[llvm] [X86] Try Folding icmp of v8i32 -> fcmp of v8f32 on AVX (PR #82290)

[via llvm-commits]

================
@@ -41125,6 +41255,154 @@ static SDValue combineShuffle(SDNode *N, SelectionDAG &DAG,
   return SDValue();
 }
 
+// Simplify a decomposed (sext (setcc)). Assumes prior check that
+// bitwidth(sext)==bitwidth(setcc operands).
+static SDValue simplifySExtOfDecomposedSetCCImpl(
----------------
goldsteinn wrote:

Note, there are some regressions in this patch now but they should be cleared up in #84360.

Also note, the reason #84360 is second is so that we can properly test the `cmpp` code. Without this patch we never get the necessary FMF flags to actually do the `cmpp` simplifications. If you think that is unnecessary I can swap the order (#84360 does stand on its own mostly for is simplifications on `pcmpgt`).

https://github.com/llvm/llvm-project/pull/82290