Add support for casting elements in vectors for certain Std dialect type conversion operations.

[llitchev]

Summary

Added support to the Std dialect cast operations to do casts in vector types when feasible.

[bondhugula]

mlir/lib/Dialect/StandardOps/IR/Ops.cpp

224

Doc comment please (triple /). Please describe what the method referring to its args.

2335–2340

All of these changes would be inconsistent with the current documentation / spec of these ops. See Ops.td file - those have to be updated first if the semantics indeed changed. For eg. UIToFP's doc comment says:

let description = [{
    Cast from a value interpreted as unsigned integer to the corresponding
    floating-point value. If the value cannot be exactly represented, it is
    rounded using the default rounding mode. Only scalars are currently
    supported.
  }];

We can't change the behavior while having the spec/doc to the contrary.

2417

Comment would be outdated.

[llitchev]

Resolved CR feedback - updated docs.

mlir/lib/Dialect/StandardOps/IR/Ops.cpp

224

Fixed.

2335–2340

Fixed.

2417

Fixed.

[bondhugula]

I find the commit summary and the title a bit confusing. Isn't this extending standard dialect cast ops to work with vectors of those types? It's not that vector types are being cast to int/fp.

There are operations that support casting for vector types, but the code was missing. The necessary code and tests were added.

Please rephrase. Those operations existed in the LLVM dialect not in std dialect.

The test cases for a key change to a std dialect op's semantics should reside in the std dialect's test cases, and not just in the LLVM dialect ones.

mlir/include/mlir/Dialect/StandardOps/IR/Ops.td

2433–2437 ↗

(On Diff #291061)

Nit: Reflow. Please use the whole width.

3117–3119 ↗

(On Diff #291061)

Likewise.

[bondhugula]

mlir/lib/Dialect/StandardOps/IR/Ops.cpp

221

Nit: period at the end.

225

Sentence broken.

226–228

Please rephrase - this isn't really well-formed.

[llitchev]

Rewritten the title and the summary as well.
The tests are in the StandardToLLVM directory (as the rest of the similar conversion tests). This is not part of the LLVM? Or maybe I am wrong?

mlir/include/mlir/Dialect/StandardOps/IR/Ops.td

2433–2437 ↗

(On Diff #291061)

Fixed. Using no more than 80 characters.

3117–3119 ↗

(On Diff #291061)

Fixed. Using no more than 80 characters.

mlir/lib/Dialect/StandardOps/IR/Ops.cpp

221

Fixed.

225

Rewritten.

226–228

Rewritten.

[llitchev]

Addressed some CR feedback.

mlir/lib/Dialect/StandardOps/IR/Ops.cpp

for (OpOperand &operand : op->getOpOperands()) {
     if (cast && !cast.getOperand().getType().isa<UnrankedMemRefType>()) {
       operand.set(cast.getOperand());
       folded = true;
     }
   }
   return success(folded);
 }
 
+// //===----------------------------------------------------------------------===//
+// // Common cast compatibility check for vector types
+// //===----------------------------------------------------------------------===//
+
+// Type compatibility for vector casts.
+static bool areVectorCastSimpleCompatible(
+    Type a, Type b, function_ref<bool(Type, Type)> areElementsCastCompatible) {
+  if (auto va = a.dyn_cast<VectorType>())
+    if (auto vb = b.dyn_cast<VectorType>())
+      return va.getShape().equals(vb.getShape()) &&
+             areElementsCastCompatible(va.getElementType(),
+                                       vb.getElementType());
+  return false;
+}
+
 //===----------------------------------------------------------------------===//
 // AddFOp
 //===----------------------------------------------------------------------===//
 
 OpFoldResult AddFOp::fold(ArrayRef<Attribute> operands) {
   return constFoldBinaryOp<FloatAttr>(
       operands, [](APFloat a, APFloat b) { return a + b; });
 }

 //===----------------------------------------------------------------------===//
 // FPExtOp
 //===----------------------------------------------------------------------===//
 
 bool FPExtOp::areCastCompatible(Type a, Type b) {
   if (auto fa = a.dyn_cast<FloatType>())
     if (auto fb = b.dyn_cast<FloatType>())
       return fa.getWidth() < fb.getWidth();
-  if (auto va = a.dyn_cast<VectorType>())
-    if (auto vb = b.dyn_cast<VectorType>())
-      return va.getShape().equals(vb.getShape()) &&
-             areCastCompatible(va.getElementType(), vb.getElementType());
-  return false;
+  return areVectorCastSimpleCompatible(a, b, areCastCompatible);
 }
 
 //===----------------------------------------------------------------------===//
 // FPToSIOp
 //===----------------------------------------------------------------------===//
 
 bool FPToSIOp::areCastCompatible(Type a, Type b) {
-  return a.isa<FloatType>() && b.isSignlessInteger();
+  if (a.isa<FloatType>() && b.isSignlessInteger())
+    return true;
+  return areVectorCastSimpleCompatible(a, b, areCastCompatible);
 }
 
 //===----------------------------------------------------------------------===//
 // FPToUIOp
 //===----------------------------------------------------------------------===//
 
 bool FPToUIOp::areCastCompatible(Type a, Type b) {
-  return a.isa<FloatType>() && b.isSignlessInteger();
+  if (a.isa<FloatType>() && b.isSignlessInteger())
+    return true;
+  return areVectorCastSimpleCompatible(a, b, areCastCompatible);
 }
 
 //===----------------------------------------------------------------------===//
 // FPTruncOp
 //===----------------------------------------------------------------------===//
 
 bool FPTruncOp::areCastCompatible(Type a, Type b) {
   if (auto fa = a.dyn_cast<FloatType>())
     if (auto fb = b.dyn_cast<FloatType>())
       return fa.getWidth() > fb.getWidth();
-  if (auto va = a.dyn_cast<VectorType>())
-    if (auto vb = b.dyn_cast<VectorType>())
-      return va.getShape().equals(vb.getShape()) &&
-             areCastCompatible(va.getElementType(), vb.getElementType());
-  return false;
+  return areVectorCastSimpleCompatible(a, b, areCastCompatible);
 }
 
 //===----------------------------------------------------------------------===//
 // IndexCastOp
 //===----------------------------------------------------------------------===//
 
 // Index cast is applicable from index to integer and backwards.
 bool IndexCastOp::areCastCompatible(Type a, Type b) {
if (!lhs)
     return {};
   return IntegerAttr::get(lhs.getType(), lhs.getValue().srem(rhsValue));
 }
 
 //===----------------------------------------------------------------------===//
 // SIToFPOp
 //===----------------------------------------------------------------------===//
 
 // sitofp is applicable from integer types to float types.
 bool SIToFPOp::areCastCompatible(Type a, Type b) {
-  return a.isSignlessInteger() && b.isa<FloatType>();
+  if (a.isSignlessInteger() && b.isa<FloatType>())
+    return true;
+  return areVectorCastSimpleCompatible(a, b, areCastCompatible);
 }
 
 //===----------------------------------------------------------------------===//
 // SplatOp
 //===----------------------------------------------------------------------===//
 
 static LogicalResult verify(SplatOp op) {
   // TODO: we could replace this by a trait.
   if (op.getOperand().getType() !=
OpFoldResult SubIOp::fold(ArrayRef<Attribute> operands) {
   return constFoldBinaryOp<IntegerAttr>(operands,
                                         [](APInt a, APInt b) { return a - b; });
 }
 
 //===----------------------------------------------------------------------===//
 // UIToFPOp
 //===----------------------------------------------------------------------===//
 
 // uitofp is applicable from integer types to float types.
 bool UIToFPOp::areCastCompatible(Type a, Type b) {
-  return a.isSignlessInteger() && b.isa<FloatType>();
+  if (a.isSignlessInteger() && b.isa<FloatType>())
+    return true;
+  return areVectorCastSimpleCompatible(a, b, areCastCompatible);
 }
 
 //===----------------------------------------------------------------------===//
 // SubViewOp
 //===----------------------------------------------------------------------===//
 
 /// Print a list with either (1) the static integer value in `arrayAttr` if
 /// `isDynamic` evaluates to false or (2) the next value otherwise.

mlir/test/Conversion/StandardToLLVM/convert-to-llvmir.mlir

// CHECK-NEXT: = llvm.sitofp {{.*}} : !llvm.i32 to !llvm.double
   %1 = sitofp %arg0: i32 to f64
 // CHECK-NEXT: = llvm.sitofp {{.*}} : !llvm.i64 to !llvm.float
   %2 = sitofp %arg1: i64 to f32
 // CHECK-NEXT: = llvm.sitofp {{.*}} : !llvm.i64 to !llvm.double
   %3 = sitofp %arg1: i64 to f64
   return
 }
 
+// Checking conversion of integer vectors to floating point vector types.
+// CHECK-LABEL: @sitofp_vector
+func @sitofp_vector(%arg0 : vector<2xi16>, %arg1 : vector<2xi32>, %arg2 : vector<2xi64>) {
+// CHECK-NEXT: = llvm.sitofp {{.*}} : !llvm.vec<2 x i16> to !llvm.vec<2 x float>
+  %0 = sitofp %arg0: vector<2xi16> to vector<2xf32>
+// CHECK-NEXT: = llvm.sitofp {{.*}} : !llvm.vec<2 x i16> to !llvm.vec<2 x double>
+  %1 = sitofp %arg0: vector<2xi16> to vector<2xf64>
+// CHECK-NEXT: = llvm.sitofp {{.*}} : !llvm.vec<2 x i32> to !llvm.vec<2 x float>
+  %2 = sitofp %arg1: vector<2xi32> to vector<2xf32>
+// CHECK-NEXT: = llvm.sitofp {{.*}} : !llvm.vec<2 x i32> to !llvm.vec<2 x double>
+  %3 = sitofp %arg1: vector<2xi32> to vector<2xf64>
+// CHECK-NEXT: = llvm.sitofp {{.*}} : !llvm.vec<2 x i64> to !llvm.vec<2 x float>
+  %4 = sitofp %arg2: vector<2xi64> to vector<2xf32>
+// CHECK-NEXT: = llvm.sitofp {{.*}} : !llvm.vec<2 x i64> to !llvm.vec<2 x double>
+  %5 = sitofp %arg2: vector<2xi64> to vector<2xf64>
+  return
+}
+
 // Checking conversion of unsigned integer types to floating point.
 // CHECK-LABEL: @uitofp
 func @uitofp(%arg0 : i32, %arg1 : i64) {
 // CHECK-NEXT: = llvm.uitofp {{.*}} : !llvm.i32 to !llvm.float
   %0 = uitofp %arg0: i32 to f32
 // CHECK-NEXT: = llvm.uitofp {{.*}} : !llvm.i32 to !llvm.double
   %1 = uitofp %arg0: i32 to f64
 // CHECK-NEXT: = llvm.uitofp {{.*}} : !llvm.i64 to !llvm.float
// CHECK-NEXT: = llvm.fptosi {{.*}} : !llvm.float to !llvm.i64
   %1 = fptosi %arg0: f32 to i64
 // CHECK-NEXT: = llvm.fptosi {{.*}} : !llvm.double to !llvm.i32
   %2 = fptosi %arg1: f64 to i32
 // CHECK-NEXT: = llvm.fptosi {{.*}} : !llvm.double to !llvm.i64
   %3 = fptosi %arg1: f64 to i64
   return
 }
 
+// Checking conversion of floating point vectors to integer vector types.
+// CHECK-LABEL: @fptosi_vector
+func @fptosi_vector(%arg0 : vector<2xf16>, %arg1 : vector<2xf32>, %arg2 : vector<2xf64>) {
+// CHECK-NEXT: = llvm.fptosi {{.*}} : !llvm.vec<2 x half> to !llvm.vec<2 x i32>
+  %0 = fptosi %arg0: vector<2xf16> to vector<2xi32>
+// CHECK-NEXT: = llvm.fptosi {{.*}} : !llvm.vec<2 x half> to !llvm.vec<2 x i64>
+  %1 = fptosi %arg0: vector<2xf16> to vector<2xi64>
+// CHECK-NEXT: = llvm.fptosi {{.*}} : !llvm.vec<2 x float> to !llvm.vec<2 x i32>
+  %2 = fptosi %arg1: vector<2xf32> to vector<2xi32>
+// CHECK-NEXT: = llvm.fptosi {{.*}} : !llvm.vec<2 x float> to !llvm.vec<2 x i64>
+  %3 = fptosi %arg1: vector<2xf32> to vector<2xi64>
+// CHECK-NEXT: = llvm.fptosi {{.*}} : !llvm.vec<2 x double> to !llvm.vec<2 x i32>
+  %4 = fptosi %arg2: vector<2xf64> to vector<2xi32>
+// CHECK-NEXT: = llvm.fptosi {{.*}} : !llvm.vec<2 x double> to !llvm.vec<2 x i64>
+  %5 = fptosi %arg2: vector<2xf64> to vector<2xi64>
+  return
+}
+
 // Checking conversion of floating point to integer types.
 // CHECK-LABEL: @fptoui
 func @fptoui(%arg0 : f32, %arg1 : f64) {
 // CHECK-NEXT: = llvm.fptoui {{.*}} : !llvm.float to !llvm.i32
   %0 = fptoui %arg0: f32 to i32
 // CHECK-NEXT: = llvm.fptoui {{.*}} : !llvm.float to !llvm.i64
   %1 = fptoui %arg0: f32 to i64
 // CHECK-NEXT: = llvm.fptoui {{.*}} : !llvm.double to !llvm.i32
   %2 = fptoui %arg1: f64 to i32
 // CHECK-NEXT: = llvm.fptoui {{.*}} : !llvm.double to !llvm.i64
   %3 = fptoui %arg1: f64 to i64
   return
 }
 
+// Checking conversion of floating point vectors to integer vector types.
+// CHECK-LABEL: @fptoui_vector
+func @fptoui_vector(%arg0 : vector<2xf16>, %arg1 : vector<2xf32>, %arg2 : vector<2xf64>) {
+// CHECK-NEXT: = llvm.fptoui {{.*}} : !llvm.vec<2 x half> to !llvm.vec<2 x i32>
+  %0 = fptoui %arg0: vector<2xf16> to vector<2xi32>
+// CHECK-NEXT: = llvm.fptoui {{.*}} : !llvm.vec<2 x half> to !llvm.vec<2 x i64>
+  %1 = fptoui %arg0: vector<2xf16> to vector<2xi64>
+// CHECK-NEXT: = llvm.fptoui {{.*}} : !llvm.vec<2 x float> to !llvm.vec<2 x i32>
+  %2 = fptoui %arg1: vector<2xf32> to vector<2xi32>
+// CHECK-NEXT: = llvm.fptoui {{.*}} : !llvm.vec<2 x float> to !llvm.vec<2 x i64>
+  %3 = fptoui %arg1: vector<2xf32> to vector<2xi64>
+// CHECK-NEXT: = llvm.fptoui {{.*}} : !llvm.vec<2 x double> to !llvm.vec<2 x i32>
+  %4 = fptoui %arg2: vector<2xf64> to vector<2xi32>
+// CHECK-NEXT: = llvm.fptoui {{.*}} : !llvm.vec<2 x double> to !llvm.vec<2 x i64>
+  %5 = fptoui %arg2: vector<2xf64> to vector<2xi64>
+  return
+}
+
+// Checking conversion of integer vectors to floating point vector types.
+// CHECK-LABEL: @uitofp_vector
+func @uitofp_vector(%arg0 : vector<2xi16>, %arg1 : vector<2xi32>, %arg2 : vector<2xi64>) {
+// CHECK-NEXT: = llvm.uitofp {{.*}} : !llvm.vec<2 x i16> to !llvm.vec<2 x float>
+  %0 = uitofp %arg0: vector<2xi16> to vector<2xf32>
+// CHECK-NEXT: = llvm.uitofp {{.*}} : !llvm.vec<2 x i16> to !llvm.vec<2 x double>
+  %1 = uitofp %arg0: vector<2xi16> to vector<2xf64>
+// CHECK-NEXT: = llvm.uitofp {{.*}} : !llvm.vec<2 x i32> to !llvm.vec<2 x float>
+  %2 = uitofp %arg1: vector<2xi32> to vector<2xf32>
+// CHECK-NEXT: = llvm.uitofp {{.*}} : !llvm.vec<2 x i32> to !llvm.vec<2 x double>
+  %3 = uitofp %arg1: vector<2xi32> to vector<2xf64>
+// CHECK-NEXT: = llvm.uitofp {{.*}} : !llvm.vec<2 x i64> to !llvm.vec<2 x float>
+  %4 = uitofp %arg2: vector<2xi64> to vector<2xf32>
+// CHECK-NEXT: = llvm.uitofp {{.*}} : !llvm.vec<2 x i64> to !llvm.vec<2 x double>
+  %5 = uitofp %arg2: vector<2xi64> to vector<2xf64>
+  return
+}
 
 // Checking conversion of integer types to floating point.
 // CHECK-LABEL: @fptrunc
 func @fptrunc(%arg0 : f32, %arg1 : f64) {
 // CHECK-NEXT: = llvm.fptrunc {{.*}} : !llvm.float to !llvm.half
   %0 = fptrunc %arg0: f32 to f16
 // CHECK-NEXT: = llvm.fptrunc {{.*}} : !llvm.double to !llvm.half
   %1 = fptrunc %arg1: f64 to f16