diff --git a/Cargo.lock b/Cargo.lock
index 6df2102470f..ff09a08b5af 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -41,7 +41,7 @@ dependencies = [
 [[package]]
 name = "gccjit"
 version = "1.0.0"
-source = "git+https://github.com/antoyo/gccjit.rs#bdb86fb5092895ff5589726b33250010c64d93f6"
+source = "git+https://github.com/antoyo/gccjit.rs#a8997afb665dc467c1bdbddf04877143683f0cce"
 dependencies = [
  "gccjit_sys",
 ]
@@ -49,7 +49,7 @@ dependencies = [
 [[package]]
 name = "gccjit_sys"
 version = "0.0.1"
-source = "git+https://github.com/antoyo/gccjit.rs#bdb86fb5092895ff5589726b33250010c64d93f6"
+source = "git+https://github.com/antoyo/gccjit.rs#a8997afb665dc467c1bdbddf04877143683f0cce"
 dependencies = [
  "libc 0.1.12",
 ]
@@ -215,6 +215,7 @@ dependencies = [
  "ar",
  "gccjit",
  "lang_tester",
+ "smallvec",
  "target-lexicon",
  "tempfile",
 ]
@@ -228,6 +229,12 @@ dependencies = [
  "winapi-util",
 ]
 
+[[package]]
+name = "smallvec"
+version = "1.8.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f2dd574626839106c320a323308629dcb1acfc96e32a8cba364ddc61ac23ee83"
+
 [[package]]
 name = "target-lexicon"
 version = "0.10.0"
diff --git a/Cargo.toml b/Cargo.toml
index 211d19a8dc8..3ac354ea494 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -27,6 +27,7 @@ gccjit = { git = "https://github.com/antoyo/gccjit.rs" }
 # Local copy.
 #gccjit = { path = "../gccjit.rs" }
 
+smallvec = { version = "1.6.1", features = ["union", "may_dangle"] }
 target-lexicon = "0.10.0"
 
 ar = "0.8.0"
diff --git a/patches/0001-Add-stdarch-Cargo.toml-for-testing.patch b/patches/0001-Add-stdarch-Cargo.toml-for-testing.patch
new file mode 100644
index 00000000000..93c63b5dcac
--- /dev/null
+++ b/patches/0001-Add-stdarch-Cargo.toml-for-testing.patch
@@ -0,0 +1,39 @@
+From c3821e02fbd6cb5ad6e06d759fccdc9073712375 Mon Sep 17 00:00:00 2001
+From: Antoni Boucher <bouanto@zoho.com>
+Date: Tue, 7 Jun 2022 21:40:13 -0400
+Subject: [PATCH] Add stdarch Cargo.toml for testing
+
+---
+ library/stdarch/Cargo.toml | 20 ++++++++++++++++++++
+ 1 file changed, 20 insertions(+)
+ create mode 100644 library/stdarch/Cargo.toml
+
+diff --git a/library/stdarch/Cargo.toml b/library/stdarch/Cargo.toml
+new file mode 100644
+index 0000000..fbe0a95
+--- /dev/null
++++ b/library/stdarch/Cargo.toml
+@@ -0,0 +1,20 @@
++[workspace]
++members = [
++  "crates/core_arch",
++  "crates/std_detect",
++  "crates/stdarch-gen",
++  "examples/"
++]
++exclude = [
++  "crates/wasm-assert-instr-tests"
++]
++
++[profile.release]
++debug = true
++opt-level = 3
++incremental = true
++
++[profile.bench]
++debug = 1
++opt-level = 3
++incremental = true
+-- 
+2.26.2.7.g19db9cfb68.dirty
+
diff --git a/src/attributes.rs b/src/attributes.rs
new file mode 100644
index 00000000000..e570b583dec
--- /dev/null
+++ b/src/attributes.rs
@@ -0,0 +1,112 @@
+#[cfg_attr(not(feature="master"), allow(unused_imports))]
+use gccjit::FnAttribute;
+use gccjit::Function;
+use rustc_attr::InstructionSetAttr;
+use rustc_codegen_ssa::target_features::tied_target_features;
+use rustc_data_structures::fx::FxHashMap;
+use rustc_middle::ty;
+use rustc_session::Session;
+use rustc_span::symbol::sym;
+use smallvec::{smallvec, SmallVec};
+
+use crate::context::CodegenCx;
+
+// Given a map from target_features to whether they are enabled or disabled,
+// ensure only valid combinations are allowed.
+pub fn check_tied_features(sess: &Session, features: &FxHashMap<&str, bool>) -> Option<&'static [&'static str]> {
+    for tied in tied_target_features(sess) {
+        // Tied features must be set to the same value, or not set at all
+        let mut tied_iter = tied.iter();
+        let enabled = features.get(tied_iter.next().unwrap());
+        if tied_iter.any(|feature| enabled != features.get(feature)) {
+            return Some(tied);
+        }
+    }
+    None
+}
+
+// TODO(antoyo): maybe move to a new module gcc_util.
+// To find a list of GCC's names, check https://gcc.gnu.org/onlinedocs/gcc/Function-Attributes.html
+fn to_gcc_features<'a>(sess: &Session, s: &'a str) -> SmallVec<[&'a str; 2]> {
+    let arch = if sess.target.arch == "x86_64" { "x86" } else { &*sess.target.arch };
+    match (arch, s) {
+        ("x86", "sse4.2") => smallvec!["sse4.2", "crc32"],
+        ("x86", "pclmulqdq") => smallvec!["pclmul"],
+        ("x86", "rdrand") => smallvec!["rdrnd"],
+        ("x86", "bmi1") => smallvec!["bmi"],
+        ("x86", "cmpxchg16b") => smallvec!["cx16"],
+        ("x86", "avx512vaes") => smallvec!["vaes"],
+        ("x86", "avx512gfni") => smallvec!["gfni"],
+        ("x86", "avx512vpclmulqdq") => smallvec!["vpclmulqdq"],
+        // NOTE: seems like GCC requires 'avx512bw' for 'avx512vbmi2'.
+        ("x86", "avx512vbmi2") => smallvec!["avx512vbmi2", "avx512bw"],
+        // NOTE: seems like GCC requires 'avx512bw' for 'avx512bitalg'.
+        ("x86", "avx512bitalg") => smallvec!["avx512bitalg", "avx512bw"],
+        ("aarch64", "rcpc2") => smallvec!["rcpc-immo"],
+        ("aarch64", "dpb") => smallvec!["ccpp"],
+        ("aarch64", "dpb2") => smallvec!["ccdp"],
+        ("aarch64", "frintts") => smallvec!["fptoint"],
+        ("aarch64", "fcma") => smallvec!["complxnum"],
+        ("aarch64", "pmuv3") => smallvec!["perfmon"],
+        ("aarch64", "paca") => smallvec!["pauth"],
+        ("aarch64", "pacg") => smallvec!["pauth"],
+        // Rust ties fp and neon together. In LLVM neon implicitly enables fp,
+        // but we manually enable neon when a feature only implicitly enables fp
+        ("aarch64", "f32mm") => smallvec!["f32mm", "neon"],
+        ("aarch64", "f64mm") => smallvec!["f64mm", "neon"],
+        ("aarch64", "fhm") => smallvec!["fp16fml", "neon"],
+        ("aarch64", "fp16") => smallvec!["fullfp16", "neon"],
+        ("aarch64", "jsconv") => smallvec!["jsconv", "neon"],
+        ("aarch64", "sve") => smallvec!["sve", "neon"],
+        ("aarch64", "sve2") => smallvec!["sve2", "neon"],
+        ("aarch64", "sve2-aes") => smallvec!["sve2-aes", "neon"],
+        ("aarch64", "sve2-sm4") => smallvec!["sve2-sm4", "neon"],
+        ("aarch64", "sve2-sha3") => smallvec!["sve2-sha3", "neon"],
+        ("aarch64", "sve2-bitperm") => smallvec!["sve2-bitperm", "neon"],
+        (_, s) => smallvec![s],
+    }
+}
+
+/// Composite function which sets GCC attributes for function depending on its AST (`#[attribute]`)
+/// attributes.
+pub fn from_fn_attrs<'gcc, 'tcx>(
+    cx: &CodegenCx<'gcc, 'tcx>,
+    #[cfg_attr(not(feature="master"), allow(unused_variables))]
+    func: Function<'gcc>,
+    instance: ty::Instance<'tcx>,
+) {
+    let codegen_fn_attrs = cx.tcx.codegen_fn_attrs(instance.def_id());
+
+    let function_features =
+        codegen_fn_attrs.target_features.iter().map(|features| features.as_str()).collect::<Vec<&str>>();
+
+    if let Some(features) = check_tied_features(cx.tcx.sess, &function_features.iter().map(|features| (*features, true)).collect()) {
+        let span = cx.tcx
+            .get_attr(instance.def_id(), sym::target_feature)
+            .map_or_else(|| cx.tcx.def_span(instance.def_id()), |a| a.span);
+        let msg = format!("the target features {} must all be either enabled or disabled together", features.join(", "));
+        let mut err = cx.tcx.sess.struct_span_err(span, &msg);
+        err.help("add the missing features in a `target_feature` attribute");
+        err.emit();
+        return;
+    }
+
+    let mut function_features = function_features
+        .iter()
+        .flat_map(|feat| to_gcc_features(cx.tcx.sess, feat).into_iter())
+        .chain(codegen_fn_attrs.instruction_set.iter().map(|x| match x {
+            InstructionSetAttr::ArmA32 => "-thumb-mode", // TODO(antoyo): support removing feature.
+            InstructionSetAttr::ArmT32 => "thumb-mode",
+        }))
+        .collect::<Vec<_>>();
+
+    // TODO(antoyo): check if we really need global backend features. (Maybe they could be applied
+    // globally?)
+    let mut global_features = cx.tcx.global_backend_features(()).iter().map(|s| s.as_str());
+    function_features.extend(&mut global_features);
+    let target_features = function_features.join(",");
+    if !target_features.is_empty() {
+        #[cfg(feature="master")]
+        func.add_attribute(FnAttribute::Target, &target_features);
+    }
+}
diff --git a/src/base.rs b/src/base.rs
index 19c981309d7..9d81a01e7a6 100644
--- a/src/base.rs
+++ b/src/base.rs
@@ -83,15 +83,23 @@ pub fn compile_codegen_unit<'tcx>(tcx: TyCtxt<'tcx>, cgu_name: Symbol, supports_
         context.add_command_line_option("-mavx2");
         // FIXME(antoyo): the following causes an illegal instruction on vmovdqu64 in std_example on my CPU.
         // Only add if the CPU supports it.
-        /*context.add_command_line_option("-mavx512f");
         context.add_command_line_option("-msha");
         context.add_command_line_option("-mpclmul");
         context.add_command_line_option("-mfma");
         context.add_command_line_option("-mfma4");
-        context.add_command_line_option("-mavx512vpopcntdq");
-        context.add_command_line_option("-mavx512vl");
         context.add_command_line_option("-m64");
-        context.add_command_line_option("-mbmi");*/
+        context.add_command_line_option("-mbmi");
+        context.add_command_line_option("-mgfni");
+        context.add_command_line_option("-mavxvnni");
+        context.add_command_line_option("-mf16c");
+        context.add_command_line_option("-maes");
+        context.add_command_line_option("-mxsavec");
+        context.add_command_line_option("-mbmi2");
+        context.add_command_line_option("-mrtm");
+        context.add_command_line_option("-mvaes");
+        context.add_command_line_option("-mvpclmulqdq");
+        context.add_command_line_option("-mavx");
+
         for arg in &tcx.sess.opts.cg.llvm_args {
             context.add_command_line_option(arg);
         }
diff --git a/src/builder.rs b/src/builder.rs
index 4aad171cb78..867fd531f50 100644
--- a/src/builder.rs
+++ b/src/builder.rs
@@ -213,7 +213,7 @@ impl<'a, 'gcc, 'tcx> Builder<'a, 'gcc, 'tcx> {
 
                 let actual_ty = actual_val.get_type();
                 if expected_ty != actual_ty {
-                    if !actual_ty.is_vector() && !expected_ty.is_vector() && actual_ty.is_integral() && expected_ty.is_integral() {
+                    if !actual_ty.is_vector() && !expected_ty.is_vector() && (actual_ty.is_integral() && expected_ty.is_integral()) || (actual_ty.get_pointee().is_some() && expected_ty.get_pointee().is_some()) {
                         self.context.new_cast(None, actual_val, expected_ty)
                     }
                     else if on_stack_param_indices.contains(&index) {
@@ -1490,6 +1490,9 @@ impl<'a, 'gcc, 'tcx> Builder<'a, 'gcc, 'tcx> {
         let zeros = self.context.new_rvalue_from_vector(None, cond_type, &zeros);
 
         let masks = self.context.new_comparison(None, ComparisonOp::NotEquals, cond, zeros);
+        // NOTE: masks is a vector of integers, but the values can be vectors of floats, so use bitcast to make
+        // the & operation work.
+        let masks = self.bitcast_if_needed(masks, then_val.get_type());
         let then_vals = masks & then_val;
 
         let ones = vec![self.context.new_rvalue_one(element_type); num_units];
@@ -1509,6 +1512,16 @@ impl<'a, 'gcc, 'tcx> Builder<'a, 'gcc, 'tcx> {
 fn difference_or_zero<'gcc>(a: RValue<'gcc>, b: RValue<'gcc>, context: &'gcc Context<'gcc>) -> RValue<'gcc> {
     let difference = a - b;
     let masks = context.new_comparison(None, ComparisonOp::GreaterThanEquals, b, a);
+    // NOTE: masks is a vector of integers, but the values can be vectors of floats, so use bitcast to make
+    // the & operation work.
+    let a_type = a.get_type();
+    let masks =
+        if masks.get_type() != a_type {
+            context.new_bitcast(None, masks, a_type)
+        }
+        else {
+            masks
+        };
     difference & masks
 }
 
diff --git a/src/callee.rs b/src/callee.rs
index 76419b103d0..5557f886b28 100644
--- a/src/callee.rs
+++ b/src/callee.rs
@@ -4,6 +4,7 @@ use rustc_middle::ty::{self, Instance, TypeFoldable};
 use rustc_middle::ty::layout::{FnAbiOf, HasTyCtxt};
 
 use crate::abi::FnAbiGccExt;
+use crate::attributes;
 use crate::context::CodegenCx;
 
 /// Codegens a reference to a fn/method item, monomorphizing and
@@ -67,8 +68,12 @@ pub fn get_fn<'gcc, 'tcx>(cx: &CodegenCx<'gcc, 'tcx>, instance: Instance<'tcx>)
             cx.linkage.set(FunctionType::Extern);
             let func = cx.declare_fn(&sym, &fn_abi);
 
+            attributes::from_fn_attrs(cx, func, instance);
+
             // TODO(antoyo): set linkage and attributes.
-            func
+
+            // FIXME(antoyo): this is a wrong cast. That requires changing the compiler API.
+            unsafe { std::mem::transmute(func) }
         };
 
     cx.function_instances.borrow_mut().insert(instance, func);
diff --git a/src/common.rs b/src/common.rs
index ce341406eaf..d55ad87e19b 100644
--- a/src/common.rs
+++ b/src/common.rs
@@ -177,8 +177,18 @@ impl<'gcc, 'tcx> ConstMethods<'tcx> for CodegenCx<'gcc, 'tcx> {
                 }
 
                 let value = self.const_uint_big(self.type_ix(bitsize), data);
-                // TODO(bjorn3): assert size is correct
-                self.const_bitcast(value, ty)
+                let bytesize = layout.size(self).bytes();
+                if bitsize > 1 && ty.is_integral() && bytesize as u32 == ty.get_size() {
+                    // NOTE: since the intrinsic _xabort is called with a bitcast, which
+                    // is non-const, but expects a constant, do a normal cast instead of a bitcast.
+                    // FIXME(antoyo): fix bitcast to work in constant contexts.
+                    // TODO(antoyo): perhaps only use bitcast for pointers?
+                    self.context.new_cast(None, value, ty)
+                }
+                else {
+                    // TODO(bjorn3): assert size is correct
+                    self.const_bitcast(value, ty)
+                }
             }
             Scalar::Ptr(ptr, _size) => {
                 let (alloc_id, offset) = ptr.into_parts();
diff --git a/src/declare.rs b/src/declare.rs
index a619e2f7712..5f6360a7da5 100644
--- a/src/declare.rs
+++ b/src/declare.rs
@@ -79,12 +79,11 @@ impl<'gcc, 'tcx> CodegenCx<'gcc, 'tcx> {
         unsafe { std::mem::transmute(func) }
     }
 
-    pub fn declare_fn(&self, name: &str, fn_abi: &FnAbi<'tcx, Ty<'tcx>>) -> RValue<'gcc> {
+    pub fn declare_fn(&self, name: &str, fn_abi: &FnAbi<'tcx, Ty<'tcx>>) -> Function<'gcc> {
         let (return_type, params, variadic, on_stack_param_indices) = fn_abi.gcc_type(self);
         let func = declare_raw_fn(self, name, () /*fn_abi.llvm_cconv()*/, return_type, &params, variadic);
         self.on_stack_function_params.borrow_mut().insert(func, on_stack_param_indices);
-        // FIXME(antoyo): this is a wrong cast. That requires changing the compiler API.
-        unsafe { std::mem::transmute(func) }
+        func
     }
 
     pub fn define_global(&self, name: &str, ty: Type<'gcc>, is_tls: bool, link_section: Option<Symbol>) -> LValue<'gcc> {
diff --git a/src/intrinsic/llvm.rs b/src/intrinsic/llvm.rs
index f623bc5fb10..f00c1b30109 100644
--- a/src/intrinsic/llvm.rs
+++ b/src/intrinsic/llvm.rs
@@ -287,9 +287,9 @@ pub fn adjust_intrinsic_arguments<'a, 'b, 'gcc, 'tcx>(builder: &Builder<'a, 'gcc
 pub fn adjust_intrinsic_return_value<'a, 'gcc, 'tcx>(builder: &Builder<'a, 'gcc, 'tcx>, mut return_value: RValue<'gcc>, func_name: &str, args: &[RValue<'gcc>], args_adjusted: bool) -> RValue<'gcc> {
     match func_name {
         "__builtin_ia32_vfmaddss3_round" | "__builtin_ia32_vfmaddsd3_round" => {
-            let zero = builder.context.new_rvalue_zero(builder.int_type);
             #[cfg(feature="master")]
             {
+                let zero = builder.context.new_rvalue_zero(builder.int_type);
                 return_value = builder.context.new_vector_access(None, return_value, zero).to_rvalue();
             }
         },
diff --git a/src/intrinsic/simd.rs b/src/intrinsic/simd.rs
index 8f9862414e6..9c2a1401a15 100644
--- a/src/intrinsic/simd.rs
+++ b/src/intrinsic/simd.rs
@@ -1,4 +1,6 @@
-use gccjit::{BinaryOp, RValue, Type, ToRValue, ComparisonOp, UnaryOp};
+#[cfg_attr(not(feature="master"), allow(unused_imports))]
+use gccjit::{ToRValue, ComparisonOp, UnaryOp};
+use gccjit::{BinaryOp, RValue, Type};
 use rustc_codegen_ssa::base::compare_simd_types;
 use rustc_codegen_ssa::common::{TypeKind, span_invalid_monomorphization_error};
 use rustc_codegen_ssa::mir::operand::OperandRef;
@@ -216,7 +218,7 @@ pub fn generic_simd_intrinsic<'a, 'gcc, 'tcx>(bx: &mut Builder<'a, 'gcc, 'tcx>,
         let variable = bx.current_func().new_local(None, vector.get_type(), "new_vector");
         bx.llbb().add_assignment(None, variable, vector);
         let lvalue = bx.context.new_vector_access(None, variable.to_rvalue(), index);
-        // TODO: if simd_insert is constant, use BIT_REF.
+        // TODO(antoyo): if simd_insert is constant, use BIT_REF.
         bx.llbb().add_assignment(None, lvalue, value);
         return Ok(variable.to_rvalue());
     }
@@ -545,9 +547,9 @@ pub fn generic_simd_intrinsic<'a, 'gcc, 'tcx>(bx: &mut Builder<'a, 'gcc, 'tcx>,
                 },
                 (true, true) => {
                     // Algorithm from: https://codereview.stackexchange.com/questions/115869/saturated-signed-addition
-                    // TODO: improve using conditional operators if possible.
+                    // TODO(antoyo): improve using conditional operators if possible.
                     let arg_type = lhs.get_type();
-                    // TODO: convert lhs and rhs to unsigned.
+                    // TODO(antoyo): convert lhs and rhs to unsigned.
                     let sum = lhs + rhs;
                     let vector_type = arg_type.dyncast_vector().expect("vector type");
                     let unit = vector_type.get_num_units();
@@ -581,7 +583,7 @@ pub fn generic_simd_intrinsic<'a, 'gcc, 'tcx>(bx: &mut Builder<'a, 'gcc, 'tcx>,
                     // negative of the right operand. Find a proper subtraction algorithm.
                     let rhs = bx.context.new_unary_op(None, UnaryOp::Minus, arg_type, rhs);
 
-                    // TODO: convert lhs and rhs to unsigned.
+                    // TODO(antoyo): convert lhs and rhs to unsigned.
                     let sum = lhs + rhs;
                     let vector_type = arg_type.dyncast_vector().expect("vector type");
                     let unit = vector_type.get_num_units();
diff --git a/src/lib.rs b/src/lib.rs
index b8db4c9540b..b3cbf344ad1 100644
--- a/src/lib.rs
+++ b/src/lib.rs
@@ -13,6 +13,7 @@
 #![warn(unused_lifetimes)]
 
 extern crate rustc_ast;
+extern crate rustc_attr;
 extern crate rustc_codegen_ssa;
 extern crate rustc_data_structures;
 extern crate rustc_errors;
@@ -32,6 +33,7 @@ mod abi;
 mod allocator;
 mod archive;
 mod asm;
+mod attributes;
 mod back;
 mod base;
 mod builder;
@@ -302,9 +304,11 @@ pub fn target_features(sess: &Session) -> Vec<Symbol> {
         .filter(|_feature| {
             // TODO(antoyo): implement a way to get enabled feature in libgccjit.
             // Probably using the equivalent of __builtin_cpu_supports.
+            // TODO(antoyo): maybe use whatever outputs the following command:
+            // gcc -march=native -Q --help=target
             #[cfg(feature="master")]
             {
-                _feature.contains("sse") || _feature.contains("avx")
+                (_feature.contains("sse") || _feature.contains("avx")) && !_feature.contains("avx512")
             }
             #[cfg(not(feature="master"))]
             {
diff --git a/src/mono_item.rs b/src/mono_item.rs
index e21d40b6c37..60a42846bd3 100644
--- a/src/mono_item.rs
+++ b/src/mono_item.rs
@@ -5,6 +5,7 @@ use rustc_middle::ty::{self, Instance, TypeFoldable};
 use rustc_middle::ty::layout::{FnAbiOf, LayoutOf};
 use rustc_span::def_id::DefId;
 
+use crate::attributes;
 use crate::base;
 use crate::context::CodegenCx;
 use crate::type_of::LayoutGccExt;
@@ -28,9 +29,11 @@ impl<'gcc, 'tcx> PreDefineMethods<'tcx> for CodegenCx<'gcc, 'tcx> {
 
         let fn_abi = self.fn_abi_of_instance(instance, ty::List::empty());
         self.linkage.set(base::linkage_to_gcc(linkage));
-        let _decl = self.declare_fn(symbol_name, &fn_abi);
+        let decl = self.declare_fn(symbol_name, &fn_abi);
         //let attrs = self.tcx.codegen_fn_attrs(instance.def_id());
 
+        attributes::from_fn_attrs(self, decl, instance);
+
         // TODO(antoyo): call set_link_section() to allow initializing argc/argv.
         // TODO(antoyo): set unique comdat.
         // TODO(antoyo): use inline attribute from there in linkage.set() above.
diff --git a/test.sh b/test.sh
index 32fe8cf0586..594b63dfd0d 100755
--- a/test.sh
+++ b/test.sh
@@ -107,7 +107,7 @@ else
 fi
 
 if (( $build_only == 1 )); then
-    echo "Since it's `build-only`, exiting..."
+    echo "Since it's 'build-only', exiting..."
     exit
 fi
 
@@ -324,7 +324,7 @@ function all() {
 }
 
 if [ ${#funcs[@]} -eq 0 ]; then
-    echo "No command passed, running `--all`..."
+    echo "No command passed, running '--all'..."
     all
 else
     for t in ${funcs[@]}; do