From 0ea26ca45b56b2b4cc1839cdbccc13a38b066969 Mon Sep 17 00:00:00 2001 From: Chris Lattner Date: Wed, 20 Sep 2006 03:38:48 +0000 Subject: [PATCH] Expand 64-bit shifts more optimally if we know that the high bit of the shift amount is one or zero. For example, for: long long foo1(long long X, int C) { return X << (C|32); } long long foo2(long long X, int C) { return X << (C&~32); } we get: _foo1: movb $31, %cl movl 4(%esp), %edx andb 12(%esp), %cl shll %cl, %edx xorl %eax, %eax ret _foo2: movb $223, %cl movl 4(%esp), %eax movl 8(%esp), %edx andb 12(%esp), %cl shldl %cl, %eax, %edx shll %cl, %eax ret instead of: _foo1: subl $4, %esp movl %ebx, (%esp) movb $32, %bl movl 8(%esp), %eax movl 12(%esp), %edx movb %bl, %cl orb 16(%esp), %cl shldl %cl, %eax, %edx shll %cl, %eax xorl %ecx, %ecx testb %bl, %bl cmovne %eax, %edx cmovne %ecx, %eax movl (%esp), %ebx addl $4, %esp ret _foo2: subl $4, %esp movl %ebx, (%esp) movb $223, %cl movl 8(%esp), %eax movl 12(%esp), %edx andb 16(%esp), %cl shldl %cl, %eax, %edx shll %cl, %eax xorl %ecx, %ecx xorb %bl, %bl testb %bl, %bl cmovne %eax, %edx cmovne %ecx, %eax movl (%esp), %ebx addl $4, %esp ret git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30506 91177308-0d34-0410-b5e6-96231b3b80d8 --- lib/CodeGen/SelectionDAG/LegalizeDAG.cpp | 66 ++++++++++++++++++++++++ 1 file changed, 66 insertions(+) diff --git a/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp b/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp index aa25fd82956..cb8588c0318 100644 --- a/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp +++ b/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp @@ -3781,6 +3781,72 @@ bool SelectionDAGLegalize::ExpandShift(unsigned Opc, SDOperand Op,SDOperand Amt, return true; } } + + // Okay, the shift amount isn't constant. However, if we can tell that it is + // >= 32 or < 32, we can still simplify it, without knowing the actual value. + uint64_t Mask = NVTBits, KnownZero, KnownOne; + TLI.ComputeMaskedBits(Amt, Mask, KnownZero, KnownOne); + + // If we know that the high bit of the shift amount is one, then we can do + // this as a couple of simple shifts. + if (KnownOne & Mask) { + // Mask out the high bit, which we know is set. + Amt = DAG.getNode(ISD::AND, Amt.getValueType(), Amt, + DAG.getConstant(NVTBits-1, Amt.getValueType())); + + // Expand the incoming operand to be shifted, so that we have its parts + SDOperand InL, InH; + ExpandOp(Op, InL, InH); + switch(Opc) { + case ISD::SHL: + Lo = DAG.getConstant(0, NVT); // Low part is zero. + Hi = DAG.getNode(ISD::SHL, NVT, InL, Amt); // High part from Lo part. + return true; + case ISD::SRL: + Hi = DAG.getConstant(0, NVT); // Hi part is zero. + Lo = DAG.getNode(ISD::SRL, NVT, InH, Amt); // Lo part from Hi part. + return true; + case ISD::SRA: + Hi = DAG.getNode(ISD::SRA, NVT, InH, // Sign extend high part. + DAG.getConstant(NVTBits-1, Amt.getValueType())); + Lo = DAG.getNode(ISD::SRA, NVT, InH, Amt); // Lo part from Hi part. + return true; + } + } + + // If we know that the high bit of the shift amount is zero, then we can do + // this as a couple of simple shifts. + if (KnownZero & Mask) { + // Compute 32-amt. + SDOperand Amt2 = DAG.getNode(ISD::SUB, Amt.getValueType(), + DAG.getConstant(NVTBits, Amt.getValueType()), + Amt); + + // Expand the incoming operand to be shifted, so that we have its parts + SDOperand InL, InH; + ExpandOp(Op, InL, InH); + switch(Opc) { + case ISD::SHL: + Lo = DAG.getNode(ISD::SHL, NVT, InL, Amt); + Hi = DAG.getNode(ISD::OR, NVT, + DAG.getNode(ISD::SHL, NVT, InH, Amt), + DAG.getNode(ISD::SRL, NVT, InL, Amt2)); + return true; + case ISD::SRL: + Hi = DAG.getNode(ISD::SRL, NVT, InH, Amt); + Lo = DAG.getNode(ISD::OR, NVT, + DAG.getNode(ISD::SRL, NVT, InL, Amt), + DAG.getNode(ISD::SHL, NVT, InH, Amt2)); + return true; + case ISD::SRA: + Hi = DAG.getNode(ISD::SRA, NVT, InH, Amt); + Lo = DAG.getNode(ISD::OR, NVT, + DAG.getNode(ISD::SRL, NVT, InL, Amt), + DAG.getNode(ISD::SHL, NVT, InH, Amt2)); + return true; + } + } + return false; } -- 2.34.1