rustc: Add support for some more x86 SIMD ops

This commit adds compiler support for two basic operations needed for binding
SIMD on x86 platforms:

* First, a `nontemporal_store` intrinsic was added for the `_mm_stream_ps`, seen
  in rust-lang-nursery/stdsimd#114. This was relatively straightforward and is
  quite similar to the volatile store intrinsic.

* Next, and much more intrusively, a new type to the backend was added. The
  `x86_mmx` type is used in LLVM for a 64-bit vector register and is used in
  various intrinsics like `_mm_abs_pi8` as seen in rust-lang-nursery/stdsimd#74.
  This new type was added as a new layout option as well as having support added
  to the trans backend. The type is enabled with the `#[repr(x86_mmx)]`
  attribute which is intended to just be an implementation detail of SIMD in
  Rust.

I'm not 100% certain about how the `x86_mmx` type was added, so any extra eyes
or thoughts on that would be greatly appreciated!

src/libcore/intrinsics.rs

@@ -1387,4 +1387,9 @@ pub fn volatile_copy_nonoverlapping_memory<T>(dst: *mut T, src: *const T,
     /// # } }
     /// ```
     pub fn align_offset(ptr: *const (), align: usize) -> usize;
+
+    /// Emits a `!nontemporal` store according to LLVM (see their docs).
+    /// Probably will never become stable.
+    #[cfg(not(stage0))]
+    pub fn nontemporal_store<T>(ptr: *mut T, val: T);
 }

src/librustc_llvm/ffi.rs

@@ -587,6 +587,7 @@ pub fn LLVMStructTypeInContext(C: ContextRef,
 
     // Operations on other types
     pub fn LLVMVoidTypeInContext(C: ContextRef) -> TypeRef;
+    pub fn LLVMX86MMXTypeInContext(C: ContextRef) -> TypeRef;
     pub fn LLVMRustMetadataTypeInContext(C: ContextRef) -> TypeRef;
 
     // Operations on all values

src/librustc_trans/builder.rs

@@ -612,6 +612,29 @@ pub fn atomic_store(&self, val: ValueRef, ptr: ValueRef,
         }
     }
 
+    pub fn nontemporal_store(&self, val: ValueRef, ptr: ValueRef) -> ValueRef {
+        debug!("Store {:?} -> {:?}", Value(val), Value(ptr));
+        assert!(!self.llbuilder.is_null());
+        self.count_insn("store.nontemporal");
+        let ptr = self.check_store(val, ptr);
+        unsafe {
+            let insn = llvm::LLVMBuildStore(self.llbuilder, val, ptr);
+
+            // According to LLVM [1] building a nontemporal store must *always*
+            // point to a metadata value of the integer 1. Who knew?
+            //
+            // [1]: http://llvm.org/docs/LangRef.html#store-instruction
+            let one = C_i32(self.ccx, 1);
+            let node = llvm::LLVMMDNodeInContext(self.ccx.llcx(),
+                                                 &one,
+                                                 1);
+            llvm::LLVMSetMetadata(insn,
+                                  llvm::MD_nontemporal as c_uint,
+                                  node);
+            insn
+        }
+    }
+
     pub fn gep(&self, ptr: ValueRef, indices: &[ValueRef]) -> ValueRef {
         self.count_insn("gep");
         unsafe {

src/librustc_trans/diagnostics.rs

@@ -37,13 +37,13 @@ fn main() {
 
 #[repr(simd)]
 #[derive(Copy, Clone)]
-struct i32x1(i32);
+struct i32x2(i32, i32);
 
 extern "platform-intrinsic" {
     fn simd_add<T>(a: T, b: T) -> T;
 }
 
-unsafe { simd_add(i32x1(0), i32x1(1)); } // ok!
+unsafe { simd_add(i32x2(0, 0), i32x2(1, 2)); } // ok!
 ```
 "##,
 

src/librustc_trans/intrinsic.rs

@@ -540,6 +540,22 @@ pub fn trans_intrinsic_call<'a, 'tcx>(bcx: &Builder<'a, 'tcx>,
             }
         }
 
+        "nontemporal_store" => {
+            let tp_ty = substs.type_at(0);
+            let dst = args[0].deref(bcx.ccx);
+            let val = if let OperandValue::Ref(ptr, align) = args[1].val {
+                bcx.load(ptr, align.non_abi())
+            } else {
+                from_immediate(bcx, args[1].immediate())
+            };
+            let ptr = bcx.pointercast(dst.llval, val_ty(val).ptr_to());
+            let store = bcx.nontemporal_store(val, ptr);
+            unsafe {
+                llvm::LLVMSetAlignment(store, ccx.align_of(tp_ty).abi() as u32);
+            }
+            return
+        }
+
         _ => {
             let intr = match Intrinsic::find(&name) {
                 Some(intr) => intr,

src/librustc_trans/type_.rs

@@ -286,4 +286,8 @@ pub fn pointee_for_abi_align(ccx: &CrateContext, align: Align) -> Type {
             Type::i8(ccx)
         }
     }
+
+    pub fn x86_mmx(ccx: &CrateContext) -> Type {
+        ty!(llvm::LLVMX86MMXTypeInContext(ccx.llcx()))
+    }
 }

src/librustc_trans/type_of.rs

@@ -26,8 +26,23 @@ fn uncached_llvm_type<'a, 'tcx>(ccx: &CrateContext<'a, 'tcx>,
     match layout.abi {
         layout::Abi::Scalar(_) => bug!("handled elsewhere"),
         layout::Abi::Vector => {
-            return Type::vector(&layout.field(ccx, 0).llvm_type(ccx),
-                                layout.fields.count() as u64);
+            // LLVM has a separate type for 64-bit SIMD vectors on X86 called
+            // `x86_mmx` which is needed for some SIMD operations. As a bit of a
+            // hack (all SIMD definitions are super unstable anyway) we
+            // recognize any one-element SIMD vector as "this should be an
+            // x86_mmx" type. In general there shouldn't be a need for other
+            // one-element SIMD vectors, so it's assumed this won't clash with
+            // much else.
+            let use_x86_mmx = layout.fields.count() == 1 &&
+                layout.size.bits() == 64 &&
+                (ccx.sess().target.target.arch == "x86" ||
+                 ccx.sess().target.target.arch == "x86_64");
+            if use_x86_mmx {
+                return Type::x86_mmx(ccx)
+            } else {
+                return Type::vector(&layout.field(ccx, 0).llvm_type(ccx),
+                                    layout.fields.count() as u64);
+            }
         }
         layout::Abi::ScalarPair(..) => {
             return Type::struct_(ccx, &[

src/librustc_typeck/check/intrinsic.rs

@@ -318,6 +318,10 @@ pub fn check_intrinsic_type<'a, 'tcx>(tcx: TyCtxt<'a, 'tcx, 'tcx>,
                 (0, vec![ptr_ty, tcx.types.usize], tcx.types.usize)
             },
 
+            "nontemporal_store" => {
+                (1, vec![ tcx.mk_mut_ptr(param(0)), param(0) ], tcx.mk_nil())
+            }
+
             ref other => {
                 struct_span_err!(tcx.sess, it.span, E0093,
                                 "unrecognized intrinsic function: `{}`",

src/test/codegen/nontemporal.rs

+// Copyright 2017 The Rust Project Developers. See the COPYRIGHT
+// file at the top-level directory of this distribution and at
+// http://rust-lang.org/COPYRIGHT.
+//
+// Licensed under the Apache License, Version 2.0 <LICENSE-APACHE or
+// http://www.apache.org/licenses/LICENSE-2.0> or the MIT license
+// <LICENSE-MIT or http://opensource.org/licenses/MIT>, at your
+// option. This file may not be copied, modified, or distributed
+// except according to those terms.
+
+// compile-flags: -O
+
+#![feature(core_intrinsics)]
+#![crate_type = "lib"]
+
+#[no_mangle]
+pub fn a(a: &mut u32, b: u32) {
+    // CHECK-LABEL: define void @a
+    // CHECK: store i32 %b, i32* %a, align 4, !nontemporal
+    unsafe {
+        std::intrinsics::nontemporal_store(a, b);
+    }
+}

src/test/codegen/x86_mmx.rs

+// Copyright 2017 The Rust Project Developers. See the COPYRIGHT
+// file at the top-level directory of this distribution and at
+// http://rust-lang.org/COPYRIGHT.
+//
+// Licensed under the Apache License, Version 2.0 <LICENSE-APACHE or
+// http://www.apache.org/licenses/LICENSE-2.0> or the MIT license
+// <LICENSE-MIT or http://opensource.org/licenses/MIT>, at your
+// option. This file may not be copied, modified, or distributed
+// except according to those terms.
+
+// ignore-arm
+// ignore-aarch64
+// ignore-emscripten
+// compile-flags: -O
+
+#![feature(repr_simd)]
+#![crate_type="lib"]
+
+#[repr(simd)]
+#[derive(Clone, Copy)]
+pub struct i8x8(u64);
+
+#[no_mangle]
+pub fn a(a: &mut i8x8, b: i8x8) -> i8x8 {
+    // CHECK-LABEL: define x86_mmx @a(x86_mmx*{{.*}}, x86_mmx{{.*}})
+    // CHECK: store x86_mmx %b, x86_mmx* %a
+    // CHECK: ret x86_mmx %b
+    *a = b;
+    return b
+}