[BOLT][AArch64] Remove nops in functions with defined control flow

When a function has an indirect branch with unknown control flow, we
preserve nops in order to keep all instruction offsets (from the start
of the function) the same in case the indirect branch is used by a
PC-relative jump table. However, when we know the control flow of the
function, we should be able to safely remove nops.

Author: maksfb

bolt/lib/Core/BinaryFunction.cpp

@@ -795,7 +795,6 @@ BinaryFunction::processIndirectBranch(MCInst &Instruction, unsigned Size,
 
   auto Begin = Instructions.begin();
   if (BC.isAArch64()) {
-    PreserveNops = BC.HasRelocations;
     // Start at the last label as an approximation of the current basic block.
     // This is a heuristic, since the full set of labels have yet to be
     // determined
@@ -2300,6 +2299,10 @@ Error BinaryFunction::buildCFG(MCPlusBuilder::AllocatorIdTy AllocatorId) {
       BC.errs() << "BOLT-WARNING: failed to post-process indirect branches for "
                 << *this << '\n';
     }
+
+    if (BC.isAArch64())
+      PreserveNops = BC.HasRelocations;
+
     // In relocation mode we want to keep processing the function but avoid
     // optimizing it.
     setSimple(false);

bolt/test/AArch64/remove-nops.s

@@ -0,0 +1,28 @@
+## Verify that llvm-bolt removes nop instructions from functions with indirect
+## branches that have defined control flow.
+
+# REQUIRES: system-linux
+
+# RUN: llvm-mc -filetype=obj -triple aarch64-unknown-unknown %s -o %t.o
+# RUN: %clang %cflags --target=aarch64-unknown-linux %t.o -o %t.exe -Wl,-q
+# RUN: llvm-bolt %t.exe -o %t.bolt --print-normalized 2>&1 | FileCheck %s
+# RUN: llvm-objdump -d --disassemble-symbols=_start %t.bolt \
+# RUN:   | FileCheck %s --check-prefix=CHECK-OBJDUMP
+
+# CHECK-OBJDUMP-LABEL: _start
+# CHECK-OBJDUMP-NOT: nop
+
+  .section .text
+  .align 4
+  .globl _start
+  .type  _start, %function
+_start:
+# CHECK-LABEL: Binary Function "_start"
+  nop
+# CHECK-NOT: nop
+  br      x0
+# CHECK: br x0 # TAILCALL
+.size _start, .-_start
+
+## Force relocation mode.
+  .reloc 0, R_AARCH64_NONE