Re: [PATCH, rs6000] Fix PR target/67808, LRA ICE on double to long double conversion

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

Ok, after spending the day on going down the rabbit hole of trying to optimize
just about every, here are my patches.

Note, I simplified the constraints to eliminate some rare possibilities, like
optimizing converting from double to long double if the double happened to be
in a GPR register and the long double value was stored in memory (but there
never was an optimization for having the double value be in a GPR and the long
double value to also be a GPR).

I also separated the VSX case from the non-VSX case. This is to simplify things
at the RTL level (non-VSX must load up 0.0 to put into the lower word, while
VSX can use the XXLXOR instruction to clear the register).

I dropped support in the insns for extending the DFmode value to TFmode that is
located in memory directly. Now, the compiler builds the whole value in FPRs
and then does the store. This simplifies the code somewhat, and SPE/ieeequad
paths require the value to be in registers, which might lead to other lra
bugs. In the case of just doing one conversion in straight line code, it just
changes the register allocation somewhat (allocate 1 TFmode pseudo instead of
1-2 DFmode psuedos).

I have bootstrapped the compiler on little endian power8 with no regressions. I
have built the test case with various options (-mlra vs. no -mlra, 32-bit,
64-bit, power5/power6/power7/power8), and it all builds correctly.

Is this patch ok to apply to the trunk?

I would like to apply this patch to GCC 5.x as well. However, in doing the
patch, this patch touches areas that I've been working on for IEEE 128-bit
floating point support, and so the patch will need to be reworked for GCC
5.x.  Is it ok to install in the trunk?

In addition, I will need to modify this area again with the next IEEE 128-bit
floating point support patch, but I wanted to separate this patch out so that
it could be considered by itself, and back ported to GCC 5.x.

[gcc]
2015-10-05  Michael Meissner  <meissner@linux.vnet.ibm.com>
	    Peter Bergner  <bergner@vnet.ibm.com>

	PR target/67808
	* config/rs6000/rs6000.md (extenddftf2): In the expander, only
	allow registers, but provide insns for the combiner to create for
	loads from memory. Separate VSX code from non-VSX code. For
	non-VSX code, combine extenddftf2_fprs into extenddftf2 and rename
	externaldftf2_internal to externaldftf2_fprs. Reorder constraints
	so that registers come before memory operations. Drop support from
	converting DFmode to TFmode, if the DFmode value is in a GPR
	register, and the TFmode is in memory.
	(extenddftf2_fprs): Likewise.
	(extenddftf2_internal): Likewise.
	(extenddftf2_vsx): Likewise.
	(extendsftf2): In the expander, only allow registers, but provide
	insns for the combiner to create for stores and loads.

[gcc/testsuite]
2015-10-05  Michael Meissner  <meissner@linux.vnet.ibm.com>
	    Peter Bergner <bergner@vnet.ibm.com>

	PR target/67808
	* gcc.target/powerpc/pr67808.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: pr67808.patch02b
Description: Text document