[llvm-commits] [llvm] r163814 - /llvm/trunk/utils/TableGen/AsmWriterEmitter.cpp

Manman Ren mren at apple.com
Thu Sep 13 13:36:08 PDT 2012


I should change short to unsigned short or uint16_t. 32768U should be able to fit into 16-bit.

Thanks,
Manman

On Sep 13, 2012, at 1:26 PM, Matt Beaumont-Gay <matthewbg at google.com> wrote:

> Sorry, I should have include a line number. The problem is the OpInfo2
> array, which has type short[].
> 
> On Thu, Sep 13, 2012 at 1:20 PM, Manman Ren <mren at apple.com> wrote:
>> 
>> I will change the unsigned to uint32_t and see whether that will fix the problem.
>> 
>> Thanks,
>> Manman
>> 
>> On Sep 13, 2012, at 11:10 AM, Matt Beaumont-Gay <matthewbg at google.com> wrote:
>> 
>>> Hi Manman,
>>> 
>>> This breaks the tablegen'd headers when building with -std=c++11:
>>> error: constant expression evaluates to 32768 which cannot be narrowed
>>> to type 'short' [-Wc++11-narrowing]
>>>   32768U,     // VFMADDSSr132m
>>>   ^~~~~~
>>> 
>>> On Thu, Sep 13, 2012 at 10:43 AM, Manman Ren <mren at apple.com> wrote:
>>>> Author: mren
>>>> Date: Thu Sep 13 12:43:46 2012
>>>> New Revision: 163814
>>>> 
>>>> URL: http://llvm.org/viewvc/llvm-project?rev=163814&view=rev
>>>> Log:
>>>> AsmWriterEmitter: increase the number of bits for OpcodeInfo from 32-bit to
>>>> 48-bit if necessary, in order to reduce the generated code size.
>>>> 
>>>> We have 900 cases not covered by OpcodeInfo in ATT AsmWriter and more in Intel
>>>> AsmWriter and ARM AsmWriter.
>>>> 
>>>> This patch reduced the clang Release build size by 50k, running on a Mac Pro.
>>>> 
>>>> Modified:
>>>>   llvm/trunk/utils/TableGen/AsmWriterEmitter.cpp
>>>> 
>>>> Modified: llvm/trunk/utils/TableGen/AsmWriterEmitter.cpp
>>>> URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/utils/TableGen/AsmWriterEmitter.cpp?rev=163814&r1=163813&r2=163814&view=diff
>>>> ==============================================================================
>>>> --- llvm/trunk/utils/TableGen/AsmWriterEmitter.cpp (original)
>>>> +++ llvm/trunk/utils/TableGen/AsmWriterEmitter.cpp Thu Sep 13 12:43:46 2012
>>>> @@ -313,7 +313,9 @@
>>>> 
>>>>  /// OpcodeInfo - This encodes the index of the string to use for the first
>>>>  /// chunk of the output as well as indices used for operand printing.
>>>> -  std::vector<unsigned> OpcodeInfo;
>>>> +  /// To reduce the number of unhandled cases, we expand the size from 32-bit
>>>> +  /// to 32+16 = 48-bit.
>>>> +  std::vector<std::pair<unsigned, uint16_t> > OpcodeInfo;
>>>> 
>>>>  // Add all strings to the string table upfront so it can generate an optimized
>>>>  // representation.
>>>> @@ -354,7 +356,7 @@
>>>>    }
>>>> 
>>>>    // Bias offset by one since we want 0 as a sentinel.
>>>> -    OpcodeInfo.push_back(Idx+1);
>>>> +    OpcodeInfo.push_back(std::make_pair(Idx+1, 0));
>>>>  }
>>>> 
>>>>  // Figure out how many bits we used for the string index.
>>>> @@ -362,7 +364,7 @@
>>>> 
>>>>  // To reduce code size, we compactify common instructions into a few bits
>>>>  // in the opcode-indexed table.
>>>> -  unsigned BitsLeft = 32-AsmStrBits;
>>>> +  unsigned BitsLeft = 32+16-AsmStrBits;
>>>> 
>>>>  std::vector<std::vector<std::string> > TableDrivenOperandPrinters;
>>>> 
>>>> @@ -380,6 +382,13 @@
>>>>    // ceil(log2(numentries)).
>>>>    unsigned NumBits = Log2_32_Ceil(UniqueOperandCommands.size());
>>>> 
>>>> +    // Check whether these Bits will fit in the first 32 bits.
>>>> +    if (BitsLeft > 16 && NumBits > BitsLeft - 16)
>>>> +      // We don't have enough bits in the first 32 bits, and we skip the
>>>> +      // left-over bits.
>>>> +      BitsLeft = 16;
>>>> +    bool UseSecond = (BitsLeft <= 16);
>>>> +
>>>>    // If we don't have enough bits for this operand, don't include it.
>>>>    if (NumBits > BitsLeft) {
>>>>      DEBUG(errs() << "Not enough bits to densely encode " << NumBits
>>>> @@ -390,8 +399,13 @@
>>>>    // Otherwise, we can include this in the initial lookup table.  Add it in.
>>>>    BitsLeft -= NumBits;
>>>>    for (unsigned i = 0, e = InstIdxs.size(); i != e; ++i)
>>>> -      if (InstIdxs[i] != ~0U)
>>>> -        OpcodeInfo[i] |= InstIdxs[i] << (BitsLeft+AsmStrBits);
>>>> +      // Update the first 32 bits or the second 16 bits.
>>>> +      if (InstIdxs[i] != ~0U) {
>>>> +        if (UseSecond)
>>>> +          OpcodeInfo[i].second |= InstIdxs[i] << BitsLeft;
>>>> +        else
>>>> +          OpcodeInfo[i].first |= InstIdxs[i] << (BitsLeft-16+AsmStrBits);
>>>> +      }
>>>> 
>>>>    // Remove the info about this operand.
>>>>    for (unsigned i = 0, e = NumberedInstructions.size(); i != e; ++i) {
>>>> @@ -413,13 +427,25 @@
>>>> 
>>>>  O<<"  static const unsigned OpInfo[] = {\n";
>>>>  for (unsigned i = 0, e = NumberedInstructions.size(); i != e; ++i) {
>>>> -    O << "    " << OpcodeInfo[i] << "U,\t// "
>>>> +    O << "    " << OpcodeInfo[i].first << "U,\t// "
>>>>      << NumberedInstructions[i]->TheDef->getName() << "\n";
>>>>  }
>>>>  // Add a dummy entry so the array init doesn't end with a comma.
>>>>  O << "    0U\n";
>>>>  O << "  };\n\n";
>>>> 
>>>> +  if (BitsLeft < 16) {
>>>> +    // Add a second OpInfo table only when it is necessary.
>>>> +    O<<"  static const short OpInfo2[] = {\n";
>>>> +    for (unsigned i = 0, e = NumberedInstructions.size(); i != e; ++i) {
>>>> +      O << "    " << OpcodeInfo[i].second << "U,\t// "
>>>> +        << NumberedInstructions[i]->TheDef->getName() << "\n";
>>>> +    }
>>>> +    // Add a dummy entry so the array init doesn't end with a comma.
>>>> +    O << "    0U\n";
>>>> +    O << "  };\n\n";
>>>> +  }
>>>> +
>>>>  // Emit the string itself.
>>>>  O << "  const char AsmStrs[] = {\n";
>>>>  StringTable.emit(O, printChar);
>>>> @@ -428,12 +454,14 @@
>>>>  O << "  O << \"\\t\";\n\n";
>>>> 
>>>>  O << "  // Emit the opcode for the instruction.\n"
>>>> -    << "  unsigned Bits = OpInfo[MI->getOpcode()];\n"
>>>> -    << "  assert(Bits != 0 && \"Cannot print this instruction.\");\n"
>>>> +    << "  unsigned Bits = OpInfo[MI->getOpcode()];\n";
>>>> +  if (BitsLeft < 16)
>>>> +    O << "  unsigned short Bits2 = OpInfo2[MI->getOpcode()];\n";
>>>> +  O << "  assert(Bits != 0 && \"Cannot print this instruction.\");\n"
>>>>    << "  O << AsmStrs+(Bits & " << (1 << AsmStrBits)-1 << ")-1;\n\n";
>>>> 
>>>>  // Output the table driven operand information.
>>>> -  BitsLeft = 32-AsmStrBits;
>>>> +  BitsLeft = 32+16-AsmStrBits;
>>>>  for (unsigned i = 0, e = TableDrivenOperandPrinters.size(); i != e; ++i) {
>>>>    std::vector<std::string> &Commands = TableDrivenOperandPrinters[i];
>>>> 
>>>> @@ -442,6 +470,11 @@
>>>>    unsigned NumBits = Log2_32_Ceil(Commands.size());
>>>>    assert(NumBits <= BitsLeft && "consistency error");
>>>> 
>>>> +    // Check whether these Bits will fit in the first 32 bits.
>>>> +    if (BitsLeft > 16 && NumBits > BitsLeft - 16)
>>>> +      BitsLeft = 16;
>>>> +    bool UseSecond = (BitsLeft <= 16);
>>>> +
>>>>    // Emit code to extract this field from Bits.
>>>>    BitsLeft -= NumBits;
>>>> 
>>>> @@ -450,7 +483,8 @@
>>>> 
>>>>    if (Commands.size() == 2) {
>>>>      // Emit two possibilitys with if/else.
>>>> -      O << "  if ((Bits >> " << (BitsLeft+AsmStrBits) << ") & "
>>>> +      O << (UseSecond ? "  if ((Bits2 >> " : "  if ((Bits >> ")
>>>> +        << (UseSecond ? BitsLeft : (BitsLeft-16+AsmStrBits)) << ") & "
>>>>        << ((1 << NumBits)-1) << ") {\n"
>>>>        << Commands[1]
>>>>        << "  } else {\n"
>>>> @@ -460,7 +494,8 @@
>>>>      // Emit a single possibility.
>>>>      O << Commands[0] << "\n\n";
>>>>    } else {
>>>> -      O << "  switch ((Bits >> " << (BitsLeft+AsmStrBits) << ") & "
>>>> +      O << (UseSecond ? "  switch ((Bits2 >> " : "  switch ((Bits >> ")
>>>> +        << (UseSecond ? BitsLeft : (BitsLeft-16+AsmStrBits)) << ") & "
>>>>        << ((1 << NumBits)-1) << ") {\n"
>>>>        << "  default:   // unreachable.\n";
>>>> 
>>>> 
>>>> 
>>>> _______________________________________________
>>>> llvm-commits mailing list
>>>> llvm-commits at cs.uiuc.edu
>>>> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
>> 




More information about the llvm-commits mailing list