Re: [Patch AArch64] Add support for vectorizable standard math patterns.
On 27/11/12 20:47, Mike Stump wrote: On Nov 27, 2012, at 8:51 AM, James Greenhalgh james.greenha...@arm.com wrote: In particular, we add support for vectorizing across: ceil (), ceilf (), lceil (), We add testcases ensuring that each of the expected functions are vectorized. As the i386 and rs6000 backends both ostensibly support these optimisations we add these tests to the generic testsuites, but only wire them up for AArch64. As a target may support any subset of these vectorizations we need a check_effective_target macro for each of them. Because of this change to the generic test code I've CCed Janis Johnson and Mike Stump. Gosh… leaves a bad taste in my mouth.I see why you did it that way… So, let me just ping folks, anyone see a better way to do this? If no one admits to having a better solution, I'll approve it. Let's give them a few days to come up with something concrete, if they fail, Ok. I've been pondering this one for a bit. I wonder if having a single function check_effective_target_vect_libcall(func) where func is the name of the function you needed to check, might be cleaner. The idea is that the one function might access an array (set up on the first call) and then just use info exists et_vect_libcalls(func) to do the body of the check. Not entirely sure my expect foo is good enough to code this up, though. R.
RE: [Patch AArch64] Add support for vectorizable standard math patterns.
Is this patch OK to commit? OK /Marcus Thanks Marcus, I've also back-ported this to AArch64-4.7-branch. The back-port was clean and passed a regression run on aarch64-none-elf with no regressions. Thanks, James
Re: [Patch AArch64] Add support for vectorizable standard math patterns.
On 27/11/12 16:51, James Greenhalgh wrote: Hi, This patch adds support for vectorizing across some of the rounding functions in the C math library to the AArch64 back-end. In particular, we add support for vectorizing across: ceil (), ceilf (), lceil (), floor (), floorf (), lfloor (), round (), roundf (), nearbyint (), nearbyintf (), trunc (), truncf () We add testcases ensuring that each of the expected functions are vectorized. As the i386 and rs6000 backends both ostensibly support these optimisations we add these tests to the generic testsuites, but only wire them up for AArch64. As a target may support any subset of these vectorizations we need a check_effective_target macro for each of them. Because of this change to the generic test code I've CCed Janis Johnson and Mike Stump. Is this patch OK to commit? OK /Marcus
Re: [Patch AArch64] Add support for vectorizable standard math patterns.
On 27/11/12 20:47, Mike Stump wrote: On Nov 27, 2012, at 8:51 AM, James Greenhalgh james.greenha...@arm.com wrote: In particular, we add support for vectorizing across: ceil (), ceilf (), lceil (), We add testcases ensuring that each of the expected functions are vectorized. As the i386 and rs6000 backends both ostensibly support these optimisations we add these tests to the generic testsuites, but only wire them up for AArch64. As a target may support any subset of these vectorizations we need a check_effective_target macro for each of them. Because of this change to the generic test code I've CCed Janis Johnson and Mike Stump. Gosh… leaves a bad taste in my mouth.I see why you did it that way… So, let me just ping folks, anyone see a better way to do this? If no one admits to having a better solution, I'll approve it. Let's give them a few days to come up with something concrete, if they fail, Ok. Mike, Are you happy to go with the proposal from James or would like to give folk more time to think this one over? Cheers /Marcus
Re: [Patch AArch64] Add support for vectorizable standard math patterns.
On Dec 3, 2012, at 8:43 AM, Marcus Shawcroft marcus.shawcr...@arm.com wrote: Mike, Are you happy to go with the proposal from James or would like to give folk more time to think this one over? Good to go. If people want to improve it, if they see a way, then can submit a patch anytime they want. I didn't see anyone lay down a clearly better way to do this.
Re: [Patch AArch64] Add support for vectorizable standard math patterns.
On Nov 27, 2012, at 8:51 AM, James Greenhalgh james.greenha...@arm.com wrote: In particular, we add support for vectorizing across: ceil (), ceilf (), lceil (), We add testcases ensuring that each of the expected functions are vectorized. As the i386 and rs6000 backends both ostensibly support these optimisations we add these tests to the generic testsuites, but only wire them up for AArch64. As a target may support any subset of these vectorizations we need a check_effective_target macro for each of them. Because of this change to the generic test code I've CCed Janis Johnson and Mike Stump. Gosh… leaves a bad taste in my mouth.I see why you did it that way… So, let me just ping folks, anyone see a better way to do this? If no one admits to having a better solution, I'll approve it. Let's give them a few days to come up with something concrete, if they fail, Ok.
Re: [Patch AArch64] Add support for vectorizable standard math patterns.
On Tue, Nov 27, 2012 at 8:47 PM, Mike Stump mikest...@comcast.net wrote: On Nov 27, 2012, at 8:51 AM, James Greenhalgh james.greenha...@arm.com wrote: In particular, we add support for vectorizing across: ceil (), ceilf (), lceil (), We add testcases ensuring that each of the expected functions are vectorized. As the i386 and rs6000 backends both ostensibly support these optimisations we add these tests to the generic testsuites, but only wire them up for AArch64. As a target may support any subset of these vectorizations we need a check_effective_target macro for each of them. Because of this change to the generic test code I've CCed Janis Johnson and Mike Stump. Gosh… leaves a bad taste in my mouth.I see why you did it that way… So, let me just ping folks, anyone see a better way to do this? If no one admits to having a better solution, I'll approve it. Let's give them a few days to come up with something concrete, if they fail, Ok. If it's any consolation I expect the ARM port to use a subset of these, because in the AArch32 world on v8 we only have the single precision vector variants of these instructions and not the double precision vector variants once we can get hold of the intrinsic decls we need :) . So, if someone's thinking who's going to be a consumer of a subset of these, here's one port that's likely to need this. That doesn't preclude finding a better way of doing this :) regards Ramana
