target-cpu shouldn't default to native when cross-compiling

[kornelski]

• The --mca option doesn't work on ARM due to assuming Intel asm syntax.

I've tried working around this by adding cross compilation --target=x86_64-apple-darwin, but this results in no output at all, except:

'apple-m3' is not a recognized processor for this target (ignoring processor)

I needed to add --target-cpu=skylake --mca-arg=-skip-unsupported-instructions=parse-failure to get it working.

---

When crosscompiling with llvm-mca you need to specify which CPU to use for timing information - it defaults to your current platform and error messages can be confusing.

[pacak]

Is it working if you pass --att?

[pacak]

Can you check if master works for you on ARM without having to pass anything special?

[kornelski]

When not cross-compiling, --mca on ARM works without any extra args. Works with --att too.

[pacak]

Cool. Thanks for confirming. I'll release the next version soon.

…

On Sun, Feb 16, 2025, 16:50 Kornel ***@***.***> wrote: Yes, it works now on ARM without any extra args. Works with --att too. — Reply to this email directly, view it on GitHub <#374 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAAQFI2QB23VP5OQIVMNF7D2QEB3XAVCNFSM6AAAAABW32R3D6VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDMNRRGYZTGOJXGM> . You are receiving this because you commented.Message ID: ***@***.***> [image: kornelski]*kornelski* left a comment (pacak/cargo-show-asm#374) <#374 (comment)> Yes, it works now on ARM without any extra args. Works with --att too. — Reply to this email directly, view it on GitHub <#374 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAAQFI2QB23VP5OQIVMNF7D2QEB3XAVCNFSM6AAAAABW32R3D6VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDMNRRGYZTGOJXGM> . You are receiving this because you commented.Message ID: ***@***.***>

[pacak]

What about when you cross compile? It worked for me from x86 to aarch, but I had to specify the processor because llvm-mca was defaulting to my current one... Not sure if this is easy to fix though. It wanted a concrete processor for all the timings stuff.

…

On Sun, Feb 16, 2025, 16:51 Michael ***@***.***> wrote: Cool. Thanks for confirming. I'll release the next version soon. On Sun, Feb 16, 2025, 16:50 Kornel ***@***.***> wrote: > Yes, it works now on ARM without any extra args. Works with --att too. > > — > Reply to this email directly, view it on GitHub > <#374 (comment)>, > or unsubscribe > <https://github.com/notifications/unsubscribe-auth/AAAQFI2QB23VP5OQIVMNF7D2QEB3XAVCNFSM6AAAAABW32R3D6VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDMNRRGYZTGOJXGM> > . > You are receiving this because you commented.Message ID: > ***@***.***> > [image: kornelski]*kornelski* left a comment (pacak/cargo-show-asm#374) > <#374 (comment)> > > Yes, it works now on ARM without any extra args. Works with --att too. > > — > Reply to this email directly, view it on GitHub > <#374 (comment)>, > or unsubscribe > <https://github.com/notifications/unsubscribe-auth/AAAQFI2QB23VP5OQIVMNF7D2QEB3XAVCNFSM6AAAAABW32R3D6VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDMNRRGYZTGOJXGM> > . > You are receiving this because you commented.Message ID: > ***@***.***> >

[kornelski]

However, when cross-compiling from ARM to x86-64, it still needs --target-cpu=<something made by intel> and it needs --att

cargo r -- asm --mca --target=x86_64-apple-darwin -vv --lib 'asm::llvm::Llvm as cargo_show_asm::Dumpable>::dump_range'
    Finished `dev` profile [unoptimized + debuginfo] target(s) in 0.02s
     Running `target/debug/cargo-asm asm --mca --target=x86_64-apple-darwin -vv --lib 'asm::llvm::Llvm as cargo_show_asm::Dumpable>::dump_range'`
Found sysroot: /Users/kornel/.rustup/toolchains/nightly-aarch64-apple-darwin
Running: "/Users/kornel/.rustup/toolchains/nightly-aarch64-apple-darwin/bin/cargo" "rustc" "--message-format=json-render-diagnostics" "--color" "always" "-v" "-v" "--manifest-path" "/Users/kornel/www/cargo-show-asm/Cargo.toml" "--config" "profile.release.strip=false" "--package" "cargo-show-asm" "--lib" "--target" "x86_64-apple-darwin" "--release" "--" "--emit" "asm" "-C" "llvm-args=-x86-asm-syntax=intel" "-Cdebuginfo=2" "-Ccodegen-units=1"
       Fresh cargo-show-asm v0.2.47 (/Users/kornel/www/cargo-show-asm)
    Finished `release` profile [optimized] target(s) in 0.01s

Artifact files: ["/Users/kornel/www/cargo-show-asm/target/x86_64-apple-darwin/release/libcargo_show_asm.rlib", "/Users/kornel/www/cargo-show-asm/target/x86_64-apple-darwin/release/deps/libcargo_show_asm-defff629b9f991b5.rmeta"]
Working with file: /Users/kornel/www/cargo-show-asm/target/x86_64-apple-darwin/release/deps/cargo_show_asm-defff629b9f991b5.s
running "llvm-mca" "--mtriple" "x86_64-apple-darwin"
'apple-m3' is not a recognized processor for this target (ignoring processor)
'apple-m3' is not a recognized processor for this target (ignoring processor)

[kornelski]

works: asm --mca --att --target=x86_64-apple-darwin -vv --lib 'Llvm as cargo_show_asm::Dumpable>::dump_range' --target-cpu=skylake

but with --intel (or just no --att)

running "llvm-mca" "--mtriple" "x86_64-apple-darwin" "--mcpu" "skylake"
<stdin>:4:1: error: unknown use of instruction mnemonic without a size suffix
mov rbp, rsp
^
<stdin>:11:1: error: unknown use of instruction mnemonic without a size suffix
sub rsp, 120
^
<stdin>:13:1: error: unknown use of instruction mnemonic without a size suffix
test rcx, rcx
^

[pacak]

However, when cross-compiling from ARM to x86-64, it still needs --target-cpu=<something made by intel>

Yeah, same problem when cross compiling from x86 to ARM. While I can pick some concrete processor for cross-compilation - timings are not going to match. Maybe a better approach is to detect cross-compilation case and print some suggestions... Going to experiment a bit more before releasing then.

[kornelski]

llvm-mca is inherently CPU-specific, so picking a CPU is unavoidable.

For cross-compilation to x86_64, I recommend using x86-64-v3 as the CPU name. That's the generic baseline.

For aarch64, cortex-a76 is probably the lowest CPU that MCA makes sense for.

[pacak]

Yeah, probably. Will try to make something in a few days.

[RReverser]

Even though I'm not doing cross-compilation, just using target-cpu=native, running into a similar issue on Windows (x86_64-pc-windows-msvc):

...
<stdin>:35:1: error: unknown use of instruction mnemonic without a size suffix
test dl, dl
^
<stdin>:38:1: error: unknown use of instruction mnemonic without a size suffix
cmp ax, 64
^
<stdin>:42:13: error: invalid operand for instruction
vmovd xmm0, r14d
            ^~~~
<stdin>:43:20: error: invalid operand for instruction
vpbroadcastw xmm4, xmm0
                   ^~~~
<stdin>:44:13: error: invalid operand for instruction
vmovd xmm0, ebx
            ^~~
<stdin>:45:20: error: invalid operand for instruction
vpbroadcastw xmm3, xmm0
                   ^~~~
<stdin>:46:13: error: invalid operand for instruction
vmovd xmm0, edi
            ^~~
...

Adding explicit --att fixes this bulk of issues, but then getting

^
<stdin>:565:1: error: Unfinished frame!

Adding --mca-arg=-skip-unsupported-instructions=parse-failure works around that issue too, but it's a bit unclear why it's needed - perhaps newer LLVM producing something cargo-show-asm can't parse?

[pacak]

Those error messages are coming from llvm-mca. Probably it wants something different. Not sure how to debug it - I don't have access to a Windows machine.

[RReverser]

Ah, you're right, I thought it's coming from the parser in this crate.

That just leaves passing the --att then, same as original issue.