broken LLVM module found: musttail call must precede a ret with an optional bitcast

[Sobeston]

Found on the latest master build.

Code:

const expect = @import("std").testing.expect;

fn fibonacciTailInternal(n: u16, a: u16, b: u16) u16 {
    if (n == 0) return a;
    if (n == 1) return b;
    return @call(
        .{ .modifier = .always_tail },
        fibonacciTailInternal,
        .{ n - 1, b, a + b },
    );
}

fn fibonacciTail(n: u16) u16 {
    return fibonacciTailInternal(n, 0, 1);
}

test "forced tail call" {
    try expect(fibonacciTail(10) == 55);
}

Output:

> zig test main.zig
Code Generation [676/921] std.math.log2_int... broken LLVM module found: musttail call must precede a ret with an optional bitcast
  %21 = musttail call fastcc i16 @fibonacciTailInternal(i16 %13, i16 %15, i16 %19), !dbg !3846

This is a bug in the Zig compiler.thread 178562 panic: 
Unable to dump stack trace: debug info stripped
Aborted (core dumped)

[semarie]

in your example, fibonacciTail function generate the following llvm-ir:

; Function Attrs: nobuiltin nounwind sspstrong
define internal fastcc i16 @fibonacciTail(i16 %0) unnamed_addr #1 !dbg !1773 {
Entry:
  %result = alloca i16, align 2
  %n = alloca i16, align 2
  store i16 %0, i16* %n, align 2
  call void @llvm.dbg.declare(metadata i16* %n, metadata !1777, metadata !DIExpression()), !dbg !1778
  %1 = load i16, i16* %n, align 2, !dbg !1779
  %2 = call fastcc i16 @fibonacciTailInternal(i16 %1, i16 0, i16 1), !dbg !1781
  store i16 %2, i16* %result, align 2, !dbg !1781
  %3 = load i16, i16* %result, align 2, !dbg !1782
  ret i16 %3, !dbg !1782
}

zig (stage1) seems to allocate a result variable at beginning of the function, call fibonacciTailInternal, store the result in result, and next load result in %3 to finally return %3.

The same construct is present in fibonacciTailInternal: the store %result + load %result dance before ret is the cause of the failure. LLVM needs ret to follow call to be able to do tail call.

For reference, the generated IR for the last part of the call in fibonacciTailInternal is (when .always_tail isn't used):

  %21 = call fastcc i16 @fibonacciTailInternal(i16 %13, i16 %15, i16 %19), !dbg !2082
  store i16 %21, i16* %result, align 2, !dbg !2082
  %22 = load i16, i16* %result, align 2, !dbg !2083
  ret i16 %22, !dbg !2083