Bump stablehlo to f7f8e4e35 and drop LLVM local reverts (#18668)

Continuing from @hanhanW 's #18659:

Stablehlo cherry-picks:
1. openxla/stablehlo#2572
2. openxla/stablehlo#2573

Torch-mlir cherry-picks:
1. llvm/torch-mlir#3755

---------

Signed-off-by: hanhanW <hanhan0912@gmail.com>
Signed-off-by: Benoit Jacob <jacob.benoit.1@gmail.com>
Co-authored-by: hanhanW <hanhan0912@gmail.com>

compiler/plugins/input/StableHLO/Conversion/ConvertCollectives.cpp

@@ -491,15 +491,15 @@ struct AllGatherOpConversion final
         op.getReplicaGroups(), op.getUseGlobalDeviceIds(), rewriter);
 
     // Get the collective element type attribute.
-    auto resultType = cast<RankedTensorType>(op.getResult().getType());
+    auto resultType = cast<RankedTensorType>(op.getResult(0).getType());
     IREE::Flow::CollectiveElementTypeAttr elementTypeAttr =
         IREE::Flow::getCollectiveElementTypeAttr(resultType);
     if (!elementTypeAttr) {
       return rewriter.notifyMatchFailure(
           op, "unsupported element type for collective op");
     }
     uint64_t allGatherDim = op.getAllGatherDim();
-    Value gatherInput = adaptor.getOperand();
+    Value gatherInput = adaptor.getOperands()[0];
     SmallVector<int64_t> gatherResultShape(resultType.getShape());
 
     // When all_gather_dim != 0, we need to transpose between 0 and
@@ -513,7 +513,7 @@ struct AllGatherOpConversion final
     // Create an empty tensor for the result.
     Value target = rewriter.create<tensor::EmptyOp>(
         loc, gatherResultShape,
-        getElementTypeOrSelf(adaptor.getOperand().getType()));
+        getElementTypeOrSelf(adaptor.getOperands()[0].getType()));
     Value gatherResult = rewriter.create<IREE::Flow::CollectiveAllGatherOp>(
         op.getLoc(), elementTypeAttr, target, gatherInput, channel);
 
@@ -585,7 +585,7 @@ struct AllReduceOpConversion final
     auto reductionOpAttr =
         IREE::Flow::CollectiveReductionOpAttr::get(op.getContext(), *redOp);
 
-    auto inputType = cast<RankedTensorType>(op.getOperand().getType());
+    auto inputType = cast<RankedTensorType>(op.getOperand(0).getType());
 
     // Get the collective element type attribute.
     IREE::Flow::CollectiveElementTypeAttr elementTypeAttr =
@@ -597,10 +597,11 @@ struct AllReduceOpConversion final
     // Create an empty tensor for the result.
     ArrayRef<int64_t> inputShape = inputType.getShape();
     Value target = rewriter.create<tensor::EmptyOp>(
-        loc, inputShape, getElementTypeOrSelf(adaptor.getOperand().getType()));
+        loc, inputShape,
+        getElementTypeOrSelf(adaptor.getOperands()[0].getType()));
     auto allReduceOp = rewriter.create<IREE::Flow::CollectiveAllReduceOp>(
         op.getLoc(), reductionOpAttr, elementTypeAttr, target,
-        adaptor.getOperand(), channel);
+        adaptor.getOperands()[0], channel);
     rewriter.replaceOp(op, allReduceOp.getResult());
     return success();
   }
@@ -676,7 +677,7 @@ struct AllToAllOpConversion final
         op.getReplicaGroups(), /*useGlobalDeviceIds=*/std::nullopt, rewriter);
 
     // Get the collective element type attribute.
-    auto resultType = cast<RankedTensorType>(op.getType());
+    auto resultType = cast<RankedTensorType>(op.getType(0));
     IREE::Flow::CollectiveElementTypeAttr elementTypeAttr =
         IREE::Flow::getCollectiveElementTypeAttr(resultType);
     if (!elementTypeAttr) {
@@ -687,7 +688,7 @@ struct AllToAllOpConversion final
     uint64_t splitDim = op.getSplitDimension();
     uint64_t concatDim = op.getConcatDimension();
     uint64_t splitCount = op.getSplitCount();
-    Value allToAllInput = adaptor.getOperand();
+    Value allToAllInput = adaptor.getOperands()[0];
 
     // When splitDim != 0, we need to transpose splitDim to 0 before and after
     // the all-to-all.

compiler/plugins/input/StableHLO/Conversion/Preprocessing/DotGeneralToDot.cpp

@@ -198,7 +198,7 @@ struct GeneralDotRemoveBatch final
 
     auto dot = rewriter.create<mlir::stablehlo::DotGeneralOp>(
         op.getLoc(), ty.clone(ty.getShape().drop_front()), lhs, rhs,
-        newDimNumbers, op.getPrecisionConfigAttr());
+        newDimNumbers, op.getPrecisionConfigAttr(), op.getAlgorithmAttr());
     rewriter.replaceOpWithNewOp<mlir::stablehlo::ReshapeOp>(op, ty,
                                                             dot.getResult());
     return success();

compiler/plugins/input/StableHLO/Conversion/Preprocessing/EinsumToDotGeneral.cpp

@@ -141,7 +141,7 @@ struct EinsumToDotGeneralPattern final
     auto dotGeneralOp = rewriter.create<mlir::stablehlo::DotGeneralOp>(
         einsum.getLoc(), dotGeneralResultType, einsum.getLhs(), einsum.getRhs(),
         dimNumbers,
-        /*precision_config=*/ArrayAttr{});
+        /*precision_config=*/ArrayAttr{}, mlir::stablehlo::DotAlgorithmAttr{});
 
     if (isNaturalOrder) {
       // The dot_general is already in an appropriate result order.

compiler/plugins/input/StableHLO/Conversion/Preprocessing/StableHLOToStableHLO.cpp

@@ -424,7 +424,7 @@ struct TransposeReshapeGenericDotGeneral final
 
     auto newOp = rewriter.create<mlir::stablehlo::DotGeneralOp>(
         op.getLoc(), newResultType, lhs, rhs, dimensionNumbers,
-        op.getPrecisionConfigAttr());
+        op.getPrecisionConfigAttr(), op.getAlgorithmAttr());
 
     // Copy over unknown attributes as we currently rely on it to let user tune
     // lowering parameters.

compiler/plugins/input/StableHLO/Conversion/StableHLOCustomCalls.cpp

@@ -200,7 +200,8 @@ struct HouseholderReflectorRewriter final
           auto dotNums = mlir::stablehlo::DotDimensionNumbersAttr::get(
               b.getContext(), batch, batch, lhsContract, rhsContract);
           Value dot = b.create<mlir::stablehlo::DotGeneralOp>(
-              householder0.getType(), args[0], householder, dotNums, nullptr);
+              householder0.getType(), args[0], householder, dotNums, nullptr,
+              mlir::stablehlo::DotAlgorithmAttr{});
           b.create<scf::YieldOp>(loc, dot);
         });
 

compiler/plugins/input/StableHLO/Conversion/test/legalize_control_flow.mlir

@@ -150,7 +150,7 @@ func.func @conditional_nested(%arg0: tensor<f32>, %arg1: tensor<f32>) -> tensor<
 func.func @case2(%arg0 : tensor<i32>, %arg1 : tensor<4xf32>, %arg2 : tensor<4xf32>) -> tensor<4xf32> {
 
   // CHECK-NEXT: %[[VAL_3:.*]] = stablehlo.constant dense<0> : tensor<i32>
-  // CHECK: %[[VAL_4:.*]] = stablehlo.compare EQ, %[[VAL_0]], %[[VAL_3]], NOTYPE : (tensor<i32>, tensor<i32>) -> tensor<i1>
+  // CHECK: %[[VAL_4:.*]] = stablehlo.compare EQ, %[[VAL_0]], %[[VAL_3]] : (tensor<i32>, tensor<i32>) -> tensor<i1>
   // CHECK: %[[VAL_5:.*]] = tensor.extract %[[VAL_4]][] : tensor<i1>
   // CHECK: %[[VAL_6:.*]] = scf.if %[[VAL_5]] -> (tensor<4xf32>) {
   %1 = "stablehlo.case"(%arg0) ({
@@ -180,7 +180,7 @@ func.func @case2(%arg0 : tensor<i32>, %arg1 : tensor<4xf32>, %arg2 : tensor<4xf3
 func.func @case3(%arg0 : tensor<i32>, %arg1 : tensor<4xf32>, %arg2 : tensor<4xf32>, %arg3 : tensor<4xf32>) -> tensor<4xf32> {
 
   // CHECK-NEXT: %[[VAL_4:.*]] = stablehlo.constant dense<0> : tensor<i32>
-  // CHECK: %[[VAL_5:.*]] = stablehlo.compare EQ, %[[VAL_0]], %[[VAL_4]], NOTYPE : (tensor<i32>, tensor<i32>) -> tensor<i1>
+  // CHECK: %[[VAL_5:.*]] = stablehlo.compare EQ, %[[VAL_0]], %[[VAL_4]] : (tensor<i32>, tensor<i32>) -> tensor<i1>
   // CHECK: %[[VAL_6:.*]] = tensor.extract %[[VAL_5]][] : tensor<i1>
   // CHECK: %[[VAL_7:.*]] = scf.if %[[VAL_6]] -> (tensor<4xf32>) {
   %1 = "stablehlo.case"(%arg0) ({
@@ -191,7 +191,7 @@ func.func @case3(%arg0 : tensor<i32>, %arg1 : tensor<4xf32>, %arg2 : tensor<4xf3
 
   // CHECK: } else {
   // CHECK-NEXT:   %[[VAL_9:.*]] = stablehlo.constant dense<1> : tensor<i32>
-  // CHECK:   %[[VAL_10:.*]] = stablehlo.compare EQ, %[[VAL_0]], %[[VAL_9]], NOTYPE : (tensor<i32>, tensor<i32>) -> tensor<i1>
+  // CHECK:   %[[VAL_10:.*]] = stablehlo.compare EQ, %[[VAL_0]], %[[VAL_9]] : (tensor<i32>, tensor<i32>) -> tensor<i1>
   // CHECK:   %[[VAL_11:.*]] = tensor.extract %[[VAL_10]][] : tensor<i1>
   // CHECK:   %[[VAL_12:.*]] = scf.if %[[VAL_11]] -> (tensor<4xf32>) {
   }, {

compiler/src/iree/compiler/Tools/BUILD.bazel

@@ -113,6 +113,7 @@ iree_compiler_cc_library(
         "@llvm-project//mlir:MLProgramDialect",
         "@llvm-project//mlir:MathDialect",
         "@llvm-project//mlir:MemRefDialect",
+        "@llvm-project//mlir:QuantOps",
         "@llvm-project//mlir:SCFDialect",
         "@llvm-project//mlir:SCFToGPU",
         "@llvm-project//mlir:SCFTransforms",

compiler/src/iree/compiler/Tools/init_mlir_dialects.h

@@ -40,7 +40,7 @@
 #include "mlir/Dialect/NVGPU/IR/NVGPUDialect.h"
 #include "mlir/Dialect/PDL/IR/PDL.h"
 #include "mlir/Dialect/PDLInterp/IR/PDLInterp.h"
-#include "mlir/Dialect/Quant/QuantOps.h"
+#include "mlir/Dialect/Quant/IR/Quant.h"
 #include "mlir/Dialect/SCF/IR/SCF.h"
 #include "mlir/Dialect/SCF/TransformOps/SCFTransformOps.h"
 #include "mlir/Dialect/SPIRV/IR/SPIRVDialect.h"
@@ -81,7 +81,7 @@ inline void registerMlirDialects(DialectRegistry &registry) {
                   pdl::PDLDialect,
                   pdl_interp::PDLInterpDialect,
                   scf::SCFDialect,
-                  quant::QuantizationDialect,
+                  quant::QuantDialect,
                   spirv::SPIRVDialect,
                   arm_neon::ArmNeonDialect,
                   arm_sve::ArmSVEDialect,