[BOLT] Fix Linux kernel static keys handling

[FLZ101]

Fix the following issues:

• Cursor.tell() + (int32_t)DE.getU32(Cursor) is undefined

• In bolt/test/X86/linux-static-keys.s, __start___jump_table
  is unaligned:

  ffffffff800001fd g     O .rodata        0000000000000000 __start___jump_table
  

  This makes the test case undeterministic.

[llvmbot]

@llvm/pr-subscribers-bolt

Author: Franklin (FLZ101)

Changes

Fix the following issues:

• Cursor.tell() + (int32_t)DE.getU32(Cursor) is undefined

• bolt/test/X86/linux-static-keys.s generates a binary whose .rodata is unaligned:

  3 .rodata       00000020  ffffffff800001fd  ffffffff800001fd  000001fd  2**0
                  CONTENTS, ALLOC, LOAD, READONLY, DATA
  

---

Full diff: https://github.com/llvm/llvm-project/pull/119557.diff

2 Files Affected:

• (modified) bolt/lib/Rewrite/LinuxKernelRewriter.cpp (+4-3)
• (modified) bolt/test/X86/linux-static-keys.s (+3-2)

diff --git a/bolt/lib/Rewrite/LinuxKernelRewriter.cpp b/bolt/lib/Rewrite/LinuxKernelRewriter.cpp
index 03b414b71caca7..7409c33ec84e18 100644
--- a/bolt/lib/Rewrite/LinuxKernelRewriter.cpp
+++ b/bolt/lib/Rewrite/LinuxKernelRewriter.cpp
@@ -1660,12 +1660,13 @@ Error LinuxKernelRewriter::readStaticKeysJumpTable() {
   DataExtractor::Cursor Cursor(StaticKeysJumpTableAddress - SectionAddress);
   uint32_t EntryID = 0;
   while (Cursor && Cursor.tell() < Stop->getAddress() - SectionAddress) {
+    uint64_t _;
     const uint64_t JumpAddress =
-        SectionAddress + Cursor.tell() + (int32_t)DE.getU32(Cursor);
+        (_ = SectionAddress + Cursor.tell(), _ + (int32_t)DE.getU32(Cursor));
     const uint64_t TargetAddress =
-        SectionAddress + Cursor.tell() + (int32_t)DE.getU32(Cursor);
+        (_ = SectionAddress + Cursor.tell(), _ + (int32_t)DE.getU32(Cursor));
     const uint64_t KeyAddress =
-        SectionAddress + Cursor.tell() + (int64_t)DE.getU64(Cursor);
+        (_ = SectionAddress + Cursor.tell(), _ + (int64_t)DE.getU64(Cursor));
 
     // Consume the status of the cursor.
     if (!Cursor)
diff --git a/bolt/test/X86/linux-static-keys.s b/bolt/test/X86/linux-static-keys.s
index 0bd17a375d8824..465e865458d902 100644
--- a/bolt/test/X86/linux-static-keys.s
+++ b/bolt/test/X86/linux-static-keys.s
@@ -35,13 +35,13 @@ _start:
 .L0:
   jmp L1
 # CHECK:      jit
-# CHECK-SAME: # ID: 1 {{.*}} # Likely: 0 # InitValue: 1
+# CHECK-SAME: # ID: 1 {{.*}} # Likely: 1 # InitValue: 0
   nop
 L1:
   .nops 5
   jmp .L0
 # CHECK:      jit
-# CHECK-SAME: # ID: 2 {{.*}} # Likely: 1 # InitValue: 1
+# CHECK-SAME: # ID: 2 {{.*}} # Likely: 0 # InitValue: 0
 
 ## Check that a branch profile associated with a NOP is handled properly when
 ## dynamic branch is created.
@@ -65,6 +65,7 @@ foo:
   .rodata
   .globl __start___jump_table
   .type __start___jump_table, %object
+  .align 8
 __start___jump_table:
 
   .long .L0 - . # Jump address

[maksfb]

Thank you for identifying the problem, actually two of them. I don't think we are currently affected by the order of evaluation, but it's the proper fix.

The other one with the test case is caused by the key address field, alignment just exposes it. I'v put up #119771 with the fix for the field.

[FLZ101]

Since the test case is fixed by #119771. Close this PR for now.