tests: use identical(NA_real_+0, NA_real_) to skip tests that depend on NA payload propagation (#7702)

* froll.Rraw: skip tests that depend on NA payload

RISC-V chips do not propagate NaN payload through arithmetic operations.

* Better test for when to skip NA tests

Co-Authored-By: Michael Chirico <chiricom@google.com>

* Drop unused check from frollBatch.Rraw

* Test R behaviour to stop NA-related test

* Fix typo

---------

Co-authored-by: Michael Chirico <chiricom@google.com>

Author: aitap

inst/tests/froll.Rraw

@@ -9,11 +9,9 @@ if (exists("test.data.table", .GlobalEnv, inherits=FALSE)) {
   froll = data.table:::froll
 }
 
-exact_NaN = isTRUE(capabilities()["long.double"]) && identical(as.integer(.Machine$longdouble.digits), 64L)
+exact_NaN = identical(NA_real_+0, NA_real_)
 if (!exact_NaN) {
-  cat("\n**** Skipping 8 NaN/NA algo='exact' tests because .Machine$longdouble.digits==", .Machine$longdouble.digits, " (!=64); e.g. under valgrind\n\n", sep="")
-  # for Matt when he runs valgrind it is 53, but 64 when running regular R
-  # froll.c uses long double and appears to require full long double accuracy in the algo='exact'
+  cat("\n**** Skipping 10 NaN/NA algo='exact' tests because NaN payload doesn't propagate through arithmetic operations\n\n")
 }
 
 ## rolling features
@@ -1456,8 +1454,10 @@ test(6001.731, between(frollvar(y, 3)[4L], 0, 1e-7))
 test(6001.732, between(frollsd(y, 3)[4L], 0, 1e-7))
 test(6001.733, frollvar(y, c(3,3,3,3), adaptive=TRUE)[4L], 0)
 test(6001.734, frollsd(y, c(3,3,3,3), adaptive=TRUE)[4L], 0)
-test(6001.740, frollvar(c(1.5,2.5,2,NA), c(3,3)), list(c(NA,NA,0.25,NA), c(NA,NA,0.25,NA)), output="running sequentially, because outer parallelism has been used", options=c(datatable.verbose=TRUE)) # ensure no nested parallelism in rolling functions #7352
-test(6001.741, frollsd(c(1.5,2.5,2,NA), c(3,3)), list(c(NA,NA,0.5,NA), c(NA,NA,0.5,NA)), output="running sequentially, because outer parallelism has been used", options=c(datatable.verbose=TRUE))
+if (exact_NaN) {
+  test(6001.740, frollvar(c(1.5,2.5,2,NA), c(3,3)), list(c(NA,NA,0.25,NA), c(NA,NA,0.25,NA)), output="running sequentially, because outer parallelism has been used", options=c(datatable.verbose=TRUE)) # ensure no nested parallelism in rolling functions #7352
+  test(6001.741, frollsd(c(1.5,2.5,2,NA), c(3,3)), list(c(NA,NA,0.5,NA), c(NA,NA,0.5,NA)), output="running sequentially, because outer parallelism has been used", options=c(datatable.verbose=TRUE))
+}
 test(6001.742, frollvar(c(1.5,2.5,2,1.5), c(3,3)), list(c(NA,NA,0.25,0.25), c(NA,NA,0.25,0.25)), notOutput="running sequentially, because outer parallelism has been used", options=c(datatable.verbose=TRUE)) # no NA - no fallback to exact
 test(6001.743, frollsd(c(1.5,2.5,2,1.5), c(3,3)), list(c(NA,NA,0.5,0.5), c(NA,NA,0.5,0.5)), notOutput="running sequentially, because outer parallelism has been used", options=c(datatable.verbose=TRUE))
 test(6001.744, frollvar(c(1.5,2.5,2,NA), 3), c(NA,NA,0.25,NA), notOutput="running sequentially, because outer parallelism has been used", options=c(datatable.verbose=TRUE)) # not vectorized - no outer parallelism

inst/tests/frollBatch.Rraw

@@ -9,14 +9,6 @@ if (exists("test.data.table", .GlobalEnv, inherits=FALSE)) {
   froll = data.table:::froll
 }
 
-exact_NaN = isTRUE(capabilities()["long.double"]) && identical(as.integer(.Machine$longdouble.digits), 64L)
-if (!exact_NaN) {
-  cat("\n**** Skipping 7 NaN/NA algo='exact' tests because .Machine$longdouble.digits==", .Machine$longdouble.digits, " (!=64); e.g. under valgrind\n\n", sep="")
-  # for Matt when he runs valgrind it is 53, but 64 when running regular R
-  # froll.c uses long double and appears to require full long double accuracy in the algo='exact'
-}
-
-
 ## batch validation
 set.seed(108)
 makeNA = function(x, ratio=0.1, nf=FALSE) {

inst/tests/tests.Rraw

@@ -144,6 +144,11 @@ if (!test_longdouble) {
   cat("\n**** Full long double accuracy is not available. Tests using this will be skipped.\n\n")
   # e.g. under valgrind, longdouble.digits==53; causing these to fail: 1262, 1729.04, 1729.08, 1729.09, 1729.11, 1729.13, 1830.7; #4639
 }
+exact_NA = identical(NA_real_ + 0, NA_real_)
+if (!exact_NA) {
+  cat("\n**** NaN payload does not propagate through arithmetic operations. Tests requiring NA not to decay to NaN will be skipped.\n\n")
+  # e.g. on riscv64
+}
 
 tt = Sys.getenv("TZ", unset=NA)
 TZnotUTC = !identical(tt,"") && !is_utc(tt)
@@ -1957,12 +1962,9 @@ basemean = base::mean  # to isolate time of `::` itself
 ans3 = DT[,list(basemean(x),basemean(y)),by=list(grp1,grp2)]
 test(646, ans1, ans2)
 test(647, ans1, ans3)
-if (test_longdouble) {
+if (exact_NA) {
   test(648, anyNA(ans1$V1) && !any(is.nan(ans1$V1)))
-  # used to error with `valgrind` because of the 'long double' usage in gsumm.c (although I wonder if we need long double precision).
-  # it doesn't seem to error under valgrind anymore so the test_longdouble may be removable
-  # http://valgrind.org/docs/manual/manual-core.html#manual-core.limits
-  # http://comments.gmane.org/gmane.comp.debugging.valgrind/10340
+  # Valgrind may have had NaN payload propagation problems in the past; RISC-V will in the future
 }
 ans1 = DT[,list(mean(x,na.rm=TRUE),mean(y,na.rm=TRUE)),by=list(grp1,grp2)]
 ans2 = DT[,list(mean.default(x,na.rm=TRUE),mean.default(y,na.rm=TRUE)),by=list(grp1,grp2)]