[PATCH] D88475: [InstCombine] visitTrunc - trunc (lshr (sext A), C) --> (ashr A, C) non-uniform support

[Roman Lebedev via Phabricator via llvm-commits]

lebedev.ri accepted this revision.
lebedev.ri added a comment.
This revision is now accepted and ready to land.

LGTM



================
Comment at: llvm/lib/Transforms/InstCombine/InstCombineCasts.cpp:830-855
+  Constant *C;
+  if (match(Src, m_LShr(m_SExt(m_Value(A)), m_Constant(C)))) {
     unsigned AWidth = A->getType()->getScalarSizeInBits();
     unsigned MaxShiftAmt = SrcWidth - std::max(DestWidth, AWidth);
 
     // If the shift is small enough, all zero bits created by the shift are
     // removed by the trunc.
----------------
lebedev.ri wrote:
> 
Err, that should of course be 
```
    // Note that since we will perform the shift in narrower type, we need to clamp the shift amount.
    Constant *MaxAmt = ConstantInt::get(SrcTy, AWidth - 1, false);
    Constant *ShAmt = ConstantExpr::getUMin(C, MaxAmt);
    ShAmt = ConstantExpr::getTrunc(ShAmt, A->getType());
```


================
Comment at: llvm/test/Transforms/InstCombine/cast.ll:1563
-; ALL-NEXT:    [[B:%.*]] = sext <2 x i8> [[A:%.*]] to <2 x i32>
-; ALL-NEXT:    [[C:%.*]] = lshr <2 x i32> [[B]], <i32 6, i32 undef>
-; ALL-NEXT:    [[D:%.*]] = trunc <2 x i32> [[C]] to <2 x i8>
----------------
It would be kind of good to preserve `undef`, but there is no nice way to do that currently.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D88475/new/

https://reviews.llvm.org/D88475