[llvm] 377b0e2 - [TableGen] Don't elide bitconverts in PatFrag fragments.

Simon Tatham via llvm-commits llvm-commits at lists.llvm.org
Mon Feb 17 01:31:00 PST 2020


Author: Simon Tatham
Date: 2020-02-17T09:30:45Z
New Revision: 377b0e2b06f90c4edbd320cf9d833af5979f8f5d

URL: https://github.com/llvm/llvm-project/commit/377b0e2b06f90c4edbd320cf9d833af5979f8f5d
DIFF: https://github.com/llvm/llvm-project/commit/377b0e2b06f90c4edbd320cf9d833af5979f8f5d.diff

LOG: [TableGen] Don't elide bitconverts in PatFrag fragments.

Summary:
In the DAG pattern backend, `SimplifyTree` simplifies a pattern by
removing bitconverts between two identical types. But that function is
also run on the fragments list in instances of `PatFrags`, in which
the types haven't been specified yet. So the input and output of the
bitconvert always evaluate to the empty set of types, which makes them
compare equal. So the test always passes, and bitconverts are
unconditionally removed from the PatFrag RHS.

Fixed by spotting the empty type set and using it to inhibit the
optimization.

Reviewers: nhaehnle, hfinkel

Reviewed By: nhaehnle

Subscribers: llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D74627

Added: 
    llvm/test/TableGen/simplify-patfrag.td

Modified: 
    llvm/utils/TableGen/CodeGenDAGPatterns.cpp

Removed: 
    


################################################################################
diff  --git a/llvm/test/TableGen/simplify-patfrag.td b/llvm/test/TableGen/simplify-patfrag.td
new file mode 100644
index 000000000000..693658317d5d
--- /dev/null
+++ b/llvm/test/TableGen/simplify-patfrag.td
@@ -0,0 +1,38 @@
+// RUN: llvm-tblgen -gen-dag-isel -I %p/../../include %s 2>&1 | FileCheck %s
+
+include "llvm/Target/Target.td"
+
+// Minimal Target definition
+def DemoInstrInfo : InstrInfo;
+def Demo : Target {
+  let InstructionSet = DemoInstrInfo;
+}
+
+// Some registers which can hold ints or floats
+foreach i = 0-7 in
+  def "R" # i: Register<"r" # i>;
+def GPR : RegisterClass<"Demo", [i32, f32], 32, (sequence "R%u", 0, 7)>;
+
+// Instruction to convert an int to a float
+def i2f : Instruction {
+  let Size = 2;
+  let OutOperandList = (outs GPR:$dst);
+  let InOperandList = (ins GPR:$src);
+  let AsmString = "i2f $dst, $src";
+}
+
+// Some kind of special type-conversion node supported by this target
+def specialconvert : SDNode<"TEST_TARGET_ISD::SPECIAL_CONVERT", SDTUnaryOp>;
+
+// A PatFrags that matches either bitconvert or the special version
+def anyconvert : PatFrags<(ops node:$src),
+                          [(bitconvert node:$src),
+                           (specialconvert node:$src)]>;
+
+// And a rule that matches that PatFrag and turns it into i2f
+def : Pat<(f32 (anyconvert (i32 GPR:$val))), (i2f GPR:$val)>;
+
+// CHECK: SwitchOpcode{{.*}}ISD::BITCAST
+// CHECK: MorphNodeTo1{{.*}}i2f
+// CHECK: SwitchOpcode{{.*}}TEST_TARGET_ISD::SPECIAL_CONVERT
+// CHECK: MorphNodeTo1{{.*}}i2f

diff  --git a/llvm/utils/TableGen/CodeGenDAGPatterns.cpp b/llvm/utils/TableGen/CodeGenDAGPatterns.cpp
index 703cd515a071..0ea90d7dbb65 100644
--- a/llvm/utils/TableGen/CodeGenDAGPatterns.cpp
+++ b/llvm/utils/TableGen/CodeGenDAGPatterns.cpp
@@ -2918,8 +2918,15 @@ static bool SimplifyTree(TreePatternNodePtr &N) {
 
   // If we have a bitconvert with a resolved type and if the source and
   // destination types are the same, then the bitconvert is useless, remove it.
+  //
+  // We make an exception if the types are completely empty. This can come up
+  // when the pattern being simplified is in the Fragments list of a PatFrags,
+  // so that the operand is just an untyped "node". In that situation we leave
+  // bitconverts unsimplified, and simplify them later once the fragment is
+  // expanded into its true context.
   if (N->getOperator()->getName() == "bitconvert" &&
       N->getExtType(0).isValueTypeByHwMode(false) &&
+      !N->getExtType(0).empty() &&
       N->getExtType(0) == N->getChild(0)->getExtType(0) &&
       N->getName().empty()) {
     N = N->getChildShared(0);


        


More information about the llvm-commits mailing list