Merge pull request #3228 from rtfeldman/illegal_instruction_PENTIUM_N3700

WIP don't require avx, avx2, sse, sse4.2 on all x86-64

.cargo/config

@@ -9,6 +9,10 @@ test-gen-wasm = "test -p roc_gen_wasm -p test_gen --no-default-features --featur
 # lto=fat       Spend extra effort on link-time optimization across crates
 rustflags = ["-Copt-level=s", "-Clto=fat"]
 
+[target.'cfg(not(target = "wasm32-unknown-unknown"))']
+# Sets the avx, avx2, sse2 and sse4.2 target-features correctly based on your CPU.
+rustflags = ["-Ctarget-cpu=native"]
+
 [env]
 # Gives us the path of the workspace root for use in cargo tests without having
 # to compute it per-package.

Earthfile

@@ -146,6 +146,8 @@ build-nightly-release:
 # compile everything needed for benchmarks and output a self-contained dir from which benchmarks can be run.
 prep-bench-folder:
     FROM +copy-dirs
+    # to make use of avx, avx2, sse2, sse4.2... instructions
+    ENV RUSTFLAGS="-C link-arg=-fuse-ld=lld -C target-cpu=native"
     ARG BENCH_SUFFIX=branch
     RUN cargo criterion -V
     RUN --mount=type=cache,target=$SCCACHE_DIR cd cli && cargo criterion --no-run

compiler/builtins/build.rs

@@ -164,7 +164,9 @@ where
                 };
 
                 // flaky test error that only occurs sometimes inside MacOS ci run
-                if error_str.contains("unable to build stage1 zig object: FileNotFound") {
+                if error_str.contains("unable to build stage1 zig object: FileNotFound")
+                    || error_str.contains("unable to save cached ZIR code")
+                {
                     run_command(path, command_str, args)
                 } else {
                     panic!("{} failed: {}", command_str, error_str);

compiler/collections/src/small_string_interner.rs

@@ -251,14 +251,27 @@ impl SmallStringInterner {
 
 #[allow(dead_code)]
 fn find_i16_slice(slice: &[i16], key: i16) -> Option<usize> {
-    #[cfg(target_arch = "x86_64")]
+    // run with RUSTFLAGS="-C target-cpu=native" to enable
+    #[cfg(all(
+        target_arch = "x86_64",
+        target_feature = "avx",
+        target_feature = "avx2"
+    ))]
     return find_i16_slice_x86_64(slice, key);
 
-    #[cfg(not(target_arch = "x86_64"))]
+    #[cfg(not(all(
+        target_arch = "x86_64",
+        target_feature = "avx",
+        target_feature = "avx2"
+    )))]
     return find_i16_slice_fallback(slice, key);
 }
 
-#[cfg(target_arch = "x86_64")]
+#[cfg(all(
+    target_arch = "x86_64",
+    target_feature = "avx",
+    target_feature = "avx2"
+))]
 fn find_i16_slice_x86_64(slice: &[i16], key: i16) -> Option<usize> {
     use std::arch::x86_64::*;
 
@@ -332,7 +345,11 @@ mod test {
         assert!(interner.find_and_update("c", "cd").is_some());
     }
 
-    #[cfg(target_arch = "x86_64")]
+    #[cfg(all(
+        target_arch = "x86_64",
+        target_feature = "avx",
+        target_feature = "avx2"
+    ))]
     #[test]
     fn find_test_1() {
         use super::find_i16_slice;

compiler/parse/src/blankspace.rs

@@ -254,7 +254,12 @@ fn fast_eat_spaces(state: &State) -> FastSpaceState {
                 index += 1;
 
                 // try to use SIMD instructions explicitly
-                #[cfg(target_arch = "x86_64")]
+                // run with RUSTFLAGS="-C target-cpu=native" to enable
+                #[cfg(all(
+                    target_arch = "x86_64",
+                    target_feature = "sse2",
+                    target_feature = "sse4.2"
+                ))]
                 {
                     use std::arch::x86_64::*;
 
@@ -287,7 +292,11 @@ fn fast_eat_spaces(state: &State) -> FastSpaceState {
                     }
                 }
 
-                #[cfg(not(target_arch = "x86_64"))]
+                #[cfg(not(all(
+                    target_arch = "x86_64",
+                    target_feature = "sse2",
+                    target_feature = "sse4.2"
+                )))]
                 {
                     while index < length {
                         match bytes[index] {
@@ -431,7 +440,11 @@ fn eat_line_comment<'a>(
 
         let loop_start = index;
 
-        #[cfg(target_arch = "x86_64")]
+        #[cfg(all(
+            target_arch = "x86_64",
+            target_feature = "sse2",
+            target_feature = "sse4.2"
+        ))]
         {
             use std::arch::x86_64::*;
 
@@ -519,7 +532,11 @@ fn eat_line_comment<'a>(
             }
         }
 
-        #[cfg(not(target_arch = "x86_64"))]
+        #[cfg(not(all(
+            target_arch = "x86_64",
+            target_feature = "sse2",
+            target_feature = "sse4.2"
+        )))]
         while index < length {
             match bytes[index] {
                 b'\t' => unreachable!(),