[llvm] [Support][ValueTraking] Improve KnownFPClass for fadd. Handle infinity signs and strict sign cases (PR #190559)

[Max Graey via llvm-commits]

================
@@ -340,10 +340,7 @@ define i1 @uitofp_add(i32 %arg0) {
 
 define i1 @uitofp_add_big(i1024 %arg0) {
 ; CHECK-LABEL: @uitofp_add_big(
-; CHECK-NEXT:    [[OP:%.*]] = uitofp i1024 [[ARG0:%.*]] to double
-; CHECK-NEXT:    [[ADD:%.*]] = fadd double [[OP]], [[OP]]
-; CHECK-NEXT:    [[TMP:%.*]] = fcmp ord double [[ADD]], [[ADD]]
-; CHECK-NEXT:    ret i1 [[TMP]]
+; CHECK-NEXT:    ret i1 true
----------------
MaxGraey wrote:

proof (but for half instead double) : https://alive2.llvm.org/ce/z/S9B8cu

https://github.com/llvm/llvm-project/pull/190559