[llvm] [ValueTracking] Compute knownbits from known fp classes (PR #86409)

[via llvm-commits]

================
@@ -1195,5 +1195,199 @@ define i1 @extract_value_smul_fail(i8 %xx, i8 %yy) {
   ret i1 %r
 }
 
+define i1 @test_sign_pos(float %x) {
+; CHECK-LABEL: @test_sign_pos(
+; CHECK-NEXT:    ret i1 true
+;
+  %fabs = call float @llvm.fabs.f32(float %x)
+  %y = bitcast float %fabs to i32
+  %sign = icmp sgt i32 %y, -1
+  ret i1 %sign
+}
+
+define i1 @test_sign_pos_half(half %x) {
+; CHECK-LABEL: @test_sign_pos_half(
+; CHECK-NEXT:    ret i1 true
+;
+  %fabs = call half @llvm.fabs.f16(half %x)
+  %y = bitcast half %fabs to i16
+  %sign = icmp sgt i16 %y, -1
+  ret i1 %sign
+}
+
+define i1 @test_sign_pos_half_non_elementwise(<2 x half> %x) {
+; CHECK-LABEL: @test_sign_pos_half_non_elementwise(
+; CHECK-NEXT:    [[FABS:%.*]] = call <2 x half> @llvm.fabs.v2f16(<2 x half> [[X:%.*]])
+; CHECK-NEXT:    [[Y:%.*]] = bitcast <2 x half> [[FABS]] to i32
+; CHECK-NEXT:    [[SIGN:%.*]] = icmp sgt i32 [[Y]], -1
+; CHECK-NEXT:    ret i1 [[SIGN]]
+;
+  %fabs = call <2 x half> @llvm.fabs.v2f16(<2 x half> %x)
+  %y = bitcast <2 x half> %fabs to i32
----------------
goldsteinn wrote:

As a potential todo, we could handle vec of smaller type -> bigger type. The only case that is properly tricky is bigger type -> vec of small type.

https://github.com/llvm/llvm-project/pull/86409