Every other test in tests/issues/run.sh passes -microarch:native
through $COMMON. These two lines spell their flags out, as the test is
a directory and $COMMON has -file, and left it off. On a CI virtual
machine whose CPU is older than the default x86-64-v2, `odin test`
refuses to run without it, which fails the NetBSD job at random.
Calling these failed to compile: `_mm256_round_pd` passed a u8 where
the intrinsic takes a u32, `_mm256_blend_pd` misspelt its parameter,
`_mm256_dp_ps` took its mask as an i8, `_mm_permute_ps/pd` used
SSE/SSE2 helpers without enabling them, and `_mm256_permute2f128_pd`
did not return its result.
The debug compile unit was made from the first file of the initial
package. When build tags exclude all of its files, as for the
Windows-only tests/issues/test_issue_7547.odin on other targets, there
is none, and `odin test -debug` crashed on an out-of-bounds index.
Without `-debug` such a test builds and reports that there are no
tests to run, which `tests/issues/run.sh` relies on.
Fall back to the package's path for the compile unit's file.
A procedure with `@(disabled=true)` adds no dependency where it is
used, so that the calls to it, which are removed, do not pull it in.
Taking its value, e.g. `rawptr(assert)` with `-disable-assert`, then
referenced a procedure that was never generated, and LLVM asserted
"missing procedure 'assert'".
Add the dependency for every use that is not the callee of a call.
Casting a polymorphic procedure to `rawptr`, or to a procedure type
it cannot be specialized to, was accepted. There is no procedure to
take the address of, so the value came out nil, or the build hung, as
with `x := rawptr(avl.init_cmp)`. Assigning one to `rawptr` already
errors; make `cast` and `transmute` match. Casting to a procedure type
that specializes it still works, as assignment does.
Map keys of type `string16` used `default_hasher_string`, which reads
the length as a byte count; for a `string16` it counts u16 units, so
only the first half of the key was hashed. `cstring16` keys took the
same path and were read as a pointer and a length, which crashes.
Add `default_hasher_string16` and `default_hasher_cstring16` to the
runtime and use them for the generated key hashers, which are also
what the runtime and `intrinsics.type_hasher_proc` use.
Slicing a `[dynamic; N]T` skipped the bounds check when both indices
were constant, as is done for fixed arrays. But its length is only
known at runtime, so `a[0:3]` with a length of 2 silently returned
stale elements. Always check against the length, as for `[dynamic]T`.
The checker lets a multi-value call fill several elements of a matrix
or #simd literal, as it does for arrays, but the constant part of the
literal was built as if each element were one value. A matrix literal
like `matrix[2,2]f32{three(), 4}` hit an assertion, and a #simd
literal put the constants after the call in the wrong lanes.
Skip over the values of a call when placing the constants, as the
array case already does.
`_mm_cvtsd_ss` passed two `__m128d` to `llvm.x86.sse2.cvtsd2ss`,
which takes `(<4 x float>, <2 x double>)`, so LLVM failed with
"Intrinsic has incorrect argument type". `_mm_cvtss_sd` used
`llvm.x86.sse2.cvtss2sd`, which LLVM no longer has. Both wrappers
also took the wrong operand types; use Intel's (a: __m128, b: __m128d)
and (a: __m128d, b: __m128), and express `_mm_cvtss_sd` as a lane
conversion as clang does.
`_mm_prefetch` called `llvm.prefetch` without its pointer mangling and
`__rdtscp` declared `llvm.x86.rdtscp` as taking a pointer when it
returns `{i64, i32}`; both were rejected by LLVM too.
LLVM has dropped many x86 intrinsics in favour of generic IR, and
auto-upgrades them only when reading bitcode. Odin declares them
directly, so the calls became undefined external symbols and failed
to link, e.g. `undefined reference to llvm.x86.sse41.pmaxsb`.
Express the affected wrappers the way clang's headers now do: scalar
lane ops, simd.min/max/abs/sqrt, shuffles for the immediate blends,
masked or narrowed multiplies, conversions and an unaligned store.
`_addcarryx_*` now use `llvm.x86.addcarry.*`, as clang does.
`_mm_sqrt_ss` took an unused second argument; it now takes one, as
in Intel's definition. It could not have been used before.
init left the round start untouched. A context that skipped zeroing
with `= ---`, or one reused after TurboSHAKE, kept a stale value, so
SHA-3 and SHAKE ran too few rounds and returned wrong hashes silently.
bbodi