Do not always enable avx2

Author: antoyo

src/abi.rs

@@ -3,6 +3,7 @@ use rustc_codegen_ssa::traits::{AbiBuilderMethods, BaseTypeMethods};
 use rustc_data_structures::fx::FxHashSet;
 use rustc_middle::bug;
 use rustc_middle::ty::Ty;
+#[cfg(feature = "master")]
 use rustc_session::config;
 use rustc_target::abi::call::{ArgAttributes, CastTarget, FnAbi, PassMode, Reg, RegKind};
 

src/base.rs

@@ -1,15 +1,12 @@
 use std::collections::HashSet;
 use std::env;
-use std::sync::Arc;
 use std::time::Instant;
 
 use gccjit::{
     Context,
     FunctionType,
     GlobalKind,
 };
-#[cfg(feature="master")]
-use gccjit::TargetInfo;
 use rustc_middle::dep_graph;
 use rustc_middle::ty::TyCtxt;
 #[cfg(feature="master")]
@@ -22,8 +19,7 @@ use rustc_codegen_ssa::traits::DebugInfoMethods;
 use rustc_session::config::DebugInfo;
 use rustc_span::Symbol;
 
-#[cfg(not(feature="master"))]
-use crate::TargetInfo;
+use crate::LockedTargetInfo;
 use crate::GccContext;
 use crate::builder::Builder;
 use crate::context::CodegenCx;
@@ -70,7 +66,7 @@ pub fn linkage_to_gcc(linkage: Linkage) -> FunctionType {
     }
 }
 
-pub fn compile_codegen_unit(tcx: TyCtxt<'_>, cgu_name: Symbol, target_info: Arc<TargetInfo>) -> (ModuleCodegen<GccContext>, u64) {
+pub fn compile_codegen_unit(tcx: TyCtxt<'_>, cgu_name: Symbol, target_info: LockedTargetInfo) -> (ModuleCodegen<GccContext>, u64) {
     let prof_timer = tcx.prof.generic_activity("codegen_module");
     let start_time = Instant::now();
 
@@ -89,7 +85,7 @@ pub fn compile_codegen_unit(tcx: TyCtxt<'_>, cgu_name: Symbol, target_info: Arc<
     // the time we needed for codegenning it.
     let cost = time_to_codegen.as_secs() * 1_000_000_000 + time_to_codegen.subsec_nanos() as u64;
 
-    fn module_codegen(tcx: TyCtxt<'_>, (cgu_name, target_info): (Symbol, Arc<TargetInfo>)) -> ModuleCodegen<GccContext> {
+    fn module_codegen(tcx: TyCtxt<'_>, (cgu_name, target_info): (Symbol, LockedTargetInfo)) -> ModuleCodegen<GccContext> {
         let cgu = tcx.codegen_unit(cgu_name);
         // Instantiate monomorphizations without filling out definitions yet...
         let context = Context::default();
@@ -111,14 +107,20 @@ pub fn compile_codegen_unit(tcx: TyCtxt<'_>, cgu_name: Symbol, target_info: Arc<
         // TODO(antoyo): only set on x86 platforms.
         context.add_command_line_option("-masm=intel");
 
-        let features = ["64", "avxvnni", "bmi", "sse2", "avx", "avx2", "sha", "fma", "fma4", "gfni", "f16c", "aes", "bmi2", "pclmul", "rtm",
+        // TODO: instead of setting the features manually, set the correct -march flag.
+        let features = ["64", "avxvnni", "bmi", "sse2", "avx2", "sha", "fma", "fma4", "gfni", "f16c", "aes", "bmi2", "pclmul", "rtm",
             "vaes", "vpclmulqdq", "xsavec",
         ];
 
         for feature in &features {
             add_cpu_feature_flag(feature);
         }
 
+        // NOTE: we always enable AVX because the equivalent of llvm.x86.sse2.cmp.pd in GCC for
+        // SSE2 is multiple builtins, so we use the AVX __builtin_ia32_cmppd instead.
+        // FIXME(antoyo): use the proper builtins for llvm.x86.sse2.cmp.pd and similar.
+        context.add_command_line_option("-mavx");
+
         for arg in &tcx.sess.opts.cg.llvm_args {
             context.add_command_line_option(arg);
         }

src/intrinsic/llvm.rs

@@ -236,6 +236,7 @@ pub fn adjust_intrinsic_arguments<'a, 'b, 'gcc, 'tcx>(builder: &Builder<'a, 'gcc
                 let arg2 = builder.context.new_cast(None, arg2, arg2_type);
                 args = vec![new_args[0], arg2].into();
             },
+            // These builtins are sent one more argument than needed.
             "__builtin_prefetch" => {
                 let mut new_args = args.to_vec();
                 new_args.pop();

src/lib.rs

@@ -73,6 +73,7 @@ mod type_of;
 
 use std::any::Any;
 use std::sync::Arc;
+use std::sync::Mutex;
 #[cfg(not(feature="master"))]
 use std::sync::atomic::AtomicBool;
 #[cfg(not(feature="master"))]
@@ -135,9 +136,24 @@ impl TargetInfo {
     }
 }
 
+#[derive(Clone, Debug)]
+pub struct LockedTargetInfo {
+    info: Arc<Mutex<TargetInfo>>,
+}
+
+impl LockedTargetInfo {
+    fn cpu_supports(&self, feature: &str) -> bool {
+        self.info.lock().expect("lock").cpu_supports(feature)
+    }
+
+    fn supports_128bit_int(&self) -> bool {
+        self.info.lock().expect("lock").supports_128bit_int()
+    }
+}
+
 #[derive(Clone)]
 pub struct GccCodegenBackend {
-    target_info: Arc<TargetInfo>,
+    target_info: LockedTargetInfo,
 }
 
 impl CodegenBackend for GccCodegenBackend {
@@ -146,6 +162,17 @@ impl CodegenBackend for GccCodegenBackend {
     }
 
     fn init(&self, sess: &Session) {
+        #[cfg(feature="master")]
+        {
+            let target_cpu = target_cpu(sess);
+
+            // Get the second TargetInfo with the correct CPU features by setting the arch.
+            let context = Context::default();
+            context.add_command_line_option(&format!("-march={}", target_cpu));
+
+            *self.target_info.info.lock().expect("lock") = context.get_target_info();
+        }
+
         #[cfg(feature="master")]
         gccjit::set_global_personality_function_name(b"rust_eh_personality\0");
         if sess.lto() == Lto::Thin {
@@ -161,7 +188,7 @@ impl CodegenBackend for GccCodegenBackend {
             let _int128_ty = check_context.new_c_type(CType::UInt128t);
             // NOTE: we cannot just call compile() as this would require other files than libgccjit.so.
             check_context.compile_to_file(gccjit::OutputKind::Assembler, temp_file.to_str().expect("path to str"));
-            self.target_info.supports_128bit_integers.store(check_context.get_last_error() == Ok(None), Ordering::SeqCst);
+            self.target_info.info.lock().expect("lock").supports_128bit_integers.store(check_context.get_last_error() == Ok(None), Ordering::SeqCst);
         }
     }
 
@@ -217,7 +244,7 @@ impl ExtraBackendMethods for GccCodegenBackend {
     }
 
     fn compile_codegen_unit(&self, tcx: TyCtxt<'_>, cgu_name: Symbol) -> (ModuleCodegen<Self::Module>, u64) {
-        base::compile_codegen_unit(tcx, cgu_name, Arc::clone(&self.target_info))
+        base::compile_codegen_unit(tcx, cgu_name, self.target_info.clone())
     }
 
     fn target_machine_factory(&self, _sess: &Session, _opt_level: OptLevel, _features: &[String]) -> TargetMachineFactoryFn<Self> {
@@ -306,23 +333,18 @@ impl WriteBackendMethods for GccCodegenBackend {
 #[no_mangle]
 pub fn __rustc_codegen_backend() -> Box<dyn CodegenBackend> {
     #[cfg(feature="master")]
-    let target_info = {
-        // Get the native arch and check whether the target supports 128-bit integers.
-        let context = Context::default();
-        let arch = context.get_target_info().arch().unwrap();
-
-        // Get the second TargetInfo with the correct CPU features by setting the arch.
+    let info = {
+        // Check whether the target supports 128-bit integers.
         let context = Context::default();
-        context.add_command_line_option(&format!("-march={}", arch.to_str().unwrap()));
-        Arc::new(context.get_target_info())
+        Arc::new(Mutex::new(context.get_target_info()))
     };
     #[cfg(not(feature="master"))]
-    let target_info = Arc::new(TargetInfo {
+    let info = Arc::new(Mutex::new(TargetInfo {
         supports_128bit_integers: AtomicBool::new(false),
-    });
+    }));
 
     Box::new(GccCodegenBackend {
-        target_info,
+        target_info: LockedTargetInfo { info },
     })
 }
 
@@ -356,7 +378,7 @@ pub fn target_cpu(sess: &Session) -> &str {
     }
 }
 
-pub fn target_features(sess: &Session, allow_unstable: bool, target_info: &Arc<TargetInfo>) -> Vec<Symbol> {
+pub fn target_features(sess: &Session, allow_unstable: bool, target_info: &LockedTargetInfo) -> Vec<Symbol> {
     supported_target_features(sess)
         .iter()
         .filter_map(