[DAGCombiner] Replace gathers with a zero mask with the passthru value

[bkramer]

Summary

These can be created by the legalizer when splitting a larger gather.
Also make isNullOrNullSplat a bit smarter by ignoring truncation in
BUILD_VECTORs.

[craig.topper]

llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp

8092 ↗

(On Diff #201999)

I wonder if we should just use ISD::isBuildVectorAllZeros here.

llvm/test/CodeGen/AArch64/vecreduce-bool.ll

100 ↗

(On Diff #201999)

I assume this was caused by the isNullOrNullSplat change?

[bkramer]

Use isBuildVectorAllZeros, drop isNullOrNullSplat change.

[bkramer]

llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp

8092 ↗

(On Diff #201999)

That's indeed simpler. Why do we have so many of these things? :(

llvm/test/CodeGen/AArch64/vecreduce-bool.ll

100 ↗

(On Diff #201999)

Yeah. It's a minor improvement. But I dropped this for now.

[craig.topper]

LGTM

llvm/trunk/lib/CodeGen/SelectionDAG/DAGCombiner.cpp

if (Mask.getOpcode() == ISD::SETCC) {
     AddToWorklist(Hi.getNode());
 
     return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Lo, Hi);
   }
   return SDValue();
 }
 
 SDValue DAGCombiner::visitMGATHER(SDNode *N) {
-  if (Level >= AfterLegalizeTypes)
-    return SDValue();
-
   MaskedGatherSDNode *MGT = cast<MaskedGatherSDNode>(N);
   SDValue Mask = MGT->getMask();
   SDLoc DL(N);
 
+  // Zap gathers with a zero mask.
+  if (ISD::isBuildVectorAllZeros(Mask.getNode()))
+    return CombineTo(N, MGT->getPassThru(), MGT->getChain());
+
+  if (Level >= AfterLegalizeTypes)
+    return SDValue();
+
   // If the MGATHER result requires splitting and the mask is provided by a
   // SETCC, then split both nodes and its operands before legalization. This
   // prevents the type legalizer from unrolling SETCC into scalar comparisons
   // and enables future optimizations (e.g. min/max pattern matching on X86).
 
   if (Mask.getOpcode() != ISD::SETCC)
     return SDValue();
 

llvm/trunk/test/CodeGen/X86/avx2-masked-gather.ll

 ; NOGATHER-NEXT:    vmovapd %xmm1, %xmm0
 ; NOGATHER-NEXT:    retq
 entry:
   %ld  = load <2 x double*>, <2 x double*>* %ptr
   %res = call <2 x double> @llvm.masked.gather.v2double(<2 x double*> %ld, i32 0, <2 x i1> %masks, <2 x double> %passthro)
   ret <2 x double> %res
 }
 
+
+define <2 x double> @masked_gather_zeromask(<2 x double*>* %ptr, <2 x double> %dummy, <2 x double> %passthru) {
+; X86-LABEL: masked_gather_zeromask:
+; X86:       # %bb.0: # %entry
+; X86-NEXT:    vmovaps %xmm1, %xmm0
+; X86-NEXT:    retl
+;
+; X64-LABEL: masked_gather_zeromask:
+; X64:       # %bb.0: # %entry
+; X64-NEXT:    vmovaps %xmm1, %xmm0
+; X64-NEXT:    retq
+;
+; NOGATHER-LABEL: masked_gather_zeromask:
+; NOGATHER:       # %bb.0: # %entry
+; NOGATHER-NEXT:    vmovaps %xmm1, %xmm0
+; NOGATHER-NEXT:    retq
+entry:
+  %ld  = load <2 x double*>, <2 x double*>* %ptr
+  %res = call <2 x double> @llvm.masked.gather.v2double(<2 x double*> %ld, i32 0, <2 x i1> zeroinitializer, <2 x double> %passthru)
+  ret <2 x double> %res
+}