Tablegen type comparison LE should be LT or equal.

[petecoup]

Summary

This fixes case where one could have LE(A, B) AND LE(B, A) for different sized
vector types A,B.

[petecoup]

llvm/utils/TableGen/CodeGenDAGPatterns.cpp

517

Krzysztof, I believe you made this change? I was not able to find these trunc patterns in the X86 backend, and all the default backends seem to build/test fine with my change here. Can you comment on whether this looks OK to you?

[craig.topper]

The X86 patterns are probably heavily hidden in tablegen classes, but they should be easily visible in X86GenDAGISel.inc. Does that file generate the same before and after this change?

/*515111*/    /*Scope*/ 26, /*->515138*/
/*515112*/      OPC_CheckChild0Type, MVT::v4i32,
/*515114*/      OPC_SwitchType /*2 cases */, 9, MVT::v16i8,// ->515126
/*515117*/        OPC_CheckPatternPredicate, 2, // (Subtarget->hasAVX512()) && (Subtarget->hasVLX())
/*515119*/        OPC_MorphNodeTo1, TARGET_VAL(X86::VPMOVDBZ128rr), 0,
                      MVT::v16i8, 1/*#Ops*/, 0,·
                  // Src: (X86vtrunc:{ *:[v16i8] } VR128X:{ *:[v4i32] }:$src) - Complexity = 3
                  // Dst: (VPMOVDBZ128rr:{ *:[v16i8] } VR128X:{ *:[v4i32] }:$src)

[kparzysz]

llvm/utils/TableGen/CodeGenDAGPatterns.cpp

517

I think it was originally < or ==, but then I encountered those issues in the x86 backend, so I changed it to <. In any case, the original intent was <=, so if it works now, I'm fine with it.

[petecoup]

@craig.topper Yes, X86GenDAGIsel.inc generates the same before/after this change.

llvm/utils/TableGen/CodeGenDAGPatterns.cpp

for (unsigned M : Modes) {
     }
   }
 
   auto LT = [](MVT A, MVT B) -> bool {
     return A.getScalarSizeInBits() < B.getScalarSizeInBits() ||
            (A.getScalarSizeInBits() == B.getScalarSizeInBits() &&
             A.getSizeInBits() < B.getSizeInBits());
   };
-  auto LE = [](MVT A, MVT B) -> bool {
+  auto LE = [&LT](MVT A, MVT B) -> bool {
     // This function is used when removing elements: when a vector is compared
     // to a non-vector, it should return false (to avoid removal).
     if (A.isVector() != B.isVector())
       return false;
 
-    // Note on the < comparison below:
-    // X86 has patterns like
-    //   (set VR128X:$dst, (v16i8 (X86vtrunc (v4i32 VR128X:$src1)))),
-    // where the truncated vector is given a type v16i8, while the source
-    // vector has type v4i32. They both have the same size in bits.
-    // The minimal type in the result is obviously v16i8, and when we remove
-    // all types from the source that are smaller-or-equal than v8i16, the
-    // only source type would also be removed (since it's equal in size).
-    return A.getScalarSizeInBits() <= B.getScalarSizeInBits() ||
-           A.getSizeInBits() < B.getSizeInBits();
+    return LT(A, B) || (A.getScalarSizeInBits() == B.getScalarSizeInBits() &&
+                        A.getSizeInBits() == B.getSizeInBits());
   };
 
   for (unsigned M : Modes) {
     TypeSetByHwMode::SetType &S = Small.get(M);
     TypeSetByHwMode::SetType &B = Big.get(M);
     // MinS = min scalar in Small, remove all scalars from Big that are
     // smaller-or-equal than MinS.
     auto MinS = min_if(S.begin(), S.end(), isScalar, LT);