[PATCH] D146512: [TLI] SimplifySetCC - Fold ~X s>/u< ~Y --> Y s>/u< X

[Jun Zhang via Phabricator via llvm-commits]

junaire marked an inline comment as done.
junaire added inline comments.


================
Comment at: llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp:4982
+  // ~X u< ~Y --> Y u< X
+  if ((Cond == ISD::SETGT || Cond == ISD::SETULE) && N0.getOpcode() == ISD::XOR &&
+      N0.getValueType().isInteger()) {
----------------
RKSimon wrote:
> Take a look at https://github.com/llvm/llvm-project/blob/3a8f161a3401edeb58e018e2d389dd2413a6417f/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp#L6501
> 
> This shows its safe to handle all icmp predicates - and handles cases where one operand is a constant which can be constant folded
Thanks for the review! However, I'm unsure what is the right way to say 'all icmp predicates' and I'm worried about why vectors don't fold correctly.

If I run `llc -mcpu=btver2 test.ll` to:
```define <2 x i64> @cmp_sgt_not(<2 x i64> %a, <2 x i64> %b) {                                                                                                                                                         %na = xor <2 x i64> %a, <i64 -1, i64 -1>                                                                                                                                                                          %nb = xor <2 x i64> %b, <i64 -1, i64 -1>                                                                                                                                                                          %c = icmp sgt <2 x i64> %na, %nb                                                                                                                                                                                  %r = sext <2 x i1> %c to <2 x i64>                                                                                                                                                                                ret <2 x i64> %r                                                                                                                                                                                                }                                                                                                                                                                                                                 
```
It does fold to:
```
vpcmpgtq        %xmm0, %xmm1, %xmm0                                                                                                                                                                               retq                                                                                                                                                                                                      
```

But for `ule` it doesn't work as well.

Let me know if you think it's necessary to add more tests for the remaining preds!


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D146512/new/

https://reviews.llvm.org/D146512