llvmbot
commented
2 months ago @llvm/issue-subscribers-mlir
Author: Alankar Misra (alankarmisra)
In the following example, I arith.addi and arith.floordivsi with constant vectors and vectors loaded from a memref. The add will succeed in lowering to llvm in both situations but floordivsi fails to lower to llvm with vectors loaded from memrefs for some reason.
```mlir
module @bug {
func.func @main() {
%number = arith.constant dense<10> : vector<1xi32>
%resolved_floordivsi = arith.floordivsi %number, %number : vector<1xi32>
%alloca = memref.alloca() : memref<1xi32>
%idx = arith.constant 0 : index
vector.store %number, %alloca[%idx] : memref<1xi32>, vector<1xi32>
%loadedNum = vector.load %alloca[%idx] : memref<1xi32>, vector<1xi32>
%unresolved_floordivsi = arith.floordivsi %loadedNum, %number: vector<1xi32>
vector.print %resolved_floordivsi : vector<1xi32>
vector.print %unresolved_floordivsi : vector<1xi32>
%resolved_add = arith.addi %number, %number : vector<1xi32>
%resolved_add_too = arith.addi %loadedNum, %number : vector<1xi32>
vector.print %resolved_add : vector<1xi32>
vector.print %resolved_add_too : vector<1xi32>
return
}
}
```
Running
```bash
mlir-opt input.mlir \
-lower-affine \
-convert-vector-to-llvm \
-convert-arith-to-llvm \
-convert-func-to-llvm \
-convert-index-to-llvm \
--convert-to-llvm \
-o out.mlir
```
gives me
```mlir
module @bug {
llvm.func @main() {
%0 = llvm.mlir.constant(dense<20> : vector<1xi32>) : vector<1xi32>
%1 = llvm.mlir.constant(0 : index) : i64
%2 = builtin.unrealized_conversion_cast %1 : i64 to index
%3 = llvm.mlir.constant(dense<10> : vector<1xi32>) : vector<1xi32>
%4 = llvm.mlir.constant(dense<1> : vector<1xi32>) : vector<1xi32>
%5 = llvm.mlir.constant(1 : index) : i64
%6 = llvm.mlir.constant(1 : index) : i64
%7 = llvm.alloca %5 x i32 : (i64) -> !llvm.ptr
%8 = llvm.mlir.undef : !llvm.struct<(ptr, ptr, i64, array<1 x i64>, array<1 x i64>)>
%9 = llvm.insertvalue %7, %8[0] : !llvm.struct<(ptr, ptr, i64, array<1 x i64>, array<1 x i64>)>
%10 = llvm.insertvalue %7, %9[1] : !llvm.struct<(ptr, ptr, i64, array<1 x i64>, array<1 x i64>)>
%11 = llvm.mlir.constant(0 : index) : i64
%12 = llvm.insertvalue %11, %10[2] : !llvm.struct<(ptr, ptr, i64, array<1 x i64>, array<1 x i64>)>
%13 = llvm.insertvalue %5, %12[3, 0] : !llvm.struct<(ptr, ptr, i64, array<1 x i64>, array<1 x i64>)>
%14 = llvm.insertvalue %6, %13[4, 0] : !llvm.struct<(ptr, ptr, i64, array<1 x i64>, array<1 x i64>)>
%15 = llvm.mlir.constant(0 : i64) : i64
%16 = llvm.extractelement %3[%15 : i64] : vector<1xi32>
%17 = llvm.extractvalue %14[1] : !llvm.struct<(ptr, ptr, i64, array<1 x i64>, array<1 x i64>)>
%18 = llvm.getelementptr %17[%1] : (!llvm.ptr, i64) -> !llvm.ptr, i32
llvm.store %16, %18 : i32, !llvm.ptr
%19 = llvm.extractvalue %14[1] : !llvm.struct<(ptr, ptr, i64, array<1 x i64>, array<1 x i64>)>
%20 = llvm.getelementptr %19[%1] : (!llvm.ptr, i64) -> !llvm.ptr, i32
%21 = llvm.load %20 : !llvm.ptr -> i32
%22 = llvm.mlir.undef : vector<1xi32>
%23 = llvm.mlir.constant(0 : i32) : i32
%24 = llvm.insertelement %21, %22[%23 : i32] : vector<1xi32>
%25 = llvm.shufflevector %24, %22 [0] : vector<1xi32>
%26 = arith.floordivsi %25, %3 : vector<1xi32>
vector.print %4 : vector<1xi32>
vector.print %26 : vector<1xi32>
%27 = llvm.add %25, %3 : vector<1xi32>
vector.print %0 : vector<1xi32>
vector.print %27 : vector<1xi32>
llvm.return
}
}
```
%4 is the result for the first arith.floordivsi for (10, 10) and correctly folds to 1
%26 is the result for the second arith.floordivsi and does not lower to llvm
%0 is the result for the first arith.add for (10, 10) and correctly folds to 20
%27 is the result for the second arith.add and lowers to llvm.add
```bash
mlir-opt --version
Homebrew LLVM version 18.1.8
Optimized build.
```
chadlonso
In the following example, I arith.addi and arith.floordivsi with constant vectors and vectors loaded from a memref. The add will succeed in lowering to llvm in both situations but floordivsi fails to lower to llvm with vectors loaded from memrefs for some reason.
Running
gives me
%4 is the result for the first arith.floordivsi for (10, 10) and correctly folds to 1 %26 is the result for the second arith.floordivsi and does not lower to llvm %0 is the result for the first arith.add for (10, 10) and correctly folds to 20 %27 is the result for the second arith.add and lowers to llvm.add