[BPF] __bpf_trap declaration emitted even if not called

[nikic]

For this IR:

target triple = "bpf"
    
define i32 @test(i8 %x) {  
entry:
  %0 = and i8 %x, 3
  switch i8 %0, label %default.unreachable4 [
    i8 0, label %return
    i8 1, label %sw.bb1
    i8 2, label %sw.bb2
    i8 3, label %sw.bb3
  ]

sw.bb1:                                           ; preds = %entry
  br label %return

sw.bb2:                                           ; preds = %entry
  br label %return

sw.bb3:                                           ; preds = %entry
  br label %return

default.unreachable4:                             ; preds = %entry
  unreachable

return:                                           ; preds = %entry, %sw.bb3, %sw.bb2, %sw.bb1
  %retval.0 = phi i32 [ 12, %sw.bb1 ], [ 43, %sw.bb2 ], [ 54, %sw.bb3 ], [ 32, %entry ]
  ret i32 %retval.0
}

llc produces:

	.file	"<stdin>"
	.text
	.globl	test                            # -- Begin function test
	.p2align	3
	.type	test,@function
test:                                   # @test
	.cfi_startproc
# %bb.0:                                # %entry
	w1 &= 3
	if w1 s> 1 goto LBB0_3
# %bb.1:                                # %entry
	w0 = 32
	if w1 == 0 goto LBB0_6
# %bb.2:                                # %entry
	w0 = 12
	goto LBB0_6
LBB0_3:                                 # %entry
	if w1 == 2 goto LBB0_5
# %bb.4:                                # %entry
	w0 = 54
	goto LBB0_6
LBB0_5:                                 # %sw.bb2
	w0 = 43
LBB0_6:                                 # %return
	exit
.Lfunc_end0:
	.size	test, .Lfunc_end0-test
	.cfi_endproc
                                        # -- End function
	.weak	__bpf_trap

Note that it declares .weak __bpf_trap, even though the function is never called. This results in a libbpf: extern weak function __bpf_trap is unsupported error on kernels that don't have support for __bpf_trap. On kernels that do have it, there is no error because there is no call.

The reason is that the default.unreachable4 gets lowered to a call to __bpf_trap originally and then gets eliminated as an unused block.

[nikic]

cc @yonghong-song

[nikic]

Independent of this specific issue, I also wonder whether it would make sense to gate __bpf_trap usage on the v4 CPU, so it's not used on old kernels?

[yonghong-song]

There is an bpf backend option (-bpf-disable-trap-unreachable) to disable this feature. User can use this.

Another thing is BPF backend can remove ".weak __bpf_trap" if BPF backend can confirm there is no actual reference to __bpf_trap in the code. Do you think this approach could work?

Regarding cpu v4 suggestion, even with modern kernel, people may still use cpu v1/v3 for bpf programs. Yes, I know some people still use cpu v1 for bpf programs with modern kernel. Guarding with cpu v4 can avoid this issue but may prevent to warn users in a proper way.

[nikic]

Another thing is BPF backend can remove ".weak __bpf_trap" if BPF backend can confirm there is no actual reference to __bpf_trap in the code. Do you think this approach could work?

Won't know for sure until we try it, but I would expect that to work...

[yonghong-song]

Another thing is BPF backend can remove ".weak __bpf_trap" if BPF backend can confirm there is no actual reference to __bpf_trap in the code. Do you think this approach could work?

Won't know for sure until we try it, but I would expect that to work...

Okay, I think it should work and I will give a try and send a pull request soon.

[tuliom]

Confirmed fixed with #166003.
In case someone else tries to backport this fix to LLVM 21, this commit depends on #149715.