1 //===- P9InstrResources.td - P9 Instruction Resource Defs -*- tablegen -*-==//
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
7 //===----------------------------------------------------------------------===//
9 // This file defines the resources required by P9 instructions. This is part of
10 // the P9 processor model used for instruction scheduling. This file should
11 // contain all the instructions that may be used on Power 9. This is not
12 // just instructions that are new on Power 9 but also instructions that were
13 // available on earlier architectures and are still used in Power 9.
15 // The makeup of the P9 CPU is modeled as follows:
16 // - Each CPU is made up of two superslices.
17 // - Each superslice is made up of two slices. Therefore, there are 4 slices
19 // - Up to 6 instructions can be dispatched to each CPU. Three per superslice.
21 // - One CY (Crypto) unit P9_CY_*
22 // - One DFU (Decimal Floating Point and Quad Precision) unit P9_DFU_*
23 // - Two PM (Permute) units. One on each superslice. P9_PM_*
24 // - Two DIV (Fixed Point Divide) units. One on each superslize. P9_DIV_*
25 // - Four ALU (Fixed Point Arithmetic) units. One on each slice. P9_ALU_*
26 // - Four DP (Floating Point) units. One on each slice. P9_DP_*
27 // This also includes fixed point multiply add.
28 // - Four AGEN (Address Generation) units. One for each slice. P9_AGEN_*
29 // - Four Load/Store Queues. P9_LS_*
30 // - Each set of instructions will require a number of these resources.
31 //===----------------------------------------------------------------------===//
33 // Two cycle ALU vector operation that uses an entire superslice.
34 // Uses both ALU units (the even ALUE and odd ALUO units), two pipelines
35 // (EXECE, EXECO) and 1 dispatch (DISP) to the given superslice.
36 def : InstRW<[P9_ALUE_2C, P9_ALUO_2C, IP_EXECE_1C, IP_EXECO_1C, DISP_1C],
38 (instregex "VADDU(B|H|W|D)M$"),
39 (instregex "VAND(C)?$"),
40 (instregex "VEXTS(B|H|W)2(D|W)(s)?$"),
41 (instregex "V_SET0(B|H)?$"),
42 (instregex "VS(R|L)(B|H|W|D)$"),
43 (instregex "VSUBU(B|H|W|D)M$"),
44 (instregex "VPOPCNT(B|H)$"),
45 (instregex "VRL(B|H|W|D)$"),
46 (instregex "VSRA(B|H|W|D)$"),
47 (instregex "XV(N)?ABS(D|S)P$"),
48 (instregex "XVCPSGN(D|S)P$"),
49 (instregex "XV(I|X)EXP(D|S)P$"),
50 (instregex "VRL(D|W)(MI|NM)$"),
51 (instregex "VMRG(E|O)W$"),
86 // Restricted Dispatch ALU operation for 3 cycles. The operation runs on a
87 // single slice. However, since it is Restricted, it requires all 3 dispatches
88 // (DISP) for that superslice.
89 def : InstRW<[P9_ALU_3C, IP_EXEC_1C, DISP_3SLOTS_1C],
91 (instregex "TABORT(D|W)C(I)?$"),
92 (instregex "MTFSB(0|1)$"),
93 (instregex "MFFSC(D)?RN(I)?$"),
94 (instregex "CMPRB(8)?$"),
95 (instregex "TD(I)?$"),
96 (instregex "TW(I)?$"),
97 (instregex "FCMPU(S|D)$"),
98 (instregex "XSTSTDC(S|D)P$"),
104 // Standard Dispatch ALU operation for 3 cycles. Only one slice used.
105 def : InstRW<[P9_ALU_3C, IP_EXEC_1C, DISP_1C],
107 (instregex "XSMAX(C|J)?DP$"),
108 (instregex "XSMIN(C|J)?DP$"),
109 (instregex "XSCMP(EQ|EXP|GE|GT|O|U)DP$"),
110 (instregex "CNT(L|T)Z(D|W)(8)?(o)?$"),
111 (instregex "POPCNT(D|W)$"),
112 (instregex "CMPB(8)?$"),
113 (instregex "SETB(8)?$"),
121 // Standard Dispatch ALU operation for 2 cycles. Only one slice used.
122 def : InstRW<[P9_ALU_2C, IP_EXEC_1C, DISP_1C],
124 (instregex "S(L|R)D$"),
125 (instregex "SRAD(I)?$"),
126 (instregex "EXTSWSLI_32_64$"),
127 (instregex "MFV(S)?RD$"),
128 (instregex "MTV(S)?RD$"),
129 (instregex "MTV(S)?RW(A|Z)$"),
130 (instregex "CMP(WI|LWI|W|LW)(8)?$"),
131 (instregex "CMP(L)?D(I)?$"),
132 (instregex "SUBF(I)?C(8)?$"),
133 (instregex "ANDI(S)?o(8)?$"),
134 (instregex "ADDC(8)?$"),
135 (instregex "ADDIC(8)?(o)?$"),
136 (instregex "ADD(8|4)(o)?$"),
137 (instregex "ADD(E|ME|ZE)(8)?(o)?$"),
138 (instregex "SUBF(E|ME|ZE)?(8)?(o)?$"),
139 (instregex "NEG(8)?(o)?$"),
140 (instregex "POPCNTB$"),
141 (instregex "ADD(I|IS)?(8)?$"),
142 (instregex "LI(S)?(8)?$"),
143 (instregex "(X)?OR(I|IS)?(8)?(o)?$"),
144 (instregex "NAND(8)?(o)?$"),
145 (instregex "AND(C)?(8)?(o)?$"),
146 (instregex "NOR(8)?(o)?$"),
147 (instregex "OR(C)?(8)?(o)?$"),
148 (instregex "EQV(8)?(o)?$"),
149 (instregex "EXTS(B|H|W)(8)?(_32)?(_64)?(o)?$"),
150 (instregex "ADD(4|8)(TLS)?(_)?$"),
151 (instregex "NEG(8)?$"),
152 (instregex "ADDI(S)?toc(HA|L)(8)?$"),
174 // Restricted Dispatch ALU operation for 2 cycles. The operation runs on a
175 // single slice. However, since it is Restricted, it requires all 3 dispatches
176 // (DISP) for that superslice.
177 def : InstRW<[P9_ALU_2C, IP_EXEC_1C, DISP_3SLOTS_1C],
179 (instregex "RLDC(L|R)$"),
180 (instregex "RLWIMI(8)?$"),
181 (instregex "RLDIC(L|R)(_32)?(_64)?$"),
182 (instregex "M(F|T)OCRF(8)?$"),
183 (instregex "CR(6)?(UN)?SET$"),
184 (instregex "CR(N)?(OR|AND)(C)?$"),
185 (instregex "S(L|R)W(8)?$"),
186 (instregex "RLW(INM|NM)(8)?$"),
187 (instregex "F(N)?ABS(D|S)$"),
188 (instregex "FNEG(D|S)$"),
189 (instregex "FCPSGN(D|S)$"),
190 (instregex "SRAW(I)?$"),
191 (instregex "ISEL(8)?$"),
202 // Three cycle ALU vector operation that uses an entire superslice.
203 // Uses both ALU units (the even ALUE and odd ALUO units), two pipelines
204 // (EXECE, EXECO) and 1 dispatch (DISP) to the given superslice.
205 def : InstRW<[P9_ALUE_3C, P9_ALUO_3C, IP_EXECE_1C, IP_EXECO_1C, DISP_1C],
207 (instregex "M(T|F)VSCR$"),
208 (instregex "VCMPNEZ(B|H|W)$"),
209 (instregex "VCMPEQU(B|H|W|D)$"),
210 (instregex "VCMPNE(B|H|W)$"),
211 (instregex "VABSDU(B|H|W)$"),
212 (instregex "VADDU(B|H|W)S$"),
213 (instregex "VAVG(S|U)(B|H|W)$"),
214 (instregex "VCMP(EQ|GE|GT)FP(o)?$"),
215 (instregex "VCMPBFP(o)?$"),
216 (instregex "VC(L|T)Z(B|H|W|D)$"),
217 (instregex "VADDS(B|H|W)S$"),
218 (instregex "V(MIN|MAX)FP$"),
219 (instregex "V(MIN|MAX)(S|U)(B|H|W|D)$"),
287 // 7 cycle DP vector operation that uses an entire superslice.
288 // Uses both DP units (the even DPE and odd DPO units), two pipelines (EXECE,
289 // EXECO) and all three dispatches (DISP) to the given superslice.
290 def : InstRW<[P9_DPE_7C, P9_DPO_7C, IP_EXECE_1C, IP_EXECO_1C, DISP_1C],
397 // 5 cycle Restricted DP operation. One DP unit, one EXEC pipeline and all three
398 // dispatch units for the superslice.
399 def : InstRW<[P9_DP_5C, IP_EXEC_1C, DISP_3SLOTS_1C],
401 (instregex "MADD(HD|HDU|LD|LD8)$"),
402 (instregex "MUL(HD|HW|LD|LI|LI8|LW)(U)?$")
405 // 7 cycle Restricted DP operation. One DP unit, one EXEC pipeline and all three
406 // dispatch units for the superslice.
407 def : InstRW<[P9_DP_7C, IP_EXEC_1C, DISP_3SLOTS_1C],
410 (instregex "FRI(N|P|Z|M)(D|S)$"),
411 (instregex "FRE(S)?$"),
412 (instregex "FADD(S)?$"),
413 (instregex "FMSUB(S)?$"),
414 (instregex "FMADD(S)?$"),
415 (instregex "FSUB(S)?$"),
416 (instregex "FCFID(U)?(S)?$"),
417 (instregex "FCTID(U)?(Z)?$"),
418 (instregex "FCTIW(U)?(Z)?$"),
419 (instregex "FRSQRTE(S)?$"),
448 // 7 cycle Restricted DP operation and one 3 cycle ALU operation.
449 // These operations can be done in parallel. The DP is restricted so we need a
450 // full 4 dispatches.
451 def : InstRW<[P9_DP_7C, P9_ALU_3C, IP_EXEC_1C, IP_EXEC_1C,
452 DISP_3SLOTS_1C, DISP_1C],
454 (instregex "FSEL(D|S)o$")
457 // 5 Cycle Restricted DP operation and one 2 cycle ALU operation.
458 def : InstRW<[P9_DPOpAndALUOp_7C, IP_EXEC_1C, IP_EXEC_1C,
459 DISP_3SLOTS_1C, DISP_1C],
461 (instregex "MUL(H|L)(D|W)(U)?o$")
464 // 7 cycle Restricted DP operation and one 3 cycle ALU operation.
465 // These operations must be done sequentially.The DP is restricted so we need a
466 // full 4 dispatches.
467 def : InstRW<[P9_DPOpAndALU2Op_10C, IP_EXEC_1C, IP_EXEC_1C,
468 DISP_3SLOTS_1C, DISP_1C],
470 (instregex "FRI(N|P|Z|M)(D|S)o$"),
471 (instregex "FRE(S)?o$"),
472 (instregex "FADD(S)?o$"),
473 (instregex "FSUB(S)?o$"),
474 (instregex "F(N)?MSUB(S)?o$"),
475 (instregex "F(N)?MADD(S)?o$"),
476 (instregex "FCFID(U)?(S)?o$"),
477 (instregex "FCTID(U)?(Z)?o$"),
478 (instregex "FCTIW(U)?(Z)?o$"),
479 (instregex "FMUL(S)?o$"),
480 (instregex "FRSQRTE(S)?o$"),
484 // 7 cycle DP operation. One DP unit, one EXEC pipeline and 1 dispatch units.
485 def : InstRW<[P9_DP_7C, IP_EXEC_1C, DISP_1C],
520 // Three Cycle PM operation. Only one PM unit per superslice so we use the whole
521 // superslice. That includes both exec pipelines (EXECO, EXECE) and one
523 def : InstRW<[P9_PM_3C, IP_EXECO_1C, IP_EXECE_1C, DISP_1C],
525 (instregex "LVS(L|R)$"),
526 (instregex "VSPLTIS(W|H|B)$"),
527 (instregex "VSPLT(W|H|B)(s)?$"),
528 (instregex "V_SETALLONES(B|H)?$"),
529 (instregex "VEXTRACTU(B|H|W)$"),
530 (instregex "VINSERT(B|H|W|D)$"),
628 // 12 Cycle DFU operation. Only one DFU unit per CPU so we use a whole
629 // superslice. That includes both exec pipelines (EXECO, EXECE) and one
631 def : InstRW<[P9_DFU_12C, IP_EXECE_1C, IP_EXECO_1C, DISP_1C],
652 // 23 Cycle DFU operation. Only one DFU unit per CPU so we use a whole
653 // superslice. That includes both exec pipelines (EXECO, EXECE) and one
655 def : InstRW<[P9_DFU_23C, IP_EXECE_1C, IP_EXECO_1C, DISP_1C],
660 // 24 Cycle DFU operation. Only one DFU unit per CPU so we use a whole
661 // superslice. That includes both exec pipelines (EXECO, EXECE) and one
663 def : InstRW<[P9_DFU_24C, IP_EXECE_1C, IP_EXECO_1C, DISP_1C],
677 // 37 Cycle DFU operation. Only one DFU unit per CPU so we use a whole
678 // superslice. That includes both exec pipelines (EXECO, EXECE) and one
680 def : InstRW<[P9_DFU_37C, IP_EXECE_1C, IP_EXECO_1C, DISP_1C],
685 // 58 Cycle DFU operation. Only one DFU unit per CPU so we use a whole
686 // superslice. That includes both exec pipelines (EXECO, EXECE) and one
688 def : InstRW<[P9_DFU_58C, IP_EXECE_1C, IP_EXECO_1C, DISP_1C],
694 // 76 Cycle DFU operation. Only one DFU unit per CPU so we use a whole
695 // superslice. That includes both exec pipelines (EXECO, EXECE) and all three
697 def : InstRW<[P9_DFU_76C, IP_EXECE_1C, IP_EXECO_1C, DISP_1C],
703 // 6 Cycle Load uses a single slice.
704 def : InstRW<[P9_LS_6C, IP_AGEN_1C, DISP_1C],
706 (instregex "LXVL(L)?")
709 // 5 Cycle Load uses a single slice.
710 def : InstRW<[P9_LS_5C, IP_AGEN_1C, DISP_1C],
712 (instregex "LVE(B|H|W)X$"),
713 (instregex "LVX(L)?"),
714 (instregex "LXSI(B|H)ZX$"),
728 // 4 Cycle Load uses a single slice.
729 def : InstRW<[P9_LS_4C, IP_AGEN_1C, DISP_1C],
731 (instregex "DCB(F|T|ST)(EP)?$"),
732 (instregex "DCBZ(L)?(EP)?$"),
733 (instregex "DCBTST(EP)?$"),
734 (instregex "CP_COPY(8)?$"),
735 (instregex "CP_PASTE(8)?$"),
736 (instregex "ICBI(EP)?$"),
737 (instregex "ICBT(LS)?$"),
738 (instregex "LBARX(L)?$"),
739 (instregex "LBZ(CIX|8|X|X8|XTLS|XTLS_32)?(_)?$"),
740 (instregex "LD(ARX|ARXL|BRX|CIX|X|XTLS)?(_)?$"),
741 (instregex "LH(A|B)RX(L)?(8)?$"),
742 (instregex "LHZ(8|CIX|X|X8|XTLS|XTLS_32)?(_)?$"),
743 (instregex "LWARX(L)?$"),
744 (instregex "LWBRX(8)?$"),
745 (instregex "LWZ(8|CIX|X|X8|XTLS|XTLS_32)?(_)?$"),
757 // 4 Cycle Restricted load uses a single slice but the dispatch for the whole
759 def : InstRW<[P9_LS_4C, IP_AGEN_1C, DISP_3SLOTS_1C],
766 // Cracked Load Instructions.
767 // Load instructions that can be done in parallel.
768 def : InstRW<[P9_LS_4C, P9_LS_4C, IP_AGEN_1C, IP_AGEN_1C,
779 // Cracked Load Instruction.
780 // Requires Load and ALU pieces totaling 6 cycles. The Load and ALU
781 // operations can be run in parallel.
782 def : InstRW<[P9_LS_4C, P9_ALU_2C, IP_EXEC_1C, IP_AGEN_1C,
783 DISP_PAIR_1C, DISP_PAIR_1C],
785 (instregex "L(W|H)ZU(X)?(8)?$")
788 // Cracked TEND Instruction.
789 // Requires Load and ALU pieces totaling 6 cycles. The Load and ALU
790 // operations can be run in parallel.
791 def : InstRW<[P9_LS_4C, P9_ALU_2C, IP_EXEC_1C, IP_AGEN_1C,
798 // Cracked Store Instruction
799 // Consecutive Store and ALU instructions. The store is restricted and requires
801 def : InstRW<[P9_StoreAndALUOp_3C, IP_EXEC_1C, IP_EXEC_1C, IP_AGEN_1C,
802 DISP_3SLOTS_1C, DISP_1C],
804 (instregex "ST(B|H|W|D)CX$")
807 // Cracked Load Instruction.
808 // Two consecutive load operations for a total of 8 cycles.
809 def : InstRW<[P9_LoadAndLoadOp_8C, IP_AGEN_1C, IP_AGEN_1C,
815 // Cracked Load instruction.
816 // Requires consecutive Load and ALU pieces totaling 6 cycles. The Load and ALU
817 // operations cannot be done at the same time and so their latencies are added.
818 def : InstRW<[P9_LoadAndALUOp_6C, IP_EXEC_1C, IP_AGEN_1C,
821 (instregex "LHA(X)?(8)?$"),
822 (instregex "CP_PASTE(8)?o$"),
823 (instregex "LWA(X)?(_32)?$"),
827 // Cracked Restricted Load instruction.
828 // Requires consecutive Load and ALU pieces totaling 6 cycles. The Load and ALU
829 // operations cannot be done at the same time and so their latencies are added.
830 // Full 6 dispatches are required as this is both cracked and restricted.
831 def : InstRW<[P9_LoadAndALUOp_6C, IP_EXEC_1C, IP_AGEN_1C,
832 DISP_3SLOTS_1C, DISP_3SLOTS_1C],
837 // Cracked Load instruction.
838 // Requires consecutive Load and ALU pieces totaling 7 cycles. The Load and ALU
839 // operations cannot be done at the same time and so their latencies are added.
840 // Full 4 dispatches are required as this is a cracked instruction.
841 def : InstRW<[P9_LoadAndALUOp_7C, IP_AGEN_1C, IP_EXEC_1C, DISP_1C, DISP_1C],
847 // Cracked Load instruction.
848 // Requires consecutive Load (4 cycles) and ALU (3 cycles) pieces totaling 7
849 // cycles. The Load and ALU operations cannot be done at the same time and so
850 // their latencies are added.
851 // Full 6 dispatches are required as this is a restricted instruction.
852 def : InstRW<[P9_LoadAndALU2Op_7C, IP_AGEN_1C, IP_EXEC_1C,
853 DISP_3SLOTS_1C, DISP_3SLOTS_1C],
859 // Cracked Load instruction.
860 // Requires consecutive Load and ALU pieces totaling 8 cycles. The Load and ALU
861 // operations cannot be done at the same time and so their latencies are added.
862 // Full 4 dispatches are required as this is a cracked instruction.
863 def : InstRW<[P9_LoadAndALU2Op_8C, IP_AGEN_1C, IP_EXEC_1C, DISP_1C, DISP_1C],
871 // Cracked 3-Way Load Instruction
872 // Load with two ALU operations that depend on each other
873 def : InstRW<[P9_LoadAndALUOp_6C, P9_ALU_2C, IP_AGEN_1C, IP_EXEC_1C, IP_EXEC_1C,
874 DISP_PAIR_1C, DISP_PAIR_1C, DISP_1C],
876 (instregex "LHAU(X)?(8)?$"),
880 // Cracked Load that requires the PM resource.
881 // Since the Load and the PM cannot be done at the same time the latencies are
882 // added. Requires 8 cycles. Since the PM requires the full superslice we need
883 // both EXECE, EXECO pipelines as well as 1 dispatch for the PM. The Load
884 // requires the remaining 1 dispatch.
885 def : InstRW<[P9_LoadAndPMOp_8C, IP_AGEN_1C, IP_EXECE_1C, IP_EXECO_1C,
893 // Single slice Restricted store operation. The restricted operation requires
894 // all three dispatches for the superslice.
895 def : InstRW<[P9_LS_1C, IP_EXEC_1C, IP_AGEN_1C, DISP_3SLOTS_1C],
897 (instregex "STF(S|D|IWX|SX|DX)$"),
898 (instregex "STXS(D|DX|SPX|IWX|IBX|IHX|SP)(v)?$"),
899 (instregex "STW(8)?$"),
900 (instregex "(D|X)FSTORE(f32|f64)$"),
901 (instregex "ST(W|H|D)BRX$"),
902 (instregex "ST(B|H|D)(8)?$"),
903 (instregex "ST(B|W|H|D)(CI)?X(TLS|TLS_32)?(8)?(_)?$"),
911 // Vector Store Instruction
912 // Requires the whole superslice and therefore requires one dispatch
913 // as well as both the Even and Odd exec pipelines.
914 def : InstRW<[P9_LS_1C, IP_EXECE_1C, IP_EXECO_1C, IP_AGEN_1C, DISP_1C],
916 (instregex "STVE(B|H|W)X$"),
917 (instregex "STVX(L)?$"),
918 (instregex "STXV(B16X|H8X|W4X|D2X|L|LL|X)?$")
921 // 5 Cycle DIV operation. Only one DIV unit per superslice so we use the whole
922 // superslice. That includes both exec pipelines (EXECO, EXECE) and two
924 def : InstRW<[P9_DIV_5C, IP_EXECE_1C, IP_EXECO_1C, DISP_EVEN_1C],
926 (instregex "MTCTR(8)?(loop)?$"),
927 (instregex "MTLR(8)?$")
930 // 12 Cycle DIV operation. Only one DIV unit per superslice so we use the whole
931 // superslice. That includes both exec pipelines (EXECO, EXECE) and two
933 def : InstRW<[P9_DIV_12C, IP_EXECE_1C, IP_EXECO_1C, DISP_EVEN_1C],
935 (instregex "M(T|F)VRSAVE(v)?$"),
936 (instregex "M(T|F)PMR$"),
937 (instregex "M(T|F)TB(8)?$"),
938 (instregex "MF(SPR|CTR|LR)(8)?$"),
939 (instregex "M(T|F)MSR(D)?$"),
940 (instregex "MTSPR(8)?$")
943 // 16 Cycle DIV operation. Only one DIV unit per superslice so we use the whole
944 // superslice. That includes both exec pipelines (EXECO, EXECE) and two
946 def : InstRW<[P9_DIV_16C_8, IP_EXECO_1C, IP_EXECE_1C, DISP_EVEN_1C],
953 // 24 Cycle DIV operation. Only one DIV unit per superslice so we use the whole
954 // superslice. That includes both exec pipelines (EXECO, EXECE) and two
956 def : InstRW<[P9_DIV_24C_8, IP_EXECO_1C, IP_EXECE_1C, DISP_EVEN_1C],
967 // 40 Cycle DIV operation. Only one DIV unit per superslice so we use the whole
968 // superslice. That includes both exec pipelines (EXECO, EXECE) and all three
970 def : InstRW<[P9_DIV_40C_8, IP_EXECO_1C, IP_EXECE_1C, DISP_EVEN_1C],
976 // Cracked DIV and ALU operation. Requires one full slice for the ALU operation
977 // and one full superslice for the DIV operation since there is only one DIV per
978 // superslice. Latency of DIV plus ALU is 26.
979 def : InstRW<[P9_IntDivAndALUOp_18C_8, IP_EXECE_1C, IP_EXECO_1C, IP_EXEC_1C,
980 DISP_EVEN_1C, DISP_1C],
982 (instregex "DIVW(U)?(O)?o$")
985 // Cracked DIV and ALU operation. Requires one full slice for the ALU operation
986 // and one full superslice for the DIV operation since there is only one DIV per
987 // superslice. Latency of DIV plus ALU is 26.
988 def : InstRW<[P9_IntDivAndALUOp_26C_8, IP_EXECE_1C, IP_EXECO_1C, IP_EXEC_1C,
989 DISP_EVEN_1C, DISP_1C],
997 // Cracked DIV and ALU operation. Requires one full slice for the ALU operation
998 // and one full superslice for the DIV operation since there is only one DIV per
999 // superslice. Latency of DIV plus ALU is 42.
1000 def : InstRW<[P9_IntDivAndALUOp_42C_8, IP_EXECE_1C, IP_EXECO_1C, IP_EXEC_1C,
1001 DISP_EVEN_1C, DISP_1C],
1007 // CR access instructions in _BrMCR, IIC_BrMCRX.
1009 // Cracked, restricted, ALU operations.
1010 // Here the two ALU ops can actually be done in parallel and therefore the
1011 // latencies are not added together. Otherwise this is like having two
1012 // instructions running together on two pipelines and 6 dispatches. ALU ops are
1014 def : InstRW<[P9_ALU_2C, P9_ALU_2C, IP_EXEC_1C, IP_EXEC_1C,
1015 DISP_3SLOTS_1C, DISP_3SLOTS_1C],
1021 // Cracked ALU operations.
1022 // Here the two ALU ops can actually be done in parallel and therefore the
1023 // latencies are not added together. Otherwise this is like having two
1024 // instructions running together on two pipelines and 2 dispatches. ALU ops are
1026 def : InstRW<[P9_ALU_2C, P9_ALU_2C, IP_EXEC_1C, IP_EXEC_1C,
1029 (instregex "ADDC(8)?o$"),
1030 (instregex "SUBFC(8)?o$")
1033 // Cracked ALU operations.
1034 // Two ALU ops can be done in parallel.
1035 // One is three cycle ALU the ohter is a two cycle ALU.
1036 // One of the ALU ops is restricted the other is not so we have a total of
1038 def : InstRW<[P9_ALU_2C, P9_ALU_3C, IP_EXEC_1C, IP_EXEC_1C,
1039 DISP_3SLOTS_1C, DISP_1C],
1041 (instregex "F(N)?ABS(D|S)o$"),
1042 (instregex "FCPSGN(D|S)o$"),
1043 (instregex "FNEG(D|S)o$"),
1047 // Cracked ALU operations.
1048 // Here the two ALU ops can actually be done in parallel and therefore the
1049 // latencies are not added together. Otherwise this is like having two
1050 // instructions running together on two pipelines and 2 dispatches.
1051 // ALU ops are 3 cycles each.
1052 def : InstRW<[P9_ALU_3C, P9_ALU_3C, IP_EXEC_1C, IP_EXEC_1C,
1058 // Cracked Restricted ALU operations.
1059 // Here the two ALU ops can actually be done in parallel and therefore the
1060 // latencies are not added together. Otherwise this is like having two
1061 // instructions running together on two pipelines and 6 dispatches.
1062 // ALU ops are 3 cycles each.
1063 def : InstRW<[P9_ALU_3C, P9_ALU_3C, IP_EXEC_1C, IP_EXEC_1C,
1064 DISP_3SLOTS_1C, DISP_3SLOTS_1C],
1066 (instregex "MTFSF(b|o)?$"),
1067 (instregex "MTFSFI(o)?$")
1070 // Cracked instruction made of two ALU ops.
1071 // The two ops cannot be done in parallel.
1072 // One of the ALU ops is restricted and takes 3 dispatches.
1073 def : InstRW<[P9_ALUOpAndALUOp_4C, IP_EXEC_1C, IP_EXEC_1C,
1074 DISP_3SLOTS_1C, DISP_1C],
1076 (instregex "RLD(I)?C(R|L)o$"),
1077 (instregex "RLW(IMI|INM|NM)(8)?o$"),
1078 (instregex "SLW(8)?o$"),
1079 (instregex "SRAW(I)?o$"),
1080 (instregex "SRW(8)?o$"),
1085 // Cracked instruction made of two ALU ops.
1086 // The two ops cannot be done in parallel.
1087 // Both of the ALU ops are restricted and take 3 dispatches.
1088 def : InstRW<[P9_ALU2OpAndALU2Op_6C, IP_EXEC_1C, IP_EXEC_1C,
1089 DISP_3SLOTS_1C, DISP_3SLOTS_1C],
1091 (instregex "MFFS(L|CE|o)?$")
1094 // Cracked ALU instruction composed of three consecutive 2 cycle loads for a
1095 // total of 6 cycles. All of the ALU operations are also restricted so each
1096 // takes 3 dispatches for a total of 9.
1097 def : InstRW<[P9_ALUOpAndALUOpAndALUOp_6C, IP_EXEC_1C, IP_EXEC_1C, IP_EXEC_1C,
1098 DISP_3SLOTS_1C, DISP_3SLOTS_1C, DISP_3SLOTS_1C],
1100 (instregex "MFCR(8)?$")
1103 // Cracked instruction made of two ALU ops.
1104 // The two ops cannot be done in parallel.
1105 def : InstRW<[P9_ALUOpAndALUOp_4C, IP_EXEC_1C, IP_EXEC_1C, DISP_1C, DISP_1C],
1107 (instregex "EXTSWSLI_32_64o$"),
1108 (instregex "SRAD(I)?o$"),
1115 // 33 Cycle DP Instruction Restricted. Takes one slice and 3 dispatches.
1116 def : InstRW<[P9_DP_33C_8, IP_EXEC_1C, DISP_3SLOTS_1C],
1121 // 33 Cycle DP Instruction Restricted and Cracked with 3 Cycle ALU.
1122 def : InstRW<[P9_DPOpAndALU2Op_36C_8, IP_EXEC_1C, IP_EXEC_1C,
1123 DISP_3SLOTS_1C, DISP_1C],
1128 // 36 Cycle DP Instruction.
1129 // Instruction can be done on a single slice.
1130 def : InstRW<[P9_DP_36C_10, IP_EXEC_1C, DISP_1C],
1135 // 36 Cycle DP Instruction Restricted. Takes one slice and 3 dispatches.
1136 def : InstRW<[P9_DP_36C_10, IP_EXEC_1C, DISP_3SLOTS_1C],
1141 // 36 Cycle DP Vector Instruction.
1142 def : InstRW<[P9_DPE_36C_10, P9_DPO_36C_10, IP_EXECE_1C, IP_EXECO_1C,
1148 // 27 Cycle DP Vector Instruction.
1149 def : InstRW<[P9_DPE_27C_10, P9_DPO_27C_10, IP_EXECE_1C, IP_EXECO_1C,
1155 // 36 Cycle DP Instruction Restricted and Cracked with 3 Cycle ALU.
1156 def : InstRW<[P9_DPOpAndALU2Op_39C_10, IP_EXEC_1C, IP_EXEC_1C,
1157 DISP_3SLOTS_1C, DISP_1C],
1162 // 26 Cycle DP Instruction.
1163 def : InstRW<[P9_DP_26C_5, IP_EXEC_1C, DISP_1C],
1168 // 26 Cycle DP Instruction Restricted. Takes one slice and 3 dispatches.
1169 def : InstRW<[P9_DP_26C_5, IP_EXEC_1C, DISP_3SLOTS_1C],
1174 // 26 Cycle DP Instruction Restricted and Cracked with 3 Cycle ALU.
1175 def : InstRW<[P9_DPOpAndALU2Op_29C_5, IP_EXEC_1C, IP_EXEC_1C,
1176 DISP_3SLOTS_1C, DISP_1C],
1181 // 33 Cycle DP Instruction. Takes one slice and 1 dispatch.
1182 def : InstRW<[P9_DP_33C_8, IP_EXEC_1C, DISP_1C],
1187 // 22 Cycle DP Instruction Restricted. Takes one slice and 3 dispatches.
1188 def : InstRW<[P9_DP_22C_5, IP_EXEC_1C, DISP_3SLOTS_1C],
1193 // 22 Cycle DP Instruction Restricted and Cracked with 2 Cycle ALU.
1194 def : InstRW<[P9_DPOpAndALU2Op_25C_5, IP_EXEC_1C, IP_EXEC_1C,
1195 DISP_3SLOTS_1C, DISP_1C],
1200 // 22 Cycle DP Instruction. Takes one slice and 1 dispatch.
1201 def : InstRW<[P9_DP_22C_5, IP_EXEC_1C, DISP_1C],
1206 // 24 Cycle DP Vector Instruction. Takes one full superslice.
1207 // Includes both EXECE, EXECO pipelines and 1 dispatch for the given
1209 def : InstRW<[P9_DPE_24C_8, P9_DPO_24C_8, IP_EXECE_1C, IP_EXECO_1C,
1215 // 33 Cycle DP Vector Instruction. Takes one full superslice.
1216 // Includes both EXECE, EXECO pipelines and 1 dispatch for the given
1218 def : InstRW<[P9_DPE_33C_8, P9_DPO_33C_8, IP_EXECE_1C, IP_EXECO_1C,
1224 // Instruction cracked into three pieces. One Load and two ALU operations.
1225 // The Load and one of the ALU ops cannot be run at the same time and so the
1226 // latencies are added together for 6 cycles. The remainaing ALU is 2 cycles.
1227 // Both the load and the ALU that depends on it are restricted and so they take
1228 // a total of 7 dispatches. The final 2 dispatches come from the second ALU op.
1229 // The two EXEC pipelines are for the 2 ALUs while the AGEN is for the load.
1230 def : InstRW<[P9_LoadAndALU2Op_7C, P9_ALU_2C,
1231 IP_AGEN_1C, IP_EXEC_1C, IP_EXEC_1C,
1232 DISP_3SLOTS_1C, DISP_3SLOTS_1C, DISP_1C],
1234 (instregex "LF(SU|SUX)$")
1237 // Cracked instruction made up of a Store and an ALU. The ALU does not depend on
1238 // the store and so it can be run at the same time as the store. The store is
1240 def : InstRW<[P9_LS_1C, P9_ALU_2C, IP_AGEN_1C, IP_EXEC_1C, IP_EXEC_1C,
1241 DISP_3SLOTS_1C, DISP_1C],
1243 (instregex "STF(S|D)U(X)?$"),
1244 (instregex "ST(B|H|W|D)U(X)?(8)?$")
1247 // Cracked instruction made up of a Load and an ALU. The ALU does not depend on
1248 // the load and so it can be run at the same time as the load.
1249 def : InstRW<[P9_LS_4C, P9_ALU_2C, IP_AGEN_1C, IP_EXEC_1C,
1250 DISP_PAIR_1C, DISP_PAIR_1C],
1252 (instregex "LBZU(X)?(8)?$"),
1253 (instregex "LDU(X)?$")
1256 // Cracked instruction made up of a Load and an ALU. The ALU does not depend on
1257 // the load and so it can be run at the same time as the load. The load is also
1258 // restricted. 3 dispatches are from the restricted load while the other two
1259 // are from the ALU. The AGEN pipeline is from the load and the EXEC pipeline
1260 // is required for the ALU.
1261 def : InstRW<[P9_LS_4C, P9_ALU_2C, IP_AGEN_1C, IP_EXEC_1C,
1262 DISP_3SLOTS_1C, DISP_1C],
1264 (instregex "LF(DU|DUX)$")
1267 // Crypto Instructions
1269 // 6 Cycle CY operation. Only one CY unit per CPU so we use a whole
1270 // superslice. That includes both exec pipelines (EXECO, EXECE) and one
1272 def : InstRW<[P9_CY_6C, IP_EXECO_1C, IP_EXECE_1C, DISP_1C],
1274 (instregex "VPMSUM(B|H|W|D)$"),
1275 (instregex "V(N)?CIPHER(LAST)?$"),
1279 // Branch Instructions
1282 def : InstRW<[P9_BR_2C, DISP_BR_1C],
1284 (instregex "BCCCTR(L)?(8)?$"),
1285 (instregex "BCCL(A|R|RL)?$"),
1286 (instregex "BCCTR(L)?(8)?(n)?$"),
1287 (instregex "BD(N)?Z(8|A|Am|Ap|m|p)?$"),
1288 (instregex "BD(N)?ZL(A|Am|Ap|R|R8|RL|RLm|RLp|Rm|Rp|m|p)?$"),
1289 (instregex "BL(_TLS|_NOP)?$"),
1290 (instregex "BL8(_TLS|_NOP|_NOP_TLS|_TLS_)?$"),
1291 (instregex "BLA(8|8_NOP)?$"),
1292 (instregex "BLR(8|L)?$"),
1293 (instregex "TAILB(A)?(8)?$"),
1294 (instregex "TAILBCTR(8)?$"),
1295 (instregex "gBC(A|Aat|CTR|CTRL|L|LA|LAat|LR|LRL|Lat|at)?$"),
1296 (instregex "BCLR(L)?(n)?$"),
1297 (instregex "BCTR(L)?(8)?$"),
1311 // Five Cycle Branch with a 2 Cycle ALU Op
1312 // Operations must be done consecutively and not in parallel.
1313 def : InstRW<[P9_BROpAndALUOp_7C, IP_EXEC_1C, DISP_BR_1C, DISP_1C],
1318 // Special Extracted Instructions For Atomics
1321 def : InstRW<[P9_LS_1C, P9_LS_1C, P9_LS_4C, P9_LS_4C, P9_LS_4C,
1322 IP_EXEC_1C, IP_EXEC_1C, IP_AGEN_1C, IP_AGEN_1C, IP_AGEN_1C,
1323 IP_AGEN_1C, IP_AGEN_1C, DISP_1C, DISP_3SLOTS_1C,
1324 DISP_3SLOTS_1C, DISP_1C, DISP_1C, DISP_1C],
1326 (instregex "L(D|W)AT$")
1330 def : InstRW<[P9_LS_1C, P9_LS_4C, P9_LS_4C, IP_EXEC_1C, IP_AGEN_1C, IP_AGEN_1C,
1331 IP_AGEN_1C, DISP_1C, DISP_3SLOTS_1C, DISP_1C],
1333 (instregex "ST(D|W)AT$")
1336 // Signal Processing Engine (SPE) Instructions
1337 // These instructions are not supported on Power 9
1346 (instregex "EVADD(I)?W$"),
1347 (instregex "EVADD(SM|SS|UM|US)IAAW$"),
1348 (instregex "EVAND(C)?$"),
1349 (instregex "EVCMP(EQ|GTS|GTU|LTS|LTU)$"),
1350 (instregex "EVCNTL(S|Z)W$"),
1351 (instregex "EVDIVW(S|U)$"),
1352 (instregex "EVEXTS(B|H)$"),
1353 (instregex "EVLD(H|W|D)(X)?$"),
1354 (instregex "EVLHH(E|OS|OU)SPLAT(X)?$"),
1355 (instregex "EVLWHE(X)?$"),
1356 (instregex "EVLWHO(S|U)(X)?$"),
1357 (instregex "EVLW(H|W)SPLAT(X)?$"),
1358 (instregex "EVMERGE(HI|LO|HILO|LOHI)$"),
1359 (instregex "EVMHEG(S|U)M(F|I)A(A|N)$"),
1360 (instregex "EVMHES(M|S)(F|I)(A|AA|AAW|ANW)?$"),
1361 (instregex "EVMHEU(M|S)I(A|AA|AAW|ANW)?$"),
1362 (instregex "EVMHOG(U|S)M(F|I)A(A|N)$"),
1363 (instregex "EVMHOS(M|S)(F|I)(A|AA|AAW|ANW)?$"),
1364 (instregex "EVMHOU(M|S)I(A|AA|ANW|AAW)?$"),
1365 (instregex "EVMWHS(M|S)(F|FA|I|IA)$"),
1366 (instregex "EVMWHUMI(A)?$"),
1367 (instregex "EVMWLS(M|S)IA(A|N)W$"),
1368 (instregex "EVMWLU(M|S)I(A|AA|AAW|ANW)?$"),
1369 (instregex "EVMWSM(F|I)(A|AA|AN)?$"),
1370 (instregex "EVMWSSF(A|AA|AN)?$"),
1371 (instregex "EVMWUMI(A|AA|AN)?$"),
1372 (instregex "EV(N|X)?OR(C)?$"),
1373 (instregex "EVR(LW|LWI|NDW)$"),
1374 (instregex "EVSLW(I)?$"),
1375 (instregex "EVSPLAT(F)?I$"),
1376 (instregex "EVSRW(I)?(S|U)$"),
1377 (instregex "EVST(DD|DH|DW|WHE|WHO|WWE|WWO)(X)?$"),
1378 (instregex "EVSUBF(S|U)(M|S)IAAW$"),
1379 (instregex "EVSUB(I)?FW$")
1380 )> { let Unsupported = 1; }
1382 // General Instructions without scheduling support.
1385 (instregex "(H)?RFI(D)?$"),
1386 (instregex "DSS(ALL)?$"),
1387 (instregex "DST(ST)?(T)?(64)?$"),
1388 (instregex "ICBL(C|Q)$"),
1389 (instregex "L(W|H|B)EPX$"),
1390 (instregex "ST(W|H|B)EPX$"),
1391 (instregex "(L|ST)FDEPX$"),
1392 (instregex "M(T|F)SR(IN)?$"),
1393 (instregex "M(T|F)DCR$"),
1394 (instregex "NOP_GT_PWR(6|7)$"),
1395 (instregex "TLB(IA|IVAX|SX|SX2|SX2D|LD|LI|RE|RE2|WE|WE2)$"),
1396 (instregex "WRTEE(I)?$"),
1415 )> { let Unsupported = 1; }