[PATCH] PR target/65240, Fix Power{7,8} insn constraint issue with -O3 -ffast-math

[Michael Meissner <meissner at linux dot vnet dot ibm dot com>]

This patch fixes PR 65240, which was a latent bug that was introduced when I
added the -mupper-regs support to the PowerPC compiler.  In the PowerPC
compiler, if you use -ffast-math, the compiler allows floating point constants
in move RTLs until register allocation time, in order to allow the
optimizations that replace division by a constant with multiplication by the
reciprocal.  If the register being loaded up is a traditional Altivec register,
the load will fail, since there is no D-FORM (register+offset) addressing mode
for the traditional altivec registers.

I had had a define_split to try and handle this situation, but it didn't work
all of the time.  I rewrote the insns, so there is an UNSPEC in it to prevent
over optimization.

I have done bootstrap builds and make check's on both a big endian power7
system, and a little endian power8 system with no regressions in the tests.  I
have also built both power7 and power8 spec 2006 (v5) versions, and Spec
built.  I have run the Spec floating point tests on a power7 box, and all of
the tests performed about at the same speed as before the bug fix was done,
with the exception of gromacs, which is 4% faster with the fix.

Looking at the static instruction counts, the big change is to fold the addi
instruction into the load of the constant if it is loading the value to a
traditional floating point register (previously, it would load up the constant
with addis/addi and then do an indirect load).  I do see other instructions
change, including a few more floating point loads, but with the exception of
gromacs, most of the changes are in the noise level.

Is this ok to install into the trunk?  At present, it is not needed for GCC
4.9, since that branch does not have the -mupper-regs support.

[gcc]
2015-03-05  Michael Meissner  <meissner@linux.vnet.ibm.com>

	PR target/65240
	* config/rs6000/rs6000.md (UNSPEC_FUSION_FP_CONSTANT): New unspec.
	(Vsx_load): New mode attribute to give appropriate VSX load
	instruction.
	(move floating point constant define_split): Use an UNSPEC to make
	sure that the load of the constant is kept as a memory address,
	instead of being converted back into a move of the constant.
	(load_<mode>_constant): Likewise.

[gcc/testsuite]
2015-03-05  Michael Meissner  <meissner@linux.vnet.ibm.com>

	PR target/65240
	* gcc.target/powerpc/pr65240.c: New test.

-- 
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: pr65240.patch03b
Description: Text document