Re: [PATCH] x86: guard against straight-line speculation past RET

[Jan Beulich <jbeulich@xxxxxxxx>]

On 07.09.2020 11:25, Jan Beulich wrote:
> On 04.09.2020 20:18, Andrew Cooper wrote:
>> Clang doesn't actually expand the macro for ret instructions, so a Clang
>> build of Xen only ends up getting protected in the assembly files.
>>
>> The following experiment demonstrates the issue:
>>
>> $ cat ret.c
>> asm (".macro ret\n\t"
>>      ".error \"foo\"\n\t"
>>      ".endm\n\t");
>> void foo(void) {}
>>
>> $ gcc -O3 -c ret.c -o ret.o && objdump -d ret.o
>> /tmp/ccf8hkyN.s: Assembler messages:
>> /tmp/ccf8hkyN.s:16: Error: foo
>>
>> $ clang-10 -O3 -c ret.c -o ret.o && objdump -d ret.o
>>
>> ret.o:     file format elf64-x86-64
>>
>>
>> Disassembly of section .text:
>>
>> 0000000000000000 <foo>:
>>    0:    c3                       retq
>>
>>
>> Worse, -no-integrated-as doesn't immediately help, even though it
>> invokes $(AS).
>>
>> I tracked that down to the difference between ret and retq, which
>> highlights an assumption about GCC which may not remain true in the future.
>>
>> Adding a second macro covering retq fixes the scenario in combination
>> with -no-integrated-as.
> 
> Ah, yes, I should of course have thought of retq. Albeit as per
> above - generated code looks fine here when using clang 5.

I'm sorry, I can indeed see this part of the issue. I did look at
the wrong build tree when putting together the earlier reply.

Jan