Rollup merge of #48983 - gnzlbg:red, r=alexcrichton

add intrinsics for portable packed simd vector reductions

Adds the following portable vector reduction intrinsics:

* fn simd_reduce_add<T, U>(x: T) -> U;
* fn simd_reduce_mul<T, U>(x: T) -> U;
* fn simd_reduce_min<T, U>(x: T) -> U;
* fn simd_reduce_max<T, U>(x: T) -> U;
* fn simd_reduce_and<T, U>(x: T) -> U;
* fn simd_reduce_or<T, U>(x: T) -> U;
* fn simd_reduce_xor<T, U>(x: T) -> U;

I've also added:

* fn simd_reduce_all<T>(x: T) -> bool;
* fn simd_reduce_any<T>(x: T) -> bool;

These produce better code that what we are currently producing in `stdsimd`, but the code is still not optimal due to this LLVM bug:  https://bugs.llvm.org/show_bug.cgi?id=36702

r? @alexcrichton

8 files changed, 728 insertions, 3 deletions

diff --git a/src/librustc_llvm/ffi.rs b/src/librustc_llvm/ffi.rs
index 0ec5700f5f3..c0cdd212770 100644
--- a/src/librustc_llvm/ffi.rs
+++ b/src/librustc_llvm/ffi.rs
@@ -621,6 +621,7 @@ extern "C" {
     pub fn LLVMConstIntGetSExtValue(ConstantVal: ValueRef) -> c_longlong;
     pub fn LLVMRustConstInt128Get(ConstantVal: ValueRef, SExt: bool,
                                   high: *mut u64, low: *mut u64) -> bool;
+    pub fn LLVMConstRealGetDouble (ConstantVal: ValueRef, losesInfo: *mut Bool) -> f64;
 
 
     // Operations on composite constants
@@ -1201,6 +1202,46 @@ extern "C" {
                                 Name: *const c_char)
                                 -> ValueRef;
 
+    pub fn LLVMRustBuildVectorReduceFAdd(B: BuilderRef,
+                                         Acc: ValueRef,
+                                         Src: ValueRef)
+                                         -> ValueRef;
+    pub fn LLVMRustBuildVectorReduceFMul(B: BuilderRef,
+                                         Acc: ValueRef,
+                                         Src: ValueRef)
+                                         -> ValueRef;
+    pub fn LLVMRustBuildVectorReduceAdd(B: BuilderRef,
+                                        Src: ValueRef)
+                                        -> ValueRef;
+    pub fn LLVMRustBuildVectorReduceMul(B: BuilderRef,
+                                        Src: ValueRef)
+                                        -> ValueRef;
+    pub fn LLVMRustBuildVectorReduceAnd(B: BuilderRef,
+                                        Src: ValueRef)
+                                        -> ValueRef;
+    pub fn LLVMRustBuildVectorReduceOr(B: BuilderRef,
+                                       Src: ValueRef)
+                                       -> ValueRef;
+    pub fn LLVMRustBuildVectorReduceXor(B: BuilderRef,
+                                        Src: ValueRef)
+                                        -> ValueRef;
+    pub fn LLVMRustBuildVectorReduceMin(B: BuilderRef,
+                                        Src: ValueRef,
+                                        IsSigned: bool)
+                                        -> ValueRef;
+    pub fn LLVMRustBuildVectorReduceMax(B: BuilderRef,
+                                        Src: ValueRef,
+                                        IsSigned: bool)
+                                        -> ValueRef;
+    pub fn LLVMRustBuildVectorReduceFMin(B: BuilderRef,
+                                         Src: ValueRef,
+                                         IsNaN: bool)
+                                         -> ValueRef;
+    pub fn LLVMRustBuildVectorReduceFMax(B: BuilderRef,
+                                         Src: ValueRef,
+                                         IsNaN: bool)
+                                         -> ValueRef;
+
     pub fn LLVMBuildIsNull(B: BuilderRef, Val: ValueRef, Name: *const c_char) -> ValueRef;
     pub fn LLVMBuildIsNotNull(B: BuilderRef, Val: ValueRef, Name: *const c_char) -> ValueRef;
     pub fn LLVMBuildPtrDiff(B: BuilderRef,
@@ -1567,6 +1608,7 @@ extern "C" {
     pub fn LLVMRustWriteValueToString(value_ref: ValueRef, s: RustStringRef);
 
     pub fn LLVMIsAConstantInt(value_ref: ValueRef) -> ValueRef;
+    pub fn LLVMIsAConstantFP(value_ref: ValueRef) -> ValueRef;
 
     pub fn LLVMRustPassKind(Pass: PassRef) -> PassKind;
     pub fn LLVMRustFindAndCreatePass(Pass: *const c_char) -> PassRef;
diff --git a/src/librustc_trans/builder.rs b/src/librustc_trans/builder.rs
index d4e05a18e3a..3f5a9a54ff1 100644
--- a/src/librustc_trans/builder.rs
+++ b/src/librustc_trans/builder.rs
@@ -955,6 +955,147 @@ impl<'a, 'tcx> Builder<'a, 'tcx> {
         }
     }
 
+    pub fn vector_reduce_fadd_fast(&self, acc: ValueRef, src: ValueRef) -> ValueRef {
+        self.count_insn("vector.reduce.fadd_fast");
+        unsafe {
+            // FIXME: add a non-fast math version once
+            // https://bugs.llvm.org/show_bug.cgi?id=36732
+            // is fixed.
+            let instr = llvm::LLVMRustBuildVectorReduceFAdd(self.llbuilder, acc, src);
+            if instr.is_null() {
+                bug!("LLVMRustBuildVectorReduceFAdd is not available in LLVM version < 5.0");
+            }
+            llvm::LLVMRustSetHasUnsafeAlgebra(instr);
+            instr
+        }
+    }
+    pub fn vector_reduce_fmul_fast(&self, acc: ValueRef, src: ValueRef) -> ValueRef {
+        self.count_insn("vector.reduce.fmul_fast");
+        unsafe {
+            // FIXME: add a non-fast math version once
+            // https://bugs.llvm.org/show_bug.cgi?id=36732
+            // is fixed.
+            let instr = llvm::LLVMRustBuildVectorReduceFMul(self.llbuilder, acc, src);
+            if instr.is_null() {
+                bug!("LLVMRustBuildVectorReduceFMul is not available in LLVM version < 5.0");
+            }
+            llvm::LLVMRustSetHasUnsafeAlgebra(instr);
+            instr
+        }
+    }
+    pub fn vector_reduce_add(&self, src: ValueRef) -> ValueRef {
+        self.count_insn("vector.reduce.add");
+        unsafe {
+            let instr = llvm::LLVMRustBuildVectorReduceAdd(self.llbuilder, src);
+            if instr.is_null() {
+                bug!("LLVMRustBuildVectorReduceAdd is not available in LLVM version < 5.0");
+            }
+            instr
+        }
+    }
+    pub fn vector_reduce_mul(&self, src: ValueRef) -> ValueRef {
+        self.count_insn("vector.reduce.mul");
+        unsafe {
+            let instr = llvm::LLVMRustBuildVectorReduceMul(self.llbuilder, src);
+            if instr.is_null() {
+                bug!("LLVMRustBuildVectorReduceMul is not available in LLVM version < 5.0");
+            }
+            instr
+        }
+    }
+    pub fn vector_reduce_and(&self, src: ValueRef) -> ValueRef {
+        self.count_insn("vector.reduce.and");
+        unsafe {
+            let instr = llvm::LLVMRustBuildVectorReduceAnd(self.llbuilder, src);
+            if instr.is_null() {
+                bug!("LLVMRustBuildVectorReduceAnd is not available in LLVM version < 5.0");
+            }
+            instr
+        }
+    }
+    pub fn vector_reduce_or(&self, src: ValueRef) -> ValueRef {
+        self.count_insn("vector.reduce.or");
+        unsafe {
+            let instr = llvm::LLVMRustBuildVectorReduceOr(self.llbuilder, src);
+            if instr.is_null() {
+                bug!("LLVMRustBuildVectorReduceOr is not available in LLVM version < 5.0");
+            }
+            instr
+        }
+    }
+    pub fn vector_reduce_xor(&self, src: ValueRef) -> ValueRef {
+        self.count_insn("vector.reduce.xor");
+        unsafe {
+            let instr = llvm::LLVMRustBuildVectorReduceXor(self.llbuilder, src);
+            if instr.is_null() {
+                bug!("LLVMRustBuildVectorReduceXor is not available in LLVM version < 5.0");
+            }
+            instr
+        }
+    }
+    pub fn vector_reduce_fmin(&self, src: ValueRef) -> ValueRef {
+        self.count_insn("vector.reduce.fmin");
+        unsafe {
+            let instr = llvm::LLVMRustBuildVectorReduceFMin(self.llbuilder, src, true);
+            if instr.is_null() {
+                bug!("LLVMRustBuildVectorReduceFMin is not available in LLVM version < 5.0");
+            }
+            instr
+        }
+    }
+    pub fn vector_reduce_fmax(&self, src: ValueRef) -> ValueRef {
+        self.count_insn("vector.reduce.fmax");
+        unsafe {
+            let instr = llvm::LLVMRustBuildVectorReduceFMax(self.llbuilder, src, true);
+            if instr.is_null() {
+                bug!("LLVMRustBuildVectorReduceFMax is not available in LLVM version < 5.0");
+            }
+            instr
+        }
+    }
+    pub fn vector_reduce_fmin_fast(&self, src: ValueRef) -> ValueRef {
+        self.count_insn("vector.reduce.fmin_fast");
+        unsafe {
+            let instr = llvm::LLVMRustBuildVectorReduceFMin(self.llbuilder, src, false);
+            if instr.is_null() {
+                bug!("LLVMRustBuildVectorReduceFMin is not available in LLVM version < 5.0");
+            }
+            llvm::LLVMRustSetHasUnsafeAlgebra(instr);
+            instr
+        }
+    }
+    pub fn vector_reduce_fmax_fast(&self, src: ValueRef) -> ValueRef {
+        self.count_insn("vector.reduce.fmax_fast");
+        unsafe {
+            let instr = llvm::LLVMRustBuildVectorReduceFMax(self.llbuilder, src, false);
+            if instr.is_null() {
+                bug!("LLVMRustBuildVectorReduceFMax is not available in LLVM version < 5.0");
+            }
+            llvm::LLVMRustSetHasUnsafeAlgebra(instr);
+            instr
+        }
+    }
+    pub fn vector_reduce_min(&self, src: ValueRef, is_signed: bool) -> ValueRef {
+        self.count_insn("vector.reduce.min");
+        unsafe {
+            let instr = llvm::LLVMRustBuildVectorReduceMin(self.llbuilder, src, is_signed);
+            if instr.is_null() {
+                bug!("LLVMRustBuildVectorReduceMin is not available in LLVM version < 5.0");
+            }
+            instr
+        }
+    }
+    pub fn vector_reduce_max(&self, src: ValueRef, is_signed: bool) -> ValueRef {
+        self.count_insn("vector.reduce.max");
+        unsafe {
+            let instr = llvm::LLVMRustBuildVectorReduceMax(self.llbuilder, src, is_signed);
+            if instr.is_null() {
+                bug!("LLVMRustBuildVectorReduceMax is not available in LLVM version < 5.0");
+            }
+            instr
+        }
+    }
+
     pub fn extract_value(&self, agg_val: ValueRef, idx: u64) -> ValueRef {
         self.count_insn("extractvalue");
         assert_eq!(idx as c_uint as u64, idx);
diff --git a/src/librustc_trans/common.rs b/src/librustc_trans/common.rs
index c71c0cc0ebf..e83e73c8ae7 100644
--- a/src/librustc_trans/common.rs
+++ b/src/librustc_trans/common.rs
@@ -269,6 +269,19 @@ pub fn const_get_elt(v: ValueRef, idx: u64) -> ValueRef {
     }
 }
 
+pub fn const_get_real(v: ValueRef) -> Option<(f64, bool)> {
+    unsafe {
+        if is_const_real(v) {
+            let mut loses_info: llvm::Bool = ::std::mem::uninitialized();
+            let r = llvm::LLVMConstRealGetDouble(v, &mut loses_info as *mut llvm::Bool);
+            let loses_info = if loses_info == 1 { true } else { false };
+            Some((r, loses_info))
+        } else {
+            None
+        }
+    }
+}
+
 pub fn const_to_uint(v: ValueRef) -> u64 {
     unsafe {
         llvm::LLVMConstIntGetZExtValue(v)
@@ -281,6 +294,13 @@ pub fn is_const_integral(v: ValueRef) -> bool {
     }
 }
 
+pub fn is_const_real(v: ValueRef) -> bool {
+    unsafe {
+        !llvm::LLVMIsAConstantFP(v).is_null()
+    }
+}
+
+
 #[inline]
 fn hi_lo_to_u128(lo: u64, hi: u64) -> u128 {
     ((hi as u128) << 64) | (lo as u128)
diff --git a/src/librustc_trans/intrinsic.rs b/src/librustc_trans/intrinsic.rs
index 4cef7470c62..c3de9e0ffcc 100644
--- a/src/librustc_trans/intrinsic.rs
+++ b/src/librustc_trans/intrinsic.rs
@@ -1018,14 +1018,22 @@ fn generic_simd_intrinsic<'a, 'tcx>(
                          name, $($fmt)*));
         }
     }
-    macro_rules! require {
-        ($cond: expr, $($fmt: tt)*) => {
-            if !$cond {
+    macro_rules! return_error {
+        ($($fmt: tt)*) => {
+            {
                 emit_error!($($fmt)*);
                 return Err(());
             }
         }
     }
+
+    macro_rules! require {
+        ($cond: expr, $($fmt: tt)*) => {
+            if !$cond {
+                return_error!($($fmt)*);
+            }
+        };
+    }
     macro_rules! require_simd {
         ($ty: expr, $position: expr) => {
             require!($ty.is_simd(), "expected SIMD {} type, found non-SIMD `{}`", $position, $ty)
@@ -1145,6 +1153,161 @@ fn generic_simd_intrinsic<'a, 'tcx>(
         return Ok(bx.extract_element(args[0].immediate(), args[1].immediate()))
     }
 
+    macro_rules! arith_red {
+        ($name:tt : $integer_reduce:ident, $float_reduce:ident, $ordered:expr) => {
+            if name == $name {
+                require!(ret_ty == in_elem,
+                         "expected return type `{}` (element of input `{}`), found `{}`",
+                         in_elem, in_ty, ret_ty);
+                return match in_elem.sty {
+                    ty::TyInt(_) | ty::TyUint(_) => {
+                        let r = bx.$integer_reduce(args[0].immediate());
+                        if $ordered {
+                            // if overflow occurs, the result is the
+                            // mathematical result modulo 2^n:
+                            if name.contains("mul") {
+                                Ok(bx.mul(args[1].immediate(), r))
+                            } else {
+                                Ok(bx.add(args[1].immediate(), r))
+                            }
+                        } else {
+                            Ok(bx.$integer_reduce(args[0].immediate()))
+                        }
+                    },
+                    ty::TyFloat(f) => {
+                        // ordered arithmetic reductions take an accumulator
+                        let acc = if $ordered {
+                            let acc = args[1].immediate();
+                            // FIXME: https://bugs.llvm.org/show_bug.cgi?id=36734
+                            // * if the accumulator of the fadd isn't 0, incorrect
+                            //   code is generated
+                            // * if the accumulator of the fmul isn't 1, incorrect
+                            //   code is generated
+                            match const_get_real(acc) {
+                                None => return_error!("accumulator of {} is not a constant", $name),
+                                Some((v, loses_info)) => {
+                                    if $name.contains("mul") && v != 1.0_f64 {
+                                        return_error!("accumulator of {} is not 1.0", $name);
+                                    } else if $name.contains("add") && v != 0.0_f64 {
+                                        return_error!("accumulator of {} is not 0.0", $name);
+                                    } else if loses_info {
+                                        return_error!("accumulator of {} loses information", $name);
+                                    }
+                                }
+                            }
+                            acc
+                        } else {
+                            // unordered arithmetic reductions do not:
+                            match f.bit_width() {
+                                32 => C_undef(Type::f32(bx.cx)),
+                                64 => C_undef(Type::f64(bx.cx)),
+                                v => {
+                                    return_error!(r#"
+unsupported {} from `{}` with element `{}` of size `{}` to `{}`"#,
+                                        $name, in_ty, in_elem, v, ret_ty
+                                    )
+                                }
+                            }
+
+                        };
+                        Ok(bx.$float_reduce(acc, args[0].immediate()))
+                    }
+                    _ => {
+                        return_error!(
+                            "unsupported {} from `{}` with element `{}` to `{}`",
+                            $name, in_ty, in_elem, ret_ty
+                        )
+                    },
+                }
+            }
+        }
+    }
+
+    arith_red!("simd_reduce_add_ordered": vector_reduce_add, vector_reduce_fadd_fast, true);
+    arith_red!("simd_reduce_mul_ordered": vector_reduce_mul, vector_reduce_fmul_fast, true);
+    arith_red!("simd_reduce_add_unordered": vector_reduce_add, vector_reduce_fadd_fast, false);
+    arith_red!("simd_reduce_mul_unordered": vector_reduce_mul, vector_reduce_fmul_fast, false);
+
+    macro_rules! minmax_red {
+        ($name:tt: $int_red:ident, $float_red:ident) => {
+            if name == $name {
+                require!(ret_ty == in_elem,
+                         "expected return type `{}` (element of input `{}`), found `{}`",
+                         in_elem, in_ty, ret_ty);
+                return match in_elem.sty {
+                    ty::TyInt(_i) => {
+                        Ok(bx.$int_red(args[0].immediate(), true))
+                    },
+                    ty::TyUint(_u) => {
+                        Ok(bx.$int_red(args[0].immediate(), false))
+                    },
+                    ty::TyFloat(_f) => {
+                        Ok(bx.$float_red(args[0].immediate()))
+                    }
+                    _ => {
+                        return_error!("unsupported {} from `{}` with element `{}` to `{}`",
+                                      $name, in_ty, in_elem, ret_ty)
+                    },
+                }
+            }
+
+        }
+    }
+
+    minmax_red!("simd_reduce_min": vector_reduce_min, vector_reduce_fmin);
+    minmax_red!("simd_reduce_max": vector_reduce_max, vector_reduce_fmax);
+
+    minmax_red!("simd_reduce_min_nanless": vector_reduce_min, vector_reduce_fmin_fast);
+    minmax_red!("simd_reduce_max_nanless": vector_reduce_max, vector_reduce_fmax_fast);
+
+    macro_rules! bitwise_red {
+        ($name:tt : $red:ident, $boolean:expr) => {
+            if name == $name {
+                let input = if !$boolean {
+                    require!(ret_ty == in_elem,
+                             "expected return type `{}` (element of input `{}`), found `{}`",
+                             in_elem, in_ty, ret_ty);
+                    args[0].immediate()
+                } else {
+                    match in_elem.sty {
+                        ty::TyInt(_) | ty::TyUint(_) => {},
+                        _ => {
+                            return_error!("unsupported {} from `{}` with element `{}` to `{}`",
+                                          $name, in_ty, in_elem, ret_ty)
+                        }
+                    }
+
+                    // boolean reductions operate on vectors of i1s:
+                    let i1 = Type::i1(bx.cx);
+                    let i1xn = Type::vector(&i1, in_len as u64);
+                    bx.trunc(args[0].immediate(), i1xn)
+                };
+                return match in_elem.sty {
+                    ty::TyInt(_) | ty::TyUint(_) => {
+                        let r = bx.$red(input);
+                        Ok(
+                            if !$boolean {
+                                r
+                            } else {
+                                bx.zext(r, Type::bool(bx.cx))
+                            }
+                        )
+                    },
+                    _ => {
+                        return_error!("unsupported {} from `{}` with element `{}` to `{}`",
+                                      $name, in_ty, in_elem, ret_ty)
+                    },
+                }
+            }
+        }
+    }
+
+    bitwise_red!("simd_reduce_and": vector_reduce_and, false);
+    bitwise_red!("simd_reduce_or": vector_reduce_or, false);
+    bitwise_red!("simd_reduce_xor": vector_reduce_xor, false);
+    bitwise_red!("simd_reduce_all": vector_reduce_and, true);
+    bitwise_red!("simd_reduce_any": vector_reduce_or, true);
+
     if name == "simd_cast" {
         require_simd!(ret_ty, "return");
         let out_len = ret_ty.simd_size(tcx);
diff --git a/src/librustc_typeck/check/intrinsic.rs b/src/librustc_typeck/check/intrinsic.rs
index 59cf8a0358e..99707a4a3c0 100644
--- a/src/librustc_typeck/check/intrinsic.rs
+++ b/src/librustc_typeck/check/intrinsic.rs
@@ -361,6 +361,14 @@ pub fn check_platform_intrinsic_type<'a, 'tcx>(tcx: TyCtxt<'a, 'tcx, 'tcx>,
         "simd_insert" => (2, vec![param(0), tcx.types.u32, param(1)], param(0)),
         "simd_extract" => (2, vec![param(0), tcx.types.u32], param(1)),
         "simd_cast" => (2, vec![param(0)], param(1)),
+        "simd_reduce_all" | "simd_reduce_any" => (1, vec![param(0)], tcx.types.bool),
+        "simd_reduce_add_ordered" | "simd_reduce_mul_ordered"
+            => (2, vec![param(0), param(1)], param(1)),
+        "simd_reduce_add_unordered" | "simd_reduce_mul_unordered" |
+        "simd_reduce_and" | "simd_reduce_or"  | "simd_reduce_xor" |
+        "simd_reduce_min" | "simd_reduce_max" |
+        "simd_reduce_min_nanless" | "simd_reduce_max_nanless"
+            => (2, vec![param(0)], param(1)),
         name if name.starts_with("simd_shuffle") => {
             match name["simd_shuffle".len()..].parse() {
                 Ok(n) => {
diff --git a/src/rustllvm/RustWrapper.cpp b/src/rustllvm/RustWrapper.cpp
index 27d5496f576..a5644d6f9e2 100644
--- a/src/rustllvm/RustWrapper.cpp
+++ b/src/rustllvm/RustWrapper.cpp
@@ -21,6 +21,8 @@
 
 #if LLVM_VERSION_GE(5, 0)
 #include "llvm/ADT/Optional.h"
+#else
+#include <cstdlib>
 #endif
 
 //===----------------------------------------------------------------------===
@@ -1395,3 +1397,98 @@ LLVMRustModuleCost(LLVMModuleRef M) {
   auto f = unwrap(M)->functions();
   return std::distance(std::begin(f), std::end(f));
 }
+
+// Vector reductions:
+#if LLVM_VERSION_GE(5, 0)
+extern "C" LLVMValueRef
+LLVMRustBuildVectorReduceFAdd(LLVMBuilderRef B, LLVMValueRef Acc, LLVMValueRef Src) {
+    return wrap(unwrap(B)->CreateFAddReduce(unwrap(Acc),unwrap(Src)));
+}
+extern "C" LLVMValueRef
+LLVMRustBuildVectorReduceFMul(LLVMBuilderRef B, LLVMValueRef Acc, LLVMValueRef Src) {
+    return wrap(unwrap(B)->CreateFMulReduce(unwrap(Acc),unwrap(Src)));
+}
+extern "C" LLVMValueRef
+LLVMRustBuildVectorReduceAdd(LLVMBuilderRef B, LLVMValueRef Src) {
+    return wrap(unwrap(B)->CreateAddReduce(unwrap(Src)));
+}
+extern "C" LLVMValueRef
+LLVMRustBuildVectorReduceMul(LLVMBuilderRef B, LLVMValueRef Src) {
+    return wrap(unwrap(B)->CreateMulReduce(unwrap(Src)));
+}
+extern "C" LLVMValueRef
+LLVMRustBuildVectorReduceAnd(LLVMBuilderRef B, LLVMValueRef Src) {
+    return wrap(unwrap(B)->CreateAndReduce(unwrap(Src)));
+}
+extern "C" LLVMValueRef
+LLVMRustBuildVectorReduceOr(LLVMBuilderRef B, LLVMValueRef Src) {
+    return wrap(unwrap(B)->CreateOrReduce(unwrap(Src)));
+}
+extern "C" LLVMValueRef
+LLVMRustBuildVectorReduceXor(LLVMBuilderRef B, LLVMValueRef Src) {
+    return wrap(unwrap(B)->CreateXorReduce(unwrap(Src)));
+}
+extern "C" LLVMValueRef
+LLVMRustBuildVectorReduceMin(LLVMBuilderRef B, LLVMValueRef Src, bool IsSigned) {
+    return wrap(unwrap(B)->CreateIntMinReduce(unwrap(Src), IsSigned));
+}
+extern "C" LLVMValueRef
+LLVMRustBuildVectorReduceMax(LLVMBuilderRef B, LLVMValueRef Src, bool IsSigned) {
+    return wrap(unwrap(B)->CreateIntMaxReduce(unwrap(Src), IsSigned));
+}
+extern "C" LLVMValueRef
+LLVMRustBuildVectorReduceFMin(LLVMBuilderRef B, LLVMValueRef Src, bool NoNaN) {
+   return wrap(unwrap(B)->CreateFPMinReduce(unwrap(Src), NoNaN));
+}
+extern "C" LLVMValueRef
+LLVMRustBuildVectorReduceFMax(LLVMBuilderRef B, LLVMValueRef Src, bool NoNaN) {
+  return wrap(unwrap(B)->CreateFPMaxReduce(unwrap(Src), NoNaN));
+}
+
+#else
+
+extern "C" LLVMValueRef
+LLVMRustBuildVectorReduceFAdd(LLVMBuilderRef, LLVMValueRef, LLVMValueRef) {
+  return nullptr;
+}
+extern "C" LLVMValueRef
+LLVMRustBuildVectorReduceFMul(LLVMBuilderRef, LLVMValueRef, LLVMValueRef) {
+  return nullptr;
+}
+extern "C" LLVMValueRef
+LLVMRustBuildVectorReduceAdd(LLVMBuilderRef, LLVMValueRef) {
+  return nullptr;
+}
+extern "C" LLVMValueRef
+LLVMRustBuildVectorReduceMul(LLVMBuilderRef, LLVMValueRef) {
+  return nullptr;
+}
+extern "C" LLVMValueRef
+LLVMRustBuildVectorReduceAnd(LLVMBuilderRef, LLVMValueRef) {
+  return nullptr;
+}
+extern "C" LLVMValueRef
+LLVMRustBuildVectorReduceOr(LLVMBuilderRef, LLVMValueRef) {
+  return nullptr;
+}
+extern "C" LLVMValueRef
+LLVMRustBuildVectorReduceXor(LLVMBuilderRef, LLVMValueRef) {
+  return nullptr;
+}
+extern "C" LLVMValueRef
+LLVMRustBuildVectorReduceMin(LLVMBuilderRef, LLVMValueRef, bool) {
+  return nullptr;
+}
+extern "C" LLVMValueRef
+LLVMRustBuildVectorReduceMax(LLVMBuilderRef, LLVMValueRef, bool) {
+  return nullptr;
+}
+extern "C" LLVMValueRef
+LLVMRustBuildVectorReduceFMin(LLVMBuilderRef, LLVMValueRef, bool) {
+  return nullptr;
+}
+extern "C" LLVMValueRef
+LLVMRustBuildVectorReduceFMax(LLVMBuilderRef, LLVMValueRef, bool) {
+  return nullptr;
+}
+#endif
diff --git a/src/test/compile-fail/simd-intrinsic-generic-reduction.rs b/src/test/compile-fail/simd-intrinsic-generic-reduction.rs
new file mode 100644
index 00000000000..57e4bb76a6c
--- /dev/null
+++ b/src/test/compile-fail/simd-intrinsic-generic-reduction.rs
@@ -0,0 +1,82 @@
+// Copyright 2015 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.
+
+// min-llvm-version 5.0
+// ignore-emscripten
+
+// Test that the simd_reduce_{op} intrinsics produce ok-ish error
+// messages when misused.
+
+#![feature(repr_simd, platform_intrinsics)]
+#![allow(non_camel_case_types)]
+
+#[repr(simd)]
+#[derive(Copy, Clone)]
+pub struct f32x4(pub f32, pub f32, pub f32, pub f32);
+
+#[repr(simd)]
+#[derive(Copy, Clone)]
+pub struct u32x4(pub u32, pub u32, pub u32, pub u32);
+
+
+extern "platform-intrinsic" {
+    fn simd_reduce_add_ordered<T, U>(x: T, y: U) -> U;
+    fn simd_reduce_mul_ordered<T, U>(x: T, y: U) -> U;
+    fn simd_reduce_and<T, U>(x: T) -> U;
+    fn simd_reduce_or<T, U>(x: T) -> U;
+    fn simd_reduce_xor<T, U>(x: T) -> U;
+    fn simd_reduce_all<T>(x: T) -> bool;
+    fn simd_reduce_any<T>(x: T) -> bool;
+}
+
+fn main() {
+    let x = u32x4(0, 0, 0, 0);
+    let z = f32x4(0.0, 0.0, 0.0, 0.0);
+
+    unsafe {
+        simd_reduce_add_ordered(z, 0_f32);
+        simd_reduce_mul_ordered(z, 1_f32);
+
+        simd_reduce_add_ordered(z, 2_f32);
+        //~^ ERROR accumulator of simd_reduce_add_ordered is not 0.0
+        simd_reduce_mul_ordered(z, 3_f32);
+        //~^ ERROR accumulator of simd_reduce_mul_ordered is not 1.0
+
+        let _: f32 = simd_reduce_and(x);
+        //~^ ERROR expected return type `u32` (element of input `u32x4`), found `f32`
+        let _: f32 = simd_reduce_or(x);
+        //~^ ERROR expected return type `u32` (element of input `u32x4`), found `f32`
+        let _: f32 = simd_reduce_xor(x);
+        //~^ ERROR expected return type `u32` (element of input `u32x4`), found `f32`
+
+        let _: f32 = simd_reduce_and(z);
+        //~^ ERROR unsupported simd_reduce_and from `f32x4` with element `f32` to `f32`
+        let _: f32 = simd_reduce_or(z);
+        //~^ ERROR unsupported simd_reduce_or from `f32x4` with element `f32` to `f32`
+        let _: f32 = simd_reduce_xor(z);
+        //~^ ERROR unsupported simd_reduce_xor from `f32x4` with element `f32` to `f32`
+
+        let _: bool = simd_reduce_all(z);
+        //~^ ERROR unsupported simd_reduce_all from `f32x4` with element `f32` to `bool`
+        let _: bool = simd_reduce_any(z);
+        //~^ ERROR unsupported simd_reduce_any from `f32x4` with element `f32` to `bool`
+
+        foo(0_f32);
+    }
+}
+
+#[inline(never)]
+unsafe fn foo(x: f32) {
+    let z = f32x4(0.0, 0.0, 0.0, 0.0);
+    simd_reduce_add_ordered(z, x);
+    //~^ ERROR accumulator of simd_reduce_add_ordered is not a constant
+    simd_reduce_mul_ordered(z, x);
+    //~^ ERROR accumulator of simd_reduce_mul_ordered is not a constant
+}
diff --git a/src/test/run-pass/simd-intrinsic-generic-reduction.rs b/src/test/run-pass/simd-intrinsic-generic-reduction.rs
new file mode 100644
index 00000000000..9a1214d3b35
--- /dev/null
+++ b/src/test/run-pass/simd-intrinsic-generic-reduction.rs
@@ -0,0 +1,172 @@
+// Copyright 2015 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.
+
+// min-llvm-version 5.0
+// ignore-emscripten
+
+// Test that the simd_reduce_{op} intrinsics produce the correct results.
+
+#![feature(repr_simd, platform_intrinsics)]
+#[allow(non_camel_case_types)]
+
+#[repr(simd)]
+#[derive(Copy, Clone)]
+struct i32x4(pub i32, pub i32, pub i32, pub i32);
+
+#[repr(simd)]
+#[derive(Copy, Clone)]
+struct u32x4(pub u32, pub u32, pub u32, pub u32);
+
+#[repr(simd)]
+#[derive(Copy, Clone)]
+struct f32x4(pub f32, pub f32, pub f32, pub f32);
+
+#[repr(simd)]
+#[derive(Copy, Clone)]
+struct b8x4(pub i8, pub i8, pub i8, pub i8);
+
+#[repr(simd)]
+#[derive(Copy, Clone)]
+struct b8x16(
+    pub i8, pub i8, pub i8, pub i8,
+    pub i8, pub i8, pub i8, pub i8,
+    pub i8, pub i8, pub i8, pub i8,
+    pub i8, pub i8, pub i8, pub i8
+);
+
+extern "platform-intrinsic" {
+    fn simd_reduce_add_unordered<T, U>(x: T) -> U;
+    fn simd_reduce_mul_unordered<T, U>(x: T) -> U;
+    fn simd_reduce_add_ordered<T, U>(x: T, acc: U) -> U;
+    fn simd_reduce_mul_ordered<T, U>(x: T, acc: U) -> U;
+    fn simd_reduce_min<T, U>(x: T) -> U;
+    fn simd_reduce_max<T, U>(x: T) -> U;
+    fn simd_reduce_min_nanless<T, U>(x: T) -> U;
+    fn simd_reduce_max_nanless<T, U>(x: T) -> U;
+    fn simd_reduce_and<T, U>(x: T) -> U;
+    fn simd_reduce_or<T, U>(x: T) -> U;
+    fn simd_reduce_xor<T, U>(x: T) -> U;
+    fn simd_reduce_all<T>(x: T) -> bool;
+    fn simd_reduce_any<T>(x: T) -> bool;
+}
+
+fn main() {
+    unsafe {
+        let x = i32x4(1, -2, 3, 4);
+        let r: i32 = simd_reduce_add_unordered(x);
+        assert_eq!(r, 6_i32);
+        let r: i32 = simd_reduce_mul_unordered(x);
+        assert_eq!(r, -24_i32);
+        let r: i32 = simd_reduce_add_ordered(x, -1);
+        assert_eq!(r, 5_i32);
+        let r: i32 = simd_reduce_mul_ordered(x, -1);
+        assert_eq!(r, 24_i32);
+
+        let r: i32 = simd_reduce_min(x);
+        assert_eq!(r, -2_i32);
+        let r: i32 = simd_reduce_max(x);
+        assert_eq!(r, 4_i32);
+
+        let x = i32x4(-1, -1, -1, -1);
+        let r: i32 = simd_reduce_and(x);
+        assert_eq!(r, -1_i32);
+        let r: i32 = simd_reduce_or(x);
+        assert_eq!(r, -1_i32);
+        let r: i32 = simd_reduce_xor(x);
+        assert_eq!(r, 0_i32);
+
+        let x = i32x4(-1, -1, 0, -1);
+        let r: i32 = simd_reduce_and(x);
+        assert_eq!(r, 0_i32);
+        let r: i32 = simd_reduce_or(x);
+        assert_eq!(r, -1_i32);
+        let r: i32 = simd_reduce_xor(x);
+        assert_eq!(r, -1_i32);
+    }
+
+    unsafe {
+        let x = u32x4(1, 2, 3, 4);
+        let r: u32 = simd_reduce_add_unordered(x);
+        assert_eq!(r, 10_u32);
+        let r: u32 = simd_reduce_mul_unordered(x);
+        assert_eq!(r, 24_u32);
+        let r: u32 = simd_reduce_add_ordered(x, 1);
+        assert_eq!(r, 11_u32);
+        let r: u32 = simd_reduce_mul_ordered(x, 2);
+        assert_eq!(r, 48_u32);
+
+        let r: u32 = simd_reduce_min(x);
+        assert_eq!(r, 1_u32);
+        let r: u32 = simd_reduce_max(x);
+        assert_eq!(r, 4_u32);
+
+        let t = u32::max_value();
+        let x = u32x4(t, t, t, t);
+        let r: u32 = simd_reduce_and(x);
+        assert_eq!(r, t);
+        let r: u32 = simd_reduce_or(x);
+        assert_eq!(r, t);
+        let r: u32 = simd_reduce_xor(x);
+        assert_eq!(r, 0_u32);
+
+        let x = u32x4(t, t, 0, t);
+        let r: u32 = simd_reduce_and(x);
+        assert_eq!(r, 0_u32);
+        let r: u32 = simd_reduce_or(x);
+        assert_eq!(r, t);
+        let r: u32 = simd_reduce_xor(x);
+        assert_eq!(r, t);
+    }
+
+    unsafe {
+        let x = f32x4(1., -2., 3., 4.);
+        let r: f32 = simd_reduce_add_unordered(x);
+        assert_eq!(r, 6_f32);
+        let r: f32 = simd_reduce_mul_unordered(x);
+        assert_eq!(r, -24_f32);
+        // FIXME: only works correctly for accumulator, 0:
+        // https://bugs.llvm.org/show_bug.cgi?id=36734
+        let r: f32 = simd_reduce_add_ordered(x, 0.);
+        assert_eq!(r, 6_f32);
+        // FIXME: only works correctly for accumulator, 1:
+        // https://bugs.llvm.org/show_bug.cgi?id=36734
+        let r: f32 = simd_reduce_mul_ordered(x, 1.);
+        assert_eq!(r, -24_f32);
+
+        let r: f32 = simd_reduce_min(x);
+        assert_eq!(r, -2_f32);
+        let r: f32 = simd_reduce_max(x);
+        assert_eq!(r, 4_f32);
+        let r: f32 = simd_reduce_min_nanless(x);
+        assert_eq!(r, -2_f32);
+        let r: f32 = simd_reduce_max_nanless(x);
+        assert_eq!(r, 4_f32);
+    }
+
+    unsafe {
+        let x = b8x4(!0, !0, !0, !0);
+        let r: bool = simd_reduce_all(x);
+        assert_eq!(r, true);
+        let r: bool = simd_reduce_any(x);
+        assert_eq!(r, true);
+
+        let x = b8x4(!0, !0, 0, !0);
+        let r: bool = simd_reduce_all(x);
+        assert_eq!(r, false);
+        let r: bool = simd_reduce_any(x);
+        assert_eq!(r, true);
+
+        let x = b8x4(0, 0, 0, 0);
+        let r: bool = simd_reduce_all(x);
+        assert_eq!(r, false);
+        let r: bool = simd_reduce_any(x);
+        assert_eq!(r, false);
+    }
+}