forked from OSchip/llvm-project
33 lines
1.9 KiB
Plaintext
33 lines
1.9 KiB
Plaintext
* Writing out ELF files is close to working but the following needs to
|
|
be examined more closely:
|
|
- ELF files are written with the wrong E_MACHINE value because
|
|
ELFObjectWriter::WriteHeader function does not yet support
|
|
target specific E_MACHINE values.
|
|
- ELF relocation records are incorrect because the function
|
|
ELFObjectWriter::RecordRelocation is hard coded for X86/X86-64.
|
|
- Relocations use 2-byte / 4-byte to terminology in reference to
|
|
the size of the immediate value being changed. The Xilinx
|
|
terminology seems to be (???) 4-byte / 8-byte in reference
|
|
to the number of bytes of instructions that are being changed.
|
|
- BRLID and like instructions are always assumed to use a 4-byte
|
|
immediate value for the relocation and BEQID and like instructions
|
|
are always assumed to use a 2-byte immediate value for the relocation.
|
|
I think this means that conditional branches like BEQID can only
|
|
branch += 32768 bytes (~8192 instructions). We should allow conditional
|
|
branches to use 4-byte relocations but I'm not sure how to do that
|
|
right now.
|
|
- Relocation records for indirect calls are not being generated
|
|
correctly. These should emit and IMM 0 directly before the ORI
|
|
instruction that loads the register (just like when a BRLID
|
|
instruction is used instead of an ORI).
|
|
|
|
* Code generation seems to work relatively well now but the following
|
|
needs to be examined more closely:
|
|
- The stack layout needs to be examined to make sure it meets
|
|
the standard, especially in regards to var arg functions.
|
|
- The delay slot filler is ad hoc but seems to work. Load and
|
|
store instructions were prevented from being moved to delay
|
|
slots but I'm not sure that is necessary.
|
|
- The processor itineraries are copied from a different backend
|
|
and need to be updated to model the MicroBlaze correctly.
|