[AsmPrinter] Fix Crash when Emitting Global Constant of small bit width when targeting Big Endian arch

[HazyFish]

Summary

For Big Endian, the function emitGlobalConstantLargeInt tries to right shift Realigned by an amount ExtraBitSize in place. However, if the constant to emit has a bit width less than 64 and the bit width is not a multiple of 8, the shift amount will be greater than the bit width of Realigned, which causes assertion error described in issue issue #59055.

This patch fixes the issue by avoiding right shift when bit width is under 64 to avoid the assertion error.

[efriedma]

llvm/lib/CodeGen/AsmPrinter/AsmPrinter.cpp

3273

Please upload patches with context (use "git diff -U100000", or use Arcanist).

Maybe it's clearer to just write:

if (BitWidth >= 64)
  Realigned.lshrInPlace(ExtraBitsSize);

llvm/test/CodeGen/AArch64/aarch64_be-global-const.ll

7

If you can't make update_llc_test_checks.py generate something appropriate, please write an appropriate CHECK line by hand.

Maybe it makes sense to just add a RUN line to an existing test, instead of adding a separate test file?

[HazyFish]

Address feedback

[HazyFish]

Align with upstream

[Peter]

arc land

llvm/lib/CodeGen/AsmPrinter/AsmPrinter.cpp

static void emitGlobalConstantVector(const DataLayout &DL,
   } else {
     for (unsigned I = 0, E = CV->getType()->getNumElements(); I != E; ++I) {
       emitGlobalAliasInline(AP, DL.getTypeAllocSize(CV->getType()) * I, AliasList);
       emitGlobalConstantImpl(DL, CV->getOperand(I), AP);
     }
     EmittedSize =
         DL.getTypeAllocSize(ElementType) * CV->getType()->getNumElements();
   }
 
   unsigned Size = DL.getTypeAllocSize(CV->getType());
   if (unsigned Padding = Size - EmittedSize)
     AP.OutStreamer->emitZeros(Padding);
 }
 
 static void emitGlobalConstantStruct(const DataLayout &DL,
                                      const ConstantStruct *CS, AsmPrinter &AP,
                                      const Constant *BaseCV, uint64_t Offset,
if (DL.isBigEndian()) {
       // The most significant chunk is chunkN and it should be emitted first.
       // However, due to the alignment issue chunkN contains useless bits.
       // Realign the chunks so that they contain only useful information:
       // ExtraBits     0       1       (BitWidth / 64) - 1
       //       chu[nk1 chu][nk2 chu] ... [nkN-1 chunkN]
       ExtraBitsSize = alignTo(ExtraBitsSize, 8);
       ExtraBits = Realigned.getRawData()[0] &
         (((uint64_t)-1) >> (64 - ExtraBitsSize));
+      if (BitWidth >= 64)
         Realigned.lshrInPlace(ExtraBitsSize);
     } else
       ExtraBits = Realigned.getRawData()[BitWidth / 64];
   }
 
   // We don't expect assemblers to support integer data directives
   // for more than 64 bits, so we emit the data in at most 64-bit
   // quantities at a time.
   const uint64_t *RawData = Realigned.getRawData();

llvm/test/CodeGen/AArch64/aarch64_be-global-const.ll

+; RUN: llc -mtriple=aarch64_be < %s | FileCheck %s
+
+; CHECK-LABEL: G:
+; CHECK: .byte 11
+; CHECK: .size G, 1
+@G = global <4 x i1> <i1 true, i1 false, i1 true, i1 true>