Auto merge of #135258 - oli-obk:push-ktzskvxuwnlt, r=<try>

Use llvm.memset.p0i8.* to initialize all same-bytes arrays

It doesn't affect tests, LLVM seems smart enough for it, but then I wonder why we have the zero case at all (it was introduced in #43488, maybe LLVM wasn't smart enough then). So let's run perf to see if there's any build time effect, and if no, I'll remove the zero special case and also run perf.

Author: bors

compiler/rustc_codegen_ssa/src/mir/rvalue.rs

@@ -93,23 +93,37 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
                     return;
                 }
 
-                if let OperandValue::Immediate(v) = cg_elem.val {
+                let try_init_all_same = |bx: &mut Bx, v| {
                     let start = dest.val.llval;
                     let size = bx.const_usize(dest.layout.size.bytes());
 
-                    // Use llvm.memset.p0i8.* to initialize all zero arrays
-                    if bx.cx().const_to_opt_u128(v, false) == Some(0) {
-                        let fill = bx.cx().const_u8(0);
-                        bx.memset(start, fill, size, dest.val.align, MemFlags::empty());
-                        return;
+                    // Use llvm.memset.p0i8.* to initialize all same byte arrays
+                    if let Some(int) = bx.cx().const_to_opt_u128(v, false) {
+                        let bytes = &int.to_le_bytes()[..cg_elem.layout.size.bytes_usize()];
+                        let first = bytes[0];
+                        if bytes[1..].iter().all(|&b| b == first) {
+                            let fill = bx.cx().const_u8(first);
+                            bx.memset(start, fill, size, dest.val.align, MemFlags::empty());
+                            return true;
+                        }
                     }
 
                     // Use llvm.memset.p0i8.* to initialize byte arrays
                     let v = bx.from_immediate(v);
                     if bx.cx().val_ty(v) == bx.cx().type_i8() {
                         bx.memset(start, v, size, dest.val.align, MemFlags::empty());
-                        return;
+                        return true;
+                    }
+                    false
+                };
+
+                match cg_elem.val {
+                    OperandValue::Immediate(v) => {
+                        if try_init_all_same(bx, v) {
+                            return;
+                        }
                     }
+                    _ => (),
                 }
 
                 let count = self

tests/codegen/slice-init.rs

@@ -63,6 +63,34 @@ pub fn nonzero_integer_array() {
     opaque(&x);
 }
 
+const N: usize = 100;
+
+// FIXME: The two bytes of the u16 are the same, so we should
+// just use memset, too.
+// CHECK-LABEL: @u16_init_one_bytes
+#[no_mangle]
+pub fn u16_init_one_bytes() -> [u16; N] {
+    // CHECK-NOT: select
+    // CHECK-NOT: br
+    // CHECK-NOT: switch
+    // CHECK-NOT: icmp
+    // CHECK: call void @llvm.memset.p0
+    [const { u16::from_be_bytes([1, 1]) }; N]
+}
+
+// FIXME: undef bytes can just be initialized with the same value as the
+// defined bytes, if the defines bytes are all the same.
+// CHECK-LABEL: @option_none_init
+#[no_mangle]
+pub fn option_none_init() -> [Option<u8>; N] {
+    // CHECK-NOT: select
+    // CHECK: br label %repeat_loop_header{{.*}}
+    // CHECK-NOT: switch
+    // CHECK: icmp
+    // CHECK-NOT: call void @llvm.memset.p0
+    [None; N]
+}
+
 // Use an opaque function to prevent rustc from removing useless drops.
 #[inline(never)]
 pub fn opaque(_: impl Sized) {}