https://gcc.gnu.org/bugzilla/show_bug.cgi?id=90823
Bug ID: 90823
Summary: PowerPC command line switches don't work with #pragma
CPU target or target attribute
Product: gcc
Version: 10.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: meissner at gcc dot gnu.org
Target Milestone: ---
I was working on the patch to temporarily disable -mpcrel in GCC 10 until the
full pc-relative support is in place. When I went to update the tests that
test the pc-relative support that is in there to add the -mpcrel option, I
discovered that the tests that change the target to power9 did not work
(because -mpcrel requires the 'future' processor).
This would be true of other switches like -mpower9-vector.
If you compile the following code with -O2 -mpower9-vector:
double def (long a, long b) { return (double) (a % b); }
#pragma GCC target ("cpu=power9")
double p9 (long a, long b) { return (double) (a % b); }
#pragma GCC target ("cpu=power8")
double p8 (long a, long b) { return (double) (a % b); }
#pragma GCC target ("cpu=power7")
double p7 (long a, long b) { return (double) (a % b); }
Each of the functions will get compiled for the -mcpu=power9 target because the
-mpower9-vector from the command line is applied after the #pragma GCC target
stuff is done, and using -mpower9-vector turns on all of the other power9
options.
If instead you compile it using -O2 -mcpu=power9, then it generates the
expected code (the 'def' and 'p9' functions use modsd/mtvsrd, the 'p8' function
does divide/multiply/subtract and mtvsrd, and the 'p7' function does
divide/multiply/subtract, store, and load).
