diff options
author | Daniel Baumann <daniel.baumann@progress-linux.org> | 2024-04-17 12:02:58 +0000 |
---|---|---|
committer | Daniel Baumann <daniel.baumann@progress-linux.org> | 2024-04-17 12:02:58 +0000 |
commit | 698f8c2f01ea549d77d7dc3338a12e04c11057b9 (patch) | |
tree | 173a775858bd501c378080a10dca74132f05bc50 /library/stdarch/crates/stdarch-test | |
parent | Initial commit. (diff) | |
download | rustc-698f8c2f01ea549d77d7dc3338a12e04c11057b9.tar.xz rustc-698f8c2f01ea549d77d7dc3338a12e04c11057b9.zip |
Adding upstream version 1.64.0+dfsg1.upstream/1.64.0+dfsg1
Signed-off-by: Daniel Baumann <daniel.baumann@progress-linux.org>
Diffstat (limited to 'library/stdarch/crates/stdarch-test')
-rw-r--r-- | library/stdarch/crates/stdarch-test/Cargo.toml | 24 | ||||
-rw-r--r-- | library/stdarch/crates/stdarch-test/src/disassembly.rs | 193 | ||||
-rw-r--r-- | library/stdarch/crates/stdarch-test/src/lib.rs | 202 | ||||
-rw-r--r-- | library/stdarch/crates/stdarch-test/src/wasm.rs | 55 |
4 files changed, 474 insertions, 0 deletions
diff --git a/library/stdarch/crates/stdarch-test/Cargo.toml b/library/stdarch/crates/stdarch-test/Cargo.toml new file mode 100644 index 000000000..9ac1057be --- /dev/null +++ b/library/stdarch/crates/stdarch-test/Cargo.toml @@ -0,0 +1,24 @@ +[package] +name = "stdarch-test" +version = "0.1.0" +authors = ["Alex Crichton <alex@alexcrichton.com>"] +edition = "2018" + +[dependencies] +assert-instr-macro = { path = "../assert-instr-macro" } +simd-test-macro = { path = "../simd-test-macro" } +cc = "1.0" +lazy_static = "1.0" +rustc-demangle = "0.1.8" +cfg-if = "0.1" + +# We use a crates.io dependency to disassemble wasm binaries to look for +# instructions for `#[assert_instr]`. Note that we use an `=` dependency here +# instead of a floating dependency because the text format for wasm changes over +# time, and we want to make updates to this explicit rather than automatically +# picking up updates which might break CI with new instruction names. +[target.'cfg(target_arch = "wasm32")'.dependencies] +wasmprinter = "=0.2.24" + +[features] +default = [] diff --git a/library/stdarch/crates/stdarch-test/src/disassembly.rs b/library/stdarch/crates/stdarch-test/src/disassembly.rs new file mode 100644 index 000000000..3ace6b20e --- /dev/null +++ b/library/stdarch/crates/stdarch-test/src/disassembly.rs @@ -0,0 +1,193 @@ +//! Disassembly calling function for most targets. + +use crate::Function; +use std::{collections::HashSet, env, process::Command, str}; + +// Extracts the "shim" name from the `symbol`. +fn normalize(mut symbol: &str) -> String { + // Remove trailing colon: + if symbol.ends_with(':') { + symbol = &symbol[..symbol.len() - 1]; + } + if symbol.ends_with('>') { + symbol = &symbol[..symbol.len() - 1]; + } + if let Some(idx) = symbol.find('<') { + symbol = &symbol[idx + 1..]; + } + + let mut symbol = rustc_demangle::demangle(symbol).to_string(); + symbol = match symbol.rfind("::h") { + Some(i) => symbol[..i].to_string(), + None => symbol.to_string(), + }; + + // Remove Rust paths + if let Some(last_colon) = symbol.rfind(':') { + symbol = (&symbol[last_colon + 1..]).to_string(); + } + + // Normalize to no leading underscore to handle platforms that may + // inject extra ones in symbol names. + while symbol.starts_with('_') { + symbol.remove(0); + } + // Windows/x86 has a suffix such as @@4. + if let Some(idx) = symbol.find("@@") { + symbol = (&symbol[..idx]).to_string(); + } + symbol +} + +pub(crate) fn disassemble_myself() -> HashSet<Function> { + let me = env::current_exe().expect("failed to get current exe"); + + let disassembly = if cfg!(target_os = "windows") && cfg!(target_env = "msvc") { + let target = if cfg!(target_arch = "x86_64") { + "x86_64-pc-windows-msvc" + } else if cfg!(target_arch = "x86") { + "i686-pc-windows-msvc" + } else { + panic!("disassembly unimplemented") + }; + let mut cmd = cc::windows_registry::find(target, "dumpbin.exe") + .expect("failed to find `dumpbin` tool"); + let output = cmd + .arg("/DISASM") + .arg(&me) + .output() + .expect("failed to execute dumpbin"); + println!( + "{}\n{}", + output.status, + String::from_utf8_lossy(&output.stderr) + ); + assert!(output.status.success()); + // Windows does not return valid UTF-8 output: + String::from_utf8_lossy(Vec::leak(output.stdout)) + } else if cfg!(target_os = "windows") { + panic!("disassembly unimplemented") + } else { + let objdump = env::var("OBJDUMP").unwrap_or_else(|_| "objdump".to_string()); + let add_args = if cfg!(target_os = "macos") && cfg!(target_arch = "aarch64") { + // Target features need to be enabled for LLVM objdump on Macos ARM64 + vec!["--mattr=+v8.6a,+crypto,+tme"] + } else { + vec![] + }; + let output = Command::new(objdump.clone()) + .arg("--disassemble") + .arg("--no-show-raw-insn") + .args(add_args) + .arg(&me) + .output() + .unwrap_or_else(|_| panic!("failed to execute objdump. OBJDUMP={}", objdump)); + println!( + "{}\n{}", + output.status, + String::from_utf8_lossy(&output.stderr) + ); + assert!(output.status.success()); + + String::from_utf8_lossy(Vec::leak(output.stdout)) + }; + + parse(&disassembly) +} + +fn parse(output: &str) -> HashSet<Function> { + let mut lines = output.lines(); + + println!( + "First 100 lines of the disassembly input containing {} lines:", + lines.clone().count() + ); + for line in output.lines().take(100) { + println!("{}", line); + } + + let mut functions = HashSet::new(); + let mut cached_header = None; + while let Some(header) = cached_header.take().or_else(|| lines.next()) { + if !header.ends_with(':') || !header.contains("stdarch_test_shim") { + continue; + } + eprintln!("header: {}", header); + let symbol = normalize(header); + eprintln!("normalized symbol: {}", symbol); + let mut instructions = Vec::new(); + while let Some(instruction) = lines.next() { + if instruction.ends_with(':') { + cached_header = Some(instruction); + break; + } + if instruction.is_empty() { + cached_header = None; + break; + } + let mut parts = if cfg!(target_env = "msvc") { + // Each line looks like: + // + // > $addr: ab cd ef $instr.. + // > 00 12 # this line os optional + if instruction.starts_with(" ") { + continue; + } + instruction + .split_whitespace() + .skip(1) + .skip_while(|s| s.len() == 2 && usize::from_str_radix(s, 16).is_ok()) + .map(std::string::ToString::to_string) + .skip_while(|s| *s == "lock") // skip x86-specific prefix + .collect::<Vec<String>>() + } else { + // objdump with --no-show-raw-insn + // Each line of instructions should look like: + // + // $rel_offset: $instruction... + instruction + .split_whitespace() + .skip(1) + .skip_while(|s| *s == "lock") // skip x86-specific prefix + .map(std::string::ToString::to_string) + .collect::<Vec<String>>() + }; + + if cfg!(target_arch = "aarch64") { + // Normalize [us]shll.* ..., #0 instructions to the preferred form: [us]xtl.* ... + // as LLVM objdump does not do that. + // See https://developer.arm.com/documentation/ddi0602/latest/SIMD-FP-Instructions/UXTL--UXTL2--Unsigned-extend-Long--an-alias-of-USHLL--USHLL2- + // and https://developer.arm.com/documentation/ddi0602/latest/SIMD-FP-Instructions/SXTL--SXTL2--Signed-extend-Long--an-alias-of-SSHLL--SSHLL2- + // for details. + match (parts.first(), parts.last()) { + (Some(instr), Some(last_arg)) + if (instr.starts_with("ushll.") || instr.starts_with("sshll.")) + && last_arg == "#0" => + { + assert_eq!(parts.len(), 4); + let mut new_parts = Vec::with_capacity(3); + let new_instr = format!("{}{}{}", &instr[..1], "xtl", &instr[5..]); + new_parts.push(new_instr); + new_parts.push(parts[1].clone()); + new_parts.push(parts[2][0..parts[2].len() - 1].to_owned()); // strip trailing comma + parts = new_parts; + } + _ => {} + }; + } + instructions.push(parts.join(" ")); + } + let function = Function { + name: symbol, + instrs: instructions, + }; + assert!(functions.insert(function)); + } + + eprintln!("all found functions dump:"); + for k in &functions { + eprintln!(" f: {}", k.name); + } + + functions +} diff --git a/library/stdarch/crates/stdarch-test/src/lib.rs b/library/stdarch/crates/stdarch-test/src/lib.rs new file mode 100644 index 000000000..078736c66 --- /dev/null +++ b/library/stdarch/crates/stdarch-test/src/lib.rs @@ -0,0 +1,202 @@ +//! Runtime support needed for testing the stdarch crate. +//! +//! This basically just disassembles the current executable and then parses the +//! output once globally and then provides the `assert` function which makes +//! assertions about the disassembly of a function. +#![feature(bench_black_box)] // For black_box +#![deny(rust_2018_idioms)] +#![allow(clippy::missing_docs_in_private_items, clippy::print_stdout)] + +#[macro_use] +extern crate lazy_static; +#[macro_use] +extern crate cfg_if; + +pub use assert_instr_macro::*; +pub use simd_test_macro::*; +use std::{cmp, collections::HashSet, env, hash, hint::black_box, str}; + +cfg_if! { + if #[cfg(target_arch = "wasm32")] { + pub mod wasm; + use wasm::disassemble_myself; + } else { + mod disassembly; + use crate::disassembly::disassemble_myself; + } +} + +lazy_static! { + static ref DISASSEMBLY: HashSet<Function> = disassemble_myself(); +} + +#[derive(Debug)] +struct Function { + name: String, + instrs: Vec<String>, +} +impl Function { + fn new(n: &str) -> Self { + Self { + name: n.to_string(), + instrs: Vec::new(), + } + } +} + +impl cmp::PartialEq for Function { + fn eq(&self, other: &Self) -> bool { + self.name == other.name + } +} +impl cmp::Eq for Function {} + +impl hash::Hash for Function { + fn hash<H: hash::Hasher>(&self, state: &mut H) { + self.name.hash(state) + } +} + +/// Main entry point for this crate, called by the `#[assert_instr]` macro. +/// +/// This asserts that the function at `fnptr` contains the instruction +/// `expected` provided. +pub fn assert(shim_addr: usize, fnname: &str, expected: &str) { + // Make sure that the shim is not removed + black_box(shim_addr); + + //eprintln!("shim name: {}", fnname); + let function = &DISASSEMBLY + .get(&Function::new(fnname)) + .unwrap_or_else(|| panic!("function \"{}\" not found in the disassembly", fnname)); + //eprintln!(" function: {:?}", function); + + let mut instrs = &function.instrs[..]; + while instrs.last().map_or(false, |s| s == "nop") { + instrs = &instrs[..instrs.len() - 1]; + } + + // Look for `expected` as the first part of any instruction in this + // function, e.g., tzcntl in tzcntl %rax,%rax. + // + // There are two cases when the expected instruction is nop: + // 1. The expected intrinsic is compiled away so we can't + // check for it - aka the intrinsic is not generating any code. + // 2. It is a mark, indicating that the instruction will be + // compiled into other instructions - mainly because of llvm + // optimization. + let found = expected == "nop" || instrs.iter().any(|s| s.starts_with(expected)); + + // Look for subroutine call instructions in the disassembly to detect whether + // inlining failed: all intrinsics are `#[inline(always)]`, so calling one + // intrinsic from another should not generate subroutine call instructions. + let inlining_failed = if cfg!(target_arch = "x86_64") || cfg!(target_arch = "wasm32") { + instrs.iter().any(|s| s.starts_with("call ")) + } else if cfg!(target_arch = "x86") { + instrs.windows(2).any(|s| { + // On 32-bit x86 position independent code will call itself and be + // immediately followed by a `pop` to learn about the current address. + // Let's not take that into account when considering whether a function + // failed inlining something. + s[0].starts_with("call ") && s[1].starts_with("pop") // FIXME: original logic but does not match comment + }) + } else if cfg!(target_arch = "aarch64") { + instrs.iter().any(|s| s.starts_with("bl ")) + } else { + // FIXME: Add detection for other archs + false + }; + + let instruction_limit = std::env::var("STDARCH_ASSERT_INSTR_LIMIT") + .ok() + .map_or_else( + || match expected { + // `cpuid` returns a pretty big aggregate structure, so exempt + // it from the slightly more restrictive 22 instructions below. + "cpuid" => 30, + + // Apparently, on Windows, LLVM generates a bunch of + // saves/restores of xmm registers around these intstructions, + // which exceeds the limit of 20 below. As it seems dictated by + // Windows's ABI (I believe?), we probably can't do much + // about it. + "vzeroall" | "vzeroupper" if cfg!(windows) => 30, + + // Intrinsics using `cvtpi2ps` are typically "composites" and + // in some cases exceed the limit. + "cvtpi2ps" => 25, + // core_arch/src/arm_shared/simd32 + // vfmaq_n_f32_vfma : #instructions = 26 >= 22 (limit) + "usad8" | "vfma" | "vfms" => 27, + "qadd8" | "qsub8" | "sadd8" | "sel" | "shadd8" | "shsub8" | "usub8" | "ssub8" => 29, + // core_arch/src/arm_shared/simd32 + // vst1q_s64_x4_vst1 : #instructions = 22 >= 22 (limit) + "vld3" => 23, + // core_arch/src/arm_shared/simd32 + // vld4q_lane_u32_vld4 : #instructions = 31 >= 22 (limit) + "vld4" => 32, + // core_arch/src/arm_shared/simd32 + // vst1q_s64_x4_vst1 : #instructions = 40 >= 22 (limit) + "vst1" => 41, + // core_arch/src/arm_shared/simd32 + // vst4q_u32_vst4 : #instructions = 26 >= 22 (limit) + "vst4" => 27, + + // Temporary, currently the fptosi.sat and fptoui.sat LLVM + // intrinsics emit unnecessary code on arm. This can be + // removed once it has been addressed in LLVM. + "fcvtzu" | "fcvtzs" | "vcvt" => 64, + + // core_arch/src/arm_shared/simd32 + // vst1q_p64_x4_nop : #instructions = 33 >= 22 (limit) + "nop" if fnname.contains("vst1q_p64") => 34, + + // Original limit was 20 instructions, but ARM DSP Intrinsics + // are exactly 20 instructions long. So, bump the limit to 22 + // instead of adding here a long list of exceptions. + _ => 22, + }, + |v| v.parse().unwrap(), + ); + let probably_only_one_instruction = instrs.len() < instruction_limit; + + if found && probably_only_one_instruction && !inlining_failed { + return; + } + + // Help debug by printing out the found disassembly, and then panic as we + // didn't find the instruction. + println!("disassembly for {}: ", fnname,); + for (i, instr) in instrs.iter().enumerate() { + println!("\t{:2}: {}", i, instr); + } + + if !found { + panic!( + "failed to find instruction `{}` in the disassembly", + expected + ); + } else if !probably_only_one_instruction { + panic!( + "instruction found, but the disassembly contains too many \ + instructions: #instructions = {} >= {} (limit)", + instrs.len(), + instruction_limit + ); + } else if inlining_failed { + panic!( + "instruction found, but the disassembly contains subroutine \ + call instructions, which hint that inlining failed" + ); + } +} + +pub fn assert_skip_test_ok(name: &str) { + if env::var("STDARCH_TEST_EVERYTHING").is_err() { + return; + } + panic!("skipped test `{}` when it shouldn't be skipped", name); +} + +// See comment in `assert-instr-macro` crate for why this exists +pub static mut _DONT_DEDUP: *const u8 = std::ptr::null(); diff --git a/library/stdarch/crates/stdarch-test/src/wasm.rs b/library/stdarch/crates/stdarch-test/src/wasm.rs new file mode 100644 index 000000000..bf411c121 --- /dev/null +++ b/library/stdarch/crates/stdarch-test/src/wasm.rs @@ -0,0 +1,55 @@ +//! Disassembly calling function for `wasm32` targets. + +use crate::Function; +use std::collections::HashSet; + +pub(crate) fn disassemble_myself() -> HashSet<Function> { + // Use `std::env::args` to find the path to our executable. Assume the + // environment is configured such that we can read that file. Read it and + // use the `wasmprinter` crate to transform the binary to text, then search + // the text for appropriately named functions. + let me = std::env::args() + .next() + .expect("failed to find current wasm file"); + let output = wasmprinter::print_file(&me).unwrap(); + + let mut ret: HashSet<Function> = HashSet::new(); + let mut lines = output.lines().map(|s| s.trim()); + while let Some(line) = lines.next() { + // If this isn't a function, we don't care about it. + if !line.starts_with("(func ") { + continue; + } + + let mut function = Function { + name: String::new(), + instrs: Vec::new(), + }; + + // Empty functions will end in `))` so there's nothing to do, otherwise + // we'll have a bunch of following lines which are instructions. + // + // Lines that have an imbalanced `)` mark the end of a function. + if !line.ends_with("))") { + while let Some(line) = lines.next() { + function.instrs.push(line.to_string()); + if !line.starts_with("(") && line.ends_with(")") { + break; + } + } + } + // The second element here split on whitespace should be the name of + // the function, skipping the type/params/results + function.name = line.split_whitespace().nth(1).unwrap().to_string(); + if function.name.starts_with("$") { + function.name = function.name[1..].to_string() + } + + if !function.name.contains("stdarch_test_shim") { + continue; + } + + assert!(ret.insert(function)); + } + return ret; +} |