[CIR] Implement folder for VecShuffleDynamicOp #142315
Conversation
llvmbot
added
clang
|
@llvm/pr-subscribers-clang @llvm/pr-subscribers-clangir Author: Amr Hesham (AmrDeveloper) ChangesThis change adds a folder for the VecShuffleDynamicOp Issue #136487 Full diff: https://github.com/llvm/llvm-project/pull/142315.diff 4 Files Affected:
diff --git a/clang/include/clang/CIR/Dialect/IR/CIROps.td b/clang/include/clang/CIR/Dialect/IR/CIROps.td
index 07851610a2abd..4442c3a5607ae 100644
--- a/clang/include/clang/CIR/Dialect/IR/CIROps.td
+++ b/clang/include/clang/CIR/Dialect/IR/CIROps.td
@@ -2188,6 +2188,7 @@ def VecShuffleDynamicOp : CIR_Op<"vec.shuffle.dynamic",
}];
let hasVerifier = 1;
+ let hasFolder = 1;
}
#endif // CLANG_CIR_DIALECT_IR_CIROPS_TD
diff --git a/clang/lib/CIR/Dialect/IR/CIRDialect.cpp b/clang/lib/CIR/Dialect/IR/CIRDialect.cpp
index 36f050de9f8bb..90b32950e4774 100644
--- a/clang/lib/CIR/Dialect/IR/CIRDialect.cpp
+++ b/clang/lib/CIR/Dialect/IR/CIRDialect.cpp
@@ -1579,6 +1579,38 @@ OpFoldResult cir::VecExtractOp::fold(FoldAdaptor adaptor) {
// VecShuffleDynamicOp
//===----------------------------------------------------------------------===//
+OpFoldResult cir::VecShuffleDynamicOp::fold(FoldAdaptor adaptor) {
+ mlir::Attribute vec = adaptor.getVec();
+ mlir::Attribute indices = adaptor.getIndices();
+ if (mlir::isa_and_nonnull<cir::ConstVectorAttr>(vec) &&
+ mlir::isa_and_nonnull<cir::ConstVectorAttr>(indices)) {
+ auto vecAttr = mlir::cast<cir::ConstVectorAttr>(vec);
+ auto indicesAttr = mlir::cast<cir::ConstVectorAttr>(indices);
+ auto vecTy = cast<cir::VectorType>(vecAttr.getType());
+
+ mlir::ArrayAttr vecElts = vecAttr.getElts();
+ mlir::ArrayAttr indicesElts = indicesAttr.getElts();
+
+ const uint64_t numElements = vecElts.size();
+
+ SmallVector<mlir::Attribute, 16> elements;
+ elements.reserve(numElements);
+
+ const uint64_t maskBits = llvm::NextPowerOf2(numElements - 1) - 1;
+ for (uint64_t i = 0; i < numElements; i++) {
+ cir::IntAttr idxAttr = mlir::cast<cir::IntAttr>(indicesElts[i]);
+ uint64_t idxValue = idxAttr.getUInt();
+ uint64_t newIdx = idxValue & maskBits;
+ elements.push_back(vecElts[newIdx]);
+ }
+
+ return cir::ConstVectorAttr::get(
+ vecTy, mlir::ArrayAttr::get(getContext(), elements));
+ }
+
+ return {};
+}
+
LogicalResult cir::VecShuffleDynamicOp::verify() {
// The number of elements in the two input vectors must match.
if (getVec().getType().getSize() !=
diff --git a/clang/lib/CIR/Dialect/Transforms/CIRCanonicalize.cpp b/clang/lib/CIR/Dialect/Transforms/CIRCanonicalize.cpp
index fb000adee04c6..7d03e374c27e8 100644
--- a/clang/lib/CIR/Dialect/Transforms/CIRCanonicalize.cpp
+++ b/clang/lib/CIR/Dialect/Transforms/CIRCanonicalize.cpp
@@ -138,10 +138,10 @@ void CIRCanonicalizePass::runOnOperation() {
assert(!cir::MissingFeatures::complexRealOp());
assert(!cir::MissingFeatures::complexImagOp());
assert(!cir::MissingFeatures::callOp());
- // CastOp, UnaryOp and VecExtractOp are here to perform a manual `fold` in
- // applyOpPatternsGreedily.
+ // CastOp, UnaryOp, VecExtractOp and VecShuffleDynamicOp are here to perform
+ // a manual `fold` in applyOpPatternsGreedily.
if (isa<BrOp, BrCondOp, CastOp, ScopeOp, SwitchOp, SelectOp, UnaryOp,
- VecExtractOp>(op))
+ VecExtractOp, VecShuffleDynamicOp>(op))
ops.push_back(op);
});
diff --git a/clang/test/CIR/Transforms/vector-shuffle-dynamic-fold.cir b/clang/test/CIR/Transforms/vector-shuffle-dynamic-fold.cir
new file mode 100644
index 0000000000000..db5d6bff04d5b
--- /dev/null
+++ b/clang/test/CIR/Transforms/vector-shuffle-dynamic-fold.cir
@@ -0,0 +1,18 @@
+// RUN: cir-opt %s -cir-canonicalize -o - | FileCheck %s
+
+!s32i = !cir.int<s, 32>
+
+module {
+ cir.func @fold_shuffle_dynamic_vector_op_test() {
+ %alloca = cir.alloca !cir.vector<4 x !s32i>, !cir.ptr<!cir.vector<4 x !s32i>>, ["r", init]
+ %vec = cir.const #cir.const_vector<[#cir.int<1> : !s32i, #cir.int<2> : !s32i, #cir.int<3> : !s32i, #cir.int<4> : !s32i]> : !cir.vector<4 x !s32i>
+ %indices = cir.const #cir.const_vector<[#cir.int<8> : !s32i, #cir.int<7> : !s32i, #cir.int<6> : !s32i, #cir.int<5> : !s32i]> : !cir.vector<4 x !s32i>
+ %new_vec = cir.vec.shuffle.dynamic %vec : !cir.vector<4 x !s32i>, %indices : !cir.vector<4 x !s32i>
+ cir.store align(16) %new_vec, %alloca : !cir.vector<4 x !s32i>, !cir.ptr<!cir.vector<4 x !s32i>>
+ cir.return
+ }
+
+ // CHECK: %[[NEW_VEC:.*]] = cir.const #cir.const_vector<[#cir.int<1> : !s32i, #cir.int<4> : !s32i, #cir.int<3> : !s32i, #cir.int<2> : !s32i]> : !cir.vector<4 x !s32i>
+}
+
+
|
| vecTy, mlir::ArrayAttr::get(getContext(), elements)); | ||
| } | ||
|
|
||
| return {}; |
There was a problem hiding this comment.
Should this branch have a 'not yet implemented' thing here?
There was a problem hiding this comment.
As far as I understood, no, it should be null because that means we can't fold VecShuffleDynamicOp with non Const vec operands
| for (uint64_t i = 0; i < numElements; i++) { | ||
| cir::IntAttr idxAttr = mlir::cast<cir::IntAttr>(indicesElts[i]); |
There was a problem hiding this comment.
This might use getAsValueRange() on indicesElts to remove cast and use range for?
| mlir::isa_and_nonnull<cir::ConstVectorAttr>(indices)) { | ||
| auto vecAttr = mlir::cast<cir::ConstVectorAttr>(vec); | ||
| auto indicesAttr = mlir::cast<cir::ConstVectorAttr>(indices); | ||
| auto vecTy = cast<cir::VectorType>(vecAttr.getType()); |
There was a problem hiding this comment.
| auto vecTy = cast<cir::VectorType>(vecAttr.getType()); | |
| auto vecTy = mlir::cast<cir::VectorType>(vecAttr.getType()); |
| for (const mlir::APInt &idxAttr : | ||
| indicesElts.getAsValueRange<cir::IntAttr, mlir::APInt>()) { |
There was a problem hiding this comment.
does not work simply: for (const auto &idxAttr : indicesElts.getAsValueRange<cir::IntAttr>()) {?
There was a problem hiding this comment.
or maybe use just getAsRange ie.:
for (const auto &idxAttr : indicesElts.getAsRange<cir::IntAttr>()) {
and original uint64_t idxValue = idxAttr.getUInt();
There was a problem hiding this comment.
does not work simply: for (const auto &idxAttr : indicesElts.getAsValueRangecir::IntAttr()) {?
Yes, it reports “Cannot form a reference to void”
I will use getAsRange
AmrDeveloper
force-pushed
the
folder_vec_dyn_shuffle
branch
from
b3a2fed to
ec7d646
Compare
| cir.return | ||
| } | ||
|
|
||
| // CHECK: %[[NEW_VEC:.*]] = cir.const #cir.const_vector<[#cir.int<1> : !s32i, #cir.int<4> : !s32i, #cir.int<3> : !s32i, #cir.int<2> : !s32i]> : !cir.vector<4 x !s32i> |
There was a problem hiding this comment.
Please change the testcase to return the vector instead of storing it and have a CHECK line that checks the return value is NEW_VEC.
| cir.return | ||
| } | ||
|
|
||
| // CHECK: %[[NEW_VEC:.*]] = cir.const #cir.const_vector<[#cir.int<1> : !s32i, #cir.int<4> : !s32i, #cir.int<3> : !s32i, #cir.int<2> : !s32i]> : !cir.vector<4 x !s32i> |
There was a problem hiding this comment.
It's not obvious to me from this check what the full form of the transformed function will be. I assume we're getting rid of some dead instructions here, but that's not clear from the check. I'd suggest showing the full transformed function with CHECK-NEXT to verify that no unnecessary artifacts were left behind.
It isn't obvious at all why "8, 7, 6, 5" yielded the result here. I see that those values are ANDed with 3 to use "0, 3, 2, 1" as the shuffle indexes. Maybe add another test with indexes that don't require the mask and add a comment explaining the masking here.
This change adds a folder for the VecShuffleDynamicOp
Issue #136487