Cranelift: formalize, document, and audit binary representation of booleans

[cfallin]

In #3202 and a number of past issues as well, we've repeatedly encountered some uncertainty as to (i) how boolean values of varying widths are stored in registers, (ii) whether they can be loaded/stored or not and what representation they take in memory (see e.g. #3102 where we addressed this for trampoline args), and (iii) how this interacts with boolean<->integer conversion ops like bint and bmask.

The understanding we seem to generally agree on is that booleans of width > 1 bit have all bits set (true) or all bits cleared (false). And a b1 is stored in a register with the same invariant as other integer values: the upper bits (in this case everything above the LSB) are undefined.

However, this is far from being confusion-free and unambiguous; thus, it seems reasonable to (i) agree definitively on how bools are represented in registers and memory, and (ii) audit our compliance to whatever invariants we decide on.

[cfallin]

cc @abrown @afonso360 @bjorn3 @alexcrichton @akirilov-arm for recent folks who've discussed this topic.

[cfallin]

I'd like to propose the following for loads and stores:

• Define loads and stores of boolean values. We haven't done this in the past, but it is a somewhat incongruous hole in our semantics, IMHO, and we do load and store booleans sometimes behind the scenes (see: trampoline args). I think it would be better if we formally say what it means to do so.
• Let's define the size of a b1 to be one byte, and others (b8 and up) as the size of the equivalent integer type.
• A store of a b1 stores 0 or 1 (i.e., it is zero-extended to 8 bits); a store of a b8..b128 stores all zeroes or all ones.
• A load of a boolean type has undefined behavior if the value loaded from memory is not canonical as per the above bullet point. better option, see below

The last is probably the most controversial; disallowing loads and stores of bools avoids the semantic question, which may have been why this was done (@sunfishcode comments?). But we already allow bitcasts, so we already have some notion of the bits underlying a bool, so IMHO it's better to allow loading/storing those bits.

An alternative is to canonicalize a boolean when loading. For a b1 this is a mask (and op); for a b8 and up we could just compare to zero and if the loaded value is nonzero (any bit is set), canonicalize to true (all bits set). I actually prefer this option, after a bit of thought: it gives well-defined semantics and avoids illegal bool values (e.g. b128 with half the bits set) that might cause other undefined behavior.

Then for bitcasts:

• A raw_bitcast or bitcast from a bool to an int produces the same value that a store would produce.
• A raw_bitcast or bitcast from an int to a bool canonicalizes just as a load does.

All of the above should fully pin down how we store the values in registers as well: we (i) have a well-defined integer interpretation of the bool, (ii) always store it in canonicalized form (all zeroes or all ones in defined bits), and (iii) keep to the same "upper bits undefined" invariant as for integers. With one exception: for a b1 type I think that we should maintain the lowest 8 bits defined, i.e. lowest byte of 0x00 or 0x01 only, so that a store can be done without a mask.

Thoughts?

[cfallin]

I've just added this to the agenda for Monday's public Cranelift biweekly meeting as well.

[abrown]

I think a side effect of going doing the path above would be that raw_bitcast is no longer necessary. bitcast says: "A bitcast is equivalent to storing one type and loading the other type from the same address." If we can load and store booleans I think that eliminates the reason for raw_bitcast's existence (?).

[cfallin]

That's a good point; the distinction between the two is pretty confusing, so if we could remove raw_bitcast as well and have only one kind of bitcast, all the better.

[fitzgen]

A store of a b1 stores 0 or 1 (i.e., it is zero-extended to 8 bits); a store of a b8..b128 stores all zeroes or all ones.

Why not have b1 store all ones? The inconsistency between b1 and b8..b128 is a bit of a surprise.

[cfallin]

Why not have b1 store all ones?

So my thought process is: b1 is, semantically, one bit wide; so "all ones" is just 0x01. But because memory is byte-oriented, we have to store a whole byte, so we zero-extend. I definitely recognize this is kind of subjective and "all ones in all bits that are written" is just as much a valid perspective though.

Pragmatic reasons: b1 will be most common, and generating a zero or one is slightly easier than zero or -1 (all ones), e.g. the one instruction cset on aarch64; and bint (bool to int) from a b1 with a zero-or-one value is a no-op (just a copy) while it's a masking operation with zero or -1.

[cfallin]

I should probably add another possibility too: we could define all of b1..b128 to use 0 for false and 1 for true.

The main reason for the all-ones representation, to my understanding, was to be able to use e.g. a b128 as a SIMD mask after a raw_bitcast (or at least, that's the way it has been explained to me -- many issues back I think someone linked some code that had assumed this). But this is sort of getting a desired higher-level semantic behavior implicitly by setting up the low-level definitions in a certain way; it would be better to require the "turn this bool into an all-ones mask" use-case to use bmask, which already exists for this purpose. Then we get to use the (IMHO) slightly more intuitive 0/1 values which work well for most cases, and we have a fully consistent behavior. We would just need to look for use-cases that assume all-ones currently and ensure they use bmask instead.

This probably needs a sanity check from a SIMD perspective though -- @abrown / @jlb6740 / @akirilov-arm, thoughts?

[cfallin]

Ah, here is one example of a dependence on all-ones in cg_clif: link (linked by @afonso360 in #3003)

[abrown]

I would mention that the all-ones/all-zeroes representation for SIMD is not just because using the output of a comparison as a mask is convenient, it's also that the Wasm SIMD spec specifies that representation (but obviously without the boolean type) and that machines (x64, aarch64) represent things this way as well.

[cfallin]

@abrown this has me thinking further: the existing Wasm-SIMD implementation (lowering in cranelift-wasm specifically) does not use b128 (or any other boolean type), right?

So one fundamental question I have is: what is the use-case for boolean types wider than 1 bit? If for an actual SIMD use-case we represent the mask as an I8X16 (that's what I gather from here and the implementation of optionally_bitcast_vector), then it seems that the question of what Wasm-SIMD masks require is separate from the question of what b128 means.

In other words, what we decide here doesn't impact Wasm-SIMD, because Wasm-SIMD doesn't depend on CLIF-level bool types; so it's better to base on decision on whatever makes the most semantic sense and leads to the least confusion, IMHO.

If b128 is just "a 1-bit-narrow value stored in a much wider space", then IMHO it makes as much sense to zero-extend it (as one would zero-extend an i8 stored in an i128 worth of space) as sign-extend it.

Argued from another perspective: If we say that booleans (abstractly) have two possible values, then storing such a value in anything wider than 1 bit grants us the freedom to choose the bit representation for each of those two possible values. In other words, b128 is still just semantically true or false; the bit-patterns that represent those are invisible at the CLIF-semantics level.

So I'm starting to convince myself a bit more than the 0/1 approach is cleanest, and then we need to look for uses of raw_bitcast (as in cg_clif) to peek under the hood and see the actual bits, and convert those uses into bmask or bint as appropriate. But I'm (still) very curious how that sits with others...