Add a target lowering hook to control whether it's worthwhile to compress fp constant.

For x86, if sse2 is available, it's not a good idea since cvtss2sd is slower than a movsd load and it prevents load folding. On x87, it's important to shrink fp constant since fldt is very expensive.

llvm-svn: 47931

include/llvm/Target/TargetLowering.h

@@ -467,6 +467,11 @@ public:
     assert(0 && "Unsupported extended type!");
   }
 
+  /// ShouldShrinkFPConstant - If true, then instruction selection should
+  /// seek to shrink the FP constant of the specified type to a smaller type
+  /// in order to save space and / or reduce runtime.
+  virtual bool ShouldShrinkFPConstant(MVT::ValueType VT) const { return true; }
+
   /// hasTargetDAGCombine - If true, the target has custom DAG combine
   /// transformations that it can perform for the specified node.
   bool hasTargetDAGCombine(ISD::NodeType NT) const {

lib/CodeGen/SelectionDAG/LegalizeDAG.cpp

@@ -504,7 +504,8 @@ static SDOperand ExpandConstantFP(ConstantFPSDNode *CFP, bool UseCP,
     if (CFP->isValueValidForType(SVT, CFP->getValueAPF()) &&
         // Only do this if the target has a native EXTLOAD instruction from
         // smaller type.
-        TLI.isLoadXLegal(ISD::EXTLOAD, SVT)) {
+        TLI.isLoadXLegal(ISD::EXTLOAD, SVT) &&
+        TLI.ShouldShrinkFPConstant(VT)) {
       const Type *SType = MVT::getTypeForValueType(SVT);
       LLVMC = cast<ConstantFP>(ConstantExpr::getFPTrunc(LLVMC, SType));
       VT = SVT;

lib/Target/X86/X86ISelLowering.h

@@ -434,6 +434,16 @@ namespace llvm {
     virtual bool isVectorClearMaskLegal(std::vector<SDOperand> &BVOps,
                                         MVT::ValueType EVT,
                                         SelectionDAG &DAG) const;
+
+    /// ShouldShrinkFPConstant - If true, then instruction selection should
+    /// seek to shrink the FP constant of the specified type to a smaller type
+    /// in order to save space and / or reduce runtime.
+    virtual bool ShouldShrinkFPConstant(MVT::ValueType VT) const {
+      // Don't shrink FP constpool if SSE2 is available since cvtss2sd is more
+      // expensive than a straight movsd. On the other hand, it's important to
+      // shrink long double fp constant since fldt is very slow.
+      return !X86ScalarSSEf64 || VT == MVT::f80;
+    }
     
     /// IsEligibleForTailCallOptimization - Check whether the call is eligible
     /// for tail call optimization. Target which want to do tail call

test/CodeGen/X86/shrink-fp-const1.ll

@@ -0,0 +1,7 @@
+; RUN: llvm-as < %s | llc -march=x86-64 -mattr=+sse2 | not grep cvtss2sd
+; PR1264
+
+define double @foo(double %x) {
+        %y = mul double %x, 5.000000e-01
+        ret double %y
+}