[flang]Pass to add vscale range attribute #68103
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Add vscale range attirbute for the Scalable Vector Extension (SVE) if provided on the command-line (options in a previous commit) If no command-line option is provided, if the target-feature of SVE is specified and the architecture is AArch64, it defualts to 128-2048. in other words a vscale-min of 1, vscale-max of 16. A pass is used to add the atribute to all functions. The vectorizer will use this attribute to generate the SVE instruction to match the range specified. The attribute is harmless if there is no vectorizable operations in the function.
llvmbot
added
flang:driver
flang
|
@llvm/pr-subscribers-flang-fir-hlfir @llvm/pr-subscribers-flang-driver ChangesAdd vscale range attirbute for the Scalable Vector Extension (SVE) if provided on the command-line (options in a previous commit) If no command-line option is provided, if the target-feature of SVE is specified and the architecture is AArch64, it defualts to 128-2048. in other words a vscale-min of 1, vscale-max of 16. A pass is used to add the atribute to all functions. The vectorizer will use this attribute to generate the SVE instruction to match the range specified. The attribute is harmless if there is no vectorizable operations in the function. Full diff: https://github.com/llvm/llvm-project/pull/68103.diff 8 Files Affected:
diff --git a/flang/include/flang/Optimizer/Transforms/Passes.h b/flang/include/flang/Optimizer/Transforms/Passes.h
index 8aeb3e373298e88..64882c8ec406b0a 100644
--- a/flang/include/flang/Optimizer/Transforms/Passes.h
+++ b/flang/include/flang/Optimizer/Transforms/Passes.h
@@ -80,6 +80,10 @@ std::unique_ptr<mlir::Pass> createOMPFunctionFilteringPass();
std::unique_ptr<mlir::OperationPass<mlir::ModuleOp>>
createOMPMarkDeclareTargetPass();
+std::unique_ptr<mlir::Pass> createVScaleAttrPass();
+std::unique_ptr<mlir::Pass>
+createVScaleAttrPass(std::pair<unsigned, unsigned> vscaleAttr);
+
// declarative passes
#define GEN_PASS_REGISTRATION
#include "flang/Optimizer/Transforms/Passes.h.inc"
diff --git a/flang/include/flang/Optimizer/Transforms/Passes.td b/flang/include/flang/Optimizer/Transforms/Passes.td
index 9474edf13ce4639..f5022beaaeec7f9 100644
--- a/flang/include/flang/Optimizer/Transforms/Passes.td
+++ b/flang/include/flang/Optimizer/Transforms/Passes.td
@@ -326,4 +326,18 @@ def OMPFunctionFiltering : Pass<"omp-function-filtering"> {
];
}
+def VScaleAttr : Pass<"vscale-attr", "mlir::func::FuncOp"> {
+ let summary = "Add vscale_range attribute to functions";
+ let description = [{
+ Set an attribute for the vscale range on functions, to allow scalable
+ vector operations to be used on Arm processors.
+ }];
+ let options = [
+ Option<"vscaleRange", "vscale-range",
+ "std::pair<unsigned, unsigned>", /*default=*/"std::pair<unsigned, unsigned>{}",
+ "vector scale range">,
+ ];
+ let constructor = "::fir::createVScaleAttrPass()";
+}
+
#endif // FLANG_OPTIMIZER_TRANSFORMS_PASSES
diff --git a/flang/include/flang/Tools/CLOptions.inc b/flang/include/flang/Tools/CLOptions.inc
index 0b210bec9f785c8..f7737a5b79f08fd 100644
--- a/flang/include/flang/Tools/CLOptions.inc
+++ b/flang/include/flang/Tools/CLOptions.inc
@@ -202,6 +202,9 @@ inline void createDefaultFIROptimizerPassPipeline(
if (pc.LoopVersioning)
pm.addPass(fir::createLoopVersioningPass());
+ if (pc.VScaleMin != 0)
+ pm.addPass(fir::createVScaleAttrPass({pc.VScaleMin,pc.VScaleMax}));
+
pm.addPass(mlir::createCSEPass());
if (pc.StackArrays)
diff --git a/flang/include/flang/Tools/CrossToolHelpers.h b/flang/include/flang/Tools/CrossToolHelpers.h
index 8453828995ad015..a15f43aac55d068 100644
--- a/flang/include/flang/Tools/CrossToolHelpers.h
+++ b/flang/include/flang/Tools/CrossToolHelpers.h
@@ -42,6 +42,8 @@ struct MLIRToLLVMPassPipelineConfig {
bool LoopVersioning = false; ///< Run the version loop pass.
llvm::codegenoptions::DebugInfoKind DebugInfo =
llvm::codegenoptions::NoDebugInfo; ///< Debug info generation.
+ unsigned VScaleMin = 0; ///< SVE vector range minimum.
+ unsigned VScaleMax = 0; ///< SVE vector range minimum.
};
struct OffloadModuleOpts {
diff --git a/flang/lib/Frontend/FrontendActions.cpp b/flang/lib/Frontend/FrontendActions.cpp
index fa12e37607cf110..67f2fab2804b9f9 100644
--- a/flang/lib/Frontend/FrontendActions.cpp
+++ b/flang/lib/Frontend/FrontendActions.cpp
@@ -695,6 +695,21 @@ void CodeGenAction::lowerHLFIRToFIR() {
}
}
+// TODO: We should get this from TargetInfo. However, that depends on
+// too much of clang, so for now, replicate the functionality.
+static std::optional<std::pair<unsigned, unsigned>>
+getVScaleRange(CompilerInstance& ci, const Fortran::frontend::LangOptions &langOpts) {
+ if (langOpts.VScaleMin || langOpts.VScaleMax)
+ return std::pair<unsigned, unsigned>(
+ langOpts.VScaleMin ? langOpts.VScaleMin : 1, langOpts.VScaleMax);
+
+ std::string featuresStr = getTargetFeatures(ci);
+ if (featuresStr.find("+sve") != std::string::npos)
+ return std::pair<unsigned, unsigned>(1, 16);
+
+ return std::nullopt;
+}
+
// Lower the previously generated MLIR module into an LLVM IR module
void CodeGenAction::generateLLVMIR() {
assert(mlirModule && "The MLIR module has not been generated yet.");
@@ -715,6 +730,18 @@ void CodeGenAction::generateLLVMIR() {
MLIRToLLVMPassPipelineConfig config(level, opts);
+ const auto targetOpts = ci.getInvocation().getTargetOpts();
+ const llvm::Triple triple(targetOpts.triple);
+
+ // Only get the vscale range if AArch64.
+ if (triple.isAArch64()) {
+ auto langOpts = ci.getInvocation().getLangOpts();
+ if (auto vsr = getVScaleRange(ci, langOpts)) {
+ config.VScaleMin = vsr->first;
+ config.VScaleMax = vsr->second;
+ }
+ }
+
// Create the pass pipeline
fir::createMLIRToLLVMPassPipeline(pm, config);
(void)mlir::applyPassManagerCLOptions(pm);
diff --git a/flang/lib/Optimizer/Transforms/CMakeLists.txt b/flang/lib/Optimizer/Transforms/CMakeLists.txt
index 3d2b7e5eaeade0a..428c4c2a1e64408 100644
--- a/flang/lib/Optimizer/Transforms/CMakeLists.txt
+++ b/flang/lib/Optimizer/Transforms/CMakeLists.txt
@@ -19,6 +19,7 @@ add_flang_library(FIRTransforms
OMPEarlyOutlining.cpp
OMPFunctionFiltering.cpp
OMPMarkDeclareTarget.cpp
+ VScaleAttr.cpp
DEPENDS
FIRDialect
diff --git a/flang/lib/Optimizer/Transforms/VScaleAttr.cpp b/flang/lib/Optimizer/Transforms/VScaleAttr.cpp
new file mode 100644
index 000000000000000..601a937de37be9b
--- /dev/null
+++ b/flang/lib/Optimizer/Transforms/VScaleAttr.cpp
@@ -0,0 +1,90 @@
+//===- VScaleAttr.cpp -------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+//===----------------------------------------------------------------------===//
+/// \file
+/// This pass adds a `vscale_range` attribute to function definitions.
+/// The attribute is used for scalable vector operations on Arm processors
+/// and should only be run on processors that support this feature. [It is
+/// likely harmless to run it on something else, but it is also not valuable].
+//===----------------------------------------------------------------------===//
+
+#include "flang/ISO_Fortran_binding_wrapper.h"
+#include "flang/Optimizer/Builder/BoxValue.h"
+#include "flang/Optimizer/Builder/FIRBuilder.h"
+#include "flang/Optimizer/Builder/Runtime/Inquiry.h"
+#include "flang/Optimizer/Dialect/FIRDialect.h"
+#include "flang/Optimizer/Dialect/FIROps.h"
+#include "flang/Optimizer/Dialect/FIRType.h"
+#include "flang/Optimizer/Dialect/Support/FIRContext.h"
+#include "flang/Optimizer/Dialect/Support/KindMapping.h"
+#include "flang/Optimizer/Transforms/Passes.h"
+#include "mlir/Dialect/LLVMIR/LLVMAttrs.h"
+#include "mlir/Dialect/LLVMIR/LLVMDialect.h"
+#include "mlir/IR/Matchers.h"
+#include "mlir/IR/TypeUtilities.h"
+#include "mlir/Pass/Pass.h"
+#include "mlir/Transforms/DialectConversion.h"
+#include "mlir/Transforms/GreedyPatternRewriteDriver.h"
+#include "mlir/Transforms/RegionUtils.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/Support/raw_ostream.h"
+
+#include <algorithm>
+
+namespace fir {
+#define GEN_PASS_DECL_VSCALEATTR
+#define GEN_PASS_DEF_VSCALEATTR
+#include "flang/Optimizer/Transforms/Passes.h.inc"
+} // namespace fir
+
+#define DEBUG_TYPE "vscale-attr"
+
+namespace {
+
+class VScaleAttrPass : public fir::impl::VScaleAttrBase<VScaleAttrPass> {
+public:
+ VScaleAttrPass(const fir::VScaleAttrOptions &options) {
+ vscaleRange = options.vscaleRange;
+ }
+ VScaleAttrPass() {}
+ void runOnOperation() override;
+};
+
+} // namespace
+
+void VScaleAttrPass::runOnOperation() {
+ LLVM_DEBUG(llvm::dbgs() << "=== Begin " DEBUG_TYPE " ===\n");
+ mlir::func::FuncOp func = getOperation();
+
+ LLVM_DEBUG(llvm::dbgs() << "Func-name:" << func.getSymName() << "\n");
+
+ auto context = &getContext();
+
+ auto intTy = mlir::IntegerType::get(context, 32);
+
+ assert(vscaleRange.first && "VScaleRange minimum should be non-zero");
+
+ func->setAttr("vscale_range",
+ mlir::LLVM::VScaleRangeAttr::get(
+ context, mlir::IntegerAttr::get(intTy, vscaleRange.first),
+ mlir::IntegerAttr::get(intTy, vscaleRange.second)));
+
+ LLVM_DEBUG(llvm::dbgs() << "=== End " DEBUG_TYPE " ===\n");
+}
+
+std::unique_ptr<mlir::Pass>
+fir::createVScaleAttrPass(std::pair<unsigned, unsigned> vscaleAttr) {
+ VScaleAttrOptions opts;
+ opts.vscaleRange = vscaleAttr;
+ return std::make_unique<VScaleAttrPass>(opts);
+}
+
+std::unique_ptr<mlir::Pass> fir::createVScaleAttrPass() {
+ return std::make_unique<VScaleAttrPass>();
+}
diff --git a/flang/test/Lower/Arm/arm-sve-vector-bits-vscale-range.f90 b/flang/test/Lower/Arm/arm-sve-vector-bits-vscale-range.f90
new file mode 100644
index 000000000000000..1c413477826e5b9
--- /dev/null
+++ b/flang/test/Lower/Arm/arm-sve-vector-bits-vscale-range.f90
@@ -0,0 +1,23 @@
+! RUN: %flang_fc1 -triple aarch64-none-linux-gnu -target-feature +sve -mvscale-min=1 -mvscale-max=1 -emit-llvm -o - %s | FileCheck %s -D#VBITS=1
+! RUN: %flang_fc1 -triple aarch64-none-linux-gnu -target-feature +sve -mvscale-min=2 -mvscale-max=2 -emit-llvm -o - %s | FileCheck %s -D#VBITS=2
+! RUN: %flang_fc1 -triple aarch64-none-linux-gnu -target-feature +sve -mvscale-min=4 -mvscale-max=4 -emit-llvm -o - %s | FileCheck %s -D#VBITS=4
+! RUN: %flang_fc1 -triple aarch64-none-linux-gnu -target-feature +sve -mvscale-min=8 -mvscale-max=8 -emit-llvm -o - %s | FileCheck %s -D#VBITS=8
+! RUN: %flang_fc1 -triple aarch64-none-linux-gnu -target-feature +sve -mvscale-min=16 -mvscale-max=16 -emit-llvm -o - %s | FileCheck %s -D#VBITS=16
+! RUN: %flang_fc1 -triple aarch64-none-linux-gnu -target-feature +sve2 -mvscale-min=1 -mvscale-max=1 -emit-llvm -o - %s | FileCheck %s -D#VBITS=1
+! RUN: %flang_fc1 -triple aarch64-none-linux-gnu -target-feature +sve2 -mvscale-min=2 -mvscale-max=2 -emit-llvm -o - %s | FileCheck %s -D#VBITS=2
+! RUN: %flang_fc1 -triple aarch64-none-linux-gnu -target-feature +sve -mvscale-min=1 -emit-llvm -o - %s | FileCheck %s -D#VBITS=1 --check-prefix=CHECK-NOMAX
+! RUN: %flang_fc1 -triple aarch64-none-linux-gnu -target-feature +sve -mvscale-min=2 -emit-llvm -o - %s | FileCheck %s -D#VBITS=2 --check-prefix=CHECK-NOMAX
+! RUN: %flang_fc1 -triple aarch64-none-linux-gnu -target-feature +sve -mvscale-min=4 -emit-llvm -o - %s | FileCheck %s -D#VBITS=4 --check-prefix=CHECK-NOMAX
+! RUN: %flang_fc1 -triple aarch64-none-linux-gnu -target-feature +sve -mvscale-min=8 -emit-llvm -o - %s | FileCheck %s -D#VBITS=8 --check-prefix=CHECK-NOMAX
+! RUN: %flang_fc1 -triple aarch64-none-linux-gnu -target-feature +sve -mvscale-min=16 -emit-llvm -o - %s | FileCheck %s -D#VBITS=16 --check-prefix=CHECK-NOMAX
+! RUN: %flang_fc1 -triple aarch64-none-linux-gnu -target-feature +sve2 -mvscale-min=1 -mvscale-max=0 -emit-llvm -o - %s | FileCheck %s --check-prefix=CHECK-UNBOUNDED
+! RUN: %flang_fc1 -triple aarch64-none-linux-gnu -target-feature +sve -mvscale-min=1 -mvscale-max=0 -emit-llvm -o - %s | FileCheck %s --check-prefix=CHECK-UNBOUNDED
+! RUN: %flang_fc1 -triple aarch64-none-linux-gnu -target-feature +sve -emit-llvm -o - %s | FileCheck %s --check-prefix=CHECK-NONE
+
+! CHECK-LABEL: @func_() #0
+! CHECK: attributes #0 = {{{.*}} vscale_range([[#VBITS]],[[#VBITS]]) {{.*}}}
+! CHECK-NOMAX: attributes #0 = {{{.*}} vscale_range([[#VBITS]],0) {{.*}}}
+! CHECK-UNBOUNDED: attributes #0 = {{{.*}} vscale_range(1,0) {{.*}}}
+! CHECK-NONE: attributes #0 = {{{.*}} vscale_range(1,16) {{.*}}}
+subroutine func
+end subroutine func
|
|
✅ With the latest revision this PR passed the C/C++ code formatter. |
kiranchandramohan
requested review from
banach-space,
tblah,
jeanPerier and
davemgreen
jeanPerier
reviewed
Oct 4, 2023
There was a problem hiding this comment.
Pass plugin-in looks ok to me (please address the formatting issue from the build bot). I only wonder a bit if this should not be done in the codegen/target rewrite pipeline rather than in createDefaultFIROptimizerPassPipeline though (unless the attribute matters to MLIR/FIR optimizations like inlining ?)
I do not know about enough LLVM vscale range attribute to review the logic of where to set it and with which values, so best if you get approval from an SVE knowledgeable person here.
tblah
reviewed
Oct 4, 2023
There was a problem hiding this comment.
Looks good. But please address @jeanPerier's comment. This looks to me like it should be more of a codegen thing as well.
Also change wording to more generic for pass description
tblah
approved these changes
Oct 4, 2023
|
This commit breaks ppc64le-flang-rhel-clang build. $ cat end.f90
end
$ flang-new -fc1 -triple aarch64-none-linux-gnu -emit-llvm -o - end.f90
error: unable to create target: 'No available targets are compatible with triple "aarch64-none-linux-gnu"' |
Ah. I'm not sure why the clang test works (or isn't being run), as far as I understand, the same test (except for it using clang rather than flang) I'll spend a few minutes trying to find a fix, and if I can't, I'll revert. |
Fix here: |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Add vscale range attirbute for the Scalable Vector Extension (SVE) if provided on the command-line (options in a previous commit)
If no command-line option is provided, if the target-feature of SVE is specified and the architecture is AArch64, it defualts to 128-2048. in other words a vscale-min of 1, vscale-max of 16.
A pass is used to add the atribute to all functions. The vectorizer will use this attribute to generate the SVE instruction to match the range specified. The attribute is harmless if there is no vectorizable operations in the function.