Implement callx instruction #584
Conversation
dthaler
force-pushed
the
callx
branch
2 times, most recently
from
13c247e to
4bd4861
Compare
These instructions are generated by both clang (under -O0 or -O1) and gcc (if the experimental -mxbpf option is passed to the compiler), but are not supported by Linux. Per mailing list discussion at https://mailarchive.ietf.org/arch/msg/bpf/CDQjTO8R8gdPdfeKVnoxWco8_Lw/ the intent is to support them eventually, and with this PR, ebpf-for-windows can support them now. This will also unblock the ability to use versions of clang later than clang-11 by using -O1 instead of -O2 (which generates correlated branches PREVAIL can't deal with). The mailing list discussion at https://mailarchive.ietf.org/arch/msg/bpf/Vx1H3ViPUWoGKNssCO22lOIjyXU/ agreed that inst.dst is where to put the register to use going forward and clang v.19 will do so but earlier versions of clang used inst.imm. Signed-off-by: Dave Thaler <[email protected]>
There was a problem hiding this comment.
Looks good. I will merge after the style comments are addressed.
I'd do some of it differently, but it can probably wait for a future PR:
Instead of a dedicated Callx struct, we can have a Value function as a field in Call.
Other fields in Call go either into a Signature struct or left as platform details (name).
The value of the register will not be replaced by multiple instruction, but with the meet-over-signatures of possible values (signature, as a type, is itself an abstract domain; meet can be as simple as "either a single value or BOTTOM").
get_assertions() will not be needed as a general interface to the assertions module; we only need it for callx, and we pass instruction because we don't have a "signature" struct/abstract domain.
Signed-off-by: Dave Thaler <[email protected]>
Signed-off-by: Dave Thaler <[email protected]>
|
@elazarg I believe I've now addressed your comments. Please re-review. |
"callx" instructions are generated by both clang (under -O0 or -O1) and gcc (if the
experimental -mxbpf option is passed to the compiler), but are not supported by Linux.
Per mailing list discussion at
https://mailarchive.ietf.org/arch/msg/bpf/CDQjTO8R8gdPdfeKVnoxWco8_Lw/ the intent is for the Linux verifier to support them eventually, and with this PR, PREVAIL and dependent projects like ebpf-for-windows can support them now. This will also unblock the ability to use versions of clang later than clang-11 by using -O1 instead of -O2, which generates correlated branches PREVAIL can't deal with.
The approach taken by this PR is:
integer value at the time of the instruction. This covers the common case.
been done with the normal BPF_CALL instruction.
The mailing list discussion at
https://mailarchive.ietf.org/arch/msg/bpf/Vx1H3ViPUWoGKNssCO22lOIjyXU/
agreed that inst.dst is where to put the register to use going forward.
Clang v.19 will do so but earlier versions of clang all used inst.imm.