From: Thomas Karpiniec Date: Wed, 15 Jun 2022 08:19:56 +0000 (+1000) Subject: Fix up non-test clippy lints X-Git-Tag: v0.3.0~5 X-Git-Url: https://code.octet-stream.net/hashgood/commitdiff_plain/3036c7308170f6c68961e44e61e06727eef78327?hp=12939f1c8ec8ecba887c4ea7dca40337ad6832d9 Fix up non-test clippy lints --- diff --git a/src/main.rs b/src/main.rs index fd965da..f0fc474 100644 --- a/src/main.rs +++ b/src/main.rs @@ -41,11 +41,11 @@ impl Opt { fn get_paste(&self) -> bool { #[cfg(feature = "paste")] { - return self.paste; + self.paste } #[cfg(not(feature = "paste"))] { - return false; + false } } } diff --git a/src/verify.rs b/src/verify.rs index 25e565f..c9ded29 100644 --- a/src/verify.rs +++ b/src/verify.rs @@ -61,15 +61,15 @@ fn get_from_clipboard() -> Result { filename: None, bytes, }; - return Ok(CandidateHashes { + Ok(CandidateHashes { alg, hashes: vec![candidate], source: VerificationSource::Clipboard, - }); + }) } #[cfg(not(feature = "paste"))] { - return Err("Paste not implemented".to_owned()); + Err("Paste not implemented".to_owned()) } } @@ -96,7 +96,7 @@ fn get_from_file(path: &Path) -> Result { let line = line.trim().to_owned(); // Does our first line look like a raw hash on its own? If so, use that - if let Some(candidate) = read_raw_candidate_from_file(&line, &path) { + if let Some(candidate) = read_raw_candidate_from_file(&line, path) { return Ok(candidate); } @@ -105,7 +105,7 @@ fn get_from_file(path: &Path) -> Result { let full_lines = vec![Ok(line)].into_iter().chain(reader.lines()); // Does the entire file look like a coreutils-style digests file? (SHA1SUMS, etc.) - if let Some(candidate) = read_coreutils_digests_from_file(full_lines, &path) { + if let Some(candidate) = read_coreutils_digests_from_file(full_lines, path) { return Ok(candidate); } @@ -148,51 +148,49 @@ where { let mut hashes = vec![]; let mut alg: Option = None; - for l in lines { - if let Ok(l) = l { - let l = l.as_ref().trim(); - // Allow (ignore) blank lines - if l.is_empty() { - continue; - } - // Expected format - // - let (line_alg, bytes, filename) = match l - .find(' ') - .and_then(|space_pos| { - // Char before filename should be space for text or * for binary - match l.chars().nth(space_pos + 1) { - Some(' ') | Some('*') => (l.get(..space_pos)).zip(l.get(space_pos + 2..)), - _ => None, - } - }) - .and_then(|(maybe_hash, filename)| { - // Filename should be in this position without extra whitespace - if filename.trim() == filename { - try_parse_hash(maybe_hash).map(|(alg, bytes)| (alg, bytes, filename)) - } else { - None - } - }) { - Some(t) => t, - None => { - // if we have a line with content we cannot parse, this is an error - return None; + for l in lines.flatten() { + let l = l.as_ref().trim(); + // Allow (ignore) blank lines + if l.is_empty() { + continue; + } + // Expected format + // + let (line_alg, bytes, filename) = match l + .find(' ') + .and_then(|space_pos| { + // Char before filename should be space for text or * for binary + match l.chars().nth(space_pos + 1) { + Some(' ') | Some('*') => (l.get(..space_pos)).zip(l.get(space_pos + 2..)), + _ => None, } - }; - if alg.is_some() && alg != Some(line_alg) { - // Different algorithms in the same digest file are not supported + }) + .and_then(|(maybe_hash, filename)| { + // Filename should be in this position without extra whitespace + if filename.trim() == filename { + try_parse_hash(maybe_hash).map(|(alg, bytes)| (alg, bytes, filename)) + } else { + None + } + }) { + Some(t) => t, + None => { + // if we have a line with content we cannot parse, this is an error return None; - } else { - // If we are the first line, we define the overall algorithm - alg = Some(line_alg); } - // So far so good - create an entry for this line - hashes.push(CandidateHash { - bytes, - filename: Some(filename.to_owned()), - }); + }; + if alg.is_some() && alg != Some(line_alg) { + // Different algorithms in the same digest file are not supported + return None; + } else { + // If we are the first line, we define the overall algorithm + alg = Some(line_alg); } + // So far so good - create an entry for this line + hashes.push(CandidateHash { + bytes, + filename: Some(filename.to_owned()), + }); } // It is a failure if we got zero hashes or we somehow don't know the algorithm