xref: /llvm-project/bolt/test/runtime/X86/shrinkwrapping-alignment.s (revision d648aa1b8e937de1648524e1f1016b53b29ba2a4)
16b4eb0b9SAmir Ayupov# This reproduces a bug with shrink wrapping when trying to move
26b4eb0b9SAmir Ayupov# push-pops in a function where we are not allowed to modify the
36b4eb0b9SAmir Ayupov# stack layout for alignment reasons. In this bug, we failed to
46b4eb0b9SAmir Ayupov# propagate alignment requirement upwards in the call graph for
56b4eb0b9SAmir Ayupov# some functions when there is a cycle in the call graph.
66b4eb0b9SAmir Ayupov
76b4eb0b9SAmir Ayupov# REQUIRES: system-linux
86b4eb0b9SAmir Ayupov
96b4eb0b9SAmir Ayupov# RUN: llvm-mc -filetype=obj -triple x86_64-unknown-unknown \
106b4eb0b9SAmir Ayupov# RUN:   %s -o %t.o
116b4eb0b9SAmir Ayupov# RUN: link_fdata %s %t.o %t.fdata
126b4eb0b9SAmir Ayupov# Delete our BB symbols so BOLT doesn't mark them as entry points
136b4eb0b9SAmir Ayupov# RUN: llvm-strip --strip-unneeded %t.o
14487570fbSAmir Ayupov# RUN: %clang %cflags -no-pie %t.o -o %t.exe -Wl,-q
156b4eb0b9SAmir Ayupov
16*d648aa1bSMaksim Panchenko# RUN: llvm-bolt %t.exe --relocs=1 --frame-opt=all --print-finalized \
17*d648aa1bSMaksim Panchenko# RUN:   --lite=0 --print-only=main --data %t.fdata -o %t.out | FileCheck %s
186b4eb0b9SAmir Ayupov
196b4eb0b9SAmir Ayupov# RUN: %t.out
206b4eb0b9SAmir Ayupov
216b4eb0b9SAmir Ayupov# CHECK: BOLT-INFO: Shrink wrapping moved 1 spills inserting load/stores and 0 spills inserting push/pops
226b4eb0b9SAmir Ayupov
236b4eb0b9SAmir Ayupov  .text
246b4eb0b9SAmir Ayupov  .globl bar
256b4eb0b9SAmir Ayupov  .type bar, %function
266b4eb0b9SAmir Ayupov  .p2align 4
276b4eb0b9SAmir Ayupovbar:
286b4eb0b9SAmir Ayupov# FDATA: 0 [unknown] 0 1 bar 0 0 510
296b4eb0b9SAmir Ayupov  pushq %rbp
306b4eb0b9SAmir Ayupov  movq  %rsp, %rbp
316b4eb0b9SAmir Ayupov  pushq %rbx                  # We save rbx here, but there is an
326b4eb0b9SAmir Ayupov                              # opportunity to move it to .BB2
336b4eb0b9SAmir Ayupov  subq  $0x18, %rsp
346b4eb0b9SAmir Ayupov  cmpl  $0x2, %edi
356b4eb0b9SAmir Ayupov.J1:
366b4eb0b9SAmir Ayupov  jb    .BBend
376b4eb0b9SAmir Ayupov# FDATA: 1 bar #.J1# 1 bar #.BB2# 0 10
386b4eb0b9SAmir Ayupov# FDATA: 1 bar #.J1# 1 bar #.BBend# 0 500
396b4eb0b9SAmir Ayupov.BB2:
406b4eb0b9SAmir Ayupov  movq $2, %rbx               # Use rbx in a cold block
416b4eb0b9SAmir Ayupov  xorq %rax, %rax
426b4eb0b9SAmir Ayupov  movb mystring, %al
436b4eb0b9SAmir Ayupov  addq %rbx, %rax
446b4eb0b9SAmir Ayupov  movb %al, mystring
456b4eb0b9SAmir Ayupov  leaq mystring, %rdi
466b4eb0b9SAmir Ayupov  #callq puts
476b4eb0b9SAmir Ayupov
486b4eb0b9SAmir Ayupov.BBend:
496b4eb0b9SAmir Ayupov  addq $0x18, %rsp
506b4eb0b9SAmir Ayupov  pop %rbx                    # Restore
516b4eb0b9SAmir Ayupov  xorq %rax, %rax
526b4eb0b9SAmir Ayupov  cmpq  $0x0, %rax
536b4eb0b9SAmir Ayupov  jne  .BBnever
546b4eb0b9SAmir Ayupov  jmp  .BBbarend
556b4eb0b9SAmir Ayupov.BBnever:
566b4eb0b9SAmir Ayupov  # This is a path that is never executed, but we add a call to main here
576b4eb0b9SAmir Ayupov  # to force a cycle in the call graph and to require us to have an aligned
586b4eb0b9SAmir Ayupov  # stack
596b4eb0b9SAmir Ayupov  callq main
606b4eb0b9SAmir Ayupov.BBbarend:
616b4eb0b9SAmir Ayupov  leaveq
626b4eb0b9SAmir Ayupov  retq
636b4eb0b9SAmir Ayupov  .size bar, .-bar
646b4eb0b9SAmir Ayupov
656b4eb0b9SAmir Ayupov# Frame alignedness information needs to be transmitted from foo to main to bar
666b4eb0b9SAmir Ayupov  .globl  main
676b4eb0b9SAmir Ayupov  .type main, %function
686b4eb0b9SAmir Ayupov  .p2align  4
696b4eb0b9SAmir Ayupovmain:
706b4eb0b9SAmir Ayupov  # Call a function that requires an aligned stack
716b4eb0b9SAmir Ayupov  callq foo
726b4eb0b9SAmir Ayupov  # Call a function that can be shrink-wrapped
736b4eb0b9SAmir Ayupov  callq bar
746b4eb0b9SAmir Ayupov  retq
756b4eb0b9SAmir Ayupov  .size main, .-main
766b4eb0b9SAmir Ayupov
776b4eb0b9SAmir Ayupov# Frame alignedness information needs to be transmitted from foo to main to bar
786b4eb0b9SAmir Ayupov  .globl  foo
796b4eb0b9SAmir Ayupov  .type foo, %function
806b4eb0b9SAmir Ayupov  .p2align  4
816b4eb0b9SAmir Ayupovfoo:
826b4eb0b9SAmir Ayupov  # Use an instruction that requires an aligned stack
836b4eb0b9SAmir Ayupov  movdqa -0x10(%rsp), %xmm0
846b4eb0b9SAmir Ayupov  retq
856b4eb0b9SAmir Ayupov  .size foo, .-foo
866b4eb0b9SAmir Ayupov
876b4eb0b9SAmir Ayupov  .data
886b4eb0b9SAmir Ayupovmystring: .asciz "0 is rbx mod 10 contents in decimal\n"
89