bump LLVM to acc6f3e9c1af6c7445aae6f10d4b016ac84112d3 (#15296)
https://github.com/llvm/llvm-project/pull/70218 just missed the last
integrate, and cherry-picks are frowned upon.
The good thing with just missing an integrate is that just bumping the
submodule shouldn't be too hard still. I just had to fix up one small
thing in CollapseDimensions.
ci-extra:build_test_all_windows,build_test_all_macos_arm64,build_test_all_macos_x86_64
diff --git a/compiler/src/iree/compiler/Codegen/Common/MaterializeUserConfigs.cpp b/compiler/src/iree/compiler/Codegen/Common/MaterializeUserConfigs.cpp
index da0f5eb..918b43a 100644
--- a/compiler/src/iree/compiler/Codegen/Common/MaterializeUserConfigs.cpp
+++ b/compiler/src/iree/compiler/Codegen/Common/MaterializeUserConfigs.cpp
@@ -190,9 +190,15 @@
/// the same time. Because the strategy is rooted on the variant op, the
/// strategy can change the translation info on the exports if needed, else
/// back to default IREE codegen.
+ StringRef entryPoint = libraryFunc->getLeafReference();
+ Operation *transformRoot = transform::detail::findTransformEntryPoint(
+ variantOp, *transformLibrary, entryPoint);
+ if (!transformRoot) {
+ return;
+ }
if (failed(transform::applyTransformNamedSequence(
- variantOp, *transformLibrary, options.enableExpensiveChecks(true),
- libraryFunc->getLeafReference()))) {
+ variantOp, transformRoot, *transformLibrary,
+ options.enableExpensiveChecks(true)))) {
return signalPassFailure();
}
diff --git a/compiler/src/iree/compiler/Codegen/LLVMCPU/DispatchABI.cpp b/compiler/src/iree/compiler/Codegen/LLVMCPU/DispatchABI.cpp
index 6524e5e..324ec17 100644
--- a/compiler/src/iree/compiler/Codegen/LLVMCPU/DispatchABI.cpp
+++ b/compiler/src/iree/compiler/Codegen/LLVMCPU/DispatchABI.cpp
@@ -1011,7 +1011,8 @@
// Load the placeholder global ordinal.
Value globalPtr = builder.create<LLVM::AddressOfOp>(loc, globalOp);
- Value ordinalValue = builder.create<LLVM::LoadOp>(loc, globalPtr);
+ Value ordinalValue =
+ builder.create<LLVM::LoadOp>(loc, globalOp.getType(), globalPtr);
// Load constant from the executable constants struct.
auto constantsPtrValue =
@@ -1051,7 +1052,7 @@
// Load the placeholder global ordinal.
Value globalPtr = builder.create<LLVM::AddressOfOp>(loc, globalOp);
- return builder.create<LLVM::LoadOp>(loc, globalPtr);
+ return builder.create<LLVM::LoadOp>(loc, globalOp.getType(), globalPtr);
}
std::pair<Value, Value> HALDispatchABI::loadImportFunc(Operation *forOp,
diff --git a/compiler/src/iree/compiler/Codegen/LLVMCPU/test/hal_executable_constants.mlir b/compiler/src/iree/compiler/Codegen/LLVMCPU/test/hal_executable_constants.mlir
index 0df1be1..45341e0 100644
--- a/compiler/src/iree/compiler/Codegen/LLVMCPU/test/hal_executable_constants.mlir
+++ b/compiler/src/iree/compiler/Codegen/LLVMCPU/test/hal_executable_constants.mlir
@@ -3,7 +3,7 @@
// CHECK-DAG: llvm.mlir.global internal @__constant_ordinal_foo() {{.+}}hal.executable.constant.key = "foo"{{.+}} : i32
// CHECK-LABEL: llvm.func @constant_values
func.func @constant_values() {
- // CHECK: %[[FOO_ORDINAL_PTR:.+]] = llvm.mlir.addressof @__constant_ordinal_foo : !llvm.ptr<i32>
+ // CHECK: %[[FOO_ORDINAL_PTR:.+]] = llvm.mlir.addressof @__constant_ordinal_foo : !llvm.ptr
// CHECK: %[[FOO_ORDINAL:.+]] = llvm.load %[[FOO_ORDINAL_PTR]]
// CHECK: %[[FOO_PTR:.+]] = llvm.getelementptr %{{.+}}[%[[FOO_ORDINAL]]]
// CHECK: %[[FOO:.+]] = llvm.load %[[FOO_PTR]]
diff --git a/compiler/src/iree/compiler/Codegen/LLVMGPU/ConvertToLLVM.cpp b/compiler/src/iree/compiler/Codegen/LLVMGPU/ConvertToLLVM.cpp
index 589d9dc..bf9926c 100644
--- a/compiler/src/iree/compiler/Codegen/LLVMGPU/ConvertToLLVM.cpp
+++ b/compiler/src/iree/compiler/Codegen/LLVMGPU/ConvertToLLVM.cpp
@@ -76,8 +76,7 @@
loc, IntegerType::get(builder.getContext(), 64),
builder.getI64IntegerAttr(offset));
Value shiftedPtr = builder.create<LLVM::GEPOp>(
- loc, globalPtr.getType(),
- LLVM::LLVMPointerType::get(globalOp.getContext()), globalPtr,
+ loc, globalPtr.getType(), global.getGlobalType(), globalPtr,
ValueRange({zero, offsetValue}));
addressOfOp.replaceAllUsesWith(shiftedPtr);
addressOfOp.erase();
diff --git a/compiler/src/iree/compiler/Codegen/LLVMGPU/test/convert_to_nvvm.mlir b/compiler/src/iree/compiler/Codegen/LLVMGPU/test/convert_to_nvvm.mlir
index fdc4ce6..3a63bac 100644
--- a/compiler/src/iree/compiler/Codegen/LLVMGPU/test/convert_to_nvvm.mlir
+++ b/compiler/src/iree/compiler/Codegen/LLVMGPU/test/convert_to_nvvm.mlir
@@ -212,18 +212,18 @@
}
// CHECK: llvm.mlir.global external @__dynamic_shared_memory__() {addr_space = 3 : i32, alignment = 16 : i64} : !llvm.array<0 x i8>
// CHECK-LABEL: llvm.func @shared_memory_lowering() {
-// CHECK: %{{.*}} = llvm.mlir.addressof @__dynamic_shared_memory__ : !llvm.ptr<array<0 x i8>, 3>
+// CHECK: %{{.*}} = llvm.mlir.addressof @__dynamic_shared_memory__ : !llvm.ptr<3>
// CHECK-NEXT: %{{.*}} = llvm.mlir.constant(0 : i64) : i64
// CHECK-NEXT: %{{.*}} = llvm.mlir.constant(0 : i64) : i64
-// CHECK-NEXT: %{{.*}} = llvm.getelementptr %{{.*}} : (!llvm.ptr<array<0 x i8>, 3>, i64, i64) -> !llvm.ptr<array<0 x i8>, 3>
-// CHECK: %{{.*}} = llvm.mlir.addressof @__dynamic_shared_memory__ : !llvm.ptr<array<0 x i8>, 3>
+// CHECK-NEXT: %{{.*}} = llvm.getelementptr %{{.*}} : (!llvm.ptr<3>, i64, i64) -> !llvm.ptr<3>
+// CHECK: %{{.*}} = llvm.mlir.addressof @__dynamic_shared_memory__ : !llvm.ptr<3>
// CHECK-NEXT: %{{.*}} = llvm.mlir.constant(0 : i64) : i64
// CHECK-NEXT: %{{.*}} = llvm.mlir.constant(2048 : i64) : i64
-// CHECK-NEXT: %{{.*}} = llvm.getelementptr %{{.*}} : (!llvm.ptr<array<0 x i8>, 3>, i64, i64) -> !llvm.ptr<array<0 x i8>, 3>
-// CHECK: %{{.*}} = llvm.mlir.addressof @__dynamic_shared_memory__ : !llvm.ptr<array<0 x i8>, 3>
+// CHECK-NEXT: %{{.*}} = llvm.getelementptr %{{.*}} : (!llvm.ptr<3>, i64, i64) -> !llvm.ptr<3>
+// CHECK: %{{.*}} = llvm.mlir.addressof @__dynamic_shared_memory__ : !llvm.ptr<3>
// CHECK-NEXT: %{{.*}} = llvm.mlir.constant(0 : i64) : i64
// CHECK-NEXT: %{{.*}} = llvm.mlir.constant(4096 : i64) : i64
-// CHECK-NEXT: %{{.*}} = llvm.getelementptr %{{.*}} : (!llvm.ptr<array<0 x i8>, 3>, i64, i64) -> !llvm.ptr<array<0 x i8>, 3>
+// CHECK-NEXT: %{{.*}} = llvm.getelementptr %{{.*}} : (!llvm.ptr<3>, i64, i64) -> !llvm.ptr<3>
// -----
@@ -240,7 +240,7 @@
func.func @shared_memory_dealloc_elision() {
%f0 = arith.constant 0.0 : f32
%c0 = arith.constant 0 : index
- // CHECK: llvm.mlir.addressof @__dynamic_shared_memory__ : !llvm.ptr<array<0 x i8>, 3>
+ // CHECK: llvm.mlir.addressof @__dynamic_shared_memory__ : !llvm.ptr<3>
%0 = memref.alloc() : memref<1xf32, #gpu.address_space<workgroup>>
memref.store %f0, %0[%c0] : memref<1xf32, #gpu.address_space<workgroup>>
// CHECK-NOT: free
@@ -277,14 +277,14 @@
}
// CHECK-LABEL: llvm.mlir.global external @__dynamic_shared_memory__() {addr_space = 3 : i32, alignment = 16 : i64} : !llvm.array<0 x i8>
// CHECK-LABEL: llvm.func @shared_memory_lowering_aligned_alloc() {
-// CHECK: %{{.*}} = llvm.mlir.addressof @__dynamic_shared_memory__ : !llvm.ptr<array<0 x i8>, 3>
+// CHECK: %{{.*}} = llvm.mlir.addressof @__dynamic_shared_memory__ : !llvm.ptr<3>
// CHECK-NEXT: %{{.*}} = llvm.mlir.constant(0 : i64) : i64
// CHECK-NEXT: %{{.*}} = llvm.mlir.constant(0 : i64) : i64
-// CHECK-NEXT: %{{.*}} = llvm.getelementptr %{{.*}} : (!llvm.ptr<array<0 x i8>, 3>, i64, i64) -> !llvm.ptr<array<0 x i8>, 3>
-// CHECK: %{{.*}} = llvm.mlir.addressof @__dynamic_shared_memory__ : !llvm.ptr<array<0 x i8>, 3>
+// CHECK-NEXT: %{{.*}} = llvm.getelementptr %{{.*}} : (!llvm.ptr<3>, i64, i64) -> !llvm.ptr<3>
+// CHECK: %{{.*}} = llvm.mlir.addressof @__dynamic_shared_memory__ : !llvm.ptr<3>
// CHECK-NEXT: %{{.*}} = llvm.mlir.constant(0 : i64) : i64
// CHECK-NEXT: %{{.*}} = llvm.mlir.constant(4 : i64) : i64
-// CHECK-NEXT: %{{.*}} = llvm.getelementptr %{{.*}} : (!llvm.ptr<array<0 x i8>, 3>, i64, i64) -> !llvm.ptr<array<0 x i8>, 3>
+// CHECK-NEXT: %{{.*}} = llvm.getelementptr %{{.*}} : (!llvm.ptr<3>, i64, i64) -> !llvm.ptr<3>
// -----
@@ -392,7 +392,7 @@
}
// CHECK: llvm.mlir.global external @__dynamic_shared_memory__() {addr_space = 3 : i32, alignment = 16 : i64} : !llvm.array<0 x i8>
// CHECK-LABEL: llvm.func @shared_memory_lowering_index() {
-// CHECK: %{{.*}} = llvm.mlir.addressof @__dynamic_shared_memory__ : !llvm.ptr<array<0 x i8>, 3>
+// CHECK: %{{.*}} = llvm.mlir.addressof @__dynamic_shared_memory__ : !llvm.ptr<3>
// CHECK-NEXT: %{{.*}} = llvm.mlir.constant(0 : i64) : i64
// CHECK-NEXT: %{{.*}} = llvm.mlir.constant(0 : i64) : i64
-// CHECK-NEXT: %{{.*}} = llvm.getelementptr %{{.*}} : (!llvm.ptr<array<0 x i8>, 3>, i64, i64) -> !llvm.ptr<array<0 x i8>, 3>
+// CHECK-NEXT: %{{.*}} = llvm.getelementptr %{{.*}} : (!llvm.ptr<3>, i64, i64) -> !llvm.ptr<3>
diff --git a/compiler/src/iree/compiler/Dialect/Flow/Transforms/CollapseDimensions.cpp b/compiler/src/iree/compiler/Dialect/Flow/Transforms/CollapseDimensions.cpp
index 19a0a43..bf0336c 100644
--- a/compiler/src/iree/compiler/Dialect/Flow/Transforms/CollapseDimensions.cpp
+++ b/compiler/src/iree/compiler/Dialect/Flow/Transforms/CollapseDimensions.cpp
@@ -421,8 +421,8 @@
rewriter.setInsertionPoint(genericOp.value());
FailureOr<SmallVector<Value>> maybeReplacements =
- mlir::linalg::collapseGenericOpIterationDims(genericOp.value(),
- collapseIndices, rewriter);
+ mlir::linalg::collapseOpIterationDims(genericOp.value(), collapseIndices,
+ rewriter);
if (failed(maybeReplacements))
return false;
rewriter.replaceOp(genericOp.value(), maybeReplacements.value());
diff --git a/compiler/src/iree/compiler/Dialect/Flow/Transforms/CollapseReductionDims.cpp b/compiler/src/iree/compiler/Dialect/Flow/Transforms/CollapseReductionDims.cpp
index 407777b..0bd3185 100644
--- a/compiler/src/iree/compiler/Dialect/Flow/Transforms/CollapseReductionDims.cpp
+++ b/compiler/src/iree/compiler/Dialect/Flow/Transforms/CollapseReductionDims.cpp
@@ -8,6 +8,7 @@
#include "iree/compiler/Dialect/Flow/Transforms/Passes.h"
#include "iree/compiler/Dialect/Flow/Transforms/RegionOpUtils.h"
#include "mlir/Dialect/Linalg/IR/Linalg.h"
+#include "mlir/Dialect/Linalg/IR/LinalgInterfaces.h"
#include "mlir/Dialect/Linalg/Transforms/Transforms.h"
#include "mlir/Transforms/GreedyPatternRewriteDriver.h"
@@ -44,19 +45,19 @@
}
static SmallVector<ReassociationIndices>
-collapseDimensions(linalg::GenericOp genericOp) {
+collapseDimensions(linalg::LinalgOp linalgOp) {
SmallVector<ReassociationIndices> collapseIndices;
- if (!isNonNullAndOutsideDispatch(genericOp)) {
+ if (!isNonNullAndOutsideDispatch(linalgOp)) {
return collapseIndices;
}
SmallVector<unsigned> reductionDims;
- genericOp.getReductionDims(reductionDims);
+ linalgOp.getReductionDims(reductionDims);
if (reductionDims.size() < 2)
return collapseIndices;
- for (AffineMap map : genericOp.getIndexingMapsArray()) {
+ for (AffineMap map : linalgOp.getIndexingMapsArray()) {
if (!hasContiguousDims(map, reductionDims))
return collapseIndices;
}
diff --git a/compiler/src/iree/compiler/InputConversion/TOSA/Passes.cpp b/compiler/src/iree/compiler/InputConversion/TOSA/Passes.cpp
index a559350..ae907a7 100644
--- a/compiler/src/iree/compiler/InputConversion/TOSA/Passes.cpp
+++ b/compiler/src/iree/compiler/InputConversion/TOSA/Passes.cpp
@@ -49,7 +49,10 @@
iree_compiler::createTosaToLinalgExt());
passManager.addNestedPass<func::FuncOp>(mlir::createCanonicalizerPass());
- tosa::addTosaToLinalgPasses(passManager, TosaToLinalgOptions());
+ TosaToLinalgNamedOptions tosaToLinalgNamedOptions;
+ tosaToLinalgNamedOptions.preferConv2DKernelLayoutHWCF = true;
+ tosa::addTosaToLinalgPasses(passManager, TosaToLinalgOptions(),
+ tosaToLinalgNamedOptions);
passManager.addNestedPass<func::FuncOp>(
iree_compiler::createConverti48Toi64());
diff --git a/llvm-external-projects/iree-dialects/test/Dialect/iree_linalg_ext/decompose_softmax.mlir b/llvm-external-projects/iree-dialects/test/Dialect/iree_linalg_ext/decompose_softmax.mlir
index ea0717c..381669f 100644
--- a/llvm-external-projects/iree-dialects/test/Dialect/iree_linalg_ext/decompose_softmax.mlir
+++ b/llvm-external-projects/iree-dialects/test/Dialect/iree_linalg_ext/decompose_softmax.mlir
@@ -11,7 +11,7 @@
// CHECK: func.func @softmax(%[[ARG0:[a-zA-Z0-9_]+]]: tensor<2x16x32xf32>) -> tensor<2x16x32xf32> {
// CHECK: %[[D0:.+]] = tensor.empty() : tensor<2x16x32xf32>
// CHECK: %[[D1:.+]] = tensor.empty() : tensor<2x16xf32>
-// CHECK: %[[CST:.+]] = arith.constant -1.401300e-45 : f32
+// CHECK: %[[CST:.+]] = arith.constant -3.40282347E+38 : f32
// CHECK: %[[D2:.+]] = linalg.fill ins(%[[CST]] : f32) outs(%[[D1]] : tensor<2x16xf32>) -> tensor<2x16xf32>
// CHECK: %[[D3:.+]] = linalg.generic {indexing_maps = [#[[MAP]], #[[MAP1]]], iterator_types = ["parallel",
// CHECK-SAME: "parallel", "reduction"]} ins(%[[ARG0]] : tensor<2x16x32xf32>) outs(%[[D2]] : tensor<2x16xf32>) {
diff --git a/third_party/llvm-project b/third_party/llvm-project
index 6aaa03a..acc6f3e 160000
--- a/third_party/llvm-project
+++ b/third_party/llvm-project
@@ -1 +1 @@
-Subproject commit 6aaa03a0232b89682e0745433bdfb4d785f64a01
+Subproject commit acc6f3e9c1af6c7445aae6f10d4b016ac84112d3
