[BOLT] Fix shrinkwrapping bug when changing frame alignment

Summary:
This fixes a bug with shrink wrapping when trying to move
push-pops in a function where we are not allowed to modify the
stack layout for alignment reasons. In this bug, we failed to
propagate alignment requirement upwards in the call graph from
function A to B when: (1) there is a cycle in the call graph and
(2) the distance from A to B is greater than 1 in the call graph
and (3) there is a node in the path from A to B, not including
A or B, that does not access parameters in the stack.

(cherry picked from FBD25315977)
This commit is contained in:
Rafael Auler 2020-12-03 20:09:32 -08:00 committed by Maksim Panchenko
parent e067f2adf4
commit d2f68039bc
2 changed files with 89 additions and 1 deletions

View File

@ -363,7 +363,7 @@ bool FrameAnalysis::updateArgsTouchedFor(const BinaryFunction &BF, MCInst &Inst,
FunctionsRequireAlignment.insert(&BF);
}
if (Iter == ArgsTouchedMap.end())
return false;
return Changed;
if (CurOffset == StackPointerTracking::EMPTY ||
CurOffset == StackPointerTracking::SUPERPOSITION) {

View File

@ -0,0 +1,88 @@
# This reproduces a bug with shrink wrapping when trying to move
# push-pops in a function where we are not allowed to modify the
# stack layout for alignment reasons. In this bug, we failed to
# propagate alignment requirement upwards in the call graph for
# some functions when there is a cycle in the call graph.
# REQUIRES: system-linux
# RUN: llvm-mc -filetype=obj -triple x86_64-unknown-unknown \
# RUN: %s -o %t.o
# RUN: link_fdata %s %t.o %t.fdata
# Delete our BB symbols so BOLT doesn't mark them as entry points
# RUN: strip --strip-unneeded %t.o
# RUN: %host_cc %t.o -o %t.exe -Wl,-q
# RUN: llvm-bolt %t.exe -relocs=1 -frame-opt=all -print-finalized \
# RUN: -lite=0 -print-only=main -data %t.fdata -o %t.out | FileCheck %s
# RUN: %t.out
# CHECK: BOLT-INFO: Shrink wrapping moved 1 spills inserting load/stores and 0 spills inserting push/pops
.text
.globl bar
.type bar, %function
.p2align 4
bar:
# FDATA: 0 [unknown] 0 1 bar 0 0 510
pushq %rbp
movq %rsp, %rbp
pushq %rbx # We save rbx here, but there is an
# opportunity to move it to .BB2
subq $0x18, %rsp
cmpl $0x2, %edi
.J1:
jb .BBend
# FDATA: 1 bar #.J1# 1 bar #.BB2# 0 10
# FDATA: 1 bar #.J1# 1 bar #.BBend# 0 500
.BB2:
movq $2, %rbx # Use rbx in a cold block
xorq %rax, %rax
movb mystring, %al
addq %rbx, %rax
movb %al, mystring
leaq mystring, %rdi
#callq puts
.BBend:
addq $0x18, %rsp
pop %rbx # Restore
xorq %rax, %rax
cmpq $0x0, %rax
jne .BBnever
jmp .BBbarend
.BBnever:
# This is a path that is never executed, but we add a call to main here
# to force a cycle in the call graph and to require us to have an aligned
# stack
callq main
.BBbarend:
leaveq
retq
.size bar, .-bar
# Frame alignedness information needs to be transmitted from foo to main to bar
.globl main
.type main, %function
.p2align 4
main:
# Call a function that requires an aligned stack
callq foo
# Call a function that can be shrink-wrapped
callq bar
retq
.size main, .-main
# Frame alignedness information needs to be transmitted from foo to main to bar
.globl foo
.type foo, %function
.p2align 4
foo:
# Use an instruction that requires an aligned stack
movdqa -0x10(%rsp), %xmm0
retq
.size foo, .-foo
.data
mystring: .asciz "0 is rbx mod 10 contents in decimal\n"