Merge commit 'dbee13661efa269cb4cd57bb4c6b99a19732b484' into sync_cg_clif-2020-12-27

[rust.git] / compiler / rustc_codegen_cranelift / src / intrinsics / simd.rs

diff --git a/compiler/rustc_codegen_cranelift/src/intrinsics/simd.rs b/compiler/rustc_codegen_cranelift/src/intrinsics/simd.rs
index 2e31c4669e25bb16be06813d419695493377566e..e0eb5c59590ffba54aeffe2931d933875cf9f093 100644 (file)
--- a/compiler/rustc_codegen_cranelift/src/intrinsics/simd.rs
+++ b/compiler/rustc_codegen_cranelift/src/intrinsics/simd.rs
@@ -35,30 +35,33 @@ pub(super) fn codegen_simd_intrinsic_call<'tcx>(
             });
         };
 
-        // FIXME support float comparisons
         simd_eq, (c x, c y) {
             validate_simd_type!(fx, intrinsic, span, x.layout().ty);
-            simd_cmp!(fx, Equal(x, y) -> ret);
+            simd_cmp!(fx, Equal|Equal(x, y) -> ret);
         };
         simd_ne, (c x, c y) {
             validate_simd_type!(fx, intrinsic, span, x.layout().ty);
-            simd_cmp!(fx, NotEqual(x, y) -> ret);
+            simd_cmp!(fx, NotEqual|NotEqual(x, y) -> ret);
         };
         simd_lt, (c x, c y) {
             validate_simd_type!(fx, intrinsic, span, x.layout().ty);
-            simd_cmp!(fx, UnsignedLessThan|SignedLessThan(x, y) -> ret);
+            simd_cmp!(fx, UnsignedLessThan|SignedLessThan|LessThan(x, y) -> ret);
         };
         simd_le, (c x, c y) {
             validate_simd_type!(fx, intrinsic, span, x.layout().ty);
-            simd_cmp!(fx, UnsignedLessThanOrEqual|SignedLessThanOrEqual(x, y) -> ret);
+            simd_cmp!(fx, UnsignedLessThanOrEqual|SignedLessThanOrEqual|LessThanOrEqual(x, y) -> ret);
         };
         simd_gt, (c x, c y) {
             validate_simd_type!(fx, intrinsic, span, x.layout().ty);
-            simd_cmp!(fx, UnsignedGreaterThan|SignedGreaterThan(x, y) -> ret);
+            simd_cmp!(fx, UnsignedGreaterThan|SignedGreaterThan|GreaterThan(x, y) -> ret);
         };
         simd_ge, (c x, c y) {
             validate_simd_type!(fx, intrinsic, span, x.layout().ty);
-            simd_cmp!(fx, UnsignedGreaterThanOrEqual|SignedGreaterThanOrEqual(x, y) -> ret);
+            simd_cmp!(
+                fx,
+                UnsignedGreaterThanOrEqual|SignedGreaterThanOrEqual|GreaterThanOrEqual
+                (x, y) -> ret
+            );
         };
 
         // simd_shuffle32<T, U>(x: T, y: T, idx: [u32; 32]) -> U
@@ -70,11 +73,11 @@ pub(super) fn codegen_simd_intrinsic_call<'tcx>(
             assert_eq!(x.layout(), y.layout());
             let layout = x.layout();
 
-            let (lane_type, lane_count) = lane_type_and_count(fx.tcx, layout);
-            let (ret_lane_type, ret_lane_count) = lane_type_and_count(fx.tcx, ret.layout());
+            let (lane_count, lane_ty) = layout.ty.simd_size_and_type(fx.tcx);
+            let (ret_lane_count, ret_lane_ty) = ret.layout().ty.simd_size_and_type(fx.tcx);
 
-            assert_eq!(lane_type, ret_lane_type);
-            assert_eq!(n, ret_lane_count);
+            assert_eq!(lane_ty, ret_lane_ty);
+            assert_eq!(u64::from(n), ret_lane_count);
 
             let total_len = lane_count * 2;
 
@@ -102,14 +105,14 @@ pub(super) fn codegen_simd_intrinsic_call<'tcx>(
             };
 
             for &idx in &indexes {
-                assert!(idx < total_len, "idx {} out of range 0..{}", idx, total_len);
+                assert!(u64::from(idx) < total_len, "idx {} out of range 0..{}", idx, total_len);
             }
 
             for (out_idx, in_idx) in indexes.into_iter().enumerate() {
-                let in_lane = if in_idx < lane_count {
-                    x.value_field(fx, mir::Field::new(in_idx.try_into().unwrap()))
+                let in_lane = if u64::from(in_idx) < lane_count {
+                    x.value_field(fx, mir::Field::new(in_idx.into()))
                 } else {
-                    y.value_field(fx, mir::Field::new((in_idx - lane_count).try_into().unwrap()))
+                    y.value_field(fx, mir::Field::new(usize::from(in_idx) - usize::try_from(lane_count).unwrap()))
                 };
                 let out_lane = ret.place_field(fx, mir::Field::new(out_idx));
                 out_lane.write_cvalue(fx, in_lane);
@@ -128,7 +131,7 @@ pub(super) fn codegen_simd_intrinsic_call<'tcx>(
             };
 
             let idx = idx_const.val.try_to_bits(Size::from_bytes(4 /* u32*/)).unwrap_or_else(|| panic!("kind not scalar: {:?}", idx_const));
-            let (_lane_type, lane_count) = lane_type_and_count(fx.tcx, base.layout());
+            let (lane_count, _lane_ty) = base.layout().ty.simd_size_and_type(fx.tcx);
             if idx >= lane_count.into() {
                 fx.tcx.sess.span_fatal(fx.mir.span, &format!("[simd_insert] idx {} >= lane_count {}", idx, lane_count));
             }
@@ -143,14 +146,21 @@ pub(super) fn codegen_simd_intrinsic_call<'tcx>(
             let idx_const = if let Some(idx_const) = crate::constant::mir_operand_get_const_val(fx, idx) {
                 idx_const
             } else {
-                fx.tcx.sess.span_fatal(
+                fx.tcx.sess.span_warn(
                     span,
                     "Index argument for `simd_extract` is not a constant",
                 );
+                let res = crate::trap::trap_unimplemented_ret_value(
+                    fx,
+                    ret.layout(),
+                    "Index argument for `simd_extract` is not a constant",
+                );
+                ret.write_cvalue(fx, res);
+                return;
             };
 
             let idx = idx_const.val.try_to_bits(Size::from_bytes(4 /* u32*/)).unwrap_or_else(|| panic!("kind not scalar: {:?}", idx_const));
-            let (_lane_type, lane_count) = lane_type_and_count(fx.tcx, v.layout());
+            let (lane_count, _lane_ty) = v.layout().ty.simd_size_and_type(fx.tcx);
             if idx >= lane_count.into() {
                 fx.tcx.sess.span_fatal(fx.mir.span, &format!("[simd_extract] idx {} >= lane_count {}", idx, lane_count));
             }
@@ -202,9 +212,10 @@ pub(super) fn codegen_simd_intrinsic_call<'tcx>(
             assert_eq!(a.layout(), c.layout());
             let layout = a.layout();
 
-            let (_lane_layout, lane_count) = lane_type_and_count(fx.tcx, layout);
-            let (ret_lane_layout, ret_lane_count) = lane_type_and_count(fx.tcx, ret.layout());
+            let (lane_count, _lane_ty) = layout.ty.simd_size_and_type(fx.tcx);
+            let (ret_lane_count, ret_lane_ty) = ret.layout().ty.simd_size_and_type(fx.tcx);
             assert_eq!(lane_count, ret_lane_count);
+            let ret_lane_layout = fx.layout_of(ret_lane_ty);
 
             for lane in 0..lane_count {
                 let lane = mir::Field::new(lane.try_into().unwrap());
@@ -228,11 +239,42 @@ pub(super) fn codegen_simd_intrinsic_call<'tcx>(
             simd_flt_binop!(fx, fmax(x, y) -> ret);
         };
 
+        simd_reduce_add_ordered | simd_reduce_add_unordered, (c v) {
+            validate_simd_type!(fx, intrinsic, span, v.layout().ty);
+            simd_reduce(fx, v, ret, |fx, lane_layout, a, b| {
+                if lane_layout.ty.is_floating_point() {
+                    fx.bcx.ins().fadd(a, b)
+                } else {
+                    fx.bcx.ins().iadd(a, b)
+                }
+            });
+        };
+
+        simd_reduce_mul_ordered | simd_reduce_mul_unordered, (c v) {
+            validate_simd_type!(fx, intrinsic, span, v.layout().ty);
+            simd_reduce(fx, v, ret, |fx, lane_layout, a, b| {
+                if lane_layout.ty.is_floating_point() {
+                    fx.bcx.ins().fmul(a, b)
+                } else {
+                    fx.bcx.ins().imul(a, b)
+                }
+            });
+        };
+
+        simd_reduce_all, (c v) {
+            validate_simd_type!(fx, intrinsic, span, v.layout().ty);
+            simd_reduce_bool(fx, v, ret, |fx, a, b| fx.bcx.ins().band(a, b));
+        };
+
+        simd_reduce_any, (c v) {
+            validate_simd_type!(fx, intrinsic, span, v.layout().ty);
+            simd_reduce_bool(fx, v, ret, |fx, a, b| fx.bcx.ins().bor(a, b));
+        };
+
         // simd_fabs
         // simd_saturating_add
         // simd_bitmask
         // simd_select
-        // simd_reduce_add_{,un}ordered
         // simd_rem
     }
 }