Support z separate/no-separate code

[quic-areg]

-z separate-code puts the text segment in its
own pages disjoint from any other segments.

[quic-seaswara]

Can we import the tests from lld for this as part of this PR ?

[quic-seaswara]

Do we have other options that are not compatible with -z seperate-code ?

How does using PHDR's, MEMORY, linker scripts, Adding LMA addresses, Using linker special directives (BYTE, QUAD, SQUAD, etc), orphan section placements work with this option ? Can you explore more with using the bfd linker ?

How does -z relro work this option ?

We might need to add a FAQ also for cases that people might want to consider this option for image size constraints, adding a new segment can cause increase in image size for example.

These analysis could be seperate bugs, and it might deserve an epic as applicable.

Does these -z options translate to reproduce command line option appropriately ?

Can we document incompatibilities ? and also document this option in detail.

I would prefer a test that can build and run in addition to a lit unit test.

You can decide to make some of these as follow ups with some discretion.

[partaror]

This option is affecting all the targets, however, the test is added just for x86. Can we add a test in the common/Standalone subdirectory as well?

[quic-seaswara]

I took some time to write test cases, it looks like there is some behavioral difference with these switches with linker script, omagic support, and shared library support.

@quic-areg can you please look into this usage, though we dont need to copy exact bfd behavior in cases where the linker gets confused and creates a segment that mixes R/RW/RWX.

Please see this test and let me know how we plan to address this both in diagnostics and other comments we have raised.

cat > 1.c << \!
int bss[10] = { 0 };
const int mydata[] = { 1, 2, 3, 4};
int foo() { return 0; }
int bar() { return 0; }
int baz() { return 0; }
!

cat > script.t << \!
SECTIONS {
  .foo : { *(.text.foo) }
  .rodata : { *(.rodata*) }
  .data : { *(.data*) }
  .bar : { *(.text.bar) }
  .baz : { *(.text.baz) }
  .bss : { *(.bss.*) }
}
!

clang -c -target x86_64 1.c -fPIC
echo "****************************************"
echo "No linker script ....."
echo "****************************************"
ld.bfd 1.o -z separate-code -o nols.1
readelf -l -W nols.1
echo "****************************************"
echo "No linker script ....omagic."
echo "****************************************"
ld.bfd 1.o -z separate-code -o nols.omagic -N
readelf -l -W nols.omagic
echo "****************************************"
echo "No linker script ....partial link."
echo "****************************************"
ld.bfd 1.o -z separate-code -o nols.partiallink -r
readelf -S -W nols.partiallink
echo "****************************************"
echo "No linker script ....shared library"
echo "****************************************"
ld.bfd 1.o -z separate-code -shared -o nols.shared.1
readelf -l -W nols.shared.1
echo "****************************************"
echo "No linker script ....."
echo "****************************************"
ld.bfd 1.o -z no-separate-code -o nols.no.1
readelf -l -W nols.no.1
echo "****************************************"
echo "No linker script ....omagic."
echo "****************************************"
ld.bfd 1.o -z no-separate-code -o nols.no.omagic -N
readelf -l -W nols.no.omagic
echo "****************************************"
echo "No linker script ....shared library"
echo "****************************************"
ld.bfd 1.o -z no-separate-code -shared -o nols.no.shared.1
readelf -l -W nols.no.shared.1
echo "****************************************"
echo "linker script ....."
echo "****************************************"
ld.bfd 1.o -z separate-code -o nols.1  -T script.t
readelf -l -W nols.1
echo "No linker script ....omagic."
ld.bfd 1.o -z separate-code -o nols.omagic -N -T script.t
readelf -l -W nols.omagic
echo "No linker script ....partial link."
ld.bfd 1.o -z separate-code -o nols.partiallink -r -T script.t
readelf -S -W nols.partiallink
echo "No linker script ....shared library"
ld.bfd 1.o -z separate-code -shared -o nols.shared.1 -T script.t
readelf -l -W nols.shared.1
echo "No linker script ....."
ld.bfd 1.o -z no-separate-code -o nols.no.1 -T script.t
readelf -l -W nols.no.1
echo "No linker script ....omagic."
ld.bfd 1.o -z no-separate-code -o nols.no.omagic -N -T script.t
readelf -l -W nols.no.omagic
echo "No linker script ....shared library"
ld.bfd 1.o -z no-separate-code -shared -o nols.no.shared.1 -T script.t
readelf -l -W nols.no.shared.1

[quic-areg]

it looks like there is some behavioral difference with these switches with linker script, omagic support, and shared library support

It is expected that readelf -l would differ with the absence/presence of these options.

mixing -z separate-code with partial linking, omagic, or linker scripts does not make sense. We could warn for these combinations though neither lld nor bfd currently do.

[quic-seaswara]

it looks like there is some behavioral difference with these switches with linker script, omagic support, and shared library support

It is expected that readelf -l would differ with the absence/presence of these options.

mixing -z separate-code with partial linking, omagic, or linker scripts does not make sense. We could warn for these combinations though neither lld nor bfd currently do.

Mixing -z seperate-code with omagic makes the linker choose omagic behavior, if you observe.

For shared libraries -z separate-code is not effective, if you see.

Since -z relro also bumps the page alignment for certain sections when they are seen in the program layout, doesn't it conflict ?

I am not sure why it does not make sense with linker scripts, does -z separate-code option always create aligned segments to page boundaries ?

[quic-areg]

it looks like there is some behavioral difference with these switches with linker script, omagic support, and shared library support

It is expected that readelf -l would differ with the absence/presence of these options.
mixing -z separate-code with partial linking, omagic, or linker scripts does not make sense. We could warn for these combinations though neither lld nor bfd currently do.

Mixing -z seperate-code with omagic makes the linker choose omagic behavior, if you observe.

For shared libraries -z separate-code is not effective, if you see.

Since -z relro also bumps the page alignment for certain sections when they are seen in the program layout, doesn't it conflict ?

I am not sure why it does not make sense with linker scripts, does -z separate-code option always create aligned segments to page boundaries ?

Mixing -z seperate-code with omagic makes the linker choose omagic behavior, if you observe.

Yes I observed this. It makes no sense to request an RWX segment (omagic) along with a separate code segment. The options should not be mixed. both eld,lld,and bfd behave the same way when they are mixed (omagic takes precedence), if you observe:

ld.eld 1.o --omagic -z separate-code --omagic -o e.out --rosegment --no-align-segments 
ld.lld 1.o --omagic -z separate-code --omagic -o l.out  
ld.bfd 1.o --omagic -z separate-code --omagic -o b.out 
llvm-readelf -l e.out l.out b.out

  LOAD           0x000000 0x0000000000400000 0x0000000000400000 0x000138 0x000168 RWE 0x10

  LOAD           0x0000b0 0x00000000002000b0 0x00000000002000b0 0x0000b8 0x0000e8 RWE 0x10

  LOAD           0x0000b0 0x00000000004000b0 0x00000000004000b0 0x0000b8 0x0000e8 RWE 0x10

For shared libraries -z separate-code is not effective, if you see.

I guess I am missing something. Can you show an example?

here's my example, if you see

ld.eld -shared 1.o -z separate-code -o e.out --rosegment --no-align-segments
ld.lld -shared 1.o -z separate-code -o l.out
ld.bfd -shared 1.o -z separate-code -o b.out
llvm-readelf -l e.out l.out b.out

  LOAD           0x000000 0x0000000000000000 0x0000000000000000 0x000190 0x000190 R E 0x1000
  LOAD           0x001000 0x0000000000001000 0x0000000000001000 0x0000e4 0x0000e4 R   0x1000
  LOAD           0x002000 0x0000000000002000 0x0000000000002000 0x000028 0x000028 R E 0x1000
  LOAD           0x003000 0x0000000000003000 0x0000000000003000 0x000088 0x000088 R   0x1000
  LOAD           0x003088 0x0000000000003088 0x0000000000003088 0x0000e0 0x000110 RW  0x1000

  LOAD           0x000000 0x0000000000000000 0x0000000000000000 0x00039c 0x00039c R   0x1000
  LOAD           0x001000 0x0000000000001000 0x0000000000001000 0x000028 0x000028 R E 0x1000
  LOAD           0x002000 0x0000000000002000 0x0000000000002000 0x000070 0x001000 RW  0x1000
  LOAD           0x002070 0x0000000000003070 0x0000000000003070 0x000000 0x000028 RW  0x1000

  LOAD           0x000000 0x0000000000000000 0x0000000000000000 0x0002d8 0x0002d8 R   0x1000
  LOAD           0x001000 0x0000000000001000 0x0000000000001000 0x000028 0x000028 R E 0x1000
  LOAD           0x002000 0x0000000000002000 0x0000000000002000 0x000088 0x000088 R   0x1000
  LOAD           0x002f40 0x0000000000003f40 0x0000000000003f40 0x0000c0 0x0000e8 RW  0x1000

The code segments are all in their own pages; it looks 'effective' to me. The other segments are allowed to overlap.

why are you adding --rosegment and --no-align-segments. can we just do -z separate-code?

eld page-aligns segments by default when there is no linker script present, so the effect of -z separate-code cannot be seen. eld also does not emit a separate ro segment, which the other linkers do. This is important because we need more segments to show that r/rw segments may overlap and that only code segments are aligned. this cannot be seen when there are only two segments, which I believe is the default

Since -z relro also bumps the page alignment for certain sections when they are seen in the program layout, doesn't it conflict ?

No I don't think so. Do you have an example?

I am not sure why it does not make sense with linker scripts, does -z separate-code option always create aligned segments to page boundaries ?

If the user has a linker script then they should own the layout/page alignment decisions. eg we page align segments by default when there is no linker script, but this is not turned on with linker scripts. lld does not honor -z separate-code with linker scripts for example.

does -z separate-code option always create aligned segments to page boundaries ?

starting code segments at page aligned offsets is one way to ensure that they are not co-resident on the same page as any other segment, and that seems to be how the other linkers implement it.