llvm/test/CodeGen/X86/stack-adjust-128.ll

; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
; RUN: llc < %s -mtriple=x86_64-unknown-linux-gnu -verify-machineinstrs | FileCheck %s --check-prefix=LINUX
; RUN: llc < %s -mtriple=x86_64-pc-windows-msvc -verify-machineinstrs | FileCheck %s --check-prefix=WIN64

; A 128-byte stack adjustment should be emitted as `add rsp, -128` in the
; prologue and `sub rsp, -128` in the epilogue: 128 misses the sign-extended
; 8-bit immediate form, but -128 fits it, saving three bytes each. Windows
; epilogues keep the canonical `add rsp, imm` because the Win64 v1 unwinder
; recognizes epilogues by disassembling for it; prologues are located by
; SizeOfProlog and never disassembled, so they may flip.

declare void @escape(ptr)

; One value lives across the call, forcing a single CSR push; the 120 bytes
; of locals then round up to a 128-byte adjustment on Linux.
define ptr @frame128(ptr %p) {
; LINUX-LABEL: frame128:
; LINUX:       # %bb.0:
; LINUX-NEXT:    pushq %rbx
; LINUX-NEXT:    .cfi_def_cfa_offset 16
; LINUX-NEXT:    addq $-128, %rsp
; LINUX-NEXT:    .cfi_def_cfa_offset 144
; LINUX-NEXT:    .cfi_offset %rbx, -16
; LINUX-NEXT:    movq %rdi, %rbx
; LINUX-NEXT:    leaq {{[0-9]+}}(%rsp), %rdi
; LINUX-NEXT:    callq escape@PLT
; LINUX-NEXT:    movq %rbx, %rax
; LINUX-NEXT:    subq $-128, %rsp
; LINUX-NEXT:    .cfi_def_cfa_offset 16
; LINUX-NEXT:    popq %rbx
; LINUX-NEXT:    .cfi_def_cfa_offset 8
; LINUX-NEXT:    retq
;
; WIN64-LABEL: frame128:
; WIN64:       # %bb.0:
; WIN64-NEXT:    pushq %rsi
; WIN64-NEXT:    .seh_pushreg %rsi
; WIN64-NEXT:    subq $160, %rsp
; WIN64-NEXT:    .seh_stackalloc 160
; WIN64-NEXT:    .seh_endprologue
; WIN64-NEXT:    movq %rcx, %rsi
; WIN64-NEXT:    leaq {{[0-9]+}}(%rsp), %rcx
; WIN64-NEXT:    callq escape
; WIN64-NEXT:    movq %rsi, %rax
; WIN64-NEXT:    .seh_startepilogue
; WIN64-NEXT:    addq $160, %rsp
; WIN64-NEXT:    popq %rsi
; WIN64-NEXT:    .seh_endepilogue
; WIN64-NEXT:    retq
; WIN64-NEXT:    .seh_endproc
  %buf = alloca [120 x i8]
  call void @escape(ptr %buf)
  ret ptr %p
}

; The same shape with 96 bytes of locals reaches a 128-byte adjustment on
; Win64 (32-byte shadow area included): the prologue flips to `add rsp, -128`
; but the epilogue must keep `add rsp, 128`.
define ptr @frame128_win(ptr %p) {
; LINUX-LABEL: frame128_win:
; LINUX:       # %bb.0:
; LINUX-NEXT:    pushq %rbx
; LINUX-NEXT:    .cfi_def_cfa_offset 16
; LINUX-NEXT:    subq $96, %rsp
; LINUX-NEXT:    .cfi_def_cfa_offset 112
; LINUX-NEXT:    .cfi_offset %rbx, -16
; LINUX-NEXT:    movq %rdi, %rbx
; LINUX-NEXT:    movq %rsp, %rdi
; LINUX-NEXT:    callq escape@PLT
; LINUX-NEXT:    movq %rbx, %rax
; LINUX-NEXT:    addq $96, %rsp
; LINUX-NEXT:    .cfi_def_cfa_offset 16
; LINUX-NEXT:    popq %rbx
; LINUX-NEXT:    .cfi_def_cfa_offset 8
; LINUX-NEXT:    retq
;
; WIN64-LABEL: frame128_win:
; WIN64:       # %bb.0:
; WIN64-NEXT:    pushq %rsi
; WIN64-NEXT:    .seh_pushreg %rsi
; WIN64-NEXT:    addq $-128, %rsp
; WIN64-NEXT:    .seh_stackalloc 128
; WIN64-NEXT:    .seh_endprologue
; WIN64-NEXT:    movq %rcx, %rsi
; WIN64-NEXT:    leaq {{[0-9]+}}(%rsp), %rcx
; WIN64-NEXT:    callq escape
; WIN64-NEXT:    movq %rsi, %rax
; WIN64-NEXT:    .seh_startepilogue
; WIN64-NEXT:    addq $128, %rsp
; WIN64-NEXT:    popq %rsi
; WIN64-NEXT:    .seh_endepilogue
; WIN64-NEXT:    retq
; WIN64-NEXT:    .seh_endproc
  %buf = alloca [96 x i8]
  call void @escape(ptr %buf)
  ret ptr %p
}

; An adjacent size keeps the plain 32-bit immediate.
define ptr @frame144(ptr %p) {
; LINUX-LABEL: frame144:
; LINUX:       # %bb.0:
; LINUX-NEXT:    pushq %rbx
; LINUX-NEXT:    .cfi_def_cfa_offset 16
; LINUX-NEXT:    subq $144, %rsp
; LINUX-NEXT:    .cfi_def_cfa_offset 160
; LINUX-NEXT:    .cfi_offset %rbx, -16
; LINUX-NEXT:    movq %rdi, %rbx
; LINUX-NEXT:    leaq {{[0-9]+}}(%rsp), %rdi
; LINUX-NEXT:    callq escape@PLT
; LINUX-NEXT:    movq %rbx, %rax
; LINUX-NEXT:    addq $144, %rsp
; LINUX-NEXT:    .cfi_def_cfa_offset 16
; LINUX-NEXT:    popq %rbx
; LINUX-NEXT:    .cfi_def_cfa_offset 8
; LINUX-NEXT:    retq
;
; WIN64-LABEL: frame144:
; WIN64:       # %bb.0:
; WIN64-NEXT:    pushq %rsi
; WIN64-NEXT:    .seh_pushreg %rsi
; WIN64-NEXT:    subq $176, %rsp
; WIN64-NEXT:    .seh_stackalloc 176
; WIN64-NEXT:    .seh_endprologue
; WIN64-NEXT:    movq %rcx, %rsi
; WIN64-NEXT:    leaq {{[0-9]+}}(%rsp), %rcx
; WIN64-NEXT:    callq escape
; WIN64-NEXT:    movq %rsi, %rax
; WIN64-NEXT:    .seh_startepilogue
; WIN64-NEXT:    addq $176, %rsp
; WIN64-NEXT:    popq %rsi
; WIN64-NEXT:    .seh_endepilogue
; WIN64-NEXT:    retq
; WIN64-NEXT:    .seh_endproc
  %buf = alloca [136 x i8]
  call void @escape(ptr %buf)
  ret ptr %p
}