[BUG Report] -dead_strip, strips prefix data unconditionally on macOS

Project Infrastructure LLVM Dev List Archives
1

Hi,

I just came across a rather annoying behavior with llvm 3.9. Assuming the following
samle code in test.ll:

; Lets have some global int x = 4
@x = global i32 10, align 4
; and two strings "p = %d\n" for the prefix data,
; as well as "x = %d\n" to print the (global) x value.
@.str = private unnamed_addr constant [8 x i8] c"x = %d\0A\00", align 1
@.str2 = private unnamed_addr constant [8 x i8] c"p = %d\0A\00", align 1

; declare printf, we'll use this later for printf style debugging.
declare i32 @printf(i8*, ...)

; define a main function.
define i32 @main() prefix i32 123 {
  ; obtain a i32 pointer to the main function.
  ; the prefix data is right before that pointer.
  %main = bitcast i32 ()* @main to i32*

  ; use the gep, to cmpute the start of the prefix data.
  %prefix_ptr = getelementptr inbounds i32, i32* %main, i32 -1
  ; and load it.
  %prefix_val = load i32, i32* %prefix_ptr

  ; print that value.
  %ret = call i32 (i8*, ...) @printf(i8* getelementptr inbounds ([8 x i8], [8 x i8]* @.str2, i32 0, i32 0), i32 %prefix_val)

  ; similarly let's do the same with the global x.
  %1 = alloca i32, align 4
  store i32 0, i32* %1, align 4
  %2 = load i32, i32* @x, align 4
  %3 = call i32 (i8*, ...) @printf(i8* getelementptr inbounds ([8 x i8], [8 x i8]* @.str, i32 0, i32 0), i32 %2)
  ret i32 0
}

gives the following result (expected)

   $ clang test.ll
   $ ./a.out
   p = 123
   x = 10

however, with -dead_strip on macOS, we see the following:

   $ clang test.ll -dead_strip
   $ ./a.out
   p = 0
   x = 10

Thus I believe we are incorrectly stripping prefix data when linking with -dead_strip on macOS.

As I do not have a bugzilla account, and hence cannot post this as a proper bug report.

Cheers,
Moritz

2

AFAIK, this cannot actually work on Apple platforms, because its object file format (Mach-O) doesn’t use sections to determine the ranges of code/data to keep together, but instead infers boundaries based on the range between global symbols in the symbol table.

So, the symbol pointing to the beginning of @main necessarily makes that be a section boundary.

I think the best that could be done in LLVM is to not emit the “.subsections_via_symbols” asm directive (effectively disabling dead stripping on that object) if any prefix data exists. Currently it emits that flag unconditionally for MachO.

3

That is in theory what omitting the .subsections_via_symbols directive is supposed to do, but in an experiment I ran a year or two ago I found that the Mach-O linker was still dead stripping on symbol boundaries with this directive omitted.

In any case, a more precise approach has more recently (~a few months ago) become possible. There is a relatively new asm directive called .altentry that, as I understand it, tells the linker to disregard a given symbol as a section boundary (LLVM already uses this for aliases pointing into the middle of a global). So what you would do is to use .altentry on the function symbol, with an internal symbol appearing before the prefix data to ensure that it is not considered part of the body of the previous function.

Peter

4

Even disassembling dead-code stripping, how can you guarantee that the linker won’t change the order in which atoms are laid out?

5

Hi James,

this is the route we just went. We simply post process the generated assembly and strip out the
.subsections_via_symbols. Quite ugly, I must say. However we did some post asm mangling already,
so this was rather straight forward to implement: https://phabricator.haskell.org/D3287

However the drawback is that on any mach-o platform, you now don’t get the benefit of -dead_code
stripping if you use prefix data. And this might for, say iOS, not be such a great deal.

Cheers,
  Moritz

6

Hi Peter,

I’ve just experimented with this a bit:

Say we would end up with the following assembly:

  .section __TEXT,__text
  .globl _main

  .long 1
_main:
  inc %eax
  ret

  .globl _main.dsp
  .alt_entry _main.dsp
_main.dsp = _main-4

  .subsections_via_symbols

(e.g. we inject the .alt_entry after the fact, pointing to the start of the prefix data)

this will yield:

$ clang test.s -dead_strip
ld: warning: N_ALT_ENTRY bit set on first atom in section __TEXT/__text

And the prefix data will be stripped again.

E.g. what you end up getting is:

$ otool -vVtdj a.out
a.out:
_main:
0000000100000fb5 ff c0 incl %eax
0000000100000fb7 c3 retq

instead of what we’d like to get:

otool -vVtdj a.out
a.out:
_main.dsp:
0000000100000fb1 01 00 addl %eax, (%rax)
0000000100000fb3 00 00 addb %al, (%rax)
_main:
0000000100000fb5 ff c0 incl %eax
0000000100000fb7 c3 retq

.alt_entry’s are not dead_strip protected, and this makes sense I guess, as if the alt_entry is never
actually called visibly from anywhere, it’s probably not needed. However there is the .no_daed_strip
directive. Thus if we graft this slightly different:

  .section __TEXT,__text
  .globl _main

  .long 1
_main:
  inc %eax
  ret

  .no_dead_strip _main.dsp
  .alt_entry _main.dsp
_main.dsp = _main-4

  .subsections_via_symbols

we still get a warning, but it won’t get stripped. At that point however, we don’t need the .alt_entry
anymore (and can drop the warning).

Thus, I’d propose that for functions with prefix_data, a second symbol with .no_dead_strip is emitted
for the prefix data entry point.

Cheers,
Moritz

7

Hi Peter,

I’ve just experimented with this a bit:

Say we would end up with the following assembly:

  .section __TEXT,__text
  .globl _main

  .long 1
_main:
  inc %eax
  ret

  .globl _main.dsp
  .alt_entry _main.dsp

What happens if you try ".alt_entry _main" instead? The alt_entry is
supposed to be bound to the atom appearing *before* it.

_main.dsp = _main-4

  .subsections_via_symbols

(e.g. we inject the .alt_entry after the fact, pointing to the start of
the prefix data)

this will yield:

$ clang test.s -dead_strip
ld: warning: N_ALT_ENTRY bit set on first atom in section __TEXT/__text

And the prefix data will be stripped again.

E.g. what you end up getting is:

$ otool -vVtdj a.out
a.out:
_main:
0000000100000fb5 ff c0 incl %eax
0000000100000fb7 c3 retq

instead of what we’d like to get:

otool -vVtdj a.out
a.out:
_main.dsp:
0000000100000fb1 01 00 addl %eax, (%rax)
0000000100000fb3 00 00 addb %al, (%rax)
_main:
0000000100000fb5 ff c0 incl %eax
0000000100000fb7 c3 retq

.alt_entry’s are not dead_strip protected, and this makes sense I guess,
as if the alt_entry is never
actually called visibly from anywhere, it’s probably not needed. However
there is the .no_daed_strip

directive. Thus if we graft this slightly different:

  .section __TEXT,__text
  .globl _main

  .long 1
_main:
  inc %eax
  ret

  .no_dead_strip _main.dsp
  .alt_entry _main.dsp
_main.dsp = _main-4

  .subsections_via_symbols

we still get a warning, but it won’t get stripped. At that point however,
we don’t need the .alt_entry
anymore (and can drop the warning).

Thus, I’d propose that for functions with prefix_data, a second symbol
with .no_dead_strip is emitted
for the prefix data entry point.

I don't think that is sufficient. I believe that the linker is allowed to
move the function away from the prefix data even if the function is not
dead stripped.

Peter

8

Thank you Peter!

That seems to do the trick!

$ cat test.s
  .section __TEXT,__text
  .globl _main

  .long 1
_main:
  inc %eax
  ret

  .alt_entry _main
_main.dsp = _main-4

  .subsections_via_symbols

$ clang test.s -dead_strip
$ otool -vVtdj a.out
a.out:
_main.dsp:
0000000100000fb1 01 00 addl %eax, (%rax)
0000000100000fb3 00 00 addb %al, (%rax)
_main:
0000000100000fb5 ff c0 incl %eax
0000000100000fb7 c3 retq

However, now I need to figure out how to generate this from
llvm via an alias.

@.str = private unnamed_addr constant [8 x i8] c"p = %d\0A\00", align 1

%struct.prefix = type { i32 }
@main.dsp = alias i32, i32* getelementptr (%struct.prefix, %struct.prefix* bitcast (i32 ()* @main to %struct.prefix*), i32 -1, i32 0)

declare i32 @printf(i8*, ...)

define i32 @main() prefix %struct.prefix { i32 123 } {
  %main = bitcast i32 ()* @main to i32*
  %prefix_ptr = getelementptr inbounds i32, i32* %main, i32 -1
  %prefix_val = load i32, i32* %prefix_ptr
  %ret = call i32 (i8*, ...) @printf(i8* getelementptr inbounds ([8 x i8], [8 x i8]* @.str, i32 0, i32 0), i32 %prefix_val)
  ret i32 0
}

generates the following for me:

        .section __TEXT,__text,regular,pure_instructions
        .macosx_version_min 10, 12
        .globl _main
        .p2align 4, 0x90
        .long 123 ## @main
                                        ## 0x7b
_main:
        .cfi_startproc
## BB#0:
        pushq %rax
Ltmp0:
        .cfi_def_cfa_offset 16
        movl _main-4(%rip), %esi
        leaq L_.str(%rip), %rdi
        xorl %eax, %eax
        callq _printf
        xorl %eax, %eax
        popq %rcx
        retq
        .cfi_endproc

        .section __TEXT,__cstring,cstring_literals
L_.str: ## @.str
        .asciz "p = %d\n"

        .globl _main.dsp
        .alt_entry _main.dsp
_main.dsp = _main-4
.subsections_via_symbols

any ideas how to get the .alt_entry right?

Cheers,
Moritz

9

Firstly, do you need “main.dsp” defined as an external symbol, or can all external references go via “main”? If the answer is the latter, that will make the solution simpler.

If only the latter, you will need to make a change to LLVM here: http://llvm-cs.pcc.me.uk/lib/CodeGen/AsmPrinter/AsmPrinter.cpp#650

Basically you would need to add a hook to the TargetLoweringObjectFile class that allows the object format to control how prefix data is emitted. For Mach-O you would emit a label for a dummy internal symbol, followed by the prefix data and then an alt_entry directive for the function symbol. All other object formats would just emit the prefix data.

Peter

10

Peter,

thanks again! Yes, we only need to refer to main, but we must ensure that
the prefix data is not stripped.

I’ll have a look at the AsmPrinter.

Another idea that came to mind is abusing the prologue data. And simply
injecting the prefix data into the prologue data. Then adding the *real*
entry_point as an alt_entry after the prologue data.

Right now we have.

.- - - - -. <- main.dsp

Prefix |
- - - - -| <- main
Body |

'- - - - -'

with Prologue, I believe:

.- - - - - -. <- main

Prologue |
- - - - - -| <- alt_entry main_alt
Body |

'- - - - - -'

This could probably work today. However prologue data says it needs a special
format. Do you happen to know, if that format is important as well, if we
would never ever go through main, but only through main_alt?

Cheers,
Moritz

11

I suspect that the format isn’t important if you do that, but I wouldn’t recommend it, at least because inlining (and other inter-procedural optimizations) are not expected to work correctly if you produce IR like that.

Peter

12

Hi Peter,

we though that might cause troubles as well. My current interim solution looks
like the following:

@.str = private unnamed_addr constant [8 x i8] c"p = %d\0A\00", align 1

%struct.prefix = type { i32 }
declare i32 @printf(i8*, ...)

define i32 @main() prefix %struct.prefix { i32 123 } {
  %main = bitcast i32 ()* @main to i32*
  %prefix_ptr = getelementptr inbounds i32, i32* %main, i32 -1
  %prefix_val = load i32, i32* %prefix_ptr
  %ret = call i32 (i8*, ...) @printf(i8* getelementptr inbounds ([8 x i8], [8 x i8]* @.str, i32 0, i32 0), i32 %prefix_val)
  ret i32 0
}

@main.dsp_helper = alias i32, i32* getelementptr (%struct.prefix, %struct.prefix* bitcast (i32 ()* @main to %struct.prefix*), i32 -1, i32 0)

module asm ".alt_entry _main"
module asm "_main.dsp = _main.dsp_helper"

This still generated the "ld: warning: N_ALT_ENTRY bit set on first atom in section __TEXT/__text” warning during linking,
but seems to work as expected, and computes the prefix data size through llvm.

Any drawbacks we should be aware of, using this until the prefix data with -dead_strip is fixed in a release built of llvm?

Cheers,
Moritz

13

That will set alt_entry on every atom you define in the text section (dsp_helper becomes alt_entry by virtue of being an alias). I think that would prevent dead stripping in text altogether because every atom will be linked to the previous one.

Peter

14

Oh, that’s great that it’s possible to implement properly, now. Does it actually work for

It’d be cool if LLVM hooked up its generic section handling support to this feature now, so that the only global symbols that didn’t get marked as .alt_entry were those at the beginning of what llvm would consider sections.

Then apple platforms could behave sanely, like all other platforms do, only with -f{function,data}-sections defaulted to on instead of off.

(That is, if you specify -fno-function-sections -fno-data-sections, it could mark nearly everything as .alt_entry – except the first symbol in the object file)

15

(Oops forgot to finish the sentence)

… symbols in data sections, as well as text?

16

Oh, that's great that it's possible to implement properly, now. Does it
actually work for

(Oops forgot to finish the sentence)

... symbols in data sections, as well as text?

I can't see anything in the linker that is specific to text. At some point
I want to try removing some of the Darwin-specific hacks I have in the
LowerTypeTests pass, which is already creating multiple symbols pointing
into the same global on every other platform.

It'd be cool if LLVM hooked up its generic section handling support to
this feature now, so that the only global symbols that *didn't* get marked
as .alt_entry were those at the beginning of what llvm would consider
sections.

Then apple platforms could behave sanely, like all other platforms do,
only with -f{function,data}-sections defaulted to on instead of off.

(That is, if you specify -fno-function-sections -fno-data-sections, it
could mark nearly *everything* as .alt_entry -- except the first symbol in
the object file)

That sounds like a great idea to me!

Peter

17

What is the advantage of not using -f{function,data}-sections? (i.e. what isn’t sane about it?)

I’m asking because I was told the only reason not to use it all the time on ELF is that it makes intermediate object files larger.

Thanks,

18

My point was more that it’d be great if symbols and sections and dead-code stripping, in general, would work more like they do elsewhere. ISTM that’d make things cleaner generally. No special case for offset aliases would be needed, nor for features like prefix data. They’d just work by virtue of generic handling for llvm sections turning into the appropriate symbol flags. (It’s a shame that defining the start of a subsection wasn’t made opt-IN for a symbol in macho, but oh well.)

Now, one of the consequences of hooking up the infrastructure that was is that -fno-{function,data}-sections would automatically work. But I was thinking of that independently of whether a user might want to actually specify that.

However – there is actually another possible reason to disable it: when it’s disabled, you avoid inserting relocations for references to local data and functions. That can sometimes allow the use of shorter instructions to reference nearby data – so, in fact, it’s possible for -f{function,data}-sections to make your FINAL binary larger than it would be without.

19

I mean, yes, it would be great if MachO were a better object file format. But it has a limited number of sections, so .subsections_via_symbols and atoms work around that problem. The only way to do better in the long term is to use a better object file format, either new or existing. Every so often somebody takes a look at that, but so far Apple has continued to extend MachO as necessary to support their use cases.

20

No – that’s the whole topic of discussion here… With the appropriate automatic emission of alt_entry symbol flags in llvm, it seems that you ought to be able to get the desired behaviors even with the MachO object file format being what it is.