Invalid LLVM coverage data produced with crate brotli_decompressor

[catenacyber]

I tried this code:

use std::io;
use brotli_decompressor;

fn main() {
    match brotli_decompressor::BrotliDecompress(&mut io::stdin(), &mut io::stdout()) {
        Ok(_) => {},
        Err(e) => panic!("Error {:?}", e),
    }
}

compiled with RUSTFLAGS="-Zinstrument-coverage"
and Cargo.toml having

[dependencies]
brotli-decompressor = "2.3.1"

I expected to see this happen:
Running echo 123 | ./target/debug/rustcov then llvm-profdata merge -j=1 -sparse, then llvm-cov should give me a coverage report

Instead, this happened:
llvm-cov fails with `Failed to load coverage: Malformed instrumentation profile data

Meta

rustc --version --verbose:

rustc 1.52.0-nightly (476acbf1e 2021-03-03)
binary: rustc
commit-hash: 476acbf1e9965b5e95c90f0d7d658709812b7003
commit-date: 2021-03-03
host: x86_64-unknown-linux-gnu
release: 1.52.0-nightly
LLVM version: 11.0.1

Backtrace

There is no backtrace here as the failure happens with llvm-cov
I used the following clang

clang --version
clang version 12.0.0 (https://github.com/llvm/llvm-project.git 4918a3d138b907a571f496661b5367e090e1e8bb)

Using this patch in llvm

diff --git a/llvm/lib/ProfileData/Coverage/CoverageMappingReader.cpp b/llvm/lib/ProfileData/Coverage/CoverageMappingReader.cpp
index 1acdcb4bebb9..b985ea11138d 100644
--- a/llvm/lib/ProfileData/Coverage/CoverageMappingReader.cpp
+++ b/llvm/lib/ProfileData/Coverage/CoverageMappingReader.cpp
@@ -544,8 +544,11 @@ class VersionedCovMapFuncRecordReader : public CovMapFuncRecordReader {
       StringRef FuncName;
       if (Error Err = CFR->template getFuncName<Endian>(ProfileNames, FuncName))
         return Err;
-      if (FuncName.empty())
-        return make_error<InstrProfError>(instrprof_error::malformed);
+      if (FuncName.empty()) {
+       FuncName = "ossfuzz_llvm_rust_coverage_warning";
+       printf("function with hash %lx has no name", FuncHash);
+        //return make_error<InstrProfError>(instrprof_error::eof);
+      }
       ++CovMapNumUsedRecords;
       Records.emplace_back(Version, FuncName, FuncHash, Mapping,
                            FileRange.StartingIndex, FileRange.Length);

I get more information with, as llvm-cov prints
function with hash 66eebda9c13062cc has no name

cc @richkadel
Might end up as a duplicate of #82144
But I do not see any optimization in cargo build -vv

[richkadel]

I believe I have a working fix for this. It is apparently not a dup of #82144 but may be related.

The fix I applied is a change in rustc_typeck/src/collect.rs function codegen_fn_attrs():

                } else if list_contains_name(&items[..], sym::always) {
                    if tcx.sess.opts.debugging_opts.instrument_coverage {
                        InlineAttr::Hint
                    } else {
                        InlineAttr::Always
                    }

I want to make a note of this, but I'm also investigating an idea that may (possibly) allow LLVM inlining.

If that fix works, I may be able to enable LLVM optimization, and this workaround would not be necessary.

What's happening is, #[inline(always)] causes LLVM codegen to force inlining of the function, regardless of optimization settings (apparently).

I'm not sure if there is an LLVM flag we can apply today, to turn that off, or if that's even a good idea. But this fix would seem to disable inlining of all Rust functions, as long as LLVM optimization is NOT applied.

The sample code works with this fix.

I'll be submitting a PR in the near future with either this fix or a better one, if I confirm I can enable LLVM inlining after all.

cc: @tmandry @wesleywiser

[catenacyber]

Thanks, looking forward to the fix