The following, from the documentation as of the time of writing, illustrates
the problem:
```zig
// Compile time coercion of float to int
test "implicit cast to comptime_int" {
var f: f32 = 54.0 / 5;
_ = f;
}
```
It is not clear how to unify the types of 54.0 and 5 to perform the
division. We can either
- cast 54.0 to comptime_int resulting in @as(comptime_int, 10), which is
casted to @as(f32, 10), or
- cast 5 to comptime_float resulting in @as(comptime_float, 10.8), which
is casted to @as(f32, 10.8)
Since the two resulting values are different, a compiler error is appropriate.
If we know that casting to either type will result in the same value we
don't need to error. For instance, 10.0 / 2 is okay, as is 10 / 2.0.
Fixes: #12364
The `Value.eql` function has to test for value equality *as-if* the lhs
value parameter is coerced into the type of the rhs. For tagged unions,
there was a problematic case when the lhs was an anonymous struct,
because in such case the value is empty_struct_value and the type
contains all the value information. But the only type available in the
function was the rhs type.
So the fix involved making `Value.eqlAdvanced` also accept the lhs type,
and then enhancing the logic to handle the case of the `.anon_struct` tag.
closes#12418
Tests run locally:
* test-behavior
* test-cases
When lowering a struct type to an LLVM struct type, keep track of
whether there are any underaligned fields. If so, then make it a packed
llvm struct. This works because we already insert manual padding bytes
regardless.
We could unconditionally use an LLVM packed struct; the reason we bother
checking for underaligned fields is that it is a conservative choice, in
case LLVM handles packed structs less optimally. A future improvement
could simplify this code by unconditionally using packed LLVM structs
and then make sure measure perf is unaffected.
closes#12190
`validateExternType` does not require the type to be resolved so we can
check it earlier. Only doing it in `resolveTypeFully` lead to worse or
missing compile errors.
It did not handle properly when the dummy operand was a comptime_int; it
was crashing in coerce because comptime_int is supposed to be
comptime-known. So when calling coerceResultPtr, we pass the actual
operand, not a dummy operand, which means it will have the proper
comptime value when necessary.
Previously, the logic for analyzing coerce_result_ptr would generate
invalid bitcast instructions which did not include coercion logic, such
as optional wrapping, resulting in miscompilations.
Now, the logic of resolve_inferred_alloc goes back over all the
placeholders inserted by coerce_result_ptr, and replaces them with logic
doing the proper coercions.
Closes#12045
For calling convention ABI purposes, integer attributes and return
values need to have an LLVM attribute signext or zeroext added
sometimes. This commit implements that logic.
It also implements a proof-of-concept of moving the F16T type from
being a compiler_rt hack to being how the compiler lowers f16 in
functions that need to match certain calling conventions.
Closes#12054
Previously, struct types, alignment values, and initialization
expressions were all lowered into the same ZIR body, which caused false
positive "depends on itself" errors when the initialization expression
depended on the size of the struct.
This also uses ResultLoc.coerced_ty for struct field alignment and
initialization values. The resulting ZIR encoding ends up being roughly
the same, neither smaller nor larger than previously.
Closes#12029
