[PATCH] D28622: DAGCombiner: Allow negating ConstantFP after legalize

[Matt Arsenault via Phabricator via llvm-commits]

arsenm created this revision.
arsenm added a subscriber: llvm-commits.
Herald added a reviewer: tstellarAMD.
Herald added subscribers: nhaehnle, wdng.

https://reviews.llvm.org/D28622

Files:
  include/llvm/ADT/APFloat.h
  lib/CodeGen/SelectionDAG/DAGCombiner.cpp
  test/CodeGen/AMDGPU/fneg-combines.ll
  unittests/ADT/APFloatTest.cpp


Index: unittests/ADT/APFloatTest.cpp
===================================================================
--- unittests/ADT/APFloatTest.cpp
+++ unittests/ADT/APFloatTest.cpp
@@ -2829,6 +2829,28 @@
   EXPECT_TRUE(PSmallestNormalized.bitwiseIsEqual(abs(MSmallestNormalized)));
 }
 
+TEST(APFloatTest, neg) {
+  APFloat One = APFloat(APFloat::IEEEsingle(), "1.0");
+  APFloat NegOne = APFloat(APFloat::IEEEsingle(), "-1.0");
+  APFloat Zero = APFloat::getZero(APFloat::IEEEsingle(), false);
+  APFloat NegZero = APFloat::getZero(APFloat::IEEEsingle(), true);
+  APFloat Inf = APFloat::getInf(APFloat::IEEEsingle(), false);
+  APFloat NegInf = APFloat::getInf(APFloat::IEEEsingle(), true);
+  APFloat QNaN = APFloat::getNaN(APFloat::IEEEsingle(), false);
+  APFloat NegQNaN = APFloat::getNaN(APFloat::IEEEsingle(), true);
+
+  EXPECT_TRUE(NegOne.bitwiseIsEqual(neg(One)));
+  EXPECT_TRUE(One.bitwiseIsEqual(neg(NegOne)));
+  EXPECT_TRUE(NegZero.bitwiseIsEqual(neg(Zero)));
+  EXPECT_TRUE(Zero.bitwiseIsEqual(neg(NegZero)));
+  EXPECT_TRUE(NegInf.bitwiseIsEqual(neg(Inf)));
+  EXPECT_TRUE(Inf.bitwiseIsEqual(neg(NegInf)));
+  EXPECT_TRUE(NegInf.bitwiseIsEqual(neg(Inf)));
+  EXPECT_TRUE(Inf.bitwiseIsEqual(neg(NegInf)));
+  EXPECT_TRUE(NegQNaN.bitwiseIsEqual(neg(QNaN)));
+  EXPECT_TRUE(QNaN.bitwiseIsEqual(neg(NegQNaN)));
+}
+
 TEST(APFloatTest, ilogb) {
   EXPECT_EQ(-1074, ilogb(APFloat::getSmallest(APFloat::IEEEdouble(), false)));
   EXPECT_EQ(-1074, ilogb(APFloat::getSmallest(APFloat::IEEEdouble(), true)));
Index: test/CodeGen/AMDGPU/fneg-combines.ll
===================================================================
--- test/CodeGen/AMDGPU/fneg-combines.ll
+++ test/CodeGen/AMDGPU/fneg-combines.ll
@@ -1235,8 +1235,7 @@
 
 ; GCN-LABEL: {{^}}v_fneg_sin_f32:
 ; GCN: {{buffer|flat}}_load_dword [[A:v[0-9]+]]
-; GCN: v_mov_b32_e32 [[K:v[0-9]+]], 0x3e22f983
-; GCN: v_mul_f32_e64 [[MUL:v[0-9]+]], [[K]], -[[A]]
+; GCN: v_mul_f32_e32 [[MUL:v[0-9]+]], 0xbe22f983, [[A]]
 ; GCN: v_fract_f32_e32 [[FRACT:v[0-9]+]], [[MUL]]
 ; GCN: v_sin_f32_e32 [[RESULT:v[0-9]+]], [[FRACT]]
 ; GCN: buffer_store_dword [[RESULT]]
Index: lib/CodeGen/SelectionDAG/DAGCombiner.cpp
===================================================================
--- lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -607,10 +607,16 @@
 
   switch (Op.getOpcode()) {
   default: return false;
-  case ISD::ConstantFP:
-    // Don't invert constant FP values after legalize.  The negated constant
-    // isn't necessarily legal.
-    return LegalOperations ? 0 : 1;
+  case ISD::ConstantFP: {
+    if (!LegalOperations)
+      return 1;
+
+    // Don't invert constant FP values after legalize, unless the target the
+    // negated constant is necessarily legal.
+    EVT VT = Op.getValueType();
+    return TLI.isOperationLegal(ISD::ConstantFP, VT) ||
+      TLI.isFPImmLegal(neg(cast<ConstantFPSDNode>(Op)->getValueAPF()), VT);
+  }
   case ISD::FADD:
     // FIXME: determine better conditions for this xform.
     if (!Options->UnsafeFPMath) return 0;
Index: include/llvm/ADT/APFloat.h
===================================================================
--- include/llvm/ADT/APFloat.h
+++ include/llvm/ADT/APFloat.h
@@ -1101,6 +1101,12 @@
   return X;
 }
 
+/// \brief Returns the negated value of the argument.
+inline APFloat neg(APFloat X) {
+  X.changeSign();
+  return X;
+}
+
 /// Implements IEEE minNum semantics. Returns the smaller of the 2 arguments if
 /// both are not NaN. If either argument is a NaN, returns the other argument.
 LLVM_READONLY


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D28622.84165.patch
Type: text/x-patch
Size: 3565 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20170112/1d9ff4cd/attachment.bin>