Support fno-plt like gcc

[hstk30-hw]

Hi, I see https://reviews.llvm.org/D39079 add the support for fno-plt in frontend. But has no affect on asm code.

https://discourse.llvm.org/t/bug-fno-plt-has-no-effect/57986

If I want to support fno-plt in clang like gcc, Where should I start?

[hstk30-hw]

CC @tmsri @MaskRay

[tmsri]

Sorry, it's been a while. Trying to understand what you want to achieve here, Is it that you have a PLT call in inline asm that you want to avoid with -fno-plt?

[hstk30-hw]

I want to achieve you proposed in gcc https://gcc.gnu.org/legacy-ml/gcc-patches/2015-05/msg00001.html @tmsri

https://godbolt.org/z/Yeor6TTeE

For now, the fno-plt in clang seem just add NonLazyBind attribute to the function, seem not affect in the downstream.

[MaskRay]

mips has -mno-plt. Some arches such as x86/aarch64 have -fno-plt. in Clang, -fno-plt is only implemented for x86. I can add the AArch64 support.

I have some notes on https://maskray.me/blog/2021-09-19-all-about-procedure-linkage-table#fno-plt . -fno-plt is significantly worse on RISC architectures due to a longer code sequence.

[wzssyqa]

gcc support -mplt only for TARGET_ABSOLUTE_ABICALLS.

#define TARGET_ABSOLUTE_ABICALLS
(TARGET_ABICALLS
&& !TARGET_SHARED
&& TARGET_EXPLICIT_RELOCS
&& !ABI_HAS_64BIT_SYMBOLS)

For normal PIC ELFs, all of them are no-plt.

I think that it makes no sense to support it on LLVM.

[MaskRay]

gcc support -mplt only for TARGET_ABSOLUTE_ABICALLS.

#define TARGET_ABSOLUTE_ABICALLS (TARGET_ABICALLS && !TARGET_SHARED && TARGET_EXPLICIT_RELOCS && !ABI_HAS_64BIT_SYMBOLS)

For normal PIC ELFs, all of them are no-plt.

I think that it makes no sense to support it on LLVM.

mips's situation is different. -mno-plt is generally the default except certain -fno-pic cases that default to -mplt (most o32/n32 and n64 -msym32).
GCC's mips -mplt implementation confuses problematic copy relocations with PLT, which is a flaw but can be fixed.

For most other architectures (probably all supported by LLVM), -fplt is the default and having -fno-plt could be interesting in some cases. The calls will be similar to __dllspec(dllimport) and can have an interesting trade-off for C library functions when dynamically linked.

[Kojoley]

Some arches such as x86/aarch64 have -fno-plt. in Clang, -fno-plt is only implemented for x86.

It doesn't seem to be implemented for 32bit x86 and x86 GlobalISel.

https://godbolt.org/z/aa6jqraPY

void bar(void);

void foo(void)
{
    bar();
}

; clang -m32 -O3 -fpic -fno-plt
foo():                                # @foo()
        push    ebx
        sub     esp, 8
        call    .L0$pb
.L0$pb:
        pop     ebx
.Ltmp0:
        add     ebx, offset _GLOBAL_OFFSET_TABLE_+(.Ltmp0-.L0$pb)
        call    bar()@PLT
        add     esp, 8
        pop     ebx
        ret

; gcc -m32 -O3 -fpic -fno-plt
foo():
        call    __x86.get_pc_thunk.ax
        add     eax, OFFSET FLAT:_GLOBAL_OFFSET_TABLE_
        jmp     [DWORD PTR _Z3barv@GOT[eax]]
__x86.get_pc_thunk.ax:
        mov     eax, DWORD PTR [esp]
        ret

clang -O3 -fpic -fno-plt -mllvm -global-isel
fatal error: error in backend: cannot select: %0:gr64(p0) = G_GLOBAL_VALUE @_Z3barv (in function: _Z3foov)
clang -m32 -O3 -fpic -fno-plt -mllvm -global-isel
fatal error: error in backend: cannot select: %0:gr32(p0) = G_GLOBAL_VALUE @_Z3barv (in function: _Z3foov)

[llvmbot]

@llvm/issue-subscribers-backend-aarch64

Author: None (hstk30-hw)

Hi, I see https://reviews.llvm.org/D39079 add the support for `fno-plt` in frontend. But has no affect on asm code.

https://discourse.llvm.org/t/bug-fno-plt-has-no-effect/57986

If I want to support fno-plt in clang like gcc, Where should I start?

[hstk30-hw]

-fno-plt is significantly worse on RISC architectures due to a longer code sequence.

I test it in AArch64, gets 1.x% improvement, and branch-miss is down.
In AArch64 use 3 Instructions compare with 1 instruction in X86.
Branch-prediction is still the key :)

Thx @MaskRay