Fix a few AMDGPU related issues

src/librustc_codegen_llvm/attributes.rs

@@ -16,6 +16,7 @@ use rustc::hir::def_id::{DefId, LOCAL_CRATE};
 use rustc::session::Session;
 use rustc::session::config::Sanitizer;
 use rustc::ty::TyCtxt;
+use rustc::ty::layout::HasTyCtxt;
 use rustc::ty::query::Providers;
 use rustc_data_structures::sync::Lrc;
 use rustc_data_structures::fx::FxHashMap;
@@ -32,12 +33,16 @@ use value::Value;
 
 /// Mark LLVM function to use provided inline heuristic.
 #[inline]
-pub fn inline(val: &'ll Value, inline: InlineAttr) {
+pub fn inline(cx: &CodegenCx<'ll, '_>, val: &'ll Value, inline: InlineAttr) {
     use self::InlineAttr::*;
     match inline {
         Hint   => Attribute::InlineHint.apply_llfn(Function, val),
         Always => Attribute::AlwaysInline.apply_llfn(Function, val),
-        Never  => Attribute::NoInline.apply_llfn(Function, val),
+        Never  => {
+            if cx.tcx().sess.target.target.arch != "amdgpu" {
+                Attribute::NoInline.apply_llfn(Function, val);
+            }
+        },
         None   => {
             Attribute::InlineHint.unapply_llfn(Function, val);
             Attribute::AlwaysInline.unapply_llfn(Function, val);
@@ -142,7 +147,7 @@ pub fn from_fn_attrs(
     let codegen_fn_attrs = id.map(|id| cx.tcx.codegen_fn_attrs(id))
         .unwrap_or(CodegenFnAttrs::new());
 
-    inline(llfn, codegen_fn_attrs.inline);
+    inline(cx, llfn, codegen_fn_attrs.inline);
 
     // The `uwtable` attribute according to LLVM is:
     //

src/librustc_codegen_llvm/builder.rs

@@ -496,6 +496,14 @@ impl Builder<'a, 'll, 'tcx> {
 
 
     pub fn range_metadata(&self, load: &'ll Value, range: Range<u128>) {
+        if self.sess().target.target.arch == "amdgpu" {
+            // amdgpu/LLVM does something weird and thinks a i64 value is
+            // split into a v2i32, halving the bitwidth LLVM expects,
+            // tripping an assertion. So, for now, just disable this
+            // optimization.
+            return;
+        }
+
         unsafe {
             let llty = val_ty(load);
             let v = [

src/librustc_codegen_llvm/callee.rs

@@ -96,7 +96,7 @@ pub fn get_fn(
         debug!("get_fn: not casting pointer!");
 
         if instance.def.is_inline(tcx) {
-            attributes::inline(llfn, attributes::InlineAttr::Hint);
+            attributes::inline(cx, llfn, attributes::InlineAttr::Hint);
         }
         attributes::from_fn_attrs(cx, llfn, Some(instance.def.def_id()));
 

src/librustc_codegen_llvm/mono_item.rs

@@ -180,7 +180,7 @@ fn predefine_fn<'a, 'tcx>(cx: &CodegenCx<'a, 'tcx>,
 
     debug!("predefine_fn: mono_ty = {:?} instance = {:?}", mono_ty, instance);
     if instance.def.is_inline(cx.tcx) {
-        attributes::inline(lldecl, attributes::InlineAttr::Hint);
+        attributes::inline(cx, lldecl, attributes::InlineAttr::Hint);
     }
     attributes::from_fn_attrs(cx, lldecl, Some(instance.def.def_id()));
 

src/librustc_llvm/build.rs

@@ -81,7 +81,7 @@ fn main() {
     let is_crossed = target != host;
 
     let mut optional_components =
-        vec!["x86", "arm", "aarch64", "mips", "powerpc",
+        vec!["x86", "arm", "aarch64", "amdgpu", "mips", "powerpc",
              "systemz", "jsbackend", "webassembly", "msp430", "sparc", "nvptx"];
 
     let mut version_cmd = Command::new(&llvm_config);

src/librustc_target/abi/call/amdgpu.rs

@@ -0,0 +1,42 @@
+// Copyright 2018 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.
+
+use abi::call::{ArgType, FnType, };
+use abi::{HasDataLayout, LayoutOf, TyLayout, TyLayoutMethods};
+
+fn classify_ret_ty<'a, Ty, C>(_tuncx: C, ret: &mut ArgType<'a, Ty>)
+  where Ty: TyLayoutMethods<'a, C> + Copy,
+        C: LayoutOf<Ty = Ty, TyLayout = TyLayout<'a, Ty>> + HasDataLayout
+{
+  ret.extend_integer_width_to(32);
+}
+
+fn classify_arg_ty<'a, Ty, C>(_cx: C, arg: &mut ArgType<'a, Ty>)
+  where Ty: TyLayoutMethods<'a, C> + Copy,
+        C: LayoutOf<Ty = Ty, TyLayout = TyLayout<'a, Ty>> + HasDataLayout
+{
+  arg.extend_integer_width_to(32);
+}
+
+pub fn compute_abi_info<'a, Ty, C>(cx: C, fty: &mut FnType<'a, Ty>)
+  where Ty: TyLayoutMethods<'a, C> + Copy,
+        C: LayoutOf<Ty = Ty, TyLayout = TyLayout<'a, Ty>> + HasDataLayout
+{
+  if !fty.ret.is_ignore() {
+    classify_ret_ty(cx, &mut fty.ret);
+  }
+
+  for arg in &mut fty.args {
+    if arg.is_ignore() {
+      continue;
+    }
+    classify_arg_ty(cx, arg);
+  }
+}

src/librustc_target/abi/call/mod.rs

@@ -13,6 +13,7 @@ use abi::{HasDataLayout, LayoutOf, TyLayout, TyLayoutMethods};
 use spec::HasTargetSpec;
 
 mod aarch64;
+mod amdgpu;
 mod arm;
 mod asmjs;
 mod hexagon;
@@ -503,6 +504,7 @@ impl<'a, Ty> FnType<'a, Ty> {
                 x86_64::compute_abi_info(cx, self);
             },
             "aarch64" => aarch64::compute_abi_info(cx, self),
+            "amdgpu" => amdgpu::compute_abi_info(cx, self),
             "arm" => arm::compute_abi_info(cx, self),
             "mips" => mips::compute_abi_info(cx, self),
             "mips64" => mips64::compute_abi_info(cx, self),