Linking fails with "relocation truncated to fit" with aggressive inlining - linker

To work around a Rust compiler bug in the AVR backend, I have marked a lot of my functions as #[inline(always)], by just adding the annotations until enough case-of-case optimizations etc. would fire that I no longer run into the issue.
However, with these annotations, linking now fails with lots of relocation truncated to fit messages:
target/avr-atmega328p/release/deps/chip8_avr-41c427b8d446a439.o: In function `LBB5_18':
chip8_avr.cgu-0.rs:(.text.main+0x432): relocation truncated to fit: R_AVR_7_PCREL against `no symbol'
target/avr-atmega328p/release/deps/chip8_avr-41c427b8d446a439.o: In function `LBB5_23':
chip8_avr.cgu-0.rs:(.text.main+0x45c): relocation truncated to fit: R_AVR_7_PCREL against `no symbol'
target/avr-atmega328p/release/deps/chip8_avr-41c427b8d446a439.o: In function `LBB5_31':
chip8_avr.cgu-0.rs:(.text.main+0x4ae): relocation truncated to fit: R_AVR_7_PCREL against `no symbol'
target/avr-atmega328p/release/deps/chip8_avr-41c427b8d446a439.o: In function `LBB5_34':
chip8_avr.cgu-0.rs:(.text.main+0x4d2): relocation truncated to fit: R_AVR_7_PCREL against `no symbol'
target/avr-atmega328p/release/deps/chip8_avr-41c427b8d446a439.o: In function `LBB5_146':
chip8_avr.cgu-0.rs:(.text.main+0x58a): relocation truncated to fit: R_AVR_7_PCREL against `no symbol'
chip8_avr.cgu-0.rs:(.text.main+0x58e): relocation truncated to fit: R_AVR_7_PCREL against `no symbol'
chip8_avr.cgu-0.rs:(.text.main+0x592): relocation truncated to fit: R_AVR_7_PCREL against `no symbol'
target/avr-atmega328p/release/deps/chip8_avr-41c427b8d446a439.o: In function `LBB5_153':
chip8_avr.cgu-0.rs:(.text.main+0x59a): relocation truncated to fit: R_AVR_7_PCREL against `no symbol'
chip8_avr.cgu-0.rs:(.text.main+0x59e): relocation truncated to fit: R_AVR_7_PCREL against `no symbol'
chip8_avr.cgu-0.rs:(.text.main+0x5a6): relocation truncated to fit: R_AVR_7_PCREL against `no symbol'
chip8_avr.cgu-0.rs:(.text.main+0x5aa): additional relocation overflows omitted from the output
collect2: error: ld returned 1 exit status
This SO answer implies that large intra-function jumps are something that the compiler needs to be prepared for. What is the equivalent setting on Rust?

After looking at disassemblies in enough detail, it turns out that these relocation targets are all in offsets of branch instructions, not (short) jumps; e.g. at 0x08:
00000000 <_ZN12chip8_engine7opcodes6decode17haab3c6c935229a6aE>:
0: e8 2f mov r30, r24
2: f9 2f mov r31, r25
4: 80 e0 ldi r24, 0x00 ; 0
6: 61 30 cpi r22, 0x01 ; 1
8: 01 f4 brne .+0 ; 0xa <_ZN12chip8_engine7opcodes6decode17haab3c6c935229a6aE+0xa>
a: 81 83 std Z+1, r24 ; 0x01
c: 82 83 std Z+2, r24 ; 0x02
e: 81 e0 ldi r24, 0x01 ; 1
00000010 <LBB0_2>:
10: 80 83 st Z, r24
12: 08 95 ret
The AVR fork of the Rust compiler currently generates these branches with empty (.+0) offsets, then tries to use the linker to fill them in. For large enough functions, these intra-function offsets can become larger than what fits into a single branch instruction.
I've reported this as a compiler bug in the AVR fork of Rust. One potential solution that came up there was to get the linker to generate a two-step branch (a branch to a jump) for cases where the offset doesn't fit; the other is to not use the linker at all: since the branches in question are intra-function, the relative addresses should be known at compile time, allowing for generating a two-step branch when needed.

Related

Will an executable access shared-libraries' global variable via GOT?

I was learning dynamic linking recently and gave it a try:
dynamic.c
int global_variable = 10;
int XOR(int a) {
return global_variable;
}
test.c
#include <stdio.h>
extern int global_variable;
extern int XOR(int);
int main() {
global_variable = 3;
printf("%d\n", XOR(0x10));
}
The compiling commands are:
clang -shared -fPIC -o dynamic.so dynamic.c
clang -o test test.c dynamic.so
I was expecting that in executable test the main function will access global_variable via GOT. However, on the contrary, the global_variable is placed in test's data section and XOR in dynamic.so access the global_variable indirectly.
Could anyone tell me why the compiler didn't ask the test to access global_variable via GOT, but asked the shared object file to do so?
Part of the point of a shared library is that one copy gets loaded into memory, and multiple processes can access that one copy. But every program has its own copy of each of the library's variables. If they were accessed relative to the library's GOT then those would instead be shared among the processes using the library, just like the functions are.
There are other possibilities, but it is clean and consistent for each executable to provide for itself all the variables it needs. That requires the library functions to access all of its variables with static storage duration (not just external ones) indirectly, relative to the program. This is ordinary dynamic linking, just going the opposite direction from what you usually think of.
Turns out my clang produced PIC by default so it messed with results.
I will leave updated answer here, and the original can be read below it.
After digging a bit more into the topic i have noticed that compilation of test.c does not generate a .got section by itself. You can check it by compiling the executable into an object file and omitting the linking step for now (-c option):
clang -c -o test.o test.c
If you inspect the sections of resulting object file with readelf -S you will notice that there is no .got in there:
Section Headers:
[Nr] Name Type Address Offset
Size EntSize Flags Link Info Align
[ 0] NULL 0000000000000000 00000000
0000000000000000 0000000000000000 0 0 0
[ 1] .text PROGBITS 0000000000000000 00000040
0000000000000035 0000000000000000 AX 0 0 1
[ 2] .rela.text RELA 0000000000000000 00000210
0000000000000060 0000000000000018 I 11 1 8
[ 3] .data PROGBITS 0000000000000000 00000075
0000000000000000 0000000000000000 WA 0 0 1
[ 4] .bss NOBITS 0000000000000000 00000075
0000000000000000 0000000000000000 WA 0 0 1
[ 5] .rodata PROGBITS 0000000000000000 00000075
0000000000000004 0000000000000000 A 0 0 1
[ 6] .comment PROGBITS 0000000000000000 00000079
0000000000000013 0000000000000001 MS 0 0 1
[ 7] .note.GNU-stack PROGBITS 0000000000000000 0000008c
0000000000000000 0000000000000000 0 0 1
[ 8] .note.gnu.pr[...] NOTE 0000000000000000 00000090
0000000000000030 0000000000000000 A 0 0 8
[ 9] .eh_frame PROGBITS 0000000000000000 000000c0
0000000000000038 0000000000000000 A 0 0 8
[10] .rela.eh_frame RELA 0000000000000000 00000270
0000000000000018 0000000000000018 I 11 9 8
[11] .symtab SYMTAB 0000000000000000 000000f8
00000000000000d8 0000000000000018 12 4 8
[12] .strtab STRTAB 0000000000000000 000001d0
000000000000003e 0000000000000000 0 0 1
[13] .shstrtab STRTAB 0000000000000000 00000288
0000000000000074 0000000000000000 0 0 1
This means that the entirety of .got section present in the test executable actually comes from dynamic.so, as it is PIC and uses GOT.
Would it be possible to compile dynamic.so as non-PIC as well? Turns out it apparently used to be 10 years ago (the article compiles examples to 32-bits, they dont have to work on 64 bits!). Linked article describes how a non-PIC shared library was relocated at load time - basically, every time an address that needed to be relocated after loading was present in machine code, it was instead set to zeroes and a relocation of a certain type was set in the library. During loading of the library the loader filled the zeros with actual runtime address of data/code that was needed. It is important to note that it cannot be applied in your though as 64-bit shared libraries cannot be made out of non-PIC (Source).
If you compile dynamic.so as a shared 32-bit library instead and do not use the -fPIC option (you usually need special repositories enabled to compile 32-bit code and have 32-bit libc installed):
gcc -m32 dynamic.c -shared -o dynamic.so
You will notice that:
// readelf -s dynamic.so
(... lots of output)
27: 00004010 4 OBJECT GLOBAL DEFAULT 19 global_variable
// readelf -S dynamic.so
(... lots of output)
[17] .got PROGBITS 00003ff0 002ff0 000010 04 WA 0 0 4
[18] .got.plt PROGBITS 00004000 003000 00000c 04 WA 0 0 4
[19] .data PROGBITS 0000400c 00300c 000008 00 WA 0 0 4
[20] .bss NOBITS 00004014 003014 000004 00 WA 0 0 1
global_variable is at offset 0x4010 which is inside .data section. Also, while .got is present (at offset 0x3ff0), it only contains relocations coming from other sources than your code:
// readelf -r
Offset Info Type Sym.Value Sym. Name
00003f28 00000008 R_386_RELATIVE
00003f2c 00000008 R_386_RELATIVE
0000400c 00000008 R_386_RELATIVE
00003ff0 00000106 R_386_GLOB_DAT 00000000 _ITM_deregisterTM[...]
00003ff4 00000206 R_386_GLOB_DAT 00000000 __cxa_finalize#GLIBC_2.1.3
00003ff8 00000306 R_386_GLOB_DAT 00000000 __gmon_start__
00003ffc 00000406 R_386_GLOB_DAT 00000000 _ITM_registerTMCl[...]
This article introduces GOT as part of introduction on PIC, and i have found that to be the case in plenty of places, which would suggest that indeed GOT is only used by PIC code although i am not 100% sure of it and i recommend researching the topic more.
What does this mean for you? A section in the first article i linked called "Extra credit #2" contains an explanation for a similar scenario. Although it is 10 years old, uses 32-bit code and the shared library is non-PIC it shares some similarities with your situation and might explain the problem you presented in your question.
Also keep in mind that (although similar) -fPIE and -fPIC are two separate options with slightly different effects and that if your executable during inspection is not loaded at 0x400000 then it probably is compiled as PIE without your knowledge which might also have impact on results. In the end it all boils down to what data is to be shared between processes, what data/code can be loaded at arbitrary address, what has to be loaded at fixed address etc. Hope this helps.
Also two other answers on Stack Overflow which seem relevant to me: here and here. Both the answers and comments.
Original answer:
I tried reproducing your problem with exactly the same code and compilation commands as the ones you provided, but it seems like both main and XOR use the GOT to access the global_variable. I will answer by providing example output of commands that i used to inspect the data flow. If your outputs differ from mine, it means there is some other difference between our environments (i mean a big difference, if only addresses/values are different then its ok). Best way to find that difference is for you to provide commands you originally used as well as their output.
First step is to check what address is accessed whenever a write or read to global_variable happens. For that we can use objdump -D -j .text test command to disassemble the code and look at the main function:
0000000000001150 <main>:
1150: 55 push %rbp
1151: 48 89 e5 mov %rsp,%rbp
1154: 48 8b 05 8d 2e 00 00 mov 0x2e8d(%rip),%rax # 3fe8 <global_variable>
115b: c7 00 03 00 00 00 movl $0x3,(%rax)
1161: bf 10 00 00 00 mov $0x10,%edi
1166: e8 d5 fe ff ff call 1040 <XOR#plt>
116b: 89 c6 mov %eax,%esi
116d: 48 8d 3d 90 0e 00 00 lea 0xe90(%rip),%rdi # 2004 <_IO_stdin_used+0x4>
1174: b0 00 mov $0x0,%al
1176: e8 b5 fe ff ff call 1030 <printf#plt>
117b: 31 c0 xor %eax,%eax
117d: 5d pop %rbp
117e: c3 ret
117f: 90 nop
Numbers in the first column are not absolute addresses - instead they are offsets relative to the base address at which the executable will be loaded. For the sake of explanation i will refer to them as "offsets".
The assembly at offset 0x115b and 0x1161 comes directly from the line global_variable = 3; in your code. To confirm that, you could compile the program with -g for debug symbols and invoke objdump with -S. This will display source code above corresponding assembly.
We will focus on what these two instructions are doing. First instruction is a mov of 8 bytes from a location in memory to the rax register. The location in memory is given as relative to the current rip value, offset by a constant 0x2e8d. Objdump already calculated the value for us, and it is equal to 0x3fe8. So this will take 8 bytes present in memory at the 0x3fe8 offset and store them in the rax register.
Next instruction is again a mov, the suffix l tells us that data size is 4 bytes this time. It stores a 4 byte integer with value equal to 0x3 in the location pointed to by the current value of rax (not in the rax itself! brackets around a register such as those in (%rax) signify that the location in the instruction is not the register itself, but rather where its contents are pointing to!).
To summarize, we read a pointer to a 4 byte variable from a certain location at offset 0x3fe8 and later store an immediate value of 0x3 at the location specified by said pointer. Now the question is: where does that offset of 0x3fe8 come from?
It actually comes from GOT. To show the contents of the .got section we can use the objdump -s -j .got test command. -s means we want to focus on actual raw contents of the section, without any disassembling. The output in my case is:
test: file format elf64-x86-64
Contents of section .got:
3fd0 00000000 00000000 00000000 00000000 ................
3fe0 00000000 00000000 00000000 00000000 ................
3ff0 00000000 00000000 00000000 00000000 ................
The whole section is obviously set to zero, as GOT is populated with data after loading the program into memory, but what is important is the address range. We can see that .got starts at 0x3fd0 offset and ends at 0x3ff0. This means it also includes the 0x3fe8 offset - which means the location of global_variable is indeed stored in GOT.
Another way of finding this information is to use readelf -S test to show sections of the executable file and scroll down to the .got section:
[Nr] Name Type Address Offset
Size EntSize Flags Link Info Align
(...lots of sections...)
[22] .got PROGBITS 0000000000003fd0 00002fd0
0000000000000030 0000000000000008 WA 0 0 8
Looking at the Address and Size columns, we can see that the section is loaded at offset 0x3fd0 in memory and its size is 0x30 - which corresponds to what objdump displayed. Note that in readelf ouput "Offset" is actually the offset into the file form which the program is loaded - not the offset in memory that we are interested in.
by issuing the same commands on the dynamic.so library we get similar results:
00000000000010f0 <XOR>:
10f0: 55 push %rbp
10f1: 48 89 e5 mov %rsp,%rbp
10f4: 89 7d fc mov %edi,-0x4(%rbp)
10f7: 48 8b 05 ea 2e 00 00 mov 0x2eea(%rip),%rax # 3fe8 <global_variable##Base-0x38>
10fe: 8b 00 mov (%rax),%eax
1100: 5d pop %rbp
1101: c3 ret
So we see that both main and XOR use GOT to find the location of global_variable.
As for the location of global_variable we need to run the program to populate GOT. For that we can use GDB. We can run our program in GDB by invoking it this way:
LD_LIBRARY_PATH="$LD_LIBRARY_PATH:." gdb ./test
LD_LIBRARY_PATH environment variable tells linker where to look for shared objects, so we extend it to include the current directory "." so that it may find dynamic.so.
After the GDB loads our code, we may invoke break main to set up a breakpoint at main and run to run the program. The program execution should pause at the beginning of the main function, giving us a view into our executable after it was fully loaded into memory, with GOT populated.
Running disassemble main in this state will show us the actual absolute offsets into memory:
Dump of assembler code for function main:
0x0000555555555150 <+0>: push %rbp
0x0000555555555151 <+1>: mov %rsp,%rbp
=> 0x0000555555555154 <+4>: mov 0x2e8d(%rip),%rax # 0x555555557fe8
0x000055555555515b <+11>: movl $0x3,(%rax)
0x0000555555555161 <+17>: mov $0x10,%edi
0x0000555555555166 <+22>: call 0x555555555040 <XOR#plt>
0x000055555555516b <+27>: mov %eax,%esi
0x000055555555516d <+29>: lea 0xe90(%rip),%rdi # 0x555555556004
0x0000555555555174 <+36>: mov $0x0,%al
0x0000555555555176 <+38>: call 0x555555555030 <printf#plt>
0x000055555555517b <+43>: xor %eax,%eax
0x000055555555517d <+45>: pop %rbp
0x000055555555517e <+46>: ret
End of assembler dump.
(gdb)
Our 0x3fe8 offset has turned into an absolute address of equal to 0x555555557fe8. We may again check that this location comes from the .got section by issuing maintenance info sections inside GDB, which will list a long list of sections and their memory mappings. For me .got is placed in this address range:
[21] 0x555555557fd0->0x555555558000 at 0x00002fd0: .got ALLOC LOAD DATA HAS_CONTENTS
Which contains 0x555555557fe8.
To finally inspect the address of global_variable itself we may examine the contents of that memory by issuing x/xag 0x555555557fe8. Arguments xag of the x command deal with the size, format and type of data being inspected - for explanation invoke help x in GDB. On my machine the command returns:
0x555555557fe8: 0x7ffff7fc4020 <global_variable>
On your machine it may only display the address and the data, without the "<global_variable>" helper, which probably comes from an extension i have installed called pwndbg. It is ok, because the value at that address is all we need. We now know that the global_variable is located in memory under the address 0x7ffff7fc4020. Now we may issue info proc mappings in GDB to find out what address range does this address belong to. My output is pretty long, but among all the ranges listed there is one of interest to us:
0x7ffff7fc4000 0x7ffff7fc5000 0x1000 0x3000 /home/user/test_got/dynamic.so
The address is inside of that memory area, and GDB tells us that it comes from the dynamic.so library.
In case any of the outputs of said commands are different for you (change in a value is ok - i mean a fundamental difference like addresses not belonging to certain address ranges etc.) please provide more information about what exactly did you do to come to the conclusion that global_variable is stored in the .data section - what commands did you invoke and what outputs they produced.

gdb incorrectly resolving stack variable location

While debugging a small kernel I am writing for fun/learning experience, I encountered a somewhat puzzling issue with gdb where it is apparently not correctly resolving local variable addresses on the stack. My investigation so far suggests that the debugging symbols are correct but somehow gdb still reads from a wrong memory location when displaying the contents of that variable.
The relevant C code in question is:
typedef union
{
uint16_t packed;
struct __attribute__((packed))
{
uint8_t PhysicalLimit;
uint8_t LinearLimit;
} limits;
} MemAddrLimits;
void KernelMain32()
{
ClearScreen();
SimplePrint("kernelMain32");
MemAddrLimits memAddr;
memAddr.packed = GetMemoryAddressLimits();
for (;;) {}
}
where GetMemoryAddressLimits() returns the memory address widths provided by the cpuid instruction in a 2-byte integer (0x3028 currently for my tests). However, when stepping through this function using gdb to print the value of memAddr does not show the right result:
gdb> p memAddr
$1 = {packed = 0, limits = {PhysicalLimit = 0 '\000', LinearLimit = 0 '\000'}}
gdb> info locals
memAddr = {packed = 0, limits = {PhysicalLimit = 0 '\000', LinearLimit = 0 '\000'}}
gdb> info addr memAddr prints Symbol "memAddr" is a variable at frame base reg $ebp offset 8+-18. i.e., memAddr is located at ebp-10 and, indeed, inspecting that address shows the expected content:
gdb> x/hx $ebp-10
0x8ffee: 0x3028
In contrast gdb> p &memAddr gives a value of (MemAddrLimits *) 0x7f6 at which location the memory is zeroed.
When declaring memAddr as a uint16_t instead of my union type these issues do not occur. In that case we get
gdb> info addr memAddr
Symbol "memAddr" is multi-location:
Range 0x8b95-0x8b97: a variable in $eax
.
However, the result is still (also) written to ebp-10, i.e., the disassembly of the function is identical - the only difference is in debug symbols.
Am I missing something here or has someone a good idea on what might be going wrong in this case?
More Details
Program versions and build flags
Using gcc (Ubuntu 9.3.0-10ubuntu2) 9.3.0 and GNU gdb (Ubuntu 9.1-0ubuntu1) 9.1.
Compiling with flags
-ffreestanding -m32 -fcf-protection=none -fno-pie -fno-pic -O0 -gdwarf-2 -fvar-tracking -fvar-tracking-assignments
and linking with
-m elf_i386 -nodefaultlibs -nostartfiles -Ttext 0x7c00 -e start -g
The linking phase produces the kernel.elf which I postprocess to extract the raw executable binary as well as a symbols file to load into gdb. So far, this has been working well for me.
There's obviously more code involved in the binary than what I have shown, most of which written in assembly, which shouldn't be relevant here.
Compiled Files
gcc generates the following code (snippet from objdump -d kernel.elf):
00008b74 <KernelMain32>:
8b74: 55 push ebp
8b75: 89 e5 mov ebp,esp
8b77: 83 ec 18 sub esp,0x18
8b7a: e8 f0 fe ff ff call 8a6f <ClearScreen>
8b7f: 68 41 8c 00 00 push 0x8c41
8b84: e8 7a ff ff ff call 8b03 <SimplePrint>
8b89: 83 c4 04 add esp,0x4
8b8c: e8 0f 00 00 00 call 8ba0 <GetMemoryAddressLimits>
8b91: 66 89 45 f6 mov WORD PTR [ebp-0xa],ax
8b95: eb fe jmp 8b95 <KernelMain32+0x21>
From that we can see that memAddr is indeed located at ebp-10 on the stack, consistent to what gdb> info addr memAddr told us.
Dwarf information (objdump --dwarf kernel.elf):
<1><4ff>: Abbrev Number: 20 (DW_TAG_subprogram)
<500> DW_AT_external : 1
<501> DW_AT_name : (indirect string, offset: 0x23c): KernelMain32
<505> DW_AT_decl_file : 2
<506> DW_AT_decl_line : 79
<507> DW_AT_decl_column : 6
<508> DW_AT_low_pc : 0x8b74
<50c> DW_AT_high_pc : 0x8b97
<510> DW_AT_frame_base : 0x20 (location list)
<514> DW_AT_GNU_all_call_sites: 1
<515> DW_AT_sibling : <0x544>
<2><519>: Abbrev Number: 21 (DW_TAG_variable)
<51a> DW_AT_name : (indirect string, offset: 0x2d6): memAddr
<51e> DW_AT_decl_file : 2
<51f> DW_AT_decl_line : 86
<520> DW_AT_decl_column : 19
<521> DW_AT_type : <0x4f3>
<525> DW_AT_location : 2 byte block: 91 6e (DW_OP_fbreg: -18)
and relevant snippet from objdump --dwarf=loc kernel.elf:
Offset Begin End Expression
00000000 <End of list>
objdump: Warning: There is an overlap [0x8 - 0x0] in .debug_loc section.
00000000 <End of list>
objdump: Warning: There is a hole [0x8 - 0x20] in .debug_loc section.
00000020 00008b74 00008b75 (DW_OP_breg4 (esp): 4)
0000002c 00008b75 00008b77 (DW_OP_breg4 (esp): 8)
00000038 00008b77 00008b97 (DW_OP_breg5 (ebp): 8)
00000044 <End of list>
[...]
These all seem to be what I'd expect. (I'm not sure if the warnings in the last one have significance, though).
Additional Note
If I change compilation flag -gdwarf-2 to just -g I get
gdb> p &memAddr
$1 = (MemAddrLimits *) 0x8ffde
gdb> info addr memAddr
Symbol "memAddr" is a complex DWARF expression:
0: DW_OP_fbreg -18
.
gdb> p memAddr
$2 = {packed = 0, limits = {PhysicalLimit = 0 '\000', LinearLimit = 0 '\000'}}
gdb> p/x $ebp-10
$3 = 0x8ffee
So memAddr is still not resolved correctly but p &memAddr at least is in the stack frame and not somewhere completely different. However, info addr memAddr seems to have problems now...
After some more investigation, I have tracked this to being due to remote debugging 32-bit code (my kernel not yet having switched to long mode) on a x86-64 qemu emulated system.
If I debug the same code with qemu-system-i386 everything works just as it should.

GCC behavior for unresolved weak functions

Consider the simple program below:
__attribute__((weak)) void weakf(void);
int main(int argc, char *argv[])
{
weakf();
}
When compiling this with gcc and running it on a Linux PC, it segfaults. When running it on ARM CM0 (arm-none-eabi-gcc), the linker replace the undefined symbol by a jump to the following instruction and a nop.
Where is this behavior documented? Is there possible ways to change it through command line options? I have been through GCC and LD documentations, there is no information about that.
If I check the ARM compiler doc however, this is clearly explained.
man nm
I was reading some docs and happened to come across a related quote for this:
man nm
says:
"V"
"v" The symbol is a weak object. When a weak defined symbol is linked with a normal defined symbol, the normal defined symbol is used with no error. When a weak undefined symbol is linked and
the symbol is not defined, the value of the weak symbol becomes zero with no error. On some systems, uppercase indicates that a default value has been specified.
"W"
"w" The symbol is a weak symbol that has not been specifically tagged as a weak object symbol. When a weak defined symbol is linked with a normal defined symbol, the normal defined symbol is
used with no error. When a weak undefined symbol is linked and the symbol is not defined, the value of the symbol is determined in a system-specific manner without error. On some systems,
uppercase indicates that a default value has been specified.
nm is part of Binutils, which GCC uses under the hood, so this should be canonical enough.
Then, example on your source file:
main.c
__attribute__((weak)) void weakf(void);
int main(int argc, char *argv[])
{
weakf();
}
we do:
gcc -O0 -ggdb3 -std=c99 -Wall -Wextra -pedantic -o main.out main.c
nm main.out
which contains:
w weakf
and so it is a system-specific value. I can't find where the per-system behavior is defined however. I don't think you can do better than reading Binutils source here.
v would be fixed to 0, but that is used for undefined variables (which are objects): How to make weak linking work with GCC?
Then:
gdb -batch -ex 'disassemble/rs main' main.out
gives:
Dump of assembler code for function main:
main.c:
4 {
0x0000000000001135 <+0>: 55 push %rbp
0x0000000000001136 <+1>: 48 89 e5 mov %rsp,%rbp
0x0000000000001139 <+4>: 48 83 ec 10 sub $0x10,%rsp
0x000000000000113d <+8>: 89 7d fc mov %edi,-0x4(%rbp)
0x0000000000001140 <+11>: 48 89 75 f0 mov %rsi,-0x10(%rbp)
5 weakf();
0x0000000000001144 <+15>: e8 e7 fe ff ff callq 0x1030 <weakf#plt>
0x0000000000001149 <+20>: b8 00 00 00 00 mov $0x0,%eax
6 }
0x000000000000114e <+25>: c9 leaveq
0x000000000000114f <+26>: c3 retq
End of assembler dump.
which means it gets resolved at the PLT.
Then since I don't fully understand PLT, I experimentally verify that it resolves to address 0 and segfaults:
gdb -nh -ex run -ex bt main.out
I'm supposing the same happens on ARM, it must just set it to 0 as well.
On ARM with gcc this code does not work for me (test on armv7 with gcc Debian 4.6.3-14+rpi1). It looks like the arm compiler toolchain has a different behavior.
I do not found useful documentation for this behavior. It seems that the weakf equals NULL if it's undefine at link time.
So I sugest you to test it:
if (weakf == NULL) printf ("weakf not found\n");
else weakf();

meaning of R_386_32/R_386_PC32 in .rel.text section of elf

to understand the concept of relocation, i wrote a simple chk.c program as follows :
1 #include<stdio.h>
2 main(){
3 int x,y,sum;
4 x = 3;
5 y = 4;
6 sum = x + y;
7 printf("sum = %d\n",sum);
8 }
its equivalent assembly code, using "objdump -d chk.o" is :
00000000 <main>:
0: 55 push %ebp
1: 89 e5 mov %esp,%ebp
3: 83 e4 f0 and $0xfffffff0,%esp
6: 83 ec 20 sub $0x20,%esp
9: c7 44 24 1c 03 00 00 movl $0x3,0x1c(%esp)
10: 00
11: c7 44 24 18 04 00 00 movl $0x4,0x18(%esp)
18: 00
19: 8b 44 24 18 mov 0x18(%esp),%eax
1d: 8b 54 24 1c mov 0x1c(%esp),%edx
21: 8d 04 02 lea (%edx,%eax,1),%eax
24: 89 44 24 14 mov %eax,0x14(%esp)
28: b8 00 00 00 00 mov $0x0,%eax
2d: 8b 54 24 14 mov 0x14(%esp),%edx
31: 89 54 24 04 mov %edx,0x4(%esp)
35: 89 04 24 mov %eax,(%esp)
38: e8 fc ff ff ff call 39 <main+0x39>
3d: c9 leave
3e: c3 ret
and .rel.text section seen using readelf is as follows :
Relocation section '.rel.text' at offset 0x360 contains 2 entries:
Offset Info Type Sym.Value Sym. Name
00000029 00000501 R_386_32 00000000 .rodata
00000039 00000902 R_386_PC32 00000000 printf
i have following questions based on this :
1) from 2nd entry in .rel.text section, i am able to understand that value at offset 0x39 in .text section (which is 0xfcffffff here) has to be replaced with address of a symbol associated with index 9 of symbol table (& that comes out to be printf). But i am not able to clearly understand the meaning of 0x02 (ELF32_R_TYPE) here. What does R_386_PC32 specify here ? Can anyone please explain its meaning clearly.
2) i am also not able to understand the 1st entry. what needs to be replaced at offset of 0x29 in .text section and why is not clear here. Again i want to know the meaning of R_386_32 here. i found one pdf elf_format.pdf, but i am not able to clearly understand the meaning of "Type" in .rel.text section from that.
3) Also i want to know the meaning of assembly inst "lea (%edx,%eax,1),%eax". Though i found a very good link (What's the purpose of the LEA instruction?) describing the meaning of lea, but the format of lea (what are 3 arg's inside brackets) is not clear.
if anyone can clearly explain the answers of above questions, it will be greatly appreciated. i am still struggling to find the answers to these questions,though i have tried a lot with google.
one more question. i have shown the symbol table entries for both offset 5 and 9 below.
Num: Value Size Type Bind Vis Ndx Name
5: 00000000 0 SECTION LOCAL DEFAULT 5
9: 00000000 0 NOTYPE GLOBAL DEFAULT UND printf'
The info field for the first entry in .rel.text table is 0x05 which indicates the index of symbol table. I have shown the symbol table entry for index 5 above, but not able to understand how that tells us that it is for .rodata .
1), 2): R_386_32 is a relocation that places the absolute 32-bit address of the symbol into the specified memory location. R_386_PC32 is a relocation that places the PC-relative 32-bit address of the symbol into the specified memory location. R_386_32 is useful for static data, as shown here, since the compiler just loads the relocated symbol address into some register and then treats it as a pointer. R_386_PC32 is useful for function references since it can be used as an immediate argument to call. See elf_machdep.c for an example of how the relocations are processed.
3) lea (%edx,%eax,1),%eax means simply %eax = %edx + 1*%eax if expressed in C syntax. Here, it's basically being used as a substitute for the add opcode.
EDIT: Here's an example.
Suppose your code gets loaded into memory starting at 0x401000, that the string "sum = %d\n" ends up at 0x401800 (at the start of the .rodata section), and that printf is at 0x1400ab80, in libc.
Then, the R_386_32 relocation at 0x29 will place the bytes 00 18 40 00 at 0x401029 (simply copying the absolute address of the symbol), making the instruction at 0x401028
401028: b8 00 18 40 00 mov $0x401800,%eax
The R_386_PC32 relocation at 0x39 places the bytes 43 9b c0 13 at 0x401039 (the value 0x1400ab80 - 0x40103d = 0x13c09b43 in hex), making that instruction
401038: e8 43 9b c0 13 call $0x1400ab80 <printf>
We subtract 0x40103d to account for the value of %pc (which is the address of the instruction after call).
The first relocation entry is to get the pointer to your format string ("sum = ...") in the process of setting up the call to printf. Since the .rodata section is relocated as well as the .text section, references to strings and other constant data will need fixups.
With that in mind, it would appear that the R_386_32 relocations deal with data, and the R_386_PC32 with code addresses, but the ELF spec (of which I do not have a handy copy) probably explains the various details.
The lea instruction is what the compiler chose to perform the addition for this routine. It could have chosen add or a couple other possibilities, but that form of lea seems to be used quite often for certain cases, because it can combine an addition with a multiplication. The result of the instruction is lea (%edx,%eax,1),%eax is that %eax will get the value of %edx + 1 * %eax. The 1 can be replaced with a restricted set of small integers. The original purpose of the lea instruction was "Load Effective Address" - to take a base pointer, an index, and a size and yield the address of an element in an array. But, as you can see, compilers can choose to use it for other things, as well...

What do R_X86_64_32S and R_X86_64_64 relocation mean?

Got the following error when I tried to compile a C application in 64-bit FreeBSD:
relocation R_X86_64_32S can not be used when making a shared object; recompile with -fPIC
What is R_X86_64_32S relocation and what is R_X86_64_64?
I've googled about the error, and it's possible causes - It'd be great if anyone could tell what R_X86_64_32S really means.
The R_X86_64_32S and R_X86_64_64 are names of relocation types, for code compiled for the amd64 architecture. You can look all of them up in the amd64 ABI.
According to it, R_X86_64_64 is broken down to:
R_X86_64 - all names are prefixed with this
64 - Direct 64 bit relocation
and R_X86_64_32S to:
R_X86_64 - prefix
32S - truncate value to 32 bits and sign-extend
which basically means "the value of the symbol pointed to by this relocation, plus any addend", in both cases. For R_X86_64_32S the linker then verifies that the generated value sign-extends to the original 64-bit value.
Now, in an executable file, the code and data segments are given a specified virtual base address. The executable code is not shared, and each executable gets its own fresh address space. This means that the compiler knows exactly where the data section will be, and can reference it directly. Libraries, on the other hand, can only know that their data section will be at a specified offset from the base address; the value of that base address can only be known at runtime. Hence, all libraries must be produced with code that can execute no matter where it is put into memory, known as position independent code (or PIC for short).
Now when it comes to resolving your problem, the error message speaks for itself.
For any of this to make sense, you must first:
see a minimal example of relocation: https://stackoverflow.com/a/30507725/895245
understand the basic structure of an ELF file: https://stackoverflow.com/a/30648229/895245
Standards
R_X86_64_64, R_X86_64_32 and R_X86_64_32S are all defined by the System V AMD ABI, which contains the AMD64 specifics of the ELF file format.
They are all possible values for the ELF32_R_TYPE field of a relocation entry, specified in the System V ABI 4.1 (1997) which specifies the architecture neutral parts of the ELF format. That standard only specifies the field, but not it's arch dependant values.
Under 4.4.1 "Relocation Types" we see the summary table:
Name Field Calculation
------------ ------ -----------
R_X86_64_64 word64 A + S
R_X86_64_32 word32 A + S
R_X86_64_32S word32 A + S
We will explain this table later.
And the note:
The R_X86_64_32 and R_X86_64_32S relocations truncate the computed value to 32-bits. The linker must verify that the generated value for the R_X86_64_32 (R_X86_64_32S) relocation zero-extends (sign-extends) to the original 64-bit value.
Example of R_X86_64_64 and R_X86_64_32
Let's first look into R_X86_64_64 and R_X86_64_32:
.section .text
/* Both a and b contain the address of s. */
a: .long s
b: .quad s
s:
Then:
as --64 -o main.o main.S
objdump -dzr main.o
Contains:
0000000000000000 <a>:
0: 00 00 add %al,(%rax)
0: R_X86_64_32 .text+0xc
2: 00 00 add %al,(%rax)
0000000000000004 <b>:
4: 00 00 add %al,(%rax)
4: R_X86_64_64 .text+0xc
6: 00 00 add %al,(%rax)
8: 00 00 add %al,(%rax)
a: 00 00 add %al,(%rax)
Tested on Ubuntu 14.04, Binutils 2.24.
Ignore the disassembly for now (which is meaningless since this is data), and look only to the labels, bytes and relocations.
The first relocation:
0: R_X86_64_32 .text+0xc
Which means:
0: acts on byte 0 (label a)
R_X86_64_: prefix used by all relocation types of the AMD64 system V ABI
32: the 64-bit address of the label s is truncated to a 32 bit address because we only specified a .long (4 bytes)
.text: we are on the .text section
0xc: this is the addend, which is a field of the relocation entry
The address of the relocation is calculated as:
A + S
Where:
A: the addend, here 0xC
S: the value of the symbol before relocation, here 00 00 00 00 == 0
Therefore, after relocation, the new address will be 0xC == 12 bytes into the .text section.
This is exactly what we expect, since s comes after a .long (4 bytes) and a .quad (8 bytes).
R_X86_64_64 is analogous, but simpler, since here there is no need to truncate the address of s. This is indicated by the standard through word64 instead of word32 on the Field column.
R_X86_64_32S vs R_X86_64_32
The difference between R_X86_64_32S vs R_X86_64_32 is when the linker will complain "with relocation truncated to fit":
32: complains if the truncated after relocation value does not zero extend the old value, i.e. the truncated bytes must be zero:
E.g.: FF FF FF FF 80 00 00 00 to 80 00 00 00 generates a complaint because FF FF FF FF is not zero.
32S: complains if the truncated after relocation value does not sign extend the old value.
E.g.: FF FF FF FF 80 00 00 00 to 80 00 00 00 is fine, because the last bit of 80 00 00 00 and the truncated bits are all 1.
See also: What does this GCC error "... relocation truncated to fit..." mean?
R_X86_64_32S can be generated with:
.section .text
.global _start
_start:
mov s, %eax
s:
Then:
as --64 -o main.o main.S
objdump -dzr main.o
Gives:
0000000000000000 <_start>:
0: 8b 04 25 00 00 00 00 mov 0x0,%eax
3: R_X86_64_32S .text+0x7
Now we can observe the "relocation" truncated to fit on 32S with a linker script:
SECTIONS
{
. = 0xFFFFFFFF80000000;
.text :
{
*(*)
}
}
Now:
ld -Tlink.ld a.o
Is fine, because: 0xFFFFFFFF80000000 gets truncated into 80000000, which is a sign extension.
But if we change the linker script to:
. = 0xFFFF0FFF80000000;
It now generates the error, because that 0 made it not be a sign extension anymore.
Rationale for using 32S for memory access but 32 for immediates: When is it better for an assembler to use sign extended relocation like R_X86_64_32S instead of zero extension like R_X86_64_32?
R_X86_64_32S and PIE (position independent executables
R_X86_64_32S cannot be used in position independent executables, e.g. done with gcc -pie, otherwise link fails with:
relocation R_X86_64_32S against `.text' can not be used when making a PIE object; recompile with -fPIC
l
I have provided a minimal example explaining it at: What is the -fPIE option for position-independent executables in gcc and ld?
That means that compiled a shared object without using -fPIC flag as you should:
gcc -shared foo.c -o libfoo.so # Wrong
You need to call
gcc -shared -fPIC foo.c -o libfoo.so # Right
Under ELF platform (Linux) shared objects are compiled with position independent code - code that can run from any location in memory, if this flag is not given, the code that is generated is position dependent, so it is not possible to use this shared object.
I ran into this problem and found this answer didn't help me. I was trying to link a static library together with a shared library. I also investigated putting the -fPIC switch earlier on the command line (as advised in answers elsewhere).
The only thing that fixed the problem, for me, was changing the static library to shared. I suspect the error message about -fPIC can happen due to a number of causes but fundamentally what you want to look at is how your libraries are being built, and be suspicious of libraries that are being built in different ways.
In my case the issue arose because the program to compile expected to find shared libraries in a remote directory, while only the corresponding static libraries were there in a mistake.
Actually, this relocation error was a file-not-found error in disguise.
I have detailed how I coped with it in this other thread https://stackoverflow.com/a/42388145/5459638
The above answer demonstrates what these relocations are, and I found building x86_64 objects with GCC -mcmodel=large flag can prevent R_X86_64_32S because the compiler has no assumption on the relocated address in this model.
In the following case:
extern int myarr[];
int test(int i)
{
return myarr[i];
}
Built with gcc -O2 -fno-pie -c test_array.c and disassemble with objdump -drz test_array.o, we have:
0: 48 63 ff movslq %edi,%rdi
3: 8b 04 bd 00 00 00 00 mov 0x0(,%rdi,4),%eax
6: R_X86_64_32S myarr
a: c3 ret
With -mcmodel=large, i.e. gcc -mcmodel=large -O2 -fno-pie -c test_array.c, we have:
0: 48 b8 00 00 00 00 00 movabs $0x0,%rax
7: 00 00 00
2: R_X86_64_64 myarr
a: 48 63 ff movslq %edi,%rdi
d: 8b 04 b8 mov (%rax,%rdi,4),%eax
10: c3 ret

Resources