[BOLT] Support map other function entry address

[linsinan1995]

Allow BOLT to map the old address to a new binary address if the old address is the entry of the function.

[llvmbot]

@llvm/pr-subscribers-bolt

Author: sinan (linsinan1995)

Changes

Allow BOLT to map the old address to a new binary address if the old address is the entry of the function.

---

Full diff: https://github.com/llvm/llvm-project/pull/101466.diff

2 Files Affected:

• (modified) bolt/lib/Rewrite/RewriteInstance.cpp (+8)
• (added) bolt/test/X86/dynamic-relocs-on-entry.s (+25)

diff --git a/bolt/lib/Rewrite/RewriteInstance.cpp b/bolt/lib/Rewrite/RewriteInstance.cpp
index 9077869fe4955..e2273472c15de 100644
--- a/bolt/lib/Rewrite/RewriteInstance.cpp
+++ b/bolt/lib/Rewrite/RewriteInstance.cpp
@@ -5509,6 +5509,14 @@ uint64_t RewriteInstance::getNewFunctionOrDataAddress(uint64_t OldAddress) {
   if (const BinaryFunction *BF =
           BC->getBinaryFunctionContainingAddress(OldAddress)) {
     if (BF->isEmitted()) {
+      // If OldAddress is the another entry point of
+      // the function, then BOLT could get the new address.
+      if (BF->isMultiEntry()) {
+        for (auto &BB : *BF)
+          if (BB.isEntryPoint() &&
+              (BF->getAddress() + BB.getOffset()) == OldAddress)
+            return BF->getOutputAddress() + BB.getOffset();
+      }
       BC->errs() << "BOLT-ERROR: unable to get new address corresponding to "
                     "input address 0x"
                  << Twine::utohexstr(OldAddress) << " in function " << *BF
diff --git a/bolt/test/X86/dynamic-relocs-on-entry.s b/bolt/test/X86/dynamic-relocs-on-entry.s
new file mode 100644
index 0000000000000..7f47955440dd4
--- /dev/null
+++ b/bolt/test/X86/dynamic-relocs-on-entry.s
@@ -0,0 +1,25 @@
+// This test examines whether BOLT can correctly update when
+// dynamic relocation points to other entry points of the
+// function.
+
+# RUN: %clang %cflags -fPIC -pie %s -o %t.exe -nostdlib -Wl,-q
+# RUN: llvm-bolt %t.exe -o %t.bolt | FileCheck %s
+
+    .text
+    .type   chain, @function
+chain:
+    movq    $1, %rax
+Lable:
+    ret
+    .size   chain, .-chain
+    .type   _start, @function
+    .global _start
+_start:
+    jmpq    *.Lfoo(%rip)
+    ret
+    .size   _start, .-_start
+  .data
+.Lfoo:
+  .quad Lable
+
+# CHECK-NOT: BOLT-ERROR

[linsinan1995]

take the assembly code attached in the test as an example

    .text
    .type   chain, @function
chain:
    movq    $1, %rax
Lable:
    ret
    .size   chain, .-chain
    .type   _start, @function
    .global _start
_start:
    jmpq    *.Lfoo(%rip)
    ret
    .size   _start, .-_start
  .data
.Lfoo:
  .quad Lable

(1) if -no-pie, there is not a dynamic relocation for Label, then everything is fine.

gcc test.s -nostdlib -nodefaultlibs -Wl,-q
llvm-bolt a.out -o a.opt

(2) if it is pie, then an error occurs

gcc test.s -nostdlib -nodefaultlibs -Wl,-q -fPIC -fPIE -pie
llvm-bolt a.out -o a.opt

BOLT-ERROR: unable to get new address corresponding to input address 0x1007 in function chain/1(*2). Consider adding this function to --skip-funcs=...

This change can make the case (2) work

[yota9]

LGTM thanks!
JFYI next time you can use objdump with -R in order to skip extra file creation for readelf+objdump. It's a minor thing so I don't really mind anyway.

[ayermolo]

Looks like an X86 test is failing.

[yota9]

I think you need either remove # here or (probably even better) make var setting above like [[#%x,ADDR:]]

[yota9]

Test passed, thanks

[linsinan1995]

Thanks for reviewing && suggestions @ayermolo @yota9

I will merge the code by Wednesday if there are no new comments.

[yota9]

Also could you please push this one and #69136 to 19.x branch? Thanks!

[linsinan1995]

/cherry-pick 734c048

[llvmbot]

/pull-request #102282