[mlir][arith] Canonicalize sitofp(truncf) -> sitofp, and uitofp. #139925
Conversation
|
@llvm/pr-subscribers-mlir-arith Author: Christian Sigg (chsigg) ChangesAdd a canonicalization patterns that simplifies This assumes that the destination type of truncf is representable by the intermediate type. Note that the truncf semantics requires that the destination type is narrower than the source type, so this is true for all types I can possibly think of, but one could probably construct an artificial counter example. Somewhat related: #128096 Full diff: https://github.com/llvm/llvm-project/pull/139925.diff 4 Files Affected:
diff --git a/mlir/include/mlir/Dialect/Arith/IR/ArithOps.td b/mlir/include/mlir/Dialect/Arith/IR/ArithOps.td
index d50b6aeca15c9..599b3b982ec7f 100644
--- a/mlir/include/mlir/Dialect/Arith/IR/ArithOps.td
+++ b/mlir/include/mlir/Dialect/Arith/IR/ArithOps.td
@@ -1273,6 +1273,7 @@ def Arith_TruncFOp :
];
let hasFolder = 1;
+ let hasCanonicalizer = 1;
let hasVerifier = 1;
let assemblyFormat = [{ $in ($roundingmode^)?
(`fastmath` `` $fastmath^)?
diff --git a/mlir/lib/Dialect/Arith/IR/ArithCanonicalization.td b/mlir/lib/Dialect/Arith/IR/ArithCanonicalization.td
index 7e212df9029d1..7be73c4343639 100644
--- a/mlir/lib/Dialect/Arith/IR/ArithCanonicalization.td
+++ b/mlir/lib/Dialect/Arith/IR/ArithCanonicalization.td
@@ -419,6 +419,20 @@ def TruncIShrUIMulIToMulUIExtended :
(ValueWiderThan $mul, $x),
(TruncationMatchesShiftAmount $mul, $x, $c0)]>;
+//===----------------------------------------------------------------------===//
+// TruncIOp
+//===----------------------------------------------------------------------===//
+
+// truncf(sitofp(x)) -> sitofp(x).
+def TruncFSIToFPToSIToFP :
+ Pat<(Arith_TruncFOp:$tr (Arith_SIToFPOp:$fp $x), $rmf, $fmf),
+ (Arith_SIToFPOp $x)>;
+
+// truncf(sitofp(x)) -> sitofp(x).
+def TruncFUIToFPToUIToFP :
+ Pat<(Arith_TruncFOp:$tr (Arith_UIToFPOp:$fp $x), $rmf, $fmf),
+ (Arith_UIToFPOp $x)>;
+
//===----------------------------------------------------------------------===//
// MulFOp
//===----------------------------------------------------------------------===//
diff --git a/mlir/lib/Dialect/Arith/IR/ArithOps.cpp b/mlir/lib/Dialect/Arith/IR/ArithOps.cpp
index 3b308716c84dc..41f2d0f3425e2 100644
--- a/mlir/lib/Dialect/Arith/IR/ArithOps.cpp
+++ b/mlir/lib/Dialect/Arith/IR/ArithOps.cpp
@@ -1552,6 +1552,11 @@ OpFoldResult arith::TruncFOp::fold(FoldAdaptor adaptor) {
});
}
+void arith::TruncFOp::getCanonicalizationPatterns(RewritePatternSet &patterns,
+ MLIRContext *context) {
+ patterns.add<TruncFSIToFPToSIToFP, TruncFUIToFPToUIToFP>(context);
+}
+
bool arith::TruncFOp::areCastCompatible(TypeRange inputs, TypeRange outputs) {
return checkWidthChangeCast<std::less, FloatType>(inputs, outputs);
}
diff --git a/mlir/test/Dialect/Arith/canonicalize.mlir b/mlir/test/Dialect/Arith/canonicalize.mlir
index d62c5b18fd041..a5dab73a62fac 100644
--- a/mlir/test/Dialect/Arith/canonicalize.mlir
+++ b/mlir/test/Dialect/Arith/canonicalize.mlir
@@ -753,6 +753,16 @@ func.func @truncExtf3(%arg0: f32) -> f16 {
return %truncf : f16
}
+// CHECK-LABEL: @truncSitofp
+// CHECK-NOT: truncf
+// CHECK: %[[SITOFP:.*]] = arith.sitofp %[[ARG0:.*]] : i32 to f32
+// CHECK: return %[[SITOFP]]
+func.func @truncSitofp(%arg0: i32) -> f32 {
+ %sitofp = arith.sitofp %arg0 : i32 to f64
+ %trunc = arith.truncf %sitofp : f64 to f32
+ return %trunc : f32
+}
+
// TODO: We should also add a test for not folding arith.extf on information loss.
// This may happen when extending f8E5M2FNUZ to f16.
|
|
@llvm/pr-subscribers-mlir Author: Christian Sigg (chsigg) ChangesAdd a canonicalization patterns that simplifies This assumes that the destination type of truncf is representable by the intermediate type. Note that the truncf semantics requires that the destination type is narrower than the source type, so this is true for all types I can possibly think of, but one could probably construct an artificial counter example. Somewhat related: #128096 Full diff: https://github.com/llvm/llvm-project/pull/139925.diff 4 Files Affected:
diff --git a/mlir/include/mlir/Dialect/Arith/IR/ArithOps.td b/mlir/include/mlir/Dialect/Arith/IR/ArithOps.td
index d50b6aeca15c9..599b3b982ec7f 100644
--- a/mlir/include/mlir/Dialect/Arith/IR/ArithOps.td
+++ b/mlir/include/mlir/Dialect/Arith/IR/ArithOps.td
@@ -1273,6 +1273,7 @@ def Arith_TruncFOp :
];
let hasFolder = 1;
+ let hasCanonicalizer = 1;
let hasVerifier = 1;
let assemblyFormat = [{ $in ($roundingmode^)?
(`fastmath` `` $fastmath^)?
diff --git a/mlir/lib/Dialect/Arith/IR/ArithCanonicalization.td b/mlir/lib/Dialect/Arith/IR/ArithCanonicalization.td
index 7e212df9029d1..7be73c4343639 100644
--- a/mlir/lib/Dialect/Arith/IR/ArithCanonicalization.td
+++ b/mlir/lib/Dialect/Arith/IR/ArithCanonicalization.td
@@ -419,6 +419,20 @@ def TruncIShrUIMulIToMulUIExtended :
(ValueWiderThan $mul, $x),
(TruncationMatchesShiftAmount $mul, $x, $c0)]>;
+//===----------------------------------------------------------------------===//
+// TruncIOp
+//===----------------------------------------------------------------------===//
+
+// truncf(sitofp(x)) -> sitofp(x).
+def TruncFSIToFPToSIToFP :
+ Pat<(Arith_TruncFOp:$tr (Arith_SIToFPOp:$fp $x), $rmf, $fmf),
+ (Arith_SIToFPOp $x)>;
+
+// truncf(sitofp(x)) -> sitofp(x).
+def TruncFUIToFPToUIToFP :
+ Pat<(Arith_TruncFOp:$tr (Arith_UIToFPOp:$fp $x), $rmf, $fmf),
+ (Arith_UIToFPOp $x)>;
+
//===----------------------------------------------------------------------===//
// MulFOp
//===----------------------------------------------------------------------===//
diff --git a/mlir/lib/Dialect/Arith/IR/ArithOps.cpp b/mlir/lib/Dialect/Arith/IR/ArithOps.cpp
index 3b308716c84dc..41f2d0f3425e2 100644
--- a/mlir/lib/Dialect/Arith/IR/ArithOps.cpp
+++ b/mlir/lib/Dialect/Arith/IR/ArithOps.cpp
@@ -1552,6 +1552,11 @@ OpFoldResult arith::TruncFOp::fold(FoldAdaptor adaptor) {
});
}
+void arith::TruncFOp::getCanonicalizationPatterns(RewritePatternSet &patterns,
+ MLIRContext *context) {
+ patterns.add<TruncFSIToFPToSIToFP, TruncFUIToFPToUIToFP>(context);
+}
+
bool arith::TruncFOp::areCastCompatible(TypeRange inputs, TypeRange outputs) {
return checkWidthChangeCast<std::less, FloatType>(inputs, outputs);
}
diff --git a/mlir/test/Dialect/Arith/canonicalize.mlir b/mlir/test/Dialect/Arith/canonicalize.mlir
index d62c5b18fd041..a5dab73a62fac 100644
--- a/mlir/test/Dialect/Arith/canonicalize.mlir
+++ b/mlir/test/Dialect/Arith/canonicalize.mlir
@@ -753,6 +753,16 @@ func.func @truncExtf3(%arg0: f32) -> f16 {
return %truncf : f16
}
+// CHECK-LABEL: @truncSitofp
+// CHECK-NOT: truncf
+// CHECK: %[[SITOFP:.*]] = arith.sitofp %[[ARG0:.*]] : i32 to f32
+// CHECK: return %[[SITOFP]]
+func.func @truncSitofp(%arg0: i32) -> f32 {
+ %sitofp = arith.sitofp %arg0 : i32 to f64
+ %trunc = arith.truncf %sitofp : f64 to f32
+ return %trunc : f32
+}
+
// TODO: We should also add a test for not folding arith.extf on information loss.
// This may happen when extending f8E5M2FNUZ to f16.
|
|
So the one thing I'm dubious about here is post- |
|
What is consuming the IR post |
|
Maybe this canonicalization is fine, but if we're in a contest where doesn't exist but , undoing the transformation of the former to the latter would make things more annoying in lowering. Maybe we want to use a fastmath flag on the sitofp to indicate when this transformation is allowed? |
If the pass is a "codegen prepare" pass and is generating non-canonical IR intentional to make the lowering easier, then no it is not reasonable to expect that canonicalization can run there.
That's not a good argument against the canonicalization to me: the lowering must support Something more fundamental to me is rather if we know numerically that or if the rounding can differ when removing the intermediate value? The intermediate is higher precision, but still can it always represent exactly all of the values of the smaller one in all cases? We only need a fast math flag if our contraction changes the numerics. |
Good point, it probably can differ. |
Add a canonicalization pattern that simplifies `truncf(sitofp(x))` to `sitofp(x)`. This assumes that the destination type of truncf is representable by the source type. Note that the truncf semantics requires that the destination type is narrower than the source type, so this is true for all fp types I can possibly think of, but one could probably construct a artificial counter example.
chsigg
force-pushed
the
piper_export_cl_758681496
branch
from
8215575 to
bd58ab7
Compare
Add a canonicalization patterns that simplifies
truncf(sitofp(x))tositofp(x)andtruncf(uitofp(x))touitofp(x).This assumes that the destination type of truncf is representable by the intermediate type.
Note that the truncf semantics requires that the destination type is narrower than the source type, so this is true for all types I can possibly think of, but one could probably construct an artificial counter example.
Somewhat related: #128096