[ELF][LTO] Add baseline test for invalid relocations against runtime calls by arichardson · Pull Request #127286 · llvm/llvm-project (original) (raw)

@llvm/pr-subscribers-lld-elf

@llvm/pr-subscribers-lld

Author: Alexander Richardson (arichardson)

Changes

This can happen when using a LTO build of compiler-rt for ARM and the
program uses 64-bit division.
The 64-bit division function in compiler-rt (__aeabi_ldivmod) is written
in assembly and calls the C function __divmoddi4, which works fine normally.
However, when building with LTO the call inside __aeabi_ldivmod is replaced
with a jump to address zero, which then crashes the program.

Building with -pie generates an error instead of a jump to address zero,
and surprisingly just declaring the __aeabi_ldivmod function (but not
calling it) in the input IR also avoids this issue.

Full diff: https://github.com/llvm/llvm-project/pull/127286.diff

1 Files Affected:

diff --git a/lld/test/ELF/lto/arm-rtlibcall.ll b/lld/test/ELF/lto/arm-rtlibcall.ll new file mode 100644 index 0000000000000..89b0a1eb0e90e --- /dev/null +++ b/lld/test/ELF/lto/arm-rtlibcall.ll @@ -0,0 +1,122 @@ +; REQUIRES: arm +;; https://github.com/llvm/llvm-project/issues/127284 +;; Test for LTO optimizing out references to symbols that are pulled in by +;; compiler-generated libcalls (post LTO). +;; The problem here is that the call to __aeabi_ldivmod is generated post-LTO, +;; during ISel, so aeabi_ldivmod.o is only marked as required afterwards but by +;; that time we have decided that all the callees of __aeabi_ldivmod are not +;; needed and have been marked as ABS zero symbols. +; RUN: rm -rf %t && split-file %s %t && cd %t +; RUN: llvm-as divmoddi4.ll -o divmoddi4.bc +; RUN: llvm-mc -filetype=obj -triple=armv7-none-unknown-eabi aeabi_ldivmod.s -o aeabi_ldivmod.o +;; With an explicit __aebi_ldivmod call in the input IR this works as expected: +; RUN: llvm-as main-explicit.ll -o main-explicit-ldivmod.bc +; RUN: ld.lld main-explicit-ldivmod.bc --start-lib aeabi_ldivmod.o divmoddi4.bc --end-lib -o test.exe -Bstatic +; RUN: llvm-objdump -d -r -t test.exe | FileCheck %s --check-prefix=GOOD-DUMP +; GOOD-DUMP-LABEL: SYMBOL TABLE: +; GOOD-DUMP: [[#]] g F .text [[#]] _start +; GOOD-DUMP: [[#]] g F .text 00000024 __aeabi_ldivmod +; GOOD-DUMP: [[#]] g F .text [[#]] __divmoddi4 +; GOOD-DUMP-LABEL: <__aeabi_ldivmod>: +; GOOD-DUMP: bl 0x20140 <__divmoddi4> @ imm = #0x28 + +; But if the call is generated by ISel, we end up with an invalid reference: +; RUN: llvm-as main-implicit.ll -o main-implicit-ldivmod.bc +; RUN: ld.lld main-implicit-ldivmod.bc --start-lib aeabi_ldivmod.o divmoddi4.bc --end-lib -o test.exe -Bstatic +; RUN: llvm-objdump -d -r -t test.exe | FileCheck %s --check-prefix=BAD-DUMP +;; We jump to address zero here and __divmoddi4 ends up being an absolute symbol: +; BAD-DUMP-LABEL: SYMBOL TABLE: +; BAD-DUMP: [[#]] g F .text [[#]] _start +; BAD-DUMP: [[#]] g F .text 00000024 __aeabi_ldivmod +; BAD-DUMP: [[#]] g ABS 00000000 __divmoddi4 +; BAD-DUMP-LABEL: <__aeabi_ldivmod>: +; BAD-DUMP: bl 0x0 <__divmoddi4> @ imm = #-0x200fc +;; Linking with -pie complains about the invalid relocation (and even points back to the source files) +; RUN: not ld.lld main-implicit-ldivmod.bc --start-lib aeabi_ldivmod.o divmoddi4.bc --end-lib -o test.exe --no-undefined -pie --no-dynamic-linker 2>&1 | FileCheck %s --check-prefix=PIE-ERROR +PIE-ERROR: ld.lld: error: relocation R_ARM_CALL cannot refer to absolute symbol: __divmoddi4 +PIE-ERROR-NEXT: >>> defined in divmoddi4.bc +PIE-ERROR-NEXT: >>> referenced by aeabi_ldivmod.o:(__aeabi_ldivmod) + +;; Interestingly, just declaring __aeabi_ldivmod is sufficient to not run into this issue. +; RUN: llvm-as main-declared.ll -o main-declared-ldivmod.bc +; RUN: ld.lld main-declared-ldivmod.bc --start-lib aeabi_ldivmod.o divmoddi4.bc --end-lib -o test.exe -Bstatic +; RUN: llvm-objdump -d -r -t test.exe | FileCheck %s --check-prefix=GOOD-DUMP + +;--- divmoddi4.ll +target datalayout = "e-m:e-p:32:32-Fi8-i64:64-v128:64:128-a:0:32-n32-S64" +target triple = "armv7-none-unknown-eabi" + +; Adding it to llvm.used does not appears to have any effect! +; @llvm.used = appending global [1 x ptr] [ptr @__divmoddi4], section "llvm.metadata" + +; Stub version of the real __divmoddi4 +define i64 @__divmoddi4(i64 %a, i64 %b, ptr writeonly %rem) #0 align 32 { +entry:

+.size __aeabi_ldivmod, . - __aeabi_ldivmod + +;--- main-implicit.ll +target datalayout = "e-m:e-p:32:32-Fi8-i64:64-v128:64:128-a:0:32-n32-S64" +target triple = "armv7-none-unknown-eabi" + +define dso_local i64 @_start(i64 %num, i64 %denom) local_unnamed_addr #0 { +entry: