[llvm] [AArch64] Support varargs for preserve_nonecc (PR #99434)

[via llvm-commits]

================
@@ -0,0 +1,105 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc -mtriple=aarch64 < %s | FileCheck %s
+
+%va_list = type { ptr, ptr, ptr, i32, i32 }
+
+define preserve_nonecc i32 @callee(i32 %a1, i32 %a2, i32 %a3, i32 %a4, i32 %a5, ...) nounwind noinline ssp {
+; CHECK-LABEL: callee:
+; CHECK:       // %bb.0: // %entry
+; CHECK-NEXT:    sub sp, sp, #192
+; CHECK-NEXT:    mov x8, #-24 // =0xffffffffffffffe8
+; CHECK-NEXT:    mov x9, sp
+; CHECK-NEXT:    add x10, sp, #136
+; CHECK-NEXT:    movk x8, #65408, lsl #32
+; CHECK-NEXT:    add x9, x9, #128
+; CHECK-NEXT:    stp x6, x7, [sp, #144]
+; CHECK-NEXT:    stp x9, x8, [sp, #176]
+; CHECK-NEXT:    add x9, x10, #24
+; CHECK-NEXT:    add x10, sp, #192
+; CHECK-NEXT:    mov w8, #-24 // =0xffffffe8
+; CHECK-NEXT:    str x5, [sp, #136]
+; CHECK-NEXT:    stp q0, q1, [sp]
+; CHECK-NEXT:    stp q2, q3, [sp, #32]
+; CHECK-NEXT:    stp q4, q5, [sp, #64]
+; CHECK-NEXT:    stp q6, q7, [sp, #96]
+; CHECK-NEXT:    stp x10, x9, [sp, #160]
+; CHECK-NEXT:    tbz w8, #31, .LBB0_3
+; CHECK-NEXT:  // %bb.1: // %maybe_reg
+; CHECK-NEXT:    add w9, w8, #8
+; CHECK-NEXT:    cmp w9, #0
+; CHECK-NEXT:    str w9, [sp, #184]
+; CHECK-NEXT:    b.gt .LBB0_3
+; CHECK-NEXT:  // %bb.2: // %in_reg
+; CHECK-NEXT:    ldr x9, [sp, #168]
+; CHECK-NEXT:    add x8, x9, w8, sxtw
+; CHECK-NEXT:    b .LBB0_4
+; CHECK-NEXT:  .LBB0_3: // %on_stack
+; CHECK-NEXT:    ldr x8, [sp, #160]
+; CHECK-NEXT:    add x9, x8, #8
+; CHECK-NEXT:    str x9, [sp, #160]
+; CHECK-NEXT:  .LBB0_4: // %end
+; CHECK-NEXT:    ldr w0, [x8]
+; CHECK-NEXT:    add sp, sp, #192
+; CHECK-NEXT:    ret
+entry:
+  %args = alloca %va_list, align 8
+  call void @llvm.va_start(ptr %args)
+  %gr_offs_p = getelementptr inbounds %va_list, ptr %args, i32 0, i32 3
+  %gr_offs = load i32, ptr %gr_offs_p, align 8
+  %0 = icmp sge i32 %gr_offs, 0
+  br i1 %0, label %on_stack, label %maybe_reg
+
+maybe_reg:
+  %new_reg_offs = add i32 %gr_offs, 8
+  store i32 %new_reg_offs, ptr %gr_offs_p, align 8
+  %inreg = icmp sle i32 %new_reg_offs, 0
+  br i1 %inreg, label %in_reg, label %on_stack
+
+in_reg:
+  %reg_top_p = getelementptr inbounds %va_list, ptr %args, i32 0, i32 1
+  %reg_top = load ptr, ptr %reg_top_p, align 8
+  %reg = getelementptr inbounds i8, ptr %reg_top, i32 %gr_offs
+  br label %end
+
+on_stack:
+  %stack_p = getelementptr inbounds %va_list, ptr %args, i32 0, i32 0
+  %stack = load ptr, ptr %stack_p, align 8
+  %new_stack = getelementptr inbounds i8, ptr %stack, i64 8
+  store ptr %new_stack, ptr %stack_p, align 8
+  br label %end
+
+end:
+  %p = phi ptr [ %reg, %in_reg ], [ %stack, %on_stack ]
+  %10 = load i32, ptr %p, align 8
+  call void @llvm.va_end.p0(ptr %args)
+  ret i32 %10
+}
+
+declare void @llvm.va_start(ptr) nounwind
+declare void @llvm.va_end(ptr) nounwind
+
+define i32 @caller() nounwind ssp {
+; CHECK-LABEL: caller:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    sub sp, sp, #32
+; CHECK-NEXT:    mov w8, #10 // =0xa
+; CHECK-NEXT:    mov w9, #9 // =0x9
+; CHECK-NEXT:    mov w0, #1 // =0x1
+; CHECK-NEXT:    mov w1, #2 // =0x2
+; CHECK-NEXT:    mov w2, #3 // =0x3
+; CHECK-NEXT:    mov w3, #4 // =0x4
+; CHECK-NEXT:    mov w4, #5 // =0x5
+; CHECK-NEXT:    mov w5, #6 // =0x6
+; CHECK-NEXT:    mov w6, #7 // =0x7
+; CHECK-NEXT:    mov w7, #8 // =0x8
+; CHECK-NEXT:    str x30, [sp, #16] // 8-byte Folded Spill
+; CHECK-NEXT:    str w8, [sp, #8]
+; CHECK-NEXT:    str w9, [sp]
+; CHECK-NEXT:    bl callee
+; CHECK-NEXT:    ldr x30, [sp, #16] // 8-byte Folded Reload
+; CHECK-NEXT:    add sp, sp, #32
+; CHECK-NEXT:    ret
+  %r = tail call i32 (i32, i32, i32, i32, i32, ...) @callee(i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10)
----------------
antangelo wrote:

Thanks for catching that! I was diffing output in the preserve_none case against the C calling convention while testing and missed changing it back.
The only effect it has on the generated code is the spilling of registers that are not saved in preserve_none. The function arguments are passed exactly the same otherwise.

https://github.com/llvm/llvm-project/pull/99434