Incorrect `<section>_end` symbol values due to incorrect handling of orphan sections

[parth-07]

We may have incorrect values of <section>_end symbols due to incorrect handling of orphan sections.

Reproducer:

#!/usr/bin/env bash

cat >1.c <<\EOF
int a[8000] = { 1 };
int b[100] = { 0 };
__attribute__((section(".temp"))) int c = 10;
int main() { return 0; }
EOF

cat >script.t <<\EOF
SECTIONS {
  .text (0x1000) : { *(.text) }
  .temp (0x2000) : { *(.temp) }
  __temp_end = .;
  __bss_start = .;
  . = ALIGN(1K);
  .bss : { *(.bss) }
}
EOF

clang -o 1.o -target x86_64-linux-pc-gnu 1.c -c
LDs=(ld.eld)
SFs=(eld)

for i in "${!SFs[@]}"; do
  ${LDs[$i]} -o 1.${SFs[$i]}.out 1.o -T script.t -Map 1.${SFs[$i]}.map.txt
done

llvm-readelf -S 1.eld.out | grep .temp
llvm-readelf -s 1.eld.out | grep temp_end

Output:

[ 3] .temp             PROGBITS        0000000000002000 002000 000004 00  WA  0   0  4
12: 0000000000009d10     0 NOTYPE  GLOBAL DEFAULT   ABS __temp_end

__temp_end should be 0x2004 instead of 0x9d10.

[quic-seaswara]

We have differences in how we do orphan section handling w.r.t GNU linker. The behavior is very loosely documented on where the placement of orphan sections need to be placed in the layout.

our implementation aligns with placing orphan section that have the parity of section permissions.

When the orphan section is placed section assignments that are moved appropriately.

Please document this behavior.

[quic-seaswara]

We can come back to this issue, if there is a case that indeed breaks that relies on this behavior.

[parth-07]

Another example where the current placement of orphan sections results in unexpected layout:

#!/usr/bin/env bash

cat >1.c <<\EOF
int bar() { return 3; }
int foo() { return bar(); }

int u = 1;
EOF

cat >script.t <<\EOF
SECTIONS {
  . = u;
  .text : {
    . = . + 0xfda;
    *(.text*)
  }
  u = 0x3000;
  DATA_START = DEFINED(DATA_START) ? DATA_START : .;
  . = ALIGN(DATA_START, 0x1000);
  .data : {
    *(.data)
  }
}
EOF

LDs=(ld.eld ld.lld aarch64-none-elf-ld.bfd)
SFs=(eld lld bfd)

clang -o 1.o -target aarch64-unknown-elf 1.c -c -ffunction-sections

for i in "${!SFs[@]}"; do
  ${LDs[$i]} -o 1.${SFs[$i]}.out 1.o -T script.t
done

In the case of both lld and bfd, .data is aligned to 0x1000, whereas, for eld, it is not aligned because of unexpected placement of .ARM.exidx.text.bar and .ARM.exidx.text.foo section in between . = ALIGN(DATA_START, 0x1000) and .data : { ... }