X86-64 -tailcallopt calling convention specification seems out-of-sync with regular cc

[llvmbot]

Bugzilla Link	4278
Resolution	FIXED
Resolved on	Jun 12, 2009 11:31
Version	unspecified
OS	Linux
Reporter	LLVM Bugzilla Contributor

Extended Description

In CC_X86_64_TailCall (X86CallingConv.td), there are several differences from CC_X86_64_C that are not explained (AFAICT) by the need to support tail calls.

• It doesn't support fp80 arguments (causing a "Formal argument #​0 has unhandled type f80" abort).
• v2f32 doesn't seem to be handled either.
• The code that puts MMX vectors in XMM registers on Darwin doesn't check CCIfSubtarget<"hasSSE2()"> like it does for regular calls.

Also, the comment for the integer registers is a copy-paste-o; there are only 5 of them for tail calls, not 6.

[llvmbot]

Remove CC_X86_64_TailCall calling convention.
This patch removes the CC_X86_64_TailCall calling convention because it is no longer needed and out of sync with the standard x86-64 calling convention. Potential function pointers will now be passed in R11 when tail calling and tail call opt is enabled. (Only tested with make check)

[llvmbot]

CC_X86_64_TailCall was a copy of CC_X86_64_C and I have not updated it since i copied it, hence the differences. Because we can use register R11 for passing the function pointer (R11 is caller saved and not used for passing parameters, System V ABI designates it as scratch register for PLT code) when tail calling there is no need for CC_X86_64_TailCall (the only difference between CC_X86_64_TailCall/CC_X86_64_C was that CC_X86_64_TailCall reserved register R9 for storing a potential function pointer). The above patch removes CC_X86_64_TailCall.

Once i have tested it i will commit it to the repository.

[llvmbot]

Fixed see: http://llvm.org/viewvc/llvm-project?view=rev&revision=73233.

Thank you Frits for bringing this to my attention!