[flang] [llvm] [mlir] [llvm][mlir][OpenMP] Support translation for linear clause in omp.wsloop (PR #139386)

[Tom Eccles via llvm-commits]

================
@@ -124,6 +124,146 @@ class PreviouslyReportedError
 
 char PreviouslyReportedError::ID = 0;
 
+/*
+ * Custom class for processing linear clause for omp.wsloop
+ * and omp.simd. Linear clause translation requires setup,
+ * initialization, update, and finalization at varying
+ * basic blocks in the IR. This class helps maintain
+ * internal state to allow consistent translation in
+ * each of these stages.
+ */
+
+class LinearClauseProcessor {
+
+private:
+  SmallVector<llvm::Value *> linearPreconditionVars;
+  SmallVector<llvm::Value *> linearLoopBodyTemps;
+  SmallVector<llvm::AllocaInst *> linearOrigVars;
+  SmallVector<llvm::Value *> linearOrigVal;
+  SmallVector<llvm::Value *> linearSteps;
+  llvm::BasicBlock *linearFinalizationBB;
+  llvm::BasicBlock *linearExitBB;
+  llvm::BasicBlock *linearLastIterExitBB;
+
+public:
+  // Allocate space for linear variabes
+  void createLinearVar(llvm::IRBuilderBase &builder,
+                       LLVM::ModuleTranslation &moduleTranslation,
+                       mlir::Value &linearVar) {
+    if (llvm::AllocaInst *linearVarAlloca = dyn_cast<llvm::AllocaInst>(
+            moduleTranslation.lookupValue(linearVar))) {
+      linearPreconditionVars.push_back(builder.CreateAlloca(
+          linearVarAlloca->getAllocatedType(), nullptr, ".linear_var"));
+      llvm::Value *linearLoopBodyTemp = builder.CreateAlloca(
+          linearVarAlloca->getAllocatedType(), nullptr, ".linear_result");
+      linearOrigVal.push_back(moduleTranslation.lookupValue(linearVar));
+      linearLoopBodyTemps.push_back(linearLoopBodyTemp);
+      linearOrigVars.push_back(linearVarAlloca);
+    }
----------------
tblah wrote:

Shouldn't there be some sort of error reporting to catch cases where the linear variable is not an alloca?

What about if the variable was passed as a function argument?

Nothing in the OpenMP MLIR dialect currently guarantees or documents that these variables are passed by-address into the clause, that's just how flang happens to work. I would be open to adding this requirement (as we have done this already for privatization and some reductions). But if you depend on this it does need to be documented. This is supposed to be generic code which supports any valid OpenMP MLIR code.

https://github.com/llvm/llvm-project/pull/139386