[Intel-gfx] [PATCH 0/3] Claw back the IS_<platform> optimisation
Tvrtko Ursulin
tursulin at ursulin.net
Thu Dec 8 09:49:56 UTC 2016
From: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
Reorganise the IS_<platform> code a bit inspired by the pre-processor
code generation as used by Chris in the self test patches.
This avoids some source code duplication and makes maintenance easier.
In the last patch try to claw back the multiple IS_<platform> optimisation,
as the commit message says:
If we use only power of two values for the platform enum
values we can let the compiler optimize some common
checks to a single conditional.
For example code like this:
if (IS_VALLEYVIEW(dev_priv) || IS_CHERRYVIEW(dev_priv))
Goes from this:
5c3c5: 8b 83 d8 06 00 00 mov 0x6d8(%rbx),%eax
5c3cb: 83 f8 12 cmp $0x12,%eax
5c3ce: 0f 84 f3 00 00 00 je 5c4c7 <fw_domain_init+0x1a7>
5c3d4: 83 f8 15 cmp $0x15,%eax
5c3d7: 0f 84 ea 00 00 00 je 5c4c7 <fw_domain_init+0x1a7>
To this:
5c1d5: f7 83 d8 06 00 00 00 testl $0x240000,0x6d8(%rbx)
5c1dc: 00 24 00
5c1df: 0f 85 da 00 00 00 jne 5c2bf <fw_domain_init+0x18f>
It is not much but there is value in this that as long as we
have to have conditions like the above sprinkled troughout the
code, we can at least have the generate binary a bit smarter.
Until we get to more than 32 platforms there is no downside
to this approach.
Patch series saves 1890 bytes of binary:
text data bss dec hex filename
1103970 26273 580 1130823 114147 i915.ko.0
1102080 26273 580 1128933 1139e5 i915.ko.1
Tvrtko Ursulin (3):
drm/i915: Introduct i915_platforms.h
drm/i915: Generate all IS_<platform> macros
drm/i915: Number the platform enum strategically
drivers/gpu/drm/i915/i915_drv.h | 69 +++++++++-----------------------
drivers/gpu/drm/i915/i915_platforms.h | 34 ++++++++++++++++
drivers/gpu/drm/i915/intel_device_info.c | 38 ++++--------------
3 files changed, 61 insertions(+), 80 deletions(-)
create mode 100644 drivers/gpu/drm/i915/i915_platforms.h
--
2.7.4
More information about the Intel-gfx
mailing list