[llvm] DAG: Handle lowering unordered compare with inf (PR #100378)

[Serge Pavlov via llvm-commits]

================
@@ -240,18 +240,22 @@ entry:
 define i1 @isfinite_f(float %x) {
 ; X86-LABEL: isfinite_f:
 ; X86:       # %bb.0: # %entry
-; X86-NEXT:    movl $2147483647, %eax # imm = 0x7FFFFFFF
-; X86-NEXT:    andl {{[0-9]+}}(%esp), %eax
-; X86-NEXT:    cmpl $2139095040, %eax # imm = 0x7F800000
-; X86-NEXT:    setl %al
+; X86-NEXT:    flds {{[0-9]+}}(%esp)
----------------
spavloff wrote:

It looks like for targets like X86 this behavior is bad. As it requires materialization of infinity value, it produces worse code.

Is there an example when this patch is profitable?

Can we confine this behavior to the profitable cases?

https://github.com/llvm/llvm-project/pull/100378