[DAGCombiner] combine vextract (v1iX extract_subvector(vNiX, Idx)) into vextract(vNiX,Idx).

[eladcohen]

Summary

Combine vextract (v1iX extract_subvector(vNiX, Idx)) into vextract(vNiX,Idx).
This case appeared in AVX512 after fixing pr33349 in r310552.

Usually we don't get these cases since v1iX are illegal in most targets and then these
(v1iX extract_subvector(vNiX, Idx)) get scalarized into (iX vextract(vNiX,Idx)).
But for targets like AVX512 where such type is legal we end up with a redundant DAG
node which can be combined.

[RKSimon]

lib/CodeGen/SelectionDAG/DAGCombiner.cpp

13790

Would this be better in SelectionDAG::getNode() ?

[eladcohen]

lib/CodeGen/SelectionDAG/DAGCombiner.cpp

13790

Sounds good. I'll move it.

[eladcohen]

Fixed Simon's comments.

[RKSimon]

LGTM with a couple of minors

lib/CodeGen/SelectionDAG/SelectionDAG.cpp

4378 ↗

(On Diff #110876)

vector

4381 ↗

(On Diff #110876)

Move this up to the rest of the comment

[eladcohen]

Fixed Simon's comments. Thanks for the review!

lib/CodeGen/SelectionDAG/DAGCombiner.cpp

if (ConstEltNo &&
     // Sometimes build_vector's scalar input types do not match result type.
     if (NVT == InEltVT)
       return Elt;
 
     // TODO: It may be useful to truncate if free if the build_vector implicitly
     // converts.
   }
 
+  // vextract (v1iX extract_subvector(vNiX, Idx)) -> vextract(vNiX,Idx)
+  if (InVec.getOpcode() == ISD::EXTRACT_SUBVECTOR &&
+      VT.getVectorNumElements() == 1) {
+    return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SDLoc(N), NVT,
+                       InVec->getOperand(0), InVec->getOperand(1));
+  }
+
   // extract_vector_elt (v2i32 (bitcast i64:x)), EltTrunc -> i32 (trunc i64:x)
   bool isLE = DAG.getDataLayout().isLittleEndian();
   unsigned EltTrunc = isLE ? 0 : VT.getVectorNumElements() - 1;
   if (ConstEltNo && InVec.getOpcode() == ISD::BITCAST && InVec.hasOneUse() &&
       ConstEltNo->getZExtValue() == EltTrunc && VT.isInteger()) {
     SDValue BCSrc = InVec.getOperand(0);
     if (BCSrc.getValueType().isScalarInteger())
       return DAG.getNode(ISD::TRUNCATE, SDLoc(N), NVT, BCSrc);

test/CodeGen/X86/pr33349.ll

 ;
 ; SKX-LABEL: test:
 ; SKX:       # BB#0: # %bb
 ; SKX-NEXT:    vpslld $31, %xmm0, %xmm0
 ; SKX-NEXT:    vptestmd %xmm0, %xmm0, %k0
 ; SKX-NEXT:    kshiftrw $2, %k0, %k1
 ; SKX-NEXT:    kshiftlw $15, %k1, %k2
 ; SKX-NEXT:    kshiftrw $15, %k2, %k2
-; SKX-NEXT:    kshiftlw $15, %k2, %k2
-; SKX-NEXT:    kshiftrw $15, %k2, %k2
 ; SKX-NEXT:    kmovd %k2, %eax
 ; SKX-NEXT:    testb $1, %al
 ; SKX-NEXT:    fld1
 ; SKX-NEXT:    fldz
 ; SKX-NEXT:    fld %st(0)
 ; SKX-NEXT:    fcmovne %st(2), %st(0)
 ; SKX-NEXT:    kshiftlw $14, %k1, %k1
 ; SKX-NEXT:    kshiftrw $15, %k1, %k1
-; SKX-NEXT:    kshiftlw $15, %k1, %k1
-; SKX-NEXT:    kshiftrw $15, %k1, %k1
 ; SKX-NEXT:    kmovd %k1, %eax
 ; SKX-NEXT:    testb $1, %al
 ; SKX-NEXT:    fld %st(1)
 ; SKX-NEXT:    fcmovne %st(3), %st(0)
 ; SKX-NEXT:    kshiftlw $15, %k0, %k1
 ; SKX-NEXT:    kshiftrw $15, %k1, %k1
-; SKX-NEXT:    kshiftlw $15, %k1, %k1
-; SKX-NEXT:    kshiftrw $15, %k1, %k1
 ; SKX-NEXT:    kmovd %k1, %eax
 ; SKX-NEXT:    testb $1, %al
 ; SKX-NEXT:    fld %st(2)
 ; SKX-NEXT:    fcmovne %st(4), %st(0)
 ; SKX-NEXT:    kshiftlw $14, %k0, %k0
 ; SKX-NEXT:    kshiftrw $15, %k0, %k0
-; SKX-NEXT:    kshiftlw $15, %k0, %k0
-; SKX-NEXT:    kshiftrw $15, %k0, %k0
 ; SKX-NEXT:    kmovd %k0, %eax
 ; SKX-NEXT:    testb $1, %al
 ; SKX-NEXT:    fxch %st(3)
 ; SKX-NEXT:    fcmovne %st(4), %st(0)
 ; SKX-NEXT:    fstp %st(4)
 ; SKX-NEXT:    fxch %st(3)
 ; SKX-NEXT:    fstpt 10(%rdi)
 ; SKX-NEXT:    fxch %st(1)