[llvm-objdump] Print Mnemonic Histogram

[SjoerdMeijer]

Summary

This adds new option --mnemonic-hist to print a histogram of all (static) instructions/mnemonics. For example:

$ llvm-objdump --triple=thumbv7 -d --mnemonic-hist
..
Instruction histogram:
         ldr:   120 (27.3973%)
         mov:    96 (21.9178%)
         blx:    56 (12.7854%)
         bl:    31 (7.07763%)
         str:    26 (5.93607%)
         add:    18 (4.10959%)
         b:    12 (2.73973%)
         sub:    10 (2.28311%)
         cmp:     9 (2.05479%)
         ..

I am probably interested in printing more information, like the encoding width, but this seems a good first step.

[DavidSpickett]

I don't know what bar there is to add stuff like this but I assume it's quite low given that it's self contained and not trying to match any equivalent GNU options. I can see the use for comparing files certainly.

llvm/test/tools/llvm-objdump/ARM/mnemonic-hist1.test

17

If you're going to pad the names to make them all line up then you should test something that isn't 3 characters. Also, and this makes the test bigger but check that the number padding works. Maybe just make one appear 10 times, unless you can figure out a macro to emit 100 of something.

Also do you expect an ordering between the 2 instructions that both have 1? As in should it be alphabetical between those 2 or is it going to depend on order of appearance in the object file.

Edit: Putting them in a std::map sorts them so you would always see mul then sub even if sub appears first in the object, correct?

Can you comment in the test file what things it's looking for.

llvm/tools/llvm-objdump/llvm-objdump.cpp

1491

std::ignore here? Or std::get<0>(IP->getMnemonic(&Inst)) directly.

2171

Probably makes no difference but why not const std::pair<std::string, unsigned> &.

[tschuett]

llvm/tools/llvm-objdump/llvm-objdump.cpp

2171

Maybe std::stable_sort to make it deterministic?

[MaskRay]

I don't know what bar there is to add stuff like this but I assume it's quite low given that it's self contained and not trying to match any equivalent GNU options. I can see the use for comparing files certainly.

I think there are two bars:

• whether we can add an option that GNU objdump doesn't have: in my understanding this is low, but we really want to avoid potential conflict with GNU. This means we usually cannot add a short option alias as GNU objdump may take a short option for anything in the future (e.g. llvm-nm -U conflict I happen to observe today). A descriptive long option usually has no semantic conflict concern.
• whether we should add an option: I think the bar is actually quite high. An option needs to be sufficiently useful to justify its addition.

I am probably interested in printing more information, like the encoding width, but this seems a good first step.

If we add an option, we want to provide some stability guarantee. Removing an option has an extremely high bar as that may break some users.
So when adding an option we need to very careful. If there is a plan to extend the option, it's best to design it well upfront.
I know that sometimes iterative process may be unavoidable due to engineering reality, in that case perhaps we can call an option experimental.
A new option needs a documentation update in llvm/docs/CommandGuide/llvm-objdump.rst

That said, for this case I wonder why we can't just use

llvm-objdump -d --no-show-raw-insn --no-leading-addr =cat | awk '/^ / {freq[$1]++; tot++} END{ for(i in freq) print i "\t" freq[i] "\t" (freq[i]/tot) }'

My reasoning is that different users may have vastly different statistics needs.
If it is difficult to come up with something that meets many's needs, I'd more like the user to compose utilities, but I think I can be persuaded the other way if you can come up with something useful which seems to benefit a lot from built-in support.

I happen to be more concerned with the disassembly stuff because the code (as you may have noticed) is quite difficult to maintain: you can find numerous toggles adapting its behavior. if (Disassembled && MnemonicHistogram) { will not be the thing that will bring everything crashing down but I'd hope that adding any stuff in this area should take more attention.

[SjoerdMeijer]

Thanks all for looking at this!

I understand the concerns about adding/removing options. What I will do is leave a message on discourse to possibly draw some attention to this and check if anyone else has objections/preferences.

In the mean time I will work on a new revision of this which will distinguish between narrow and wide instruction encodings, and see how that fits in. That will serve two purposes. First, it's what I would really need for my use case, although in its current form it's already useful. Second, although not impossible, that would be more difficult to achieve with a one-liner on the command line. That would require a script, and that could be the alternative of this.

Personally, I prefer the built-in option because of its convenience. I could never remember that one-liner command line, would need to save it to a file, which I then don't have available on the different systems I work on etc. A python script contributed to llvm/utils for example would have this problem a lot less, but still wouldn't be as convenient as a built-in option.

[SjoerdMeijer]

https://discourse.llvm.org/t/llvm-objdump-print-instruction-histogram/62333

[SjoerdMeijer]

I haven't addressed feedback yet, but have just added a few things to help with the discussion where this belongs.
This now also prints the number of bytes per instruction, so for thumb it now distinguishes between the narrow/wide encodings, and I have also added printing a histogram of immediates. This will look now look like this:

  Mnemonic histogram:
  Mnemonic Freq             Bytes
  ldr (T1): 112 ( 25.5708%)   224 ( 22.0472%)
  mov (T1):  95 ( 21.6895%)   190 ( 18.7008%)
  blx (T1):  56 ( 12.7854%)   112 ( 11.0236%)
   bl (T2):  31 ( 7.07763%)   124 ( 12.2047%)
  str (T1):  23 ( 5.25114%)    46 ( 4.52756%)
    b (T1):  11 ( 2.51142%)    22 ( 2.16535%)
  add (T1):  11 ( 2.51142%)    22 ( 2.16535%)
  sub (T1):  10 ( 2.28311%)    20 (  1.9685%)
  cmp (T1):   8 ( 1.82648%)    16 (  1.5748%)
  ldr (T2):   8 ( 1.82648%)    32 ( 3.14961%)
  add (T2):   7 ( 1.59817%)    28 ( 2.75591%)
  asr (T1):   6 ( 1.36986%)    12 (  1.1811%)
 ldrb (T1):   5 ( 1.14155%)    10 (0.984252%)
 ..
Immediate histogram:
      14:       425
       0:        40
      -4:        32
       3:        26
       1:        21
      10:        15
       2:        14
       8:        11
       5:         9
      16:         8
      18:         6

llvm/test/tools/llvm-objdump/ARM/mnemonic-hist0.test

+#RUN: llvm-mc -filetype=obj -triple=thumbv7 %s -o - | \
+#RUN: llvm-objdump --triple=thumbv7 -d --mnemonic-hist - | FileCheck %s
+
+        .text
+        .thumb
+foo:
+
+#CHECK: Mnemonic histogram:{{[[:space:]]}}
+#CHECK: {{[[:space:]]}}
+#CHECK-NOT: {{.*}}:

llvm/test/tools/llvm-objdump/ARM/mnemonic-hist1.test

+#RUN: llvm-mc -filetype=obj -triple=thumbv7 %s -o - | \
+#RUN: llvm-objdump --triple=thumbv7 -d --mnemonic-hist - | FileCheck %s
+
+        .syntax unified
+        .text
+        .thumb
+
+foo:
+        add r0,r1,r2
+        sub r2,r1,r3
+        add r4,r0,r2
+        mul r5,r4,r1
+
+#CHECK:Mnemonic histogram:
+#CHECK: add: 2 (50%)
+#CHECK: mul: 1 (25%)
+#CHECK: sub: 1 (25%)

llvm/tools/llvm-objdump/ObjdumpOpts.td

 
 def syms : Flag<["--"], "syms">,
   HelpText<"Display the symbol table">;
 def : Flag<["-"], "t">, Alias<syms>, HelpText<"Alias for --syms">;
 
 def symbolize_operands : Flag<["--"], "symbolize-operands">,
   HelpText<"Symbolize instruction operands when disassembling">;
 
+def mnemonic_hist: Flag<["--"], "mnemonic-hist">,
+  HelpText<"Display a histogram of all mnemonics">;
+
 def dynamic_syms : Flag<["--"], "dynamic-syms">,
   HelpText<"Display the contents of the dynamic symbol table">;
 def : Flag<["-"], "T">, Alias<dynamic_syms>,
   HelpText<"Alias for --dynamic-syms">;
 
 def triple_EQ : Joined<["--"], "triple=">,
   HelpText<"Target triple to disassemble for, "
             "see --version for available targets">;

llvm/tools/llvm-objdump/llvm-objdump.cpp

 
 static uint64_t StartAddress;
 static bool HasStartAddressFlag;
 static uint64_t StopAddress = UINT64_MAX;
 static bool HasStopAddressFlag;
 
 bool objdump::SymbolTable;
 static bool SymbolizeOperands;
+static bool MnemonicHistogram;
+std::map<std::string, unsigned> Hist;
+
 static bool DynamicSymbolTable;
 std::string objdump::TripleName;
 bool objdump::UnwindInfo;
 static bool Wide;
 std::string objdump::Prefix;
 uint32_t objdump::PrefixStrip;
 
 DebugVarsFormat objdump::DbgVariables = DVDisabled;
for (unsigned SI = 0, SE = Symbols.size(); SI != SE; ++SI) {
 
           IP->setCommentStream(CommentStream);
 
           PIP.printInst(
               *IP, Disassembled ? &Inst : nullptr, Bytes.slice(Index, Size),
               {SectionAddr + Index + VMAAdjustment, Section.getIndex()}, FOS,
               "", *STI, &SP, Obj->getFileName(), &Rels, LVP);
 
+          if (Disassembled && MnemonicHistogram) {
+            const char *Mnemonic;
+            int Bytes;
+            std::tie(Mnemonic, Bytes) = IP->getMnemonic(&Inst);
+            Hist[Mnemonic]++;
+          }
+
           IP->setCommentStream(llvm::nulls());
 
           // If disassembly has failed, avoid analysing invalid/incomplete
           // instruction information. Otherwise, try to resolve the target
           // address (jump target or memory operand address) and print it on the
           // right of the instruction.
           if (Disassembled && MIA) {
             // Branch targets are printed just after the instructions.
if (Demangle)
     SymName = demangle(SymName);
 
   if (O->isXCOFF() && SymbolDescription)
     SymName = getXCOFFSymbolDescription(createSymbolInfo(O, Symbol), SymName);
 
   outs() << ' ' << SymName << '\n';
 }
 
+static void printMnemonicHistogram(const ObjectFile *O) {
+  std::vector<std::pair<std::string, unsigned>> List;
+  for (auto &I : Hist)
+    List.push_back(I);
+
+  std::sort(List.begin(), List.end(), [] (std::pair<std::string, unsigned> &L,
+                                          std::pair<std::string, unsigned> &R) {
+                                            return L.second > R.second; });
+
+  unsigned Total = 0;
+  unsigned MaxStrLen = 0;
+  for (auto &I : List) {
+    Total += I.second;
+    MaxStrLen = I.first.length() > MaxStrLen ? I.first.length() : MaxStrLen;
+  }
+
+  unsigned MaxDigits = List.empty() ? 0 :
+      unsigned(std::log10(List.front().second) + 1);
+ 
+  outs() << "\nMnemonic histogram:\n\n";
+  for (auto &I : List) {
+    outs() << format("%*s", MaxStrLen + 1, I.first.c_str());
+    outs() << ":";
+    outs() << format("%*d", MaxDigits + 1, I.second);
+    outs() << " (";
+    outs() << format("%g", ((float)I.second / (float)Total) * 100);
+    outs() << "%)\n";
+  }
+}
+
 static void printUnwindInfo(const ObjectFile *O) {
   outs() << "Unwind info:\n\n";
 
   if (const COFFObjectFile *Coff = dyn_cast<COFFObjectFile>(O))
     printCOFFUnwindInfo(Coff);
   else if (const MachOObjectFile *MachO = dyn_cast<MachOObjectFile>(O))
     printMachOUnwindInfo(MachO);
   else
static void dumpObject(ObjectFile *O, const Archive *A = nullptr,
   if (Relocations && !Disassemble)
     printRelocations(O);
   if (DynamicRelocations)
     printDynamicRelocations(O);
   if (SectionContents)
     printSectionContents(O);
   if (Disassemble)
     disassembleObject(O, Relocations);
+  if (MnemonicHistogram)
+    printMnemonicHistogram(O);
   if (UnwindInfo)
     printUnwindInfo(O);
 
   // Mach-O specific options:
   if (ExportsTrie)
     printExportsTrie(O);
   if (Rebase)
     printRebaseTable(O);
static void parseObjdumpOptions(const llvm::opt::InputArgList &InputArgs) {
   ShowLMA = InputArgs.hasArg(OBJDUMP_show_lma);
   PrintSource = InputArgs.hasArg(OBJDUMP_source);
   parseIntArg(InputArgs, OBJDUMP_start_address_EQ, StartAddress);
   HasStartAddressFlag = InputArgs.hasArg(OBJDUMP_start_address_EQ);
   parseIntArg(InputArgs, OBJDUMP_stop_address_EQ, StopAddress);
   HasStopAddressFlag = InputArgs.hasArg(OBJDUMP_stop_address_EQ);
   SymbolTable = InputArgs.hasArg(OBJDUMP_syms);
   SymbolizeOperands = InputArgs.hasArg(OBJDUMP_symbolize_operands);
+  MnemonicHistogram = InputArgs.hasArg(OBJDUMP_mnemonic_hist);
   DynamicSymbolTable = InputArgs.hasArg(OBJDUMP_dynamic_syms);
   TripleName = InputArgs.getLastArgValue(OBJDUMP_triple_EQ).str();
   UnwindInfo = InputArgs.hasArg(OBJDUMP_unwind_info);
   Wide = InputArgs.hasArg(OBJDUMP_wide);
   Prefix = InputArgs.getLastArgValue(OBJDUMP_prefix).str();
   parseIntArg(InputArgs, OBJDUMP_prefix_strip, PrefixStrip);
   if (const opt::Arg *A = InputArgs.getLastArg(OBJDUMP_debug_vars_EQ)) {
     DbgVariables = StringSwitch<DebugVarsFormat>(A->getValue())