[mlir][vector] Propagate scalability to gather/scatter ptrs vector

[c-rhodes]

In convert-vector-to-llvm the first operand (vector of pointers holding
all memory addresses to read) to the masked.gather (and scatter)
intrinsic has a fixed vector type.

This may result in intrinsics where the scalable flag has been dropped:

  %0 = llvm.intr.masked.gather %1, %2, %3 {alignment = 4 : i32}
    : (!llvm.vec<4 x ptr>, vector<[4]xi1>, vector<[4]xi32>) -> vector<[4]xi32>

Fortunately the operand is overloaded on the result type so we end up
with the correct IR when lowering to LLVM, but this is still incorrect.
This patch fixes it by propagating scalability.

[llvmbot]

@llvm/pr-subscribers-mlir-llvm

@llvm/pr-subscribers-mlir

Author: Cullen Rhodes (c-rhodes)

Changes

In convert-vector-to-llvm the first operand (vector of pointers holding
all memory addresses to read) to the masked.gather (and scatter)
intrinsic has a fixed vector type.

This may result in intrinsics where the scalable flag has been dropped:

  %0 = llvm.intr.masked.gather %1, %2, %3 {alignment = 4 : i32}
    : (!llvm.vec<4 x ptr>, vector<[4]xi1>, vector<[4]xi32>) -> vector<[4]xi32>

Fortunately the operand is overloaded on the result type so we end up
with the correct IR when lowering to LLVM, but this is still incorrect.
This patch fixes it by propagating scalability.

---

Full diff: https://github.com/llvm/llvm-project/pull/97584.diff

2 Files Affected:

• (modified) mlir/lib/Conversion/VectorToLLVM/ConvertVectorToLLVM.cpp (+11-10)
• (modified) mlir/test/Conversion/VectorToLLVM/vector-to-llvm.mlir (+25)

diff --git a/mlir/lib/Conversion/VectorToLLVM/ConvertVectorToLLVM.cpp b/mlir/lib/Conversion/VectorToLLVM/ConvertVectorToLLVM.cpp
index 0eac55255b133..77bdacbc46990 100644
--- a/mlir/lib/Conversion/VectorToLLVM/ConvertVectorToLLVM.cpp
+++ b/mlir/lib/Conversion/VectorToLLVM/ConvertVectorToLLVM.cpp
@@ -102,11 +102,13 @@ static LogicalResult isMemRefTypeSupported(MemRefType memRefType,
 static Value getIndexedPtrs(ConversionPatternRewriter &rewriter, Location loc,
                             const LLVMTypeConverter &typeConverter,
                             MemRefType memRefType, Value llvmMemref, Value base,
-                            Value index, uint64_t vLen) {
+                            Value index, VectorType vectorType) {
   assert(succeeded(isMemRefTypeSupported(memRefType, typeConverter)) &&
          "unsupported memref type");
   auto pType = MemRefDescriptor(llvmMemref).getElementPtrType();
-  auto ptrsType = LLVM::getFixedVectorType(pType, vLen);
+  auto ptrsType =
+      LLVM::getVectorType(pType, vectorType.getDimSize(0),
+                          /*isScalable=*/vectorType.getScalableDims()[0]);
   return rewriter.create<LLVM::GEPOp>(
       loc, ptrsType, typeConverter.convertType(memRefType.getElementType()),
       base, index);
@@ -288,9 +290,9 @@ class VectorGatherOpConversion
     if (!isa<LLVM::LLVMArrayType>(llvmNDVectorTy)) {
       auto vType = gather.getVectorType();
       // Resolve address.
-      Value ptrs = getIndexedPtrs(rewriter, loc, *this->getTypeConverter(),
-                                  memRefType, base, ptr, adaptor.getIndexVec(),
-                                  /*vLen=*/vType.getDimSize(0));
+      Value ptrs =
+          getIndexedPtrs(rewriter, loc, *this->getTypeConverter(), memRefType,
+                         base, ptr, adaptor.getIndexVec(), vType);
       // Replace with the gather intrinsic.
       rewriter.replaceOpWithNewOp<LLVM::masked_gather>(
           gather, typeConverter->convertType(vType), ptrs, adaptor.getMask(),
@@ -305,8 +307,7 @@ class VectorGatherOpConversion
       // Resolve address.
       Value ptrs = getIndexedPtrs(
           rewriter, loc, typeConverter, memRefType, base, ptr,
-          /*index=*/vectorOperands[0],
-          LLVM::getVectorNumElements(llvm1DVectorTy).getFixedValue());
+          /*index=*/vectorOperands[0], cast<VectorType>(llvm1DVectorTy));
       // Create the gather intrinsic.
       return rewriter.create<LLVM::masked_gather>(
           loc, llvm1DVectorTy, ptrs, /*mask=*/vectorOperands[1],
@@ -343,9 +344,9 @@ class VectorScatterOpConversion
     VectorType vType = scatter.getVectorType();
     Value ptr = getStridedElementPtr(loc, memRefType, adaptor.getBase(),
                                      adaptor.getIndices(), rewriter);
-    Value ptrs = getIndexedPtrs(
-        rewriter, loc, *this->getTypeConverter(), memRefType, adaptor.getBase(),
-        ptr, adaptor.getIndexVec(), /*vLen=*/vType.getDimSize(0));
+    Value ptrs =
+        getIndexedPtrs(rewriter, loc, *this->getTypeConverter(), memRefType,
+                       adaptor.getBase(), ptr, adaptor.getIndexVec(), vType);
 
     // Replace with the scatter intrinsic.
     rewriter.replaceOpWithNewOp<LLVM::masked_scatter>(
diff --git a/mlir/test/Conversion/VectorToLLVM/vector-to-llvm.mlir b/mlir/test/Conversion/VectorToLLVM/vector-to-llvm.mlir
index 09b79708a9ab2..6f8145b618b71 100644
--- a/mlir/test/Conversion/VectorToLLVM/vector-to-llvm.mlir
+++ b/mlir/test/Conversion/VectorToLLVM/vector-to-llvm.mlir
@@ -2248,6 +2248,19 @@ func.func @gather_op(%arg0: memref<?xf32>, %arg1: vector<3xi32>, %arg2: vector<3
 
 // -----
 
+func.func @gather_op_scalable(%arg0: memref<?xf32>, %arg1: vector<[3]xi32>, %arg2: vector<[3]xi1>, %arg3: vector<[3]xf32>) -> vector<[3]xf32> {
+  %0 = arith.constant 0: index
+  %1 = vector.gather %arg0[%0][%arg1], %arg2, %arg3 : memref<?xf32>, vector<[3]xi32>, vector<[3]xi1>, vector<[3]xf32> into vector<[3]xf32>
+  return %1 : vector<[3]xf32>
+}
+
+// CHECK-LABEL: func @gather_op_scalable
+// CHECK: %[[P:.*]] = llvm.getelementptr %{{.*}}[%{{.*}}] : (!llvm.ptr, vector<[3]xi32>) -> !llvm.vec<? x 3 x ptr>, f32
+// CHECK: %[[G:.*]] = llvm.intr.masked.gather %[[P]], %{{.*}}, %{{.*}} {alignment = 4 : i32} : (!llvm.vec<? x 3 x ptr>, vector<[3]xi1>, vector<[3]xf32>) -> vector<[3]xf32>
+// CHECK: return %[[G]] : vector<[3]xf32>
+
+// -----
+
 func.func @gather_op_global_memory(%arg0: memref<?xf32, 1>, %arg1: vector<3xi32>, %arg2: vector<3xi1>, %arg3: vector<3xf32>) -> vector<3xf32> {
   %0 = arith.constant 0: index
   %1 = vector.gather %arg0[%0][%arg1], %arg2, %arg3 : memref<?xf32, 1>, vector<3xi32>, vector<3xi1>, vector<3xf32> into vector<3xf32>
@@ -2351,6 +2364,18 @@ func.func @scatter_op(%arg0: memref<?xf32>, %arg1: vector<3xi32>, %arg2: vector<
 
 // -----
 
+func.func @scatter_op_scalable(%arg0: memref<?xf32>, %arg1: vector<[3]xi32>, %arg2: vector<[3]xi1>, %arg3: vector<[3]xf32>) {
+  %0 = arith.constant 0: index
+  vector.scatter %arg0[%0][%arg1], %arg2, %arg3 : memref<?xf32>, vector<[3]xi32>, vector<[3]xi1>, vector<[3]xf32>
+  return
+}
+
+// CHECK-LABEL: func @scatter_op_scalable
+// CHECK: %[[P:.*]] = llvm.getelementptr %{{.*}}[%{{.*}}] : (!llvm.ptr, vector<[3]xi32>) -> !llvm.vec<? x 3 x ptr>, f32
+// CHECK: llvm.intr.masked.scatter %{{.*}}, %[[P]], %{{.*}} {alignment = 4 : i32} : vector<[3]xf32>, vector<[3]xi1> into !llvm.vec<? x 3 x ptr>
+
+// -----
+
 func.func @scatter_op_index(%arg0: memref<?xindex>, %arg1: vector<3xindex>, %arg2: vector<3xi1>, %arg3: vector<3xindex>) {
   %0 = arith.constant 0: index
   vector.scatter %arg0[%0][%arg1], %arg2, %arg3 : memref<?xindex>, vector<3xindex>, vector<3xi1>, vector<3xindex>

[banach-space]

Makes sense, thanks for the fix! LGTM

[dcaballe]

Thanks! Is there a way to check this in the verifier?

[c-rhodes]

Thanks! Is there a way to check this in the verifier?

I've added verifiers for gather/scatter intrinsic to check this. Generally the intrinsics are quite loose with constraints, seems like a bit of a rabbit hole.

[dcaballe]

Awesome, thanks!