register write (or
28// forward, if it allows it). This mean that a lot of instructions (including
29// loads) actually take 1 cycle (amazingly).
30//
31// Each MVE instruction needs to take 2 beats, each performing 64bits of the
32// 128bit vector operation. So long as the beats are to different pipelines,
33// the execution of the first-beat-of-the-second-instruction can overlap with
34// the second-beat-of-the-first. For example a sequence of VLDR;VADD;VMUL;VSTR
35// can look like this is a pipeline:
36//          1    2    3    4    5
37// LD/ST  : VLDR VLDR      VSTR VSTR
38// INTEGER:      VADD VADD
39// FP/MUL :           VMUL VMUL
40//
41// But a sequence of VLDR;VLDRB;VADD;VSTR because the loads cannot overlap,
42// looks like:
43//          1     2     3     4     5    6
44// LD/ST  : VLDR  VLDR  VLDRB VLDRB VSTR VSTR
45// INTEGER:                   VADD  VADD
46//
47// For this schedule, we currently model latencies and pipelines well for each
48// instruction. MVE instruction take two beats, modelled using
49// ResourceCycles=[2].
50//
51//
52// Dual Issue
53//
54// Cortex-M55 can dual issue two 16-bit T1 instructions providing one is one of
55// NOPs, ITs, Brs, ADDri/SUBri, UXTB/H, SXTB/H and MOVri's. NOPs and IT's are
56// not relevant (they will not appear when scheduling), Brs are only at the end
57// of the block. The others are more useful, and where the problems arise.
58//
59// The first problem comes from the fact that we will only be seeing Thumb2
60// instructions at the point in the pipeline where we do the scheduling. The
61// Thumb2SizeReductionPass has not been run yet. Especially pre-ra scheduling
62// (where the scheduler has the most freedom) we can only really guess at which
63// instructions will become thumb1 instructions. We are quite optimistic, and
64// may get some things wrong as a result.
65//
66// The other problem is one of telling llvm what to do exactly. The way we
67// attempt to meld this is:
68//  Set IssueWidth to 2 to allow 2 instructions per cycle.
69//  All instructions we cannot dual issue are "SingleIssue=1" (MVE/FP and T2
70//    instructions)
71//  We guess at another set of instructions that will become T1 instruction.
72//    These become the primary instruction in a dual issue pair (the normal
73//    one). These use normal resources and latencies, but set SingleIssue = 0.
74//  We guess at another set of instructions that will be shrank down into T1 DI
75//    instructions (add, sub, mov's, etc), which become the secondary. These
76//    don't use a resource, and set SingleIssue = 0.
77//
78// So our guessing is a bit rough. It may be possible to improve this by moving
79// T2SizeReduction pass earlier in the pipeline, for example, so that at least
80// Post-RA scheduling sees what is T1/T2. It may also be possible to write a
81// custom instruction matcher for more accurately guess at T1 instructions.
82
83
84def CortexM55Model : SchedMachineModel {
85  let MicroOpBufferSize = 0;      // Explicitly set to zero since M55 is in-order.
86  let IssueWidth = 2;             // There is some dual-issue support in M55.
87  let MispredictPenalty = 3;      // Default is 10
88  let LoadLatency = 4;            // Default is 4
89  let PostRAScheduler = 1;
90  let FullInstRWOverlapCheck = 1;
91
92  let CompleteModel = 0;
93  let UnsupportedFeatures = [IsARM, HasNEON, HasDotProd, HasMatMulInt8, HasZCZ,
94                             IsNotMClass, HasV8, HasV8_3a, HasTrustZone, HasDFB,
95                             IsWindows];
96}
97
98
99let SchedModel = CortexM55Model in {
100
101//===----------------------------------------------------------------------===//
102// Define each kind of processor resource and number available.
103
104// Modeling each pipeline as a ProcResource using the BufferSize = 0 since
105// M55 is in-order.
106def M55UnitALU : ProcResource<1> { let BufferSize = 0; } // Int ALU
107def M55UnitVecALU : ProcResource<1> { let BufferSize = 0; } // MVE integer pipe
108def M55UnitVecFPALU : ProcResource<1> { let BufferSize = 0; } // MVE float pipe
109def M55UnitLoadStore : ProcResource<1> { let BufferSize = 0; } // MVE load/store pipe
110def M55UnitVecSys : ProcResource<1> { let BufferSize = 0; } // MVE control/sys pipe
111
112// Some VMOV's can go down either pipeline. FIXME: This M55Write2IntFPE2 is
113// intended to model the VMOV taking either Int or FP for 2 cycles. It is not
114// clear if the llvm scheduler is using it like we want though.
115def M55UnitVecIntFP: ProcResGroup<[M55UnitVecALU, M55UnitVecFPALU]>;
116
117
118//===----------------------------------------------------------------------===//
119// Subtarget-specific SchedWrite types which both map the ProcResources and
120// set the latency.
121
122//=====//
123// ALU //
124//=====//
125
126// Generic writes for Flags, GRPs and other extra operands (eg post-inc, vadc flags, vaddlv etc)
127def M55WriteLat0  : SchedWriteRes<[]>  { let Latency = 0; let NumMicroOps = 0; }
128def M55WriteLat1  : SchedWriteRes<[]>  { let Latency = 1; let NumMicroOps = 0; }
129def M55WriteLat2  : SchedWriteRes<[]>  { let Latency = 2; let NumMicroOps = 0; }
130
131// DX instructions are ALU instructions that take a single cycle. The
132// instructions that may be shrank to T1 (and can be dual issued) are
133// SingleIssue = 0. The others are SingleIssue = 1.
134let SingleIssue = 0, Latency = 1 in {
135    def : WriteRes<WriteALU, [M55UnitALU]>;
136    def : WriteRes<WriteCMP, [M55UnitALU]>;
137    def : WriteRes<WriteBr, [M55UnitALU]>;
138    def : WriteRes<WriteBrL, [M55UnitALU]>;
139    def : WriteRes<WriteBrTbl, [M55UnitALU]>;
140    def : WriteRes<WriteST, [M55UnitALU]>;
141    def M55WriteDX_DI : SchedWriteRes<[M55UnitALU]>;
142}
143let SingleIssue = 1, Latency = 1 in {
144    def : WriteRes<WritePreLd, [M55UnitALU]>;
145    def M55WriteDX_SI : SchedWriteRes<[M55UnitALU]>;
146}
147
148def : InstRW<[M55WriteDX_SI], (instregex "t2BF[CI]", "t2CPS", "t2DBG",
149          "t2MRS", "t2MSR", "t2SEL", "t2SG", "t2TT")>;
150def : InstRW<[M55WriteDX_SI], (instregex "t2SUBS_PC_LR", "COPY")>;
151def : InstRW<[M55WriteDX_SI], (instregex "t2CS(EL|INC|INV|NEG)")>;
152// Thumb 2 instructions that could be reduced to a thumb 1 instruction and can
153// be dual issued with one of the above. This list is optimistic.
154def : InstRW<[M55WriteDX_DI], (instregex "t2ADDC?rr$", "t2ADDrr$",
155           "t2ADDSrr$", "t2ANDrr$", "t2ASRr[ir]$", "t2BICrr$", "t2CMNzrr$",
156           "t2CMPr[ir]$", "t2EORrr$", "t2LSLr[ir]$", "t2LSRr[ir]$", "t2MVNr$",
157           "t2ORRrr$", "t2REV(16|SH)?$", "t2RORrr$", "t2RSBr[ir]$", "t2RSBSri$",
158           "t2SBCrr$", "t2SUBS?rr$", "t2TEQrr$", "t2TSTrr$", "t2STRi12$",
159           "t2STRs$", "t2STRBi12$", "t2STRBs$", "t2STRHi12$", "t2STRHs$",
160           "t2STR_POST$", "t2STMIA$", "t2STMIA_UPD$", "t2STMDB$", "t2STMDB_UPD$")>;
161def : InstRW<[M55WriteDX_DI], (instregex "t2SETPAN$", "tADC$", "tADDhirr$",
162           "tADDrSP$", "tADDrSPi$", "tADDrr$", "tADDspi$", "tADDspr$", "tADR$",
163           "tAND$", "tASRri$", "tASRrr$", "tBIC$", "tBKPT$", "tCBNZ$", "tCBZ$",
164           "tCMNz$", "tCMPhir$", "tCMPi8$", "tCMPr$", "tCPS$", "tEOR$", "tHINT$",
165           "tHLT$", "tLSLri$", "tLSLrr$", "tLSRri$", "tLSRrr$", "tMOVSr$",
166           "tMUL$", "tMVN$", "tORR$", "tPICADD$", "tPOP$", "tPUSH$", "tREV$",
167           "tREV16$", "tREVSH$", "tROR$", "tRSB$", "tSBC$", "tSETEND$",
168           "tSTMIA_UPD$", "tSTRBi$", "tSTRBr$", "tSTRHi$", "tSTRHr$", "tSTRi$",
169           "tSTRr$", "tSTRspi$", "tSUBrr$", "tSUBspi$", "tSVC$", "tTRAP$",
170           "tTST$", "tUDF$")>;
171def : InstRW<[M55WriteDX_DI], (instregex "tB$", "tBLXNSr$", "tBLXr$", "tBX$",
172           "tBXNS$", "tBcc$")>;
173
174
175// CX instructions take 2 (or more) cycles. Again T1 instructions may be dual
176// issues (SingleIssue = 0)
177let SingleIssue = 0, Latency = 2 in {
178    def : WriteRes<WriteLd, [M55UnitALU]>;
179    def M55WriteCX_DI  : SchedWriteRes<[M55UnitALU]>;
180}
181let SingleIssue = 1, Latency = 2 in {
182    def : WriteRes<WriteALUsi, [M55UnitALU]>;
183    def : WriteRes<WriteALUsr, [M55UnitALU]>;
184    def : WriteRes<WriteALUSsr, [M55UnitALU]>;
185    def : WriteRes<WriteCMPsi, [M55UnitALU]>;
186    def : WriteRes<WriteCMPsr, [M55UnitALU]>;
187    def : WriteRes<WriteDIV, [M55UnitALU]>;
188    def M55WriteCX_SI  : SchedWriteRes<[M55UnitALU]>;
189}
190
191def : SchedAlias<WriteMUL16, M55WriteCX_SI>;
192def : SchedAlias<WriteMUL32, M55WriteCX_SI>;
193def : SchedAlias<WriteMUL64Lo, M55WriteCX_SI>;
194def : WriteRes<WriteMUL64Hi, []> { let Latency = 2; }
195def : SchedAlias<WriteMAC16, M55WriteCX_SI>;
196def : SchedAlias<WriteMAC32, M55WriteCX_SI>;
197def : SchedAlias<WriteMAC64Lo, M55WriteCX_SI>;
198def : WriteRes<WriteMAC64Hi, []> { let Latency = 2; }
199
200def : InstRW<[M55WriteCX_SI], (instregex "t2CDP", "t2CLREX", "t2[DI][MS]B",
201           "t2MCR", "t2MOVSs[ir]", "t2MRC", "t2MUL", "t2STC")>;
202def : InstRW<[M55WriteCX_SI], (instregex "t2Q", "t2[SU](ADD|ASX|BFX|DIV)",
203           "t2[SU]H(ADD|ASX|SUB|SAX)", "t2SM[LM]", "t2S(SAT|SUB|SAX)", "t2UQ",
204           "t2USA", "t2USUB", "t2UXTA[BH]")>;
205def : InstRW<[M55WriteCX_SI], (instregex "t2LD[AC]", "t2STL", "t2STRD")>;
206def : InstRW<[M55WriteCX_SI], (instregex "MVE_[SU]Q?R?SH[LR]$")>;
207def : InstRW<[M55WriteCX_SI, M55WriteLat2], (instregex "MVE_ASRL", "MVE_LSLL",
208            "MVE_LSRL", "MVE_[SU]Q?R?SH[LR]L")>;
209// This may be higher in practice, but that likely doesn't make a difference
210// for scheduling
211def : InstRW<[M55WriteCX_SI], (instregex "t2CLRM")>;
212
213def : InstRW<[M55WriteCX_DI], (instregex "t2LDR[BH]?i12$", "t2LDRS?[BH]?s$",
214           "t2LDM")>;
215def : InstRW<[M55WriteCX_DI], (instregex "tLDM", "tLDRBi$", "tLDRBr$",
216           "tLDRHi$", "tLDRHr$", "tLDRSB$", "tLDRSH$", "tLDRi$", "tLDRpci$",
217           "tLDRr$", "tLDRspi$")>;
218
219// Dual Issue instructions
220let Latency = 1, SingleIssue = 0 in {
221    def : WriteRes<WriteNoop, []>;
222    def M55WriteDI : SchedWriteRes<[]>;
223}
224
225def : InstRW<[M55WriteDI], (instregex "tADDi[38]$", "tSUBi[38]$", "tMOVi8$",
226           "tMOVr$", "tUXT[BH]$", "tSXT[BH]$")>;
227// Thumb 2 instructions that could be reduced to a dual issuable Thumb 1
228// instruction above.
229def : InstRW<[M55WriteDI], (instregex "t2ADDS?ri$", "t2MOV[ir]$", "t2MOVi16$",
230           "t2MOVr$", "t2SUBS?ri$", "t2[US]XT[BH]$")>;
231def : InstRW<[M55WriteDI], (instregex "t2IT", "IT")>;
232
233
234def : InstRW<[M55WriteLat0], (instregex "t2LoopDec")>;
235
236// Forwarding
237
238// No forwarding in the ALU normally
239def : ReadAdvance<ReadALU, 0>;
240def : ReadAdvance<ReadALUsr, 0>;
241def : ReadAdvance<ReadMUL, 0>;
242def : ReadAdvance<ReadMAC, 0>;
243
244//=============//
245// MVE and VFP //
246//=============//
247
248// The Writes that take ResourceCycles=[2] are MVE instruction, the others VFP.
249
250let SingleIssue = 1, Latency = 1 in {
251  def M55WriteLSE2 : SchedWriteRes<[M55UnitLoadStore]>;
252  def M55WriteIntE2 : SchedWriteRes<[M55UnitVecALU]>;
253  def M55WriteFloatE2 : SchedWriteRes<[M55UnitVecFPALU]>;
254  def M55WriteSysE2 : SchedWriteRes<[M55UnitVecSys]>;
255
256  def M55Write2LSE2 : SchedWriteRes<[M55UnitLoadStore]> { let ResourceCycles=[2]; }
257  def M55Write2IntE2 : SchedWriteRes<[M55UnitVecALU]> { let ResourceCycles=[2]; }
258  def M55Write2FloatE2 : SchedWriteRes<[M55UnitVecFPALU]> { let ResourceCycles=[2]; }
259  def M55Write2IntFPE2 : SchedWriteRes<[M55UnitVecIntFP]> { let ResourceCycles=[2]; }
260}
261
262let SingleIssue = 1, Latency = 2 in {
263  def M55WriteLSE3 : SchedWriteRes<[M55UnitLoadStore]>;
264  def M55WriteIntE3 : SchedWriteRes<[M55UnitVecALU]>;
265  def M55WriteFloatE3 : SchedWriteRes<[M55UnitVecFPALU]>;
266
267  def M55Write2LSE3 : SchedWriteRes<[M55UnitLoadStore]> { let ResourceCycles=[2]; }
268  def M55Write2IntE3 : SchedWriteRes<[M55UnitVecALU]> { let ResourceCycles=[2]; }
269  def M55Write2FloatE3 : SchedWriteRes<[M55UnitVecFPALU]> { let ResourceCycles=[2]; }
270}
271
272let SingleIssue = 1, Latency = 3 in {
273  def M55Write2IntE3Plus1 : SchedWriteRes<[M55UnitVecALU]> { let ResourceCycles=[2]; }
274
275  // Same as M55Write2IntE3/M55Write2FloatE3 above, but longer latency and no forwarding into stores
276  def M55Write2IntE4NoFwd : SchedWriteRes<[M55UnitVecALU]> { let ResourceCycles=[2]; }
277  def M55Write2FloatE4NoFwd : SchedWriteRes<[M55UnitVecFPALU]> { let ResourceCycles=[2]; }
278}
279let SingleIssue = 1, Latency = 4 in {
280  def M55Write2IntE3Plus2 : SchedWriteRes<[M55UnitVecALU]> { let ResourceCycles=[2]; }
281  def M55WriteFloatE3Plus2 : SchedWriteRes<[M55UnitVecFPALU]>;
282}
283let SingleIssue = 1, Latency = 9 in {
284  def M55WriteFloatE3Plus7 : SchedWriteRes<[M55UnitVecFPALU]>;
285}
286let SingleIssue = 1, Latency = 15 in {
287  def M55WriteFloatE3Plus13 : SchedWriteRes<[M55UnitVecFPALU]>;
288}
289let SingleIssue = 1, Latency = 16 in {
290  def M55WriteFloatE3Plus14 : SchedWriteRes<[M55UnitVecFPALU]>;
291}
292let SingleIssue = 1, Latency = 21 in {
293  def M55WriteFloatE3Plus19 : SchedWriteRes<[M55UnitVecFPALU]>;
294}
295// VMUL (Double precision) + VADD (Double precision)
296let SingleIssue = 1, Latency = 24 in {
297  def M55WriteFloatE3Plus22 : SchedWriteRes<[M55UnitVecFPALU]>;
298}
299let SingleIssue = 1, Latency = 30 in {
300  def M55WriteFloatE3Plus28 : SchedWriteRes<[M55UnitVecFPALU]>;
301}
302let SingleIssue = 1, Latency = 36 in {
303  def M55WriteFloatE3Plus34 : SchedWriteRes<[M55UnitVecFPALU]>;
304}
305
306def M55Read0 : SchedReadAdvance<0>;
307def M55Read1 : SchedReadAdvance<1, [M55Write2LSE3, M55Write2IntE3, M55Write2FloatE3]>;
308def M55GatherQRead : SchedReadAdvance<-4>;
309
310// MVE instructions
311
312// Loads and Stores of different kinds
313
314// Normal loads
315def : InstRW<[M55Write2LSE2], (instregex "MVE_VLDR(B|H|W)(S|U)(8|16|32)$")>;
316// Pre/post inc loads
317def : InstRW<[M55WriteLat1, M55Write2LSE2], (instregex "MVE_VLDR(B|H|W)(S|U)(8|16|32)_(post|pre)$")>;
318// Gather loads
319def : InstRW<[M55Write2LSE3, M55Read0, M55GatherQRead], (instregex "MVE_VLDR(B|H|W|D)(S|U)(8|16|32|64)_rq")>;
320def : InstRW<[M55Write2LSE3, M55GatherQRead], (instregex "MVE_VLDR(B|H|W|D)(S|U)(8|16|32|64)_qi$")>;
321def : InstRW<[M55WriteLat1, M55Write2LSE3, M55GatherQRead], (instregex "MVE_VLDR(W|D)U(32|64)_qi_pre$")>;
322// Interleaving loads
323def : InstRW<[M55Write2LSE2], (instregex "MVE_VLD[24][0-3]_(8|16|32)$")>;
324// Interleaving loads with wb
325def : InstRW<[M55Write2LSE2, M55WriteLat1], (instregex "MVE_VLD[24][0-3]_(8|16|32)_wb$")>;
326
327// Normal stores
328def : InstRW<[M55Write2LSE2, M55Read1], (instregex "MVE_VSTR(B|H|W)U?(8|16|32)$")>;
329// Pre/post inc stores
330def : InstRW<[M55Write2LSE2, M55Read1], (instregex "MVE_VSTR(B|H|W)U?(8|16|32)_(post|pre)$")>;
331// Scatter stores
332def : InstRW<[M55Write2LSE2, M55Read0, M55Read0, M55GatherQRead], (instregex "MVE_VSTR(B|H|W|D)(8|16|32|64)_rq")>;
333def : InstRW<[M55Write2LSE2, M55Read0, M55GatherQRead], (instregex "MVE_VSTR(B|H|W|D)(8|16|32|64)_qi")>;
334// Interleaving stores
335def : InstRW<[M55Write2LSE2], (instregex "MVE_VST(2|4)")>;
336
337// Integer pipe operations
338
339def : InstRW<[M55Write2IntE3Plus1], (instregex "MVE_VABAV")>;
340def : InstRW<[M55Write2IntE2], (instregex "MVE_VABD(u|s)")>;
341def : InstRW<[M55Write2IntE2], (instregex "MVE_VABS(u|s)")>;
342def : InstRW<[M55Write2IntE3], (instregex "MVE_VADC")>;
343def : InstRW<[M55Write2IntE2], (instregex "MVE_VADD(_qr_)?i")>;
344def : InstRW<[M55Write2IntE2], (instregex "MVE_VAND")>;
345def : InstRW<[M55Write2IntE2], (instregex "MVE_VBIC")>;
346def : InstRW<[M55Write2IntE2], (instregex "MVE_VBRSR")>;
347def : InstRW<[M55Write2IntE2], (instregex "MVE_VCADDi")>;
348def : InstRW<[M55Write2IntE2], (instregex "MVE_VCLS")>;
349def : InstRW<[M55Write2IntE2], (instregex "MVE_VCLZ")>;
350def : InstRW<[M55Write2IntE2], (instregex "MVE_V(D|I)?W?DUP")>;
351def : InstRW<[M55Write2IntE2], (instregex "MVE_VEOR")>;
352def : InstRW<[M55Write2IntE2], (instregex "MVE_VHADD")>;
353def : InstRW<[M55Write2IntE2], (instregex "MVE_VHCADD")>;
354def : InstRW<[M55Write2IntE2], (instregex "MVE_VHSUB")>;
355def : InstRW<[M55Write2IntE2], (instregex "MVE_V(MAX|MIN)A?(s|u)")>;
356def : InstRW<[M55Write2IntE3], (instregex "MVE_V(MAX|MIN)A?V(s|u)8")>;
357def : InstRW<[M55Write2IntE3Plus1], (instregex "MVE_V(MAX|MIN)A?V(s|u)16")>;
358def : InstRW<[M55Write2IntE3Plus2], (instregex "MVE_V(MAX|MIN)A?V(s|u)32")>;
359def : InstRW<[M55Write2IntE4NoFwd], (instregex "MVE_VMOVN")>;
360def : InstRW<[M55Write2IntE2], (instregex "MVE_VMOVL")>;
361def : InstRW<[M55Write2IntE3], (instregex "MVE_VMULL[BT]p")>;
362def : InstRW<[M55Write2IntE2], (instregex "MVE_VMVN")>;
363def : InstRW<[M55Write2IntE2], (instregex "MVE_VNEG(u|s)")>;
364def : InstRW<[M55Write2IntE2], (instregex "MVE_VORN")>;
365def : InstRW<[M55Write2IntE2], (instregex "MVE_VORR")>;
366def : InstRW<[M55Write2IntE2], (instregex "MVE_VPSEL")>;
367def : InstRW<[M55Write2IntE2], (instregex "MQPRCopy")>;
368def : InstRW<[M55Write2IntE2], (instregex "MVE_VQABS")>;
369def : InstRW<[M55Write2IntE2], (instregex "MVE_VQADD")>;
370def : InstRW<[M55Write2IntE4NoFwd], (instregex "MVE_VQMOV")>;
371def : InstRW<[M55Write2IntE2], (instregex "MVE_VQNEG")>;
372def : InstRW<[M55Write2IntE2], (instregex "MVE_VSHL")>;
373def : InstRW<[M55Write2IntE3], (instregex "MVE_V[QR]SHL")>;
374def : InstRW<[M55Write2IntE3], (instregex "MVE_VQRSHL")>;
375def : InstRW<[M55Write2IntE4NoFwd], (instregex "MVE_VQ?R?SHRU?N")>;
376def : InstRW<[M55Write2IntE2], (instregex "MVE_VSHR_")>;
377def : InstRW<[M55Write2IntE3], (instregex "MVE_VRSHR_")>;
378def : InstRW<[M55Write2IntE2], (instregex "MVE_VQSUB")>;
379def : InstRW<[M55Write2IntE2], (instregex "MVE_VREV")>;
380def : InstRW<[M55Write2IntE2], (instregex "MVE_VRHADD")>;
381def : InstRW<[M55Write2IntE3], (instregex "MVE_VSBC")>;
382def : InstRW<[M55Write2IntE2], (instregex "MVE_VSLI")>;
383def : InstRW<[M55Write2IntE2], (instregex "MVE_VSRI")>;
384def : InstRW<[M55Write2IntE2], (instregex "MVE_VSUB(_qr_)?i")>;
385
386// FP/Mul pipe operations.
387
388def : InstRW<[M55Write2FloatE2], (instregex "MVE_VABDf")>;
389def : InstRW<[M55Write2FloatE2], (instregex "MVE_VABSf")>;
390def : InstRW<[M55Write2FloatE2], (instregex "MVE_VADDf")>;
391def : InstRW<[M55Write2FloatE3], (instregex "MVE_VADD_qr_f")>;
392def : InstRW<[M55Write2FloatE3, M55WriteLat1], (instregex "MVE_VADDLV")>;
393def : InstRW<[M55Write2FloatE3], (instregex "MVE_VADDV")>;
394def : InstRW<[M55Write2FloatE2], (instregex "MVE_VCADDf")>;
395def : InstRW<[M55Write2FloatE3], (instregex "MVE_VCMLA")>;
396def : InstRW<[M55Write2FloatE3], (instregex "MVE_VCMUL")>;
397def : InstRW<[M55Write2FloatE2], (instregex "MVE_VCMP(i|s|u)", "MVE_VPTv(4|8|16)(i|s|u)")>;
398def : InstRW<[M55Write2FloatE2], (instregex "MVE_VCMPf", "MVE_VPTv(4|8)f")>;
399def : InstRW<[M55Write2FloatE3], (instregex "MVE_VCVTf16(u|s)16")>;
400def : InstRW<[M55Write2FloatE3], (instregex "MVE_VCVTf32(u|s)32")>;
401def : InstRW<[M55Write2FloatE3], (instregex "MVE_VCVT(u|s)16f16")>;
402def : InstRW<[M55Write2FloatE3], (instregex "MVE_VCVT(u|s)32f32")>;
403def : InstRW<[M55Write2FloatE4NoFwd], (instregex "MVE_VCVTf16f32")>;
404def : InstRW<[M55Write2FloatE3], (instregex "MVE_VCVTf32f16")>;
405def : InstRW<[M55Write2FloatE3], (instregex "MVE_VFM(A|S)")>;
406def : InstRW<[M55Write2FloatE2], (instregex "MVE_V(MIN|MAX)NM")>;
407def : InstRW<[M55Write2FloatE2], (instregex "MVE_VMOV_from_lane")>;
408def : InstRW<[M55Write2FloatE2], (instregex "MVE_VMOV_rr_q")>;
409def : InstRW<[M55Write2FloatE3], (instregex "MVE_VMOVi")>;
410def : InstRW<[M55Write2FloatE3], (instregex "MVE_VMUL(_qr_)?[if]")>;
411def : InstRW<[M55Write2FloatE3], (instregex "MVE_VQ?R?D?MULH")>;
412def : InstRW<[M55Write2FloatE3], (instregex "MVE_VQ?D?MULL[TB]?[su]")>;
413def : InstRW<[M55Write2FloatE3], (instregex "MVE_VQDMULL_qr_")>;
414def : InstRW<[M55Write2FloatE3], (instregex "MVE_VQ?R?D?ML(A|S)[^L]")>;
415def : InstRW<[M55Write2FloatE3, M55WriteLat1], (instregex "MVE_VR?ML(A|S)L")>;
416def : InstRW<[M55Write2FloatE2], (instregex "MVE_VNEGf")>;
417def : InstRW<[M55Write2FloatE3], (instregex "MVE_VRINTf")>;
418def : InstRW<[M55Write2FloatE2], (instregex "MVE_VSUBf")>;
419def : InstRW<[M55Write2FloatE3], (instregex "MVE_VSUB_qr_f")>;
420
421// Some VMOV's can go down either pipeline.
422def : InstRW<[M55Write2IntFPE2], (instregex "MVE_VMOV_to_lane", "MVE_VMOV_q_rr")>;
423
424def : InstRW<[M55WriteSysE2], (instregex "MVE_VCTP")>;
425def : InstRW<[M55WriteSysE2], (instregex "MVE_VPNOT")>;
426def : InstRW<[M55WriteSysE2], (instregex "MVE_VPST")>;
427
428
429// VFP instructions
430
431def : SchedAlias<WriteFPCVT, M55WriteFloatE3>;
432def : SchedAlias<WriteFPMOV, M55WriteFloatE3>;
433def : SchedAlias<WriteFPALU32, M55WriteFloatE3>;
434def : SchedAlias<WriteFPALU64, M55WriteFloatE3Plus13>;
435def : SchedAlias<WriteFPMUL32, M55WriteFloatE3>;
436def : SchedAlias<WriteFPMUL64, M55WriteFloatE3Plus19>;
437def : SchedAlias<WriteFPMAC32, M55WriteFloatE3Plus2>;
438def : SchedAlias<WriteFPMAC64, M55WriteFloatE3Plus34>;
439def : SchedAlias<WriteFPDIV32, M55WriteFloatE3Plus14>;
440def : SchedAlias<WriteFPDIV64, M55WriteFloatE3Plus28>;
441def : SchedAlias<WriteFPSQRT32, M55WriteFloatE3Plus14>;
442def : SchedAlias<WriteFPSQRT64, M55WriteFloatE3Plus28>;
443def : ReadAdvance<ReadFPMUL, 0>;
444def : ReadAdvance<ReadFPMAC, 0>;
445
446def : InstRW<[M55WriteLSE3], (instregex "VLD")>;
447def : InstRW<[M55WriteLSE2], (instregex "VST")>;
448def : InstRW<[M55WriteLSE3], (instregex "VLLD", "VLST")>;
449
450def : InstRW<[M55WriteFloatE3], (instregex "VABS(H|S|D)")>;
451def : InstRW<[M55WriteFloatE3], (instregex "VCVT(A|M|N|P|R|X|Z)(S|U)(H|S|D)")>;
452def : InstRW<[M55WriteFloatE3], (instregex "VCVT(B|T)(DH|HD)")>;
453def : InstRW<[M55WriteFloatE2], (instregex "VCMPZ?(E|H|S|D)")>;
454def : InstRW<[M55WriteFloatE3Plus7], (instregex "VDIVH")>;
455def : InstRW<[M55WriteFloatE3], (instregex "VFN?M(A|S)(H|S)")>; // VFMA
456def : InstRW<[M55WriteFloatE3Plus22], (instregex "VFN?M(A|S)D")>; // VFMA
457def : InstRW<[M55WriteFloatE3], (instregex "VFP_V(MAX|MIN)NM")>;
458def : InstRW<[M55WriteFloatE3], (instregex "VINSH$", "VMOVH$", "VMOVHR$", "VMOVSR$", "VMOVDRR$")>; // VINS, VMOVX, to-FP reg movs
459def : InstRW<[M55WriteFloatE2], (instregex "VMOVD$", "VMOVS$", "VMOVR")>; // Other VMOV's
460def : InstRW<[M55WriteFloatE2], (instregex "FCONSTH", "FCONSTS", "FCONSTD")>;
461def : InstRW<[M55WriteFloatE2], (instregex "VGETLNi32", "VSETLNi32")>;
462def : InstRW<[M55WriteFloatE2], (instregex "VMSR", "VMRS")>;
463def : InstRW<[M55WriteFloatE3Plus2], (instregex "VN?ML(A|S)H")>; // VMLA
464def : InstRW<[M55WriteFloatE3], (instregex "VNEG(H|S|D)")>;
465def : InstRW<[M55WriteFloatE3], (instregex "VRINT(A|M|N|P|R|X|Z)(H|S|D)")>;
466def : InstRW<[M55WriteFloatE3], (instregex "VSEL..(H|S|D)")>;
467def : InstRW<[M55WriteFloatE3Plus7], (instregex "VSQRTH")>;
468
469def : WriteRes<WriteVLD1, []>;
470def : WriteRes<WriteVLD2, []>;
471def : WriteRes<WriteVLD3, []>;
472def : WriteRes<WriteVLD4, []>;
473def : WriteRes<WriteVST1, []>;
474def : WriteRes<WriteVST2, []>;
475def : WriteRes<WriteVST3, []>;
476def : WriteRes<WriteVST4, []>;
477
478}
479