[CIR] [Lowering] [X86_64] Support VAArg for LongDouble

[ChuanqiXu9]

Recommit #1101

I am not sure what happened. But that merged PR doesn't show in the git log. Maybe the stacked PR may not get successed? But after all, we need to land it again.

Following off are original commit messages:

---

This is the following of #1100.

After #1100, when we want to use LongDouble for VAArg, we will be in trouble due to details in X86_64's ABI and this patch tries to address this.

The practical impact the patch is, after this patch, with #1088 and a small following up fix, we can build and run all C's benchmark in SpecCPU 2017. I think it is a milestone.

[github-actions]

✅ With the latest revision this PR passed the C/C++ code formatter.

[smeenai]

Yeah, looks like the stacked PR was incorrectly merged into a user branch instead of main. This looks good after you fix the formatting.

[ChuanqiXu9]

Yeah, looks like the stacked PR was incorrectly merged into a user branch instead of main. This looks good after you fix the formatting.

Done

[smeenai]

The test failure looks related, though I don't know why it would only fail on Windows.

[ChuanqiXu9]

It turns out to be a weird ordering mismatch. In my environment, I see:

cir.va.start %4 : !cir.ptr<!ty___va_list_tag> loc(#loc23)
    %5 = cir.cast(array_to_ptrdecay, %2 : !cir.ptr<!cir.array<!ty___va_list_tag x 1>>), !cir.ptr<!ty___va_list_tag> loc(#loc21)
    %6 = cir.get_member %5[2] {name = "overflow_arg_area"} : !cir.ptr<!ty___va_list_tag> -> !cir.ptr<!cir.ptr<!void>> loc(#loc21)
    %7 = cir.load %6 : !cir.ptr<!cir.ptr<!void>>, !cir.ptr<!void> loc(#loc21)
    %8 = cir.cast(bitcast, %7 : !cir.ptr<!void>), !cir.ptr<!u8i> loc(#loc21)
    %9 = cir.const #cir.int<15> : !u32i loc(#loc21)
    %10 = cir.ptr_stride(%8 : !cir.ptr<!u8i>, %9 : !u32i), !cir.ptr<!u8i> loc(#loc21)

where the case come before the const.

but the output in the above link is:

cir.va.start %4 : !cir.ptr<!ty___va_list_tag> loc(#loc23) 
# |             66:  %5 = cir.cast(array_to_ptrdecay, %2 : !cir.ptr<!cir.array<!ty___va_list_tag x 1>>), !cir.ptr<!ty___va_list_tag> loc(#loc21) 
# |             67:  %6 = cir.get_member %5[2] {name = "overflow_arg_area"} : !cir.ptr<!ty___va_list_tag> -> !cir.ptr<!cir.ptr<!void>> loc(#loc21) 
# |             68:  %7 = cir.load %6 : !cir.ptr<!cir.ptr<!void>>, !cir.ptr<!void> loc(#loc21) 
# |             69:  %8 = cir.const #cir.int<15> : !u32i loc(#loc21) 
# |             70:  %9 = cir.cast(bitcast, %7 : !cir.ptr<!void>), !cir.ptr<!u8i> loc(#loc21) 
# | check:122'0                                                  X~~~~~~~~~~~~~~~~~~~~~~~~~~~~ error: no match found
# |             71:  %10 = cir.ptr_stride(%9 : !cir.ptr<!u8i>, %8 : !u32i), !cir.ptr<!u8i> loc(#loc21) 

but now the const comes before the cast.

hmm.. although we can try to use DAG to check it. But I prefer to not debugging with CI (I don't have windows environment). So I prefer to skip this test on windows.

[bcardosolopes]

But I prefer to not debugging with CI (I don't have windows environment). So I prefer to skip this test on windows.

Doesn't seem like we have another option? If DAG works why not use it? Disabling stuff for other platforms will create tech debt for no good reason here

[smeenai]

I think it's pretty strange that we're getting different IR output on different machines. We're specifying a target triple, so it should be fully deterministic, right?

[bcardosolopes]

I think it's pretty strange that we're getting different IR output on different machines

True, if this is exposing non-deterministic behavior, looks like the right opportunity to understand and fix (or at least if we understand and it's more complex an issue can be created and this could be addressed on a follow up PR)

[ChuanqiXu9]

But I prefer to not debugging with CI (I don't have windows environment). So I prefer to skip this test on windows.

Doesn't seem like we have another option? If DAG works why not use it? Disabling stuff for other platforms will create tech debt for no good reason here

I was afraid to debug with CI. But I added the DAG now. Let's see what happens.

[ChuanqiXu9]

The CI looks green now.