BOLT-ERROR: PC-relative operand can't be evaluated:

[Sameeranjoshi]

I am trying to run SPEC CPU 2017 rate runs, using llvm-bolt for which I am trying to generate the bolt cfg specifically for the ICC binaries.

Here's how I am running it.

llvm-bolt  -update-debug-sections -print-cfg=1 perlbench_r_base.ic20.ROME-BOLT -o icc.bolt > icc.cfg

I get a few errors as

BOLT-WARNING: split function detected on input : __assert_fail_base.cold/1. The support is limited in relocation mode.
BOLT-ERROR: symbol seen in the middle of the function __BOLT_FDE_FUNCat47d47f. Skipping.
BOLT-ERROR: symbol seen in the middle of the function __BOLT_FDE_FUNCat4b1869. Skipping.
BOLT-ERROR: symbol seen in the middle of the function __BOLT_FDE_FUNCat4b1882. Skipping.
BOLT-ERROR: symbol seen in the middle of the function __BOLT_FDE_FUNCat4b1903. Skipping.
BOLT-ERROR: PC-relative operand can't be evaluated:
  **movq  %cs:0xf3700(%rip), %r8**
<MCInst #1784 MOV64rm <MCOperand Reg:128> <MCOperand Reg:55> <MCOperand Imm:1> <MCOperand Reg:0> <MCOperand Imm:997120> <MCOperand Reg:12>>
BOLT-ERROR: cannot handle PC-relative operand at 0x4016d6. Skipping function flow_cost/1(*2).

Below flags were added to the SPEC config specifically for running with bolt.

BOLT_FLAGS = -g -Wl,-emit-relocs -fno-inline-functions

Debugging the error, seems to come from BinaryFunction::disassemble() function, which flags that there could be potential error in disassembling it at MC layer from the error.

0. Can there be more inputs and details on what's going wrong here?
1. Is it the compiled binary which has issues ?
2. or is there any issue with disassembling in BOLT ?

I also also verified the rela.txt in binary.

readelf perlbench_r_base.ic20.ROME-BOLT -S | grep "rela"
  [10] .rela.dyn         RELA             0000000000401c68  00001c68
  [11] .rela.plt         RELA             0000000000401d70  00001d70
  [29] .rela.text        RELA             0000000000000000  00287eb0
  [30] .rela.eh_frame    RELA             0000000000000000  003b35b0
  [31] .rela.init        RELA             0000000000000000  003c4170
  [33] .rela.data        RELA             0000000000000000  003c41f0
  [34] .rela.fini_array  RELA             0000000000000000  003c6b60
  [35] .rela.init_array  RELA             0000000000000000  003c6b78
  [42] .rela.rodata      RELA             0000000000000000  00771840
  [43] .rela.debug_loc   RELA             0000000000000000  007db6f8
  [44] .rela.debug_info  RELA             0000000000000000  0083f568
  [45] .rela.debug_range RELA             0000000000000000  00afa940
  [46] .rela.debug_line  RELA             0000000000000000  00b6dae0
  [47] .rela.data.rel.ro RELA             0000000000000000  00b7d518
  [48] .rela.data.rel.ro RELA             0000000000000000  00b7dae8

Thanks again for helping for all the related queries.

[Sameeranjoshi]

Soft reminder/ping !

[maksfb]

I have a good idea on what's going on. The problem is related to the disassembly of the instruction that you see in the dump. I believe the code is specific to ICC. You can try to manually disable processing of the function containing the instruction with -skip-funcs= option, while I'm working on the fix.

You mention that you've added -fno-inline-functions option specifically for BOLT. Why is that?

[Sameeranjoshi]

It was to disable function in-lining for optimizations and to be explicit and not have optimizations triggered, we are working on an internal use-case[1] for performance measurement using great framework which BOLT provides at binary and MC layer of LLVM.

[1] https://lists.llvm.org/pipermail/llvm-dev/2020-October/145915.html

Thanks for looking into the issue.