Alex Crichton (Aug 16 2024 at 18:00): Alex Crichton (Aug 16 2024 at 18:00):I'm curious to explore some 128-bit stuff a bit and right now I'm looking at input such as:
function u0:0(i64, i64) -> i64, i64 tail {
block0(v0: i64, v1: i64):
v10 = load.i64 v0
v18 = uextend.i128 v1
v20 = uextend.i128 v10
v14 = iadd v18, v20
v15, v16 = isplit v14
jump block1
block1:
return v15, v16
}
which after https://github.com/bytecodealliance/wasmtime/pull/9136 produces:
0: 55 pushq %rbp
1: 48 89 e5 movq %rsp, %rbp
4: 4c 8b 0f movq (%rdi), %r9
7: 48 31 c9 xorq %rcx, %rcx
a: 48 89 f0 movq %rsi, %rax
d: 4c 01 c8 addq %r9, %rax
10: 48 83 d1 00 adcq $0, %rcx
14: 48 89 ec movq %rbp, %rsp
17: 5d popq %rbp
18: c3 retq
notably what I'm trying to get happen is to sink the movq (%rdi), %r9 into addq %r9, %rax. Currently the sinking is being prevented by this line but I don't fully understand the comment there. It seems like if this heuristic is needed then it's probably a conservative one because load sinking should be ok in this specific scenario at least.
Which leads me to my question: could this force_multiple be explained a bit more? Is there an example of how this could go wrong if force_multiple were ignored? (if I set it to false unconditionally no filetests break so I wasn't able to learn more via that). Furthermore is there a way to perhaps make this check less conservative to allow load sinking in the above case?
fitzgen (he/him) (Aug 16 2024 at 18:03):blame says it is a fix for https://github.com/bytecodealliance/wasmtime/issues/3953
fitzgen (he/him) (Aug 16 2024 at 18:03):cc @Chris Fallin @Andrew Brown
Chris Fallin (Aug 16 2024 at 18:04):yep, that; the case you show in CLIF above is a special case where a single instruction is responsible for both of the uses, in theory we could detect that, we'd want to think it through very carefully though
Chris Fallin (Aug 16 2024 at 18:05):(nit: not "heuristic" but "invariant" or "rule"; heuristics can be violated without correctness issues)
Alex Crichton (Aug 16 2024 at 18:08):hm ok, I still don't really understand what this is doing but sounds like I should assume this can't easily change
Chris Fallin (Aug 16 2024 at 18:09):so the basic idea of the analysis is to prove that an instruction is used only (or at most) once; that's the condition we need to sink it
Chris Fallin (Aug 16 2024 at 18:09):(we could do better if we reasoned about location and sunk it to the first of several locations it's used if it dominates all the others, but that's way more complex)
Chris Fallin (Aug 16 2024 at 18:10):when there is a single output, this is reasonably straightforward: we do a transitive fixpoint algorithm where we generate a "Multiple" at any op that's used multiple times, and propagate that backward
Chris Fallin (Aug 16 2024 at 18:10):(because matching/sinking can go arbitrarily high up the tree, we need a transitive notion of "used more than once")
Chris Fallin (Aug 16 2024 at 18:10):when there are multiple outputs, even if each output is used only once, the op itself is used more than once
Chris Fallin (Aug 16 2024 at 18:11):we could refine the lattice more and encode logic for 0+0 -> 0, 0+1 / 1+0 -> 1, 1+1 -> many, 1+many / many+1 -> many
Chris Fallin (Aug 16 2024 at 18:11):and then add a special case for 1+1 (if same 1) -> 1
Chris Fallin (Aug 16 2024 at 18:11):but that's... a lot of subtlety and I don't have confidence enough in it to do it right now
Chris Fallin (Aug 16 2024 at 18:11):hopefully makes more sense?
Alex Crichton (Aug 16 2024 at 18:12):One part I don't understand though is why multiple-results matter here. ISLE can't match instructions with multiple results in patterns afaik so we can't possibly go up the tree and somehow duplicate the operands of a multi-result instruction, right?
Chris Fallin (Aug 16 2024 at 18:13):well, when building the analysis we try not to couple current implementation limits such as "ISLE can't match instructions with multiple results" with the invariants elsewhere; that could easily lead to very subtle bugs later when we expand its capability, and in any case leads to very hard to reason about invariants
Alex Crichton (Aug 16 2024 at 18:13):the specific case of isplit returning two things and then ret consuming both is a bit of a red herring as it was the first test case I came up with, the actual benchmark I think stores the two results of isplit into memory or just one or something like that
Chris Fallin (Aug 16 2024 at 18:13):I think the right fix here, if we want to make the analysis more powerful, is to do as noted above -- recognize the special case
Chris Fallin (Aug 16 2024 at 18:13):or alternately, have a notion of "rooted value" that stops propagation of multiplicity
Chris Fallin (Aug 16 2024 at 18:14):then we need to guarantee that this is never matched through
Chris Fallin (Aug 16 2024 at 18:15):the key there would be to have a single source of truth for that notion/kind of instruction and consult it in both places
Alex Crichton (Aug 16 2024 at 18:17):I'm still not understanding though because isn't it structurally not possible to match through a multi-result instruction?
Alex Crichton (Aug 16 2024 at 18:17):in that aren't we already guaranteed that doesn't happen?
Chris Fallin (Aug 16 2024 at 18:18):it currently isn't possible, but let's not bake in that assumption, is what I'm saying; it's an implementation limit not something fundamental
Chris Fallin (Aug 16 2024 at 18:18):the "rooted instruction" notion above is my attempt to reify it as an actual concept we can rely on :-)
Chris Fallin (Aug 16 2024 at 18:18):said another way: can we describe accurately what the result of the analysis is?
Alex Crichton (Aug 16 2024 at 18:19):ok yeah that makes sense, I'm just trying to affirm my current understanding of the world to understand what you're saying we want to move towards
Alex Crichton (Aug 16 2024 at 18:19):e.g. I don't know what you mean by "rooted instruction"
Chris Fallin (Aug 16 2024 at 18:19):if the answer has to invoke implementation limits and just-so arguments why it won't cause an issue, that's a very bad place to be in
Chris Fallin (Aug 16 2024 at 18:19):vs. having a semantic concept and defining things around it
Alex Crichton (Aug 16 2024 at 18:19):and it's difficult for me to reason about matching through things when that's currently just a hypothetical for multi-result instructions
Chris Fallin (Aug 16 2024 at 18:20):"rooted instruction" is a concept that reifies the notion of "thing that cannot be matched through"
Chris Fallin (Aug 16 2024 at 18:20):rooted because we need to actually codegen it at the root of a matching tree
Chris Fallin (Aug 16 2024 at 18:20):once we have that guarantee, semantically, we can define multiplicity in terms of it
Chris Fallin (Aug 16 2024 at 18:20):basically my whole thing here is: we need to define multiplicity succinctly and clearly so we can think about it. if our definition includes reasoning about matching/lowering algorithm stuff, we've failed
Chris Fallin (Aug 16 2024 at 18:21):(in the sense that, that's a recipe for muddy thinking and further bugs)
Alex Crichton (Aug 16 2024 at 18:21):sorry I'm also trying to catch up because it sounds like you're trying to rebut a change I'm proposing but I'm not proposing anything and am trying to better understand what we have today
Chris Fallin (Aug 16 2024 at 18:22):basically I'm justifying why we have what we have today against the "why can't we remove force_multiple"
Chris Fallin (Aug 16 2024 at 18:24):as succinctly as I can put it: multiplicity has been defined (in this analysis) to mean "is an instruction referenced multiple times, or referenced by another instruction that is referenced multiple times". The algorithm approximates that in the presence of multiple-output instructions by the use of the force_multiple mechanism
Chris Fallin (Aug 16 2024 at 18:24):what I'm then proposing is that we can change that definition to change the answer to your question ("why can't we remove" -> "now we can") by introducing this notion of rooted instructions
Alex Crichton (Aug 16 2024 at 18:27):ok that makes sense to me yeah, thanks!
Chris Fallin (Aug 16 2024 at 18:27):cool. happy to review if you want to take a stab at this (I unfortunately don't have the time or mental energy atm to do it myself!)
Alex Crichton (Aug 16 2024 at 18:27):is this function, get_value_as_source_or_const, a suitable "choke point" where the concept of a rooted instruction might be able to be inserted?
Alex Crichton (Aug 16 2024 at 18:28):I see that for side-effecting instructions it already ensures there's only one output but I suspect more would be needed
Chris Fallin (Aug 16 2024 at 18:29):yep, that's indeed where we have all the conditions for "seeing through" a use, so let's add conditions there (for pure and impure cases) for "is rooted" and avoid seeing it as any more than an opaque value if so
Alex Crichton (Aug 16 2024 at 18:29):how come it would be necessary for the "pure" case?
Alex Crichton (Aug 16 2024 at 18:30):oh wait let me try to answer myself
Alex Crichton (Aug 16 2024 at 18:30):let's say that you v0 = load ... ; v1, v2 = isplit v0 - if you "look through" the isplit, despite it being pure, you could "look through" a use of v1 and v2 and possibly generate two loads
Alex Crichton (Aug 16 2024 at 18:30):is that right?
Chris Fallin (Aug 16 2024 at 18:30):yep exactly
Chris Fallin (Aug 16 2024 at 18:31):so we cut things at the isplit, it's "rooted" now, it becomes just a single use for the load, and we can't look through it
Chris Fallin (Aug 16 2024 at 18:33):(I'm not sure if that addresses your original desire to get more load-sinking; the loads you want sunk are actually single-use and the splits happen as consumers of the things like iadds that consume loads, right?)
Alex Crichton (Aug 16 2024 at 18:34):so as perhaps the most simple change, it would perhaps look like:
force_multipleget_value_as_source_or_const to return InputSourceInst::None if the instruction has multiple outputs Alex Crichton (Aug 16 2024 at 18:34):yeah what specifically I was seeing was that any operand to isplit would transitively mark all operands as Multiple which thwarted load sinking elsewhere
Alex Crichton (Aug 16 2024 at 18:36):the downside of this "simple" change though is that if a multi-result instruction has a dead result, e.g. the second return value of isplit, it in theory should allow seeing through the isplit but it doesn't?
Chris Fallin (Aug 16 2024 at 18:36):that delta is I think the logical outcome of the concept of "rooted" being defined for now as any multi-output; I was going to suggest a more generalized thing but then it would result in silly dead code; as long as we have comments where force_multiple is noting that we explicitly do not look through multi-output insts when matching, and add a comment where we block such seeing-through noting that it is load-bearing
Chris Fallin (Aug 16 2024 at 18:37):another alternative is to make the multiplicity analysis smarter: Once + Once -> Multiple, including across multiple inputs, but Once + not-used should result in Once
Chris Fallin (Aug 16 2024 at 18:37):s/multiple inputs/multiple outputs/
Alex Crichton (Aug 16 2024 at 18:37):removing force_multiple gets that for free though right?
Alex Crichton (Aug 16 2024 at 18:37):b/c if you only use one result of isplit the operand only gets a single use?
Alex Crichton (Aug 16 2024 at 18:38):hm actually wait I don't think "just" removing force_multiple enough, the DFS would have to stop propagating through a multi-result instruction
Alex Crichton (Aug 16 2024 at 18:39):b/c otherwise if you use both results of isplit it'll naturally mark the operand as used twice right? (if the only change is to remove force_multiple)
Chris Fallin (Aug 16 2024 at 18:39):right, I think we actually need the notion of "rooted"; it's actually that we force codegen at that matching root (by not seeing through it) and really do only use the input once as a result
Alex Crichton (Aug 16 2024 at 18:41):hm ok I'll try to poke at this and see if I can't bottom this out more
Chris Fallin (Aug 16 2024 at 18:41):happy to talk more in next week's CL meeting too!
Alex Crichton (Aug 16 2024 at 20:55):I've ended up with https://github.com/bytecodealliance/wasmtime/pull/9137 as an attempt of this
Last updated: Dec 23 2024 at 13:07 UTC