[AArch64][GISel] Incorrect ABI for calls with tail marker

[nikic]

; RUN: llc -O0 -mtriple=aarch64-unknown-linux-gnu < %s | FileCheck %s
declare void @func(i64, i64, i64, i64, i64, i128, i128)

define void @test(i128 %arg1, i128 %arg2) nounwind {
  tail call void @func(i64 0, i64 0, i64 0, i64 0, i64 0, i128 %arg1, i128 %arg2)
  ret void
}

This produces:

	sub	sp, sp, #64
	str	x30, [sp, #48]                  // 8-byte Folded Spill
	mov	x6, x0
	mov	x8, sp
	str	x1, [x8]
	str	x2, [x8, #16]
	str	x3, [x8, #32]
	mov	x4, xzr
	mov	x0, x4
	mov	x1, x4
	mov	x2, x4
	mov	x3, x4
	bl	func
	ldr	x30, [sp, #48]                  // 8-byte Folded Reload
	add	sp, sp, #64
	ret

Note that both of the i128 arguments get passed on the stack. Compare with the -global-isel=0 output:

	sub	sp, sp, #32
	str	x30, [sp, #16]                  // 8-byte Folded Spill
	mov	x7, x1
	mov	x6, x0
	mov	x8, sp
	str	x3, [x8, #8]
	str	x2, [x8]
	mov	x4, xzr
	mov	x0, x4
	mov	x1, x4
	mov	x2, x4
	mov	x3, x4
	bl	func
	ldr	x30, [sp, #16]                  // 8-byte Folded Reload
	add	sp, sp, #32
	ret

In this case the first i128 argument is passed in registers and only the second one on the stack.

Peculiarly, this does not happen when dropping the tail marker (which shouldn't make a difference as no actual tail call gets generated).

[llvmbot]

@llvm/issue-subscribers-backend-aarch64

Author: Nikita Popov (nikic)

```llvm ; RUN: llc -O0 -mtriple=aarch64-unknown-linux-gnu < %s | FileCheck %s declare void @func(i64, i64, i64, i64, i64, i128, i128)

define void @test(i128 %arg1, i128 %arg2) nounwind {
tail call void @func(i64 0, i64 0, i64 0, i64 0, i64 0, i128 %arg1, i128 %arg2)
ret void
}

This produces:
```asm
	sub	sp, sp, #<!-- -->64
	str	x30, [sp, #<!-- -->48]                  // 8-byte Folded Spill
	mov	x6, x0
	mov	x8, sp
	str	x1, [x8]
	str	x2, [x8, #<!-- -->16]
	str	x3, [x8, #<!-- -->32]
	mov	x4, xzr
	mov	x0, x4
	mov	x1, x4
	mov	x2, x4
	mov	x3, x4
	bl	func
	ldr	x30, [sp, #<!-- -->48]                  // 8-byte Folded Reload
	add	sp, sp, #<!-- -->64
	ret

Note that both of the i128 arguments get passed on the stack. Compare with the -global-isel=0 output:

	sub	sp, sp, #<!-- -->32
	str	x30, [sp, #<!-- -->16]                  // 8-byte Folded Spill
	mov	x7, x1
	mov	x6, x0
	mov	x8, sp
	str	x3, [x8, #<!-- -->8]
	str	x2, [x8]
	mov	x4, xzr
	mov	x0, x4
	mov	x1, x4
	mov	x2, x4
	mov	x3, x4
	bl	func
	ldr	x30, [sp, #<!-- -->16]                  // 8-byte Folded Reload
	add	sp, sp, #<!-- -->32
	ret

In this case the first i128 argument is passed in registers and only the second one on the stack.

Peculiarly, this does not happen when dropping the tail marker (which shouldn't make a difference as no actual tail call gets generated).

[nikic]

The problem seems to be that

llvm-project/llvm/lib/Target/AArch64/GISel/AArch64CallLowering.cpp

Line 857 in 078ae8c

if (!determineAssignments(CalleeAssigner, OutArgs, OutInfo)) {

when trying to determine whether a tail call is legal modifies state in some way -- looking at that function, it at least modifies the argument flags.