[llvm] [TableGen] Use a more efficient memory buffer for output (PR #123353)

[via llvm-commits]

llvmbot wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-tablegen

Author: Jay Foad (jayfoad)

<details>
<summary>Changes</summary>

TableGen writes all output to an in-memory buffer in case the
-write-if-change option is being used. Using raw_string_ostream for this
is inefficient because all the data has to be copied every time the
underlying std::string is resized. Fix this by writing to a custom
raw_ostream which stores the buffered data as the concatenation of a
vector of strings of increasing capacity. Each string in the vector is
never resized beyond its initial capacity to avoid unnecessary copying.


---
Full diff: https://github.com/llvm/llvm-project/pull/123353.diff


1 Files Affected:

- (modified) llvm/lib/TableGen/Main.cpp (+49-4) 


``````````diff
diff --git a/llvm/lib/TableGen/Main.cpp b/llvm/lib/TableGen/Main.cpp
index 55a99cbfc58acd..88bca04ec19a41 100644
--- a/llvm/lib/TableGen/Main.cpp
+++ b/llvm/lib/TableGen/Main.cpp
@@ -37,6 +37,52 @@
 #include <utility>
 using namespace llvm;
 
+class stringvec_ostream : public raw_ostream {
+  std::vector<std::string> V;
+
+  size_t Pos = 0;
+  uint64_t current_pos() const override { return Pos; }
+
+  void write_impl(const char *Ptr, size_t Size) override {
+    Pos += Size;
+
+    size_t ThisSize = std::min(Size, V.back().capacity() - V.back().size());
+    V.back().append(Ptr, ThisSize);
+    Ptr += ThisSize;
+    Size -= ThisSize;
+
+    if (Size != 0) {
+      size_t NewCapacity = std::max(Size, V.back().capacity() * 2);
+      V.emplace_back();
+      V.back().reserve(NewCapacity);
+      V.back().append(Ptr, Size);
+    }
+  }
+
+public:
+  stringvec_ostream() : V(1) { SetUnbuffered(); }
+
+  friend raw_ostream &operator<<(raw_ostream &OS,
+                                 const stringvec_ostream &RHS) {
+    for (const std::string &S : RHS.V)
+      OS << S;
+    return OS;
+  }
+
+  bool operator==(StringRef RHS) {
+    if (Pos != RHS.size())
+      return false;
+
+    size_t Offset = 0;
+    for (const std::string &S : V) {
+      if (S != RHS.slice(Offset, Offset + S.size()))
+        return false;
+      Offset += S.size();
+    }
+    return true;
+  }
+};
+
 static cl::opt<std::string>
 OutputFilename("o", cl::desc("Output filename"), cl::value_desc("filename"),
                cl::init("-"));
@@ -130,8 +176,7 @@ int llvm::TableGenMain(const char *argv0,
 
   // Write output to memory.
   Timer.startBackendTimer("Backend overall");
-  std::string OutString;
-  raw_string_ostream Out(OutString);
+  stringvec_ostream Out;
   unsigned status = 0;
   // ApplyCallback will return true if it did not apply any callback. In that
   // case, attempt to apply the MainFn.
@@ -158,7 +203,7 @@ int llvm::TableGenMain(const char *argv0,
     // aren't any.
     if (auto ExistingOrErr =
             MemoryBuffer::getFile(OutputFilename, /*IsText=*/true))
-      if (std::move(ExistingOrErr.get())->getBuffer() == OutString)
+      if (Out == std::move(ExistingOrErr.get())->getBuffer())
         WriteFile = false;
   }
   if (WriteFile) {
@@ -167,7 +212,7 @@ int llvm::TableGenMain(const char *argv0,
     if (EC)
       return reportError(argv0, "error opening " + OutputFilename + ": " +
                                     EC.message() + "\n");
-    OutFile.os() << OutString;
+    OutFile.os() << Out;
     if (ErrorsPrinted == 0)
       OutFile.keep();
   }

``````````

</details>


https://github.com/llvm/llvm-project/pull/123353