Changes to enable linking with LLVM/lld #173
Conversation
|
The rv32imac,gcc failure is likely due to old Binutils installed on the test machines. This could either be fixed by upgrading Binutils there, or maybe use "-z nostart-stop-gc" (as per the commit message) to the linker flags. |
| @@ -24,7 +18,6 @@ SECTIONS | |||
| *(.text.start) | |||
| } | |||
| } | |||
| INSERT BEFORE .text; | |||
|
|
|||
| SECTIONS | |||
| { | |||
There was a problem hiding this comment.
What is the new section table layout after these changes?
This is what currently seems to be produced:
0 .start 00000034 85000000 85000000 00010000 2**2
CONTENTS, ALLOC, LOAD, READONLY, CODE
1 .text 000073c4 85001000 85001000 00011000 2**12
CONTENTS, ALLOC, LOAD, READONLY, CODE
2 .rodata 00000ec7 850083c4 850083c4 000183c4 2**2
CONTENTS, ALLOC, LOAD, READONLY, DATA
3 .ARM.exidx 00000008 8500928c 8500928c 0001928c 2**2
CONTENTS, ALLOC, LOAD, READONLY, DATA
4 .bss 00010048 8500c000 8500c000 00019294 2**14
ALLOC
5 ._archive_cpio 01842a00 8501c048 8501c048 0002c048 2**0
CONTENTS, ALLOC, LOAD, DATA
6 .data 0000002c 8686ea48 8686ea48 0187ea48 2**2
CONTENTS, ALLOC, LOAD, DATA
7 _driver_list 00000030 8686ea74 8686ea74 0187ea74 2**2
CONTENTS, ALLOC, LOAD, DATA
There was a problem hiding this comment.
lld (unlike ld.bfd) does place .rodata before .text, the layout after the changes are:
[ 1] .rodata.str1.1 PROGBITS 0000000040a11000 001000 000b45 01 AMS 0 0 1
[ 2] .rodata PROGBITS 0000000040a11b48 001b48 000668 00 A 0 0 8
[ 3] .start PROGBITS 0000000040a121b0 0021b0 00002c 00 AX 0 0 1
[ 4] .text PROGBITS 0000000040a13000 003000 007214 00 AX 0 0 4096
[ 5] .bss NOBITS 0000000040a1b000 00b000 006090 00 WA 0 0 4096
[ 6] ._archive_cpio PROGBITS 0000000040a21090 011090 486c00 00 WA 0 0 1
[ 7] _driver_list PROGBITS 0000000040ea7c90 497c90 000060 00 WAR 0 0 8
[ 8] .data PROGBITS 0000000040ea7cf0 497cf0 000150 00 WA 0 0 8
| @@ -12,7 +12,7 @@ SECTIONS | |||
| *(.tdata .tdata.* .gnu.linkonce.td.*) | |||
| _tdata_end = . ; | |||
| } | |||
| .tbss : | |||
| .tbss (NOLOAD): | |||
There was a problem hiding this comment.
This is already implicitly present in the current toolchain right?
5 .tbss 0000002c 016c3b90 016c3b90 016b3b8f 2**3
ALLOC, THREAD_LOCAL
There was a problem hiding this comment.
Yes that's correct for lld too. I just added NOLOAD here for intentionality and to avoid potential bugs with other linkers and/or older ones.
|
Err, I may be mixing up the meaning of |
Most linkers heuristically recognise common section names and apply the proper flags to them. .[t]bss sections for example should be recognised and linkers mark them ALLOC, NOLOAD, which means that they don't occupy space in the ELF image and shouldn't be loaded at run-time, but they should be allocated space at run-time in the process image when loaded. Adding NOLOAD (similar to the seL4's kernel) is good for intentionality, and for linkers that might not automatically flag .[t]bss at NOLOAD. I've come across a bug in lld that actually allocated space in the ELF before, but I believe that got fixed. I will add that explanation to the commit message. |
kent-mcleod
added
hw-build
Like .bss, .tbss needs explicit NONLOAD to ensure they are not allocated/loaded in the ELF image. Sponsored by: DARPA. Signed-off-by: Hesham Almatary <[email protected]>
unused attribute does not prevent the section from being garbage-collected during link-time optimisation. This may trigger undefined references errors to [__start|__stop]_driver_list symbols that are expected to be emitted by the linker anyway. Adding "retain" attribute makes sure that the section and its associated symbols are kept regardless of linker's garbage collection. Another fix could be adding "nostart-stop-gc" to the linker flags, but since it is only one section (_driver_list) where its __start/__stop symbols are references, adding retain to it makes more sense. This additional functionality requires binutils version 2.36 or later. Sponsored by: DARPA. Signed-off-by: Hesham Almatary <[email protected]>
Most linkers heuristically recognise common section names and apply the proper flags to them. .[t]bss sections for example should be recognised and linkers mark them ALLOC, NOLOAD, which means that they don't occupy space in the ELF image and shouldn't be loaded at run-time, but they should be allocated space at run-time in the process image when loaded. Adding NOLOAD (similar to the seL4's kernel) is good for intentionality, and for linkers that might not automatically flag .[t]bss at NOLOAD. Sponsored by: DARPA. Signed-off-by: Hesham Almatary <[email protected]>
Linker's operators precedence [1] which follows C makes "+" higher than shifting "<<". Thus, ". + CONFIG_MAX_NODES" will be shifted by 12 in this case, which is wrong and is not the intended behavior. In lld (which properly implements C's operators precedence, unlike ld.bfd), the arithmetic results in a significant value. The outcome is either a huge ELF file size, or a linking error (e.g., relocation R_AARCH64_ADR_PREL_PG_HI21 out of range). This commit fixes the arithmetic and uses += which is more portable. [1] https://sourceware.org/binutils/docs/ld.html Sponsored by: DARPA. Signed-off-by: Hesham Almatary <[email protected]>
This allows to use the same file for both Arm and RISC-V. Furthermore, this new file is more portable and works with both LLVM/lld and GNU's ld, unlike the removed Arm's one which contains directives and sections that lld does not handle (e.g., .interp, INSERT BEFORE .hash). Signed-off-by: Hesham Almatary <[email protected]>
|
I dropped custom Arm's linker handling for sections and directives and rather added a new commit that unifies the linker script between Arm and RISC-V to include fixes for both and to make it more portable to build with lld and ld |
|
Also I modified the "retain" attribute to check if the toolchain supports it or not as this may fail on older GCC versions |
No description provided.