[llvm] [InstCombine] Fold `fcmp pred sqrt(X), 0.0 -> fcmp pred2 X, 0.0` (PR #101626)

[Matt Arsenault via llvm-commits]

================
@@ -2266,6 +2296,17 @@ define i1 @fcmp_sqrt_zero_uno(half %x) {
   ret i1 %cmp
 }
 
+; Make sure that ninf is cleared.
+define i1 @fcmp_sqrt_zero_uno_fmf(half %x) {
+; CHECK-LABEL: @fcmp_sqrt_zero_uno_fmf(
+; CHECK-NEXT:    [[CMP:%.*]] = fcmp ult half [[X:%.*]], 0xH0000
+; CHECK-NEXT:    ret i1 [[CMP]]
+;
+  %sqrt = call half @llvm.sqrt(half %x)
----------------
arsenm wrote:

Missing type suffixes on all the sqrt calls.

Also, can you keep ninf if the sqrt is also ninf? 

https://github.com/llvm/llvm-project/pull/101626