Stage2 miscompiles call to std.math.big.int.Managed.shiftLeft

[Vexu]

test {
    var p = try std.math.big.int.Managed.initSet(testing.allocator, 1);
    defer p.deinit();
    var a2 = try p.clone();
    defer a2.deinit();

    var shift: isize = 54;
    try a2.shiftLeft(a2, @intCast(usize, shift)); // segfault

    // var shift2 = @intCast(usize, shift);
    // try a2.shiftLeft(a2, shift2); // works as expected
}

The biggest difference between the two seems to be that the in the second one the a2 passed by value is copied first but the doc comments on shiftLeft says that aliasing here should be okay:

diff --git a/foo.ll b/foo.ll
index 7368ad2c8..29f41a7d7 100644
--- a/foo.ll
+++ b/foo.ll
@@ -1,7 +1,7 @@
 ; Function Attrs: nounwind
 define internal fastcc i16 @a.test_0(%builtin.StackTrace* nonnull %0) unnamed_addr #0 {
 Entry:
-  %1 = alloca i64, align 8
+  %1 = alloca %math.big.int.Managed, align 8
   %2 = alloca i64, align 8
   %3 = alloca { %math.big.int.Managed, i16 }, align 8
   %4 = alloca %math.big.int.Managed, align 8
@@ -46,9 +46,12 @@ TryCont2:                                         ; preds = %TryCont
   call void @llvm.dbg.declare(metadata %math.big.int.Managed* %4, metadata !4856, metadata !DIExpression()),
   store i64 54, i64* %2, align 8,
   call void @llvm.dbg.declare(metadata i64* %2, metadata !4857, metadata !DIExpression()),
-  %23 = load i64, i64* %2, align 8,
-  %24 = icmp sge i64 %23, 0,
-  br i1 %24, label %Then, label %Else,
+  %23 = bitcast %math.big.int.Managed* %1 to i8*,
+  %24 = bitcast %math.big.int.Managed* %4 to i8*,
+  call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 %23, i8* align 8 %24, i64 40, i1 false),
+  %25 = load i64, i64* %2, align 8,
+  %26 = icmp sge i64 %25, 0,
+  br i1 %26, label %Then, label %Else,
 
 Then:                                             ; preds = %TryCont2
   br label %Block,
@@ -58,18 +61,15 @@ Else:                                             ; preds = %TryCont2
   unreachable,
 
 Block:                                            ; preds = %Then
-  store i64 %23, i64* %1, align 8,
-  call void @llvm.dbg.declare(metadata i64* %1, metadata !4859, metadata !DIExpression()),
-  %25 = load i64, i64* %1, align 8,
-  %26 = call fastcc i16 @math.big.int.Managed.shiftLeft(%builtin.StackTrace* %0, %math.big.int.Managed* %4, %math.big.int.Managed* %4, i64 %25),
-  %27 = icmp ne i16 %26, 0,
-  br i1 %27, label %TryRet3, label %TryCont4,
+  %27 = call fastcc i16 @math.big.int.Managed.shiftLeft(%builtin.StackTrace* %0, %math.big.int.Managed* %4, %math.big.int.Managed* %1, i64 %25),
+  %28 = icmp ne i16 %27, 0,
+  br i1 %28, label %TryRet3, label %TryCont4,
 
 TryRet3:                                          ; preds = %Block
   call fastcc void @math.big.int.Managed.deinit(%math.big.int.Managed* %4),

This is blocking std.math.big.Rational.toFloat tests.