[llvm] [X86] Fix ABI for passing after i128 (PR #124134)

[Nikita Popov via llvm-commits]

================
@@ -31,15 +31,15 @@ define <2 x i128> @flag_copy_2(<2 x i128> %x, <2 x i128> %y) nounwind {
 ; CHECK-NEXT:    movq %r8, %rdi
 ; CHECK-NEXT:    {nf} sarq $63, %rdi
 ; CHECK-NEXT:    cmovoq %rdi, %rcx
-; CHECK-NEXT:    movabsq $-9223372036854775808, %r9 # imm = 0x8000000000000000
-; CHECK-NEXT:    {nf} xorq %r9, %rdi
+; CHECK-NEXT:    movabsq $-9223372036854775808, %r10 # imm = 0x8000000000000000
+; CHECK-NEXT:    {nf} xorq %r10, %rdi
 ; CHECK-NEXT:    cmovnoq %r8, %rdi
-; CHECK-NEXT:    subq {{[0-9]+}}(%rsp), %rsi
+; CHECK-NEXT:    subq %r9, %rsi
----------------
nikic wrote:

Yeah, I'm assuming that the frontend will handle `<2 x i128>` by directly using byval (https://clang.godbolt.org/z/bznzTKohz -- interestingly clang still directly returns the vector, I would have expected it to use sret, rather than relying on sret demotion in the backend...)

I *could* extend this code to also handle vectors of i128 and require the whole vector argument to be in consecutive arguments. I'm just not sure it makes sense to handle this, as there is no defined psABI for <2 x i128> in the first place, and if there were, the frontend would handle that part.

https://github.com/llvm/llvm-project/pull/124134