This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.
Index Nav: | [Date Index] [Subject Index] [Author Index] [Thread Index] | |
---|---|---|
Message Nav: | [Date Prev] [Date Next] | [Thread Prev] [Thread Next] |
Other format: | [Raw text] |
As I mentioned in the last message, my previous patch had some problems that showed up on big endian systems, using RELOAD (one of the tests that failed was the vshuf-v32qi.c test in the testsuite). Little endian and IRA did compiled the test fine. This patch fixes the problem. I went over the alternatives in the vsx_mov<mode>_{32bit,64bit} patterns, and I removed the '*' constraints, and checked all of the other constraints. Just to be sure, I build the Spec 2006 benchmark with this patch with the following variants, and all built fine (using svn id 236136 as the base revision). 32-bit power7 big endian, reload 32-bit power7 big endian, ira 64-bit power7 big endian, reload 64-bit power7 big endian, ira 64-bit power8 little endian, reload 64-bit power8 little endian, ira I also built the power8 little endian versions with subversion id 236325, and all built fine. I am having trouble with 236325 on my big endian power7 system, but it fails with both the straight trunk, and with these patches, so it is something else. The patches bootstrap and pass regression tests on both little endian power8 and big endian power7 systems. Are these patches ok to install in the trunk? After a burn-in period, will they be ok to back port to the GCC 6.2 branch? [gcc] 2016-05-17 Michael Meissner <meissner@linux.vnet.ibm.com> PR target/70915 * config/rs6000/constraints.md (wE constraint): New constraint for a vector constant that can be loaded with XXSPLTIB. (wM constraint): New constraint for a vector constant of a 1's. (wS constraint): New constraint for a vector constant that can be loaded with XXSPLTIB and a vector sign extend instruction. * config/rs6000/predicates.md (xxspltib_constant_split): New predicates for wE/wS constraints. (xxspltib_constant_nosplit): Likewise. (easy_vector_constant): Add support for constants that can be loaded via XXSPLTIB. (all_ones_constant): New predicate for vector constant with all 1's set. (splat_input_operand): Add support for ISA 3.0 word splat operations. * config/rs6000/rs6000.c (xxspltib_constant_p): New function to return if a constant can be loaded with the ISA 3.0 XXSPLTIB instruction and possibly with a sign extension. (output_vec_const_move): Add support for XXSPLTIB. If we are loading up 0/-1 into Altivec registers, prefer using VSPLTISW instead of XXLXOR/XXLORC. (rs6000_expand_vector_init): Add support for ISA 3.0 word splat operations. (rs6000_legitimize_reload_address): Likewise. (rs6000_output_move_128bit): Use output_vec_const_move to emit constants. * config/rs6000/vsx.md (VSX_M): Add TImode (if -mvsx-timode) and combine VSX_M and VSX_M2 into one iterator. (VSX_M2): Likewise. (VSINT_84): New iterators for loading constants with XXSPLTIB. (VSINT_842): Likewise. (UNSPEC_VSX_SIGN_EXTEND): New UNSPEC. (xxspltib_v16qi): New insns to load up constants with the ISA 3.0 XXSPLTIB instruction. (xxspltib_<mode>_nosplit): Likewise. (xxspltib_<mode>_split): New insn to load up constants with XXSPLTIB and a sign extend instruction. (vsx_mov<mode>): Replace single move that handled all vector types with separate 32-bit and 64-bit moves. Combine the movti_<bit> moves (when -mvsx-timode is in effect) into the main vector moves. Eliminate separate moves for <VSr> <VSa>, where the preferred register class (<VSr>) is listed first, and the secondary register class (<VSa>) is listed second with a '?' to discourage use. Prefer loading 0/-1 in any VSX register for ISA 3.0, and Altivec registers for ISA 2.06/2.07 (PR target/70915) so that if the register was involved in a slow operation, the clear/set operation does not wait for the slow operation to finish. Adjust the length attributes for 32-bit mode. Use rs6000_output_move_128bit and drop the use of the string instructions for 32-bit movti when -mvsx-timode is in effect. Use spacing so that the alternatives and attributes don't generate long lines, and put things in columns, so that it is easier to match up the operands and attributes with the insn alternatives. (vsx_mov<mode>_64bit): Likewise. (vsx_mov<mode>_32bit): Likewise. (vsx_movti_64bit): Fold movti into normal vector moves. (vsx_movti_32bit): Likewise. (vsx_splat_<mode>, V4SI/V4SF modes): Add support for ISA 3.0 word spat instructions. (vsx_splat_v4si_internal): Likewise. (vsx_splat_v4sf_internal): Likewise. (vector fusion peepholes): Use VSX_M instead of VSX_M2. (vsx_sign_extend_qi_<mode>): New ISA 3.0 instructions to sign extend vector elements. (vsx_sign_extend_hi_<mode>): Likewise. (vsx_sign_extend_si_v2di): Likewise. * config/rs6000/rs6000-protos.h (xxspltib_constant_p): Add declaration. * doc/md.texi (PowerPC constraints): Document the wE, wM, and wS constraints. Add trailing period to wL documentation. [gcc/testsuite] 2016-05-17 Michael Meissner <meissner@linux.vnet.ibm.com> * gcc.target/powerpc/p9-splat-1.c: New tests for ISA 3.0 word splat operations and the XXSPLTIB instruction. * gcc.target/powerpc/p9-splat-2.c: Likewise. * gcc.target/powerpc/p9-splat-3.c: Likewise. * gcc.target/powerpc/pr47755.c: Allow vspltisw in addition to xxlxor to clear a register. -- Michael Meissner, IBM IBM, M/S 2506R, 550 King Street, Littleton, MA 01460-6245, USA email: meissner@linux.vnet.ibm.com, phone: +1 (978) 899-4797
Attachment:
gcc-stage7.splat003b
Description: Text document
Index Nav: | [Date Index] [Subject Index] [Author Index] [Thread Index] | |
---|---|---|
Message Nav: | [Date Prev] [Date Next] | [Thread Prev] [Thread Next] |