Motivations behind this target
The Unified Extensible Firmware Interface (UEFI) now replaces BIOS in most current PCs.
Currently, there are two main options to develop UEFI applications:
Both solutions require a tedious setup process.
Thank to its internal linker, Free Pascal can become the easiest option to develop UEFI applications, as no external tools would be required to start hacking.
UEFI applications are PE binaries like the Windows target, which Free Pascal already supports. The only difference is the subsystem in the PE header.
Calling conventions under UEFI are identical to the ones under Windows on the same architecture, which Free Pascal already supports, too.
My current patch against trunk is here : http://nan.tf/download/uefi_all_03_01_2016.diff.
It is not yet a clean patch : lots of debug output inside.
I have resume my work on this target. Here is my updated patch against trunk : http://nan.tf/download/fpc_uefi_28_04_2017.diff
Currently, i am focusing on UEFI32 because i have a few cheap atom machines with that version (see https://software.intel.com/en-us/blogs/2015/07/22/why-cheap-systems-run-32-bit-uefi-on-x64-systems) and because i use Haiku as my main development platform, without a 64 bits freepascal yet.
But UEFI64 support should be easy to had with careful API declarations.
For UEFI 32 target, the build command line is :
make all OS_TARGET=uefi CPU_TARGET=i386
For UEFI 64 target, the build command line is :
make all OS_TARGET=uefi CPU_TARGET=x86_64
If you want to cross compile from a different architecture, it might be necessary to specify OS_SOURCE and CPU_SOURCE. Here is for example the command line for Haiku 32 bits.
make all OS_TARGET=uefi CPU_TARGET=x86_64 OS_SOURCE=haiku CPU_SOURCE=i386
Affectation of string constant to string variable does not work. It is currently necessary to create strings through an array of WideChar.
var s : string; begin s := 'Hello world !'; end;
Maybe it is related to Position Independent Code generation in the case of PE executables ?
Any hints in this area are welcome ;-)
With current trunk (3.1.1), this bug has been fixed. It helps a lot to be able to use WriteLn('Something to write') instead of constructing an array of WideChar and call UEFI functions as before ;-).
Other known problem :
list := TStringList.Create;
While constructing other kind of object works as expected, this one currently crash the program... I will have to investigate.
Update : it seems related to uses : without any uses in my program, it just works... With only one, it crash.
Testing UEFI applications with qemu
To avoid creating an image, i use FAT emulation to a local directory where my UEFI application is located :
qemu-system-i386-x86 -pflash /path/to/OVMF.fd -hda fat:local_directory -net none
"-net none" avoid net booting to save a few seconds in the boot process.
From UEFI shell, then write :
to switch to the directory where are located your application then you can launch it :
Everything has started after reading this article while i was lurking about UEFI development : http://blog.theincredibleholk.org/blog/2013/11/18/booting-to-rust/
What about doing that in Freepascal as well ?
It appears that Freepascal is even a better target for such a project because of its internal linker for PE targets : no need to install and configure an external one.
Proof of concept : Freepascal is able to output a PE binary that can run in an UEFI environment. It can output a static text and the version of the firmware to the screen using an UEFI method.
After observing that an annoying bug about string constant was fixed, i have resumed my work on this target.
Currently, here are the features that are implemented :
- WriteLn support (no more awkward array of WideChar allocation required) - Memory allocation
Start of x86_64 support (not functional yet). The idea is to check if the 'uses' problem is related to the use of PE32 binary format. UEFI is supposed to use PE32+ (with 64 bits relocations). UEFI specifications are not clear whether this variant is also used for UEFI32. With a x86_64 CPU and the win64 linker, this uncertainty should disappear...
Update : after a binary comparison between two different binaries, UEFI32 binaries are plain PE32 ones, with an IMAGE_OPTIONAL_HEADER with a size of 224 bytes while a UEFI64 binary contains an IMAGE_OPTIONAL_HEADER with a size of 240 bytes (some fields are 64 bits instead of 32 bits). This is confirmed in https://msdn.microsoft.com/en-us/library/windows/desktop/ms680339(v=vs.85).aspx. Afterthought, it make sense : i suspect Intel has chosen PE binary format and Application Binary Interface (ABI) like the corresponding ones in Windows to minimize the work at Microsoft and make sure they will adopt the UEFI specifications...
Some interesting development resources around UEFI :