Thread overview
SAOC LLDB D integration: 15th Weekly Update
December 31

Hi D community!

Sorry for being late. I'm here again, to describe what I've done during the
fifteenth week of Symmetry Autumn of Code.

LLVM upstream changes: LLD D demangling

I didn't work on the demangler patches but I touched on some other existing
ones, such as implementation of DW_TAG_immutable_type on the LLVM core which
had some missing pieces and added tests. (See
here)

I also added support for other demanglers other than Itanium on LLD linker.
This included the freshly added D demangler along with Rust and other future
demanglers added to LLVM core.

So now instead of:

app.d:16: error: undefined reference to '_D3app7noexistFZi'

You will have this:

app.d:16: error: undefined reference to 'app.noexist()'

This came along with my work on adding D demangler on the LLVM core. You can
read more about this change, here.

Type name dumping and value dumping

I added D type kind mapping to type name for the rest of the built-in types.

I also have found the missing part to make value dumping working. I needed to
implement two missing parts:

  • A way to discover the bit size based on the D type wrapper type kind.
  • A way to get the type information based on a type kind using
    lldb::TypeFlags

This way LLDB can understand if a certain type kind is built-in, has value, is
signed, is integer, is scalar, etc...

So finally, I can print a simple runtime boolean value:

(lldb) ta v
Global variables for app.d in app:
(bool) app.falseval = false
(bool) app.trueval = true

You can consult the source code for those changes
here.

Expanding value dumping to other built-in types

Having this implemented, I now need to compare and check if the DWARF bit size
and encoding match a certain D type kind. The implementation of other types are
not yet pushed, since I faced a problem while adding logic to platform-specific
size types, such as real.

The real problem

Since real is, according to D specification, platform-specific, I need to
accomudate the right bit size according to a certain target and discover the
right floating point encoding. This quite a challange because DWARF doesn't
specify the floating point encoding. To try to understand why, I did a bit of
research about that, and found
this mailing list
thread from 2015 about distiguish different floating point encoding in DWARF.

Right now, there is no way and it seems there is no intention to distiguish
target-specific floating point formats on DWARF, because according to them,
this should be specified on the target ABI. But what if the ABI doesn't specify
this behaviour? We should at least have a way to distiguish IEEE interchangable
format and non-interchangable formats, like 128-bit x86 SSE floating points.

Fortunately, we don't have to worry much about this, since we don't use 128-bit
in any of D implementation, although our spec say:

real: largest floating point size available

Implementation Defined: The real floating point type has at least the range
and precision of the double type. On x86 CPUs it is often implemented as
the 80 bit Extended Real type supported by the x86 FPU.

This is wrong, because, AFAIK, on x86-64 System V ABI, 128-bit floating point
is the largest available, since AMD64 CPUs are required to have at least SSE
extensions, which have support for 128-bit XMM registers to perform
floating-point operations.

So, LDC and DMD generates binaries with System V as target ABI but uses x87 FPU
instead of SSE for real, which means they are out of spec?

Anyway, according to Mathias and as I suggested, the simple way to do this is
to hardcode this according the target triple and the DWARF type name, but I
think this can be problematic for either when we support 128-bit floats or when
the ABI doesn't specify the floating point encoding format.

That said, I would like to have some thoughts on this, specially if someone
knows if there is any special case for certain targets and how DMD/LDC/GDC
interprets the D spec and target ABI spec.

What is next?

I plan to finish support for built-in type value dumping and hopefully start
implementing DIDerivedType which includes DWARF tags for const type
modifiers, alias/typedefs,...

December 31

On Friday, 31 December 2021 at 03:55:40 UTC, Luís Ferreira wrote:

>

Right now, there is no way and it seems there is no intention to distiguish
target-specific floating point formats on DWARF, because according to them,
this should be specified on the target ABI. But what if the ABI doesn't specify
this behaviour? We should at least have a way to distiguish IEEE interchangable
format and non-interchangable formats, like 128-bit x86 SSE floating points.

Fortunately, we don't have to worry much about this, since we don't use 128-bit
in any of D implementation, although our spec say:

We do support native 128-bit floats in D, unless you meant in the compiler implementation, in which case, all native floats (not just real) are banned throughout the compiler.

>
real: largest floating point size available

Implementation Defined: The real floating point type has at least the range
and precision of the double type. On x86 CPUs it is often implemented as
the 80 bit Extended Real type supported by the x86 FPU.

This is wrong, because, AFAIK, on x86-64 System V ABI, 128-bit floating point
is the largest available, since AMD64 CPUs are required to have at least SSE
extensions, which have support for 128-bit XMM registers to perform
floating-point operations.

So, LDC and DMD generates binaries with System V as target ABI but uses x87 FPU
instead of SSE for real, which means they are out of spec?

Anyway, according to Mathias and as I suggested, the simple way to do this is
to hardcode this according the target triple and the DWARF type name, but I
think this can be problematic for either when we support 128-bit floats or when
the ABI doesn't specify the floating point encoding format.

That said, I would like to have some thoughts on this, specially if someone
knows if there is any special case for certain targets and how DMD/LDC/GDC
interprets the D spec and target ABI spec.

Just have that real map to C long double and be done with it, even if the hardware may support a bigger float. You don't want to be incompatible with the system you're running on, else you'll be locked out of using the C math library.

January 05
On Fri, 2021-12-31 at 17:03 +0000, Iain Buclaw via Digitalmars-d wrote:
> We do support native 128-bit floats in D, unless you meant in the compiler implementation, in which case, all native floats (not just real) are banned throughout the compiler.

Oh ok, didn't know about that. For now, I hardcoded 64, 80 and 128 bit real type kinds. Later, if we end up finding out that `real` is intended to direct map to `long double` I may use clang::TargetInfo, which gives `long double` bit size according to a specified target triple.
> 

> Just have that `real` map to C `long double` and be done with it, even if the hardware may support a bigger float. You don't want to be incompatible with the system you're running on, else you'll be locked out of using the C math library.

Well, I don't think that directly mapping it is correct. e.g. https://godbolt.org/z/66f6v17Tn . Is this intended? Anyway, I still think we should discuss specification wording about how real is implemented for each target. Maybe worth mention `long double` if direct mapping is intended? System V ABI is specific about `long double` size and it is not the largest supported floating point, as I mentioned above.

-- 
Sincerely,
Luís Ferreira @ lsferreira.net



January 05
On Wednesday, 5 January 2022 at 04:34:31 UTC, Luís Ferreira wrote:
> On Fri, 2021-12-31 at 17:03 +0000, Iain Buclaw via Digitalmars-d wrote:
>> We do support native 128-bit floats in D, unless you meant in the compiler implementation, in which case, all native floats (not just real) are banned throughout the compiler.
>
> Oh ok, didn't know about that. For now, I hardcoded 64, 80 and 128 bit real type kinds. Later, if we end up finding out that `real` is intended to direct map to `long double` I may use clang::TargetInfo, which gives `long double` bit size according to a specified target triple.

This is always the case with gdc, so that would be highly recommended.


>> Just have that `real` map to C `long double` and be done with it, even if the hardware may support a bigger float. You don't want to be incompatible with the system you're running on, else you'll be locked out of using the C math library.
>
> Well, I don't think that directly mapping it is correct. e.g. https://godbolt.org/z/66f6v17Tn . Is this intended?

Looks like ldc is in the wrong there, real.sizeof should always be 113 on RISC-V.

https://explore.dgnu.org/z/9MsGjG

January 06
On Wed, 2022-01-05 at 21:52 +0000, Iain Buclaw via Digitalmars-d wrote:
> This is always the case with gdc, so that would be highly recommended.
> 
> Looks like ldc is in the wrong there, real.sizeof should always be 113 on RISC-V.
> 
> https://explore.dgnu.org/z/9MsGjG

Right. I'm going to write a patch to fix that then. I'm also going to create a patch on specification to clarify the wording and discuss there.

For now I'm going to stick with the hardcoded version I created, just for testing purposes, then update it to clang::TargetInfo to reflect the long double behaviour.

-- 
Sincerely,
Luís Ferreira @ lsferreira.net