-fasm-blocks: __asm { call fptr } does not use RIP-relative addressing

[dmiller423]

clang version 17.0.4 Target: x86_64-pc-windows-msvc Thread model: posix InstalledDir: C:\dev\llvm\bin

Target is a simple static library, contents: two functions one __weak_symbol attribute((naked)) containing assembly in intel format that calls a static global address in the same file (set by the second function).

The generated objects are not position independent code, using R_X86_64_32S for relocations of the functions containing asm blocks declared as weak.

tested TARGET list { x86_64-pc-freebsd12-elf, x86_64-pc-unknown-elf, x86_64-pc-linux-gnu }

--
reproduction:
clang-17.0.4 --target=%TARGET% --sysroot=C:/tc_sysroot/%TARGET%/ -fPIC -fPIE -fasm-blocks -MD -MT syscalls.c.o -MF syscalls.c.o.d -o syscalls.c.o -c syscalls.c

objects relocations were tested via readelf and custom tooling, all cases produce:

> readelf -r syscalls.c.o

Relocation section '.rela.text' at offset 0x120 contains 2 entries:
    Offset             Info             Type               Symbol's Value  Symbol's Name + Addend
000000000000000f  0000000300000002 R_X86_64_PC32          0000000000000000 .bss - 4
000000000000002f  000000030000000b R_X86_64_32S           0000000000000000 .bss + 0

rebuilding the same file/params with a previous version (clang-15.0.7):

> readelf -r syscalls.c.o

Relocation section '.rela.text' at offset 0x128 contains 2 entries:
    Offset             Info             Type               Symbol's Value  Symbol's Name + Addend
000000000000000f  0000000300000002 R_X86_64_PC32          0000000000000000 .bss - 4
000000000000002e  0000000300000002 R_X86_64_PC32          0000000000000000 .bss - 4

source:

#include <stdint.h>
#include <sys/types.h>

static uint64_t syscall_addr = 0;

int syscalls_init(uint64_t addr) {
    syscall_addr = addr;
    return 0;
}

__weak_symbol 
__attribute__((naked)) 
ssize_t read(int fd, void* buf, size_t nbytes)  {
    __asm {
            push rbp
            mov rbp, rsp
            mov eax, 3
            mov r10, rcx
            call syscall_addr
            pop rbp
            ret
    }
}

[llvmbot]

@llvm/issue-subscribers-clang-codegen

Author: David Miller (dmiller423)

`` clang version 17.0.4 Target: x86_64-pc-windows-msvc Thread model: posix InstalledDir: C:\dev\llvm\bin``

Target is a simple static library, contents: two functions one __weak_symbol attribute((naked)) containing assembly in intel format that calls a static global address in the same file (set by the second function).

The generated objects are not position independent code, using R_X86_64_32S for relocations of the functions containing asm blocks declared as weak.

tested TARGET list { x86_64-pc-freebsd12-elf, x86_64-pc-unknown-elf, x86_64-pc-linux-gnu }

--
reproduction:
clang-17.0.4 --target=%TARGET% --sysroot=C:/tc_sysroot/%TARGET%/ -fPIC -fPIE -fasm-blocks -MD -MT syscalls.c.o -MF syscalls.c.o.d -o syscalls.c.o -c syscalls.c

objects relocations were tested via readelf and custom tooling, all cases produce:

> readelf -r syscalls.c.o

Relocation section '.rela.text' at offset 0x120 contains 2 entries:
    Offset             Info             Type               Symbol's Value  Symbol's Name + Addend
000000000000000f  0000000300000002 R_X86_64_PC32          0000000000000000 .bss - 4
000000000000002f  000000030000000b R_X86_64_32S           0000000000000000 .bss + 0

rebuilding the same file/params with a previous version (clang-15.0.7):

> readelf -r syscalls.c.o

Relocation section '.rela.text' at offset 0x128 contains 2 entries:
    Offset             Info             Type               Symbol's Value  Symbol's Name + Addend
000000000000000f  0000000300000002 R_X86_64_PC32          0000000000000000 .bss - 4
000000000000002e  0000000300000002 R_X86_64_PC32          0000000000000000 .bss - 4

source:

#include <stdint.h>
#include <sys/types.h>

static uint64_t syscall_addr = 0;

int syscalls_init(uint64_t addr) {
    syscall_addr = addr;
    return 0;
}

__weak_symbol 
__attribute__((naked)) 
ssize_t read(int fd, void* buf, size_t nbytes)  {
    __asm {
            push rbp
            mov rbp, rsp
            mov eax, 3
            mov r10, rcx
            call syscall_addr
            pop rbp
            ret
    }
}

[efriedma-quic]

Related to https://reviews.llvm.org/D151863? CC @MaskRay @KanRobert @akshaykhadse

[MaskRay]

Adding back BaseReg = BaseReg ? BaseReg : 1; will produce callq *syscall_addr(%rip) instead of callq *syscall_addr. But that would break clang/test/CodeGen/ms-inline-asm{,-64}.c fixed by https://reviews.llvm.org/D151863

[llvmbot]

@llvm/issue-subscribers-backend-x86

Author: David Miller (dmiller423)

`` clang version 17.0.4 Target: x86_64-pc-windows-msvc Thread model: posix InstalledDir: C:\dev\llvm\bin``

Target is a simple static library, contents: two functions one __weak_symbol attribute((naked)) containing assembly in intel format that calls a static global address in the same file (set by the second function).

The generated objects are not position independent code, using R_X86_64_32S for relocations of the functions containing asm blocks declared as weak.

tested TARGET list { x86_64-pc-freebsd12-elf, x86_64-pc-unknown-elf, x86_64-pc-linux-gnu }

--
reproduction:
clang-17.0.4 --target=%TARGET% --sysroot=C:/tc_sysroot/%TARGET%/ -fPIC -fPIE -fasm-blocks -MD -MT syscalls.c.o -MF syscalls.c.o.d -o syscalls.c.o -c syscalls.c

objects relocations were tested via readelf and custom tooling, all cases produce:

> readelf -r syscalls.c.o

Relocation section '.rela.text' at offset 0x120 contains 2 entries:
    Offset             Info             Type               Symbol's Value  Symbol's Name + Addend
000000000000000f  0000000300000002 R_X86_64_PC32          0000000000000000 .bss - 4
000000000000002f  000000030000000b R_X86_64_32S           0000000000000000 .bss + 0

rebuilding the same file/params with a previous version (clang-15.0.7):

> readelf -r syscalls.c.o

Relocation section '.rela.text' at offset 0x128 contains 2 entries:
    Offset             Info             Type               Symbol's Value  Symbol's Name + Addend
000000000000000f  0000000300000002 R_X86_64_PC32          0000000000000000 .bss - 4
000000000000002e  0000000300000002 R_X86_64_PC32          0000000000000000 .bss - 4

source:

#include <stdint.h>
#include <sys/types.h>

static uint64_t syscall_addr = 0;

int syscalls_init(uint64_t addr) {
    syscall_addr = addr;
    return 0;
}

__weak_symbol 
__attribute__((naked)) 
ssize_t read(int fd, void* buf, size_t nbytes)  {
    __asm {
            push rbp
            mov rbp, rsp
            mov eax, 3
            mov r10, rcx
            call syscall_addr
            pop rbp
            ret
    }
}