auto merge of #15998 : luqmana/rust/nmnnbd, r=thestinger

LLVM recently added a new attribute, dereferenceable: http://reviews.llvm.org/D4449

>This patch adds a dereferencable attribute. In some sense, this is a companion to the nonnull attribute, but specifies that the pointer is known to be dereferencable in the same sense as a pointer generated by alloca is known to be dereferencable.

With rust, everywhere that we previously marked `nonnull` we can actually mark as `dereferenceable` (which implies nonnull) since we know the size. That is, except for one case: when generating calls for TyVisitor. It seems like we haven't substituted the self type (so we have `ty_param`) and just treat it as an opaque pointer so I just left that bit as nonnull.

With this, LLVM can for example hoist a load out of a loop where it previously couldn't:

```Rust
pub fn baz(c: &uint, n: uint) -> uint {
    let mut res = 0;
    for i in range(0, n) {
        if i > 0 {
            res += *c * i;
        }
    }
    res
}
```

Before:
```llvm
define i64 @baz(i64* noalias nocapture nonnull readonly, i64) unnamed_addr #0 {
entry-block:
  br label %for_loopback.outer

for_loopback.outer:                               ; preds = %then-block-33-, %entry-block
  %.ph = phi i64 [ %.lcssa, %then-block-33- ], [ 0, %entry-block ]
  %res.0.ph = phi i64 [ %8, %then-block-33- ], [ 0, %entry-block ]
  br label %for_loopback

for_exit:                                         ; preds = %for_loopback
  %res.0.ph.lcssa = phi i64 [ %res.0.ph, %for_loopback ]
  ret i64 %res.0.ph.lcssa

for_loopback:                                     ; preds = %for_loopback.outer, %for_body
  %2 = phi i64 [ %4, %for_body ], [ %.ph, %for_loopback.outer ]
  %3 = icmp ult i64 %2, %1
  br i1 %3, label %for_body, label %for_exit

for_body:                                         ; preds = %for_loopback
  %4 = add i64 %2, 1
  %5 = icmp eq i64 %2, 0
  br i1 %5, label %for_loopback, label %then-block-33-

then-block-33-:                                   ; preds = %for_body
  %.lcssa = phi i64 [ %4, %for_body ]
  %.lcssa15 = phi i64 [ %2, %for_body ]
  %6 = load i64* %0, align 8                     ; <------- this load
  %7 = mul i64 %6, %.lcssa15
  %8 = add i64 %7, %res.0.ph
  br label %for_loopback.outer
}
```

After:
```llvm
define i64 @baz(i64* noalias nocapture readonly dereferenceable(8), i64) unnamed_addr #0 {
entry-block:
  %2 = load i64* %0, align 8                    ; <------- load once instead
  br label %for_loopback.outer

for_loopback.outer:                               ; preds = %then-block-33-, %entry-block
  %.ph = phi i64 [ %.lcssa, %then-block-33- ], [ 0, %entry-block ]
  %res.0.ph = phi i64 [ %8, %then-block-33- ], [ 0, %entry-block ]
  br label %for_loopback

for_exit:                                         ; preds = %for_loopback
  %res.0.ph.lcssa = phi i64 [ %res.0.ph, %for_loopback ]
  ret i64 %res.0.ph.lcssa

for_loopback:                                     ; preds = %for_loopback.outer, %for_body
  %3 = phi i64 [ %5, %for_body ], [ %.ph, %for_loopback.outer ]
  %4 = icmp ult i64 %3, %1
  br i1 %4, label %for_body, label %for_exit

for_body:                                         ; preds = %for_loopback
  %5 = add i64 %3, 1
  %6 = icmp eq i64 %3, 0
  br i1 %6, label %for_loopback, label %then-block-33-

then-block-33-:                                   ; preds = %for_body
  %.lcssa = phi i64 [ %5, %for_body ]
  %.lcssa15 = phi i64 [ %3, %for_body ]
  %7 = mul i64 %2, %.lcssa15
  %8 = add i64 %7, %res.0.ph
  br label %for_loopback.outer
}
```

Author: bors

src/librustc/middle/trans/base.rs

@@ -281,11 +281,7 @@ pub fn decl_rust_fn(ccx: &CrateContext, fn_ty: ty::t, name: &str) -> ValueRef {
 
     let llfn = decl_fn(ccx, name, llvm::CCallConv, llfty, output);
     let attrs = get_fn_llvm_attributes(ccx, fn_ty);
-    for &(idx, attr) in attrs.iter() {
-        unsafe {
-            llvm::LLVMAddFunctionAttribute(llfn, idx as c_uint, attr);
-        }
-    }
+    attrs.apply_llfn(llfn);
 
     llfn
 }
@@ -962,7 +958,7 @@ pub fn invoke<'a>(
                               llargs.as_slice(),
                               normal_bcx.llbb,
                               landing_pad,
-                              attributes.as_slice());
+                              Some(attributes));
         return (llresult, normal_bcx);
     } else {
         debug!("calling {} at {}", llfn, bcx.llbb);
@@ -975,7 +971,7 @@ pub fn invoke<'a>(
             None => debuginfo::clear_source_location(bcx.fcx)
         };
 
-        let llresult = Call(bcx, llfn, llargs.as_slice(), attributes.as_slice());
+        let llresult = Call(bcx, llfn, llargs.as_slice(), Some(attributes));
         return (llresult, bcx);
     }
 }
@@ -1081,7 +1077,7 @@ pub fn call_lifetime_start(cx: &Block, ptr: ValueRef) {
     let llsize = C_u64(ccx, machine::llsize_of_alloc(ccx, val_ty(ptr).element_type()));
     let ptr = PointerCast(cx, ptr, Type::i8p(ccx));
     let lifetime_start = ccx.get_intrinsic(&"llvm.lifetime.start");
-    Call(cx, lifetime_start, [llsize, ptr], []);
+    Call(cx, lifetime_start, [llsize, ptr], None);
 }
 
 pub fn call_lifetime_end(cx: &Block, ptr: ValueRef) {
@@ -1095,7 +1091,7 @@ pub fn call_lifetime_end(cx: &Block, ptr: ValueRef) {
     let llsize = C_u64(ccx, machine::llsize_of_alloc(ccx, val_ty(ptr).element_type()));
     let ptr = PointerCast(cx, ptr, Type::i8p(ccx));
     let lifetime_end = ccx.get_intrinsic(&"llvm.lifetime.end");
-    Call(cx, lifetime_end, [llsize, ptr], []);
+    Call(cx, lifetime_end, [llsize, ptr], None);
 }
 
 pub fn call_memcpy(cx: &Block, dst: ValueRef, src: ValueRef, n_bytes: ValueRef, align: u32) {
@@ -1111,7 +1107,7 @@ pub fn call_memcpy(cx: &Block, dst: ValueRef, src: ValueRef, n_bytes: ValueRef,
     let size = IntCast(cx, n_bytes, ccx.int_type);
     let align = C_i32(ccx, align as i32);
     let volatile = C_bool(ccx, false);
-    Call(cx, memcpy, [dst_ptr, src_ptr, size, align, volatile], []);
+    Call(cx, memcpy, [dst_ptr, src_ptr, size, align, volatile], None);
 }
 
 pub fn memcpy_ty(bcx: &Block, dst: ValueRef, src: ValueRef, t: ty::t) {
@@ -1156,7 +1152,7 @@ fn memzero(b: &Builder, llptr: ValueRef, ty: Type) {
     let size = machine::llsize_of(ccx, ty);
     let align = C_i32(ccx, llalign_of_min(ccx, ty) as i32);
     let volatile = C_bool(ccx, false);
-    b.call(llintrinsicfn, [llptr, llzeroval, size, align, volatile], []);
+    b.call(llintrinsicfn, [llptr, llzeroval, size, align, volatile], None);
 }
 
 pub fn alloc_ty(bcx: &Block, t: ty::t, name: &str) -> ValueRef {
@@ -2040,7 +2036,7 @@ fn register_fn(ccx: &CrateContext,
 }
 
 pub fn get_fn_llvm_attributes(ccx: &CrateContext, fn_ty: ty::t)
-                              -> Vec<(uint, u64)> {
+                              -> llvm::AttrBuilder {
     use middle::ty::{BrAnon, ReLateBound};
 
     let (fn_sig, abi, has_env) = match ty::get(fn_ty).sty {
@@ -2056,31 +2052,33 @@ pub fn get_fn_llvm_attributes(ccx: &CrateContext, fn_ty: ty::t)
         _ => fail!("expected closure or function.")
     };
 
+    // Since index 0 is the return value of the llvm func, we start
+    // at either 1 or 2 depending on whether there's an env slot or not
+    let mut first_arg_offset = if has_env { 2 } else { 1 };
+    let mut attrs = llvm::AttrBuilder::new();
+    let ret_ty = fn_sig.output;
+
     // These have an odd calling convention, so we skip them for now.
     //
     // FIXME(pcwalton): We don't have to skip them; just untuple the result.
     if abi == RustCall {
-        return Vec::new()
+        return attrs;
     }
 
-    // Since index 0 is the return value of the llvm func, we start
-    // at either 1 or 2 depending on whether there's an env slot or not
-    let mut first_arg_offset = if has_env { 2 } else { 1 };
-    let mut attrs = Vec::new();
-    let ret_ty = fn_sig.output;
-
     // A function pointer is called without the declaration
     // available, so we have to apply any attributes with ABI
     // implications directly to the call instruction. Right now,
     // the only attribute we need to worry about is `sret`.
     if type_of::return_uses_outptr(ccx, ret_ty) {
-        attrs.push((1, llvm::StructRetAttribute as u64));
+        let llret_sz = llsize_of_real(ccx, type_of::type_of(ccx, ret_ty));
 
         // The outptr can be noalias and nocapture because it's entirely
-        // invisible to the program. We can also mark it as nonnull
-        attrs.push((1, llvm::NoAliasAttribute as u64));
-        attrs.push((1, llvm::NoCaptureAttribute as u64));
-        attrs.push((1, llvm::NonNullAttribute as u64));
+        // invisible to the program. We also know it's nonnull as well
+        // as how many bytes we can dereference
+        attrs.arg(1, llvm::StructRetAttribute)
+             .arg(1, llvm::NoAliasAttribute)
+             .arg(1, llvm::NoCaptureAttribute)
+             .arg(1, llvm::DereferenceableAttribute(llret_sz));
 
         // Add one more since there's an outptr
         first_arg_offset += 1;
@@ -2094,27 +2092,28 @@ pub fn get_fn_llvm_attributes(ccx: &CrateContext, fn_ty: ty::t)
                 ty::ty_str | ty::ty_vec(..) | ty::ty_trait(..) => true, _ => false
             } => {}
             ty::ty_uniq(_) => {
-                attrs.push((llvm::ReturnIndex as uint, llvm::NoAliasAttribute as u64));
+                attrs.ret(llvm::NoAliasAttribute);
             }
             _ => {}
         }
 
-        // We can also mark the return value as `nonnull` in certain cases
+        // We can also mark the return value as `dereferenceable` in certain cases
         match ty::get(ret_ty).sty {
             // These are not really pointers but pairs, (pointer, len)
             ty::ty_uniq(it) |
             ty::ty_rptr(_, ty::mt { ty: it, .. }) if match ty::get(it).sty {
                 ty::ty_str | ty::ty_vec(..) | ty::ty_trait(..) => true, _ => false
             } => {}
-            ty::ty_uniq(_) | ty::ty_rptr(_, _) => {
-                attrs.push((llvm::ReturnIndex as uint, llvm::NonNullAttribute as u64));
+            ty::ty_uniq(inner) | ty::ty_rptr(_, ty::mt { ty: inner, .. }) => {
+                let llret_sz = llsize_of_real(ccx, type_of::type_of(ccx, inner));
+                attrs.ret(llvm::DereferenceableAttribute(llret_sz));
             }
             _ => {}
         }
 
         match ty::get(ret_ty).sty {
             ty::ty_bool => {
-                attrs.push((llvm::ReturnIndex as uint, llvm::ZExtAttribute as u64));
+                attrs.ret(llvm::ZExtAttribute);
             }
             _ => {}
         }
@@ -2124,44 +2123,77 @@ pub fn get_fn_llvm_attributes(ccx: &CrateContext, fn_ty: ty::t)
         match ty::get(t).sty {
             // this needs to be first to prevent fat pointers from falling through
             _ if !type_is_immediate(ccx, t) => {
+                let llarg_sz = llsize_of_real(ccx, type_of::type_of(ccx, t));
+
                 // For non-immediate arguments the callee gets its own copy of
                 // the value on the stack, so there are no aliases. It's also
                 // program-invisible so can't possibly capture
-                attrs.push((idx, llvm::NoAliasAttribute as u64));
-                attrs.push((idx, llvm::NoCaptureAttribute as u64));
-                attrs.push((idx, llvm::NonNullAttribute as u64));
+                attrs.arg(idx, llvm::NoAliasAttribute)
+                     .arg(idx, llvm::NoCaptureAttribute)
+                     .arg(idx, llvm::DereferenceableAttribute(llarg_sz));
             }
+
             ty::ty_bool => {
-                attrs.push((idx, llvm::ZExtAttribute as u64));
+                attrs.arg(idx, llvm::ZExtAttribute);
             }
+
             // `~` pointer parameters never alias because ownership is transferred
-            ty::ty_uniq(_) => {
-                attrs.push((idx, llvm::NoAliasAttribute as u64));
-                attrs.push((idx, llvm::NonNullAttribute as u64));
+            ty::ty_uniq(inner) => {
+                let llsz = llsize_of_real(ccx, type_of::type_of(ccx, inner));
+
+                attrs.arg(idx, llvm::NoAliasAttribute)
+                     .arg(idx, llvm::DereferenceableAttribute(llsz));
+            }
+
+            // The visit glue deals only with opaque pointers so we don't
+            // actually know the concrete type of Self thus we don't know how
+            // many bytes to mark as dereferenceable so instead we just mark
+            // it as nonnull which still holds true
+            ty::ty_rptr(b, ty::mt { ty: it, mutbl }) if match ty::get(it).sty {
+                ty::ty_param(_) => true, _ => false
+            } && mutbl == ast::MutMutable => {
+                attrs.arg(idx, llvm::NoAliasAttribute)
+                     .arg(idx, llvm::NonNullAttribute);
+
+                match b {
+                    ReLateBound(_, BrAnon(_)) => {
+                        attrs.arg(idx, llvm::NoCaptureAttribute);
+                    }
+                    _ => {}
+                }
             }
+
             // `&mut` pointer parameters never alias other parameters, or mutable global data
             // `&` pointer parameters never alias either (for LLVM's purposes) as long as the
             // interior is safe
             ty::ty_rptr(b, mt) if mt.mutbl == ast::MutMutable ||
                                   !ty::type_contents(ccx.tcx(), mt.ty).interior_unsafe() => {
-                attrs.push((idx, llvm::NoAliasAttribute as u64));
-                attrs.push((idx, llvm::NonNullAttribute as u64));
+
+                let llsz = llsize_of_real(ccx, type_of::type_of(ccx, mt.ty));
+                attrs.arg(idx, llvm::NoAliasAttribute)
+                     .arg(idx, llvm::DereferenceableAttribute(llsz));
+
                 match b {
                     ReLateBound(_, BrAnon(_)) => {
-                        attrs.push((idx, llvm::NoCaptureAttribute as u64));
+                        attrs.arg(idx, llvm::NoCaptureAttribute);
                     }
                     _ => {}
                 }
             }
+
             // When a reference in an argument has no named lifetime, it's impossible for that
             // reference to escape this function (returned or stored beyond the call by a closure).
-            ty::ty_rptr(ReLateBound(_, BrAnon(_)), _) => {
-                attrs.push((idx, llvm::NoCaptureAttribute as u64));
-                attrs.push((idx, llvm::NonNullAttribute as u64));
+            ty::ty_rptr(ReLateBound(_, BrAnon(_)), mt) => {
+                let llsz = llsize_of_real(ccx, type_of::type_of(ccx, mt.ty));
+                attrs.arg(idx, llvm::NoCaptureAttribute)
+                     .arg(idx, llvm::DereferenceableAttribute(llsz));
             }
-            // & pointer parameters are never null
-            ty::ty_rptr(_, _) => {
-                attrs.push((idx, llvm::NonNullAttribute as u64));
+
+            // & pointer parameters are also never null and we know exactly how
+            // many bytes we can dereference
+            ty::ty_rptr(_, mt) => {
+                let llsz = llsize_of_real(ccx, type_of::type_of(ccx, mt.ty));
+                attrs.arg(idx, llvm::DereferenceableAttribute(llsz));
             }
             _ => ()
         }

src/librustc/middle/trans/build.rs

@@ -12,7 +12,7 @@
 #![allow(non_snake_case_functions)]
 
 use llvm;
-use llvm::{CallConv, AtomicBinOp, AtomicOrdering, AsmDialect};
+use llvm::{CallConv, AtomicBinOp, AtomicOrdering, AsmDialect, AttrBuilder};
 use llvm::{Opcode, IntPredicate, RealPredicate};
 use llvm::{ValueRef, BasicBlockRef};
 use middle::trans::common::*;
@@ -113,7 +113,7 @@ pub fn Invoke(cx: &Block,
               args: &[ValueRef],
               then: BasicBlockRef,
               catch: BasicBlockRef,
-              attributes: &[(uint, u64)])
+              attributes: Option<AttrBuilder>)
               -> ValueRef {
     if cx.unreachable.get() {
         return C_null(Type::i8(cx.ccx()));
@@ -681,13 +681,13 @@ pub fn InlineAsmCall(cx: &Block, asm: *const c_char, cons: *const c_char,
 }
 
 pub fn Call(cx: &Block, fn_: ValueRef, args: &[ValueRef],
-            attributes: &[(uint, u64)]) -> ValueRef {
+            attributes: Option<AttrBuilder>) -> ValueRef {
     if cx.unreachable.get() { return _UndefReturn(cx, fn_); }
     B(cx).call(fn_, args, attributes)
 }
 
 pub fn CallWithConv(cx: &Block, fn_: ValueRef, args: &[ValueRef], conv: CallConv,
-                    attributes: &[(uint, u64)]) -> ValueRef {
+                    attributes: Option<AttrBuilder>) -> ValueRef {
     if cx.unreachable.get() { return _UndefReturn(cx, fn_); }
     B(cx).call_with_conv(fn_, args, conv, attributes)
 }

src/librustc/middle/trans/builder.rs

@@ -11,7 +11,7 @@
 #![allow(dead_code)] // FFI wrappers
 
 use llvm;
-use llvm::{CallConv, AtomicBinOp, AtomicOrdering, AsmDialect};
+use llvm::{CallConv, AtomicBinOp, AtomicOrdering, AsmDialect, AttrBuilder};
 use llvm::{Opcode, IntPredicate, RealPredicate, False};
 use llvm::{ValueRef, BasicBlockRef, BuilderRef, ModuleRef};
 use middle::trans::base;
@@ -155,7 +155,7 @@ impl<'a> Builder<'a> {
                   args: &[ValueRef],
                   then: BasicBlockRef,
                   catch: BasicBlockRef,
-                  attributes: &[(uint, u64)])
+                  attributes: Option<AttrBuilder>)
                   -> ValueRef {
         self.count_insn("invoke");
 
@@ -174,8 +174,9 @@ impl<'a> Builder<'a> {
                                           then,
                                           catch,
                                           noname());
-            for &(idx, attr) in attributes.iter() {
-                llvm::LLVMAddCallSiteAttribute(v, idx as c_uint, attr);
+            match attributes {
+                Some(a) => a.apply_callsite(v),
+                None => {}
             }
             v
         }
@@ -777,7 +778,7 @@ impl<'a> Builder<'a> {
                                              c, noname(), False, False)
                 }
             });
-            self.call(asm, [], []);
+            self.call(asm, [], None);
         }
     }
 
@@ -802,12 +803,12 @@ impl<'a> Builder<'a> {
         unsafe {
             let v = llvm::LLVMInlineAsm(
                 fty.to_ref(), asm, cons, volatile, alignstack, dia as c_uint);
-            self.call(v, inputs, [])
+            self.call(v, inputs, None)
         }
     }
 
     pub fn call(&self, llfn: ValueRef, args: &[ValueRef],
-                attributes: &[(uint, u64)]) -> ValueRef {
+                attributes: Option<AttrBuilder>) -> ValueRef {
         self.count_insn("call");
 
         debug!("Call {} with args ({})",
@@ -820,15 +821,16 @@ impl<'a> Builder<'a> {
         unsafe {
             let v = llvm::LLVMBuildCall(self.llbuilder, llfn, args.as_ptr(),
                                         args.len() as c_uint, noname());
-            for &(idx, attr) in attributes.iter() {
-                llvm::LLVMAddCallSiteAttribute(v, idx as c_uint, attr);
+            match attributes {
+                Some(a) => a.apply_callsite(v),
+                None => {}
             }
             v
         }
     }
 
     pub fn call_with_conv(&self, llfn: ValueRef, args: &[ValueRef],
-                          conv: CallConv, attributes: &[(uint, u64)]) -> ValueRef {
+                          conv: CallConv, attributes: Option<AttrBuilder>) -> ValueRef {
         self.count_insn("callwithconv");
         let v = self.call(llfn, args, attributes);
         llvm::SetInstructionCallConv(v, conv);

src/librustc/middle/trans/closure.rs

@@ -588,7 +588,7 @@ pub fn get_wrapper_for_bare_fn(ccx: &CrateContext,
     }
     llargs.extend(args.iter().map(|arg| arg.val));
 
-    let retval = Call(bcx, fn_ptr, llargs.as_slice(), []);
+    let retval = Call(bcx, fn_ptr, llargs.as_slice(), None);
     if type_is_zero_size(ccx, f.sig.output) || fcx.llretptr.get().is_some() {
         RetVoid(bcx);
     } else {

src/librustc/middle/trans/expr.rs

@@ -553,7 +553,7 @@ fn trans_index<'a>(bcx: &'a Block<'a>,
             let expected = Call(bcx,
                                 expect,
                                 [bounds_check, C_bool(ccx, false)],
-                                []);
+                                None);
             bcx = with_cond(bcx, expected, |bcx| {
                 controlflow::trans_fail_bounds_check(bcx,
                                                      index_expr.span,