emit "align 1" metadata on loads/stores of packed structs

According to the LLVM reference:
> A value of 0 or an omitted align argument means that the operation has
the ABI alignment for the target.

So loads/stores of fields of packed structs need to have their align set
to 1. Implement that by tracking the alignment of `LvalueRef`s.

Fixes #39376.

15 files changed, 363 insertions, 223 deletions

diff --git a/src/librustc_trans/adt.rs b/src/librustc_trans/adt.rs
index bc1e07e708c..59f2104ec14 100644
--- a/src/librustc_trans/adt.rs
+++ b/src/librustc_trans/adt.rs
@@ -56,6 +56,8 @@ use monomorphize;
 use type_::Type;
 use type_of;
 
+use mir::lvalue::Alignment;
+
 /// Given an enum, struct, closure, or tuple, extracts fields.
 /// Treats closures as a struct with one variant.
 /// `empty_if_no_variants` is a switch to deal with empty enums.
@@ -279,6 +281,7 @@ pub fn trans_get_discr<'a, 'tcx>(
     bcx: &Builder<'a, 'tcx>,
     t: Ty<'tcx>,
     scrutinee: ValueRef,
+    alignment: Alignment,
     cast_to: Option<Type>,
     range_assert: bool
 ) -> ValueRef {
@@ -292,11 +295,12 @@ pub fn trans_get_discr<'a, 'tcx>(
 
     let val = match *l {
         layout::CEnum { discr, min, max, .. } => {
-            load_discr(bcx, discr, scrutinee, min, max, range_assert)
+            load_discr(bcx, discr, scrutinee, alignment, min, max, range_assert)
         }
         layout::General { discr, .. } => {
             let ptr = bcx.struct_gep(scrutinee, 0);
-            load_discr(bcx, discr, ptr, 0, def.variants.len() as u64 - 1,
+            load_discr(bcx, discr, ptr, alignment,
+                       0, def.variants.len() as u64 - 1,
                        range_assert)
         }
         layout::Univariant { .. } | layout::UntaggedUnion { .. } => C_u8(bcx.ccx, 0),
@@ -305,10 +309,10 @@ pub fn trans_get_discr<'a, 'tcx>(
             let llptrty = type_of::sizing_type_of(bcx.ccx,
                 monomorphize::field_ty(bcx.tcx(), substs,
                 &def.variants[nndiscr as usize].fields[0]));
-            bcx.icmp(cmp, bcx.load(scrutinee), C_null(llptrty))
+            bcx.icmp(cmp, bcx.load(scrutinee, alignment.to_align()), C_null(llptrty))
         }
         layout::StructWrappedNullablePointer { nndiscr, ref discrfield, .. } => {
-            struct_wrapped_nullable_bitdiscr(bcx, nndiscr, discrfield, scrutinee)
+            struct_wrapped_nullable_bitdiscr(bcx, nndiscr, discrfield, scrutinee, alignment)
         },
         _ => bug!("{} is not an enum", t)
     };
@@ -322,17 +326,19 @@ fn struct_wrapped_nullable_bitdiscr(
     bcx: &Builder,
     nndiscr: u64,
     discrfield: &layout::FieldPath,
-    scrutinee: ValueRef
+    scrutinee: ValueRef,
+    alignment: Alignment,
 ) -> ValueRef {
     let llptrptr = bcx.gepi(scrutinee,
         &discrfield.iter().map(|f| *f as usize).collect::<Vec<_>>()[..]);
-    let llptr = bcx.load(llptrptr);
+    let llptr = bcx.load(llptrptr, alignment.to_align());
     let cmp = if nndiscr == 0 { IntEQ } else { IntNE };
     bcx.icmp(cmp, llptr, C_null(val_ty(llptr)))
 }
 
 /// Helper for cases where the discriminant is simply loaded.
-fn load_discr(bcx: &Builder, ity: layout::Integer, ptr: ValueRef, min: u64, max: u64,
+fn load_discr(bcx: &Builder, ity: layout::Integer, ptr: ValueRef,
+              alignment: Alignment, min: u64, max: u64,
               range_assert: bool)
     -> ValueRef {
     let llty = Type::from_integer(bcx.ccx, ity);
@@ -348,11 +354,12 @@ fn load_discr(bcx: &Builder, ity: layout::Integer, ptr: ValueRef, min: u64, max:
         // rejected by the LLVM verifier (it would mean either an
         // empty set, which is impossible, or the entire range of the
         // type, which is pointless).
-        bcx.load(ptr)
+        bcx.load(ptr, alignment.to_align())
     } else {
         // llvm::ConstantRange can deal with ranges that wrap around,
         // so an overflow on (max + 1) is fine.
-        bcx.load_range_assert(ptr, min, max.wrapping_add(1), /* signed: */ True)
+        bcx.load_range_assert(ptr, min, max.wrapping_add(1), /* signed: */ True,
+                              alignment.to_align())
     }
 }
 
diff --git a/src/librustc_trans/asm.rs b/src/librustc_trans/asm.rs
index c95d4147018..12e4e57964f 100644
--- a/src/librustc_trans/asm.rs
+++ b/src/librustc_trans/asm.rs
@@ -20,6 +20,8 @@ use builder::Builder;
 use rustc::hir;
 use rustc::ty::Ty;
 
+use mir::lvalue::Alignment;
+
 use std::ffi::CString;
 use syntax::ast::AsmDialect;
 use libc::{c_uint, c_char};
@@ -38,7 +40,7 @@ pub fn trans_inline_asm<'a, 'tcx>(
     let mut indirect_outputs = vec![];
     for (i, (out, &(val, ty))) in ia.outputs.iter().zip(&outputs).enumerate() {
         let val = if out.is_rw || out.is_indirect {
-            Some(base::load_ty(bcx, val, ty))
+            Some(base::load_ty(bcx, val, Alignment::Packed, ty))
         } else {
             None
         };
diff --git a/src/librustc_trans/base.rs b/src/librustc_trans/base.rs
index 9bd19d5bbb3..5312cc60b09 100644
--- a/src/librustc_trans/base.rs
+++ b/src/librustc_trans/base.rs
@@ -90,6 +90,8 @@ use rustc::hir;
 use rustc::ty::layout::{self, Layout};
 use syntax::ast;
 
+use mir::lvalue::Alignment;
+
 pub struct StatRecorder<'a, 'tcx: 'a> {
     ccx: &'a CrateContext<'a, 'tcx>,
     name: Option<String>,
@@ -250,25 +252,25 @@ pub fn unsize_thin_ptr<'a, 'tcx>(
 /// Coerce `src`, which is a reference to a value of type `src_ty`,
 /// to a value of type `dst_ty` and store the result in `dst`
 pub fn coerce_unsized_into<'a, 'tcx>(bcx: &Builder<'a, 'tcx>,
-                                     src: ValueRef,
-                                     src_ty: Ty<'tcx>,
-                                     dst: ValueRef,
-                                     dst_ty: Ty<'tcx>) {
+                                     src: &LvalueRef<'tcx>,
+                                     dst: &LvalueRef<'tcx>) {
+    let src_ty = src.ty.to_ty(bcx.tcx());
+    let dst_ty = dst.ty.to_ty(bcx.tcx());
     let coerce_ptr = || {
         let (base, info) = if common::type_is_fat_ptr(bcx.ccx, src_ty) {
             // fat-ptr to fat-ptr unsize preserves the vtable
             // i.e. &'a fmt::Debug+Send => &'a fmt::Debug
             // So we need to pointercast the base to ensure
             // the types match up.
-            let (base, info) = load_fat_ptr(bcx, src, src_ty);
+            let (base, info) = load_fat_ptr(bcx, src.llval, src.alignment, src_ty);
             let llcast_ty = type_of::fat_ptr_base_ty(bcx.ccx, dst_ty);
             let base = bcx.pointercast(base, llcast_ty);
             (base, info)
         } else {
-            let base = load_ty(bcx, src, src_ty);
+            let base = load_ty(bcx, src.llval, src.alignment, src_ty);
             unsize_thin_ptr(bcx, base, src_ty, dst_ty)
         };
-        store_fat_ptr(bcx, base, info, dst, dst_ty);
+        store_fat_ptr(bcx, base, info, dst.llval, dst.alignment, dst_ty);
     };
     match (&src_ty.sty, &dst_ty.sty) {
         (&ty::TyRef(..), &ty::TyRef(..)) |
@@ -290,21 +292,22 @@ pub fn coerce_unsized_into<'a, 'tcx>(bcx: &Builder<'a, 'tcx>,
                 monomorphize::field_ty(bcx.tcx(), substs_b, f)
             });
 
-            let src = LvalueRef::new_sized_ty(src, src_ty);
-            let dst = LvalueRef::new_sized_ty(dst, dst_ty);
-
             let iter = src_fields.zip(dst_fields).enumerate();
             for (i, (src_fty, dst_fty)) in iter {
                 if type_is_zero_size(bcx.ccx, dst_fty) {
                     continue;
                 }
 
-                let src_f = src.trans_field_ptr(bcx, i);
-                let dst_f = dst.trans_field_ptr(bcx, i);
+                let (src_f, src_f_align) = src.trans_field_ptr(bcx, i);
+                let (dst_f, dst_f_align) = dst.trans_field_ptr(bcx, i);
                 if src_fty == dst_fty {
                     memcpy_ty(bcx, dst_f, src_f, src_fty, None);
                 } else {
-                    coerce_unsized_into(bcx, src_f, src_fty, dst_f, dst_fty);
+                    coerce_unsized_into(
+                        bcx,
+                        &LvalueRef::new_sized_ty(src_f, src_fty, src_f_align),
+                        &LvalueRef::new_sized_ty(dst_f, dst_fty, dst_f_align)
+                    );
                 }
             }
         }
@@ -399,7 +402,8 @@ pub fn call_assume<'a, 'tcx>(b: &Builder<'a, 'tcx>, val: ValueRef) {
 /// Helper for loading values from memory. Does the necessary conversion if the in-memory type
 /// differs from the type used for SSA values. Also handles various special cases where the type
 /// gives us better information about what we are loading.
-pub fn load_ty<'a, 'tcx>(b: &Builder<'a, 'tcx>, ptr: ValueRef, t: Ty<'tcx>) -> ValueRef {
+pub fn load_ty<'a, 'tcx>(b: &Builder<'a, 'tcx>, ptr: ValueRef,
+                         alignment: Alignment, t: Ty<'tcx>) -> ValueRef {
     let ccx = b.ccx;
     if type_is_zero_size(ccx, t) {
         return C_undef(type_of::type_of(ccx, t));
@@ -419,29 +423,31 @@ pub fn load_ty<'a, 'tcx>(b: &Builder<'a, 'tcx>, ptr: ValueRef, t: Ty<'tcx>) -> V
     }
 
     if t.is_bool() {
-        b.trunc(b.load_range_assert(ptr, 0, 2, llvm::False), Type::i1(ccx))
+        b.trunc(b.load_range_assert(ptr, 0, 2, llvm::False, alignment.to_align()),
+                Type::i1(ccx))
     } else if t.is_char() {
         // a char is a Unicode codepoint, and so takes values from 0
         // to 0x10FFFF inclusive only.
-        b.load_range_assert(ptr, 0, 0x10FFFF + 1, llvm::False)
+        b.load_range_assert(ptr, 0, 0x10FFFF + 1, llvm::False, alignment.to_align())
     } else if (t.is_region_ptr() || t.is_box()) && !common::type_is_fat_ptr(ccx, t) {
-        b.load_nonnull(ptr)
+        b.load_nonnull(ptr, alignment.to_align())
     } else {
-        b.load(ptr)
+        b.load(ptr, alignment.to_align())
     }
 }
 
 /// Helper for storing values in memory. Does the necessary conversion if the in-memory type
 /// differs from the type used for SSA values.
-pub fn store_ty<'a, 'tcx>(cx: &Builder<'a, 'tcx>, v: ValueRef, dst: ValueRef, t: Ty<'tcx>) {
+pub fn store_ty<'a, 'tcx>(cx: &Builder<'a, 'tcx>, v: ValueRef, dst: ValueRef,
+                          dst_align: Alignment, t: Ty<'tcx>) {
     debug!("store_ty: {:?} : {:?} <- {:?}", Value(dst), t, Value(v));
 
     if common::type_is_fat_ptr(cx.ccx, t) {
         let lladdr = cx.extract_value(v, abi::FAT_PTR_ADDR);
         let llextra = cx.extract_value(v, abi::FAT_PTR_EXTRA);
-        store_fat_ptr(cx, lladdr, llextra, dst, t);
+        store_fat_ptr(cx, lladdr, llextra, dst, dst_align, t);
     } else {
-        cx.store(from_immediate(cx, v), dst, None);
+        cx.store(from_immediate(cx, v), dst, dst_align.to_align());
     }
 }
 
@@ -449,24 +455,25 @@ pub fn store_fat_ptr<'a, 'tcx>(cx: &Builder<'a, 'tcx>,
                                data: ValueRef,
                                extra: ValueRef,
                                dst: ValueRef,
+                               dst_align: Alignment,
                                _ty: Ty<'tcx>) {
     // FIXME: emit metadata
-    cx.store(data, get_dataptr(cx, dst), None);
-    cx.store(extra, get_meta(cx, dst), None);
+    cx.store(data, get_dataptr(cx, dst), dst_align.to_align());
+    cx.store(extra, get_meta(cx, dst), dst_align.to_align());
 }
 
 pub fn load_fat_ptr<'a, 'tcx>(
-    b: &Builder<'a, 'tcx>, src: ValueRef, t: Ty<'tcx>
+    b: &Builder<'a, 'tcx>, src: ValueRef, alignment: Alignment, t: Ty<'tcx>
 ) -> (ValueRef, ValueRef) {
     let ptr = get_dataptr(b, src);
     let ptr = if t.is_region_ptr() || t.is_box() {
-        b.load_nonnull(ptr)
+        b.load_nonnull(ptr, alignment.to_align())
     } else {
-        b.load(ptr)
+        b.load(ptr, alignment.to_align())
     };
 
     // FIXME: emit metadata on `meta`.
-    let meta = b.load(get_meta(b, src));
+    let meta = b.load(get_meta(b, src), alignment.to_align());
 
     (ptr, meta)
 }
@@ -633,7 +640,7 @@ pub fn trans_ctor_shim<'a, 'tcx>(ccx: &CrateContext<'a, 'tcx>,
             bcx.alloca(fn_ty.ret.memory_ty(ccx), "sret_slot")
         };
         // Can return unsized value
-        let mut dest_val = LvalueRef::new_sized_ty(dest, sig.output());
+        let mut dest_val = LvalueRef::new_sized_ty(dest, sig.output(), Alignment::AbiAligned);
         dest_val.ty = LvalueTy::Downcast {
             adt_def: sig.output().ty_adt_def().unwrap(),
             substs: substs,
@@ -642,7 +649,7 @@ pub fn trans_ctor_shim<'a, 'tcx>(ccx: &CrateContext<'a, 'tcx>,
         let mut llarg_idx = fn_ty.ret.is_indirect() as usize;
         let mut arg_idx = 0;
         for (i, arg_ty) in sig.inputs().iter().enumerate() {
-            let lldestptr = dest_val.trans_field_ptr(&bcx, i);
+            let (lldestptr, _) = dest_val.trans_field_ptr(&bcx, i);
             let arg = &fn_ty.args[arg_idx];
             arg_idx += 1;
             if common::type_is_fat_ptr(bcx.ccx, arg_ty) {
@@ -662,14 +669,12 @@ pub fn trans_ctor_shim<'a, 'tcx>(ccx: &CrateContext<'a, 'tcx>,
         }
 
         if let Some(cast_ty) = fn_ty.ret.cast {
-            let load = bcx.load(bcx.pointercast(dest, cast_ty.ptr_to()));
-            let llalign = llalign_of_min(ccx, fn_ty.ret.ty);
-            unsafe {
-                llvm::LLVMSetAlignment(load, llalign);
-            }
-            bcx.ret(load)
+            bcx.ret(bcx.load(
+                bcx.pointercast(dest, cast_ty.ptr_to()),
+                Some(llalign_of_min(ccx, fn_ty.ret.ty))
+            ));
         } else {
-            bcx.ret(bcx.load(dest))
+            bcx.ret(bcx.load(dest, None))
         }
     } else {
         bcx.ret_void();
diff --git a/src/librustc_trans/builder.rs b/src/librustc_trans/builder.rs
index c113716ca5c..89aaa8b6630 100644
--- a/src/librustc_trans/builder.rs
+++ b/src/librustc_trans/builder.rs
@@ -19,9 +19,8 @@ use machine::llalign_of_pref;
 use type_::Type;
 use value::Value;
 use libc::{c_uint, c_char};
-use rustc::ty::{Ty, TyCtxt, TypeFoldable};
+use rustc::ty::TyCtxt;
 use rustc::session::Session;
-use type_of;
 
 use std::borrow::Cow;
 use std::ffi::CString;
@@ -486,11 +485,6 @@ impl<'a, 'tcx> Builder<'a, 'tcx> {
         builder.dynamic_alloca(ty, name)
     }
 
-    pub fn alloca_ty(&self, ty: Ty<'tcx>, name: &str) -> ValueRef {
-        assert!(!ty.has_param_types());
-        self.alloca(type_of::type_of(self.ccx, ty), name)
-    }
-
     pub fn dynamic_alloca(&self, ty: Type, name: &str) -> ValueRef {
         self.count_insn("alloca");
         unsafe {
@@ -511,10 +505,14 @@ impl<'a, 'tcx> Builder<'a, 'tcx> {
         }
     }
 
-    pub fn load(&self, ptr: ValueRef) -> ValueRef {
+    pub fn load(&self, ptr: ValueRef, align: Option<u32>) -> ValueRef {
         self.count_insn("load");
         unsafe {
-            llvm::LLVMBuildLoad(self.llbuilder, ptr, noname())
+            let load = llvm::LLVMBuildLoad(self.llbuilder, ptr, noname());
+            if let Some(align) = align {
+                llvm::LLVMSetAlignment(load, align as c_uint);
+            }
+            load
         }
     }
 
@@ -539,8 +537,9 @@ impl<'a, 'tcx> Builder<'a, 'tcx> {
 
 
     pub fn load_range_assert(&self, ptr: ValueRef, lo: u64,
-                             hi: u64, signed: llvm::Bool) -> ValueRef {
-        let value = self.load(ptr);
+                             hi: u64, signed: llvm::Bool,
+                             align: Option<u32>) -> ValueRef {
+        let value = self.load(ptr, align);
 
         unsafe {
             let t = llvm::LLVMGetElementType(llvm::LLVMTypeOf(ptr));
@@ -558,8 +557,8 @@ impl<'a, 'tcx> Builder<'a, 'tcx> {
         value
     }
 
-    pub fn load_nonnull(&self, ptr: ValueRef) -> ValueRef {
-        let value = self.load(ptr);
+    pub fn load_nonnull(&self, ptr: ValueRef, align: Option<u32>) -> ValueRef {
+        let value = self.load(ptr, align);
         unsafe {
             llvm::LLVMSetMetadata(value, llvm::MD_nonnull as c_uint,
                                   llvm::LLVMMDNodeInContext(self.ccx.llcx(), ptr::null(), 0));
diff --git a/src/librustc_trans/callee.rs b/src/librustc_trans/callee.rs
index c6b86c6ba48..4a8658dd2e3 100644
--- a/src/librustc_trans/callee.rs
+++ b/src/librustc_trans/callee.rs
@@ -41,6 +41,8 @@ use std::iter;
 
 use syntax_pos::DUMMY_SP;
 
+use mir::lvalue::Alignment;
+
 #[derive(Debug)]
 pub enum CalleeData {
     /// Constructor for enum variant/tuple-like-struct.
@@ -358,29 +360,27 @@ fn trans_fn_once_adapter_shim<'a, 'tcx>(
     let fn_ty = callee.direct_fn_type(bcx.ccx, &[]);
     let self_idx = fn_ty.ret.is_indirect() as usize;
     let env_arg = &orig_fn_ty.args[0];
-    let llenv = if env_arg.is_indirect() {
-        llargs[self_idx]
+    let env = if env_arg.is_indirect() {
+        LvalueRef::new_sized_ty(llargs[self_idx], closure_ty, Alignment::AbiAligned)
     } else {
-        let scratch = bcx.alloca_ty(closure_ty, "self");
+        let scratch = LvalueRef::alloca(&bcx, closure_ty, "self");
         let mut llarg_idx = self_idx;
-        env_arg.store_fn_arg(&bcx, &mut llarg_idx, scratch);
+        env_arg.store_fn_arg(&bcx, &mut llarg_idx, scratch.llval);
         scratch
     };
 
-    debug!("trans_fn_once_adapter_shim: env={:?}", Value(llenv));
+    debug!("trans_fn_once_adapter_shim: env={:?}", env);
     // Adjust llargs such that llargs[self_idx..] has the call arguments.
     // For zero-sized closures that means sneaking in a new argument.
     if env_arg.is_ignore() {
-        llargs.insert(self_idx, llenv);
+        llargs.insert(self_idx, env.llval);
     } else {
-        llargs[self_idx] = llenv;
+        llargs[self_idx] = env.llval;
     }
 
     // Call the by-ref closure body with `self` in a cleanup scope,
     // to drop `self` when the body returns, or in case it unwinds.
-    let self_scope = CleanupScope::schedule_drop_mem(
-        &bcx, LvalueRef::new_sized_ty(llenv, closure_ty)
-    );
+    let self_scope = CleanupScope::schedule_drop_mem(&bcx, env);
 
     let llfn = callee.reify(bcx.ccx);
     let llret;
@@ -512,7 +512,7 @@ fn trans_fn_pointer_shim<'a, 'tcx>(
     let llfnpointer = llfnpointer.unwrap_or_else(|| {
         // the first argument (`self`) will be ptr to the fn pointer
         if is_by_ref {
-            bcx.load(self_arg)
+            bcx.load(self_arg, None)
         } else {
             self_arg
         }
diff --git a/src/librustc_trans/glue.rs b/src/librustc_trans/glue.rs
index fdefd37549c..ba7558b8acf 100644
--- a/src/librustc_trans/glue.rs
+++ b/src/librustc_trans/glue.rs
@@ -13,7 +13,6 @@
 // Code relating to drop glue.
 
 use std;
-use std::ptr;
 use std::iter;
 
 use llvm;
@@ -41,6 +40,7 @@ use Disr;
 use builder::Builder;
 
 use syntax_pos::DUMMY_SP;
+use mir::lvalue::Alignment;
 
 pub fn trans_exchange_free_ty<'a, 'tcx>(bcx: &Builder<'a, 'tcx>, ptr: LvalueRef<'tcx>) {
     let content_ty = ptr.ty.to_ty(bcx.tcx());
@@ -199,9 +199,9 @@ pub fn implement_drop_glue<'a, 'tcx>(ccx: &CrateContext<'a, 'tcx>, g: DropGlueKi
 
     let value = get_param(llfn, 0);
     let ptr = if ccx.shared().type_is_sized(t) {
-        LvalueRef::new_sized_ty(value, t)
+        LvalueRef::new_sized_ty(value, t, Alignment::AbiAligned)
     } else {
-        LvalueRef::new_unsized_ty(value, get_param(llfn, 1), t)
+        LvalueRef::new_unsized_ty(value, get_param(llfn, 1), t, Alignment::AbiAligned)
     };
 
     let skip_dtor = match g {
@@ -216,11 +216,13 @@ pub fn implement_drop_glue<'a, 'tcx>(ccx: &CrateContext<'a, 'tcx>, g: DropGlueKi
             assert!(!skip_dtor);
             let content_ty = t.boxed_ty();
             let ptr = if !bcx.ccx.shared().type_is_sized(content_ty) {
-                let llbox = bcx.load(get_dataptr(&bcx, ptr.llval));
-                let info = bcx.load(get_meta(&bcx, ptr.llval));
-                LvalueRef::new_unsized_ty(llbox, info, content_ty)
+                let llbox = bcx.load(get_dataptr(&bcx, ptr.llval), None);
+                let info = bcx.load(get_meta(&bcx, ptr.llval), None);
+                LvalueRef::new_unsized_ty(llbox, info, content_ty, Alignment::AbiAligned)
             } else {
-                LvalueRef::new_sized_ty(bcx.load(ptr.llval), content_ty)
+                LvalueRef::new_sized_ty(
+                    bcx.load(ptr.llval, None),
+                    content_ty, Alignment::AbiAligned)
             };
             drop_ty(&bcx, ptr);
             trans_exchange_free_ty(&bcx, ptr);
@@ -231,7 +233,7 @@ pub fn implement_drop_glue<'a, 'tcx>(ccx: &CrateContext<'a, 'tcx>, g: DropGlueKi
             // versus without calling Drop::drop. Assert caller is
             // okay with always calling the Drop impl, if any.
             assert!(!skip_dtor);
-            let dtor = bcx.load(ptr.llextra);
+            let dtor = bcx.load(ptr.llextra, None);
             bcx.call(dtor, &[ptr.llval], None);
             bcx
         }
@@ -384,7 +386,7 @@ pub fn size_and_align_of_dst<'a, 'tcx>(bcx: &Builder<'a, 'tcx>, t: Ty<'tcx>, inf
             let info = bcx.pointercast(info, Type::int(bcx.ccx).ptr_to());
             let size_ptr = bcx.gepi(info, &[1]);
             let align_ptr = bcx.gepi(info, &[2]);
-            (bcx.load(size_ptr), bcx.load(align_ptr))
+            (bcx.load(size_ptr, None), bcx.load(align_ptr, None))
         }
         ty::TySlice(_) | ty::TyStr => {
             let unit_ty = t.sequence_element_type(bcx.tcx());
@@ -416,8 +418,8 @@ fn drop_structural_ty<'a, 'tcx>(
         let tcx = cx.tcx();
         for (i, field) in variant.fields.iter().enumerate() {
             let arg = monomorphize::field_ty(tcx, substs, field);
-            let field_ptr = av.trans_field_ptr(&cx, i);
-            drop_ty(&cx, LvalueRef::new_sized_ty(field_ptr, arg));
+            let (field_ptr, align) = av.trans_field_ptr(&cx, i);
+            drop_ty(&cx, LvalueRef::new_sized_ty(field_ptr, arg, align));
         }
     }
 
@@ -426,8 +428,8 @@ fn drop_structural_ty<'a, 'tcx>(
     match t.sty {
         ty::TyClosure(def_id, substs) => {
             for (i, upvar_ty) in substs.upvar_tys(def_id, cx.tcx()).enumerate() {
-                let llupvar = ptr.trans_field_ptr(&cx, i);
-                drop_ty(&cx, LvalueRef::new_sized_ty(llupvar, upvar_ty));
+                let (llupvar, align) = ptr.trans_field_ptr(&cx, i);
+                drop_ty(&cx, LvalueRef::new_sized_ty(llupvar, upvar_ty, align));
             }
         }
         ty::TyArray(_, n) => {
@@ -435,29 +437,29 @@ fn drop_structural_ty<'a, 'tcx>(
             let len = C_uint(cx.ccx, n);
             let unit_ty = t.sequence_element_type(cx.tcx());
             cx = tvec::slice_for_each(&cx, base, unit_ty, len,
-                |bb, vv| drop_ty(bb, LvalueRef::new_sized_ty(vv, unit_ty)));
+                |bb, vv| drop_ty(bb, LvalueRef::new_sized_ty(vv, unit_ty, ptr.alignment)));
         }
         ty::TySlice(_) | ty::TyStr => {
             let unit_ty = t.sequence_element_type(cx.tcx());
             cx = tvec::slice_for_each(&cx, ptr.llval, unit_ty, ptr.llextra,
-                |bb, vv| drop_ty(bb, LvalueRef::new_sized_ty(vv, unit_ty)));
+                |bb, vv| drop_ty(bb, LvalueRef::new_sized_ty(vv, unit_ty, ptr.alignment)));
         }
         ty::TyTuple(ref args, _) => {
             for (i, arg) in args.iter().enumerate() {
-                let llfld_a = ptr.trans_field_ptr(&cx, i);
-                drop_ty(&cx, LvalueRef::new_sized_ty(llfld_a, *arg));
+                let (llfld_a, align) = ptr.trans_field_ptr(&cx, i);
+                drop_ty(&cx, LvalueRef::new_sized_ty(llfld_a, *arg, align));
             }
         }
         ty::TyAdt(adt, substs) => match adt.adt_kind() {
             AdtKind::Struct => {
                 for (i, field) in adt.variants[0].fields.iter().enumerate() {
                     let field_ty = monomorphize::field_ty(cx.tcx(), substs, field);
-                    let mut field_ptr = ptr.clone();
-                    field_ptr.llval = ptr.trans_field_ptr(&cx, i);
-                    field_ptr.ty = LvalueTy::from_ty(field_ty);
-                    if cx.ccx.shared().type_is_sized(field_ty) {
-                        field_ptr.llextra = ptr::null_mut();
-                    }
+                    let (llval, align) = ptr.trans_field_ptr(&cx, i);
+                    let field_ptr = if cx.ccx.shared().type_is_sized(field_ty) {
+                        LvalueRef::new_sized_ty(llval, field_ty, align)
+                    } else {
+                        LvalueRef::new_unsized_ty(llval, ptr.llextra, field_ty, align)
+                    };
                     drop_ty(&cx, field_ptr);
                 }
             }
@@ -490,9 +492,12 @@ fn drop_structural_ty<'a, 'tcx>(
                     layout::General { .. } |
                     layout::RawNullablePointer { .. } |
                     layout::StructWrappedNullablePointer { .. } => {
-                        let lldiscrim_a = adt::trans_get_discr(&cx, t, ptr.llval, None, false);
+                        let lldiscrim_a = adt::trans_get_discr(
+                            &cx, t, ptr.llval, ptr.alignment, None, false);
                         let tcx = cx.tcx();
-                        drop_ty(&cx, LvalueRef::new_sized_ty(lldiscrim_a, tcx.types.isize));
+                        // FIXME: why are dropping an isize?
+                        drop_ty(&cx, LvalueRef::new_sized_ty(lldiscrim_a, tcx.types.isize,
+                                                             ptr.alignment));
 
                         // Create a fall-through basic block for the "else" case of
                         // the switch instruction we're about to generate. Note that
diff --git a/src/librustc_trans/intrinsic.rs b/src/librustc_trans/intrinsic.rs
index 37f63e74996..1335a87b0b7 100644
--- a/src/librustc_trans/intrinsic.rs
+++ b/src/librustc_trans/intrinsic.rs
@@ -36,6 +36,8 @@ use syntax_pos::Span;
 use std::cmp::Ordering;
 use std::iter;
 
+use mir::lvalue::Alignment;
+
 fn get_simple_intrinsic(ccx: &CrateContext, name: &str) -> Option<ValueRef> {
     let llvm_name = match name {
         "sqrtf32" => "llvm.sqrt.f32",
@@ -243,7 +245,7 @@ pub fn trans_intrinsic_call<'a, 'tcx>(bcx: &Builder<'a, 'tcx>,
                 bcx.volatile_store(llargs[2], get_meta(bcx, llargs[0]));
             } else {
                 let val = if fn_ty.args[1].is_indirect() {
-                    bcx.load(llargs[1])
+                    bcx.load(llargs[1], None) // FIXME: this is incorrect
                 } else {
                     from_immediate(bcx, llargs[1])
                 };
@@ -348,7 +350,7 @@ pub fn trans_intrinsic_call<'a, 'tcx>(bcx: &Builder<'a, 'tcx>,
             let val_ty = substs.type_at(0);
             match val_ty.sty {
                 ty::TyAdt(adt, ..) if adt.is_enum() => {
-                    adt::trans_get_discr(bcx, val_ty, llargs[0],
+                    adt::trans_get_discr(bcx, val_ty, llargs[0], Alignment::AbiAligned,
                                          Some(llret_ty), true)
                 }
                 _ => C_null(llret_ty)
@@ -547,8 +549,11 @@ pub fn trans_intrinsic_call<'a, 'tcx>(bcx: &Builder<'a, 'tcx>,
                         // destructors, and the contents are SIMD
                         // etc.
                         assert!(!bcx.ccx.shared().type_needs_drop(arg_type));
-                        let arg = LvalueRef::new_sized_ty(llarg, arg_type);
-                        (0..contents.len()).map(|i| bcx.load(arg.trans_field_ptr(bcx, i))).collect()
+                        let arg = LvalueRef::new_sized_ty(llarg, arg_type, Alignment::AbiAligned);
+                        (0..contents.len()).map(|i| {
+                            let (ptr, align) = arg.trans_field_ptr(bcx, i);
+                            bcx.load(ptr, align.to_align())
+                        }).collect()
                     }
                     intrinsics::Type::Pointer(_, Some(ref llvm_elem), _) => {
                         let llvm_elem = one(ty_to_type(bcx.ccx, llvm_elem, &mut false));
@@ -624,7 +629,7 @@ pub fn trans_intrinsic_call<'a, 'tcx>(bcx: &Builder<'a, 'tcx>,
             let ptr = bcx.pointercast(llresult, ty.ptr_to());
             bcx.store(llval, ptr, Some(type_of::align_of(ccx, ret_ty)));
         } else {
-            store_ty(bcx, llval, llresult, ret_ty);
+            store_ty(bcx, llval, llresult, Alignment::AbiAligned, ret_ty);
         }
     }
 }
@@ -780,10 +785,10 @@ fn trans_msvc_try<'a, 'tcx>(bcx: &Builder<'a, 'tcx>,
             None => bug!("msvc_try_filter not defined"),
         };
         let tok = catchpad.catch_pad(cs, &[tydesc, C_i32(ccx, 0), slot]);
-        let addr = catchpad.load(slot);
-        let arg1 = catchpad.load(addr);
+        let addr = catchpad.load(slot, None);
+        let arg1 = catchpad.load(addr, None);
         let val1 = C_i32(ccx, 1);
-        let arg2 = catchpad.load(catchpad.inbounds_gep(addr, &[val1]));
+        let arg2 = catchpad.load(catchpad.inbounds_gep(addr, &[val1]), None);
         let local_ptr = catchpad.bitcast(local_ptr, i64p);
         catchpad.store(arg1, local_ptr, None);
         catchpad.store(arg2, catchpad.inbounds_gep(local_ptr, &[val1]), None);
diff --git a/src/librustc_trans/meth.rs b/src/librustc_trans/meth.rs
index aecba2f57e5..3033ae61d20 100644
--- a/src/librustc_trans/meth.rs
+++ b/src/librustc_trans/meth.rs
@@ -36,7 +36,7 @@ pub fn get_virtual_method<'a, 'tcx>(bcx: &Builder<'a, 'tcx>,
     debug!("get_virtual_method(vtable_index={}, llvtable={:?})",
            vtable_index, Value(llvtable));
 
-    bcx.load(bcx.gepi(llvtable, &[vtable_index + VTABLE_OFFSET]))
+    bcx.load(bcx.gepi(llvtable, &[vtable_index + VTABLE_OFFSET]), None)
 }
 
 /// Generate a shim function that allows an object type like `SomeTrait` to
diff --git a/src/librustc_trans/mir/block.rs b/src/librustc_trans/mir/block.rs
index 027779aca63..356d3e7372b 100644
--- a/src/librustc_trans/mir/block.rs
+++ b/src/librustc_trans/mir/block.rs
@@ -37,7 +37,7 @@ use std::cmp;
 use super::{MirContext, LocalRef};
 use super::analyze::CleanupKind;
 use super::constant::Const;
-use super::lvalue::LvalueRef;
+use super::lvalue::{Alignment, LvalueRef};
 use super::operand::OperandRef;
 use super::operand::OperandValue::{Pair, Ref, Immediate};
 
@@ -120,7 +120,7 @@ impl<'a, 'tcx> MirContext<'a, 'tcx> {
                     bcx.cleanup_ret(cleanup_pad, None);
                 } else {
                     let ps = self.get_personality_slot(&bcx);
-                    let lp = bcx.load(ps);
+                    let lp = bcx.load(ps, None);
                     Lifetime::End.call(&bcx, ps);
                     if !bcx.sess().target.target.options.custom_unwind_resume {
                         bcx.resume(lp);
@@ -147,7 +147,8 @@ impl<'a, 'tcx> MirContext<'a, 'tcx> {
             mir::TerminatorKind::Switch { ref discr, ref adt_def, ref targets } => {
                 let discr_lvalue = self.trans_lvalue(&bcx, discr);
                 let ty = discr_lvalue.ty.to_ty(bcx.tcx());
-                let discr = adt::trans_get_discr(&bcx, ty, discr_lvalue.llval, None, true);
+                let discr = adt::trans_get_discr(&bcx, ty, discr_lvalue.llval, discr_lvalue.alignment,
+                                                 None, true);
 
                 let mut bb_hist = FxHashMap();
                 for target in targets {
@@ -179,7 +180,8 @@ impl<'a, 'tcx> MirContext<'a, 'tcx> {
 
             mir::TerminatorKind::SwitchInt { ref discr, switch_ty, ref values, ref targets } => {
                 let (otherwise, targets) = targets.split_last().unwrap();
-                let discr = bcx.load(self.trans_lvalue(&bcx, discr).llval);
+                let lv = self.trans_lvalue(&bcx, discr);
+                let discr = bcx.load(lv.llval, lv.alignment.to_align());
                 let discr = base::to_immediate(&bcx, discr, switch_ty);
                 let switch = bcx.switch(discr, llblock(self, *otherwise), values.len());
                 for (value, target) in values.iter().zip(targets) {
@@ -202,7 +204,7 @@ impl<'a, 'tcx> MirContext<'a, 'tcx> {
                         LocalRef::Operand(None) => bug!("use of return before def"),
                         LocalRef::Lvalue(tr_lvalue) => {
                             OperandRef {
-                                val: Ref(tr_lvalue.llval),
+                                val: Ref(tr_lvalue.llval, tr_lvalue.alignment),
                                 ty: tr_lvalue.ty.to_ty(bcx.tcx())
                             }
                         }
@@ -210,21 +212,22 @@ impl<'a, 'tcx> MirContext<'a, 'tcx> {
                     let llslot = match op.val {
                         Immediate(_) | Pair(..) => {
                             let llscratch = bcx.alloca(ret.original_ty, "ret");
-                            self.store_operand(&bcx, llscratch, op, None);
+                            self.store_operand(&bcx, llscratch, None, op);
                             llscratch
                         }
-                        Ref(llval) => llval
+                        Ref(llval, align) => {
+                            assert_eq!(align, Alignment::AbiAligned, "return pointer is unaligned!");
+                            llval
+                        }
                     };
-                    let load = bcx.load(bcx.pointercast(llslot, cast_ty.ptr_to()));
-                    let llalign = llalign_of_min(bcx.ccx, ret.ty);
-                    unsafe {
-                        llvm::LLVMSetAlignment(load, llalign);
-                    }
+                    let load = bcx.load(
+                        bcx.pointercast(llslot, cast_ty.ptr_to()),
+                        Some(llalign_of_min(bcx.ccx, ret.ty)));
                     load
                 } else {
                     let op = self.trans_consume(&bcx, &mir::Lvalue::Local(mir::RETURN_POINTER));
-                    if let Ref(llval) = op.val {
-                        base::load_ty(&bcx, llval, op.ty)
+                    if let Ref(llval, align) = op.val {
+                        base::load_ty(&bcx, llval, align, op.ty)
                     } else {
                         op.pack_if_pair(&bcx).immediate()
                     }
@@ -425,7 +428,7 @@ impl<'a, 'tcx> MirContext<'a, 'tcx> {
                     // The first argument is a thin destination pointer.
                     let llptr = self.trans_operand(&bcx, &args[0]).immediate();
                     let val = self.trans_operand(&bcx, &args[1]);
-                    self.store_operand(&bcx, llptr, val, None);
+                    self.store_operand(&bcx, llptr, None, val);
                     funclet_br(self, bcx, target);
                     return;
                 }
@@ -550,7 +553,7 @@ impl<'a, 'tcx> MirContext<'a, 'tcx> {
                         if let ReturnDest::IndirectOperand(dst, _) = ret_dest {
                             // Make a fake operand for store_return
                             let op = OperandRef {
-                                val: Ref(dst),
+                                val: Ref(dst, Alignment::AbiAligned),
                                 ty: sig.output(),
                             };
                             self.store_return(&bcx, ret_dest, fn_ty.ret, op);
@@ -652,33 +655,39 @@ impl<'a, 'tcx> MirContext<'a, 'tcx> {
         }
 
         // Force by-ref if we have to load through a cast pointer.
-        let (mut llval, by_ref) = match op.val {
+        let (mut llval, align, by_ref) = match op.val {
             Immediate(_) | Pair(..) => {
                 if arg.is_indirect() || arg.cast.is_some() {
                     let llscratch = bcx.alloca(arg.original_ty, "arg");
-                    self.store_operand(bcx, llscratch, op, None);
-                    (llscratch, true)
+                    self.store_operand(bcx, llscratch, None, op);
+                    (llscratch, Alignment::AbiAligned, true)
                 } else {
-                    (op.pack_if_pair(bcx).immediate(), false)
+                    (op.pack_if_pair(bcx).immediate(), Alignment::AbiAligned, false)
                 }
             }
-            Ref(llval) => (llval, true)
+            Ref(llval, Alignment::Packed) if arg.is_indirect() => {
+                // `foo(packed.large_field)`. We can't pass the (unaligned) field directly. I
+                // think that ATM (Rust 1.16) we only pass temporaries, but we shouldn't
+                // have scary latent bugs around.
+
+                let llscratch = bcx.alloca(arg.original_ty, "arg");
+                base::memcpy_ty(bcx, llscratch, llval, op.ty, Some(1));
+                (llscratch, Alignment::AbiAligned, true)
+            }
+            Ref(llval, align) => (llval, align, true)
         };
 
         if by_ref && !arg.is_indirect() {
             // Have to load the argument, maybe while casting it.
             if arg.original_ty == Type::i1(bcx.ccx) {
                 // We store bools as i8 so we need to truncate to i1.
-                llval = bcx.load_range_assert(llval, 0, 2, llvm::False);
+                llval = bcx.load_range_assert(llval, 0, 2, llvm::False, None);
                 llval = bcx.trunc(llval, arg.original_ty);
             } else if let Some(ty) = arg.cast {
-                llval = bcx.load(bcx.pointercast(llval, ty.ptr_to()));
-                let llalign = llalign_of_min(bcx.ccx, arg.ty);
-                unsafe {
-                    llvm::LLVMSetAlignment(llval, llalign);
-                }
+                llval = bcx.load(bcx.pointercast(llval, ty.ptr_to()),
+                                 align.min_with(llalign_of_min(bcx.ccx, arg.ty)));
             } else {
-                llval = bcx.load(llval);
+                llval = bcx.load(llval, align.to_align());
             }
         }
 
@@ -702,16 +711,16 @@ impl<'a, 'tcx> MirContext<'a, 'tcx> {
 
         // Handle both by-ref and immediate tuples.
         match tuple.val {
-            Ref(llval) => {
+            Ref(llval, align) => {
                 for (n, &ty) in arg_types.iter().enumerate() {
-                    let ptr = LvalueRef::new_sized_ty(llval, tuple.ty);
-                    let ptr = ptr.trans_field_ptr(bcx, n);
+                    let ptr = LvalueRef::new_sized_ty(llval, tuple.ty, align);
+                    let (ptr, align) = ptr.trans_field_ptr(bcx, n);
                     let val = if common::type_is_fat_ptr(bcx.ccx, ty) {
-                        let (lldata, llextra) = base::load_fat_ptr(bcx, ptr, ty);
+                        let (lldata, llextra) = base::load_fat_ptr(bcx, ptr, align, ty);
                         Pair(lldata, llextra)
                     } else {
                         // trans_argument will load this if it needs to
-                        Ref(ptr)
+                        Ref(ptr, align)
                     };
                     let op = OperandRef {
                         val: val,
@@ -839,15 +848,15 @@ impl<'a, 'tcx> MirContext<'a, 'tcx> {
                     return if fn_ret_ty.is_indirect() {
                         // Odd, but possible, case, we have an operand temporary,
                         // but the calling convention has an indirect return.
-                        let tmp = bcx.alloca_ty(ret_ty, "tmp_ret");
-                        llargs.push(tmp);
-                        ReturnDest::IndirectOperand(tmp, index)
+                        let tmp = LvalueRef::alloca(bcx, ret_ty, "tmp_ret");
+                        llargs.push(tmp.llval);
+                        ReturnDest::IndirectOperand(tmp.llval, index)
                     } else if is_intrinsic {
                         // Currently, intrinsics always need a location to store
                         // the result. so we create a temporary alloca for the
                         // result
-                        let tmp = bcx.alloca_ty(ret_ty, "tmp_ret");
-                        ReturnDest::IndirectOperand(tmp, index)
+                        let tmp = LvalueRef::alloca(bcx, ret_ty, "tmp_ret");
+                        ReturnDest::IndirectOperand(tmp.llval, index)
                     } else {
                         ReturnDest::DirectOperand(index)
                     };
@@ -892,7 +901,7 @@ impl<'a, 'tcx> MirContext<'a, 'tcx> {
         let in_type = val.ty;
         let out_type = dst.ty.to_ty(bcx.tcx());;
         let llalign = cmp::min(align_of(bcx.ccx, in_type), align_of(bcx.ccx, out_type));
-        self.store_operand(bcx, cast_ptr, val, Some(llalign));
+        self.store_operand(bcx, cast_ptr, Some(llalign), val);
     }
 
 
@@ -908,15 +917,15 @@ impl<'a, 'tcx> MirContext<'a, 'tcx> {
             Nothing => (),
             Store(dst) => ret_ty.store(bcx, op.immediate(), dst),
             IndirectOperand(tmp, index) => {
-                let op = self.trans_load(bcx, tmp, op.ty);
+                let op = self.trans_load(bcx, tmp, Alignment::AbiAligned, op.ty);
                 self.locals[index] = LocalRef::Operand(Some(op));
             }
             DirectOperand(index) => {
                 // If there is a cast, we have to store and reload.
                 let op = if ret_ty.cast.is_some() {
-                    let tmp = bcx.alloca_ty(op.ty, "tmp_ret");
-                    ret_ty.store(bcx, op.immediate(), tmp);
-                    self.trans_load(bcx, tmp, op.ty)
+                    let tmp = LvalueRef::alloca(bcx, op.ty, "tmp_ret");
+                    ret_ty.store(bcx, op.immediate(), tmp.llval);
+                    self.trans_load(bcx, tmp.llval, tmp.alignment, op.ty)
                 } else {
                     op.unpack_if_pair(bcx)
                 };
diff --git a/src/librustc_trans/mir/constant.rs b/src/librustc_trans/mir/constant.rs
index f92faaa0508..11668e792e3 100644
--- a/src/librustc_trans/mir/constant.rs
+++ b/src/librustc_trans/mir/constant.rs
@@ -40,6 +40,7 @@ use syntax_pos::Span;
 use std::fmt;
 use std::ptr;
 
+use super::lvalue::Alignment;
 use super::operand::{OperandRef, OperandValue};
 use super::MirContext;
 
@@ -140,7 +141,7 @@ impl<'tcx> Const<'tcx> {
             // a constant LLVM global and cast its address if necessary.
             let align = type_of::align_of(ccx, self.ty);
             let ptr = consts::addr_of(ccx, self.llval, align, "const");
-            OperandValue::Ref(consts::ptrcast(ptr, llty.ptr_to()))
+            OperandValue::Ref(consts::ptrcast(ptr, llty.ptr_to()), Alignment::AbiAligned)
         };
 
         OperandRef {
diff --git a/src/librustc_trans/mir/lvalue.rs b/src/librustc_trans/mir/lvalue.rs
index bd6e70639bb..6800fa34152 100644
--- a/src/librustc_trans/mir/lvalue.rs
+++ b/src/librustc_trans/mir/lvalue.rs
@@ -25,10 +25,52 @@ use value::Value;
 use glue;
 
 use std::ptr;
+use std::ops;
 
 use super::{MirContext, LocalRef};
 use super::operand::OperandValue;
 
+#[derive(Copy, Clone, Debug, PartialEq, Eq)]
+pub enum Alignment {
+    Packed,
+    AbiAligned,
+}
+
+impl ops::BitOr for Alignment {
+    type Output = Self;
+
+    fn bitor(self, rhs: Self) -> Self {
+        match (self, rhs) {
+            (Alignment::Packed, _) => Alignment::Packed,
+            (Alignment::AbiAligned, a) => a,
+        }
+    }
+}
+
+impl Alignment {
+    pub fn from_packed(packed: bool) -> Self {
+        if packed {
+            Alignment::Packed
+        } else {
+            Alignment::AbiAligned
+        }
+    }
+
+    pub fn to_align(self) -> Option<u32> {
+        match self {
+            Alignment::Packed => Some(1),
+            Alignment::AbiAligned => None,
+        }
+    }
+
+    pub fn min_with(self, align: u32) -> Option<u32> {
+        match self {
+            Alignment::Packed => Some(1),
+            Alignment::AbiAligned => Some(align),
+        }
+    }
+}
+
 #[derive(Copy, Clone, Debug)]
 pub struct LvalueRef<'tcx> {
     /// Pointer to the contents of the lvalue
@@ -39,25 +81,38 @@ pub struct LvalueRef<'tcx> {
 
     /// Monomorphized type of this lvalue, including variant information
     pub ty: LvalueTy<'tcx>,
+
+    /// Whether this lvalue is known to be aligned according to its layout
+    pub alignment: Alignment,
 }
 
 impl<'a, 'tcx> LvalueRef<'tcx> {
-    pub fn new_sized(llval: ValueRef, lvalue_ty: LvalueTy<'tcx>) -> LvalueRef<'tcx> {
-        LvalueRef { llval: llval, llextra: ptr::null_mut(), ty: lvalue_ty }
+    pub fn new_sized(llval: ValueRef, lvalue_ty: LvalueTy<'tcx>,
+                     alignment: Alignment) -> LvalueRef<'tcx> {
+        LvalueRef { llval: llval, llextra: ptr::null_mut(), ty: lvalue_ty, alignment: alignment }
     }
 
-    pub fn new_sized_ty(llval: ValueRef, ty: Ty<'tcx>) -> LvalueRef<'tcx> {
-        LvalueRef::new_sized(llval, LvalueTy::from_ty(ty))
+    pub fn new_sized_ty(llval: ValueRef, ty: Ty<'tcx>, alignment: Alignment) -> LvalueRef<'tcx> {
+        LvalueRef::new_sized(llval, LvalueTy::from_ty(ty), alignment)
     }
 
-    pub fn new_unsized_ty(llval: ValueRef, llextra: ValueRef, ty: Ty<'tcx>) -> LvalueRef<'tcx> {
+    pub fn new_unsized_ty(llval: ValueRef, llextra: ValueRef, ty: Ty<'tcx>, alignment: Alignment)
+                          -> LvalueRef<'tcx> {
         LvalueRef {
             llval: llval,
             llextra: llextra,
             ty: LvalueTy::from_ty(ty),
+            alignment: alignment,
         }
     }
 
+    pub fn alloca(bcx: &Builder<'a, 'tcx>, ty: Ty<'tcx>, name: &str) -> LvalueRef<'tcx> {
+        debug!("alloca({:?}: {:?})", name, ty);
+        let tmp = bcx.alloca(type_of::type_of(bcx.ccx, ty), name);
+        assert!(!ty.has_param_types());
+        Self::new_sized_ty(tmp, ty, Alignment::AbiAligned)
+    }
+
     pub fn len(&self, ccx: &CrateContext<'a, 'tcx>) -> ValueRef {
         let ty = self.ty.to_ty(ccx.tcx());
         match ty.sty {
@@ -81,10 +136,12 @@ impl<'a, 'tcx> LvalueRef<'tcx> {
         fields: &Vec<Ty<'tcx>>,
         ix: usize,
         needs_cast: bool
-    ) -> ValueRef {
+    ) -> (ValueRef, Alignment) {
         let fty = fields[ix];
         let ccx = bcx.ccx;
 
+        let alignment = self.alignment | Alignment::from_packed(st.packed);
+
         let ptr_val = if needs_cast {
             let fields = st.field_index_by_increasing_offset().map(|i| {
                 type_of::in_memory_type_of(ccx, fields[i])
@@ -101,14 +158,14 @@ impl<'a, 'tcx> LvalueRef<'tcx> {
         //   * Field is sized - pointer is properly aligned already
         if st.offsets[ix] == layout::Size::from_bytes(0) || st.packed ||
             bcx.ccx.shared().type_is_sized(fty) {
-                return bcx.struct_gep(ptr_val, st.memory_index[ix] as usize);
+                return (bcx.struct_gep(ptr_val, st.memory_index[ix] as usize), alignment);
             }
 
         // If the type of the last field is [T] or str, then we don't need to do
         // any adjusments
         match fty.sty {
             ty::TySlice(..) | ty::TyStr => {
-                return bcx.struct_gep(ptr_val, st.memory_index[ix] as usize);
+                return (bcx.struct_gep(ptr_val, st.memory_index[ix] as usize), alignment);
             }
             _ => ()
         }
@@ -117,7 +174,7 @@ impl<'a, 'tcx> LvalueRef<'tcx> {
         if !self.has_extra() {
             debug!("Unsized field `{}`, of `{:?}` has no metadata for adjustment",
                 ix, Value(ptr_val));
-            return bcx.struct_gep(ptr_val, ix);
+            return (bcx.struct_gep(ptr_val, ix), alignment);
         }
 
         // We need to get the pointer manually now.
@@ -163,11 +220,11 @@ impl<'a, 'tcx> LvalueRef<'tcx> {
         // Finally, cast back to the type expected
         let ll_fty = type_of::in_memory_type_of(bcx.ccx, fty);
         debug!("struct_field_ptr: Field type is {:?}", ll_fty);
-        bcx.pointercast(byte_ptr, ll_fty.ptr_to())
+        (bcx.pointercast(byte_ptr, ll_fty.ptr_to()), alignment)
     }
 
     /// Access a field, at a point when the value's case is known.
-    pub fn trans_field_ptr(self, bcx: &Builder<'a, 'tcx>, ix: usize) -> ValueRef {
+    pub fn trans_field_ptr(self, bcx: &Builder<'a, 'tcx>, ix: usize) -> (ValueRef, Alignment) {
         let discr = match self.ty {
             LvalueTy::Ty { .. } => 0,
             LvalueTy::Downcast { variant_index, .. } => variant_index,
@@ -186,17 +243,18 @@ impl<'a, 'tcx> LvalueRef<'tcx> {
             layout::Vector { count, .. } => {
                 assert_eq!(discr, 0);
                 assert!((ix as u64) < count);
-                bcx.struct_gep(self.llval, ix)
+                (bcx.struct_gep(self.llval, ix), self.alignment)
             }
             layout::General { discr: d, ref variants, .. } => {
                 let mut fields = adt::compute_fields(bcx.ccx, t, discr, false);
                 fields.insert(0, d.to_ty(&bcx.tcx(), false));
                 self.struct_field_ptr(bcx, &variants[discr], &fields, ix + 1, true)
             }
-            layout::UntaggedUnion { .. } => {
+            layout::UntaggedUnion { ref variants } => {
                 let fields = adt::compute_fields(bcx.ccx, t, 0, false);
                 let ty = type_of::in_memory_type_of(bcx.ccx, fields[ix]);
-                bcx.pointercast(self.llval, ty.ptr_to())
+                (bcx.pointercast(self.llval, ty.ptr_to()),
+                 self.alignment | Alignment::from_packed(variants.packed))
             }
             layout::RawNullablePointer { nndiscr, .. } |
             layout::StructWrappedNullablePointer { nndiscr,  .. } if discr as u64 != nndiscr => {
@@ -205,19 +263,19 @@ impl<'a, 'tcx> LvalueRef<'tcx> {
                 // (e.d., Result of Either with (), as one side.)
                 let ty = type_of::type_of(bcx.ccx, nullfields[ix]);
                 assert_eq!(machine::llsize_of_alloc(bcx.ccx, ty), 0);
-                bcx.pointercast(self.llval, ty.ptr_to())
+                (bcx.pointercast(self.llval, ty.ptr_to()), Alignment::Packed)
             }
             layout::RawNullablePointer { nndiscr, .. } => {
                 let nnty = adt::compute_fields(bcx.ccx, t, nndiscr as usize, false)[0];
                 assert_eq!(ix, 0);
                 assert_eq!(discr as u64, nndiscr);
                 let ty = type_of::type_of(bcx.ccx, nnty);
-                bcx.pointercast(self.llval, ty.ptr_to())
+                (bcx.pointercast(self.llval, ty.ptr_to()), self.alignment)
             }
             layout::StructWrappedNullablePointer { ref nonnull, nndiscr, .. } => {
                 assert_eq!(discr as u64, nndiscr);
                 self.struct_field_ptr(bcx, &nonnull,
-                    &adt::compute_fields(bcx.ccx, t, discr, false), ix, false)
+                     &adt::compute_fields(bcx.ccx, t, discr, false), ix, false)
             }
             _ => bug!("element access in type without elements: {} represented as {:#?}", t, l)
         }
@@ -250,7 +308,8 @@ impl<'a, 'tcx> MirContext<'a, 'tcx> {
             mir::Lvalue::Static(def_id) => {
                 let const_ty = self.monomorphized_lvalue_ty(lvalue);
                 LvalueRef::new_sized(consts::get_static(ccx, def_id),
-                                     LvalueTy::from_ty(const_ty))
+                                     LvalueTy::from_ty(const_ty),
+                                     Alignment::AbiAligned)
             },
             mir::Lvalue::Projection(box mir::Projection {
                 ref base,
@@ -264,18 +323,20 @@ impl<'a, 'tcx> MirContext<'a, 'tcx> {
                 let (llptr, llextra) = match ptr.val {
                     OperandValue::Immediate(llptr) => (llptr, ptr::null_mut()),
                     OperandValue::Pair(llptr, llextra) => (llptr, llextra),
-                    OperandValue::Ref(_) => bug!("Deref of by-Ref type {:?}", ptr.ty)
+                    OperandValue::Ref(..) => bug!("Deref of by-Ref type {:?}", ptr.ty)
                 };
                 LvalueRef {
                     llval: llptr,
                     llextra: llextra,
                     ty: projected_ty,
+                    alignment: Alignment::AbiAligned,
                 }
             }
             mir::Lvalue::Projection(ref projection) => {
                 let tr_base = self.trans_lvalue(bcx, &projection.base);
                 let projected_ty = tr_base.ty.projection_ty(tcx, &projection.elem);
                 let projected_ty = self.monomorphize(&projected_ty);
+                let align = tr_base.alignment;
 
                 let project_index = |llindex| {
                     let element = if let ty::TySlice(_) = tr_base.ty.to_ty(tcx).sty {
@@ -285,10 +346,10 @@ impl<'a, 'tcx> MirContext<'a, 'tcx> {
                         let zero = common::C_uint(bcx.ccx, 0u64);
                         bcx.inbounds_gep(tr_base.llval, &[zero, llindex])
                     };
-                    element
+                    (element, align)
                 };
 
-                let (llprojected, llextra) = match projection.elem {
+                let ((llprojected, align), llextra) = match projection.elem {
                     mir::ProjectionElem::Deref => bug!(),
                     mir::ProjectionElem::Field(ref field, _) => {
                         let llextra = if self.ccx.shared().type_is_sized(projected_ty.to_ty(tcx)) {
@@ -318,7 +379,7 @@ impl<'a, 'tcx> MirContext<'a, 'tcx> {
                     }
                     mir::ProjectionElem::Subslice { from, to } => {
                         let llindex = C_uint(bcx.ccx, from);
-                        let llbase = project_index(llindex);
+                        let (llbase, align) = project_index(llindex);
 
                         let base_ty = tr_base.ty.to_ty(bcx.tcx());
                         match base_ty.sty {
@@ -328,25 +389,26 @@ impl<'a, 'tcx> MirContext<'a, 'tcx> {
                                 let base_ty = self.monomorphized_lvalue_ty(lvalue);
                                 let llbasety = type_of::type_of(bcx.ccx, base_ty).ptr_to();
                                 let llbase = bcx.pointercast(llbase, llbasety);
-                                (llbase, ptr::null_mut())
+                                ((llbase, align), ptr::null_mut())
                             }
                             ty::TySlice(..) => {
                                 assert!(tr_base.llextra != ptr::null_mut());
                                 let lllen = bcx.sub(tr_base.llextra,
                                                     C_uint(bcx.ccx, from+to));
-                                (llbase, lllen)
+                                ((llbase, align), lllen)
                             }
                             _ => bug!("unexpected type {:?} in Subslice", base_ty)
                         }
                     }
                     mir::ProjectionElem::Downcast(..) => {
-                        (tr_base.llval, tr_base.llextra)
+                        ((tr_base.llval, align), tr_base.llextra)
                     }
                 };
                 LvalueRef {
                     llval: llprojected,
                     llextra: llextra,
                     ty: projected_ty,
+                    alignment: align,
                 }
             }
         };
@@ -357,6 +419,8 @@ impl<'a, 'tcx> MirContext<'a, 'tcx> {
     // Perform an action using the given Lvalue.
     // If the Lvalue is an empty LocalRef::Operand, then a temporary stack slot
     // is created first, then used as an operand to update the Lvalue.
+    //
+    // FIXME: this is only called from transmute; please remove it.
     pub fn with_lvalue_ref<F, U>(&mut self, bcx: &Builder<'a, 'tcx>,
                                  lvalue: &mir::Lvalue<'tcx>, f: F) -> U
     where F: FnOnce(&mut Self, LvalueRef<'tcx>) -> U
@@ -367,10 +431,9 @@ impl<'a, 'tcx> MirContext<'a, 'tcx> {
                 LocalRef::Operand(None) => {
                     let lvalue_ty = self.monomorphized_lvalue_ty(lvalue);
                     assert!(!lvalue_ty.has_erasable_regions());
-                    let lltemp = bcx.alloca_ty(lvalue_ty, "lvalue_temp");
-                    let lvalue = LvalueRef::new_sized(lltemp, LvalueTy::from_ty(lvalue_ty));
+                    let lvalue = LvalueRef::alloca(bcx, lvalue_ty, "lvalue_temp");
                     let ret = f(self, lvalue);
-                    let op = self.trans_load(bcx, lvalue.llval, lvalue_ty);
+                    let op = self.trans_load(bcx, lvalue.llval, lvalue.alignment, lvalue_ty);
                     self.locals[index] = LocalRef::Operand(Some(op));
                     ret
                 }
@@ -381,7 +444,8 @@ impl<'a, 'tcx> MirContext<'a, 'tcx> {
                     if common::type_is_zero_size(bcx.ccx, ty) {
                         // Pass an undef pointer as no stores can actually occur.
                         let llptr = C_undef(type_of(bcx.ccx, ty).ptr_to());
-                        f(self, LvalueRef::new_sized(llptr, LvalueTy::from_ty(ty)))
+                        f(self, LvalueRef::new_sized(llptr, LvalueTy::from_ty(ty),
+                                                     Alignment::AbiAligned))
                     } else {
                         bug!("Lvalue local already set");
                     }
diff --git a/src/librustc_trans/mir/mod.rs b/src/librustc_trans/mir/mod.rs
index 8a0a97a5634..e017c8f2532 100644
--- a/src/librustc_trans/mir/mod.rs
+++ b/src/librustc_trans/mir/mod.rs
@@ -38,7 +38,7 @@ use rustc_data_structures::indexed_vec::{IndexVec, Idx};
 pub use self::constant::trans_static_initializer;
 
 use self::analyze::CleanupKind;
-use self::lvalue::LvalueRef;
+use self::lvalue::{Alignment, LvalueRef};
 use rustc::mir::traversal;
 
 use self::operand::{OperandRef, OperandValue};
@@ -269,8 +269,7 @@ pub fn trans_mir<'a, 'tcx: 'a>(
 
                 debug!("alloc: {:?} ({}) -> lvalue", local, name);
                 assert!(!ty.has_erasable_regions());
-                let lltemp = bcx.alloca_ty(ty, &name.as_str());
-                let lvalue = LvalueRef::new_sized(lltemp, LvalueTy::from_ty(ty));
+                let lvalue = LvalueRef::alloca(&bcx, ty, &name.as_str());
                 if dbg {
                     let (scope, span) = mircx.debug_loc(source_info);
                     declare_local(&bcx, &mircx.debug_context, name, ty, scope,
@@ -283,12 +282,12 @@ pub fn trans_mir<'a, 'tcx: 'a>(
                 if local == mir::RETURN_POINTER && mircx.fn_ty.ret.is_indirect() {
                     debug!("alloc: {:?} (return pointer) -> lvalue", local);
                     let llretptr = llvm::get_param(llfn, 0);
-                    LocalRef::Lvalue(LvalueRef::new_sized(llretptr, LvalueTy::from_ty(ty)))
+                    LocalRef::Lvalue(LvalueRef::new_sized(llretptr, LvalueTy::from_ty(ty),
+                                                          Alignment::AbiAligned))
                 } else if lvalue_locals.contains(local.index()) {
                     debug!("alloc: {:?} -> lvalue", local);
                     assert!(!ty.has_erasable_regions());
-                    let lltemp = bcx.alloca_ty(ty, &format!("{:?}", local));
-                    LocalRef::Lvalue(LvalueRef::new_sized(lltemp, LvalueTy::from_ty(ty)))
+                    LocalRef::Lvalue(LvalueRef::alloca(&bcx, ty,  &format!("{:?}", local)))
                 } else {
                     // If this is an immediate local, we do not create an
                     // alloca in advance. Instead we wait until we see the
@@ -388,9 +387,9 @@ fn arg_local_refs<'a, 'tcx>(bcx: &Builder<'a, 'tcx>,
                 _ => bug!("spread argument isn't a tuple?!")
             };
 
-            let lltemp = bcx.alloca_ty(arg_ty, &format!("arg{}", arg_index));
+            let lvalue = LvalueRef::alloca(bcx, arg_ty, &format!("arg{}", arg_index));
             for (i, &tupled_arg_ty) in tupled_arg_tys.iter().enumerate() {
-                let dst = bcx.struct_gep(lltemp, i);
+                let dst = bcx.struct_gep(lvalue.llval, i);
                 let arg = &mircx.fn_ty.args[idx];
                 idx += 1;
                 if common::type_is_fat_ptr(bcx.ccx, tupled_arg_ty) {
@@ -409,7 +408,7 @@ fn arg_local_refs<'a, 'tcx>(bcx: &Builder<'a, 'tcx>,
             // we can create one debuginfo entry for the argument.
             arg_scope.map(|scope| {
                 let variable_access = VariableAccess::DirectVariable {
-                    alloca: lltemp
+                    alloca: lvalue.llval
                 };
                 declare_local(
                     bcx,
@@ -422,7 +421,7 @@ fn arg_local_refs<'a, 'tcx>(bcx: &Builder<'a, 'tcx>,
                 );
             });
 
-            return LocalRef::Lvalue(LvalueRef::new_sized(lltemp, LvalueTy::from_ty(arg_ty)));
+            return LocalRef::Lvalue(lvalue);
         }
 
         let arg = &mircx.fn_ty.args[idx];
@@ -469,21 +468,21 @@ fn arg_local_refs<'a, 'tcx>(bcx: &Builder<'a, 'tcx>,
             };
             return LocalRef::Operand(Some(operand.unpack_if_pair(bcx)));
         } else {
-            let lltemp = bcx.alloca_ty(arg_ty, &format!("arg{}", arg_index));
+            let lltemp = LvalueRef::alloca(bcx, arg_ty, &format!("arg{}", arg_index));
             if common::type_is_fat_ptr(bcx.ccx, arg_ty) {
                 // we pass fat pointers as two words, but we want to
                 // represent them internally as a pointer to two words,
                 // so make an alloca to store them in.
                 let meta = &mircx.fn_ty.args[idx];
                 idx += 1;
-                arg.store_fn_arg(bcx, &mut llarg_idx, base::get_dataptr(bcx, lltemp));
-                meta.store_fn_arg(bcx, &mut llarg_idx, base::get_meta(bcx, lltemp));
+                arg.store_fn_arg(bcx, &mut llarg_idx, base::get_dataptr(bcx, lltemp.llval));
+                meta.store_fn_arg(bcx, &mut llarg_idx, base::get_meta(bcx, lltemp.llval));
             } else  {
                 // otherwise, arg is passed by value, so make a
                 // temporary and store it there
-                arg.store_fn_arg(bcx, &mut llarg_idx, lltemp);
+                arg.store_fn_arg(bcx, &mut llarg_idx, lltemp.llval);
             }
-            lltemp
+            lltemp.llval
         };
         arg_scope.map(|scope| {
             // Is this a regular argument?
@@ -573,7 +572,8 @@ fn arg_local_refs<'a, 'tcx>(bcx: &Builder<'a, 'tcx>,
                 );
             }
         });
-        LocalRef::Lvalue(LvalueRef::new_sized(llval, LvalueTy::from_ty(arg_ty)))
+        LocalRef::Lvalue(LvalueRef::new_sized(llval, LvalueTy::from_ty(arg_ty),
+                                              Alignment::AbiAligned))
     }).collect()
 }
 
diff --git a/src/librustc_trans/mir/operand.rs b/src/librustc_trans/mir/operand.rs
index 28a247ee612..cb77fcbbff8 100644
--- a/src/librustc_trans/mir/operand.rs
+++ b/src/librustc_trans/mir/operand.rs
@@ -10,6 +10,7 @@
 
 use llvm::ValueRef;
 use rustc::ty::Ty;
+use rustc::ty::layout::Layout;
 use rustc::mir;
 use rustc_data_structures::indexed_vec::Idx;
 
@@ -23,6 +24,7 @@ use type_::Type;
 use std::fmt;
 
 use super::{MirContext, LocalRef};
+use super::lvalue::Alignment;
 
 /// The representation of a Rust value. The enum variant is in fact
 /// uniquely determined by the value's type, but is kept as a
@@ -31,7 +33,7 @@ use super::{MirContext, LocalRef};
 pub enum OperandValue {
     /// A reference to the actual operand. The data is guaranteed
     /// to be valid for the operand's lifetime.
-    Ref(ValueRef),
+    Ref(ValueRef, Alignment),
     /// A single LLVM value.
     Immediate(ValueRef),
     /// A pair of immediate LLVM values. Used by fat pointers too.
@@ -58,9 +60,9 @@ pub struct OperandRef<'tcx> {
 impl<'tcx> fmt::Debug for OperandRef<'tcx> {
     fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
         match self.val {
-            OperandValue::Ref(r) => {
-                write!(f, "OperandRef(Ref({:?}) @ {:?})",
-                       Value(r), self.ty)
+            OperandValue::Ref(r, align) => {
+                write!(f, "OperandRef(Ref({:?}, {:?}) @ {:?})",
+                       Value(r), align, self.ty)
             }
             OperandValue::Immediate(i) => {
                 write!(f, "OperandRef(Immediate({:?}) @ {:?})",
@@ -137,27 +139,33 @@ impl<'a, 'tcx> MirContext<'a, 'tcx> {
     pub fn trans_load(&mut self,
                       bcx: &Builder<'a, 'tcx>,
                       llval: ValueRef,
+                      align: Alignment,
                       ty: Ty<'tcx>)
                       -> OperandRef<'tcx>
     {
         debug!("trans_load: {:?} @ {:?}", Value(llval), ty);
 
         let val = if common::type_is_fat_ptr(bcx.ccx, ty) {
-            let (lldata, llextra) = base::load_fat_ptr(bcx, llval, ty);
+            let (lldata, llextra) = base::load_fat_ptr(bcx, llval, align, ty);
             OperandValue::Pair(lldata, llextra)
         } else if common::type_is_imm_pair(bcx.ccx, ty) {
+            let f_align = match *bcx.ccx.layout_of(ty) {
+                Layout::Univariant { ref variant, .. } =>
+                    Alignment::from_packed(variant.packed) | align,
+                _ => align
+            };
             let [a_ty, b_ty] = common::type_pair_fields(bcx.ccx, ty).unwrap();
             let a_ptr = bcx.struct_gep(llval, 0);
             let b_ptr = bcx.struct_gep(llval, 1);
 
             OperandValue::Pair(
-                base::load_ty(bcx, a_ptr, a_ty),
-                base::load_ty(bcx, b_ptr, b_ty)
+                base::load_ty(bcx, a_ptr, f_align, a_ty),
+                base::load_ty(bcx, b_ptr, f_align, b_ty)
             )
         } else if common::type_is_immediate(bcx.ccx, ty) {
-            OperandValue::Immediate(base::load_ty(bcx, llval, ty))
+            OperandValue::Immediate(base::load_ty(bcx, llval, align, ty))
         } else {
-            OperandValue::Ref(llval)
+            OperandValue::Ref(llval, align)
         };
 
         OperandRef { val: val, ty: ty }
@@ -212,7 +220,7 @@ impl<'a, 'tcx> MirContext<'a, 'tcx> {
         // out from their home
         let tr_lvalue = self.trans_lvalue(bcx, lvalue);
         let ty = tr_lvalue.ty.to_ty(bcx.tcx());
-        self.trans_load(bcx, tr_lvalue.llval, ty)
+        self.trans_load(bcx, tr_lvalue.llval, tr_lvalue.alignment, ty)
     }
 
     pub fn trans_operand(&mut self,
@@ -230,9 +238,9 @@ impl<'a, 'tcx> MirContext<'a, 'tcx> {
             mir::Operand::Constant(ref constant) => {
                 let val = self.trans_constant(bcx, constant);
                 let operand = val.to_operand(bcx.ccx);
-                if let OperandValue::Ref(ptr) = operand.val {
+                if let OperandValue::Ref(ptr, align) = operand.val {
                     // If this is a OperandValue::Ref to an immediate constant, load it.
-                    self.trans_load(bcx, ptr, operand.ty)
+                    self.trans_load(bcx, ptr, align, operand.ty)
                 } else {
                     operand
                 }
@@ -243,8 +251,8 @@ impl<'a, 'tcx> MirContext<'a, 'tcx> {
     pub fn store_operand(&mut self,
                          bcx: &Builder<'a, 'tcx>,
                          lldest: ValueRef,
-                         operand: OperandRef<'tcx>,
-                         align: Option<u32>) {
+                         align: Option<u32>,
+                         operand: OperandRef<'tcx>) {
         debug!("store_operand: operand={:?}, align={:?}", operand, align);
         // Avoid generating stores of zero-sized values, because the only way to have a zero-sized
         // value is through `undef`, and store itself is useless.
@@ -252,7 +260,10 @@ impl<'a, 'tcx> MirContext<'a, 'tcx> {
             return;
         }
         match operand.val {
-            OperandValue::Ref(r) => base::memcpy_ty(bcx, lldest, r, operand.ty, align),
+            OperandValue::Ref(r, Alignment::Packed) =>
+                base::memcpy_ty(bcx, lldest, r, operand.ty, Some(1)),
+            OperandValue::Ref(r, Alignment::AbiAligned) =>
+                base::memcpy_ty(bcx, lldest, r, operand.ty, align),
             OperandValue::Immediate(s) => {
                 bcx.store(base::from_immediate(bcx, s), lldest, align);
             }
diff --git a/src/librustc_trans/mir/rvalue.rs b/src/librustc_trans/mir/rvalue.rs
index 81b241b4851..be74e30ec74 100644
--- a/src/librustc_trans/mir/rvalue.rs
+++ b/src/librustc_trans/mir/rvalue.rs
@@ -33,7 +33,7 @@ use Disr;
 use super::MirContext;
 use super::constant::const_scalar_checked_binop;
 use super::operand::{OperandRef, OperandValue};
-use super::lvalue::{LvalueRef};
+use super::lvalue::LvalueRef;
 
 impl<'a, 'tcx> MirContext<'a, 'tcx> {
     pub fn trans_rvalue(&mut self,
@@ -50,7 +50,7 @@ impl<'a, 'tcx> MirContext<'a, 'tcx> {
                let tr_operand = self.trans_operand(&bcx, operand);
                // FIXME: consider not copying constants through stack. (fixable by translating
                // constants into OperandValue::Ref, why don’t we do that yet if we don’t?)
-               self.store_operand(&bcx, dest.llval, tr_operand, None);
+               self.store_operand(&bcx, dest.llval, dest.alignment.to_align(), tr_operand);
                bcx
            }
 
@@ -61,7 +61,7 @@ impl<'a, 'tcx> MirContext<'a, 'tcx> {
                     // into-coerce of a thin pointer to a fat pointer - just
                     // use the operand path.
                     let (bcx, temp) = self.trans_rvalue_operand(bcx, rvalue);
-                    self.store_operand(&bcx, dest.llval, temp, None);
+                    self.store_operand(&bcx, dest.llval, dest.alignment.to_align(), temp);
                     return bcx;
                 }
 
@@ -81,13 +81,15 @@ impl<'a, 'tcx> MirContext<'a, 'tcx> {
                         // index into the struct, and this case isn't
                         // important enough for it.
                         debug!("trans_rvalue: creating ugly alloca");
-                        let lltemp = bcx.alloca_ty(operand.ty, "__unsize_temp");
-                        base::store_ty(&bcx, llval, lltemp, operand.ty);
-                        lltemp
+                        let scratch = LvalueRef::alloca(&bcx, operand.ty, "__unsize_temp");
+                        base::store_ty(&bcx, llval, scratch.llval, scratch.alignment, operand.ty);
+                        scratch
+                    }
+                    OperandValue::Ref(llref, align) => {
+                        LvalueRef::new_sized_ty(llref, operand.ty, align)
                     }
-                    OperandValue::Ref(llref) => llref
                 };
-                base::coerce_unsized_into(&bcx, llref, operand.ty, dest.llval, cast_ty);
+                base::coerce_unsized_into(&bcx, &llref, &dest);
                 bcx
             }
 
@@ -97,7 +99,7 @@ impl<'a, 'tcx> MirContext<'a, 'tcx> {
                 let size = C_uint(bcx.ccx, size);
                 let base = base::get_dataptr(&bcx, dest.llval);
                 tvec::slice_for_each(&bcx, base, tr_elem.ty, size, |bcx, llslot| {
-                    self.store_operand(bcx, llslot, tr_elem, None);
+                    self.store_operand(bcx, llslot, dest.alignment.to_align(), tr_elem);
                 })
             }
 
@@ -111,15 +113,15 @@ impl<'a, 'tcx> MirContext<'a, 'tcx> {
                             let op = self.trans_operand(&bcx, operand);
                             // Do not generate stores and GEPis for zero-sized fields.
                             if !common::type_is_zero_size(bcx.ccx, op.ty) {
-                                let mut val = LvalueRef::new_sized(dest.llval, dest.ty);
+                                let mut val = LvalueRef::new_sized(dest.llval, dest.ty, dest.alignment);
                                 let field_index = active_field_index.unwrap_or(i);
                                 val.ty = LvalueTy::Downcast {
                                     adt_def: adt_def,
                                     substs: self.monomorphize(&substs),
                                     variant_index: disr.0 as usize,
                                 };
-                                let lldest_i = val.trans_field_ptr(&bcx, field_index);
-                                self.store_operand(&bcx, lldest_i, op, None);
+                                let (lldest_i, align) = val.trans_field_ptr(&bcx, field_index);
+                                self.store_operand(&bcx, lldest_i, align.to_align(), op);
                             }
                         }
                     },
@@ -131,6 +133,7 @@ impl<'a, 'tcx> MirContext<'a, 'tcx> {
                         } else {
                             None
                         };
+                        let alignment = dest.alignment;
                         for (i, operand) in operands.iter().enumerate() {
                             let op = self.trans_operand(&bcx, operand);
                             // Do not generate stores and GEPis for zero-sized fields.
@@ -144,7 +147,7 @@ impl<'a, 'tcx> MirContext<'a, 'tcx> {
                                     i
                                 };
                                 let dest = bcx.gepi(dest.llval, &[0, i]);
-                                self.store_operand(&bcx, dest, op, None);
+                                self.store_operand(&bcx, dest, alignment.to_align(), op);
                             }
                         }
                     }
@@ -169,7 +172,7 @@ impl<'a, 'tcx> MirContext<'a, 'tcx> {
             _ => {
                 assert!(rvalue_creates_operand(rvalue));
                 let (bcx, temp) = self.trans_rvalue_operand(bcx, rvalue);
-                self.store_operand(&bcx, dest.llval, temp, None);
+                self.store_operand(&bcx, dest.llval, dest.alignment.to_align(), temp);
                 bcx
             }
         }
@@ -228,7 +231,7 @@ impl<'a, 'tcx> MirContext<'a, 'tcx> {
                                     operand.ty, cast_ty);
                                 OperandValue::Pair(lldata, llextra)
                             }
-                            OperandValue::Ref(_) => {
+                            OperandValue::Ref(..) => {
                                 bug!("by-ref operand {:?} in trans_rvalue_operand",
                                      operand);
                             }
diff --git a/src/test/codegen/packed.rs b/src/test/codegen/packed.rs
new file mode 100644
index 00000000000..db2cd3b4165
--- /dev/null
+++ b/src/test/codegen/packed.rs
@@ -0,0 +1,29 @@
+// Copyright 2016 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: -C no-prepopulate-passes
+
+#![crate_type = "lib"]
+
+#[repr(packed)]
+pub struct Packed {
+    dealign: u8,
+    data: u32
+}
+
+// CHECK-LABEL: @write_pkd
+#[no_mangle]
+pub fn write_pkd(pkd: &mut Packed) -> u32 {
+// CHECK: %{{.*}} = load i32, i32* %{{.*}}, align 1
+// CHECK: store i32 42, i32* %{{.*}}, align 1
+    let result = pkd.data;
+    pkd.data = 42;
+    result
+}