const vector passed to codegen

7 files changed, 170 insertions, 17 deletions

diff --git a/compiler/rustc_codegen_gcc/src/common.rs b/compiler/rustc_codegen_gcc/src/common.rs
index 7a456e1c5d6..dca6b6494f9 100644
--- a/compiler/rustc_codegen_gcc/src/common.rs
+++ b/compiler/rustc_codegen_gcc/src/common.rs
@@ -160,6 +160,11 @@ impl<'gcc, 'tcx> ConstMethods<'tcx> for CodegenCx<'gcc, 'tcx> {
         self.context.new_struct_constructor(None, struct_type.as_type(), None, values)
     }
 
+    fn const_vector(&self, values: &[RValue<'gcc>]) -> RValue<'gcc> {
+        let typ = self.type_vector(values[0].get_type(), values.len() as u64);
+        self.context.new_rvalue_from_vector(None, typ, values)
+    }
+
     fn const_to_opt_uint(&self, _v: RValue<'gcc>) -> Option<u64> {
         // TODO(antoyo)
         None
diff --git a/compiler/rustc_codegen_llvm/src/common.rs b/compiler/rustc_codegen_llvm/src/common.rs
index 197bbb9ddbf..283518ed2bc 100644
--- a/compiler/rustc_codegen_llvm/src/common.rs
+++ b/compiler/rustc_codegen_llvm/src/common.rs
@@ -100,11 +100,6 @@ impl<'ll> CodegenCx<'ll, '_> {
         unsafe { llvm::LLVMConstArray2(ty, elts.as_ptr(), len) }
     }
 
-    pub fn const_vector(&self, elts: &[&'ll Value]) -> &'ll Value {
-        let len = c_uint::try_from(elts.len()).expect("LLVMConstVector elements len overflow");
-        unsafe { llvm::LLVMConstVector(elts.as_ptr(), len) }
-    }
-
     pub fn const_bytes(&self, bytes: &[u8]) -> &'ll Value {
         bytes_in_context(self.llcx, bytes)
     }
@@ -224,6 +219,11 @@ impl<'ll, 'tcx> ConstMethods<'tcx> for CodegenCx<'ll, 'tcx> {
         struct_in_context(self.llcx, elts, packed)
     }
 
+    fn const_vector(&self, elts: &[&'ll Value]) -> &'ll Value {
+        let len = c_uint::try_from(elts.len()).expect("LLVMConstVector elements len overflow");
+        unsafe { llvm::LLVMConstVector(elts.as_ptr(), len) }
+    }
+
     fn const_to_opt_uint(&self, v: &'ll Value) -> Option<u64> {
         try_as_const_integral(v).and_then(|v| unsafe {
             let mut i = 0u64;
diff --git a/compiler/rustc_codegen_ssa/src/mir/block.rs b/compiler/rustc_codegen_ssa/src/mir/block.rs
index bc3076528da..772adf13ff1 100644
--- a/compiler/rustc_codegen_ssa/src/mir/block.rs
+++ b/compiler/rustc_codegen_ssa/src/mir/block.rs
@@ -923,8 +923,12 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
                         // third argument must be constant. This is
                         // checked by the type-checker.
                         if i == 2 && intrinsic.name == sym::simd_shuffle {
+                            // FIXME: the simd_shuffle argument is actually an array,
+                            // not a vector, so we need this special hack to make sure
+                            // it is passed as an immediate. We should pass the
+                            // shuffle indices as a vector instead to avoid this hack.
                             if let mir::Operand::Constant(constant) = &arg.node {
-                                let (llval, ty) = self.simd_shuffle_indices(bx, constant);
+                                let (llval, ty) = self.immediate_const_vector(bx, constant);
                                 return OperandRef {
                                     val: Immediate(llval),
                                     layout: bx.layout_of(ty),
diff --git a/compiler/rustc_codegen_ssa/src/mir/constant.rs b/compiler/rustc_codegen_ssa/src/mir/constant.rs
index f4d974f7036..0aa85b82038 100644
--- a/compiler/rustc_codegen_ssa/src/mir/constant.rs
+++ b/compiler/rustc_codegen_ssa/src/mir/constant.rs
@@ -1,6 +1,6 @@
 use rustc_middle::mir::interpret::ErrorHandled;
 use rustc_middle::ty::layout::HasTyCtxt;
-use rustc_middle::ty::{self, Ty};
+use rustc_middle::ty::{self, Ty, ValTree};
 use rustc_middle::{bug, mir, span_bug};
 use rustc_target::abi::Abi;
 
@@ -28,7 +28,7 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
             .expect("erroneous constant missed by mono item collection")
     }
 
-    /// This is a convenience helper for `simd_shuffle_indices`. It has the precondition
+    /// This is a convenience helper for `immediate_const_vector`. It has the precondition
     /// that the given `constant` is an `Const::Unevaluated` and must be convertible to
     /// a `ValTree`. If you want a more general version of this, talk to `wg-const-eval` on zulip.
     ///
@@ -59,23 +59,42 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
         self.cx.tcx().const_eval_resolve_for_typeck(ty::ParamEnv::reveal_all(), uv, constant.span)
     }
 
-    /// process constant containing SIMD shuffle indices
-    pub fn simd_shuffle_indices(
+    /// process constant containing SIMD shuffle indices & constant vectors
+    pub fn immediate_const_vector(
         &mut self,
         bx: &Bx,
         constant: &mir::ConstOperand<'tcx>,
     ) -> (Bx::Value, Ty<'tcx>) {
         let ty = self.monomorphize(constant.ty());
+        let ty_is_simd = ty.is_simd();
+        // FIXME: ideally we'd assert that this is a SIMD type, but simd_shuffle
+        // in its current form relies on a regular array being passed as an
+        // immediate argument. This hack can be removed once that is fixed.
+        let field_ty = if ty_is_simd {
+            ty.simd_size_and_type(bx.tcx()).1
+        } else {
+            ty.builtin_index().unwrap()
+        };
+
         let val = self
             .eval_unevaluated_mir_constant_to_valtree(constant)
             .ok()
             .map(|x| x.ok())
             .flatten()
             .map(|val| {
-                let field_ty = ty.builtin_index().unwrap();
-                let values: Vec<_> = val
-                    .unwrap_branch()
-                    .iter()
+                // Depending on whether this is a SIMD type with an array field
+                // or a type with many fields (one for each elements), the valtree
+                // is either a single branch with N children, or a root node
+                // with exactly one child which then in turn has many children.
+                // So we look at the first child to determine whether it is a
+                // leaf or whether we have to go one more layer down.
+                let branch_or_leaf = val.unwrap_branch();
+                let first = branch_or_leaf.get(0).unwrap();
+                let field_iter = match first {
+                    ValTree::Branch(_) => first.unwrap_branch().iter(),
+                    ValTree::Leaf(_) => branch_or_leaf.iter(),
+                };
+                let values: Vec<_> = field_iter
                     .map(|field| {
                         if let Some(prim) = field.try_to_scalar() {
                             let layout = bx.layout_of(field_ty);
@@ -84,11 +103,11 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
                             };
                             bx.scalar_to_backend(prim, scalar, bx.immediate_backend_type(layout))
                         } else {
-                            bug!("simd shuffle field {:?}", field)
+                            bug!("field is not a scalar {:?}", field)
                         }
                     })
                     .collect();
-                bx.const_struct(&values, false)
+                if ty_is_simd { bx.const_vector(&values) } else { bx.const_struct(&values, false) }
             })
             .unwrap_or_else(|| {
                 bx.tcx().dcx().emit_err(errors::ShuffleIndicesEvaluation { span: constant.span });
diff --git a/compiler/rustc_codegen_ssa/src/mir/operand.rs b/compiler/rustc_codegen_ssa/src/mir/operand.rs
index 2bc2d0f70bf..6979478fda1 100644
--- a/compiler/rustc_codegen_ssa/src/mir/operand.rs
+++ b/compiler/rustc_codegen_ssa/src/mir/operand.rs
@@ -635,7 +635,24 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
                 self.codegen_consume(bx, place.as_ref())
             }
 
-            mir::Operand::Constant(ref constant) => self.eval_mir_constant_to_operand(bx, constant),
+            mir::Operand::Constant(ref constant) => {
+                let constant_ty = self.monomorphize(constant.ty());
+                // Most SIMD vector constants should be passed as immediates.
+                // (In particular, some intrinsics really rely on this.)
+                if constant_ty.is_simd() {
+                    // However, some SIMD types do not actually use the vector ABI
+                    // (in particular, packed SIMD types do not). Ensure we exclude those.
+                    let layout = bx.layout_of(constant_ty);
+                    if let Abi::Vector { .. } = layout.abi {
+                        let (llval, ty) = self.immediate_const_vector(bx, constant);
+                        return OperandRef {
+                            val: OperandValue::Immediate(llval),
+                            layout: bx.layout_of(ty),
+                        };
+                    }
+                }
+                self.eval_mir_constant_to_operand(bx, constant)
+            }
         }
     }
 }
diff --git a/compiler/rustc_codegen_ssa/src/traits/consts.rs b/compiler/rustc_codegen_ssa/src/traits/consts.rs
index d93b3e06ca6..c15f1fa8e56 100644
--- a/compiler/rustc_codegen_ssa/src/traits/consts.rs
+++ b/compiler/rustc_codegen_ssa/src/traits/consts.rs
@@ -30,6 +30,7 @@ pub trait ConstMethods<'tcx>: BackendTypes {
 
     fn const_str(&self, s: &str) -> (Self::Value, Self::Value);
     fn const_struct(&self, elts: &[Self::Value], packed: bool) -> Self::Value;
+    fn const_vector(&self, elts: &[Self::Value]) -> Self::Value;
 
     fn const_to_opt_uint(&self, v: Self::Value) -> Option<u64>;
     fn const_to_opt_u128(&self, v: Self::Value, sign_ext: bool) -> Option<u128>;
diff --git a/tests/codegen/const-vector.rs b/tests/codegen/const-vector.rs
new file mode 100644
index 00000000000..d368838201e
--- /dev/null
+++ b/tests/codegen/const-vector.rs
@@ -0,0 +1,107 @@
+//@ compile-flags: -C no-prepopulate-passes -Copt-level=0
+
+// This test checks that constants of SIMD type are passed as immediate vectors.
+// We ensure that both vector representations (struct with fields and struct wrapping array) work.
+#![crate_type = "lib"]
+#![feature(abi_unadjusted)]
+#![feature(const_trait_impl)]
+#![feature(repr_simd)]
+#![feature(rustc_attrs)]
+#![feature(simd_ffi)]
+#![allow(non_camel_case_types)]
+
+// Setting up structs that can be used as const vectors
+#[repr(simd)]
+#[derive(Clone)]
+pub struct i8x2(i8, i8);
+
+#[repr(simd)]
+#[derive(Clone)]
+pub struct i8x2_arr([i8; 2]);
+
+#[repr(simd)]
+#[derive(Clone)]
+pub struct f32x2(f32, f32);
+
+#[repr(simd)]
+#[derive(Clone)]
+pub struct f32x2_arr([f32; 2]);
+
+#[repr(simd, packed)]
+#[derive(Copy, Clone)]
+pub struct Simd<T, const N: usize>([T; N]);
+
+// The following functions are required for the tests to ensure
+// that they are called with a const vector
+
+extern "unadjusted" {
+    #[no_mangle]
+    fn test_i8x2(a: i8x2);
+}
+
+extern "unadjusted" {
+    #[no_mangle]
+    fn test_i8x2_two_args(a: i8x2, b: i8x2);
+}
+
+extern "unadjusted" {
+    #[no_mangle]
+    fn test_i8x2_mixed_args(a: i8x2, c: i32, b: i8x2);
+}
+
+extern "unadjusted" {
+    #[no_mangle]
+    fn test_i8x2_arr(a: i8x2_arr);
+}
+
+extern "unadjusted" {
+    #[no_mangle]
+    fn test_f32x2(a: f32x2);
+}
+
+extern "unadjusted" {
+    #[no_mangle]
+    fn test_f32x2_arr(a: f32x2_arr);
+}
+
+extern "unadjusted" {
+    #[no_mangle]
+    fn test_simd(a: Simd<i32, 4>);
+}
+
+extern "unadjusted" {
+    #[no_mangle]
+    fn test_simd_unaligned(a: Simd<i32, 3>);
+}
+
+// Ensure the packed variant of the simd struct does not become a const vector
+// if the size is not a power of 2
+// CHECK: %"Simd<i32, 3>" = type { [3 x i32] }
+
+pub fn do_call() {
+    unsafe {
+        // CHECK: call void @test_i8x2(<2 x i8> <i8 32, i8 64>
+        test_i8x2(const { i8x2(32, 64) });
+
+        // CHECK: call void @test_i8x2_two_args(<2 x i8> <i8 32, i8 64>, <2 x i8> <i8 8, i8 16>
+        test_i8x2_two_args(const { i8x2(32, 64) }, const { i8x2(8, 16) });
+
+        // CHECK: call void @test_i8x2_mixed_args(<2 x i8> <i8 32, i8 64>, i32 43, <2 x i8> <i8 8, i8 16>
+        test_i8x2_mixed_args(const { i8x2(32, 64) }, 43, const { i8x2(8, 16) });
+
+        // CHECK: call void @test_i8x2_arr(<2 x i8> <i8 32, i8 64>
+        test_i8x2_arr(const { i8x2_arr([32, 64]) });
+
+        // CHECK: call void @test_f32x2(<2 x float> <float 0x3FD47AE140000000, float 0x3FE47AE140000000>
+        test_f32x2(const { f32x2(0.32, 0.64) });
+
+        // CHECK: void @test_f32x2_arr(<2 x float> <float 0x3FD47AE140000000, float 0x3FE47AE140000000>
+        test_f32x2_arr(const { f32x2_arr([0.32, 0.64]) });
+
+        // CHECK: call void @test_simd(<4 x i32> <i32 2, i32 4, i32 6, i32 8>
+        test_simd(const { Simd::<i32, 4>([2, 4, 6, 8]) });
+
+        // CHECK: call void @test_simd_unaligned(%"Simd<i32, 3>" %1
+        test_simd_unaligned(const { Simd::<i32, 3>([2, 4, 6]) });
+    }
+}