[X86] Add demanded elts support for the inputs to pclmul intrinsic

[craig.topper]

Summary

This intrinsic uses bit 0 and bit 4 of an immediate argument to determine which bits of its inputs to read. This patch uses this information to simplify the demanded elements of the input vectors.

[RKSimon]

lib/Transforms/InstCombine/InstCombineCalls.cpp

2197 ↗

(On Diff #85239)

Add support for detecting all undef with UndefElts?

[craig.topper]

Optimize to undef if both inputs are undef.

[RKSimon]

In D28979#652966, @craig.topper wrote:

Optimize to undef if both inputs are undef.

Should we be optimizing to undef or zero? I don't have an answer but I just got stung for having written muldq/muludq optimizing to undef as generic multiplies become zero. Does clmul count as a multiply?

[craig.topper]

Convert to zero if either input is undef.

[RKSimon]

LGTM - thanks.

llvm/trunk/lib/Transforms/InstCombine/InstCombineCalls.cpp

Instruction *InstCombiner::visitCallInst(CallInst &CI) {
   case Intrinsic::x86_sse41_packusdw:
   case Intrinsic::x86_avx2_packusdw:
   case Intrinsic::x86_avx2_packuswb:
   // TODO Add support for Intrinsic::x86_avx512_mask_packus*
     if (Value *V = simplifyX86pack(*II, *this, *Builder, false))
       return replaceInstUsesWith(*II, V);
     break;
 
+  case Intrinsic::x86_pclmulqdq: {
+    if (auto *C = dyn_cast<ConstantInt>(II->getArgOperand(2))) {
+      unsigned Imm = C->getZExtValue();
+
+      bool MadeChange = false;
+      Value *Arg0 = II->getArgOperand(0);
+      Value *Arg1 = II->getArgOperand(1);
+      unsigned VWidth = Arg0->getType()->getVectorNumElements();
+      APInt DemandedElts(VWidth, 0);
+
+      APInt UndefElts1(VWidth, 0);
+      DemandedElts = (Imm & 0x01) ? 2 : 1;
+      if (Value *V = SimplifyDemandedVectorElts(Arg0, DemandedElts,
+                                                UndefElts1)) {
+        II->setArgOperand(0, V);
+        MadeChange = true;
+      }
+
+      APInt UndefElts2(VWidth, 0);
+      DemandedElts = (Imm & 0x10) ? 2 : 1;
+      if (Value *V = SimplifyDemandedVectorElts(Arg1, DemandedElts,
+                                                UndefElts2)) {
+        II->setArgOperand(1, V);
+        MadeChange = true;
+      }
+
+      // If both input elements are undef, the result is undef.
+      if (UndefElts1[(Imm & 0x01) ? 1 : 0] ||
+          UndefElts2[(Imm & 0x10) ? 1 : 0])
+        return replaceInstUsesWith(*II,
+                                   ConstantAggregateZero::get(II->getType()));
+
+      if (MadeChange)
+        return II;
+    }
+    break;
+  }
+
   case Intrinsic::x86_sse41_insertps:
     if (Value *V = simplifyX86insertps(*II, *Builder))
       return replaceInstUsesWith(*II, V);
     break;
 
   case Intrinsic::x86_sse4a_extrq: {
     Value *Op0 = II->getArgOperand(0);
     Value *Op1 = II->getArgOperand(1);