diff options
| -rw-r--r-- | include/llvm/Target/TargetLowering.h | 10 | ||||
| -rw-r--r-- | lib/Target/ARM/ARMISelLowering.h | 5 | ||||
| -rw-r--r-- | lib/Transforms/Scalar/CodeGenPrepare.cpp | 29 | ||||
| -rw-r--r-- | test/CodeGen/ARM/2012-08-30-select.ll | 5 |
4 files changed, 39 insertions, 10 deletions
diff --git a/include/llvm/Target/TargetLowering.h b/include/llvm/Target/TargetLowering.h index acf0419510..b8c070bce7 100644 --- a/include/llvm/Target/TargetLowering.h +++ b/include/llvm/Target/TargetLowering.h @@ -107,6 +107,14 @@ public: ZeroOrNegativeOneBooleanContent // All bits equal to bit 0. }; + enum SelectSupportKind { + ScalarValSelect, // The target supports scalar selects (ex: cmov). + ScalarCondVectorVal, // The target supports selects with a scalar condition + // and vector values (ex: cmov). + VectorMaskSelect // The target supports vector selects with a vector + // mask (ex: x86 blends). + }; + static ISD::NodeType getExtendForContent(BooleanContent Content) { switch (Content) { case UndefinedBooleanContent: @@ -140,6 +148,8 @@ public: /// this target. bool isSelectExpensive() const { return SelectIsExpensive; } + virtual bool isSelectSupported(SelectSupportKind kind) const { return true; } + /// isIntDivCheap() - Return true if integer divide is usually cheaper than /// a sequence of several shifts, adds, and multiplies for this target. bool isIntDivCheap() const { return IntDivIsCheap; } diff --git a/lib/Target/ARM/ARMISelLowering.h b/lib/Target/ARM/ARMISelLowering.h index 13b83de85d..757c68ee37 100644 --- a/lib/Target/ARM/ARMISelLowering.h +++ b/lib/Target/ARM/ARMISelLowering.h @@ -257,6 +257,11 @@ namespace llvm { virtual const char *getTargetNodeName(unsigned Opcode) const; + virtual bool isSelectSupported(SelectSupportKind Kind) const { + // ARM does not support scalar condition selects on vectors. + return (Kind != ScalarCondVectorVal); + } + /// getSetCCResultType - Return the value type to use for ISD::SETCC. virtual EVT getSetCCResultType(EVT VT) const; diff --git a/lib/Transforms/Scalar/CodeGenPrepare.cpp b/lib/Transforms/Scalar/CodeGenPrepare.cpp index cfca726f26..57a648f7b9 100644 --- a/lib/Transforms/Scalar/CodeGenPrepare.cpp +++ b/lib/Transforms/Scalar/CodeGenPrepare.cpp @@ -1174,17 +1174,32 @@ static bool isFormingBranchFromSelectProfitable(SelectInst *SI) { } +/// If we have a SelectInst that will likely profit from branch prediction, +/// turn it into a branch. bool CodeGenPrepare::OptimizeSelectInst(SelectInst *SI) { - // If we have a SelectInst that will likely profit from branch prediction, - // turn it into a branch. - if (DisableSelectToBranch || OptSize || !TLI || - !TLI->isPredictableSelectExpensive()) - return false; + bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(1); - if (!SI->getCondition()->getType()->isIntegerTy(1) || - !isFormingBranchFromSelectProfitable(SI)) + // Can we convert the 'select' to CF ? + if (DisableSelectToBranch || OptSize || !TLI || VectorCond) return false; + TargetLowering::SelectSupportKind SelectKind; + if (VectorCond) + SelectKind = TargetLowering::VectorMaskSelect; + else if (SI->getType()->isVectorTy()) + SelectKind = TargetLowering::ScalarCondVectorVal; + else + SelectKind = TargetLowering::ScalarValSelect; + + // Do we have efficient codegen support for this kind of 'selects' ? + if (TLI->isSelectSupported(SelectKind)) { + // We have efficient codegen support for the select instruction. + // Check if it is profitable to keep this 'select'. + if (!TLI->isPredictableSelectExpensive() || + !isFormingBranchFromSelectProfitable(SI)) + return false; + } + ModifiedDT = true; // First, we split the block containing the select into 2 blocks. diff --git a/test/CodeGen/ARM/2012-08-30-select.ll b/test/CodeGen/ARM/2012-08-30-select.ll index 4c94ac9dbf..8471be5330 100644 --- a/test/CodeGen/ARM/2012-08-30-select.ll +++ b/test/CodeGen/ARM/2012-08-30-select.ll @@ -2,9 +2,8 @@ ; rdar://12201387 ;CHECK: select_s_v_v -;CHECK: it eq -;CHECK-NEXT: moveq.w r2, #-1 -;CHECK: vbsl +;CHECK: it ne +;CHECK-NEXT: vmovne.i32 ;CHECK: bx define <16 x i8> @select_s_v_v(i32 %avail, i8* %bar) { entry: |