[llvm][ELF] Add statistics on various section sizes

[aeubanks]

Useful with other infrastructure that consume LLVM statistics to get an idea of distribution of section sizes.

The breakdown of various section types is subject to change, this is just an initial go at gather some sort of stats.

Example stats compiling X86ISelLowering.cpp (-g1):

        "elf-object-writer.AllocROBytes": 308268,
        "elf-object-writer.AllocRWBytes": 6240,
        "elf-object-writer.AllocTextBytes": 1659203,
        "elf-object-writer.DebugBytes": 3180386,
        "elf-object-writer.OtherBytes": 5862,
        "elf-object-writer.RelocationBytes": 2623440,
        "elf-object-writer.StrtabBytes": 228599,
        "elf-object-writer.SymtabBytes": 120336,
        "elf-object-writer.UnwindBytes": 85216,

[llvmbot]

@llvm/pr-subscribers-backend-x86

@llvm/pr-subscribers-mc

Author: Arthur Eubanks (aeubanks)

Changes

Useful with other infrastructure that consume LLVM statistics to get an idea of distribution of section sizes.

The breakdown of various section types is subject to change, this is just an initial go at gather some sort of stats.

---

Full diff: https://github.com/llvm/llvm-project/pull/102363.diff

2 Files Affected:

• (modified) llvm/lib/MC/ELFObjectWriter.cpp (+48-2)
• (added) llvm/test/CodeGen/X86/section-stats.ll (+7)

diff --git a/llvm/lib/MC/ELFObjectWriter.cpp b/llvm/lib/MC/ELFObjectWriter.cpp
index c40a074137ab2..4d871e4c4bf5e 100644
--- a/llvm/lib/MC/ELFObjectWriter.cpp
+++ b/llvm/lib/MC/ELFObjectWriter.cpp
@@ -14,6 +14,7 @@
 #include "llvm/ADT/DenseMap.h"
 #include "llvm/ADT/STLExtras.h"
 #include "llvm/ADT/SmallVector.h"
+#include "llvm/ADT/Statistic.h"
 #include "llvm/ADT/StringExtras.h"
 #include "llvm/ADT/StringRef.h"
 #include "llvm/ADT/Twine.h"
@@ -62,10 +63,23 @@
 
 using namespace llvm;
 
-#undef  DEBUG_TYPE
-#define DEBUG_TYPE "reloc-info"
+#define DEBUG_TYPE "elf-object-writer"
 
 namespace {
+namespace stats {
+
+STATISTIC(AllocTextBytes, "Total size of SHF_ALLOC text sections");
+STATISTIC(AllocROBytes, "Total size of SHF_ALLOC readonly sections");
+STATISTIC(AllocRWBytes, "Total size of SHF_ALLOC read-write sections");
+STATISTIC(StrtabBytes, "Total size of SHT_STRTAB sections");
+STATISTIC(SymtabBytes, "Total size of SHT_SYMTAB sections");
+STATISTIC(RelocationBytes, "Total size of relocation sections");
+STATISTIC(DynsymBytes, "Total size of SHT_DYNSYM sections");
+STATISTIC(DebugBytes, "Total size of debug info sections");
+STATISTIC(UnwindBytes, "Total size of unwind sections");
+STATISTIC(OtherBytes, "Total size of uncategorized sections");
+
+} // namespace stats
 
 struct ELFWriter;
 
@@ -951,6 +965,38 @@ void ELFWriter::writeSectionHeader(const MCAssembler &Asm) {
     else
       Size = Offsets.second - Offsets.first;
 
+    auto SectionHasFlag = [&](uint64_t Flag) -> bool {
+      return Section->getFlags() & Flag;
+    };
+    auto SectionIsType = [&](uint64_t Type) -> bool {
+      return Section->getType() == Type;
+    };
+
+    if (SectionIsType(ELF::SHT_STRTAB)) {
+      stats::StrtabBytes += Size;
+    } else if (SectionIsType(ELF::SHT_SYMTAB)) {
+      stats::SymtabBytes += Size;
+    } else if (SectionIsType(ELF::SHT_DYNSYM)) {
+      stats::DynsymBytes += Size;
+    } else if (SectionIsType(ELF::SHT_REL) || SectionIsType(ELF::SHT_RELA) ||
+               SectionIsType(ELF::SHT_RELR) || SectionIsType(ELF::SHT_CREL)) {
+      stats::RelocationBytes += Size;
+    } else if (SectionIsType(ELF::SHT_X86_64_UNWIND)) {
+      stats::UnwindBytes += Size;
+    } else if (Section->getName().starts_with(".debug")) {
+      stats::DebugBytes += Size;
+    } else if (SectionHasFlag(ELF::SHF_ALLOC)) {
+      if (SectionHasFlag(ELF::SHF_EXECINSTR)) {
+        stats::AllocTextBytes += Size;
+      } else if (SectionHasFlag(ELF::SHF_WRITE)) {
+        stats::AllocRWBytes += Size;
+      } else {
+        stats::AllocROBytes += Size;
+      }
+    } else {
+      stats::OtherBytes += Size;
+    }
+
     writeSection(GroupSymbolIndex, Offsets.first, Size, *Section);
   }
 }
diff --git a/llvm/test/CodeGen/X86/section-stats.ll b/llvm/test/CodeGen/X86/section-stats.ll
new file mode 100644
index 0000000000000..26d9cf2d60676
--- /dev/null
+++ b/llvm/test/CodeGen/X86/section-stats.ll
@@ -0,0 +1,7 @@
+; RUN: llc -o /dev/null -filetype=obj -stats %s 2>&1 | FileCheck %s
+
+; CHECK: {{[0-9+]}} elf-object-writer - Total size of SHF_ALLOC text sections
+
+define void @f() {
+    ret void
+}

[rnk]

Neat. Can you pick some LLVM object file of interest, say X86ISelLowering.cpp.o, and paste the stats output for that file, just to get a sense of what the data looks like?

I think this looks good, but I'd like a second reviewer opinion.

[aeubanks]

added numbers in the description

[zmodem]

lgtm

[MaskRay]

Perhaps check SHF_ALLOC first, then .debug, then use a switch on getType()? Then, SectionIsType could be removed.

[aeubanks]

done. I moved the check for eh_frame first since it's also SHF_ALLOC

[MaskRay]

== ".eh_frame"

The design makes it monolithic.

[MaskRay]

object writer doesn't generate SHT_RELR

[aeubanks]

done

[MaskRay]

non-x86-64 arches don't use SHT_X86_64_UNWIND. Perhaps just do a name comparison with .eh_frame

[aeubanks]

done

[MaskRay]

[[#]]. I think it's fine to test the exact number to ensure that we don't regress.

You can add a variable and test SHF_WRITE as well.

[aeubanks]

done