Skip to main content

ADR 0076: Block scalar and tile bindings

Context​

ADR 0062 recorded a single repository-wide mnemonic audit. This record preserves the accepted decisions for this family as one decision-scoped owner. The former identifiers remain only in legacy_ids and the generated ADR index; current normative meaning is owned by the affected ASL/NDF clauses.

Decisions​

Decision 018: B.IOR fields name absolute GPRs and encoded zero is zero​

Each B.IOR source and destination field accepts exactly the twenty-four absolute GPR selectors 0..23. Selectors 24..31 are reserved in B.IOR and MUST reject before effects; relative queue selectors are not legal.

Encoded selector zero names the architectural zero register. Canonical assembly and disassembly use zero, sp, a0..a7, ra, s0..s8, and x0..x3 for selectors 0..23 respectively. Numeric register names MAY be accepted as input aliases, but canonical output MUST use those ABI names.

Decision 019: the complete block schema determines B.IOR presence and arity​

The effective operation schema is selected after the complete block header has been assembled. That schema determines which of RegSrc0..RegSrc2 and RegDst are consumed. An omitted B.IOR supplies the operation-defined defaults for every consumed field. An explicitly encoded zero in a consumed field supplies the zero register and is not omission.

Fields not consumed by the selected schema MUST be encoded as zero. A nonzero unused field MUST reject before block effects. Disassembly MUST preserve the distinction between an omitted instruction and an explicitly encoded all-zero B.IOR so that reassembly preserves the exact instruction stream.

Decision 020: PTO blocks admit at most one B.IOR and permit register aliasing​

A PTO block MAY contain at most one B.IOR. Encountering a second B.IOR in the same block MUST raise Illegal Block Exception before changing pending or architectural state, and MUST preserve the first binding.

Source selectors MAY repeat. A source and destination MAY name the same GPR. Such aliasing does not itself make the block illegal; the selected operation defines when inputs are read and when an output becomes visible.

Decision 021: B.IOR is available to every schema that declares GPR operands​

B.IOR is not restricted to a separated block. It MAY appear in any active BSTART block whose effective operation schema declares GPR input or output fields. If the selected schema consumes no GPR field, an omitted or explicitly all-zero B.IOR supplies the default state, while any nonzero binding MUST reject before commit.

Decision 022: TLOAD and TSTORE assign the first two B.IOR sources​

For TLOAD and TSTORE, RegSrc0 supplies the per-PE global-memory base address and RegSrc1 supplies the row stride in bytes. Each participating PE reads the selected selectors from its own GPR file.

When B.IOR is omitted, the base defaults to zero and the row stride defaults to ceil(resolved_columns * element_bits / 8), producing dense rows. An explicitly encoded zero selector supplies a zero base or zero stride and MUST NOT select the omission default. ADR 0074 owns the byte-address formula.

Decision 023: B.IOT has exactly five Local-Tile forms​

B.IOT has exactly the five accepted forms that bind zero, one, or two ordered Local Tile sources and zero or one Local Tile destination. It has no Shared destination form, no reuse field, and no legacy four-bit size field.

Each six-bit source selector names a relative Local Tile queue entry: 0..15 select T#1..T#16, 16..31 select U#1..U#16, 32..47 select M#1..M#16, and 48..63 select N#1..N#16. Source operands are consumed in the encoded program order.

On a destination form, DstTile=0..3 selects the T, U, M, or N destination hand respectively. It does not expose a physical Tile register. TSize=1..7 declares 128 B through 8 KiB of capacity per participating PE; encoded TSize=0 is reserved on every destination form.

Decision 024: B.IOT.PE_MASK is a four-PE predicate​

Every four-bit PE_MASK value is assigned and multiple set bits are legal. All effective Local Tile bindings in one block MUST use the same nonzero mask, and any Local/Shared operands composed by that block MUST use the same mask. An operation MAY impose a stricter mask requirement as part of its own schema.

PE_MASK=0000 is a strict no-op. It MUST NOT add a binding, read a source, allocate or rename a destination, update a descriptor, produce a fault, or change the B.IOT sequence-termination state.

Decision 025: B.IOT.L terminates only the binding sequence​

The encoded L bit is the last marker for the current block's effective B.IOT sequence. It is not a source-lifetime or source-release control.

Every block that requires an effective B.IOT sequence MUST contain exactly one nonzero-mask binding with last=1, and that binding MUST be the final effective B.IOT in the block. Ending the block without that marker, placing another effective B.IOT after it, or placing more than one effective marker MUST raise Illegal Block Exception before the offending instruction or commit changes architectural or pending block state.

Decision 026: B.IOT sources persist and destinations are renamed​

Reading a Local Tile source through B.IOT MUST NOT modify, release, or invalidate its payload or descriptor. Successful block completion therefore does not imply source lifetime termination, regardless of the last bit.

A destination form selects a destination hand and requests a new allocation. Hardware MUST rename that request to a Local Tile register in the selected hand, then atomically publish the new payload and descriptor at successful block commit. Consumers refer to the renamed result through the architectural Local Tile queue model; they do not identify the physical allocation by reusing the producer's PE_MASK.

Decision 027: B.IOS binds absolute Core-private Shared registers​

Each core contains one architectural bank of 256 persistent Shared Tile registers, named S0..S255. All four PEs in that core observe the same bank; another core observes a different bank. SharedTID=0..255 is an absolute index, so encoded zero names S0 and does not mean omission.

B.IOS Sx, mask=PE_MASK is the source form and encodes TSize=0. B.IOS mask=PE_MASK, ->Sx<TSize> is the destination form and requires TSize=1..7, encoding 128 B through 8 KiB per participating PE. The role encoded by TSize MUST agree with the selected operation schema.

Decision 028: B.IOS uses an ordered four-entry binding stream​

One effective B.IOS adds one ordered Shared operand binding. A block MAY contain at most four effective Shared bindings. Two unconsumed bindings in the same block MUST NOT name the same Sx; a duplicate or fifth effective binding MUST raise Illegal Block Exception before changing the binding stream or other pending state.

PE_MASK is a four-PE predicate and multiple bits are legal. Every effective Shared binding and every composed Local binding in one block MUST use the same nonzero mask unless the selected operation defines a stricter rule. PE_MASK=0000 is a strict no-op: it adds no binding, performs no schema or duplicate check, reads no source, allocates nothing, updates no descriptor, accesses no memory, and produces no fault.

Decision 029: Shared destinations atomically update persistent state​

Each allocated Sx retains one per-PE Tile descriptor, a fixed allocation mask, an initialized-quarter mask, and persistent payload. The first nonzero destination write establishes its descriptor and allocation mask. A later destination write MAY update any subset of that allocation mask only when the descriptor is compatible; it MUST NOT expand the mask or silently replace an incompatible descriptor. Software MUST allocate another Sx for a different mask or descriptor.

A successful Shared destination operation atomically publishes the compatible descriptor update and every selected fixed-offset payload quarter. An observer MUST see either the old complete state or the new complete state, never a torn descriptor/payload update. A Shared source read does not modify the descriptor, payload, allocation mask, or initialized mask.

The architecture imposes no ordering between conflicting accesses from different PEs. Programs MUST avoid conflicting offsets or establish ordering with separate synchronization mechanisms.

Decision 030: an uninitialized Shared source reads as undefined state​

Reading an Sx with no allocated descriptor is legal. The consumer operation MUST derive a read-only temporary descriptor from its own completed schema, including its data attributes, dimensions, and any Local counterpart required by that operation. Every selected payload element is an undefined-register value.

This read MUST NOT allocate or initialize the Shared register, write back the temporary descriptor, update any Shared state, or raise an exception solely because the descriptor or selected quarter was uninitialized. Reading an allocated but uninitialized quarter follows the same undefined-register rule for that quarter.

Last updated on by Kevin Zhou