Instruction Predecoding and Decoded Instruction Caches

Mark Smotherman
Last updated: May 2025

UNDER CONSTRUCTION

Summary: ... tbd...

... intro to be written ... incomplete - lots of ideas and patents, ... will try to include the early ones ...

• decoded instruction queues
  • IBM Stretch (1961) - predecoded up to two instructions at a time after instruction fetch, held predecoded (and some pre-executed) instructions in its four-level lookahead unit

  • IBM S/370 M165 (1971) and 3033 (1978) - four-entry decoded instruction queue between IPPF ("instruction pre-processing function") and execution units; IPPF starts the fetching of memory operands, similar to IBM Stretch

• decoded instruction caches
  • LLNL S-1 (1978) - expanded 36-bit word to 56-bit icache format to reduce inst. decoding time

  • Jim Pomerene and Rudolph Rechtschaffen, "Cache memory architecture with decoding" US Patent 4,437,149 (filed Nov. 1980, granted to IBM March 1984)

  • Dave Patterson, RISC-II (ISCA 1983 paper) - "on the miss" expansion between memory and icache

  • CCI Power 6/32 (1984)
    • variable-length instuction set similar to the VAX in which the length of an instruction cannot be determined by the opcode
    • also sold as the Harris HCX-5, -7, and -9; the ICL Clan 5, 6, and 7; and, the Unisys 7000/40
    • A 1998 TUHS post by Eric Edwards says that it was implemented on five boards using AMD 2900 series bit slice processors, PLAs, and 74F series parts. A 1987 comp.arch post by "bjj" from Penn State says that the 4K-word instruction cache contained decoded instructions stored as fixed-length 73-bit words. Kevin McKean in a 1986 two-part series in UNIX/WORLD states that with the decoded instruction cache and memory interleaving turned on, the 10 MHz Power 6/32 was 11x faster than a 5 MHz VAX 11/780 on a call-and-return-intensive benchmark.

  • Yale Patt, HPS (1985)

  • AT&T CRISP (1987) - variable-length source instructions were predecoded into fixed-length entries and placed in a 32-entry DIC; also each branch wwas folded into the previous instruction's DIC entry by including a next-address field; conditional branches were handled by including a second, alternate next-address field and information for determining a misprediction

  • IBM RS/6000 (1989) - instructions are predecoded into eight general classes to assist in routing to the function units (see also US Patents 5,828,895, filed 1995, and 6,286,094, filed 1999)

  • ... lots recently ... (perhaps discuss MIPS R10000)

• instruction caches with instruction length info added

  • ... early patents ...

  • AMD K5 (1994) - adds 5 bits per instruction byte (start, end, prefix, opcode, number of Rops); see US Patent 5,758,114 (parent filed 1995, granted to AMD in 1998); only 2 bits added per byte in K8

  • ...

• instruction caches with scheduling information added
  • NS Swordfish (1991) - instruction pair dependency bit is contained in each decoded i-cache entry; it is set on i-cache refill by predecode hardware and yields LIW issue of independent instruction pairs; no bits are used in the normal instruction format; see US Patent 5,669,011 (parent filed 1990, granted to NS 1997)

  • Minagawa/Saito/Aikawa (1991) - "Pre-decoding mechanism for superscalar architecture," IEEE Pacific Rim Conf. on Comm., Comp., and Sig. Proc., pp. 22-24; on i-cache miss, a predecoder adds instruction grouping ("priority") and function unit assignment fields; see US Patent 5,377,339 (parent filed 1991, granted 1994)

  • ...

• trace caches
  • Alex Peleg and Uri Weiser, "Dynamic flow instruction cache memory organized around trace segments independent of virtual address line," US Patent 5,381,533 (parent filed 1992, granted to Intel 1995)

  • ...

• ...

(US Patents - search subclass 213 under class 712)

[History page] [Mark's homepage] [CPSC homepage] [Clemson Univ. homepage]

mark@cs.clemson.edu