[Mlir-commits] [mlir] [OpenMP][MLIR] Hoist static `alloca`s emitted by private `init` regions to the allocation IP of the construct (PR #171597)

[Tom Eccles]

================
@@ -1596,10 +1596,64 @@ static llvm::Expected<llvm::Value *> initPrivateVar(
   return phis[0];
 }
 
+/// Beginning with \p startBlock, this function visits all reachable successor
+/// blocks. For each such block, static alloca instructions (i.e. non-array
+/// allocas) are collected. Then, these collected alloca instructions are moved
+/// to the \p allocaIP insertion point.
+///
+/// This is useful in cases where, for example, more than one allocatable or
+/// array are privatized. In such cases, we allocate a number of temporary
+/// descriptors to handle the initialization logic. Additonally, for each
+/// private value, there is branching logic based on the value of the origianl
+/// private variable's allocation state. Therefore, we end up with descriptor
+/// alloca instructions preceded by conditional branches which casues runtime
+/// issues at least on the GPU.
+static void hoistStaticAllocasToAllocaIP(
+    llvm::BasicBlock *startBlock,
+    const llvm::OpenMPIRBuilder::InsertPointTy &allocaIP) {
+  llvm::SmallVector<llvm::BasicBlock *> inlinedBlocks{startBlock};
+  llvm::SmallPtrSet<llvm::BasicBlock *, 4> seenBlocks;
+  llvm::SmallVector<llvm::Instruction *> staticAllocas;
+
+  while (!inlinedBlocks.empty()) {
+    llvm::BasicBlock *curBlock = inlinedBlocks.front();
+    inlinedBlocks.erase(inlinedBlocks.begin());
+    llvm::Instruction *terminator = curBlock->getTerminator();
+
+    for (llvm::Instruction &inst : *curBlock) {
+      if (auto *allocaInst = mlir::dyn_cast<llvm::AllocaInst>(&inst)) {
+        if (!allocaInst->isArrayAllocation()) {
----------------
tblah wrote:

If I understand correctly, the issue comes from mixing allocas with the if/else branching. If the init region only has a single block then it can't be using branching.

The solution you propose using dominance info sounds like it would ultimately give us better code (even more allocas will end up in the entry block), although it might require updates to more tests.

https://github.com/llvm/llvm-project/pull/171597