blob: c472b6bbf9c6a20ecf940a2361685bd7cced9c7b [file] [log] [blame]
# In this test case, we reproduce the behavior seen in gcc where the
# base address of a jump table is decremented by some number and ends up
# at the exact addess of a jump table from another function. After
# linking, the instruction references another jump table and that
# confuses BOLT.
# We repro here the following issue:
# Before assembler: Instruction operand is: jumptable - 32
# After linking: Instruction operand is: another_jumptable
# REQUIRES: system-linux, asserts
# XFAIL: *
# RUN: llvm-mc -filetype=obj -triple x86_64-unknown-unknown %s -o %t.o
# RUN: llvm-strip --strip-unneeded %t.o
# RUN: %clang %cflags -no-pie -nostartfiles -nostdlib -lc %t.o -o %t.exe -Wl,-q
# RUN: llvm-bolt %t.exe -o %t.exe.bolt --relocs=1 --lite=0 \
# RUN: --reorder-blocks=reverse
# Useful when manually testing this. Currently we just check that
# the test does not cause BOLT to assert.
# COM: %t.exe.bolt 1 2
.file "jt-symbol-disambiguation-3.s"
.text
# ----
# Func foo contains a jump table whose start is colocated with a
# jump table reference in another function. However, the other function
# does not use the first entries of it and is merely doing arithmetics
# to save the creation of unused first entries.
# ----
.globl foo
.type foo, @function
foo:
.cfi_startproc
xor %rax,%rax
cmpq $3, %rdi
ja .LBBAD
jmpq *.JT1(,%rdi,8)
.LBB1:
movl $0x4,%eax
jmp .LBB5
.LBB2:
movl $0x5,%eax
jmp .LBB5
.LBB3:
movl $0x6,%eax
jmp .LBB5
.LBB4:
movl $0x7,%eax
.LBB5:
retq
.LBBAD:
mov $1, %rdi
callq exit@PLT
retq
.cfi_endproc
.size foo, .-foo
# ----
# Func _start scans an object with indexed access using %rax * 8 as an
# index. However, %rax is known to be at least one, so the compiler
# loads the pointer for the base address as object - 8 instead of just
# object.
# ----
.globl _start
.type _start, @function
_start:
.cfi_startproc
movq (%rsp), %rdi
callq foo
movq $1, %rdi
cmpq $7, %rax
ja .LBB10
# Here the compiler uses the knowledge that the first four entries
# of the jump table are not accessed and subtracts 32 from the base
# address of it, so it doesn't have to allocate four unused entries in
# memory. Unfortunately this can confuse BOLT since it ends up being a
# direct reference to JT1, after linker is done.
jmpq *.JT2-32(,%rax,8)
.LBB6:
xorq %rdi, %rdi
jmp .LBB10
.LBB7:
xorq %rdi, %rdi
jmp .LBB10
.LBB8:
xorq %rdi, %rdi
jmp .LBB10
.LBB9:
xorq %rdi, %rdi
jmp .LBB10
.LBB10:
callq exit@PLT
retq
.cfi_endproc
.size _start, .-_start
# ----
# Data section
# ----
.section .rodata,"a",@progbits
.p2align 3
.JT1:
.quad .LBB1
.quad .LBB2
.quad .LBB3
.quad .LBB4
.JT2:
.quad .LBB6
.quad .LBB7
.quad .LBB8
.quad .LBB9