Trevor Gross (Jul 08 2026 at 19:35): Trevor Gross (Jul 08 2026 at 19:35):Would somebody be able to point me to where x86 return ABI is figured out, specifically on Windows? I assume it has to be somewhere here but I can't find it https://github.com/bytecodealliance/wasmtime/blob/90ae123dc4f1c4d2e8c3bffaa61094d9a970d650/cranelift/codegen/src/isa/x64/abi.rs
Trevor Gross (Jul 08 2026 at 19:36):f129 is currently returned in xmm0 but LLVM recently changed to return indirectly (matching GCC) so I'm looking to do the same update here
Alex Crichton (Jul 08 2026 at 19:55):I think you'll want code somewhere around here perhaps (in that abi.rs file)
Trevor Gross (Jul 08 2026 at 20:01):That and the block below it seem to only be for argument handling (at least, based on ArgsOrRets)
Trevor Gross (Jul 08 2026 at 20:02):I was kind of trying to trace where the i128 handling would be since that's similar-ish, returning in registers on most platforms but xmm on Windows. But I couldn't seem to find it anywhere
Alex Crichton (Jul 08 2026 at 20:08):the if above that I know I've seen before in terms of handling returns-of-i128, but I'm not immediately seeing where it connects to regalloc and such
Alex Crichton (Jul 08 2026 at 20:09):it might be this loop though where reg_tys will have two elements for i64 I believe
Trevor Gross (Jul 08 2026 at 20:25):Hm hm rc_for_type doesn't do any special handling for i128 and I don't see anything specific for Windows in that loop. Maybe clif doesn't actually have specific handling for windows i128 and the frontend handles that instead?
Trevor Gross (Jul 08 2026 at 20:26):Since you're around, how do you get rustc to emit the .clif? I thought --emit=clif worked but it seems it doesn't
Alex Crichton (Jul 08 2026 at 20:26):I wouldn't be surprised yeah if it's a frontend thing, I don't think we have native tests in cranelift itself about Rust code using a plain extern "C" to call an i128-returning function
Alex Crichton (Jul 08 2026 at 20:26):for rustc emitting clif I'd ask bjorn3, I'm not sure myself
Trevor Gross (Jul 08 2026 at 20:29):@bjorn3 do you happen to know more about how/where this works?
bjorn3 (Jul 08 2026 at 20:30):--emit=llvm-ir should work
Trevor Gross (Jul 08 2026 at 20:31):Ah, it errors but still creates the files
RUSTFLAGS="-Zcodegen-backend=cranelift" cargo rustc -- --emit=llvm-ir
Compiling clif-f128 v0.1.0 (~/projects/scratch/clif-f128)
error: could not copy "/Users/tmgross/Documents/projects/scratch/clif-f128/target/debug/deps/clif_f128-d2d7b61ef0548f3f.edabakgdnj4dr38hhli25hfbo.0k4dnte.rcgu.ll" to "/Users/tmgross/Documents/projects/scratch/clif-f128/target/debug/deps/clif_f128-d2d7b61ef0548f3f.ll": No such file or directory (os error 2)
error: could not compile `clif-f128` (lib) due to 1 previous error
tmgross@quince-2 clif-f128 > fd . -e clif target
target/debug/deps/clif_f128-d2d7b61ef0548f3f.clif/
target/debug/deps/clif_f128-d2d7b61ef0548f3f.clif/pass_f128.opt.clif
target/debug/deps/clif_f128-d2d7b61ef0548f3f.clif/pass_f128.unopt.clif
target/debug/deps/clif_f128-d2d7b61ef0548f3f.clif/ret_f128.opt.clif
target/debug/deps/clif_f128-d2d7b61ef0548f3f.clif/ret_f128.unopt.clif
The windows i128 abi is implemented in rustc (through abi adjustments) and (for libcalls) in cg_clif (eg https://github.com/rust-lang/rustc_codegen_cranelift/blob/5bbb9e4f6618aa7c32bde230875fa36608e60bdf/src/abi/mod.rs#L164)
bjorn3 (Jul 08 2026 at 20:33):the rustc side: https://github.com/rust-lang/rust/blob/7ef36edc23ac10760eb9f097d32edbab2e1e8928/compiler/rustc_target/src/callconv/x86_win64.rs#L21
bjorn3 (Jul 08 2026 at 20:34):the copy error is because cg_ssa expects a single .ll file to be generated, while cg_clif produces a directory of files.
Trevor Gross (Jul 08 2026 at 20:34):Doesn't cranelift itself need knowledge of the ABI? For knowing how to emit libcalls
Trevor Gross (Jul 08 2026 at 20:34):(I'm looking at this all because of https://github.com/rust-lang/rust/pull/158918)
bjorn3 (Jul 08 2026 at 20:35):all 128bit libcalls are emitted as regular calls on the cg_clif side. Cranelift itself doesn't have libcalls for 128bit operations.
bjorn3 (Jul 08 2026 at 20:37):you could adjust the cg_clif code i linked to apply for f128 in addition to i128. together with the abi adjustment change you already do in your PR that should be enough to make it work.
Trevor Gross (Jul 08 2026 at 20:37):Ah - so e.g. f128 * f128 is changed to __multf2 in the rustc backend before even hitting cranelift's codegen?
bjorn3 (Jul 08 2026 at 20:38):__multif3, but yes: https://github.com/rust-lang/rustc_codegen_cranelift/blob/5bbb9e4f6618aa7c32bde230875fa36608e60bdf/src/codegen_f16_f128.rs#L137
Trevor Gross (Jul 08 2026 at 20:39):Alright, that all makes a lot more sense. Thanks for the help clearing this up
Last updated: Jul 29 2026 at 05:03 UTC