Extend and clean the ccall test suite.

[maleadt]

The current ccall test suite doesn't match/test all functionality in libccalltest. This PR removes some unused functionality from libccalltest, but mainly extends libccalltest and test/ccall to align the different structs and test methods.

As part of a small clean-up, I've also scoped most tests in a let...end block, and added a factor-like parameter to the struct test methods to catch cases where the ABI allocates too much memory for a certain parameter (it helped me discover issues with the ARM ABI from #14194).

I've only tested these changes on x86_64, so it's possible that it triggers some issues on other platforms...

cc @vtjnash

[edit by tkelman: fixes https://github.com//issues/13242]

[maleadt]

Also, as part of #14204 I learned that I should use the C type aliases. Should I convert the ccall test suite to use these types? I've currently kept it at how it was implemented before, except where the types differ (i.e. use Cchar instead of Char, but still use Float64 etc. because these don't differ from the C types).

[vtjnash]

no, there's no need to convert. Char corresponds to wchar_t (usually), not char (similar to how Bool corresponds to unsigned char or C++ bool, not int as is customary in C; and Int corresponds to intptr_t / size_t, not int), so that's really just a matter of misleading aliases. The Cprefix aliases are available to help with that conversion, but are only really necessary for places where platforms differ (for example, Cchar is aware of the platform-specific sign for char, Clong is aware of the number of bits in a long, etc.)

this looks like a good improvement. the previous tests were usually enough to cause failures for the x86_64 ABI, but I'm not surprised they were insufficient to catch other problems.

perhaps you already know this (as advice for #14194), but part of the original value of this file is that you can introspect how clang lowers it as a way of matching the EABI spec to the llvm attributes:
clang -emit-llvm -S -o - ccalltest.c

[maleadt]

@vtjnash OK thanks for the clarification, much appreciated!
About that clang usage, in a comment on #13752 you used --target=arm-pc-linux-elf, but code generated with that command doesn't use the VFP (I'm not sure if it's EABI). For developing the ABI, I've been using --target=arm-pc-linux-elf-eabi instead. Any idea about the exact differences?

[tkelman]

Char is not wchar_t on windows.

[maleadt]

So apparently this trips up the Win32 ABI (unsure about the Win64 one, it just seems to have halted). I guess those ABI's will need fixing before this can be merged? There's also some compilation warnings in ccalltest.c which I should look at.

[tkelman]

The win64 freeze is probably not your fault, I would restart it but there's a conflict so this needs a rebase.

[vtjnash]

Yep, good catch. It looks like we have a couple mistakes in our ABI. We had copied our ABI from LDC, and they fixed this over the summer: ldc-developers/ldc#856

• the conditional in use_sret should be size == 1 || size == 2 || size == 4 || size == 8 (instead of size <= 8 -- or this is a clang bug, but i doubt that). this change also needs to be made to the win64 abi and darwin32* abi.
• needPassByRef should use the x32 logic (minus the special case for complex)
• preferred_llvm_type logic will need to take into account the change to use_sret (ibid for win64)

[*] source: https://github.com/ldc-developers/ldc/blob/bddd95194815c24c48676e1b2132436c0e4df96d/gen/abi-x86.cpp#L67

[ViralBShah]

Bump. Does this help improve testing, given the recent arm and aarch64 abi implementations?

[maleadt]

Yeah sorry for not updating this, but I haven't found the time to fix the Windows ABI. Without that, I guess this can't get merged. Maybe somebody else could look at this?

Concerning the ARM ABI, it greatly helped me verify the the implementation and find corner cases, and it should pass this test suite without any issue. Though there are still some untested code paths (I should probably look at coverage statistics some day), this PR greatly improves testing coverage.
Lacking an ARMv8 board, I haven't tested the AArch64 ABI using this PR. Maybe @yuyichao has?

[yuyichao]

For the failing test, the llvm ir generated by clang is

; ModuleID = 'test.c'
target datalayout = "e-m:x-p:32:32-i64:64-f80:32-n8:16:32-a:0:32-S32"
target triple = "i386-pc-windows-msvc18.0.0"

%struct.struct10 = type { i8, i8, i8, i8 }

; Function Attrs: nounwind
define i32 @test_10(%struct.struct10* byval nocapture align 4 %a) #0 {
  %1 = getelementptr inbounds %struct.struct10, %struct.struct10* %a, i32 0, i32 0
  %2 = bitcast %struct.struct10* %a to i32*
  %3 = load i32, i32* %2, align 4
  %4 = trunc i32 %3 to i16
  %5 = trunc i32 %3 to i8
  %6 = add i8 %5, 1
  store i8 %6, i8* %1, align 4, !tbaa !1
  %7 = getelementptr inbounds %struct.struct10, %struct.struct10* %a, i32 0, i32 1
  %8 = lshr i16 %4, 8
  %9 = trunc i16 %8 to i8
  %10 = add i8 %9, -2
  store i8 %10, i8* %7, align 1, !tbaa !5
  %11 = getelementptr inbounds %struct.struct10, %struct.struct10* %a, i32 0, i32 2
  %12 = lshr i32 %3, 16
  %13 = trunc i32 %12 to i8
  %14 = add i8 %13, 3
  store i8 %14, i8* %11, align 2, !tbaa !6
  %15 = getelementptr inbounds %struct.struct10, %struct.struct10* %a, i32 0, i32 3
  %16 = lshr i32 %3, 24
  %17 = trunc i32 %16 to i8
  %18 = add i8 %17, -4
  store i8 %18, i8* %15, align 1, !tbaa !7
  %19 = bitcast %struct.struct10* %a to i32*
  %20 = load i32, i32* %19, align 4
  ret i32 %20
}

attributes #0 = { nounwind "disable-tail-calls"="false" "less-precise-fpmad"="false" "no-frame-pointer-elim"="false" "no-infs-fp-math"="false" "no-nans-fp-math"="false" "stack-protector-buffer-size"="8" "target-cpu"="pentium4" "target-features"="+sse,+sse2" "unsafe-fp-math"="false" "use-soft-float"="false" }

!llvm.ident = !{!0}

!0 = !{!"clang version 3.7.1 (tags/RELEASE_371/final)"}
!1 = !{!2, !3, i64 0}
!2 = !{!"", !3, i64 0, !3, i64 1, !3, i64 2, !3, i64 3}
!3 = !{!"omnipotent char", !4, i64 0}
!4 = !{!"Simple C/C++ TBAA"}
!5 = !{!2, !3, i64 1}
!6 = !{!2, !3, i64 2}
!7 = !{!2, !3, i64 3}

And it appears to be passing the argument by reference but not using sret whereas the logic in abi_win32.cpp seems to be doing neither. Where can I find a complete doc about the win32 calling convention?

[yuyichao]

From the clang output and the ldc code, it seems that the win32 abi use sret for structs that are 1, 2, 4, 8 bytes in size and use pass by reference for all structs.

[yuyichao]

Subsumed by #14906

[maleadt]

@yuyichao OK, thanks for looking into this.