[CommandLine] Reduce size of Option class

[hintonda]

Summary

Reduce size of Option class from 184 bytes to 136 bytes by
placing more member variables in Bit Field (16 bytes), and
reducing the initial sizes of Categories and Subs to 1 (32 bytes).

Saves about 48k for bin/opt.

[beanz]

This is pretty cool. Since these are usually allocated as globals, this is saving pages of dirtied memory.

One implementation thought. What if we group the 16-bit fields together and make them uint16_t then use a uint16_t as the bitfield type for the remaining 15 bits. The performance of the 16-bit fields should be better if they are aligned (not that cl::opt perf really matters).

[hintonda]

• Change type to multiple uint16_t's and reorder members.

[hintonda]

• Remove unneeded bit fields.

[hintonda]

In D62091#1507424, @beanz wrote:

This is pretty cool. Since these are usually allocated as globals, this is saving pages of dirtied memory.

One implementation thought. What if we group the 16-bit fields together and make them uint16_t then use a uint16_t as the bitfield type for the remaining 15 bits. The performance of the 16-bit fields should be better if they are aligned (not that cl::opt perf really matters).

I tried to keep the relative order as close to the original as possible, but after rereading your comment, I'm unsure if you wanted me to group the non-bit field unit16_t's together as well. Please let me know...

[beanz]

The main thing is to make sure that if you have proper integer sized members they are properly aligned and packed, which this seems to do well.

llvm/include/llvm/Support/CommandLine.h

278 ↗

(On Diff #200134)

If I'm counting right (and math isn't my strong point) we are at 15 bits used in the bitfield. I would add another 1-bit field named "Unused" just to mark out that we have one more bit before the bitfield would grow.

If the bitfield grows, it will result in 7-bytes of padding coming back in, which would be unfortunate.

[beanz]

llvm/include/llvm/Support/CommandLine.h

288 ↗

(On Diff #200134)

Side note: I wonder how many members actually get put into this. The common-case for most options is surely 0.

[beanz]

llvm/include/llvm/Support/CommandLine.h

288 ↗

(On Diff #200134)

From eyeballing (not an exhaustive search), I don't see anywhere that more than one Sub command is set. You might be able to save more space by shrinking the inline size here. If I read the code right, SmallPtrSet is 32 bytes + (sizeof (T) * N). In this case that should be 64 bytes. Changing the size to 2, would save another 16 bytes per option.

[hintonda]

llvm/include/llvm/Support/CommandLine.h

288 ↗

(On Diff #200134)

Actually, it's empty about 99% or the time and defaults to TopLevelSubCommand at that point. This is how all the options in passes are setup -- well over 1,000. Only a few tools currently use SubCommands -- the most I've seen is 4 by llvm-pdbutil.

So, I was considering something a bit more radical. Instead of each Option keeping track of SubCommand, and Category, membership, why not let the SubCommand's and Category's do it instead?

[hintonda]

llvm/include/llvm/Support/CommandLine.h

288 ↗

(On Diff #200134)

So as a quick test, I set both to 1 and reduced the overall size down to 136 bytes -- Categories always has at least 1, but rarely more, and Subs is almost always empty:

--- a/llvm/include/llvm/Support/CommandLine.h
+++ b/llvm/include/llvm/Support/CommandLine.h
@@ -283,9 +283,9 @@ public:
   StringRef ArgStr;   // The argument string itself (ex: "help", "o")
   StringRef HelpStr;  // The descriptive text message for -help
   StringRef ValueStr; // String describing what the value of this option is
-  SmallVector<OptionCategory *, 2>
+  SmallVector<OptionCategory *, 1>
       Categories;                    // The Categories this option belongs to
-  SmallPtrSet<SubCommand *, 4> Subs; // The subcommands this option belongs to.
+  SmallPtrSet<SubCommand *, 1> Subs; // The subcommands this option belongs to.

[hintonda]

• Reduce initial vector/set size to the expected value.

llvm/trunk/include/llvm/Support/CommandLine.h

class Option {
 
   virtual enum ValueExpected getValueExpectedFlagDefault() const {
     return ValueOptional;
   }
 
   // Out of line virtual function to provide home for the class.
   virtual void anchor();
 
-  int NumOccurrences = 0; // The number of times specified
+  uint16_t NumOccurrences; // The number of times specified
   // Occurrences, HiddenFlag, and Formatting are all enum types but to avoid
   // problems with signed enums in bitfields.
-  unsigned Occurrences : 3; // enum NumOccurrencesFlag
+  uint16_t Occurrences : 3; // enum NumOccurrencesFlag
   // not using the enum type for 'Value' because zero is an implementation
   // detail representing the non-value
-  unsigned Value : 2;
-  unsigned HiddenFlag : 2; // enum OptionHidden
-  unsigned Formatting : 2; // enum FormattingFlags
-  unsigned Misc : 5;
-  unsigned Position = 0;       // Position of last occurrence of the option
-  unsigned AdditionalVals = 0; // Greater than 0 for multi-valued option.
+  uint16_t Value : 2;
+  uint16_t HiddenFlag : 2; // enum OptionHidden
+  uint16_t Formatting : 2; // enum FormattingFlags
+  uint16_t Misc : 5;
+  uint16_t FullyInitialized : 1; // Has addArgument been called?
+  uint16_t Position;             // Position of last occurrence of the option
+  uint16_t AdditionalVals;       // Greater than 0 for multi-valued option.
 
 public:
   StringRef ArgStr;   // The argument string itself (ex: "help", "o")
   StringRef HelpStr;  // The descriptive text message for -help
   StringRef ValueStr; // String describing what the value of this option is
-  SmallVector<OptionCategory *, 2>
+  SmallVector<OptionCategory *, 1>
       Categories;                    // The Categories this option belongs to
-  SmallPtrSet<SubCommand *, 4> Subs; // The subcommands this option belongs to.
-  bool FullyInitialized = false; // Has addArgument been called?
+  SmallPtrSet<SubCommand *, 1> Subs; // The subcommands this option belongs to.
 
   inline enum NumOccurrencesFlag getNumOccurrencesFlag() const {
     return (enum NumOccurrencesFlag)Occurrences;
   }
 
   inline enum ValueExpected getValueExpectedFlag() const {
     return Value ? ((enum ValueExpected)Value) : getValueExpectedFlagDefault();
   }
public:
   void setMiscFlag(enum MiscFlags M) { Misc |= M; }
   void setPosition(unsigned pos) { Position = pos; }
   void addCategory(OptionCategory &C);
   void addSubCommand(SubCommand &S) { Subs.insert(&S); }
 
 protected:
   explicit Option(enum NumOccurrencesFlag OccurrencesFlag,
                   enum OptionHidden Hidden)
-      : Occurrences(OccurrencesFlag), Value(0), HiddenFlag(Hidden),
-        Formatting(NormalFormatting), Misc(0) {
+      : NumOccurrences(0), Occurrences(OccurrencesFlag), Value(0),
+        HiddenFlag(Hidden), Formatting(NormalFormatting), Misc(0),
+        FullyInitialized(false), Position(0), AdditionalVals(0) {
     Categories.push_back(&GeneralCategory);
   }
 
   inline void setNumAdditionalVals(unsigned n) { AdditionalVals = n; }
 
 public:
   virtual ~Option() = default;