[llvm-branch-commits] [llvm] [AsmPrinter][ELF] Support profile-guided section prefix for jump tables' (read-only) data sections (PR #122215)

[Mingming Liu via llvm-branch-commits]

================
@@ -2876,42 +2875,101 @@ void AsmPrinter::emitJumpTableInfo() {
       MJTI->getEntryKind() == MachineJumpTableInfo::EK_LabelDifference32 ||
           MJTI->getEntryKind() == MachineJumpTableInfo::EK_LabelDifference64,
       F);
+
+  std::vector<unsigned> JumpTableIndices;
+  if (!TM.Options.EnableStaticDataPartitioning) {
+    for (unsigned JTI = 0, JTSize = JT.size(); JTI < JTSize; ++JTI)
+      JumpTableIndices.push_back(JTI);
+    emitJumpTables(JumpTableIndices, TLOF.getSectionForJumpTable(F, TM),
+                   JTInDiffSection, *MJTI);
+    return;
+  }
+
+  // When static data partitioning is enabled, collect jump table entries that
+  // go into the same section together to reduce the amount of section switch
+  // statements.
+  //
+  // Iterate all jump tables, put hot jump table indices towards the beginning
+  // of the vector, and cold jump table indices towards the end. Meanwhile
+  // retain the relative orders of original jump tables within a hot or unlikely
+  // section by reversing the cold jump table indices.
+  int NextHotJumpTableIndex = 0, NextColdJumpTableIndex = JT.size() - 1;
+  JumpTableIndices.resize(JT.size());
+  for (unsigned JTI = 0, JTSize = JT.size(); JTI < JTSize; ++JTI) {
+    if (JT[JTI].Hotness == MachineFunctionDataHotness::Cold)
+      JumpTableIndices[NextColdJumpTableIndex--] = JTI;
+    else
+      JumpTableIndices[NextHotJumpTableIndex++] = JTI;
+  }
+
+  if (NextHotJumpTableIndex != 0) {
+    emitJumpTables(
+        ArrayRef<unsigned>(JumpTableIndices).take_front(NextHotJumpTableIndex),
+        TLOF.getSectionForJumpTable(F, TM, &JT[0]), JTInDiffSection, *MJTI);
+  }
----------------
mingmingl-llvm wrote:

> ArrayRef::take_front(0) will return an empty array, and emitJumpTables() handles this case.

The updated change removed `if (NumHotJT>0)` and `if(NumColdJT)>0)` as suggested.

In the original patch, they also gated the call to `TLOF.getSectionForJumpTable(F, TM, &MJTE)`, so that a `MCSection` won't be created if the number of JTs is zero. 

In the updated patch, `emitJumpTableImpl` reads `TM.Options.EnableStaticDataPartitioning` directly to call one of two `getSectionForJumpTable` functions.

https://github.com/llvm/llvm-project/pull/122215