Re: [PATCH 08/14] bpf/tests: Add tests for ALU operations implemented with function calls

[Yonghong Song <yhs@fb.com>]

On 7/29/21 2:17 PM, Johan Almbladh wrote:
> On Thu, Jul 29, 2021 at 1:52 AM Yonghong Song <yhs@fb.com> wrote:
>>> +             /*
>>> +              * Register (non-)clobbering test, in the case where a 32-bit
>>> +              * JIT implements complex ALU64 operations via function calls.
>>> +              */
>>> +             "INT: Register clobbering, R1 updated",
>>> +             .u.insns_int = {
>>> +                     BPF_ALU32_IMM(BPF_MOV, R0, 0),
>>> +                     BPF_ALU32_IMM(BPF_MOV, R1, 123456789),
>>> +                     BPF_ALU32_IMM(BPF_MOV, R2, 2),
>>> +                     BPF_ALU32_IMM(BPF_MOV, R3, 3),
>>> +                     BPF_ALU32_IMM(BPF_MOV, R4, 4),
>>> +                     BPF_ALU32_IMM(BPF_MOV, R5, 5),
>>> +                     BPF_ALU32_IMM(BPF_MOV, R6, 6),
>>> +                     BPF_ALU32_IMM(BPF_MOV, R7, 7),
>>> +                     BPF_ALU32_IMM(BPF_MOV, R8, 8),
>>> +                     BPF_ALU32_IMM(BPF_MOV, R9, 9),
>>> +                     BPF_ALU64_IMM(BPF_DIV, R1, 123456789),
>>> +                     BPF_JMP_IMM(BPF_JNE, R0, 0, 10),
>>> +                     BPF_JMP_IMM(BPF_JNE, R1, 1, 9),
>>> +                     BPF_JMP_IMM(BPF_JNE, R2, 2, 8),
>>> +                     BPF_JMP_IMM(BPF_JNE, R3, 3, 7),
>>> +                     BPF_JMP_IMM(BPF_JNE, R4, 4, 6),
>>> +                     BPF_JMP_IMM(BPF_JNE, R5, 5, 5),
>>> +                     BPF_JMP_IMM(BPF_JNE, R6, 6, 4),
>>> +                     BPF_JMP_IMM(BPF_JNE, R7, 7, 3),
>>> +                     BPF_JMP_IMM(BPF_JNE, R8, 8, 2),
>>> +                     BPF_JMP_IMM(BPF_JNE, R9, 9, 1),
>>> +                     BPF_ALU32_IMM(BPF_MOV, R0, 1),
>>> +                     BPF_EXIT_INSN(),
>>> +             },
>>> +             INTERNAL,
>>> +             { },
>>> +             { { 0, 1 } }
>>> +     },
>>> +     {
>>> +             "INT: Register clobbering, R2 updated",
>>> +             .u.insns_int = {
>>> +                     BPF_ALU32_IMM(BPF_MOV, R0, 0),
>>> +                     BPF_ALU32_IMM(BPF_MOV, R1, 1),
>>> +                     BPF_ALU32_IMM(BPF_MOV, R2, 2 * 123456789),
>>> +                     BPF_ALU32_IMM(BPF_MOV, R3, 3),
>>> +                     BPF_ALU32_IMM(BPF_MOV, R4, 4),
>>> +                     BPF_ALU32_IMM(BPF_MOV, R5, 5),
>>> +                     BPF_ALU32_IMM(BPF_MOV, R6, 6),
>>> +                     BPF_ALU32_IMM(BPF_MOV, R7, 7),
>>> +                     BPF_ALU32_IMM(BPF_MOV, R8, 8),
>>> +                     BPF_ALU32_IMM(BPF_MOV, R9, 9),
>>> +                     BPF_ALU64_IMM(BPF_DIV, R2, 123456789),
>>> +                     BPF_JMP_IMM(BPF_JNE, R0, 0, 10),
>>> +                     BPF_JMP_IMM(BPF_JNE, R1, 1, 9),
>>> +                     BPF_JMP_IMM(BPF_JNE, R2, 2, 8),
>>> +                     BPF_JMP_IMM(BPF_JNE, R3, 3, 7),
>>> +                     BPF_JMP_IMM(BPF_JNE, R4, 4, 6),
>>> +                     BPF_JMP_IMM(BPF_JNE, R5, 5, 5),
>>> +                     BPF_JMP_IMM(BPF_JNE, R6, 6, 4),
>>> +                     BPF_JMP_IMM(BPF_JNE, R7, 7, 3),
>>> +                     BPF_JMP_IMM(BPF_JNE, R8, 8, 2),
>>> +                     BPF_JMP_IMM(BPF_JNE, R9, 9, 1),
>>> +                     BPF_ALU32_IMM(BPF_MOV, R0, 1),
>>> +                     BPF_EXIT_INSN(),
>>> +             },
>>> +             INTERNAL,
>>> +             { },
>>> +             { { 0, 1 } }
>>> +     },
>>
>> It looks like the above two tests, "R1 updated" and "R2 updated" should
>> be very similar and the only difference is one immediate is 123456789
>> and another is 2 * 123456789. But for generated code, they all just have
>> the final immediate. Could you explain what the difference in terms of
>> jit for the above two tests?
> 
> When a BPF_CALL instruction is executed, the eBPF assembler have
> already saved any caller-saved registers that must be preserved, put
> the arguments in R1-R5, and expects a return value in R0. It is just
> for the JIT to emit the call.
> 
> Not so when an eBPF instruction is implemented by a function call,
> like ALU64 DIV in a 32-bit JIT. In this case, the function call is
> unexpected by the eBPF assembler, and must be invisible to it. Now the
> JIT must take care of saving all caller-saved registers on stack, put
> the operands in the right argument registers, put the return value in
> the destination register, and finally restore all caller-saved
> registers without overwriting the computed result.
> 
> The test checks that all other registers retain their value after such
> a hidden function call. However, one register will contain the result.
> In order to verify that all registers are saved and restored properly,
> we must vary the destination and run it two times. It is not the
> result of the operation that its tested, it is absence of possible
> side effects.
> 
> I can put a more elaborate description in the comment to explain this.

Indeed, an elaborate description as comments will be great.

> 
>>
>>> +     {
>>> +             /*
>>> +              * Test 32-bit JITs that implement complex ALU64 operations as
>>> +              * function calls R0 = f(R1, R2), and must re-arrange operands.
>>> +              */
[...]