When targeting AArch64, unable to widen cast f16 to f128

[andrewrk]

export fn entry() bool {
    var x: f16 = 12.34;
    var y: f128 = x;
}

LLVM ERROR: Unsupported library call operation!

Note that it works if you do:

export fn entry() void {
    var x: f16 = 12.34;
    var y: f32 = x;
    var z: f128 = y;
}

This issue is to:

• Work with upstream LLVM to get this working
• In the mean time, work around the issue by generating f16 -> f128 casts by taking a brief detour through f32.

[shawnl]

I don't get it. This is not a difficult function. The widening float conversion you just move the bits and zero the rest.

[LemonBoy]

There's a patch for this that's stuck in the limbo.

[mikdusan]

test is now passing for both aarch64 and powerpc64le

$ /opt/llvm-linux-x86_64-12.0.0.1-release-assert/bin/llvm-config --assertion-mode
ON
$ zig test test/behavior.zig -lc -Itest -target aarch64-linux -mcpu=baseline+v8_5a --test-cmd /opt/qemu-linux-x86_64-5.2.0/bin/qemu-aarch64 --test-cmd-bin --test-filter "float widening f16 to f128"
All 1 tests passed.
$ zig test test/behavior.zig -lc -Itest -target powerpc64le-linux --test-cmd /opt/qemu-linux-x86_64-5.2.0/bin/qemu-ppc64le --test-cmd-bin --test-filter "float widening f16 to f128"
All 1 tests passed.

however, this related test in doc is failing:

zig/doc/langref.html.in

Lines 5199 to 5209 in f693247

	test "float widening" {
	// Note: there is an open issue preventing this from working on aarch64:
	// https://github.com/ziglang/zig/issues/3282
	if (std.Target.current.cpu.arch == .aarch64) return error.SkipZigTest;
	var a: f16 = 12.34;
	var b: f32 = a;
	var c: f64 = b;
	var d: f128 = c;
	try expect(d == a);
	}