What's new
What has been added and fixed, newest first. Follow in a feed reader
-
Four PowerISA instructions under their real names
The external-PID loads and stores of the embedded PowerPC cores had pages called lbzepx, lhzepx, lwzepx and sthep, names no assembler or manual uses, and wrong encodings. They are now lbepx, lhepx, lwepx and sthepx, as binutils and Power ISA 2.07B call them, with the encodings both give. Links to the old pages lead to the new ones and searching an old name still finds the instruction. In the API each record moved to the file of its new name, such as api/v1/powerisa/lbepx.json, and the old files are gone.
See it -
Choose A64, A32 or T32 when decoding an ARM word
A word from one ARM instruction set can match another's pattern: Arm's own encoding diagrams show 865 encodings decoding to a different instruction when the set is not known, 584 of them Thumb. Each ARM entry now says which set it is, the decoder matches within the one you choose, and a word that matches nothing there is checked against the other two. With the right set chosen the figure is 24 of 6,852.
See it -
Every ARM encoding checked against Arm's own diagrams
All 1,145 ARM bit patterns were compared with the encoding diagrams in Arm's specification. 1,096 agree exactly, 13 were wrong and are fixed (AES and SHA1 crypto, dup, cpy, SVE dot products), and 23 cannot be checked against this release. The decoder's field widths now come from each record's own positions, which corrected three entries that were decoded with a misplaced mask. The same diagrams also show that the ARM decoder cannot yet tell A64, A32 and T32 words apart, and the decoder page now says so.
See it -
Search an AMD GPU instruction by either of its names
AMD gives some instructions a second name, often a rename between GPU generations: v_ffbh_i32 is v_cls_i32 in newer code. Each page now says "also written as", searching either name finds it, and a name with no page of its own redirects to the one that has. The sidebar also shows only the instructions around the current one in a long category, with the filter still searching the whole category, so AMD GPU pages are about a third of their weight.
See it -
482 more AMD GPU instructions
Instructions that AMD's own specification defines and this site had no page for are now here: RDNA 3's dual-issue v_dual pairs, the WMMA and SWMMAC matrix operations, tensor loads, export, and names older generations use, such as buffer_load_dwordx2 and v_add_nc_u32. Each page has AMD's description, its operands, and a bit layout where AMD's own identifiers confirm one. The decoder now reads 1,998 more of LLVM's assembled test instructions correctly, and still none wrongly.
See it -
A JSON API for every instruction
Look up any instruction as JSON from one URL, for example /api/v1/x86/add.json, with no key and no signup. Each lookup returns every form of the mnemonic, the records are the dataset's own public-domain ones, and a JSON Schema describes them. Examples for the command line, JavaScript, Python, an AI agent tool and a VS Code hover.
See it -
New note: what a zero is worth
How the AMD GPU decoder was checked against LLVM's assembler, how far its zero wrong answers reach in each GPU generation, and why most instructions it sees get no answer rather than a guess.
See it -
Decode AMD GPU instructions
The decoder now reads 32, 64 and 96-bit AMDGPU instructions for a GPU generation you choose, with every field broken out and register numbers where they can be read reliably. Paste bytes as LLVM prints them or dwords as llvm-objdump does. Checked against LLVM's own assembler tests: 25,923 instructions decoded, none to the wrong instruction.
See it -
Search from your browser's address bar
Add instructionsets.com as a search engine and type an instruction straight into the address bar. A single exact match opens its page, and anything else is listed in full. Pressing Enter in any search box on the site now shows every result instead of only the first ten.
See it -
This page, and a feed to follow it
Changes to the site are listed here and published as an RSS feed, so a feed reader can tell you when something is added or fixed.
See it -
PowerISA extended mnemonics show every form
69 PowerISA pages printed blank cells in their Extended Mnemonics table. Every form is shown now, and where the source gives no equivalent instruction the page says so rather than leaving a gap.
See it -
New note: measuring the reference instead of the data
How the accuracy of this reference was measured against its own vendor specifications, the 18 wrong RISC-V encodings that turned up, and the four times the audit itself was the problem.
See it -
A way to reach us from every page
Report an error, ask a question or suggest something from the Contact page. The form opens your own email app with the message ready to send.
See it -
Comparison destinations, and bit diagrams that show every field
AMDGPU comparisons (v_cmp_) now list the vcc destination they write. Bit diagrams wrap instead of scrolling sideways, which had hidden part of most ARM diagrams on a desktop screen.
See it -
Bit layouts for 1,514 AMDGPU instructions
Taken from AMD's machine-readable ISA specification, each checked by rebuilding one of AMD's own encoding identifiers from the instruction's opcode. Opcodes change between GPU generations, so each diagram names its generation and says where the others differ.
See it -
Operands for 1,635 AMDGPU instructions
Named and in order, with whether each is read or written, from AMD's machine-readable ISA specification. AMDGPU pages with no operand list went from 910 to 142.
See it -
Pages that fit a phone, and a search shortcut
Instruction pages no longer run wider than a phone screen, wide tables scroll inside their panel, Ctrl+K or / jumps to search, and table headers sort from the keyboard.
See it -
28 new cross-architecture comparisons
19 comparisons of how x86, ARM, RISC-V and PowerISA each express the same operation, including where one has no instruction for it, and 9 of NVIDIA PTX against AMDGPU.
See it -
Instruction word decoder
Paste a 32-bit ARM or RISC-V instruction word and get the instruction with every field broken out. It carries only the 908 encodings whose bit pattern rebuilds their own opcode, so it declines rather than guesses.
See it -
The whole dataset, in the public domain
Every instruction record on the site, as JSON and JSONL under CC0. Use it for anything: no permission and no attribution needed.
See it -
Accuracy measured against the vendor manuals
A published audit of how well the data agrees with the vendor specifications. It found 18 wrong RISC-V encodings, which were fixed.
See it