https://gcc.gnu.org/bugzilla/show_bug.cgi?id=89437
Bug ID: 89437
Summary: [9 regression] incorrect result for sinl (atanl (x))
Product: gcc
Version: 9.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: wilco at gcc dot gnu.org
Target Milestone: ---
A recently added optimization uses an inline expansion for sinl (atanl (x)). As
it involves computing sqrtl (x * x + 1) which can overflow for large x, there
is a check to ensure x is within the right range. The check emitted is x <=
sqrtl (LDBL_MAX). It's not clear whether that's a bug in mpfr or
build_sinatan_real, however it is safe to change the test to x < sqrtl
(LDBL_MAX).
In principle the comparison should use a much smaller constant given that x * x
+ 1 equals x * x for x > 2^113 for any long double format, so we know sinl
(atanl (x)) must be 1.0 for all larger values.
Testcase is gcc.dg/sinatan-1.c after using nextafterl instead of incorrect
nextafter for the long double. I'll post a patch to fix this.
