When using llvm opaque pointers, typed pointers and pointer bitcasts are
no longer needed. This also avoids needing packed struct layouts that
are nested inside pointers, letting us avoid computing struct layouts
in Sema that could cause unnecessary dependency loops.
This commit changes the way Zig is intended to deal with variable
declaration for exotic targets. Where previously the idea was to
enfore local/global variables to be placed into their respective
address spaces, depending on the target, this is now fixed to the
generic address space.
To facilitate this for targets where local variables _must_ be
generated into a specific address space (ex. amdgcn where locals
must be generated into the private address space), the variable
allocations (alloca) are generated into the right address space
and then addrspace-casted back to the generic address space. While this
could be less efficient in theory, LLVM will hopefull deal with figuring
out the actual correct address space for a pointer for us. HIP seems to
do the same thing in this regard.
Global variables are handled in a similar way.
These const qualifiers on pointers to opaque types do not serve any
purpose. If anything they are misleading since the underlying pointers
very likely point to objects that are in fact mutated.
This commit does not change any behavior.
This bug manifested as a segfault in stage1 when calling this function.
The C++ code looks like this:
```c++
entry->llvm_di_type = ZigLLVMCreateDebugForwardDeclType(g->dbuilder,
ZigLLVMTag_DW_structure_type(), full_name,
import ? ZigLLVMFileToScope(import->data.structure.root_struct->di_file) : nullptr,
import ? import->data.structure.root_struct->di_file : nullptr,
line);
```
There is actually no problem here - what happened is that because
cross-language LTO was enabled between zig and c++ code, and because
Zig annotated the file parameter (3rd line) as being non-null, the C++
code assumed that parameter could not be null, and eagerly dereferenced
`import->...`, causing a segfault, since it was null.
I verified that this commit fixed the problem and I also verified this
hypothesis by disabling LTO and noticing that it indeed avoided the
problem.
As part of the Opaque Pointers upgrade documentation, LLVM says that the
function LLVMGetGEPSourceElementType() can be used to obtain element
type information in lieu of LLVMGetElementType(), however, this function
actually returns the struct type, not the field type. The GEP
instruction does store the information we need, however, this is not
exposed in the C API. It seems like they accidentally exposed the wrong
field, because one would never need the struct type since one must
already pass it directly to the GEP instruction, so one will always have
it handy, whereas one will usually not have the field type handy.
Previously, the Zig ABI size and LLVM ABI size of these types disagreed
sometimes. This code also corrects the logging messages to not trigger
LLVM assertions.
Rather than lowering float negation as `0.0 - x`.
* Add AIR instruction for float negation.
* Add compiler-rt functions for f128, f80 negation
closes#11853
llvm: dump failed module when -femit-llvm-ir set
print_air:
* print fully qualified name
* use Type.fmt and Value.fmtValue, fmtDebug is useless
TypedValue
* handle anon structs and tuples
* fix bugs
The previous implementation of calling conventions was hacky and broken.
This commit reworks lowerFnParamTy into iterateParamTypes which returns
enum tags indicating how to handle each parameter. This is then used in
the three places that matter:
* lowering a function type to llvm type
* converting function parameters to the canonical type representation
(with respect to isByRef).
* converting canonical type representation to function arguments at
callsites (again with respect to isByRef).
As a result, we are one step closer to the C ABI tests passing. Before
this commit, attempting to build them crashed the compiler. I isolated
the broken function and verified that it now is lowered correctly. I
will keep working on this one piece at a time until all the C ABI tests
pass, and then I will enable all of them in the CI.
So far it's supported by the LLVM backend only. I recommend for the
other backends to wait for the resolution of #10761 before adding
support for this feature.
* std.math.snan: fix compilation error. Also make it and nan inline.
* LLVM: use a proper enum type for float op instead of enum literal.
Also various cleanups.
* LLVM: use LLVMBuildVectorSplat for vector splat AIR instruction.
- also the bindings had parameter order wrong
* LLVM: additionally handle f16 lowering. For now all targets report OK
but I think we will need to add some exceptions to this list.
Updates stage2 to manually lower softfloat operations for all unary
floating point operations and arithmetic.
Softfloat support still needs to be added for conversion operators
(float<->float and int<->float)
Notably, Value.eql and Value.hash are improved to treat NaN as equal to
itself, so that Type/Value can be hash map keys. Likewise float hashing
normalizes the float value before computing the hash.
LLVM backend: generate DIGlobalVariable's for non-function globals and
rename linkage names when exporting functions and globals.
zig_llvm.cpp: add some wrappers to convert a handful of DI classes
into DINode's since DIGlobalVariable is not a DIScope like the others.
zig_llvm.cpp: add some wrappers to allow replacing the LinkageName of
DISubprogram and DIGlobalVariable.
zig_llvm.cpp: fix DI class mixup causing nonsense reinterpret_cast.
The end result is that GDB is now usable since you now no longer need
to manually cast every global nor fully qualify every export.
Adds 2 new AIR instructions:
* dbg_var_ptr
* dbg_var_val
Sema no longer emits dbg_stmt AIR instructions when strip=true.
LLVM backend: fixed lowerPtrToVoid when calling ptrAlignment on
the element type is problematic.
LLVM backend: fixed alloca instructions improperly getting debug
location annotated, causing chaotic debug info behavior.
zig_llvm.cpp: fixed incorrect bindings for a function that should use
unsigned integers for line and column.
A bunch of C test cases regressed because the new dbg_var AIR
instructions caused their operands to be alive, exposing latent bugs.
Mostly it's just a problem that the C backend lowers mutable
and const slices to the same C type, so we need to represent that in the
C backend instead of printing two duplicate typedefs.
Previously, we did this so that we could insert a debug variable
declaration intrinsic on the alloca. But there is a dbg.value intrinsic
for declaring variables that are values.
`Module.Union.getLayout` now additionally returns a `padding` field
which tells how many bytes are between the final field end offset and
the ending offset of the union. This is used by the LLVM backend to
explicitly insert padding.
LLVM backend: lowering of unions now inserts additional padding so that
LLVM's internals will agree on the ABI size to match what ABI size zig
wants unions to be. This is an alternative to calling LLVMABISizeOfType
and LLVMABIAlignmentOfType which end up crashing when recursive struct
definitions come into play. We no longer ever call these two functions
and the bindings are deleted to avoid future footgun firings.
LLVM backend: lowering of unions now represents untagged unions
consistently. Before it was tripping an assertion.
LLVM backend: switch cases call inttoptr on the case items and condition
if necessary. Prevents tripping an LLVM assertion.
After this commit, we are no longer tripping over any LLVM assertions.
* Sema: resolve type fully when emitting an alloc AIR instruction to
avoid tripping assertion for checking struct field alignment.
* LLVM backend: keep a reference to the LLVM target data alive during
lowering so that we can ask LLVM what it thinks the ABI alignment
and size of LLVM types are. We need this in order to lower tuples and
structs so that we can put in extra padding bytes when Zig disagrees
with LLVM about the size or alignment of something.
* LLVM backend: make the LLVM struct type packed that contains the most
aligned union field and the padding. This prevents the struct from
being too big according to LLVM. In the future, we may want to
consider instead emitting unions in a "flat" manner; putting the tag,
most aligned union field, and padding all in the same struct field
space.
* LLVM backend: make structs with 2 or fewer fields return isByRef=false.
This results in more efficient codegen. This required lowering of
bitcast to sometimes store the struct into an alloca, ptrcast, and
then load because LLVM does not allow bitcasting structs.
* enable more passing behavior tests.
