Benjamin Bouvier (Feb 21 2020 at 14:47): Benjamin Bouvier (Feb 21 2020 at 14:47):Starting a new topic for general chatting, as a possible starting point for beginners and newcomers in the Cranelift stream, and so we can get a link to it and put it in README etc.
Chris Fallin (Feb 21 2020 at 21:29):Has anyone successfully cross-compiled wasmtime / Cranelift with cargo --target ...? I'm trying to build an aarch64 binary and run it with qemu-aarch64, to avoid slow CPU on my physical aarch64 machine (RPi4). I'm running into an issue where some part of the build system is invoking native (x86) cc. I've never cross-compiled with the Cargo toolchain in general and I'm curious if this is supposed to be supported...
bjorn3 (Feb 21 2020 at 21:32):What crate invokes cc?
fitzgen (he/him) (Feb 21 2020 at 21:33):cc @Alex Crichton: see cross compilation question above ^
Alex Crichton (Feb 21 2020 at 21:34):@Chris Fallin in theory it should work just fine with the right configuration, but compiling for aarch64 won't succeed for wasmtime at least because it has x86/x86_64-specific things and will liekly fail to link/resolve/etc
Alex Crichton (Feb 21 2020 at 21:35):cranelift however should be comipleable to aarch64 (I think?), but can you gist the error you're seeing?
Chris Fallin (Feb 21 2020 at 21:39):Ah, I just discovered I needed some config in ~/.cargo/config (as per japaric/rust-cross), namely
[target.aarch64-unknown-linux-gnu] linker = "aarch64-linux-gnu-gcc"
I'm now just seeing a missing symbol __chkstk in wasmtime_jit::link::link_module, which I suppose is some of the platform-specific glue you mentioned. How difficult do you think it would be to stub things out? (I just want to get far enough to try a "hello world" and shake out all my unimplemented!()'s in my new arm64 backend...)
Alex Crichton (Feb 21 2020 at 21:43):@Chris Fallin to get things working at all you can use stubs like in https://github.com/bytecodealliance/wasmtime/pull/943
Alex Crichton (Feb 21 2020 at 21:43):which will likely merge soon anyway
Alex Crichton (Feb 21 2020 at 21:43):and yeah there you need to configure the linker for cargo to pass to rustc to use
Chris Fallin (Feb 21 2020 at 21:44):Excellent, will take a look -- thanks very much!
Joey Gouly (Feb 26 2020 at 17:39):is "v0 -> v1" a regmove? I can't find where that syntax is printed
Joey Gouly (Feb 26 2020 at 17:39):hard to search for -> in rust code :P
fitzgen (he/him) (Feb 26 2020 at 17:39):its what happens when v0 becomes an alias for v1
Joey Gouly (Feb 26 2020 at 17:40):I see, I guess those are skipped when iterating over insts in a block?
Joey Gouly (Feb 26 2020 at 17:40):With for inst in f.layout.block_insts(bb) {
Joey Gouly (Feb 26 2020 at 17:42):or maybe there is something which replaces the aliases with the actual value
Dan Gohman (Feb 26 2020 at 17:42):@Joey Gouly Right. They're not instructions, they're just aliases.
fitzgen (he/him) (Feb 26 2020 at 17:43):see https://docs.rs/cranelift-codegen/0.59.0/cranelift_codegen/ir/dfg/struct.DataFlowGraph.html#method.change_to_alias and related methods
Joey Gouly (Feb 26 2020 at 17:44):is there a reason they need to stay around, rather than the instructions being updated?
Chris Fallin (Feb 26 2020 at 17:45):I think the main reason is that this design avoids the need for def-to-use linkage -- otherwise one needs to find all uses to do the rewrite.
Joey Gouly (Feb 26 2020 at 17:46):ah good point. @Chris Fallin I think some of the new backend code is wrong then, because it is not following the aliases
Chris Fallin (Feb 26 2020 at 17:47):@Joey Gouly I see, details welcome if you've found a bug :-)
Dan Gohman (Feb 26 2020 at 17:48):You can call DataFlowGraph::resolve_aliases on a value if you want to look through aliases
Chris Fallin (Feb 26 2020 at 17:50):@Joey Gouly ah, I've found at least one case in LowerCtx where it doesn't resolve aliases. I'll audit for this, but specifics of what you ran into would be useful as well, thanks!
Joey Gouly (Feb 26 2020 at 17:55):The one I found was in the edge-blocks phi parts
Joey Gouly (Feb 26 2020 at 17:56):@Chris Fallin the best way to do this would be to replace all value_regs[] lookup with a function that calls DataFlowGraph::resolve_aliases
Chris Fallin (Feb 26 2020 at 17:56):Yup, on it -- thanks for finding this!
Joey Gouly (Feb 27 2020 at 11:08):@Chris Fallin DataFlowGraph::value_def() already resolves aliases, so you did it eagerly in a few places, not a big deal though
Joey Gouly (Feb 28 2020 at 11:42):Does clif-util wasmtranslate the entire wasm module to CLIF and then compile it? Does it translate per-function or?
edit: I think it's eager, I found src/wasm.rs.
Joey Gouly (Feb 28 2020 at 16:58):@Dan Gohman why are the aliases printed per block?
Dan Gohman (Feb 28 2020 at 17:00):@Joey Gouly The aliases don't have an inherent location, so they could go anywhere.
Joey Gouly (Feb 28 2020 at 17:01):Ok, that's what I expected. Just surprising to see them in random blocks
Dan Gohman (Feb 28 2020 at 17:01):It's not always at the top of a block, although that's common because the SSA algorithm oftgen emits aliases for block parameters.
Dan Gohman (Feb 28 2020 at 17:02):For aliases of instruction values, we print them immediately after the instructions whose values they're aliasing
Joey Gouly (Feb 28 2020 at 17:03):Ah I see that now. I was mostly seeing them at the start of the block, due to large numbers of block arguments
Joey Gouly (Mar 06 2020 at 11:15):Why does rust code compiled to wasm, have vmctx params in the clif?
bjorn3 (Mar 06 2020 at 11:17):The vmctx is necessary to get for example the location of the heap
Joey Gouly (Mar 06 2020 at 11:17):Ah, for example wasmtime will setup something to pass there?
bjorn3 (Mar 06 2020 at 11:18):Yes
Joey Gouly (Mar 06 2020 at 11:28):With fn0 = u0:827 sig0, is that u0:827 actually an entity reference to another function? I thought clif functions were insulated from all other functions
bjorn3 (Mar 06 2020 at 11:32):u0:827 is a number that doesn't have any meaning during compilation. After compilation the user of cranelift-codegen (for example cranelift-module, or wasmtime) interprets it to determine what the target of a relocations is.
Joey Gouly (Mar 06 2020 at 11:33):Aha, so the user has to keep track of which functions CLIF functions are which numbers
bjorn3 (Mar 06 2020 at 11:34):Yes, except when using cranelift-module, in which case cranelift-module does it for you.
Andrew Brown (Mar 06 2020 at 18:29):Can anyone online give me a quick review of a refactoring PR? No functionality changed... https://github.com/bytecodealliance/wasmtime/pull/1246
teapotd (Apr 29 2020 at 16:15):Hello. Could someone review/merge my PRs (https://github.com/bytecodealliance/wasmtime/pull/1382, https://github.com/bytecodealliance/wasmtime/pull/1534, https://github.com/bytecodealliance/wasmtime/pull/1510)?
Andrew Brown (Apr 29 2020 at 16:19):I'll take a look at #1382
teapotd (May 26 2020 at 20:24):What are valid values for b8/16/32/64 types? According to Cranelift IR docs, they are either all zeros or all ones, but when I tried bconst.b8 true I got 1, not 0xFF.
teapotd (May 26 2020 at 20:49):On the other hand, bnot for b32 negates all bits and icmp for vector types yields masks filled with zeroes or ones on x86, as boolean vector type.
Andrew Brown (May 26 2020 at 22:35):Several larger boolean types are also defined, primarily to be used as SIMD element types. They can be stored in memory, and are represented as either all zero bits or all one bits.
Andrew Brown (May 26 2020 at 22:35):So, I read that as: the SIMD types do have a defined representation but the scalar types may not
Andrew Brown (May 26 2020 at 22:36):i.e. you have to use bint to get an integer value out of a b8 but if you extracted a lane of a b8x16 it would be either 0x00 or 0xff
teapotd (May 26 2020 at 22:43):Hm, if you create a boolean vector with vconst it's using 0/1: https://github.com/bytecodealliance/wasmtime/blob/master/cranelift/filetests/filetests/isa/x86/simd-vconst-rodata.clif#L13-L19
Andrew Brown (May 26 2020 at 22:49)::slight_smile: then something is not right: perhaps the documentation should change or perhaps vconst.BxN should change... since I don't do a lot with the BxN types I haven't noticed any issues
Andrew Brown (May 26 2020 at 22:50):And, thinking about it more, it is the documentation that is correct: it should match the Wasm SIMD spec which says
The comparison operations all compare two vectors lane-wise, and produce a mask vector with the same number of lanes as the input interpretation where the bits in each lane are 0 for false and all ones for true.
Andrew Brown (May 26 2020 at 22:51):https://github.com/WebAssembly/simd/blob/master/proposals/simd/SIMD.md#comparisons
Andrew Brown (May 26 2020 at 22:52):This should be a relatively simple fix but feel free to open an issue if it wasn't the main point of what you were getting at... your question was more about scalar types?
teapotd (May 26 2020 at 22:57):No, I was interested in vector types, as I was looking into #1187.
teapotd (May 26 2020 at 23:02):There's a vselect instruction in Cranelift that currently has no encoding and it could be encoded as blend on x86. We could then use vselect instead of bitselect directly in wasm code translator or replace it in preopt.
teapotd (May 26 2020 at 23:04):I'll make fix for vconst
Andrew Brown (May 26 2020 at 23:11):Sounds good; just a heads up that at least PBLENDW should be coming soon-ish: https://github.com/bytecodealliance/wasmtime/pull/1765/commits/66216fc1f5f133ebffdba42bf075da6b5cc3c634
Carlo Kok (Jul 06 2020 at 17:58):does anyone know why cranelift got merged into wasmtime? Just curious really.
Till Schneidereit (Jul 06 2020 at 18:07):@Carlo Kok I wrote a comment explaining the reasoning here: https://github.com/bytecodealliance/wasmtime/issues/1185#issuecomment-591050388
Carlo Kok (Jul 06 2020 at 18:07):thanks! note that I wasn't complaining. Just curious
Carlo Kok (Jul 06 2020 at 18:08):good read. Thanks
Till Schneidereit (Jul 06 2020 at 18:24):note that I wasn't complaining. Just curious
Understood, yes, but thank you for emphasizing it! :heart:
Carlo Kok (Jul 06 2020 at 18:30):is there a list of supported triples ?
Carlo Kok (Jul 06 2020 at 18:37):So feature wise, things I currently use in LLVM that cranelift doesn't have yet: Exception handling, COMDAT; So I'd say cranelift already has a fairly impressive set of features.
bjorn3 (Jul 06 2020 at 18:55):x86_64 is the most complete target. aarch64 comes next. x86 is a little worse. risc-v and arm32 are just stubs.
Carlo Kok (Jul 06 2020 at 18:55):oh arm32 isn't done at all? IT's a bit hard to see from the codegen repo
bjorn3 (Jul 06 2020 at 18:57):yes, there are literally no instruction encodings: https://github.com/bytecodealliance/wasmtime/blob/c91a9313b532067bb53eea44a182bb0c5bd2ebf4/cranelift/codegen/meta/src/isa/arm32/mod.rs#L73-L77
bjorn3 (Jul 06 2020 at 18:57):risc-v at least has some: https://github.com/bytecodealliance/wasmtime/blob/master/cranelift/codegen/meta/src/isa/riscv/encodings.rs
Carlo Kok (Jul 06 2020 at 18:59):So feature wise, things I currently use in LLVM that cranelift doesn't have yet: arm(32), Exception handling, COMDAT; So I'd say cranelift already has a fairly impressive set of features.
Carlo Kok (Jul 06 2020 at 18:59):updated :)
bjorn3 (Jul 06 2020 at 19:00):cranelift is currently transitioning to a new framework for backends. the aarch64 backend has been created using that new framework. the x86_64 backend is currently using the old framework. a new x86_64 backend is being developed, but it is still in a relatively early stage.
bjorn3 (Jul 06 2020 at 19:01):dwarf unwinding support would be useful for https://github.com/bjorn3/rustc_codegen_cranelift/ too. there is already an issue for it: https://github.com/bytecodealliance/wasmtime/issues/1677
Carlo Kok (Jul 06 2020 at 19:02):I read that yes (new framework) but i do know x86-32 works too (old backend though)
Carlo Kok (Jul 06 2020 at 19:03):yeah. Just unwinding has some concerns though: win32 is vastly different than win64; and there's exception handling itself too
Carlo Kok (Jul 06 2020 at 19:04):Exception handling and unwinding isn't that different though. At least not on the targets I saw.
bjorn3 (Jul 06 2020 at 19:05):Cranelift can already generate SEH and .eh_frame tables. They are just treated as not requiring any cleanup.
Carlo Kok (Jul 06 2020 at 19:06):ah cool
bjorn3 (Jul 06 2020 at 19:07):Wasmtime uses them to unwind back to the caller when the wasm code traps or a called rust function panics.
Carlo Kok (Jul 06 2020 at 19:14):And this works with win32 seh too?
Carlo Kok (Jul 06 2020 at 19:14):That is impressive
bjorn3 (Jul 06 2020 at 19:16):I don't know if it works on win32 SEH. As far as I know Wasmtime doesn't work on 32bit yet.
Carlo Kok (Jul 06 2020 at 19:16):Ah
Carlo Kok (Jul 08 2020 at 11:11):I want to tackle COMDAT myself once it's supported in the object code
Carlo Kok (Jul 08 2020 at 11:11):would a feature reqeust for SEH and arm32 be welcome? so there's some place to track and gauge interest?
Benjamin Bouvier (Jul 08 2020 at 11:57):hi! I think there's already an issue open for supporting arm32: https://github.com/bytecodealliance/wasmtime/issues/1173
Carlo Kok (Jul 08 2020 at 12:04):thanx! Missed that one
Joey Gouly (Jul 21 2020 at 14:37):a boolean vector's elements are meant to be all 0 or all 1 right?
Chris Fallin (Jul 21 2020 at 15:01):@Joey Gouly they at least encode false or true (i.e., are not vectors of false/true):
/// Boolean types: `B1`, `B8`, `B16`, `B32`, `B64`, and `B128`. These all encode 'true' or 'false'. The
/// larger types use redundant bits.
(from codegen/src/ir/types.rs)
Chris Fallin (Jul 21 2020 at 15:02):I'm not sure whether a true B128 is meant to be 0....001 or 1...111 though
Chris Fallin (Jul 21 2020 at 15:04):I can't seem to find anything in the old x86 backend that would imply either way... continuing to look
Joey Gouly (Jul 21 2020 at 15:05):Sorry what I meant was that each element was 0x00... or 0xff...
Chris Fallin (Jul 21 2020 at 15:05):Ah, OK, it seems that it must be numeric 0 or 1, because we use a simple zero-extension to go from e.g. B8 to I8 and the latter must be 0 or 1
Joey Gouly (Jul 21 2020 at 15:06):Look at https://github.com/bytecodealliance/wasmtime/blob/main/cranelift/codegen/src/simple_preopt.rs#L956 for example
Chris Fallin (Jul 21 2020 at 15:06):Ah, well that is interesting indeed
Chris Fallin (Jul 21 2020 at 15:08):@Benjamin Bouvier or @Dan Gohman do you know more?
Chris Fallin (Jul 21 2020 at 15:08):I suppose it has to be 0 / -1 (ff...) given the masking above -- but then that means we have to be careful with bint of the result of that must be 0 / 1
Benjamin Bouvier (Jul 21 2020 at 15:19):I seem to recall that true lanes are all ones, while false lanes are all zeroes, i don't know if it's guaranteed though
Benjamin Bouvier (Jul 21 2020 at 15:19):@Andrew Brown you might know ^?
Andrew Brown (Jul 21 2020 at 15:21):Yeah, I think what Benjamin said is correct: the vector booleans have to be all 0s or all 1s... There was an issue about this, let me dig it up.
Andrew Brown (Jul 21 2020 at 15:27):Here it is: https://github.com/bytecodealliance/wasmtime/pull/1769. This fixed a bug where boolean vector lanes were not all 0s or all 1s. The Cranelift IR documentation describing this is here: https://github.com/bytecodealliance/wasmtime/blob/main/cranelift/docs/ir.md#boolean-types.
Joey Gouly (Jul 21 2020 at 15:37):Thanks for the links Andrew!
Joey Gouly (Jul 21 2020 at 15:39):@Chris Fallin looks like aarch64's Bint lowering is wrong?
Chris Fallin (Jul 21 2020 at 15:40):Yes, given the above, it should mask down to the LSB
Chris Fallin (Jul 21 2020 at 15:40):I'll put together a patch unless you want to include this in yours
Chris Fallin (Jul 21 2020 at 15:42):Also icmp / fcmp -- they use cset which yields 0/1. If dest type is > b1 then we need a negation
Joey Gouly (Jul 21 2020 at 15:44):Vector comparisons don't use cset
Chris Fallin (Jul 21 2020 at 15:51):Two separate issues I guess -- there's the question of vector lane values (your immediate question) but then there's also the representation of scalar b8 / b16 / ...
Chris Fallin (Jul 21 2020 at 15:52):For the latter, we should also use the 0 / -1 representation
Joey Gouly (Jul 21 2020 at 15:54):Oh silly me, I was thinking you meant bNxM when you said >b1
Chris Fallin (Jul 21 2020 at 16:06):Yup no worries! I'll patch the scalar case
Joey Gouly (Jul 21 2020 at 16:10):Thanks! And the original reason I went looking into this was a mistake on my side anyway, but Im happy we found this issue due to it :)
Benjamin Bouvier (Jul 29 2020 at 13:00):@Alex Crichton Not sure where this message has gone, but: yesterday i had to pin the version of rustc nightly because staticvec doesn't compile in rustc nightly anymore, and this was breaking the build.
Benjamin Bouvier (Jul 29 2020 at 13:01):opened a PR for static vec, but it's not sufficient as staticvec's CI relies on miri, thus the latest rustc nightly version that supports miri, which is currently 2 weeks old... https://github.com/slightlyoutofphase/staticvec/pull/39
Alex Crichton (Jul 29 2020 at 13:08):Bah, I'll work on updating our CI to be green soon
Benjamin Bouvier (Jul 29 2020 at 13:20)::+1:
Benjamin Bouvier (Jul 29 2020 at 13:20):@Alex Crichton since you're around, can i bother you with more CI questions?
Alex Crichton (Jul 29 2020 at 13:21):Sure yeah what's up?
Benjamin Bouvier (Jul 29 2020 at 13:21):I'm trying to mark the cranelift filetests as should_panic with the experimental_x64 feature, but it doesn't seem to be taken into account when i run the tests, unfortunately.
Alex Crichton (Jul 29 2020 at 13:21):What's the marker look like?
Benjamin Bouvier (Jul 29 2020 at 13:21):i've added #[cfg_attr(feature = "experimental_x64", should_panic) above the test in cranelift/tests/filetests.rs, but it does nothing.
Benjamin Bouvier (Jul 29 2020 at 13:21):I'm running the tests with cargo --features experimental_x64 --all --exclude lightbeam (+ peepmatic exclusions)
Alex Crichton (Jul 29 2020 at 13:22):Is it a #[test] function?
Benjamin Bouvier (Jul 29 2020 at 13:22):yes it is
Alex Crichton (Jul 29 2020 at 13:22):And that crate has that feature in cargo.toml?
Benjamin Bouvier (Jul 29 2020 at 13:22):yes
Alex Crichton (Jul 29 2020 at 13:23):Hm weird... Wanna gist your patch and I'll take a look?
Benjamin Bouvier (Jul 29 2020 at 13:25):Sure, thanks!
Benjamin Bouvier (Jul 29 2020 at 13:25):https://paste.mozilla.org/ryemqRPh
Alex Crichton (Jul 29 2020 at 13:26):Hm and you're not getting compile.failures?
Benjamin Bouvier (Jul 29 2020 at 13:27):Nope
Alex Crichton (Jul 29 2020 at 13:27):If so then cranelift/cargo.toml may not have the feature?
Alex Crichton (Jul 29 2020 at 13:27):I'll dig a bit more in in a sec
Benjamin Bouvier (Jul 29 2020 at 13:28):it is there, though: https://github.com/bytecodealliance/wasmtime/blob/main/cranelift/Cargo.toml#L51
Benjamin Bouvier (Jul 29 2020 at 13:30):ok, so new fact: if i run the tests from wasmtime root dir, it doesn't happen. If i run from the wasmtime/cranelift directory, the compile error happens, and the test isn't run if i remove the body addends
Alex Crichton (Jul 29 2020 at 13:35):aha
Alex Crichton (Jul 29 2020 at 13:35):if you do cargo test -p cranelift-tools --feature ... that doesn't actually do what you would expect
Alex Crichton (Jul 29 2020 at 13:35):it's basically a bug in Cargo
Alex Crichton (Jul 29 2020 at 13:35):it applies the features to w/e is in the current directory, not the target of -p
Benjamin Bouvier (Jul 29 2020 at 13:36):aha! and if i run in wasmtime with --all, it'll work? let me try this!
Alex Crichton (Jul 29 2020 at 13:36):you'll need to do --manifest-path cranelift/Cargo.toml or change the cwd
Alex Crichton (Jul 29 2020 at 13:36):nah it won't workwith --all either
Alex Crichton (Jul 29 2020 at 13:36):there's a bit more info here about the unstable feature to fix this
Benjamin Bouvier (Jul 29 2020 at 13:38):oh well, thanks it works with this!
Benjamin Bouvier (Jul 29 2020 at 13:39):i'll split the CI steps then
Benjamin Bouvier (Jul 29 2020 at 14:05):@Alex Crichton well, does it mean that we should be able to run the tests with experimental_x64 in CI? i'm a bit lost regarding how to do this, then
Alex Crichton (Jul 29 2020 at 14:06):I'd probably limit this to a nightly rustc which executes Cargo with the relevant -Z flag
Alex Crichton (Jul 29 2020 at 14:06):so that way we don't have to bend over backwards for this
Benjamin Bouvier (Jul 29 2020 at 14:15):great idea!
Benjamin Bouvier (Jul 29 2020 at 14:15):but it causes new compile time errors that i don't quite understand :confused:
Benjamin Bouvier (Jul 29 2020 at 14:16):https://paste.mozilla.org/BtuHrkxb
Alex Crichton (Jul 29 2020 at 14:17):hm interesting
Alex Crichton (Jul 29 2020 at 14:17):so what's happening in cargo is that the behavior of the feature resolver is different, where features are less eagerly unified
Alex Crichton (Jul 29 2020 at 14:17):I think that means the wiggle crate depends on syn but doesn't activate the right set of features
Alex Crichton (Jul 29 2020 at 14:17):er, no
Alex Crichton (Jul 29 2020 at 14:18):hm I'll have to dig into that later
Alex Crichton (Jul 29 2020 at 14:18):that may actually be a bug in cargo? unsure
Benjamin Bouvier (Jul 29 2020 at 14:48):oops, sorry to cause more work
Alex Crichton (Jul 29 2020 at 14:51):heh no worries
Alex Crichton (Jul 29 2020 at 14:51):@Benjamin Bouvier btw it looks like wasmtime CI is green, what was the failure you were seeing again?
Benjamin Bouvier (Jul 29 2020 at 14:54):@Alex Crichton oh yeah i added a commit that pinned the version of rustc nightly, let me find that
Benjamin Bouvier (Jul 29 2020 at 14:55):https://github.com/bytecodealliance/wasmtime/commit/2c1d3704653f094a93848a87f79361bca511392b was the commit, but i can see there was another CI failure
Benjamin Bouvier (Jul 29 2020 at 14:56):https://github.com/bytecodealliance/wasmtime/runs/918219496?check_suite_focus=true was one of the CI failures with staticvec
Alex Crichton (Jul 29 2020 at 14:57):hm did you bump a staticvec version?
Alex Crichton (Jul 29 2020 at 14:57):b/c CI for the main branch came back green a few hour sago
Benjamin Bouvier (Jul 29 2020 at 15:01):no, i did pin the version of rustc nightly running in wasmtime (first link above)
Alex Crichton (Jul 29 2020 at 15:01):hm ok well I can try to help with a PR
Alex Crichton (Jul 29 2020 at 15:02):but it sounds like local changes may have necessitated a pin
Alex Crichton (Jul 29 2020 at 15:02):I don't know what to debug if main CI isn't failing
Benjamin Bouvier (Jul 29 2020 at 15:02):see also https://github.com/slightlyoutofphase/staticvec/pull/39, but it seems that it requires a miri in rustc nightly to work
Alex Crichton (Jul 29 2020 at 15:02):it makes sense that it'd break, but I don't know why main CI isn't failing in tha tcase
Benjamin Bouvier (Jul 29 2020 at 15:03):Reverting 2c1d3704653f094a93848a87f79361bca511392b should make the CI fail
Alex Crichton (Jul 29 2020 at 15:03):oooh wait I see
Alex Crichton (Jul 29 2020 at 15:03):you landed that on main
Alex Crichton (Jul 29 2020 at 15:03):sorry I see now
Benjamin Bouvier (Jul 29 2020 at 15:03):no worries
Alex Crichton (Jul 29 2020 at 15:03):I thought this was just in your PR
Benjamin Bouvier (Jul 29 2020 at 15:03):i did it a bit preemptively because it was blocking a PR landing in spidermonkey
Alex Crichton (Jul 29 2020 at 15:03):ah ok, in any case that's all I would have done, we can't do much until staticvec updates
Alex Crichton (Jul 29 2020 at 15:03):nah no worries
Alex Crichton (Jul 29 2020 at 15:03):just got mixed up what was landed where
Till Schneidereit (Jul 29 2020 at 15:05):@Benjamin Bouvier I think it'd be good to do these kinds of things in their own PR, or at least mention the pinning in the PR they're included in
Benjamin Bouvier (Jul 29 2020 at 15:07):@Till Schneidereit yeah, i mentioned it here on zulip with a mention to Alex, but this message disappeared in the meanwhile, not sure what happened...
Benjamin Bouvier (Jul 29 2020 at 15:07):i'll put a message in the PR next time
Till Schneidereit (Jul 29 2020 at 15:08):ah, yeah, the vanished message is very strange. But documenting things like this in a PR is good practice anyway, I think. Though really I'd prefer it to be its own PR, and an issue opened about reverting it at the same time
Benjamin Bouvier (Jul 29 2020 at 15:24):Right. There's also a speed vs process tradeoff here: opening a new PR would have taken two more CI runs (one for the PR, one to rebase the other PR on top of the first one). And, this is not the first time this is happening after the merger of Wasmtime and Cranelift repositories (i can recall at least two other instances, last being last week, so third time this is happening to me), and because of time zones, people working in europe are more likely to hit the "unrelated CI failures because of a rustc nightly change" before anyone else...
Alex Crichton (Jul 29 2020 at 15:25):@Benjamin Bouvier ok I opened a Cargo issue for the wiggle problem you're seeing
Alex Crichton (Jul 29 2020 at 15:25):in the meantime you can use doctest = false in crates/wiggle/macro/Cargo.toml
Alex Crichton (Jul 29 2020 at 15:26):and still use -Zfeatures=all i think
Benjamin Bouvier (Jul 29 2020 at 15:46):@Alex Crichton Thanks a lot! I can run all the tests! (with a few annotations to skip some)
Benjamin Bouvier (Jul 29 2020 at 15:47):that being said, if i try to run a particular test by adding its name to the CLI command, i run into a similar compile error as above, again
Benjamin Bouvier (Jul 29 2020 at 15:47):it's probably a later problem, though, as long as we can run something in the meanwhile
Benjamin Bouvier (Jul 31 2020 at 16:52):Just a heads up: it seems the test_x64 CI has an intermittent failure, not sure yet what causes it, and I won't be around or have time today to investigate it too much, unfortunately.
Andrew Brown (Jul 31 2020 at 17:18):I observed something similar in #2060 but thought it was due to rebasing. It's a bit hard to figure out what is going wrong because all the should_panic tests emit backtraces so there are a lot errors in the logs--is there a way to mute those somehow?
Andrew Brown (Jul 31 2020 at 17:21):@Chris Fallin, @Alex Crichton: I think #2071 looks like a good change if we overlook the intermittent CI issues--with Benjamin out, should we merge it or look into the CI issues first?
Alex Crichton (Jul 31 2020 at 17:25):seems fine by me, but I don't know much about the components in play here
Chris Fallin (Jul 31 2020 at 17:26):It seems this is good to merge, but I also think we should address the CI intermittent failure somehow; slippery slope to "manually ignore" tests
Benjamin Bouvier (Jul 31 2020 at 17:50):i managed to catch the crash in rr, that i packed so one can reproduce it on a similar CPU (machine cpuinfo is there: https://benj.me/pub/cpuinfo)
Here's the packed trace directory, to be unzipped and then run with rr replay $trace_dir: https://benj.me/pub/rr-trace.zip
Benjamin Bouvier (Jul 31 2020 at 17:51):We should probably disable the x64 CI if it's too buggy, though
Benjamin Bouvier (Jul 31 2020 at 17:51):/me disappears
Andrew Brown (Jul 31 2020 at 19:35):I see someone restarted the CI and this time the new backend step passed; I'll merge #2071 since I can use Benjamin's changes and we can talk more about troubleshooting or disabling the new backend step if it keeps happening (any objections?)
Benjamin Bouvier (Sep 03 2020 at 11:53):Fwiw, Cranelift has been shipped in Firefox and will be used for wasm compilation on aarch64 devices! Thanks to all who contributed making this possible! cc @Joey Gouly @Anton Kirilov etc.
Benjamin Bouvier (Sep 03 2020 at 11:55):(Also: these two days are company-wide PTO for Mozillians, so Mozillians' activity might be low!)
Joey Gouly (Sep 03 2020 at 12:07):@Benjamin Bouvier awesome!
bjorn3 (Sep 03 2020 at 19:14)::rocket: :confetti: :tada: :ship: :tada: :confetti: :rocket:
Joey Gouly (Sep 18 2020 at 15:59):@fitzgen (he/him) Hi, I think changing from String -> anyhow broke some of the usefulness of clif-utils errors
Joey Gouly (Sep 18 2020 at 16:01):previously when it failed, it printed out what it matched / what it missed, but I dont seem to get that now
fitzgen (he/him) (Sep 18 2020 at 16:01):oh? which ones? anyhow has lots of ability to annotate errors with extra context, and its possible I accidentally forgot to translate some string concatenation into annotated context
Joey Gouly (Sep 18 2020 at 16:02):for example: cargo run -p cranelift-tools -- test cranelift/filetests/filetests/vcode/aarch64/reftypes.clif
fitzgen (he/him) (Sep 18 2020 at 16:03):Joey Gouly said:
previously when it failed, it printed out what it matched / what it missed, but I dont seem to get that now
for filetests? its possible that I messed up the if <condition> { print() } logic when translating CLI flags, which should be unrelated to anyhow, although it was the same PR. some of them used flag_print and others used flag_verbose, and I might have messed some of them up
Joey Gouly (Sep 18 2020 at 16:03):if you delete one of the nextln checks, the output you get is: FAIL reftypes.clif: compile
fitzgen (he/him) (Sep 18 2020 at 16:03):let me look into it
Joey Gouly (Sep 18 2020 at 16:06):thanks!
fitzgen (he/him) (Sep 18 2020 at 17:55):Fix over here: https://github.com/bytecodealliance/wasmtime/pull/2207
Joey Gouly (Sep 22 2020 at 09:45):Thanks!
Joey Gouly (Oct 16 2020 at 16:17):@fitzgen (he/him) Hi, I was wondering in general what the plans for peepomatic were? Is the aim to turn it on by default?
fitzgen (he/him) (Oct 16 2020 at 16:38):@Joey Gouly
The most immediate plan is to demonstrate better codegen while not sacrificing compilation time (very much, since peepmatic is currently a little slower than the hand-written simple preopt pass) by creating peephole passes from superoptimizations synthesized by Souper. See the recent work to add clif-util souper-harvest and clif-util souper-to-peepmatic commands.
After that is successful, we can start talking about when it makes sense to turn on custom Peepmatic passes for which Cranelift embeddings. For example, it likely makes sense for Lucet before it makes sense for SpiderMonkey.
Eventually I would like to replace the hand written simple preopt with the peepmatic version. There are a few open questions here. First is the question of compile times: peepmatic is currently a little slower than the hand written pass. There's certainly more room for improvement here. But it is also easier to add new optimizations via peepmatic DSL than via open coding matches on InstructionData, and peepmatic is designed such that its performance scales as the number of optimizations in a peephole pass grows. When do we cross the line from "not worth it" to "worth it" when balancing compile times and code quality? I'm not sure. That's kind of why I've been declining to add anything to the peepmatic version of simple-preopt that isn't already in the hand-written version. At least this way we are comparing apples to apples, even if we are leaving peepmatic's biggest strengths on the floor.
This sort of ties into having a better benchmark suite as well, so we can have concrete numbers to work with. A benchmark suite that doesn't require a Web embedding. I want to start work on this soon because it is so hard to evaluate any decisions we make related to this stuff without it.
More things I'd like Peepmatic to do eventually:
vcode peephole optimizations (it is already generic over the IR it is optimizing in anticipation of this, but we need a simple reaching defs analysis so we can determine whether two uses of some virtual register are the same value or not)
clif -> vcode lowering
perhaps even instruction selection (the go compiler uses their equivalent DSL for instruction selection and I find it very nice, but other folks might have other feelings on this topic)
perhaps port LLVM's inst combine pass over to peepmatic
Joey Gouly (Oct 20 2020 at 09:33):@fitzgen (he/him) thanks for the explanation!
carado (Jan 16 2021 at 17:59):hi! i was wondering if there is documentation somewhere for cranelift_frontend's FuncInstBuilder ? it seems to be where the logic for instructions is, but docs.rs doesn't seem to have generated documentation for it (indeed, searching a function used in the example on https://docs.rs/cranelift-frontend/0.69.0/cranelift_frontend/index.html such as iconst doesn't yield any results)
carado (Jan 16 2021 at 18:00):(FuncInstBuilder being what is returned by cranelift_frontend::FunctionBuilder::ins())
bjorn3 (Jan 16 2021 at 19:45):FuncInstBuilder implements the InstBuilder trait: https://docs.rs/cranelift-codegen/0.66.0/cranelift_codegen/ir/trait.InstBuilder.html
bjorn3 (Jan 16 2021 at 19:45):@carado
carado (Jan 16 2021 at 19:51):oh cool, thanks!
Anhad Singh (Jan 25 2021 at 04:54):Hello!
I am getting (exit code: 0xc0000005, STATUS_ACCESS_VIOLATION) when trying to Linkage::Import
#[no_mangle]
pub extern "C" fn ello_world() -> i64 {
println!("Ello World!");
0
}
Nvm I fixed it. Is there a way to do overloaded functions in module.symobl() as the docs say that the one last created is accepted
Last updated: Dec 23 2024 at 12:05 UTC