[AArch64][GlobalISel] Make vector dup optimization look at last elt of ZeroVec

[paquette]

Summary

Fix an off-by-one error which made us not look at the last element of the zero vector. This caused a miscompile on 188.ammp.

[aemerson]

LGTM.

llvm/trunk/lib/Target/AArch64/AArch64InstructionSelector.cpp

bool AArch64InstructionSelector::tryOptVectorDup(MachineInstr &I) const {
 
   // The shuffle's second operand doesn't matter if the mask is all zero.
   auto *ZeroVec = getOpcodeDef(G_BUILD_VECTOR, I.getOperand(3).getReg(), MRI);
   if (!ZeroVec)
     return false;
   int64_t Zero = 0;
   if (!mi_match(ZeroVec->getOperand(1).getReg(), MRI, m_ICst(Zero)) || Zero)
     return false;
-  for (unsigned i = 1, e = ZeroVec->getNumOperands() - 1; i < e; ++i) {
+  for (unsigned i = 1, e = ZeroVec->getNumOperands(); i < e; ++i) {
     if (ZeroVec->getOperand(i).getReg() != ZeroVec->getOperand(1).getReg())
       return false; // This wasn't an all zeros vector.
   }
 
   // We're done, now find out what kind of splat we need.
   LLT VecTy = MRI.getType(I.getOperand(0).getReg());
   LLT EltTy = VecTy.getElementType();
   if (VecTy.getSizeInBits() != 128 || EltTy.getSizeInBits() < 32) {

llvm/trunk/test/CodeGen/AArch64/GlobalISel/opt-shuffle-splat.mir

 regBankSelected: true
 tracksRegLiveness: true
 body:             |
   bb.1.entry:
     liveins: $d0
 
     ; This test is exactly the same as splat_2xf64, except it adds two copies.
     ; These copies shouldn't get in the way of matching the dup pattern.
+    ; CHECK-LABEL: name: splat_2xf64_copies
+    ; CHECK: liveins: $d0
+    ; CHECK: [[COPY:%[0-9]+]]:fpr64 = COPY $d0
+    ; CHECK: [[DEF:%[0-9]+]]:fpr128 = IMPLICIT_DEF
+    ; CHECK: [[INSERT_SUBREG:%[0-9]+]]:fpr128 = INSERT_SUBREG [[DEF]], [[COPY]], %subreg.dsub
+    ; CHECK: [[DUPv2i64lane:%[0-9]+]]:fpr128 = DUPv2i64lane [[INSERT_SUBREG]], 0
+    ; CHECK: $q0 = COPY [[DUPv2i64lane]]
+    ; CHECK: RET_ReallyLR implicit $q0
     %0:fpr(s64) = COPY $d0
     %2:fpr(<2 x s64>) = G_IMPLICIT_DEF
     %6:fpr(<2 x s64>) = COPY %2
     %3:gpr(s32) = G_CONSTANT i32 0
     %5:fpr(<2 x s32>) = G_BUILD_VECTOR %3(s32), %3(s32)
     %1:fpr(<2 x s64>) = G_INSERT_VECTOR_ELT %6, %0(s64), %3(s32)
     %7:fpr(<2 x s64>) = COPY %1
     %4:fpr(<2 x s64>) = G_SHUFFLE_VECTOR %7(<2 x s64>), %2, %5(<2 x s32>)
     $q0 = COPY %4(<2 x s64>)
     RET_ReallyLR implicit $q0
+
+...
+---
+name:            not_all_zeros
+alignment:       2
+legalized:       true
+regBankSelected: true
+tracksRegLiveness: true
+body:             |
+  bb.1.entry:
+    liveins: $x0
+    ; Make sure that we don't do the optimization when it's not all zeroes.
+    ; CHECK-LABEL: name: not_all_zeros
+    ; CHECK: liveins: $x0
+    ; CHECK: [[COPY:%[0-9]+]]:gpr64 = COPY $x0
+    ; CHECK: [[DEF:%[0-9]+]]:fpr128 = IMPLICIT_DEF
+    ; CHECK: [[INSvi64gpr:%[0-9]+]]:fpr128 = INSvi64gpr [[DEF]], 0, [[COPY]]
+    ; CHECK: [[ADRP:%[0-9]+]]:gpr64common = ADRP target-flags(aarch64-page) %const.0
+    ; CHECK: [[LDRQui:%[0-9]+]]:fpr128 = LDRQui [[ADRP]], target-flags(aarch64-pageoff, aarch64-nc) %const.0
+    ; CHECK: [[REG_SEQUENCE:%[0-9]+]]:qq = REG_SEQUENCE [[INSvi64gpr]], %subreg.qsub0, [[DEF]], %subreg.qsub1
+    ; CHECK: [[TBLv16i8Two:%[0-9]+]]:fpr128 = TBLv16i8Two [[REG_SEQUENCE]], [[LDRQui]]
+    ; CHECK: $q0 = COPY [[TBLv16i8Two]]
+    ; CHECK: RET_ReallyLR implicit $q0
+    %0:gpr(s64) = COPY $x0
+    %2:fpr(<2 x s64>) = G_IMPLICIT_DEF
+    %3:gpr(s32) = G_CONSTANT i32 0
+    %6:gpr(s32) = G_CONSTANT i32 1
+    %5:fpr(<2 x s32>) = G_BUILD_VECTOR %3(s32), %6(s32)
+    %1:fpr(<2 x s64>) = G_INSERT_VECTOR_ELT %2, %0(s64), %3(s32)
+    %4:fpr(<2 x s64>) = G_SHUFFLE_VECTOR %1(<2 x s64>), %2, %5(<2 x s32>)
+    $q0 = COPY %4(<2 x s64>)
+    RET_ReallyLR implicit $q0