search

riscv-collab / riscv-gnu-toolchain

GNU toolchain for RISC-V, including GCC
Other
3.53k stars 1.16k forks source link

_GLOBAL_OFFSET_TABLE_ in .data sections #377

Closed Pinpiew closed 2 years ago

Pinpiew commented 6 years ago

the result of compile, _GLOBAL_OFFSETTABLE in .data sections

but at startup, the ins "la t0, trap_vector", may be use the _GLOBAL_OFFSETTABLE, but this time .data sections have not load to VMA.

then register t0 get value '0'

disassemble as following: (_ftext is the entry) Idx Name Size VMA LMA File off Algn 0 .text 0000316a 0000000020000000 0000000020000000 00001000 21 CONTENTS, ALLOC, LOAD, READONLY, CODE 1 .rodata 0000085c 0000000020003170 0000000020003170 00004170 23 CONTENTS, ALLOC, LOAD, READONLY, DATA 2 .data 000000f1 0000000020020000 00000000200039cc 00005000 23 CONTENTS, ALLOC, LOAD, DATA 3 .bss 000022c0 0000000020020100 0000000020003abd 00005100 24

0000000020000000 <_ftext>: 20000000: 34001073 csrw mscratch,zero 20000004: 00020297 auipc t0,0x20 20000008: 0742b283 ld t0,116(t0) # 20020078 <_GLOBAL_OFFSETTABLE+0x8> 2000000c: 30529073 csrw mtvec,t0 20000010: 30502373 csrr t1,mtvec 20000014: 00629063 bne t0,t1,20000014 <_ftext+0x14> 20000018: 00020117 auipc sp,0x20 2000001c: 06813103 ld sp,104(sp) # 20020080 <_GLOBAL_OFFSETTABLE+0x10> 20000020: f14026f3 csrr a3,mhartid 20000024: 00168693 addi a3,a3,1 20000028: 00169693 slli a3,a3,0x1 2000002c: 00c69613 slli a2,a3,0xc 20000030: 00c10133 add sp,sp,a2 20000034: 00020617 auipc a2,0x20 20000038: 06463603 ld a2,100(a2) # 20020098 <_GLOBAL_OFFSETTABLE+0x28> 2000003c: 00020697 auipc a3,0x20 20000040: 04c6b683 ld a3,76(a3) # 20020088 <_GLOBAL_OFFSETTABLE+0x18> 20000044: 00020717 auipc a4,0x20 20000048: 04c73703 ld a4,76(a4) # 20020090 <_GLOBAL_OFFSETTABLE+0x20>

jim-wilson commented 6 years ago

This doesn't look like a toolchain problem. It looks like you are writing your own OS, and your OS code is buggy. You need to load .data into memory before trying to read values from .data. You might try looking at other OS source code trees to see how they handle OS startup.

Pinpiew commented 6 years ago

it is a startup code, without OS.

if VMA == LMA， there is no problem, if not, problem may be happened.

if VMA != LMA, first must copy the .data section to VMA, this operation need linker srcipt symbal, but get the symbal address rely on the GLOBAL_OFFSET_TABLE, and then GLOBAL_OFFSET_TABLE is at the .data section. so get linker srcipt symbal address error.

why GLOBAL_OFFSET_TABLE data place at .data section? if move to .rodata section, is it better?

jim-wilson commented 6 years ago

So you are trying to load the address of the data section so you can copy it into memory, but this address was itself placed in the data section? That isn't going to work. You may need to write your code differently, or compile it differently. Though I would expect the data address to be in rodata not data, and rodata should be next to text and hence should already be in memory here. GLOBAL_OFFSET_TABLE is a pointer into the data section, we can't move that. Or maybe you just need a better linker script to make this work.

You can find an example of similar code in sifive/freedom-e-sdk in the hifive1 support where binaries get loaded into rom, and then data is copied to flash at run-time to make it writable.

lazyparser commented 3 years ago

@Pinpiew Hi, would you like to update the status of this issue? Is it safe to close now?

TommyMurphyTM1234 commented 2 years ago

This doesn't look like a toolchain problem. It looks like you are writing your own OS, and your OS code is buggy. You need to load .data into memory before trying to read values from .data. You might try looking at other OS source code trees to see how they handle OS startup.