Adding the x86 part of behavioural testing for std::arch intrinsics

ci/run.sh

@@ -91,6 +91,9 @@ fi
 # Test targets compiled with extra features.
 case ${TARGET} in
     x86_64-unknown-linux-gnu)
+        TEST_CPPFLAGS="-fuse-ld=lld -I/usr/include/x86_64-linux-gnu/"
+        TEST_CXX_COMPILER="clang++-19"
+        TEST_RUNNER="${CARGO_TARGET_X86_64_UNKNOWN_LINUX_GNU_RUNNER}"
         export STDARCH_DISABLE_ASSERT_INSTR=1
 
         export RUSTFLAGS="${RUSTFLAGS} -C target-feature=+avx"
@@ -181,6 +184,14 @@ case "${TARGET}" in
             --linker "${CARGO_TARGET_AARCH64_BE_UNKNOWN_LINUX_GNU_LINKER}" \
             --cxx-toolchain-dir "${AARCH64_BE_TOOLCHAIN}"
         ;;
+    x86_64-unknown-linux-gnu*)
+        CPPFLAGS="${TEST_CPPFLAGS}" RUSTFLAGS="${HOST_RUSTFLAGS}" RUST_LOG=warn \
+            cargo run "${INTRINSIC_TEST}" "${PROFILE}"  \
+            --bin intrinsic-test -- intrinsics_data/x86-intel.xml \
+            --runner "${TEST_RUNNER}" \
+            --cppcompiler "${TEST_CXX_COMPILER}" \
+            --target "${TARGET}"
+        ;;
      *)
         ;;
 esac

crates/intrinsic-test/Cargo.toml

@@ -19,3 +19,5 @@ pretty_env_logger = "0.5.0"
 rayon = "1.5.0"
 diff = "0.1.12"
 itertools = "0.14.0"
+quick-xml = { version = "0.37.5", features = ["serialize", "overlapped-lists"] }
+serde-xml-rs = "0.8.0"

crates/intrinsic-test/src/arm/json_parser.rs

@@ -79,8 +79,8 @@ fn json_to_intrinsic(
 ) -> Result<Intrinsic<ArmIntrinsicType>, Box<dyn std::error::Error>> {
     let name = intr.name.replace(['[', ']'], "");
 
-    let results = ArmIntrinsicType::from_c(&intr.return_type.value, target)?;
-
+    let mut results = ArmIntrinsicType::from_c(&intr.return_type.value)?;
+    results.set_metadata("target".to_string(), target.to_string());
     let args = intr
         .arguments
         .into_iter()

crates/intrinsic-test/src/arm/mod.rs

@@ -10,7 +10,6 @@ use std::fs::{self, File};
 use rayon::prelude::*;
 
 use crate::arm::config::POLY128_OSTREAM_DEF;
-use crate::common::SupportedArchitectureTest;
 use crate::common::cli::ProcessedCli;
 use crate::common::compare::compare_outputs;
 use crate::common::gen_c::{write_main_cpp, write_mod_cpp};
@@ -28,13 +27,6 @@ pub struct ArmArchitectureTest {
     cli_options: ProcessedCli,
 }
 
-fn chunk_info(intrinsic_count: usize) -> (usize, usize) {
-    let available_parallelism = std::thread::available_parallelism().unwrap().get();
-    let chunk_size = intrinsic_count.div_ceil(Ord::min(available_parallelism, intrinsic_count));
-
-    (chunk_size, intrinsic_count.div_ceil(chunk_size))
-}
-
 impl SupportedArchitectureTest for ArmArchitectureTest {
     fn create(cli_options: ProcessedCli) -> Box<Self> {
         let a32 = cli_options.target.contains("v7");
@@ -79,7 +71,7 @@ impl SupportedArchitectureTest for ArmArchitectureTest {
             .enumerate()
             .map(|(i, chunk)| {
                 let c_filename = format!("c_programs/mod_{i}.cpp");
-                let mut file = File::create(&c_filename).unwrap();
+                let mut file = fs::File::create(&c_filename).unwrap();
                 write_mod_cpp(&mut file, notice, c_target, platform_headers, chunk).unwrap();
 
                 // compile this cpp file into a .o file.
@@ -97,11 +89,12 @@ impl SupportedArchitectureTest for ArmArchitectureTest {
             .collect::<Result<(), std::io::Error>>()
             .unwrap();
 
-        let mut file = File::create("c_programs/main.cpp").unwrap();
+        let mut file = fs::File::create("c_programs/main.cpp").unwrap();
         write_main_cpp(
             &mut file,
             c_target,
             POLY128_OSTREAM_DEF,
+            Vec::from(platform_headers),
             self.intrinsics.iter().map(|i| i.name.as_str()),
         )
         .unwrap();

crates/intrinsic-test/src/arm/types.rs

@@ -1,25 +1,21 @@
+use std::collections::HashMap;
+
 use super::intrinsic::ArmIntrinsicType;
 use crate::common::cli::Language;
 use crate::common::intrinsic_helpers::{IntrinsicType, IntrinsicTypeDefinition, Sign, TypeKind};
 
 impl IntrinsicTypeDefinition for ArmIntrinsicType {
     /// Gets a string containing the typename for this type in C format.
+    /// This assumes that the metadata hashmap contains this value at the
+    /// "type" key
     fn c_type(&self) -> String {
-        let prefix = self.0.kind.c_prefix();
-        let const_prefix = if self.0.constant { "const " } else { "" };
-
-        if let (Some(bit_len), simd_len, vec_len) =
-            (self.0.bit_len, self.0.simd_len, self.0.vec_len)
-        {
-            match (simd_len, vec_len) {
-                (None, None) => format!("{const_prefix}{prefix}{bit_len}_t"),
-                (Some(simd), None) => format!("{prefix}{bit_len}x{simd}_t"),
-                (Some(simd), Some(vec)) => format!("{prefix}{bit_len}x{simd}x{vec}_t"),
-                (None, Some(_)) => todo!("{:#?}", self), // Likely an invalid case
-            }
-        } else {
-            todo!("{:#?}", self)
-        }
+        self.metadata
+            .get("type")
+            .expect("Failed to extract the C typename in Aarch!")
+            .replace("*", "")
+            .replace("  ", "")
+            .trim()
+            .to_string()
     }
 
     fn c_single_vector_type(&self) -> String {
@@ -40,7 +36,7 @@ impl IntrinsicTypeDefinition for ArmIntrinsicType {
             bit_len: Some(bl),
             simd_len,
             vec_len,
-            target,
+            metadata,
             ..
         } = &self.0
         {
@@ -50,7 +46,11 @@ impl IntrinsicTypeDefinition for ArmIntrinsicType {
                 ""
             };
 
-            let choose_workaround = language == Language::C && target.contains("v7");
+            let choose_workaround = language == Language::C
+                && metadata
+                    .get("target")
+                    .filter(|value| value.contains("v7"))
+                    .is_some();
             format!(
                 "vld{len}{quad}_{type}{size}",
                 type = match k {
@@ -102,15 +102,17 @@ impl IntrinsicTypeDefinition for ArmIntrinsicType {
         }
     }
 
-    fn from_c(s: &str, target: &str) -> Result<Self, String> {
+    fn from_c(s: &str) -> Result<Self, String> {
         const CONST_STR: &str = "const";
+        let mut metadata: HashMap<String, String> = HashMap::new();
+        metadata.insert("type".to_string(), s.to_string());
         if let Some(s) = s.strip_suffix('*') {
             let (s, constant) = match s.trim().strip_suffix(CONST_STR) {
                 Some(stripped) => (stripped, true),
                 None => (s, false),
             };
             let s = s.trim_end();
-            let temp_return = ArmIntrinsicType::from_c(s, target);
+            let temp_return = ArmIntrinsicType::from_c(s);
             temp_return.map(|mut op| {
                 op.ptr = true;
                 op.ptr_constant = constant;
@@ -151,7 +153,7 @@ impl IntrinsicTypeDefinition for ArmIntrinsicType {
                     bit_len: Some(bit_len),
                     simd_len,
                     vec_len,
-                    target: target.to_string(),
+                    metadata,
                 }))
             } else {
                 let kind = start.parse::<TypeKind>()?;
@@ -167,7 +169,7 @@ impl IntrinsicTypeDefinition for ArmIntrinsicType {
                     bit_len,
                     simd_len: None,
                     vec_len: None,
-                    target: target.to_string(),
+                    metadata,
                 }))
             }
         }

crates/intrinsic-test/src/common/constraint.rs

@@ -5,13 +5,15 @@ use std::ops::Range;
 pub enum Constraint {
     Equal(i64),
     Range(Range<i64>),
+    Set(Vec<i64>),
 }
 
 impl Constraint {
-    pub fn to_range(&self) -> Range<i64> {
+    pub fn to_vector(&self) -> Vec<i64> {
         match self {
-            Constraint::Equal(eq) => *eq..*eq + 1,
-            Constraint::Range(range) => range.clone(),
+            Constraint::Equal(eq) => vec![*eq],
+            Constraint::Range(range) => range.clone().collect::<Vec<i64>>(),
+            Constraint::Set(values) => values.clone(),
         }
     }
 }

crates/intrinsic-test/src/common/gen_c.rs

@@ -40,7 +40,7 @@ pub fn generate_c_constraint_blocks<'a, T: IntrinsicTypeDefinition + 'a>(
     };
 
     let body_indentation = indentation.nested();
-    for i in current.constraint.iter().flat_map(|c| c.to_range()) {
+    for i in current.constraint.iter().flat_map(|c| c.to_vector()) {
         let ty = current.ty.c_type();
 
         writeln!(w, "{indentation}{{")?;
@@ -140,12 +140,13 @@ pub fn write_main_cpp<'a>(
     w: &mut impl std::io::Write,
     architecture: &str,
     arch_specific_definitions: &str,
+    headers: Vec<&str>,
     intrinsics: impl Iterator<Item = &'a str> + Clone,
 ) -> std::io::Result<()> {
     writeln!(w, "#include <iostream>")?;
     writeln!(w, "#include <string>")?;
 
-    for header in ["arm_neon.h", "arm_acle.h", "arm_fp16.h"] {
+    for header in headers.iter() {
         writeln!(w, "#include <{header}>")?;
     }
 

crates/intrinsic-test/src/common/intrinsic_helpers.rs

@@ -1,3 +1,4 @@
+use std::collections::HashMap;
 use std::fmt;
 use std::ops::Deref;
 use std::str::FromStr;
@@ -121,7 +122,8 @@ pub struct IntrinsicType {
     /// A value of `None` can be assumed to be 1 though.
     pub vec_len: Option<u32>,
 
-    pub target: String,
+    // pub target: String,
+    pub metadata: HashMap<String, String>,
 }
 
 impl IntrinsicType {
@@ -133,7 +135,7 @@ impl IntrinsicType {
         if let Some(bl) = self.bit_len {
             bl
         } else {
-            unreachable!("")
+            unreachable!("{}", self.kind)
         }
     }
 
@@ -153,6 +155,14 @@ impl IntrinsicType {
         self.ptr
     }
 
+    // pub fn set_bit_len(&mut self, value: Option<u32>) {
+    //     self.bit_len = value;
+    // }
+
+    pub fn set_metadata(&mut self, key: String, value: String) {
+        self.metadata.insert(key, value);
+    }
+
     pub fn c_scalar_type(&self) -> String {
         match self.kind() {
             TypeKind::Char(_) => String::from("char"),
@@ -322,7 +332,7 @@ pub trait IntrinsicTypeDefinition: Deref<Target = IntrinsicType> {
     fn get_lane_function(&self) -> String;
 
     /// can be implemented in an `impl` block
-    fn from_c(_s: &str, _target: &str) -> Result<Self, String>
+    fn from_c(_s: &str) -> Result<Self, String>
     where
         Self: Sized;
 

crates/intrinsic-test/src/common/mod.rs

@@ -22,3 +22,10 @@ pub trait SupportedArchitectureTest {
     fn build_rust_file(&self) -> bool;
     fn compare_outputs(&self) -> bool;
 }
+
+pub fn chunk_info(intrinsic_count: usize) -> (usize, usize) {
+    let available_parallelism = std::thread::available_parallelism().unwrap().get();
+    let chunk_size = intrinsic_count.div_ceil(Ord::min(available_parallelism, intrinsic_count));
+
+    (chunk_size, intrinsic_count.div_ceil(chunk_size))
+}