From 247e185dd51261c8b8b7f572a3fa64c001e9d2d4 Mon Sep 17 00:00:00 2001
From: Nicolas Vasilache <ntv@google.com>
Date: Fri, 5 Jun 2020 08:22:15 -0400
Subject: [PATCH] [mlir][Vector] Move temporary alloc to top of the function
 alloca when lowering vector_transfers

Recently introduced allocation hoisting is quite conservative on the cases when it triggers.
This revision makes it such that the allocations for vector transfer lowerings are hoisted
to the top of the function.
This should be revisited in the context of parallelism and is a temporary workaround.

Differential Revision: https://reviews.llvm.org/D81253
---
 mlir/lib/Conversion/VectorToSCF/VectorToSCF.cpp       | 15 +++++++++++++--
 mlir/test/Conversion/VectorToSCF/vector-to-loops.mlir |  6 +++---
 2 files changed, 16 insertions(+), 5 deletions(-)

diff --git a/mlir/lib/Conversion/VectorToSCF/VectorToSCF.cpp b/mlir/lib/Conversion/VectorToSCF/VectorToSCF.cpp
index 6816bc7..329b008 100644
--- a/mlir/lib/Conversion/VectorToSCF/VectorToSCF.cpp
+++ b/mlir/lib/Conversion/VectorToSCF/VectorToSCF.cpp
@@ -196,13 +196,24 @@ Value NDTransferOpHelper<ConcreteOp>::emitInBoundsCondition(
   return inBoundsCondition;
 }
 
+// TODO: Parallelism and threadlocal considerations.
+static Value setAllocAtFunctionEntry(MemRefType memRefMinorVectorType,
+                                     Operation *op) {
+  auto &b = ScopedContext::getBuilderRef();
+  OpBuilder::InsertionGuard guard(b);
+  b.setInsertionPointToStart(&op->getParentOfType<FuncOp>().front());
+  Value res =
+      std_alloca(memRefMinorVectorType, ValueRange{}, b.getI64IntegerAttr(128));
+  return res;
+}
+
 template <>
 LogicalResult NDTransferOpHelper<TransferReadOp>::doReplace() {
   Value alloc, result;
   if (options.unroll)
     result = std_splat(vectorType, xferOp.padding());
   else
-    alloc = std_alloc(memRefMinorVectorType);
+    alloc = setAllocAtFunctionEntry(memRefMinorVectorType, op);
 
   emitLoops([&](ValueRange majorIvs, ValueRange leadingOffsets,
                 ValueRange majorOffsets, ValueRange minorOffsets,
@@ -297,7 +308,7 @@ template <>
 LogicalResult NDTransferOpHelper<TransferWriteOp>::doReplace() {
   Value alloc;
   if (!options.unroll) {
-    alloc = std_alloc(memRefMinorVectorType);
+    alloc = setAllocAtFunctionEntry(memRefMinorVectorType, op);
     std_store(xferOp.vector(),
               vector_type_cast(MemRefType::get({}, vectorType), alloc));
   }
diff --git a/mlir/test/Conversion/VectorToSCF/vector-to-loops.mlir b/mlir/test/Conversion/VectorToSCF/vector-to-loops.mlir
index d4f22d2..7391ebf 100644
--- a/mlir/test/Conversion/VectorToSCF/vector-to-loops.mlir
+++ b/mlir/test/Conversion/VectorToSCF/vector-to-loops.mlir
@@ -232,7 +232,7 @@ func @transfer_read_progressive(%A : memref<?x?xf32>, %base: index) -> vector<3x
   %f7 = constant 7.0: f32
 
   // CHECK-DAG: %[[splat:.*]] = constant dense<7.000000e+00> : vector<15xf32>
-  // CHECK-DAG: %[[alloc:.*]] = alloc() : memref<3xvector<15xf32>>
+  // CHECK-DAG: %[[alloc:.*]] = alloca() {alignment = 128 : i64} : memref<3xvector<15xf32>>
   // CHECK-DAG: %[[dim:.*]] = dim %[[A]], 0 : memref<?x?xf32>
   // CHECK: affine.for %[[I:.*]] = 0 to 3 {
   // CHECK:   %[[add:.*]] = affine.apply #[[MAP0]](%[[I]])[%[[base]]]
@@ -304,7 +304,7 @@ func @transfer_read_progressive(%A : memref<?x?xf32>, %base: index) -> vector<3x
 //  FULL-UNROLL-SAME:   %[[base:[a-zA-Z0-9]+]]: index,
 //  FULL-UNROLL-SAME:   %[[vec:[a-zA-Z0-9]+]]: vector<3x15xf32>
 func @transfer_write_progressive(%A : memref<?x?xf32>, %base: index, %vec: vector<3x15xf32>) {
-  // CHECK: %[[alloc:.*]] = alloc() : memref<3xvector<15xf32>>
+  // CHECK: %[[alloc:.*]] = alloca() {alignment = 128 : i64} : memref<3xvector<15xf32>>
   // CHECK: %[[vmemref:.*]] = vector.type_cast %[[alloc]] : memref<3xvector<15xf32>> to memref<vector<3x15xf32>>
   // CHECK: store %[[vec]], %[[vmemref]][] : memref<vector<3x15xf32>>
   // CHECK: %[[dim:.*]] = dim %[[A]], 0 : memref<?x?xf32>
@@ -359,7 +359,7 @@ func @transfer_write_progressive(%A : memref<?x?xf32>, %base: index, %vec: vecto
 //  FULL-UNROLL-SAME:   %[[vec:[a-zA-Z0-9]+]]: vector<3x15xf32>
 func @transfer_write_progressive_not_masked(%A : memref<?x?xf32>, %base: index, %vec: vector<3x15xf32>) {
   // CHECK-NOT:    scf.if
-  // CHECK-NEXT: %[[alloc:.*]] = alloc() : memref<3xvector<15xf32>>
+  // CHECK-NEXT: %[[alloc:.*]] = alloca() {alignment = 128 : i64} : memref<3xvector<15xf32>>
   // CHECK-NEXT: %[[vmemref:.*]] = vector.type_cast %[[alloc]] : memref<3xvector<15xf32>> to memref<vector<3x15xf32>>
   // CHECK-NEXT: store %[[vec]], %[[vmemref]][] : memref<vector<3x15xf32>>
   // CHECK-NEXT: affine.for %[[I:.*]] = 0 to 3 {
-- 
2.7.4