At the moment the ECF_* flags for a gimple call to a built-in function are derived from the function decl, which in turn is derived from the global command-line options. So if the compiler is run with -fno-math-errno, we always assume functions don't set errno, regardless of local optimisation options. Similarly if the compiler is run with -fmath-errno, we always assume functions set errno. This shows up in gcc.dg/lto/20110201-1_0.c, where we compile the file with -O0 and use -O2 -ffast-math for a specific function. -O2 -ffast-math is enough for us to convert cabs to sqrt as hoped, but because of the global -fmath-errno setting, we assume that the call to sqrt is not pure or const and create vops for it. This makes it appear to the gimple code that a simple sqrt optab isn't enough.
Confirmed. The solution for this specific case is to have distinct builtins for no-math-errno, like __builtin_sqrt_no_math_errno (). Note that slightly related is -frounding-math which causes math functions to become pure rather than const (and thus get a VUSE).
Author: rsandifo Date: Sat Nov 7 09:58:32 2015 New Revision: 229917 URL: https://gcc.gnu.org/viewcvs?rev=229917&root=gcc&view=rev Log: Add -fno-math-errno to gcc.dg/lto/20110201-1_0.c At the moment the ECF_* flags for a gimple call to a built-in function are derived from the function decl, which in turn is derived from the global command-line options. So if the compiler is run with -fno-math-errno, we always assume functions don't set errno, regardless of local optimization options. Similarly if the compiler is run with -fmath-errno, we always assume functions set errno. This shows up in gcc.dg/lto/20110201-1_0.c, where we compile the file with -O0 and use -O2 -ffast-math for a specific function. -O2 -ffast-math is enough for us to convert cabs to sqrt as hoped, but because of the global -fmath-errno setting, we assume that the call to sqrt is not pure or const and create vops for it. This makes it appear to the gimple code that a simple sqrt optab isn't enough. Later patches move more decisions about maths functions to gimple and think that in this case we should use: y = sqrt (x); if (!(x >= 0)) sqrt (x); // to set errno. This is being tracked as PR68235. For now the patch adds -fno-math-errno to the dg-options for this test. Tested on x86_64-linux-gnu, aarch64-linux-gnu and arm-linux-gnueabi. gcc/testsuite/ PR tree-optimization/68235 * gcc.dg/lto/20110201-1_0.c: Add -fno-math-errno. Modified: trunk/gcc/testsuite/ChangeLog trunk/gcc/testsuite/gcc.dg/lto/20110201-1_0.c
Can the bug be marked as resolved?
I'm not aware of any fix for this bug (the only commit shown is a change to testcase options, not a fix).