Implement llvm.fma.v* intrinsics

cc rust-lang#1405

Author: bjorn3

src/intrinsics/llvm.rs

@@ -51,6 +51,21 @@ pub(crate) fn codegen_llvm_intrinsic_call<'tcx>(
             });
         }
 
+        _ if intrinsic.starts_with("llvm.fma.v") => {
+            intrinsic_args!(fx, args => (x,y,z); intrinsic);
+
+            simd_trio_for_each_lane(
+                fx,
+                x,
+                y,
+                z,
+                ret,
+                &|fx, _lane_ty, _res_lane_ty, lane_x, lane_y, lane_z| {
+                    fx.bcx.ins().fma(lane_x, lane_y, lane_z)
+                },
+            );
+        }
+
         _ => {
             fx.tcx
                 .sess

src/intrinsics/mod.rs

@@ -132,6 +132,35 @@ fn simd_pair_for_each_lane<'tcx>(
     }
 }
 
+fn simd_trio_for_each_lane<'tcx>(
+    fx: &mut FunctionCx<'_, '_, 'tcx>,
+    x: CValue<'tcx>,
+    y: CValue<'tcx>,
+    z: CValue<'tcx>,
+    ret: CPlace<'tcx>,
+    f: &dyn Fn(&mut FunctionCx<'_, '_, 'tcx>, Ty<'tcx>, Ty<'tcx>, Value, Value, Value) -> Value,
+) {
+    assert_eq!(x.layout(), y.layout());
+    let layout = x.layout();
+
+    let (lane_count, lane_ty) = layout.ty.simd_size_and_type(fx.tcx);
+    let lane_layout = fx.layout_of(lane_ty);
+    let (ret_lane_count, ret_lane_ty) = ret.layout().ty.simd_size_and_type(fx.tcx);
+    let ret_lane_layout = fx.layout_of(ret_lane_ty);
+    assert_eq!(lane_count, ret_lane_count);
+
+    for lane_idx in 0..lane_count {
+        let x_lane = x.value_lane(fx, lane_idx).load_scalar(fx);
+        let y_lane = y.value_lane(fx, lane_idx).load_scalar(fx);
+        let z_lane = z.value_lane(fx, lane_idx).load_scalar(fx);
+
+        let res_lane = f(fx, lane_layout.ty, ret_lane_layout.ty, x_lane, y_lane, z_lane);
+        let res_lane = CValue::by_val(res_lane, ret_lane_layout);
+
+        ret.place_lane(fx, lane_idx).write_cvalue(fx, res_lane);
+    }
+}
+
 fn simd_reduce<'tcx>(
     fx: &mut FunctionCx<'_, '_, 'tcx>,
     val: CValue<'tcx>,