[llvm] [ValueTracking] Improve `Bitcast` handling to match SDAG (PR #125935)

[Nikita Popov via llvm-commits]

================
@@ -1376,10 +1377,32 @@ static void computeKnownBitsFromOperator(const Operator *I,
       for (unsigned i = 0; i != SubScale; ++i) {
         computeKnownBits(I->getOperand(0), SubDemandedElts.shl(i), KnownSrc, Q,
                          Depth + 1);
-        unsigned ShiftElt = Q.DL.isLittleEndian() ? i : SubScale - 1 - i;
+        unsigned ShiftElt = IsLE ? i : SubScale - 1 - i;
         Known.insertBits(KnownSrc, ShiftElt * SubBitWidth);
       }
     }
+    // Look through a cast from wider vector elements to narrow type.
+    // Examples: v2i64 -> v4i32
+    if (SubBitWidth % BitWidth == 0) {
+      unsigned SubScale = SubBitWidth / BitWidth;
+      KnownBits KnownSrc(SubBitWidth);
+      APInt SubDemandedElts =
+          APIntOps::ScaleBitMask(DemandedElts, NumElts / SubScale);
+      computeKnownBits(I->getOperand(0), SubDemandedElts, KnownSrc, Q,
+                       Depth + 1);
+
+      Known.Zero.setAllBits();
+      Known.One.setAllBits();
+      for (unsigned i = 0; i != SubScale; ++i) {
----------------
nikic wrote:

I think this was supposed to go up to NumElts, not SubScale. If one of the first two elements are non-demanded but later ones are demanded this will miscompile.

https://github.com/llvm/llvm-project/pull/125935