每个稳定 NDF ID 都有可读详情页、canonical source identity，以及对应 instruction/unit workbench 链接。
ACRC MUST implement the mnemonic-local canonical assembly, encoded legality, defaults, state and memory effects, ordering, and fault boundaries declared in this owner. The operation region below is the executable binding for every accepted decision that names this mnemonic.
ACRE MUST be the terminating scalar operation of an active SYS block and MUST act as that block's implicit stop. Request values zero and one are aliases; values two through fifteen MUST reject before recovery effects. ACRE MUST validate recovery without mutation, commit the SYS block, and only then consume and restore the complete saved context atomically.
ADD MUST implement the mnemonic-local canonical assembly, encoded legality, defaults, state and memory effects, ordering, and fault boundaries declared in this owner. The operation region below is the executable binding for every accepted decision that names this mnemonic.
ADDTPC MUST add SignExtend(imm20) shifted left by twelve to the current instruction TPC, MUST wrap at XLEN, and MUST write only through the selected Reg5 destination. It MUST NOT install a control-flow target or directly advance TPC. Encoded immediate zero MUST produce the current TPC.
AND MUST implement the mnemonic-local canonical assembly, encoded legality, defaults, state and memory effects, ordering, and fault boundaries declared in this owner. The operation region below is the executable binding for every accepted decision that names this mnemonic.
Accepted instruction completion and its architecture-visible memory events MUST be defined by the reachable ASL dispatch, completion, and memory-event owners.
B.EQ, B.NE, B.LT, B.GE, B.LTU, B.GEU, B.Z, and B.NZ are reserved two-level block-body forms. PTO scalar decode MUST reject every complete encoding in these families before operand reads or architectural effects. Their spellings MUST NOT be emitted as active PTO instructions.
PTO owns every accepted scalar, block, and Tile encoding. The declared Two-level vector, machine-block, and extended BSTART encodings, the B.TEXT out-of-line block root, and every nonzero BSTART.{FP,STD,SYS} FALL Fixup payload are reserved and MUST NOT be assigned by a future PTO instruction. Permanently deleted instruction names are assembler errors, not reservations; their former slots MAY be assigned by an active PTO instruction.
A target profile MAY provide a precise extension first-use trap. The portable default MUST remain disabled and effect-free. An enabling profile MUST define covered kinds, enable state, source and manager ACRs, the exact trap envelope, pre-effect ordering, retry state, and context-save progress.
A TLOAD or TSTORE B.IOR binding MUST encode an absolute GPR selector for the GM base and an absolute GPR selector for row stride in bytes. Each selected PE MUST resolve both selectors in its private GPR file. When B.IOR is absent, base MUST default to zero and stride MUST default to the dense physical row width in bytes; an explicitly encoded zero stride MUST remain zero. The byte address is base + row * stride + column * element size; packed four-bit columns select floor(column / 2) from each byte-aligned row base and use column parity to select the low or high nibble. Shared TSTORE Function 1 MAY use any nonzero participating PE subset. PE_MASK zero MUST have no effect. Selected PE accesses MUST be preflighted before any effect, and the architecture defines no order among them. Programmers MUST avoid conflicting GM regions.
Architecture-visible state MUST be exactly [[PTO-STATE-ARCH-GPR]], [[PTO-STATE-ARCH-TEMPORARY-QUEUES]], [[PTO-STATE-ARCH-PROGRAM-CONTROL]], [[PTO-STATE-ARCH-FAULT]], [[PTO-STATE-ARCH-MEMORY]], [[PTO-STATE-ARCH-MAINTENANCE]], [[PTO-STATE-ARCH-SYSTEM-REGISTERS]], [[PTO-STATE-ARCH-EXTENDED-SYSTEM-REGISTERS]], [[PTO-STATE-ARCH-TRAP-CONTEXT]], [[PTO-STATE-TILE-LOCAL]], [[PTO-STATE-TILE-SHARED]], [[PTO-STATE-ARCH-GQM]], and [[PTO-STATE-BLOCK-CONTROL]], and MUST change only through the accepted ASL transitions owned by those state units.
BSTART.VEC and BSTART.SFU MUST use the unchanged TEPL Mode/Function carrier. BSTART.TEPL remains an accepted compatibility spelling. Canonical assembly and disassembly MUST render BSTART.VEC or BSTART.SFU according to the selected Tile operation and MUST NOT render BSTART.TEPL.
Every direct Tile operation MUST select exactly one of VEC, TLSU, CUBE, or SFU. VEC MUST execute only elementwise operations. Specialized complex operations use SFU, global-memory and data-transfer operations use TLSU, and matrix and matrix-vector operations use CUBE.
Every direct Tile operation MUST belong to exactly one PTO instruction class: Elementwise Tile-Tile, Tile-Scalar and Immediate, Reduce and Expand, Memory and Data Movement, Matrix and Matrix-Vector, Layout and Rearrangement, or Irregular and Complex. Sync and Config is a PTO Tile programming class but has no direct Tile operation in the current binary carrier. Classification MUST remain independent of execution-engine choice.
A decoded Local consumer binds either a selected CELL range or the complete descriptor-required CELL set after LAST. Waiting is a non-faulting, no-effect state; a consumer reads only after its required set is ready.
A Local B.ASSEMBLE generation is keyed by selected PE mask and architectural destination hand/name. INIT captures one normalized parent descriptor; later writers must match that identity and use a distinct instruction instance unless they are an exact replay. Coverage and readiness are checked before LAST publication, and a fault aborts the pending working version while preserving the committed version, mapping, payload, definedness, and sources.
Each accepted Tile semantic-handler group has exactly one generated effect class. Nonrollback auxiliary effects are rejected before B.ASSEMBLE body or auxiliary effects; rollback-safe and atomic-auxiliary effects participate in the same transaction.
B.ASSEMBLE MUST decode the exact 0x53 form, preserve raw controls and the XLEN-wrapped GPR-plus-uimm11 offset, and apply only to its preceding contiguous B.IOT/B.IOS group.
A Shared B.ASSEMBLE generation MUST retain the previously published Sx object until the matching collective LAST has complete non-overlapping CELL coverage, all declared writer data is ready, every participating PE reaches the same generation ordinal with matching metadata, and no participant has faulted or been squashed. Publication MUST replace the complete Shared descriptor and payload atomically; every rejection MUST preserve the prior published generation.
A Shared destination with more than one participating PE MUST carry one B.ASSEMBLE modifier. A multi-PE standalone B.IOS destination MUST raise Fault_TileLegality before descriptor, payload, memory, or publication effects. A single-PE standalone destination retains the ordinary B.IOS behavior.
Dynamic writers carry an opaque instruction-instance plus execution-domain identity. A squash cancels every unretired contribution in that domain and preserves the older committed mapping.
B.CATR MUST be unique in an active block header. DR MUST select only operation-defined VEC, SFU, or TLSU dimension reduction. trap MUST occur only after a successful atomic commit and MUST resume the selected continuation. far MUST expose results only through initiating-core commit.
B.DATR MUST distinguish omission from every encoded zero value, MUST accept exactly the assigned DataType, Layout, CMode, and RMode values, and MUST defer nonzero field applicability to the selected Tile operation schema.
B.DIM MUST read only absolute GPR codes 0..23, MUST write the low sixteen bits of GPR plus unsigned uimm17, and MUST write each selected LB at most once across all full and compressed dimension forms in one block.
B.FPATR MUST appear exactly once in every CUBE Matrix block and MUST precede every effective B.IOR, B.IOT, or B.IOS binding. TransA and TransB MUST control only the corresponding Shared primary and scale. CScaleEn MUST control only an explicit FP32 C source on TMATMUL.ACC or TMATMULMX.ACC; bit 10 and every other reserved mode value MUST reject before descriptor state or effects. Enabled D, RowMaxOut, and GroupMaxOut results MUST publish as one atomic output group. Every assigned nonzero mode MUST execute the bit-exact FP19, conversion, activation, reduction, packing, exceptional-value, and numeric-status contract rather than an identity or implementation-selected fallback.
An ordinary B.HINT MUST occur at most once in an active block header and MUST NOT change functional results. B.HINT TRACE MUST begin an empty block and that block MUST remain active until BSTOP or a following block start.
B.IOR MUST bind only absolute GPR selectors 0..23, MUST distinguish an omitted instruction from an encoded zero selector, and MUST derive consumed fields and omission defaults from the complete selected block schema.
B.IOS MUST bind the six-bit absolute SharedTileID namespace S0..S63 and MUST reject nonzero reserved bits 27:26 before effects. SizeCode MUST denote one complete Core-wide Shared allocation in the independent 256 KiB Shared pool; decoded PEMode MUST NOT multiply that capacity. PEMode=000 MUST be a strict no-op. PEMode selects participating issuer or consumer PEs and MUST NOT imply payload quarters or offsets. One issuer publishes a complete parent; multi-PE destinations require B.ASSEMBLE. Pending sources MUST wait without payload access or binding consumption.
A participating B.IOT sequence MUST use one common decoded PE mask, MUST terminate exactly once with L, and MUST allocate a new descriptor for each destination. PEMode=000 MUST decode to zero and suppress every placement, schema, and effect check. Destination SizeCode 1..10 MUST encode 128 B through 64 KiB and MUST charge the encoded capacity independently to each selected PE's 256 KiB Local pool; codes 11..15 MUST remain reserved.
Shared producer participation, logical coverage, whole_parent_ready, and published visibility MUST remain distinct. Every Shared source consumer MUST wait/no-op before payload access until whole_parent_ready and published are both true; producer and consumer masks are independent.
A legal Local CUBE B.SUBVIEW MUST derive a bounded descriptor from the parent descriptor, XLEN offset, and encoded view capacity in CELL order. Parent lifetime and payload remain unchanged; non-CUBE, zero-valid, and out-of-range geometry is rejected with Fault_TileLegality before effects.
B.SUBVIEW MUST decode the exact 0x53 form, preserve raw fields and the XLEN-wrapped GPR-plus-uimm11 offset, and apply only to its preceding contiguous B.IOT/B.IOS group.
A Shared B.SUBVIEW source MUST evaluate GPR[RegSrc]+uimm11 in each participating PE's private GPR context. The encoded size is common, but selected PEs may materialize distinct ranges of one published parent.
BSTART MUST initialize BARG.BPC, BARG.BlockType, BARG.BPCN, BARG.TYPE, and BARG.TAKEN; BSTOP or the next BSTART MUST be the only boundary that selects the next PC from BARG.BPCN or the sequential continuation.
BCNT MUST implement the mnemonic-local canonical assembly, encoded legality, defaults, state and memory effects, ordering, and fault boundaries declared in this owner. The operation region below is the executable binding for every accepted decision that names this mnemonic.
BIC MUST implement the mnemonic-local canonical assembly, encoded legality, defaults, state and memory effects, ordering, and fault boundaries declared in this owner. The operation region below is the executable binding for every accepted decision that names this mnemonic.
BIS MUST implement the mnemonic-local canonical assembly, encoded legality, defaults, state and memory effects, ordering, and fault boundaries declared in this owner. The operation region below is the executable binding for every accepted decision that names this mnemonic.
PTO MUST reserve every ERCOV selector combination. A matching form MUST raise Fault_IllegalInstruction before reading any register or changing memory-command, trap-context, execution-context, or TPC state.
PTO MUST reserve every ESAVE selector combination. A matching form MUST raise Fault_IllegalInstruction before reading any register or changing memory-command, trap-context, execution-context, or TPC state.
MSET MUST read only absolute GPR codes 0..23. It MUST reject a complete XLEN length above 63 before effects. For an accepted nonzero length it MUST preflight the complete range and replicate RegSrc1[7:0] to every byte.
B.SUBVIEW applicability is total over accepted Tile operations; the selected operation handler retains ownership of dtype, layout, shape, and definedness legality.
PTO MUST reserve the complete XB raw encoding family. Every matching form MUST raise Fault_IllegalInstruction before decoding an operand value or changing command, block, memory, cross-block, or control-flow state.
BSE MUST implement the mnemonic-local canonical assembly, encoded legality, defaults, state and memory effects, ordering, and fault boundaries declared in this owner. The operation region below is the executable binding for every accepted decision that names this mnemonic.
BSTART.CALL MUST implement the mnemonic-local canonical assembly, encoded legality, defaults, state and memory effects, ordering, and fault boundaries declared in this owner. The operation region below is the executable binding for every accepted decision that names this mnemonic.
BSTART MUST implement the mnemonic-local canonical assembly, encoded legality, defaults, state and memory effects, ordering, and fault boundaries declared in this owner. The operation region below is the executable binding for every accepted decision that names this mnemonic.
BSTART.FP MUST accept only FALL(0), DIRECT, COND, IND, and RET, MUST initialize one FP BARG after successful predecessor retirement, and MUST reject nonzero Fixup payloads and deleted bare call forms before effects.
GMOV MUST rendezvous and preflight all four peer-resolved Local sources, MAY use any nonzero PE_MASK solely to select destination writes, and MUST perform no global-memory or Shared-register effect.
BSTART.ICALL MUST implement the mnemonic-local canonical assembly, encoded legality, defaults, state and memory effects, ordering, and fault boundaries declared in this owner. The operation region below is the executable binding for every accepted decision that names this mnemonic.
A participating BSTART.MGATHER.CAS block MUST contain explicit B.IOR, LB0, one non-terminating Local B.IOT carrying IndexTile and ExpectedTile, and one terminating Local B.IOT carrying ReplacementTile and destination. Omitted LB1, LB2, and B.DATR MUST use the MGATHER.CAS defaults.
A participating BSTART.MGATHER.MASK block MUST contain explicit B.IOR, LB0, and exactly one terminating Local B.IOT carrying IndexTile, MaskTile, and destination. Omitted LB1, LB2, and B.DATR MUST use the masked-gather defaults.
A participating BSTART.MGATHER block MUST contain explicit B.IOR, LB0, and exactly one terminating Local B.IOT carrying IndexTile and destination. Omitted LB1, LB2, and B.DATR MUST use the MGATHER defaults.
A participating BSTART.MSCATTER.MASK block MUST contain explicit B.IOR and LB0 plus exactly two Local B.IOT records: DataTile with IndexTile, followed by MaskTile with the sole last marker. Neither record has a destination.
A participating BSTART.MSCATTER block MUST contain explicit B.IOR, LB0, and exactly one terminating Local B.IOT carrying DataTile and IndexTile with no destination. Omitted LB1, LB2, and B.DATR MUST use MSCATTER defaults.
BSTART.SFU MUST implement the mnemonic-local canonical assembly, encoded legality, defaults, state and memory effects, ordering, and fault boundaries declared in this owner. The operation region below is the executable binding for every accepted decision that names this mnemonic.
BSTART.STD MUST accept only FALL(0), DIRECT, COND, IND, and RET, MUST initialize one STD BARG after successful predecessor retirement, and MUST reject nonzero Fixup payloads and deleted bare call forms before effects.
BSTART.SYS MUST accept only FALL with simm17=0, MUST publish BPC and SYS BlockType after successful predecessor retirement, and MUST treat BPCN, TYPE, and TAKEN as inapplicable non-selecting state.
BSTART.TEPL MUST implement the mnemonic-local canonical assembly, encoded legality, defaults, state and memory effects, ordering, and fault boundaries declared in this owner. The operation region below is the executable binding for every accepted decision that names this mnemonic.
BSTART.TGEMV.ACC MUST select CUBE Function 18 and TileOperation_TGEMV_ACC. Its encoded DataType is AType; omitted B.DATR MUST preserve AType as BType, omitted LB0/LB1/LB2 MUST default M/N/K independently to one, Local N and K MUST be positive and independent of per-PE TSize; M MUST equal one. AType and BType must be supported ordinary Matrix types from one numeric class. C is one explicit Local MxN accumulator source and D is a newly published destination; C and D may use one architectural Tile name. M is fixed to one and every Shared binding is illegal. TransA and TransB MUST remain zero and every Shared binding MUST reject. Every common nonzero Local PE mask is legal; mask zero is a strict no-op. C and D MUST both be explicit; direct Tile calls MUST also use different destination and accumulator TileIndex values. C MUST persist unchanged. Exactly one B.FPATR MUST close postprocess defaults and operands. Complete schema, type, shape, capacity, definedness, readiness, alias, and allocation preflight MUST precede source snapshots and effects; successful outputs MUST publish atomically and sources MUST persist.
BSTART.TGEMV.BIAS MUST select CUBE Function 17 and TileOperation_TGEMV_BIAS. Its encoded DataType is AType; omitted B.DATR MUST preserve AType as BType, omitted LB0/LB1/LB2 MUST default M/N/K independently to one, Local N and K MUST be positive and independent of per-PE TSize; M MUST equal one. AType and BType must be supported ordinary Matrix types from one numeric class. Bias is one Local row-major 1xN accumulator-type source. M is fixed to one and every Shared binding is illegal. TransA and TransB MUST remain zero and every Shared binding MUST reject. Every common nonzero Local PE mask is legal; mask zero is a strict no-op. Exactly one B.FPATR MUST close postprocess defaults and operands. Complete schema, type, shape, capacity, definedness, readiness, alias, and allocation preflight MUST precede source snapshots and effects; successful outputs MUST publish atomically and sources MUST persist.
BSTART.TGEMV MUST select CUBE Function 16 and TileOperation_TGEMV. Its encoded DataType is AType; omitted B.DATR MUST preserve AType as BType, omitted LB0/LB1/LB2 MUST default M/N/K independently to one, Local N and K MUST be positive and independent of per-PE TSize; M MUST equal one. AType and BType must be supported ordinary Matrix types from one numeric class. M is fixed to one and every Shared binding is illegal. TransA and TransB MUST remain zero and every Shared binding MUST reject. Every common nonzero Local PE mask is legal; mask zero is a strict no-op. Exactly one B.FPATR MUST close postprocess defaults and operands. Complete schema, type, shape, capacity, definedness, readiness, alias, and allocation preflight MUST precede source snapshots and effects; successful outputs MUST publish atomically and sources MUST persist.
BSTART.TGEMVMX.ACC MUST select CUBE Function 22 and TileOperation_TGEMV_MX_ACC. Its encoded DataType is AType; omitted B.DATR MUST preserve AType as BType, omitted LB0/LB1/LB2 MUST default M/N/K independently to one, Local N and K MUST be positive and independent of per-PE TSize; M MUST equal one. Each matrix side independently requires an E8M0 scale exactly when its MX input type is not FP16 or BF16. C is one explicit Local MxN accumulator source and D is a distinct newly published destination. C's encoded relative selector MUST differ from zero-extended DstTile before physical rename. M is fixed to one and every Shared binding is illegal. TransA and TransB MUST remain zero and every Shared binding MUST reject. Every common nonzero Local PE mask is legal; mask zero is a strict no-op. C and D MUST both be explicit; direct Tile calls MUST also use different destination and accumulator TileIndex values. C MUST persist unchanged. Exactly one B.FPATR MUST close postprocess defaults and operands. Complete schema, type, shape, capacity, definedness, readiness, alias, and allocation preflight MUST precede source snapshots and effects; successful outputs MUST publish atomically and sources MUST persist.
BSTART.TGEMVMX.BIAS MUST select CUBE Function 21 and TileOperation_TGEMV_MX_BIAS. Its encoded DataType is AType; omitted B.DATR MUST preserve AType as BType, omitted LB0/LB1/LB2 MUST default M/N/K independently to one, Local N and K MUST be positive and independent of per-PE TSize; M MUST equal one. Each matrix side independently requires an E8M0 scale exactly when its MX input type is not FP16 or BF16. Bias is one Local row-major 1xN accumulator- type source. M is fixed to one and every Shared binding is illegal. TransA and TransB MUST remain zero and every Shared binding MUST reject. Every common nonzero Local PE mask is legal; mask zero is a strict no-op. Exactly one B.FPATR MUST close postprocess defaults and operands. Complete schema, type, shape, capacity, definedness, readiness, alias, and allocation preflight MUST precede source snapshots and effects; successful outputs MUST publish atomically and sources MUST persist.
BSTART.TGEMVMX MUST select CUBE Function 20 and TileOperation_TGEMV_MX. Its encoded DataType is AType; omitted B.DATR MUST preserve AType as BType, omitted LB0/LB1/LB2 MUST default M/N/K independently to one, Local N and K MUST be positive and independent of per-PE TSize; M MUST equal one. Each matrix side independently requires an E8M0 scale exactly when its MX input type is not FP16 or BF16. M is fixed to one and every Shared binding is illegal. TransA and TransB MUST remain zero and every Shared binding MUST reject. Every common nonzero Local PE mask is legal; mask zero is a strict no-op. Exactly one B.FPATR MUST close postprocess defaults and operands. Complete schema, type, shape, capacity, definedness, readiness, alias, and allocation preflight MUST precede source snapshots and effects; successful outputs MUST publish atomically and sources MUST persist.
An explicit ND2M32, ND2M16, or ND2N8 B.DATR MUST select Local CUBE conversion, MUST use the BSTART DataType through DTYPE_NONE, MUST interpret LB1 as valid rows and LB0 as valid columns, and MUST reject LB2 or B.IOS.
TLOAD MUST retain B.IOR RegSrc0 as the per-PE GM base and RegSrc1 as the byte row stride, MUST preflight the complete selected footprint, and MUST publish its Local or Shared destination only after success. A singleton Shared issuer MUST load and publish the complete logical parent from that PE's private base and stride; multi-PE Shared producers MUST use B.ASSEMBLE with explicit ranges.
BSTART.TMATMUL.ACC MUST select CUBE Function 2 and TileOperation_TMATMUL_ACC. Its encoded DataType is AType; omitted B.DATR MUST preserve AType as BType, omitted LB0/LB1/LB2 MUST default M/N/K independently to one, Local M, N, and K MUST be positive per-PE dimensions independent of TSize; cooperative LB0 MUST be Core-total group_M. For every cooperative Shared-input TMATMUL, LB0 MUST be Core-total group_M; Shared A MUST have shape group_MxK and Shared B MUST have shape KxN. AType and BType must be supported ordinary Matrix types from one numeric class. C is one explicit Local MxN accumulator source and D is a newly published destination; C and D may use one architectural Tile name. Published Shared operands may replace the right group or both matrix groups; supplementary operands and destinations remain Local. Shared primaries MUST be fully published and ready in all four fixed quarters. TransA/TransB MUST apply only to the corresponding Shared primary. Every cooperative nonzero PE mask MUST be 1111; zero-row PEs MUST have no Local effect. C and D MUST both be explicit; direct Tile calls MUST also use different destination and accumulator TileIndex values. C MUST persist unchanged. Exactly one B.FPATR MUST close postprocess defaults and operands. Complete schema, type, shape, capacity, definedness, readiness, alias, and allocation preflight MUST precede source snapshots and effects; successful outputs MUST publish atomically and sources MUST persist.
BSTART.TMATMUL.BIAS MUST select CUBE Function 1 and TileOperation_TMATMUL_BIAS. Its encoded DataType is AType; omitted B.DATR MUST preserve AType as BType, omitted LB0/LB1/LB2 MUST default M/N/K independently to one, Local M, N, and K MUST be positive per-PE dimensions independent of TSize; cooperative LB0 MUST be Core-total group_M. For every cooperative Shared-input TMATMUL, LB0 MUST be Core-total group_M; Shared A MUST have shape group_MxK and Shared B MUST have shape KxN. AType and BType must be supported ordinary Matrix types from one numeric class. Bias is one Local row-major 1xN accumulator-type source. Published Shared operands may replace the right group or both matrix groups; supplementary operands and destinations remain Local. Shared primaries MUST be fully published and ready in all four fixed quarters. TransA/TransB MUST apply only to the corresponding Shared primary. Every cooperative nonzero PE mask MUST be 1111; zero-row PEs MUST have no Local effect. Exactly one B.FPATR MUST close postprocess defaults and operands. Complete schema, type, shape, capacity, definedness, readiness, alias, and allocation preflight MUST precede source snapshots and effects; successful outputs MUST publish atomically and sources MUST persist.
BSTART.TMATMUL MUST select CUBE Function 0 and TileOperation_TMATMUL. Its encoded DataType is AType; omitted B.DATR MUST preserve AType as BType, omitted LB0/LB1/LB2 MUST default M/N/K independently to one, Local M, N, and K MUST be positive per-PE dimensions independent of TSize; cooperative LB0 MUST be Core-total group_M. For every cooperative Shared-input TMATMUL, LB0 MUST be Core-total group_M; Shared A MUST have shape group_MxK and Shared B MUST have shape KxN. AType and BType must be supported ordinary Matrix types from one numeric class. Published Shared operands may replace the right group or both matrix groups; supplementary operands and destinations remain Local. Shared primaries MUST be fully published and ready in all four fixed quarters. TransA/TransB MUST apply only to the corresponding Shared primary. Every cooperative nonzero PE mask MUST be 1111; zero-row PEs MUST have no Local effect. Exactly one B.FPATR MUST close postprocess defaults and operands. Complete schema, type, shape, capacity, definedness, readiness, alias, and allocation preflight MUST precede source snapshots and effects; successful outputs MUST publish atomically and sources MUST persist.
BSTART.TMATMULMX.ACC MUST select CUBE Function 6 and TileOperation_TMATMUL_MX_ACC. Its encoded DataType is AType; omitted B.DATR MUST preserve AType as BType, omitted LB0/LB1/LB2 MUST default M/N/K independently to one, Local M, N, and K MUST be positive per-PE dimensions independent of TSize; cooperative LB0 MUST be Core-total group_M. For every cooperative Shared-input TMATMUL, LB0 MUST be Core-total group_M; Shared A MUST have shape group_MxK and Shared B MUST have shape KxN. Each matrix side independently requires an E8M0 scale exactly when its MX input type is not FP16 or BF16. C is one explicit Local MxN accumulator source and D is a distinct newly published destination. C's encoded relative selector MUST differ from zero-extended DstTile before physical rename. Published Shared operands may replace the right group or both matrix groups; supplementary operands and destinations remain Local. Shared primaries MUST be fully published and ready in all four fixed quarters. TransA/TransB MUST apply only to the corresponding Shared primary. Every cooperative nonzero PE mask MUST be 1111; zero-row PEs MUST have no Local effect. C and D MUST both be explicit; direct Tile calls MUST also use different destination and accumulator TileIndex values. C MUST persist unchanged. Exactly one B.FPATR MUST close postprocess defaults and operands. Complete schema, type, shape, capacity, definedness, readiness, alias, and allocation preflight MUST precede source snapshots and effects; successful outputs MUST publish atomically and sources MUST persist.
BSTART.TMATMULMX.BIAS MUST select CUBE Function 5 and TileOperation_TMATMUL_MX_BIAS. Its encoded DataType is AType; omitted B.DATR MUST preserve AType as BType, omitted LB0/LB1/LB2 MUST default M/N/K independently to one, Local M, N, and K MUST be positive per-PE dimensions independent of TSize; cooperative LB0 MUST be Core-total group_M. For every cooperative Shared-input TMATMUL, LB0 MUST be Core-total group_M; Shared A MUST have shape group_MxK and Shared B MUST have shape KxN. Each matrix side independently requires an E8M0 scale exactly when its MX input type is not FP16 or BF16. Bias is one Local row-major 1xN accumulator- type source. Published Shared operands may replace the right group or both matrix groups; supplementary operands and destinations remain Local. Shared primaries MUST be fully published and ready in all four fixed quarters. TransA/TransB MUST apply only to the corresponding Shared primary. Every cooperative nonzero PE mask MUST be 1111; zero-row PEs MUST have no Local effect. Exactly one B.FPATR MUST close postprocess defaults and operands. Complete schema, type, shape, capacity, definedness, readiness, alias, and allocation preflight MUST precede source snapshots and effects; successful outputs MUST publish atomically and sources MUST persist.
BSTART.TMATMULMX MUST select CUBE Function 4 and TileOperation_TMATMUL_MX. Its encoded DataType is AType; omitted B.DATR MUST preserve AType as BType, omitted LB0/LB1/LB2 MUST default M/N/K independently to one, Local M, N, and K MUST be positive per-PE dimensions independent of TSize; cooperative LB0 MUST be Core-total group_M. For every cooperative Shared-input TMATMUL, LB0 MUST be Core-total group_M; Shared A MUST have shape group_MxK and Shared B MUST have shape KxN. Each matrix side independently requires an E8M0 scale exactly when its MX input type is not FP16 or BF16. Published Shared operands may replace the right group or both matrix groups; supplementary operands and destinations remain Local. Shared primaries MUST be fully published and ready in all four fixed quarters. TransA/TransB MUST apply only to the corresponding Shared primary. Every cooperative nonzero PE mask MUST be 1111; zero-row PEs MUST have no Local effect. Exactly one B.FPATR MUST close postprocess defaults and operands. Complete schema, type, shape, capacity, definedness, readiness, alias, and allocation preflight MUST precede source snapshots and effects; successful outputs MUST publish atomically and sources MUST persist.
Canonical Function 2 Local-to-Shared MUST let one participating issuer atomically publish the complete Shared parent, while a multi-PE producer MUST use B.ASSEMBLE with explicit ranges and atomic LAST publication. Shared-to-Local MUST wait without consuming bindings until the complete parent is whole-ready and published. B.SUBVIEW is the only partial-source range mechanism. PE_MASK zero MUST have no effect.
TPREFETCH MUST preflight the complete typed, strided footprint of all four PEs before effects and MUST use the same load-event decomposition, ordering, fault, and restart rules as destination-producing TLOAD.
An explicit M322ND, M162ND, or N82ND B.DATR MUST select Local CUBE conversion, MUST use the BSTART DataType through DTYPE_NONE, MUST interpret LB1 as valid rows and LB0 as valid columns, and MUST reject LB2 or B.IOS.
TSTORE MUST interpret B.IOR RegSrc1 as a byte row stride, MUST preflight the complete selected GM footprint before any store, MUST preserve its Local or Shared source, and MUST keep a pending Shared-source block active for retry without reading payload, consuming bindings, or producing a GM effect. Function 1 Shared access MUST accept any nonzero participating PE subset; mask zero MUST have no effect.
BSTART.VEC MUST implement the mnemonic-local canonical assembly, encoded legality, defaults, state and memory effects, ordering, and fault boundaries declared in this owner. The operation region below is the executable binding for every accepted decision that names this mnemonic.
BSTOP MUST implement the mnemonic-local canonical assembly, encoded legality, defaults, state and memory effects, ordering, and fault boundaries declared in this owner. The operation region below is the executable binding for every accepted decision that names this mnemonic.
BWE MUST implement the mnemonic-local canonical assembly, encoded legality, defaults, state and memory effects, ordering, and fault boundaries declared in this owner. The operation region below is the executable binding for every accepted decision that names this mnemonic.
BWI MUST implement the mnemonic-local canonical assembly, encoded legality, defaults, state and memory effects, ordering, and fault boundaries declared in this owner. The operation region below is the executable binding for every accepted decision that names this mnemonic.
BWT MUST implement the mnemonic-local canonical assembly, encoded legality, defaults, state and memory effects, ordering, and fault boundaries declared in this owner. The operation region below is the executable binding for every accepted decision that names this mnemonic.
BXS MUST implement the mnemonic-local canonical assembly, encoded legality, defaults, state and memory effects, ordering, and fault boundaries declared in this owner. The operation region below is the executable binding for every accepted decision that names this mnemonic.
BXU MUST implement the mnemonic-local canonical assembly, encoded legality, defaults, state and memory effects, ordering, and fault boundaries declared in this owner. The operation region below is the executable binding for every accepted decision that names this mnemonic.
C.BSTART MUST open one standard block with either a direct or conditional transfer. Its signed displacement MUST be shifted left by one and added to the C.BSTART address before the retiring block is committed.
C.BSTART.FP MUST accept exactly FALL, IND, and RET. IND MUST snapshot the retiring standard-or-floating BARG.BPCN before commit. Every BrType that is not assigned to another standalone mnemonic MUST raise Fault_IllegalInstruction before effects.
C.BSTART.STD MUST accept exactly FALL, IND, and RET. IND MUST snapshot the retiring standard-or-floating BARG.BPCN before commit. BrType zero remains C.BSTOP; every other unassigned BrType MUST raise Fault_IllegalInstruction before effects.
C.BSTART.SYS MUST open the unique compressed sequential system block. It MUST NOT install a selecting BPCN, TYPE, or TAKEN continuation field.
C.BSTOP MUST implement the mnemonic-local canonical assembly, encoded legality, defaults, state and memory effects, ordering, and fault boundaries declared in this owner. The operation region below is the executable binding for every accepted decision that names this mnemonic.
C.CMP.EQI MUST implement the mnemonic-local canonical assembly, encoded legality, defaults, state and memory effects, ordering, and fault boundaries declared in this owner. The operation region below is the executable binding for every accepted decision that names this mnemonic.
C.CMP.NEI MUST implement the mnemonic-local canonical assembly, encoded legality, defaults, state and memory effects, ordering, and fault boundaries declared in this owner. The operation region below is the executable binding for every accepted decision that names this mnemonic.
C.EBREAK MUST raise software-breakpoint trap 50 and MUST zero-extend its encoded five-bit immediate into the 24-bit trap-cause field. It MUST NOT maintain a separate architectural breakpoint-tag state.
C.SETC.EQ MUST be applicable only in one active Conditional block. Across all SETC condition setters, only the first successful occurrence in that block may snapshot sources and atomically update CommitArgument and BARG.TAKEN. A rejected occurrence MUST NOT consume the shared marker.
C.SETC.NE MUST be applicable only in one active Conditional block. Across all SETC condition setters, only the first successful occurrence in that block may snapshot sources and atomically update CommitArgument and BARG.TAKEN. A rejected occurrence MUST NOT consume the shared marker.
C.SETC.TGT MUST snapshot the complete selected scalar value into BARG.BPCN. At most one successful occurrence is permitted per active Standard/Floating block; applicability and duplicate checks MUST precede source readiness.
C.SETRET MUST implement the mnemonic-local canonical assembly, encoded legality, defaults, state and memory effects, ordering, and fault boundaries declared in this owner. The operation region below is the executable binding for every accepted decision that names this mnemonic.
C.SSRGET MUST assign only direct system-register IDs 0, 1, and 16. Every other five-bit ID MUST reject before the implicit T destination effect.
CLZ MUST implement the mnemonic-local canonical assembly, encoded legality, defaults, state and memory effects, ordering, and fault boundaries declared in this owner. The operation region below is the executable binding for every accepted decision that names this mnemonic.
CTZ MUST implement the mnemonic-local canonical assembly, encoded legality, defaults, state and memory effects, ordering, and fault boundaries declared in this owner. The operation region below is the executable binding for every accepted decision that names this mnemonic.
Every Matrix ACC form MUST read explicit C and publish distinct D. C's encoded relative selector MUST differ from zero-extended DstTile before allocation, and direct Tile calls MUST use different C/D TileIndex values. C MUST persist while D, reductions, and numeric status publish atomically.
Local CUBE layouts MUST use the assigned 128-byte width-parametric CELL mappings, derive storage independently of valid M/N/K, and reject unsupported types or insufficient per-PE capacity before effects.
B.DATR Layout codes 21 through 26 MUST select ND2M32, ND2M16, ND2N8, M322ND, M162ND, and N82ND respectively. The first three are GM-to-Local CUBE loads and the last three are Local CUBE-to-GM stores. These selectors MUST NOT create distinct Tile instruction identities.
B.FPATR bit 9 MUST enable one final mathematical Local CScale source only for FP32 TMATMUL.ACC and TMATMULMX.ACC. CScale MUST be U8 CUBE_M32 [M,1], MUST not alias any output, and MUST scale C[m,n] by 2^-CScale[m,0] before RowMax, GroupMax, activation, quantization, and destination conversion. Bit 10 MUST remain reserved zero; disabled CScale MUST preserve prior results.
Every cooperative Local-A/Shared-B or Shared-A/Shared-B TMATMUL form MUST interpret LB0 as Core-total group_M in 1..128 and MUST use PE_MASK=1111. group_M<=64 selects M_per_PE=16; group_M>=65 selects M_per_PE=32; PE i owns valid_M=clamp(group_M-i*M_per_PE,0,M_per_PE). A zero-row PE MUST retain structural, collective, and Shared preflight while suppressing all compute-only Local resolution, dependency, subview, alias, allocation, generation, payload, parameter, and output effects. TGEMV remains Local-only.
A HiF4 Matrix scale MUST be one raw U32 word containing E6M2 in bits 7:0, eight E1_8 exponents in bits 15:8, and sixteen E1_16 exponents in bits 31:16. E6M2 values 00..FE MUST be finite with bias 48 and two fraction bits; FF MUST be a legal quiet NaN scale. One word scales 64 logical HiF4 lanes through the selected E1_8 plus E1_16 exponent bits.
Local Matrix primary A, C, and D MUST use one compatible CUBE_M16 or CUBE_M32 layout, primary B MUST use CUBE_N8, and their logical M/N/K dimensions MUST remain independent of per-PE TSize.
Each Matrix-MX primary side MUST independently select group-32 E8M0 scale for MX FP8/FP4 carriers or group-64 raw U32 scale for HiF4X2. HiF4X2 MUST be accepted only by Matrix-MX input roles; ordinary Matrix MUST not gain it. Each Local scale MUST use CUBE_M32 and each Shared scale MUST remain an independently bound ordinary Tile with the corresponding primary location.
Matrix scale Tiles in CUBE_M32 MUST use a two-dimensional 128-byte CellReg grid with column/K repeat fast and 32-row repeat slow. Partial final group slots and final row blocks MUST remain invalid storage tail. This generic grid MUST NOT expand primary A/C/D legality beyond one M16/M32 row block.
B.FPATR bits 7 and 8 MUST independently select logical transpose for a corresponding Shared A or B primary and its separately bound scale, while Local-side transpose MUST remain illegal and neither source is mutated.
EBREAK MUST implement the mnemonic-local canonical assembly, encoded legality, defaults, state and memory effects, ordering, and fault boundaries declared in this owner. The operation region below is the executable binding for every accepted decision that names this mnemonic.
FABS MUST implement the mnemonic-local canonical assembly, encoded legality, defaults, state and memory effects, ordering, and fault boundaries declared in this owner. The operation region below is the executable binding for every accepted decision that names this mnemonic.
FCVTA MUST implement the mnemonic-local canonical assembly, encoded legality, defaults, state and memory effects, ordering, and fault boundaries declared in this owner. The operation region below is the executable binding for every accepted decision that names this mnemonic.
FCVTM MUST implement the mnemonic-local canonical assembly, encoded legality, defaults, state and memory effects, ordering, and fault boundaries declared in this owner. The operation region below is the executable binding for every accepted decision that names this mnemonic.
FCVTN MUST implement the mnemonic-local canonical assembly, encoded legality, defaults, state and memory effects, ordering, and fault boundaries declared in this owner. The operation region below is the executable binding for every accepted decision that names this mnemonic.
FCVTP MUST implement the mnemonic-local canonical assembly, encoded legality, defaults, state and memory effects, ordering, and fault boundaries declared in this owner. The operation region below is the executable binding for every accepted decision that names this mnemonic.
FCVTZ MUST implement the mnemonic-local canonical assembly, encoded legality, defaults, state and memory effects, ordering, and fault boundaries declared in this owner. The operation region below is the executable binding for every accepted decision that names this mnemonic.
FENCE.D MUST implement the mnemonic-local canonical assembly, encoded legality, defaults, state and memory effects, ordering, and fault boundaries declared in this owner. The operation region below is the executable binding for every accepted decision that names this mnemonic.
FENCE.I MUST implement the mnemonic-local canonical assembly, encoded legality, defaults, state and memory effects, ordering, and fault boundaries declared in this owner. The operation region below is the executable binding for every accepted decision that names this mnemonic.
FENTRY MUST snapshot the inclusive R2..R23 ring before adjusting sp and MUST commit each descending stack store as one restartable memory step.
FEXIT MUST restore the inclusive R2..R23 ring and MUST retry exactly the first uncommitted load without applying the stack adjustment twice.
FMAX MUST implement the mnemonic-local canonical assembly, encoded legality, defaults, state and memory effects, ordering, and fault boundaries declared in this owner. The operation region below is the executable binding for every accepted decision that names this mnemonic.
FMIN MUST implement the mnemonic-local canonical assembly, encoded legality, defaults, state and memory effects, ordering, and fault boundaries declared in this owner. The operation region below is the executable binding for every accepted decision that names this mnemonic.
FNE MUST implement the mnemonic-local canonical assembly, encoded legality, defaults, state and memory effects, ordering, and fault boundaries declared in this owner. The operation region below is the executable binding for every accepted decision that names this mnemonic.
FNES MUST implement the mnemonic-local canonical assembly, encoded legality, defaults, state and memory effects, ordering, and fault boundaries declared in this owner. The operation region below is the executable binding for every accepted decision that names this mnemonic.
FP19 MUST use one sign bit, an eight-bit bias-127 exponent, and a ten-bit fraction. B.FPATR scales MUST be positive normal values; activation parameters MUST be positive zero or positive normal values. Subnormal and non-finite FP19 parameter carriers are illegal before architectural effects.
FRET.RA MUST snapshot and validate the pre-restore ra target before frame effects, restore the inclusive ring, then publish only that retained target.
FRET.STK MUST begin its inclusive ring at ra, load and validate stack slot zero as the target, then restore remaining slots and publish that target.
GMOV MUST complete Core4 source readiness before effects. Each selected PE reads the Local fragment chosen by its private peer_tid and writes only its same-index Local destination; partial PE masks are legal and mask zero is a strict no-op. GMOV MUST NOT access Shared tile state.
HL.ADDI MUST reconstruct its split uimm24 field exactly and apply addition with a zero-extended unsigned 24-bit immediate. The result MUST wrap modulo 2^PTO_XLEN. SrcL MUST use the complete non-consuming Reg5 source map. RegDst MUST use the common GPR, discard, U-push, and T-push destination map. The source MUST be snapshotted before destination publication. Successful execution MUST advance TPC by six bytes and MUST NOT change memory, Tile, block, reservation, privilege, numeric-status, or fault state.
HL.ADDIW MUST reconstruct its split uimm24 field exactly and apply word addition with a zero-extended unsigned 24-bit immediate. Only the low 32 bits of both operands participate; the 32-bit result MUST be sign-extended to PTO_XLEN. SrcL MUST use the complete non-consuming Reg5 source map. RegDst MUST use the common GPR, discard, U-push, and T-push destination map. The source MUST be snapshotted before destination publication. Successful execution MUST advance TPC by six bytes and MUST NOT change memory, Tile, block, reservation, privilege, numeric-status, or fault state.
HL.ADDTPC MUST add SignExtend(imm32) shifted left by twelve to the current instruction TPC, MUST wrap at XLEN, and MUST write only through the selected Reg5 destination. It MUST NOT install a control-flow target or directly advance TPC. Encoded immediate zero MUST produce the current TPC.
HL.ANDI MUST reconstruct its split simm24 field exactly and apply bitwise conjunction with a sign-extended signed 24-bit immediate. The result MUST wrap modulo 2^PTO_XLEN. SrcL MUST use the complete non-consuming Reg5 source map. RegDst MUST use the common GPR, discard, U-push, and T-push destination map. The source MUST be snapshotted before destination publication. Successful execution MUST advance TPC by six bytes and MUST NOT change memory, Tile, block, reservation, privilege, numeric-status, or fault state.
HL.ANDIW MUST reconstruct its split simm24 field exactly and apply word bitwise conjunction with a sign-extended signed 24-bit immediate. Only the low 32 bits of both operands participate; the 32-bit result MUST be sign-extended to PTO_XLEN. SrcL MUST use the complete non-consuming Reg5 source map. RegDst MUST use the common GPR, discard, U-push, and T-push destination map. The source MUST be snapshotted before destination publication. Successful execution MUST advance TPC by six bytes and MUST NOT change memory, Tile, block, reservation, privilege, numeric-status, or fault state.
HL.BFI MUST implement the mnemonic-local canonical assembly, encoded legality, defaults, state and memory effects, ordering, and fault boundaries declared in this owner. The operation region below is the executable binding for every accepted decision that names this mnemonic.
HL.CCAT MUST snapshot SrcL and SrcR, form {SrcL, SrcR}, and logically shift the concatenation right by the complete seven-bit shamt. Dst0 MUST receive the low result before Dst1 receives the high result. All shifts 0 through 127 MUST be assigned and zero-filling. Both sources MUST use the complete non-consuming Reg5 source map and both destinations MUST use the common ordered Reg5 destination map. Successful execution MUST advance TPC by six bytes and preserve all memory, Tile, block, reservation, privilege, numeric-status, and fault state.
HL.CCATW MUST snapshot SrcL and SrcR, form {SrcL[31:0], SrcR[31:0]}, and logically shift the concatenation right by the complete seven-bit shamt. Dst0 MUST receive the low result before Dst1 receives the high result. Each 32-bit result MUST be sign-extended; shifts 64 through 127 MUST produce two zeros. Both sources MUST use the complete non-consuming Reg5 source map and both destinations MUST use the common ordered Reg5 destination map. Successful execution MUST advance TPC by six bytes and preserve all memory, Tile, block, reservation, privilege, numeric-status, and fault state.
HL.DIV MUST implement the mnemonic-local canonical assembly, encoded legality, defaults, state and memory effects, ordering, and fault boundaries declared in this owner. The operation region below is the executable binding for every accepted decision that names this mnemonic.
HL.DIVU MUST implement the mnemonic-local canonical assembly, encoded legality, defaults, state and memory effects, ordering, and fault boundaries declared in this owner. The operation region below is the executable binding for every accepted decision that names this mnemonic.
HL.DIVUW MUST implement the mnemonic-local canonical assembly, encoded legality, defaults, state and memory effects, ordering, and fault boundaries declared in this owner. The operation region below is the executable binding for every accepted decision that names this mnemonic.
HL.DIVW MUST implement the mnemonic-local canonical assembly, encoded legality, defaults, state and memory effects, ordering, and fault boundaries declared in this owner. The operation region below is the executable binding for every accepted decision that names this mnemonic.
HL.LIS MUST implement the mnemonic-local canonical assembly, encoded legality, defaults, state and memory effects, ordering, and fault boundaries declared in this owner. The operation region below is the executable binding for every accepted decision that names this mnemonic.
HL.LUI MUST reconstruct its encoded 32-bit immediate, place it in result bits 63:32, clear result bits 31:0, and publish through RegDst. Encoded immediate zero MUST materialize zero.
HL.MADDW MUST interpret the low 32 bits of all three sources as signed values and MUST compute their multiply-add result modulo 2^64. RegDst0 MUST receive SignExtend(result[31:0]) before RegDst1 receives SignExtend(result[63:32]). Sources MUST be snapshotted before either write.
HL.ORI MUST reconstruct its split simm24 field exactly and apply bitwise inclusive-or with a sign-extended signed 24-bit immediate. The result MUST wrap modulo 2^PTO_XLEN. SrcL MUST use the complete non-consuming Reg5 source map. RegDst MUST use the common GPR, discard, U-push, and T-push destination map. The source MUST be snapshotted before destination publication. Successful execution MUST advance TPC by six bytes and MUST NOT change memory, Tile, block, reservation, privilege, numeric-status, or fault state.
HL.ORIW MUST reconstruct its split simm24 field exactly and apply word bitwise inclusive-or with a sign-extended signed 24-bit immediate. Only the low 32 bits of both operands participate; the 32-bit result MUST be sign-extended to PTO_XLEN. SrcL MUST use the complete non-consuming Reg5 source map. RegDst MUST use the common GPR, discard, U-push, and T-push destination map. The source MUST be snapshotted before destination publication. Successful execution MUST advance TPC by six bytes and MUST NOT change memory, Tile, block, reservation, privilege, numeric-status, or fault state.
HL.PRF.A MUST assign model 0 to L1, 1 to L2, and 2 to L3. Model values 3 through 31 MUST reject before scalar source reads or effects. Legal hints MUST form the encoded address without translation, memory event, ordering edge, data-access fault, or architecturally guaranteed placement. The effective address MUST be published through RegDst after source snapshot.
HL.PRF MUST assign model 0 to L1, 1 to L2, and 2 to L3. Model values 3 through 31 MUST reject before scalar source reads or effects. Legal hints MUST form the encoded address without translation, memory event, ordering edge, data-access fault, or architecturally guaranteed placement. The effective address MUST be discarded with no destination effect.
HL.PRFI.U MUST assign model 0 to L1, 1 to L2, and 2 to L3. Model values 3 through 31 MUST reject before scalar source reads or effects. Legal hints MUST form the encoded address without translation, memory event, ordering edge, data-access fault, or architecturally guaranteed placement. The effective address MUST be discarded with no destination effect.
HL.PRFI.UA MUST assign model 0 to L1, 1 to L2, and 2 to L3. Model values 3 through 31 MUST reject before scalar source reads or effects. Legal hints MUST form the encoded address without translation, memory event, ordering edge, data-access fault, or architecturally guaranteed placement. The effective address MUST be published through RegDst after source snapshot.
HL.QMT MUST query or atomically initialize one GQM queue, then apply event and suspend-or-restore actions in order. The s+r flag combination MUST reject before reads or effects; SrcR MUST be read only when initialization is selected.
HL.QPOP MUST atomically remove one head entry, optionally notify after success, publish data before status, and provide acquire ordering unless r=1. Empty, missing, or corrupt queues MUST report status without a queue update.
HL.QPUSH MUST atomically insert one entry at the tail or head, optionally notify after success, and provide release ordering unless r=1. Full, suspended, missing, or corrupt queues MUST report status without insertion.
HL.REM MUST snapshot both sources before destination effects. RegDst0 MUST receive the remainder before RegDst1 receives the quotient. Duplicate destinations and queue pushes MUST preserve that write order. Zero-divisor, signed-overflow, selector, and retirement behavior MUST remain identical to the corresponding quotient/remainder arithmetic contract.
HL.REMU MUST snapshot both sources before destination effects. RegDst0 MUST receive the remainder before RegDst1 receives the quotient. Duplicate destinations and queue pushes MUST preserve that write order. Zero-divisor, signed-overflow, selector, and retirement behavior MUST remain identical to the corresponding quotient/remainder arithmetic contract.
HL.REMUW MUST snapshot both sources before destination effects. RegDst0 MUST receive the remainder before RegDst1 receives the quotient. Duplicate destinations and queue pushes MUST preserve that write order. Zero-divisor, signed-overflow, selector, and retirement behavior MUST remain identical to the corresponding quotient/remainder arithmetic contract.
HL.REMW MUST snapshot both sources before destination effects. RegDst0 MUST receive the remainder before RegDst1 receives the quotient. Duplicate destinations and queue pushes MUST preserve that write order. Zero-divisor, signed-overflow, selector, and retirement behavior MUST remain identical to the corresponding quotient/remainder arithmetic contract.
HL.SD.UPO MUST implement the mnemonic-local canonical assembly, encoded legality, defaults, state and memory effects, ordering, and fault boundaries declared in this owner. The operation region below is the executable binding for every accepted decision that names this mnemonic.
HL.SD.UPR MUST implement the mnemonic-local canonical assembly, encoded legality, defaults, state and memory effects, ordering, and fault boundaries declared in this owner. The operation region below is the executable binding for every accepted decision that names this mnemonic.
HL.SETC.ANDI MUST be applicable only in one active Conditional block. Across all SETC condition setters, only the first successful occurrence in that block may snapshot sources and atomically update CommitArgument and BARG.TAKEN. A rejected occurrence MUST NOT consume the shared marker.
HL.SETC.EQI MUST be applicable only in one active Conditional block. Across all SETC condition setters, only the first successful occurrence in that block may snapshot sources and atomically update CommitArgument and BARG.TAKEN. A rejected occurrence MUST NOT consume the shared marker.
HL.SETC.GEI MUST be applicable only in one active Conditional block. Across all SETC condition setters, only the first successful occurrence in that block may snapshot sources and atomically update CommitArgument and BARG.TAKEN. A rejected occurrence MUST NOT consume the shared marker.
HL.SETC.GEUI MUST be applicable only in one active Conditional block. Across all SETC condition setters, only the first successful occurrence in that block may snapshot sources and atomically update CommitArgument and BARG.TAKEN. A rejected occurrence MUST NOT consume the shared marker.
HL.SETC.LTI MUST be applicable only in one active Conditional block. Across all SETC condition setters, only the first successful occurrence in that block may snapshot sources and atomically update CommitArgument and BARG.TAKEN. A rejected occurrence MUST NOT consume the shared marker.
HL.SETC.LTUI MUST be applicable only in one active Conditional block. Across all SETC condition setters, only the first successful occurrence in that block may snapshot sources and atomically update CommitArgument and BARG.TAKEN. A rejected occurrence MUST NOT consume the shared marker.
HL.SETC.NEI MUST be applicable only in one active Conditional block. Across all SETC condition setters, only the first successful occurrence in that block may snapshot sources and atomically update CommitArgument and BARG.TAKEN. A rejected occurrence MUST NOT consume the shared marker.
HL.SETC.ORI MUST be applicable only in one active Conditional block. Across all SETC condition setters, only the first successful occurrence in that block may snapshot sources and atomically update CommitArgument and BARG.TAKEN. A rejected occurrence MUST NOT consume the shared marker.
HL.SETRET MUST implement the mnemonic-local canonical assembly, encoded legality, defaults, state and memory effects, ordering, and fault boundaries declared in this owner. The operation region below is the executable binding for every accepted decision that names this mnemonic.
HL.SH.UPO MUST implement the mnemonic-local canonical assembly, encoded legality, defaults, state and memory effects, ordering, and fault boundaries declared in this owner. The operation region below is the executable binding for every accepted decision that names this mnemonic.
HL.SH.UPR MUST implement the mnemonic-local canonical assembly, encoded legality, defaults, state and memory effects, ordering, and fault boundaries declared in this owner. The operation region below is the executable binding for every accepted decision that names this mnemonic.
HL.SSRGET MUST implement the mnemonic-local canonical assembly, encoded legality, defaults, state and memory effects, ordering, and fault boundaries declared in this owner. The operation region below is the executable binding for every accepted decision that names this mnemonic.
HL.SSRSET MUST implement the mnemonic-local canonical assembly, encoded legality, defaults, state and memory effects, ordering, and fault boundaries declared in this owner. The operation region below is the executable binding for every accepted decision that names this mnemonic.
HL.SUBI MUST reconstruct its split uimm24 field exactly and apply subtraction with a zero-extended unsigned 24-bit immediate. The result MUST wrap modulo 2^PTO_XLEN. SrcL MUST use the complete non-consuming Reg5 source map. RegDst MUST use the common GPR, discard, U-push, and T-push destination map. The source MUST be snapshotted before destination publication. Successful execution MUST advance TPC by six bytes and MUST NOT change memory, Tile, block, reservation, privilege, numeric-status, or fault state.
HL.SUBIW MUST reconstruct its split uimm24 field exactly and apply word subtraction with a zero-extended unsigned 24-bit immediate. Only the low 32 bits of both operands participate; the 32-bit result MUST be sign-extended to PTO_XLEN. SrcL MUST use the complete non-consuming Reg5 source map. RegDst MUST use the common GPR, discard, U-push, and T-push destination map. The source MUST be snapshotted before destination publication. Successful execution MUST advance TPC by six bytes and MUST NOT change memory, Tile, block, reservation, privilege, numeric-status, or fault state.
HL.SW.UPO MUST implement the mnemonic-local canonical assembly, encoded legality, defaults, state and memory effects, ordering, and fault boundaries declared in this owner. The operation region below is the executable binding for every accepted decision that names this mnemonic.
HL.SW.UPR MUST implement the mnemonic-local canonical assembly, encoded legality, defaults, state and memory effects, ordering, and fault boundaries declared in this owner. The operation region below is the executable binding for every accepted decision that names this mnemonic.
HL.XORI MUST reconstruct its split simm24 field exactly and apply bitwise exclusive-or with a sign-extended signed 24-bit immediate. The result MUST wrap modulo 2^PTO_XLEN. SrcL MUST use the complete non-consuming Reg5 source map. RegDst MUST use the common GPR, discard, U-push, and T-push destination map. The source MUST be snapshotted before destination publication. Successful execution MUST advance TPC by six bytes and MUST NOT change memory, Tile, block, reservation, privilege, numeric-status, or fault state.
HL.XORIW MUST reconstruct its split simm24 field exactly and apply word bitwise exclusive-or with a sign-extended signed 24-bit immediate. Only the low 32 bits of both operands participate; the 32-bit result MUST be sign-extended to PTO_XLEN. SrcL MUST use the complete non-consuming Reg5 source map. RegDst MUST use the common GPR, discard, U-push, and T-push destination map. The source MUST be snapshotted before destination publication. Successful execution MUST advance TPC by six bytes and MUST NOT change memory, Tile, block, reservation, privilege, numeric-status, or fault state.
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
该稳定身份由 instruction contract 拥有；链接直接打开 canonical unit workbench。
Former independent Shared movement function encodings are reserved. Canonical TMOV/TSTORE with B.SUBVIEW or B.ASSEMBLE carry all accepted Shared movement, range, readiness, and publication semantics; Function 13 GMOV remains unchanged.
J MUST implement the mnemonic-local canonical assembly, encoded legality, defaults, state and memory effects, ordering, and fault boundaries declared in this owner. The operation region below is the executable binding for every accepted decision that names this mnemonic.
JR MUST implement the mnemonic-local canonical assembly, encoded legality, defaults, state and memory effects, ordering, and fault boundaries declared in this owner. The operation region below is the executable binding for every accepted decision that names this mnemonic.
L.BSTOP MUST implement the mnemonic-local canonical assembly, encoded legality, defaults, state and memory effects, ordering, and fault boundaries declared in this owner. The operation region below is the executable binding for every accepted decision that names this mnemonic.
LSRGET MUST read the active BARG view: ID 0 is BPC, ID 1 is BPCN for Standard/Floating blocks, and ID 2 is the canonical packed control word. IDs 3 through 4095 and inapplicable ID 1 MUST reject before destination effects.
Matrix post-processing MUST reduce the raw accumulator before conversion, select an activation-dependent multiplier before destination conversion, canonicalize special results, and publish D, enabled auxiliary outputs, and sticky flags as one non-faulting commit.
Matrix PreQuant MUST multiply by the selected FP19 scale, round and saturate at an assigned S5, S9, or S17 intermediate, add the signed offset, and only then apply final destination encoding. Shift modes MUST perform their assigned one-through-sixteen-bit ASR and saturate its S16 result.
MCOPY MUST snapshot absolute GPR destination, source, and complete XLEN length, reject wrapping or overlapping nonempty ranges, and copy forward. Each memory step MUST be an exact restart boundary and MUST NOT be repeated.
MGATHER MUST accept non-packed B8-NP, B16, B32, or B64 transfer data with S32, U32, S64, or U64 IndexTile elements. It MUST interpret each index as a full-width byte displacement, move raw carrier bits without numeric-validity rejection, probe the complete valid region before effects, and define the full physical destination using B.DATR padding outside the valid region. Packed four-bit transfer DataTypes MUST reject before allocation.
MGATHER_CAS MUST accept non-packed B8-NP, B16, B32, or B64 transfer data with S32, U32, S64, or U64 byte-displacement indices and preserve raw carrier bits without numeric-validity rejection. Each valid lane MUST perform one atomic compare-and-swap at the signed or unsigned byte displacement supplied by IndexTile and MUST place the value observed by that atomic operation in the corresponding destination element. Duplicate-address lanes MUST serialize in an implementation-defined order and MUST NOT expose a fixed row-major ordering requirement.
MGATHER_CAS MUST preflight every valid-region read and write address before its first atomic effect. On success it MUST publish one fully defined destination whose non-valid physical elements contain the selected pad value.
MaskTile MUST use packed predicate storage and every valid predicate bit MUST be defined. A zero bit MUST suppress address generation, translation, permission checks, memory access, and memory events for that lane and MUST select PadValue for the destination. A one bit MUST enable the signed-or-unsigned byte-displacement load.
MGATHER_MASK MUST preflight all and only enabled addresses before its first load effect. On success it MUST publish one fully defined destination whose disabled and non-valid physical elements contain the selected pad value.
MGATHER_MASK MUST accept non-packed B8-NP, B16, B32, or B64 transfer data with S32, U32, S64, or U64 byte-displacement indices. It MUST preserve raw carrier bits without numeric-validity rejection and MUST reject packed four-bit transfer DataTypes before allocation.
MSCATTER MUST accept non-packed B8-NP, B16, B32, or B64 transfer data with S32, U32, S64, or U64 IndexTile elements. It MUST interpret each index as a full-width byte displacement, move raw carrier bits without numeric-validity rejection, preflight every valid lane before its first store, and leave source descriptors unchanged. Packed transfer types MUST reject.
Duplicate or overlapping target addresses MUST have an implementation-defined winner. B.CATR.atomic MUST NOT impose an internal lane order; it only makes the complete block effect non-interleavable.
Duplicate enabled addresses MUST have an implementation-defined winner. B.CATR.atomic MUST NOT impose an internal enabled-lane order; it only makes the complete block effect non-interleavable.
MaskTile MUST use packed predicate storage and every valid predicate bit MUST be defined. A zero bit MUST suppress address generation, translation, permission checks, stores, and events; a one bit MUST enable the corresponding signed-or-unsigned byte-displacement store.
MSCATTER_MASK MUST accept non-packed B8-NP, B16, B32, or B64 transfer data with S32, U32, S64, or U64 byte-displacement indices. It MUST preserve raw carrier bits without numeric-validity rejection and MUST reject packed four-bit transfer DataTypes before effects.
Every valid finite floating or scale encoding MUST decompose without host floating-point arithmetic into available, sign, integer significand, and integer exponent such that its exact value is (-1)^sign * UInt(significand) * 2^exponent. Invalid internal encodings, infinities, NaNs, and integer Tile DataTypes MUST report unavailable.
Each assigned floating or scale Tile DataType MUST expose one exact carrier, lane, field-width, bias, constrained-bit, and special-value descriptor. Integer Tile DataTypes MUST report that no floating-format descriptor exists.
Numeric execution flags MUST map to CORE_STATE[36:32] as NV, DZ, OF, UF, and NX, and a successful numeric operation MUST OR its produced flags into the existing sticky status without clearing an earlier flag.
OR MUST implement the mnemonic-local canonical assembly, encoded legality, defaults, state and memory effects, ordering, and fault boundaries declared in this owner. The operation region below is the executable binding for every accepted decision that names this mnemonic.
PRF MUST form its modified and shifted register address as a non-faulting hint. Every encoded RegDst value is a non-writing alias, and legal execution MUST perform no translation, memory event, ordering edge, or destination effect.
PRFI.U MUST sign-extend its unscaled 12-bit byte displacement and form one non-faulting L1 hint address. Every encoded RegDst value is a non-writing alias, with no translation, memory event, ordering edge, or destination effect.
A release candidate MUST be the exact commit that passes the pinned ASL model, every independent AVS result, coverage, projections, and release-evidence checks.
Architecture-visible bundle-control state MUST be the state defined by [[PTO-STATE-BLOCK-CONTROL]].
Shared Tile registers MUST be the core-private state defined by [[PTO-STATE-TILE-SHARED]].
Architecture-visible execution state MUST be the state defined by [[PTO-STATE-ARCH-GPR]], [[PTO-STATE-ARCH-TEMPORARY-QUEUES]], [[PTO-STATE-ARCH-PROGRAM-CONTROL]], [[PTO-STATE-ARCH-FAULT]], [[PTO-STATE-ARCH-MEMORY]], [[PTO-STATE-ARCH-MAINTENANCE]], [[PTO-STATE-ARCH-SYSTEM-REGISTERS]], [[PTO-STATE-ARCH-EXTENDED-SYSTEM-REGISTERS]], and [[PTO-STATE-ARCH-TRAP-CONTEXT]].
Local Tile registers and their allocation masks MUST be the state defined by [[PTO-STATE-TILE-LOCAL]].
REV MUST implement the mnemonic-local canonical assembly, encoded legality, defaults, state and memory effects, ordering, and fault boundaries declared in this owner. The operation region below is the executable binding for every accepted decision that names this mnemonic.
SCVTF MUST implement the mnemonic-local canonical assembly, encoded legality, defaults, state and memory effects, ordering, and fault boundaries declared in this owner. The operation region below is the executable binding for every accepted decision that names this mnemonic.
Decisions: ADR-0024, ADR-0029. SD.U MUST snapshot scalar address/data sources, apply its mnemonic width and address-update mode, preflight access before effects, then commit the store event and any base update at the defined restart boundary.
Decisions: ADR-0030. SD.XOR MUST preflight its complete mnemonic-owned atomic access before memory, reservation, event, or destination effects and MUST commit the selected width, ordering, memory update, event, and result atomically.
Decisions: ADR-0024, ADR-0029. SDI MUST snapshot scalar address/data sources, apply its mnemonic width and address-update mode, preflight access before effects, then commit the store event and any base update at the defined restart boundary.
Decisions: ADR-0024, ADR-0029. SDI.U MUST snapshot scalar address/data sources, apply its mnemonic width and address-update mode, preflight access before effects, then commit the store event and any base update at the defined restart boundary.
SETC.AND MUST be applicable only in one active Conditional block. Across all SETC condition setters, only the first successful occurrence in that block may snapshot sources and atomically update CommitArgument and BARG.TAKEN. A rejected occurrence MUST NOT consume the shared marker.
SETC.ANDI MUST be applicable only in one active Conditional block. Across all SETC condition setters, only the first successful occurrence in that block may snapshot sources and atomically update CommitArgument and BARG.TAKEN. A rejected occurrence MUST NOT consume the shared marker.
SETC.EQ MUST be applicable only in one active Conditional block. Across all SETC condition setters, only the first successful occurrence in that block may snapshot sources and atomically update CommitArgument and BARG.TAKEN. A rejected occurrence MUST NOT consume the shared marker.
SETC.EQI MUST be applicable only in one active Conditional block. Across all SETC condition setters, only the first successful occurrence in that block may snapshot sources and atomically update CommitArgument and BARG.TAKEN. A rejected occurrence MUST NOT consume the shared marker.
SETC.GE MUST be applicable only in one active Conditional block. Across all SETC condition setters, only the first successful occurrence in that block may snapshot sources and atomically update CommitArgument and BARG.TAKEN. A rejected occurrence MUST NOT consume the shared marker.
SETC.GEI MUST be applicable only in one active Conditional block. Across all SETC condition setters, only the first successful occurrence in that block may snapshot sources and atomically update CommitArgument and BARG.TAKEN. A rejected occurrence MUST NOT consume the shared marker.
SETC.GEU MUST be applicable only in one active Conditional block. Across all SETC condition setters, only the first successful occurrence in that block may snapshot sources and atomically update CommitArgument and BARG.TAKEN. A rejected occurrence MUST NOT consume the shared marker.
SETC.GEUI MUST be applicable only in one active Conditional block. Across all SETC condition setters, only the first successful occurrence in that block may snapshot sources and atomically update CommitArgument and BARG.TAKEN. A rejected occurrence MUST NOT consume the shared marker.
SETC.LT MUST be applicable only in one active Conditional block. Across all SETC condition setters, only the first successful occurrence in that block may snapshot sources and atomically update CommitArgument and BARG.TAKEN. A rejected occurrence MUST NOT consume the shared marker.
SETC.LTI MUST be applicable only in one active Conditional block. Across all SETC condition setters, only the first successful occurrence in that block may snapshot sources and atomically update CommitArgument and BARG.TAKEN. A rejected occurrence MUST NOT consume the shared marker.
SETC.LTU MUST be applicable only in one active Conditional block. Across all SETC condition setters, only the first successful occurrence in that block may snapshot sources and atomically update CommitArgument and BARG.TAKEN. A rejected occurrence MUST NOT consume the shared marker.
SETC.LTUI MUST be applicable only in one active Conditional block. Across all SETC condition setters, only the first successful occurrence in that block may snapshot sources and atomically update CommitArgument and BARG.TAKEN. A rejected occurrence MUST NOT consume the shared marker.
SETC.NE MUST be applicable only in one active Conditional block. Across all SETC condition setters, only the first successful occurrence in that block may snapshot sources and atomically update CommitArgument and BARG.TAKEN. A rejected occurrence MUST NOT consume the shared marker.
SETC.NEI MUST be applicable only in one active Conditional block. Across all SETC condition setters, only the first successful occurrence in that block may snapshot sources and atomically update CommitArgument and BARG.TAKEN. A rejected occurrence MUST NOT consume the shared marker.
SETC.OR MUST be applicable only in one active Conditional block. Across all SETC condition setters, only the first successful occurrence in that block may snapshot sources and atomically update CommitArgument and BARG.TAKEN. A rejected occurrence MUST NOT consume the shared marker.
SETC.ORI MUST be applicable only in one active Conditional block. Across all SETC condition setters, only the first successful occurrence in that block may snapshot sources and atomically update CommitArgument and BARG.TAKEN. A rejected occurrence MUST NOT consume the shared marker.
Decisions: ADR-0026. SETC.TGT MUST snapshot its complete scalar source into the applicable active block BARG.BPCN only after placement and source readiness checks. Rejection MUST preserve the prior target and pending-block state.
Decisions: ADR-0026, ADR-0027. SETRET MUST compute pre-increment TPC plus the zero-extended unsigned halfword-scaled immediate, update architectural ra and retained return state atomically, then retire sequentially without installing a direct branch target.
Decisions: ADR-0024, ADR-0029. SH MUST snapshot scalar address/data sources, apply its mnemonic width and address-update mode, preflight access before effects, then commit the store event and any base update at the defined restart boundary.
Decisions: ADR-0024, ADR-0029. SH.PCR MUST snapshot scalar address/data sources, apply its mnemonic width and address-update mode, preflight access before effects, then commit the store event and any base update at the defined restart boundary.
Decisions: ADR-0024, ADR-0029. SH.U MUST snapshot scalar address/data sources, apply its mnemonic width and address-update mode, preflight access before effects, then commit the store event and any base update at the defined restart boundary.
Decisions: ADR-0024, ADR-0029. SHI MUST snapshot scalar address/data sources, apply its mnemonic width and address-update mode, preflight access before effects, then commit the store event and any base update at the defined restart boundary.
Decisions: ADR-0024, ADR-0029. SHI.U MUST snapshot scalar address/data sources, apply its mnemonic width and address-update mode, preflight access before effects, then commit the store event and any base update at the defined restart boundary.
Decisions: ADR-0026. SLL MUST snapshot its scalar sources, apply its mnemonic-owned width, immediate, modifier, and wrapping rule, then publish through the assigned destination or commit effect in alias-safe order.
Decisions: ADR-0026. SLLI MUST snapshot its scalar sources, apply its mnemonic-owned width, immediate, modifier, and wrapping rule, then publish through the assigned destination or commit effect in alias-safe order.
Decisions: ADR-0026. SLLIW MUST snapshot its scalar sources, apply its mnemonic-owned width, immediate, modifier, and wrapping rule, then publish through the assigned destination or commit effect in alias-safe order.
Decisions: ADR-0026. SLLW MUST snapshot its scalar sources, apply its mnemonic-owned width, immediate, modifier, and wrapping rule, then publish through the assigned destination or commit effect in alias-safe order.
Current architecture contracts MUST be owned by mnemonic or architecture ASL; catalogs and Markdown MUST remain deterministic projections or evidence.
Decisions: ADR-0026. SRA MUST snapshot its scalar sources, apply its mnemonic-owned width, immediate, modifier, and wrapping rule, then publish through the assigned destination or commit effect in alias-safe order.
Decisions: ADR-0026. SRAI MUST snapshot its scalar sources, apply its mnemonic-owned width, immediate, modifier, and wrapping rule, then publish through the assigned destination or commit effect in alias-safe order.
Decisions: ADR-0026. SRAIW MUST snapshot its scalar sources, apply its mnemonic-owned width, immediate, modifier, and wrapping rule, then publish through the assigned destination or commit effect in alias-safe order.
Decisions: ADR-0026. SRAW MUST snapshot its scalar sources, apply its mnemonic-owned width, immediate, modifier, and wrapping rule, then publish through the assigned destination or commit effect in alias-safe order.
Decisions: ADR-0026. SRL MUST snapshot its scalar sources, apply its mnemonic-owned width, immediate, modifier, and wrapping rule, then publish through the assigned destination or commit effect in alias-safe order.
Decisions: ADR-0026. SRLI MUST snapshot its scalar sources, apply its mnemonic-owned width, immediate, modifier, and wrapping rule, then publish through the assigned destination or commit effect in alias-safe order.
Decisions: ADR-0026. SRLIW MUST snapshot its scalar sources, apply its mnemonic-owned width, immediate, modifier, and wrapping rule, then publish through the assigned destination or commit effect in alias-safe order.
Decisions: ADR-0026. SRLW MUST snapshot its scalar sources, apply its mnemonic-owned width, immediate, modifier, and wrapping rule, then publish through the assigned destination or commit effect in alias-safe order.
Decisions: ADR-0031. SSRGET MUST enforce its mnemonic-owned system or maintenance access domain before effects and MUST preserve destination, queue, and system state on rejection except for the ordinary instruction-attempt trap envelope.
Decisions: ADR-0031. SSRSET MUST enforce its mnemonic-owned system or maintenance access domain before effects and MUST preserve destination, queue, and system state on rejection except for the ordinary instruction-attempt trap envelope.
Decisions: ADR-0031. SSRSWAP MUST enforce its mnemonic-owned system or maintenance access domain before effects and MUST preserve destination, queue, and system state on rejection except for the ordinary instruction-attempt trap envelope.
Decisions: ADR-0026. SUB MUST snapshot its scalar sources, apply its mnemonic-owned width, immediate, modifier, and wrapping rule, then publish through the assigned destination or commit effect in alias-safe order.
Decisions: ADR-0026. SUBI MUST snapshot its scalar sources, apply its mnemonic-owned width, immediate, modifier, and wrapping rule, then publish through the assigned destination or commit effect in alias-safe order.
Decisions: ADR-0026. SUBIW MUST snapshot its scalar sources, apply its mnemonic-owned width, immediate, modifier, and wrapping rule, then publish through the assigned destination or commit effect in alias-safe order.
Decisions: ADR-0026. SUBW MUST snapshot its scalar sources, apply its mnemonic-owned width, immediate, modifier, and wrapping rule, then publish through the assigned destination or commit effect in alias-safe order.
Decisions: ADR-0030. SW.ADD MUST preflight its complete mnemonic-owned atomic access before memory, reservation, event, or destination effects and MUST commit the selected width, ordering, memory update, event, and result atomically.
Decisions: ADR-0024, ADR-0029. SW MUST snapshot scalar address/data sources, apply its mnemonic width and address-update mode, preflight access before effects, then commit the store event and any base update at the defined restart boundary.
Decisions: ADR-0030. SW.AND MUST preflight its complete mnemonic-owned atomic access before memory, reservation, event, or destination effects and MUST commit the selected width, ordering, memory update, event, and result atomically.
Decisions: ADR-0030. SW.OR MUST preflight its complete mnemonic-owned atomic access before memory, reservation, event, or destination effects and MUST commit the selected width, ordering, memory update, event, and result atomically.
Decisions: ADR-0024, ADR-0029. SW.PCR MUST snapshot scalar address/data sources, apply its mnemonic width and address-update mode, preflight access before effects, then commit the store event and any base update at the defined restart boundary.
Decisions: ADR-0030. SW.SMAX MUST preflight its complete mnemonic-owned atomic access before memory, reservation, event, or destination effects and MUST commit the selected width, ordering, memory update, event, and result atomically.
Decisions: ADR-0030. SW.SMIN MUST preflight its complete mnemonic-owned atomic access before memory, reservation, event, or destination effects and MUST commit the selected width, ordering, memory update, event, and result atomically.
Decisions: ADR-0024, ADR-0029. SW.U MUST snapshot scalar address/data sources, apply its mnemonic width and address-update mode, preflight access before effects, then commit the store event and any base update at the defined restart boundary.
Decisions: ADR-0030. SW.UMAX MUST preflight its complete mnemonic-owned atomic access before memory, reservation, event, or destination effects and MUST commit the selected width, ordering, memory update, event, and result atomically.
Decisions: ADR-0030. SW.UMIN MUST preflight its complete mnemonic-owned atomic access before memory, reservation, event, or destination effects and MUST commit the selected width, ordering, memory update, event, and result atomically.
Decisions: ADR-0030. SW.XOR MUST preflight its complete mnemonic-owned atomic access before memory, reservation, event, or destination effects and MUST commit the selected width, ordering, memory update, event, and result atomically.
Decisions: ADR-0030. SWAPB MUST preflight its complete mnemonic-owned atomic access before memory, reservation, event, or destination effects and MUST commit the selected width, ordering, memory update, event, and result atomically.
Decisions: ADR-0030. SWAPD MUST preflight its complete mnemonic-owned atomic access before memory, reservation, event, or destination effects and MUST commit the selected width, ordering, memory update, event, and result atomically.
Decisions: ADR-0030. SWAPH MUST preflight its complete mnemonic-owned atomic access before memory, reservation, event, or destination effects and MUST commit the selected width, ordering, memory update, event, and result atomically.
Decisions: ADR-0030. SWAPW MUST preflight its complete mnemonic-owned atomic access before memory, reservation, event, or destination effects and MUST commit the selected width, ordering, memory update, event, and result atomically.
Decisions: ADR-0024, ADR-0029. SWI MUST snapshot scalar address/data sources, apply its mnemonic width and address-update mode, preflight access before effects, then commit the store event and any base update at the defined restart boundary.
Decisions: ADR-0024, ADR-0029. SWI.U MUST snapshot scalar address/data sources, apply its mnemonic width and address-update mode, preflight access before effects, then commit the store event and any base update at the defined restart boundary.
TABS MUST select VEC Mode 0 Function 15 and MUST compute typed absolute value over one ordered Local source. Element-width integer behavior, floating sign clearing, supported types, closed unary schema, PadValue, complete preflight, source snapshot, and atomic destination publication MUST follow ADR-0080 and NDF clause PTO-TABS-CONTRACT-001. Rejection MUST precede all architectural effects.
TADD MUST select VEC Mode 0 Function 0 and MUST consume exactly two ordered Local Tile sources plus one newly allocated Local Tile destination. LB0 MUST provide nonzero ValidCol; omitted LB1 MUST select ValidRow=1; omitted LB2 MUST select Col=ValidCol. Omitted B.DATR MUST select Null padding; explicit PadValue 00, 01, 10, and 11 MUST select Zero, Max, Min, and Null respectively. Both sources MUST be completely defined and match the destination shape, row-major layout, and one supported arithmetic type. Complete legality and allocation preflight MUST precede source snapshots; the valid result and padding MUST publish together after elementwise add.
TADDS MUST consume one Local source, one per-participating-PE private-GPR scalar from optional B.IOR.RegSrc0, and one renamed Local destination, and MUST compute source + scalar over the valid region. The selected DataType MUST satisfy TileA9DataTypeSupported; only PadValueOrByteId is applicable in B.DATR. Complete legality and allocation preflight MUST precede source/scalar snapshots and atomic payload, padding, status, and descriptor publication.
TAND MUST select VEC Mode 0 Function 6 and MUST compute the bitwise AND of corresponding integer elements. It MUST accept only S64/S32/S16/S8 and U64/U32/U16/U8. PadValue, shape defaults, source snapshots, and atomic destination publication MUST follow the closed Local binary Tile contract.
TANDS MUST consume one Local source, one per-participating-PE private-GPR scalar from optional B.IOR.RegSrc0, and one renamed Local destination, and MUST compute source AND scalar over the valid region. The selected DataType MUST satisfy TileVecScalarIntegerDataTypeSupported; only PadValueOrByteId is applicable in B.DATR. Complete legality and allocation preflight MUST precede source/scalar snapshots and atomic payload, padding, status, and descriptor publication.
TCI MUST select SFU Mode 3 Function 6. It MUST publish one newly allocated row-major Local S32, S16, U32, or U16 destination with ValidRow equal to one. Logical column k MUST contain start+k for ascending direction or start-k for descending direction, modulo the selected element width. Omitted B.IOR MUST select start zero and ascending direction.
TCMP MUST accept exactly the sixteen numeric source types defined by this instruction, MUST reject CMode 6 and 7 before effects, and MUST pack logical comparison i into bit i mod 8 of byte floor(i/8). The destination MUST use predicate-kind storage with source geometry and sufficient packed capacity. Complete legality preflight MUST precede source snapshots, allocation, flag updates, and the atomic predicate destination publication.
TCMPS MUST compare one Local numeric source with one private-GPR scalar, MUST accept only CMode 0 through 5, and MUST pack logical predicate i into bit i mod 8 of byte floor(i/8). Complete legality and allocation preflight MUST precede source/scalar snapshots and atomic predicate publication.
TCOLARGMAX MUST select SFU Mode 2 Function 28 and MUST compute an increasing-row TMAX scan that retains the lowest winning index over one fully defined Local source. The source geometry MUST match the B.DIM-derived shape. The destination MUST use S32 or U32 indices and equal winners MUST retain the lowest index. Only PadValueOrByteId MAY be nondefault in B.DATR. Complete preflight and source snapshot MUST precede atomic result, status, padding, and descriptor publication.
TCOLARGMIN MUST select SFU Mode 2 Function 29 and MUST compute an increasing-row TMIN scan that retains the lowest winning index over one fully defined Local source. The source geometry MUST match the B.DIM-derived shape. The destination MUST use S32 or U32 indices and equal winners MUST retain the lowest index. Only PadValueOrByteId MAY be nondefault in B.DATR. Complete preflight and source snapshot MUST precede atomic result, status, padding, and descriptor publication.
TCOLEXPAND MUST copy the one-row source bit-for-bit down rows. The complete bundle MUST use one terminating Local B.IOT, MUST reject B.IOR and B.IOS, and MUST accept only the DataTypes listed by this owner. Complete preflight and source snapshot MUST precede atomic result, status, padding, and renamed-destination publication.
TCOLEXPANDADD MUST add the full-shape source and one-row source down rows at the selected DataType. The complete bundle MUST use one terminating Local B.IOT, MUST reject B.IOR and B.IOS, and MUST accept only the DataTypes listed by this owner. Complete preflight and source snapshot MUST precede atomic result, status, padding, and renamed-destination publication.
TCOLEXPANDDIV MUST divide the full-shape source and one-row source down rows at the selected DataType. The complete bundle MUST use one terminating Local B.IOT, MUST reject B.IOR and B.IOS, and MUST accept only the DataTypes listed by this owner. Complete preflight and source snapshot MUST precede atomic result, status, padding, and renamed-destination publication.
TCOLEXPANDEXPDIF MUST subtract the one-row SrcDataType source from the full-shape source down rows and MUST exponentiate at DstDataType. The only legal (SrcDataType,DstDataType) pairs are (FP16,FP16), (BF16,BF16), (FP32,FP32), (FP16,FP32), and (BF16,FP32). BSTART selects SrcDataType; omitted B.DATR or DataType=DTYPE_NONE inherits it, while a concrete B.DATR DataType selects DstDataType. The complete bundle MUST use one terminating Local B.IOT, MUST reject B.IOR and B.IOS, and MUST reject every other source/destination pair. Complete preflight and source snapshot MUST precede atomic result, status, padding, and renamed-destination publication.
TCOLEXPANDMAX MUST select the maximum of the full-shape source and one-row source down rows at the selected DataType. The complete bundle MUST use one terminating Local B.IOT, MUST reject B.IOR and B.IOS, and MUST accept only the DataTypes listed by this owner. Complete preflight and source snapshot MUST precede atomic result, status, padding, and renamed-destination publication.
TCOLEXPANDMIN MUST select the minimum of the full-shape source and one-row source down rows at the selected DataType. The complete bundle MUST use one terminating Local B.IOT, MUST reject B.IOR and B.IOS, and MUST accept only the DataTypes listed by this owner. Complete preflight and source snapshot MUST precede atomic result, status, padding, and renamed-destination publication.
TCOLEXPANDMUL MUST multiply the full-shape source and one-row source down rows at the selected DataType. The complete bundle MUST use one terminating Local B.IOT, MUST reject B.IOR and B.IOS, and MUST accept only the DataTypes listed by this owner. Complete preflight and source snapshot MUST precede atomic result, status, padding, and renamed-destination publication.
TCOLEXPANDSUB MUST subtract the full-shape source and one-row source down rows at the selected DataType. The complete bundle MUST use one terminating Local B.IOT, MUST reject B.IOR and B.IOS, and MUST accept only the DataTypes listed by this owner. Complete preflight and source snapshot MUST precede atomic result, status, padding, and renamed-destination publication.
TCOLMAX MUST select SFU Mode 2 Function 17 and MUST compute a typed increasing-row TMAX fold initialized from row zero over one fully defined Local source. The source geometry MUST match the B.DIM-derived shape. The destination MUST use the source DataType and integer arithmetic MUST remain at the selected element width. Only PadValueOrByteId MAY be nondefault in B.DATR. Complete preflight and source snapshot MUST precede atomic result, status, padding, and descriptor publication.
TCOLMIN MUST select SFU Mode 2 Function 18 and MUST compute a typed increasing-row TMIN fold initialized from row zero over one fully defined Local source. The source geometry MUST match the B.DIM-derived shape. The destination MUST use the source DataType and integer arithmetic MUST remain at the selected element width. Only PadValueOrByteId MAY be nondefault in B.DATR. Complete preflight and source snapshot MUST precede atomic result, status, padding, and descriptor publication.
TCOLPROD MUST select SFU Mode 2 Function 19 and MUST compute a typed increasing-row left fold from one using TMUL over one fully defined Local source. The source geometry MUST match the B.DIM-derived shape. The destination MUST use the source DataType and integer arithmetic MUST remain at the selected element width. Only PadValueOrByteId MAY be nondefault in B.DATR. Complete preflight and source snapshot MUST precede atomic result, status, padding, and descriptor publication.
TCOLSUM MUST select SFU Mode 2 Function 16 and MUST compute a typed increasing-row left fold from zero using TADD over one fully defined Local source. The source geometry MUST match the B.DIM-derived shape. The destination MUST use the source DataType and integer arithmetic MUST remain at the selected element width. Only PadValueOrByteId MAY be nondefault in B.DATR. Complete preflight and source snapshot MUST precede atomic result, status, padding, and descriptor publication.
TCONCAT MUST concatenate exactly two persistent Local sources along columns. It MUST NOT accept an axis operand or a vertical concatenation form. Complete source validation and snapshot MUST precede atomic destination payload, definedness, Null-padding, and descriptor publication.
TCVT MUST accept every assigned Tile DataType, MUST distinguish omitted or DTYPE_NONE destination type from encoded FP64, MUST execute every assigned Layout transformation, and MUST reject an incompatible private/public representation before destination allocation or payload effects. Under the named hardware profile an E8M0 destination MUST accept only FP16, BF16, or FP32 sources and MUST apply PTO-TCVT-E8M0-PROFILE-001 exactly.
TCVT to E8M0 MUST accept only FP16, BF16, and FP32 sources. Positive finite values MUST round their base-two exponent under the selected RMode. Zero, negative values, and NaNs MUST produce 0xFF with NV. Positive infinity and finite range overflow or underflow MUST produce 0xFF when Sat is zero and the corresponding finite endpoint when Sat is one, with exact OF or UF plus NX status. Canonicalize MUST retain its representation role.
TDEQUANT MUST select SFU Mode 3 Function 11. It MUST accept one persistent Local S8 or U8 source and MUST publish one newly allocated Local FP32 destination. Each valid result MUST equal source minus the source-typed zero point, multiplied by a positive finite FP32 multiplier, with one FP32 rounding. Omitted B.IOR MUST select multiplier 1.0 and zero point zero. A present zero multiplier MUST be rejected before allocation or effects.
TDIV MUST retain TEPL Mode 0 Function 3 while executing on the SFU engine. It MUST divide an ordered Local numerator by denominator using the selected arithmetic type. Integer zero in the valid denominator rectangle MUST raise Illegal Block Exception before effects; floating zero MUST enter the numeric profile. Shape defaults, PadValue, snapshots, and publication MUST follow the closed binary Tile contract.
TDIVS MUST consume one Local source, one per-participating-PE private-GPR scalar from optional B.IOR.RegSrc0, and one renamed Local destination, and MUST compute source / scalar over the valid region. The selected DataType MUST satisfy TileA7DataTypeSupported; only PadValueOrByteId is applicable in B.DATR. Complete legality and allocation preflight MUST precede source/scalar snapshots and atomic payload, padding, status, and descriptor publication.
TEXP MUST retain TEPL Mode 0 Function 18 while executing on SFU. It MUST consume one persistent Local floating source and one renamed same-type destination, and MUST compute one profile natural exponential per valid element. The FP16, FP32, and BF16 DataTypes, dimension defaults, PadValue-only attribute schema, complete preflight, source snapshot, special values, status accumulation, padding, and atomic publication are mandatory.
TEXPANDS MUST read one optional B.IOR.RegSrc0 private-GPR scalar, MUST use no Tile source, and MUST fill the valid region of one renamed Local numeric destination with the scalar low-width encoding. Complete legality and allocation preflight MUST precede atomic value, padding, and descriptor publication.
TEXTRACT MUST copy one in-range rectangular window selected by two optional private-GPR offsets into one renamed Local destination. Omitted offsets MUST select zero. Complete preflight and source snapshot MUST precede atomic result, padding, and descriptor publication.
TFMA MUST select VEC Mode 0 Function 28 and MUST execute one same-type fused elementwise left multiplied by right plus addend operation. Supported types, the closed two-B.IOT schema, PadValue, exceptional floating cases, complete preflight, three-source snapshot, and atomic publication MUST follow ADR-0080 and NDF clause PTO-TFMA-CONTRACT-001. Rejection MUST precede every architectural effect.
TGATHER MUST accept every non-packed B8-NP, B16, B32, or B64 value type with every S16, U16, S32, U32, S64, or U64 index type. Each index element is a logical source-row selector at the current column. It MUST reject negative, out-of-range, or undefined references before effects, snapshot both sources, move raw value carrier bits without numeric-validity rejection, and publish one complete same-type destination with Null physical padding atomically.
TGEMV_ACC MUST select CUBE Function 18 and TileOperation_TGEMV_ACC. Its encoded DataType is AType; omitted B.DATR MUST preserve AType as BType, omitted LB0/LB1/LB2 MUST default M/N/K independently to one, Local N and K MUST be positive and independent of per-PE TSize; M MUST equal one. AType and BType must be supported ordinary Matrix types from one numeric class. C is one explicit Local MxN accumulator source and D is a newly published destination; C and D may use one architectural Tile name. M is fixed to one and every Shared binding is illegal. TransA and TransB MUST remain zero and every Shared binding MUST reject. Every common nonzero Local PE mask is legal; mask zero is a strict no-op. C and D MUST both be explicit; direct Tile calls MUST also use different destination and accumulator TileIndex values. C MUST persist unchanged. Exactly one B.FPATR MUST close postprocess defaults and operands. Complete schema, type, shape, capacity, definedness, readiness, alias, and allocation preflight MUST precede source snapshots and effects; successful outputs MUST publish atomically and sources MUST persist.
TGEMV_BIAS MUST select CUBE Function 17 and TileOperation_TGEMV_BIAS. Its encoded DataType is AType; omitted B.DATR MUST preserve AType as BType, omitted LB0/LB1/LB2 MUST default M/N/K independently to one, Local N and K MUST be positive and independent of per-PE TSize; M MUST equal one. AType and BType must be supported ordinary Matrix types from one numeric class. Bias is one Local row-major 1xN accumulator-type source. M is fixed to one and every Shared binding is illegal. TransA and TransB MUST remain zero and every Shared binding MUST reject. Every common nonzero Local PE mask is legal; mask zero is a strict no-op. Exactly one B.FPATR MUST close postprocess defaults and operands. Complete schema, type, shape, capacity, definedness, readiness, alias, and allocation preflight MUST precede source snapshots and effects; successful outputs MUST publish atomically and sources MUST persist.
TGEMV MUST select CUBE Function 16 and TileOperation_TGEMV. Its encoded DataType is AType; omitted B.DATR MUST preserve AType as BType, omitted LB0/LB1/LB2 MUST default M/N/K independently to one, Local N and K MUST be positive and independent of per-PE TSize; M MUST equal one. AType and BType must be supported ordinary Matrix types from one numeric class. M is fixed to one and every Shared binding is illegal. TransA and TransB MUST remain zero and every Shared binding MUST reject. Every common nonzero Local PE mask is legal; mask zero is a strict no-op. Exactly one B.FPATR MUST close postprocess defaults and operands. Complete schema, type, shape, capacity, definedness, readiness, alias, and allocation preflight MUST precede source snapshots and effects; successful outputs MUST publish atomically and sources MUST persist.
TGEMV_MX_ACC MUST select CUBE Function 22 and TileOperation_TGEMV_MX_ACC. Its encoded DataType is AType; omitted B.DATR MUST preserve AType as BType, omitted LB0/LB1/LB2 MUST default M/N/K independently to one, Local N and K MUST be positive and independent of per-PE TSize; M MUST equal one. Each matrix side independently requires an E8M0 scale exactly when its MX input type is not FP16 or BF16. C is one explicit Local MxN accumulator source and D is a distinct newly published destination. C's encoded relative selector MUST differ from zero-extended DstTile before physical rename. M is fixed to one and every Shared binding is illegal. TransA and TransB MUST remain zero and every Shared binding MUST reject. Every common nonzero Local PE mask is legal; mask zero is a strict no-op. C and D MUST both be explicit; direct Tile calls MUST also use different destination and accumulator TileIndex values. C MUST persist unchanged. Exactly one B.FPATR MUST close postprocess defaults and operands. Complete schema, type, shape, capacity, definedness, readiness, alias, and allocation preflight MUST precede source snapshots and effects; successful outputs MUST publish atomically and sources MUST persist.
TGEMV_MX_BIAS MUST select CUBE Function 21 and TileOperation_TGEMV_MX_BIAS. Its encoded DataType is AType; omitted B.DATR MUST preserve AType as BType, omitted LB0/LB1/LB2 MUST default M/N/K independently to one, Local N and K MUST be positive and independent of per-PE TSize; M MUST equal one. Each matrix side independently requires an E8M0 scale exactly when its MX input type is not FP16 or BF16. Bias is one Local row-major 1xN accumulator- type source. M is fixed to one and every Shared binding is illegal. TransA and TransB MUST remain zero and every Shared binding MUST reject. Every common nonzero Local PE mask is legal; mask zero is a strict no-op. Exactly one B.FPATR MUST close postprocess defaults and operands. Complete schema, type, shape, capacity, definedness, readiness, alias, and allocation preflight MUST precede source snapshots and effects; successful outputs MUST publish atomically and sources MUST persist.
TGEMV_MX MUST select CUBE Function 20 and TileOperation_TGEMV_MX. Its encoded DataType is AType; omitted B.DATR MUST preserve AType as BType, omitted LB0/LB1/LB2 MUST default M/N/K independently to one, Local N and K MUST be positive and independent of per-PE TSize; M MUST equal one. Each matrix side independently requires an E8M0 scale exactly when its MX input type is not FP16 or BF16. M is fixed to one and every Shared binding is illegal. TransA and TransB MUST remain zero and every Shared binding MUST reject. Every common nonzero Local PE mask is legal; mask zero is a strict no-op. Exactly one B.FPATR MUST close postprocess defaults and operands. Complete schema, type, shape, capacity, definedness, readiness, alias, and allocation preflight MUST precede source snapshots and effects; successful outputs MUST publish atomically and sources MUST persist.
THISTOGRAM MUST select SFU Mode 3 Function 8. It MUST consume one row-major Local U16 or U32 source, one structurally present Local U8 filter, and one mandatory B.DATR whose destination DataType is U32 and whose ByteId is operation-legal. It MUST publish one inclusive 256-bin U32 prefix histogram per source row only after complete preflight and source/filter snapshots.
B.IOT SizeCode MUST denote one selected PE's Local allocation in that PE's independent 256 KiB pool. B.IOS SizeCode MUST denote one complete Core-wide Shared allocation in the independent 256 KiB Shared pool. Local and Shared allocations MUST NOT consume one combined budget.
TIMG2COL MUST read one valid NC1HWC0 or NDC1HWC0 source descriptor, derive its matrix coordinates from posM and posK, and publish one standard Left matrix destination atomically. Spatial or logical-channel padding MUST come only from the typed source descriptor, and the source descriptor MUST remain unchanged.
TINSERT MUST snapshot one persistent old destination and one persistent insertion source, then replace only the selected in-range result window. Uncovered payload and definedness MUST remain identical to the old destination, and the renamed result MUST publish atomically.
Decisions: ADR-0031. TLB.IA MUST enforce its mnemonic-owned system or maintenance access domain before effects and MUST preserve destination, queue, and system state on rejection except for the ordinary instruction-attempt trap envelope.
Decisions: ADR-0031. TLB.IALL MUST enforce its mnemonic-owned system or maintenance access domain before effects and MUST preserve destination, queue, and system state on rejection except for the ordinary instruction-attempt trap envelope.
Decisions: ADR-0031. TLB.IAV MUST enforce its mnemonic-owned system or maintenance access domain before effects and MUST preserve destination, queue, and system state on rejection except for the ordinary instruction-attempt trap envelope.
Decisions: ADR-0031. TLB.IV MUST enforce its mnemonic-owned system or maintenance access domain before effects and MUST preserve destination, queue, and system state on rejection except for the ordinary instruction-attempt trap envelope.
CUBE TLOAD MUST derive persistent CELL geometry from explicit Layout, BSTART DataType, LB1 valid rows, and LB0 valid columns; MUST use TSize only as capacity; and MUST apply the encoded PadValue only after all valid GM reads preflight successfully.
TLOAD MUST use B.IOR RegSrc0 as the per-PE GM base and RegSrc1 as the byte row stride, MUST distinguish omission from encoded zero, MUST preflight the entire selected footprint, and MUST publish exactly one Local destination or complete Shared parent only after success. One Shared issuer loads the whole parent; a multi-PE producer uses B.ASSEMBLE and explicit writer ranges rather than implicit PE_MASK quarters.
TLOG MUST compute one same-type natural logarithm for every valid Local floating element. E4M3 zero MUST produce canonical quiet NaN 0x7F and MUST set DZ without setting OF because the operation is non-saturating.
TMATMUL_ACC MUST select CUBE Function 2 and TileOperation_TMATMUL_ACC. Its encoded DataType is AType; omitted B.DATR MUST preserve AType as BType, omitted LB0/LB1/LB2 MUST default M/N/K independently to one, Local M, N, and K MUST be positive per-PE dimensions independent of TSize; cooperative LB0 MUST be Core-total group_M. For every cooperative Shared-input TMATMUL, LB0 MUST be Core-total group_M; Shared A MUST have shape group_MxK and Shared B MUST have shape KxN. AType and BType must be supported ordinary Matrix types from one numeric class. C is one explicit Local MxN accumulator source and D is a newly published destination; C and D may use one architectural Tile name. Published Shared operands may replace the right group or both matrix groups; supplementary operands and destinations remain Local. Shared primaries MUST satisfy hardware-maintained whole-parent readiness and publication before payload access; fixed-quarter allocation or initialization masks are not prerequisites. TransA/TransB MUST apply only to the corresponding Shared primary. Every cooperative nonzero PE mask MUST be 1111; zero-row PEs MUST have no Local effect. C and D MUST both be explicit; direct Tile calls MUST also use different destination and accumulator TileIndex values. C MUST persist unchanged. Exactly one B.FPATR MUST close postprocess defaults and operands. Complete schema, type, shape, capacity, definedness, readiness, alias, and allocation preflight MUST precede source snapshots and effects; successful outputs MUST publish atomically and sources MUST persist.
TMATMUL_BIAS MUST select CUBE Function 1 and TileOperation_TMATMUL_BIAS. Its encoded DataType is AType; omitted B.DATR MUST preserve AType as BType, omitted LB0/LB1/LB2 MUST default M/N/K independently to one, Local M, N, and K MUST be positive per-PE dimensions independent of TSize; cooperative LB0 MUST be Core-total group_M. For every cooperative Shared-input TMATMUL, LB0 MUST be Core-total group_M; Shared A MUST have shape group_MxK and Shared B MUST have shape KxN. AType and BType must be supported ordinary Matrix types from one numeric class. Bias is one Local row-major 1xN accumulator-type source. Published Shared operands may replace the right group or both matrix groups; supplementary operands and destinations remain Local. Shared primaries MUST satisfy hardware-maintained whole-parent readiness and publication before payload access; fixed-quarter allocation or initialization masks are not prerequisites. TransA/TransB MUST apply only to the corresponding Shared primary. Every cooperative nonzero PE mask MUST be 1111; zero-row PEs MUST have no Local effect. Exactly one B.FPATR MUST close postprocess defaults and operands. Complete schema, type, shape, capacity, definedness, readiness, alias, and allocation preflight MUST precede source snapshots and effects; successful outputs MUST publish atomically and sources MUST persist.
TMATMUL MUST select CUBE Function 0 and TileOperation_TMATMUL. Its encoded DataType is AType; omitted B.DATR MUST preserve AType as BType, omitted LB0/LB1/LB2 MUST default M/N/K independently to one, Local M, N, and K MUST be positive per-PE dimensions independent of TSize; cooperative LB0 MUST be Core-total group_M. For every cooperative Shared-input TMATMUL, LB0 MUST be Core-total group_M; Shared A MUST have shape group_MxK and Shared B MUST have shape KxN. AType and BType must be supported ordinary Matrix types from one numeric class. Published Shared operands may replace the right group or both matrix groups; supplementary operands and destinations remain Local. Shared primaries MUST satisfy hardware-maintained whole-parent readiness and publication before payload access; fixed-quarter allocation or initialization masks are not prerequisites. TransA/TransB MUST apply only to the corresponding Shared primary. Every cooperative nonzero PE mask MUST be 1111; zero-row PEs MUST have no Local effect. Exactly one B.FPATR MUST close postprocess defaults and operands. Complete schema, type, shape, capacity, definedness, readiness, alias, and allocation preflight MUST precede source snapshots and effects; successful outputs MUST publish atomically and sources MUST persist.
TMATMUL_MX_ACC MUST select CUBE Function 6 and TileOperation_TMATMUL_MX_ACC. Its encoded DataType is AType; omitted B.DATR MUST preserve AType as BType, omitted LB0/LB1/LB2 MUST default M/N/K independently to one, Local M, N, and K MUST be positive per-PE dimensions independent of TSize; cooperative LB0 MUST be Core-total group_M. For every cooperative Shared-input TMATMUL, LB0 MUST be Core-total group_M; Shared A MUST have shape group_MxK and Shared B MUST have shape KxN. Each matrix side independently requires an E8M0 scale exactly when its MX input type is not FP16 or BF16. C is one explicit Local MxN accumulator source and D is a distinct newly published destination. C's encoded relative selector MUST differ from zero-extended DstTile before physical rename. Published Shared operands may replace the right group or both matrix groups; supplementary operands and destinations remain Local. Shared primaries MUST satisfy hardware-maintained whole-parent readiness and publication before payload access; fixed-quarter allocation or initialization masks are not prerequisites. TransA/TransB MUST apply only to the corresponding Shared primary. Every cooperative nonzero PE mask MUST be 1111; zero-row PEs MUST have no Local effect. C and D MUST both be explicit; direct Tile calls MUST also use different destination and accumulator TileIndex values. C MUST persist unchanged. Exactly one B.FPATR MUST close postprocess defaults and operands. Complete schema, type, shape, capacity, definedness, readiness, alias, and allocation preflight MUST precede source snapshots and effects; successful outputs MUST publish atomically and sources MUST persist.
TMATMUL_MX_BIAS MUST select CUBE Function 5 and TileOperation_TMATMUL_MX_BIAS. Its encoded DataType is AType; omitted B.DATR MUST preserve AType as BType, omitted LB0/LB1/LB2 MUST default M/N/K independently to one, Local M, N, and K MUST be positive per-PE dimensions independent of TSize; cooperative LB0 MUST be Core-total group_M. For every cooperative Shared-input TMATMUL, LB0 MUST be Core-total group_M; Shared A MUST have shape group_MxK and Shared B MUST have shape KxN. Each matrix side independently requires an E8M0 scale exactly when its MX input type is not FP16 or BF16. Bias is one Local row-major 1xN accumulator- type source. Published Shared operands may replace the right group or both matrix groups; supplementary operands and destinations remain Local. Shared primaries MUST satisfy hardware-maintained whole-parent readiness and publication before payload access; fixed-quarter allocation or initialization masks are not prerequisites. TransA/TransB MUST apply only to the corresponding Shared primary. Every cooperative nonzero PE mask MUST be 1111; zero-row PEs MUST have no Local effect. Exactly one B.FPATR MUST close postprocess defaults and operands. Complete schema, type, shape, capacity, definedness, readiness, alias, and allocation preflight MUST precede source snapshots and effects; successful outputs MUST publish atomically and sources MUST persist.
TMATMUL_MX MUST select CUBE Function 4 and TileOperation_TMATMUL_MX. Its encoded DataType is AType; omitted B.DATR MUST preserve AType as BType, omitted LB0/LB1/LB2 MUST default M/N/K independently to one, Local M, N, and K MUST be positive per-PE dimensions independent of TSize; cooperative LB0 MUST be Core-total group_M. For every cooperative Shared-input TMATMUL, LB0 MUST be Core-total group_M; Shared A MUST have shape group_MxK and Shared B MUST have shape KxN. Each matrix side independently requires an E8M0 scale exactly when its MX input type is not FP16 or BF16. Published Shared operands may replace the right group or both matrix groups; supplementary operands and destinations remain Local. Shared primaries MUST satisfy hardware-maintained whole-parent readiness and publication before payload access; fixed-quarter allocation or initialization masks are not prerequisites. TransA/TransB MUST apply only to the corresponding Shared primary. Every cooperative nonzero PE mask MUST be 1111; zero-row PEs MUST have no Local effect. Exactly one B.FPATR MUST close postprocess defaults and operands. Complete schema, type, shape, capacity, definedness, readiness, alias, and allocation preflight MUST precede source snapshots and effects; successful outputs MUST publish atomically and sources MUST persist.
TMAX MUST select VEC Mode 0 Function 11 and MUST compute a typed maximum over two ordered Local sources. Integer ordering MUST follow signedness. Floating NaN, signaling-NaN status, signed-zero ties, source-encoding legality, closed binary schema, PadValue, complete preflight, source snapshot, and atomic destination publication MUST follow ADR-0080 and NDF clause PTO-TMAX-CONTRACT-001. Rejection MUST precede all architectural effects.
TMAXS MUST consume one Local source, one per-participating-PE private-GPR scalar from optional B.IOR.RegSrc0, and one renamed Local destination, and MUST compute max(source, scalar) over the valid region. The selected DataType MUST satisfy TileA9DataTypeSupported; only PadValueOrByteId is applicable in B.DATR. Complete legality and allocation preflight MUST precede source/scalar snapshots and atomic payload, padding, status, and descriptor publication.
TMIN MUST select VEC Mode 0 Function 12 and MUST compute a typed minimum over two ordered Local sources. Integer ordering MUST follow signedness. Floating NaN, signaling-NaN status, signed-zero ties, source-encoding legality, closed binary schema, PadValue, complete preflight, source snapshot, and atomic destination publication MUST follow ADR-0080 and NDF clause PTO-TMIN-CONTRACT-001. Rejection MUST precede all architectural effects.
TMINS MUST consume one Local source, one per-participating-PE private-GPR scalar from optional B.IOR.RegSrc0, and one renamed Local destination, and MUST compute min(source, scalar) over the valid region. The selected DataType MUST satisfy TileA9DataTypeSupported; only PadValueOrByteId is applicable in B.DATR. Complete legality and allocation preflight MUST precede source/scalar snapshots and atomic payload, padding, status, and descriptor publication.
Local TMOV MUST copy payload and per-element definedness from one persistent Local source to one renamed Local destination after exact shape/type preflight. Shared INSERT, PUBLISH, BROADCAST, and EXTRACT modes MUST use the distinct BSTART.TMOV schemas and MUST preserve the accepted publication rules.
TMRGSORT MUST select SFU Mode 3 Function 13 and MUST accept exactly FP32 and FP16. It MUST reject a source not sorted in the selected direction before effects, MUST merge equal values from the left source first, MUST place numeric values before NaNs in either direction, and MUST set numeric invalid status when a signaling NaN is observed.
TMUL MUST select VEC Mode 0 Function 2 and MUST multiply corresponding elements of two ordered Local sources. The closed binary Tile schema, shape defaults, arithmetic DataType set, PadValue behavior, complete preflight, source snapshot, and atomic destination publication MUST match the TADD family contract. Rejection MUST precede all effects.
TMULS MUST consume one Local source, one per-participating-PE private-GPR scalar from optional B.IOR.RegSrc0, and one renamed Local destination, and MUST compute source * scalar over the valid region. The selected DataType MUST satisfy TileA7DataTypeSupported; only PadValueOrByteId is applicable in B.DATR. Complete legality and allocation preflight MUST precede source/scalar snapshots and atomic payload, padding, status, and descriptor publication.
TNEG MUST select VEC Mode 0 Function 17 and MUST negate integers modulo their element width or toggle only the floating sign bit. Supported types, closed unary schema, PadValue, complete preflight, source snapshot, and atomic destination publication MUST follow ADR-0080 and NDF clause PTO-TNEG-CONTRACT-001. Rejection MUST precede all architectural effects.
TNOT MUST select VEC Mode 0 Function 16 and MUST complement only the selected 8, 16, 32, or 64-bit integer element width. Supported types, closed unary schema, PadValue, complete preflight, source snapshot, and atomic destination publication MUST follow ADR-0080 and NDF clause PTO-TNOT-CONTRACT-001. Rejection MUST precede all architectural effects.
TOR MUST select VEC Mode 0 Function 7 and MUST compute the bitwise OR of corresponding integer elements. It MUST accept only S64/S32/S16/S8 and U64/U32/U16/U8. PadValue, shape defaults, source snapshots, and atomic destination publication MUST follow the closed Local binary Tile contract.
TORS MUST consume one Local source, one per-participating-PE private-GPR scalar from optional B.IOR.RegSrc0, and one renamed Local destination, and MUST compute source OR scalar over the valid region. The selected DataType MUST satisfy TileVecScalarIntegerDataTypeSupported; only PadValueOrByteId is applicable in B.DATR. Complete legality and allocation preflight MUST precede source/scalar snapshots and atomic payload, padding, status, and descriptor publication.
TPACK assembles two low-order byte fields into each corresponding U32 word. Widths are explicit control bytes in 1..3 with total width at most four; all unselected result bits are zero.
TPERMUTE performs a two-source per-row raw-byte lookup over Local CUBE_M16 or CUBE_M32 Cells. Every valid destination byte and its index are checked before the fresh destination is published; invalid indices reject atomically.
TPREFETCH MUST use TLOAD-equivalent typed address generation and events for all four PE-private base/stride inputs and MUST NOT allocate or write a Tile.
TQUANT MUST select SFU Mode 3 Function 10. It MUST accept one persistent Local FP32 source and MUST publish one newly allocated Local S8 or U8 destination. Each valid result MUST equal the rounded value of source times a positive finite FP32 multiplier plus a destination-typed zero point. Omitted B.IOR MUST select multiplier 1.0 and zero point zero. A present zero multiplier MUST be rejected before allocation or architectural effects.
TRECIP MUST compute one same-type reciprocal for every valid Local floating element. Either E4M3 signed zero MUST produce canonical quiet NaN 0x7F and MUST set DZ without setting OF.
TRELU MUST select VEC Mode 0 Function 23 and MUST compute a same-type rectifier over one ordered Local source. Integer signedness, floating NaN and invalid status, supported types, closed unary schema, PadValue, complete preflight, source snapshot, and atomic publication MUST follow ADR-0080 and NDF clause PTO-TRELU-CONTRACT-001. Rejection MUST precede all architectural effects.
TREM MUST retain TEPL Mode 0 Function 4 while executing on the SFU engine. Signed integer and floating modulo MUST use floor division so a nonzero result has the divisor sign; unsigned integer modulo MUST use unsigned remainder. Integer zero in the valid divisor rectangle MUST reject before effects. Shape defaults, PadValue, snapshots, and atomic publication MUST follow the closed binary Tile contract.
TREMS MUST consume one Local source, one per-participating-PE private-GPR scalar from optional B.IOR.RegSrc0, and one renamed Local destination, and MUST compute source mod scalar over the valid region. The selected DataType MUST satisfy TileA7DataTypeSupported; only PadValueOrByteId is applicable in B.DATR. Complete legality and allocation preflight MUST precede source/scalar snapshots and atomic payload, padding, status, and descriptor publication.
TROWARGMAX MUST select SFU Mode 2 Function 12 and MUST compute an increasing-column TMAX scan that retains the lowest winning index over one fully defined Local source. The source geometry MUST match the B.DIM-derived shape. The destination MUST use S32 or U32 indices and equal winners MUST retain the lowest index. Only PadValueOrByteId MAY be nondefault in B.DATR. Complete preflight and source snapshot MUST precede atomic result, status, padding, and descriptor publication.
TROWARGMIN MUST select SFU Mode 2 Function 13 and MUST compute an increasing-column TMIN scan that retains the lowest winning index over one fully defined Local source. The source geometry MUST match the B.DIM-derived shape. The destination MUST use S32 or U32 indices and equal winners MUST retain the lowest index. Only PadValueOrByteId MAY be nondefault in B.DATR. Complete preflight and source snapshot MUST precede atomic result, status, padding, and descriptor publication.
TROWEXPAND MUST copy the one-column source bit-for-bit across columns. The complete bundle MUST use one terminating Local B.IOT, MUST reject B.IOR and B.IOS, and MUST accept only the DataTypes listed by this owner. Complete preflight and source snapshot MUST precede atomic result, status, padding, and renamed-destination publication.
TROWEXPANDADD MUST add the full-shape source and one-column source across columns at the selected DataType. The complete bundle MUST use one terminating Local B.IOT, MUST reject B.IOR and B.IOS, and MUST accept only the DataTypes listed by this owner. Complete preflight and source snapshot MUST precede atomic result, status, padding, and renamed-destination publication.
TROWEXPANDDIV MUST divide the full-shape source and one-column source across columns at the selected DataType. The complete bundle MUST use one terminating Local B.IOT, MUST reject B.IOR and B.IOS, and MUST accept only the DataTypes listed by this owner. Complete preflight and source snapshot MUST precede atomic result, status, padding, and renamed-destination publication.
TROWEXPANDEXPDIF MUST subtract the one-column SrcDataType source from the full-shape source across columns and MUST exponentiate at DstDataType. The only legal (SrcDataType,DstDataType) pairs are (FP16,FP16), (BF16,BF16), (FP32,FP32), (FP16,FP32), and (BF16,FP32). BSTART selects SrcDataType; omitted B.DATR or DataType=DTYPE_NONE inherits it, while a concrete B.DATR DataType selects DstDataType. The complete bundle MUST use one terminating Local B.IOT, MUST reject B.IOR and B.IOS, and MUST reject every other source/destination pair. Complete preflight and source snapshot MUST precede atomic result, status, padding, and renamed-destination publication.
TROWEXPANDMAX MUST select the maximum of the full-shape source and the one-column source across columns at the selected DataType. The complete bundle MUST use one terminating Local B.IOT, MUST reject B.IOR and B.IOS, and MUST accept only the DataTypes listed by this owner. Complete preflight and source snapshot MUST precede atomic result, status, padding, and renamed-destination publication.
TROWEXPANDMIN MUST select the minimum of the full-shape source and the one-column source across columns at the selected DataType. The complete bundle MUST use one terminating Local B.IOT, MUST reject B.IOR and B.IOS, and MUST accept only the DataTypes listed by this owner. Complete preflight and source snapshot MUST precede atomic result, status, padding, and renamed-destination publication.
TROWEXPANDMUL MUST multiply the full-shape source and one-column source across columns at the selected DataType. The complete bundle MUST use one terminating Local B.IOT, MUST reject B.IOR and B.IOS, and MUST accept only the DataTypes listed by this owner. Complete preflight and source snapshot MUST precede atomic result, status, padding, and renamed-destination publication.
TROWEXPANDSUB MUST subtract the full-shape source and one-column source across columns at the selected DataType. The complete bundle MUST use one terminating Local B.IOT, MUST reject B.IOR and B.IOS, and MUST accept only the DataTypes listed by this owner. Complete preflight and source snapshot MUST precede atomic result, status, padding, and renamed-destination publication.
TROWMAX MUST select SFU Mode 2 Function 1 and MUST compute a typed increasing-column TMAX fold initialized from column zero over one fully defined Local source. The source geometry MUST match the B.DIM-derived shape. The destination MUST use the source DataType and integer arithmetic MUST remain at the selected element width. Only PadValueOrByteId MAY be nondefault in B.DATR. Complete preflight and source snapshot MUST precede atomic result, status, padding, and descriptor publication.
TROWMIN MUST select SFU Mode 2 Function 2 and MUST compute a typed increasing-column TMIN fold initialized from column zero over one fully defined Local source. The source geometry MUST match the B.DIM-derived shape. The destination MUST use the source DataType and integer arithmetic MUST remain at the selected element width. Only PadValueOrByteId MAY be nondefault in B.DATR. Complete preflight and source snapshot MUST precede atomic result, status, padding, and descriptor publication.
TROWPROD MUST select SFU Mode 2 Function 3 and MUST compute a typed increasing-column left fold from one using TMUL over one fully defined Local source. The source geometry MUST match the B.DIM-derived shape. The destination MUST use the source DataType and integer arithmetic MUST remain at the selected element width. Only PadValueOrByteId MAY be nondefault in B.DATR. Complete preflight and source snapshot MUST precede atomic result, status, padding, and descriptor publication.
TROWSUM MUST select SFU Mode 2 Function 0 and MUST compute a typed increasing-column left fold from zero using TADD over one fully defined Local source. The source geometry MUST match the B.DIM-derived shape. The destination MUST use the source DataType and integer arithmetic MUST remain at the selected element width. Only PadValueOrByteId MAY be nondefault in B.DATR. Complete preflight and source snapshot MUST precede atomic result, status, padding, and descriptor publication.
TRSQRT MUST compute one same-type reciprocal square root for every valid Local floating element. Either E4M3 signed zero MUST produce canonical quiet NaN 0x7F and MUST set DZ without setting OF.
TSCATTER MUST accept every non-packed B8-NP, B16, B32, or B64 value type with every S16, U16, S32, U32, S64, or U64 index type. Each index element is a logical destination-row selector at the current column. It MUST reject negative, out-of-range, or duplicate destination coordinates before effects, move raw value carrier bits without numeric-validity rejection, initialize the complete physical destination to typed zero, and publish all scattered writes atomically.
TSEL MUST consume predicate-kind storage with one low-first packed bit per logical element and MUST reject an ordinary numeric mask before effects. Both data sources and the destination MUST use one supported numeric type and matching geometry. Complete preflight and all source snapshots MUST precede exact selected-encoding and padding publication as one atomic result.
TSELS MUST read one packed predicate, one Local true source, and one private-GPR false scalar. Predicate bit one MUST copy the source encoding; bit zero MUST copy the scalar encoding. Complete preflight and all snapshots MUST precede atomic payload, padding, and descriptor publication.
TSHL MUST select VEC Mode 0 Function 9. For element width W, it MUST use the unsigned low log2(W) bits of source1 as the shift count and MUST store the low W bits of source0 shifted left, with zero carrier bits above W. Only the eight scalar integer Tile types are legal. PadValue, preflight, snapshots, and atomic publication MUST follow the closed binary contract.
TSHLS MUST consume one Local source, one per-participating-PE private-GPR scalar from optional B.IOR.RegSrc0, and one renamed Local destination, and MUST compute source << masked_count over the valid region. The selected DataType MUST satisfy TileI6DataTypeSupported; only PadValueOrByteId is applicable in B.DATR. Complete legality and allocation preflight MUST precede source/scalar snapshots and atomic payload, padding, status, and descriptor publication.
TSHR MUST select VEC Mode 0 Function 10. For element width W, it MUST use the unsigned low log2(W) bits of source1 as the shift count. Signed types MUST shift source0 arithmetically and unsigned types MUST shift logically; carrier bits above W MUST be zero. Only the eight scalar integer Tile types are legal. Preflight and publication MUST follow the closed binary contract.
TSHRS MUST consume one Local source, one per-participating-PE private-GPR scalar from optional B.IOR.RegSrc0, and one renamed Local destination, and MUST compute source >> masked_count over the valid region. The selected DataType MUST satisfy TileI6DataTypeSupported; only PadValueOrByteId is applicable in B.DATR. Complete legality and allocation preflight MUST precede source/scalar snapshots and atomic payload, padding, status, and descriptor publication.
TSHUF selects raw 32-bit words inside independent power-of-two row segments. Control mode, segment, and boundary fields are checked before publication; SELF and ZERO define out-of-segment behavior without predicates.
TSORT MUST select SFU Mode 3 Function 12 and MUST accept exactly FP32, FP16, and BF16. It MUST stably sort each independent valid-row group, MUST return the matching original within-group U32 indices, MUST place numeric values before NaNs in either direction, and MUST set numeric invalid status when a signaling NaN is observed. Omitted or zero LB0 MUST select width 32.
TSQRT MUST retain TEPL Mode 0 Function 21 while executing on SFU. It MUST consume one persistent Local floating source and one renamed same-type destination, preserve signed zero, preserve positive infinity where the format provides it, and report invalid with quiet NaN for negative nonzero inputs. Type, dimension, PadValue, preflight, snapshot, status, padding, and atomic-publication rules form one closed operation.
CUBE TSTORE MUST require an exact persistent Matrix descriptor selected by M322ND, M162ND, or N82ND, MUST preflight its full LB1 by LB0 GM footprint, and MUST store only valid elements without modifying source or padding.
TSTORE MUST use B.IOR RegSrc1 as a byte row stride, MUST snapshot and preserve its Local or Shared source, MUST preflight the complete selected GM footprint before any store, and MUST distinguish an omitted dense byte stride from an explicitly encoded zero stride. Function 1 Shared access accepts any nonzero participating PE mask, and mask zero MUST have no architectural effect.
TSUB MUST select VEC Mode 0 Function 1 and MUST subtract the ordered right Local source from the ordered left Local source. The closed binary Tile schema, shape defaults, arithmetic DataType set, PadValue behavior, complete preflight, source snapshot, and atomic destination publication MUST match the TADD family contract. Rejection MUST precede all effects.
TSUBS MUST consume one Local source, one per-participating-PE private-GPR scalar from optional B.IOR.RegSrc0, and one renamed Local destination, and MUST compute source - scalar over the valid region. The selected DataType MUST satisfy TileA9DataTypeSupported; only PadValueOrByteId is applicable in B.DATR. Complete legality and allocation preflight MUST precede source/scalar snapshots and atomic payload, padding, status, and descriptor publication.
TTRI MUST select SFU Mode 3 Function 7. It MUST publish one newly allocated row-major Local FP32, FP16, S32, S16, U32, or U16 destination. Lower orientation MUST write typed one when c <= r+diagonal; upper orientation MUST write typed one when c >= r+diagonal. Every other valid element MUST be typed zero, and the signed boundary comparison MUST NOT wrap.
TUNPACK extracts one contiguous byte field from each U32 CUBE word and zero-extends it to U32. Offset/count controls are checked before effects.
TXOR MUST select VEC Mode 0 Function 8 and MUST compute the bitwise XOR of corresponding integer elements. It MUST accept only S64/S32/S16/S8 and U64/U32/U16/U8. PadValue, shape defaults, source snapshots, and atomic destination publication MUST follow the closed Local binary Tile contract.
TXORS MUST consume one Local source, one per-participating-PE private-GPR scalar from optional B.IOR.RegSrc0, and one renamed Local destination, and MUST compute source XOR scalar over the valid region. The selected DataType MUST satisfy TileVecScalarIntegerDataTypeSupported; only PadValueOrByteId is applicable in B.DATR. Complete legality and allocation preflight MUST precede source/scalar snapshots and atomic payload, padding, status, and descriptor publication.
UCVTF MUST implement the mnemonic-local canonical assembly, encoded legality, defaults, state and memory effects, ordering, and fault boundaries declared in this owner. The operation region below is the executable binding for every accepted decision that names this mnemonic.
Decisions: ADR-0026. XOR MUST snapshot its scalar sources, apply its mnemonic-owned width, immediate, modifier, and wrapping rule, then publish through the assigned destination or commit effect in alias-safe order.
Decisions: ADR-0026. XORI MUST snapshot its scalar sources, apply its mnemonic-owned width, immediate, modifier, and wrapping rule, then publish through the assigned destination or commit effect in alias-safe order.
Decisions: ADR-0026. XORIW MUST snapshot its scalar sources, apply its mnemonic-owned width, immediate, modifier, and wrapping rule, then publish through the assigned destination or commit effect in alias-safe order.
Decisions: ADR-0026. XORW MUST snapshot its scalar sources, apply its mnemonic-owned width, immediate, modifier, and wrapping rule, then publish through the assigned destination or commit effect in alias-safe order.