[BOLT] Push code to higher addresses under options

[maksfb]

When --hot-functions-at-end is used in combination with --use-old-text, allocate code at the highest possible addresses withing old .text.

This feature is mostly useful for HHVM, where it is beneficial to have hot static code placed as close as possible to jitted code.

[llvmbot]

@llvm/pr-subscribers-bolt

Author: Maksim Panchenko (maksfb)

Changes

When --hot-functions-at-end is used in combination with --use-old-text, allocate code at the highest possible addresses withing old .text.

This feature is mostly useful for HHVM, where it is beneficial to have hot static code placed as close as possible to jitted code.

---

Full diff: https://github.com/llvm/llvm-project/pull/146180.diff

2 Files Affected:

• (modified) bolt/lib/Rewrite/RewriteInstance.cpp (+30-3)
• (added) bolt/test/code-at-high-address.test (+19)

diff --git a/bolt/lib/Rewrite/RewriteInstance.cpp b/bolt/lib/Rewrite/RewriteInstance.cpp
index 934768c244b31..241b50b0bf793 100644
--- a/bolt/lib/Rewrite/RewriteInstance.cpp
+++ b/bolt/lib/Rewrite/RewriteInstance.cpp
@@ -3918,15 +3918,42 @@ void RewriteInstance::mapCodeSections(BOLTLinker::SectionMapper MapSection) {
       return Address;
     };
 
+    // Try to allocate sections before the \p Address and return an address for
+    // the allocation of the first section, or 0 if [0, Address) range is not
+    // big enough to fit all sections.
+    auto allocateBefore = [&](uint64_t Address) -> uint64_t {
+      for (auto SI = CodeSections.rbegin(), SE = CodeSections.rend(); SI != SE;
+           ++SI) {
+        BinarySection *Section = *SI;
+        if (Section->getOutputSize() > Address)
+          return 0;
+        Address -= Section->getOutputSize();
+        Address = alignDown(Address, Section->getAlignment());
+        Section->setOutputAddress(Address);
+      }
+      return Address;
+    };
+
     // Check if we can fit code in the original .text
     bool AllocationDone = false;
     if (opts::UseOldText) {
-      const uint64_t CodeSize =
-          allocateAt(BC->OldTextSectionAddress) - BC->OldTextSectionAddress;
+      uint64_t StartAddress;
+      uint64_t EndAddress;
+      if (opts::HotFunctionsAtEnd) {
+        EndAddress = BC->OldTextSectionAddress + BC->OldTextSectionSize;
+        StartAddress = allocateBefore(EndAddress);
+      } else {
+        StartAddress = BC->OldTextSectionAddress;
+        EndAddress = allocateAt(BC->OldTextSectionAddress);
+      }
 
+      const uint64_t CodeSize = EndAddress - StartAddress;
       if (CodeSize <= BC->OldTextSectionSize) {
         BC->outs() << "BOLT-INFO: using original .text for new code with 0x"
-                   << Twine::utohexstr(opts::AlignText) << " alignment\n";
+                   << Twine::utohexstr(opts::AlignText) << " alignment";
+        if (StartAddress != BC->OldTextSectionAddress)
+          BC->outs() << " at 0x" << Twine::utohexstr(StartAddress);
+        BC->outs() << '\n';
         AllocationDone = true;
       } else {
         BC->errs()
diff --git a/bolt/test/code-at-high-address.test b/bolt/test/code-at-high-address.test
new file mode 100644
index 0000000000000..4181133c573ef
--- /dev/null
+++ b/bolt/test/code-at-high-address.test
@@ -0,0 +1,19 @@
+## Check that llvm-bolt pushes code to higher addresses under
+## --hot-functions-at-end when rewriting code in-place.
+
+REQUIRES: system-linux
+
+RUN: %clang %cflags %p/Inputs/hello.c -o %t -no-pie -Wl,-q
+RUN: llvm-bolt %t -o %t.bolt --use-old-text --align-functions=1 \
+RUN:   --no-huge-pages --align-text=1 --use-gnu-stack --hot-functions-at-end \
+RUN:   | FileCheck %s --check-prefix=CHECK-BOLT
+RUN: llvm-readelf --sections %t.bolt | FileCheck %s
+
+CHECK-BOLT: using original .text for new code with 0x1 alignment at {{.*}}
+
+## As .text is pushed higher, preceding .bolt.org.text should have non-zero
+## size.
+CHECK: .bolt.org.text PROGBITS
+CHECK-NOT: {{ 000000 }}
+CHECK-SAME: AX
+CHECK-NEXT: .text PROGBITS

[github-actions]

✅ With the latest revision this PR passed the C/C++ code formatter.

[aaupov]

General question: with use-old-text and hot-functions-at-end, what's in the space that is not used? Is it the original code, or emitted padding/traps?

[maksfb]

General question: with use-old-text and hot-functions-at-end, what's in the space that is not used? Is it the original code, or emitted padding/traps?

Remnants of the old code. It's the same as without --hot-functions-at-end, just on the other side.