All, the attached patch file contains:
1.) Updates {|v}movss/{|v}movsd (AVX/SSE) conforming to the isa
for both SSE2 and AVX code.
2.) Changes in copy detection and side effect detection for LRA
and EBO.
In doing this, I updated the auto translation back to originating forms for
movss, movsd, vmovss, vmovd so that intrinsics will map properly for all of
x8664.
Also I updated the copy expansion code to use movaps under the same target
guards that we did in the auto translation remap before. The change produces
correct code on BD and all SSE enabled x8664 architectures for expansion
patterns and intrinsics. The expanded SSE movs{s|d} instructions have been
replaced with movdq in the generic path as it correctly fits the motif as a
single source move.
These mov instructions are really merge operations that need to preserve a data
path.
These changes pass the following:
a.) No compile time failure for x86 build.
b.) The gcc regression test suite on x86/Linux with no new failures(GH
and BD).
c.) The SPEC2006 test suite at with current AMD 1 copy config at both
base and peak(GH and BD).
Thx,
m
simd_move_update.p
Description: simd_move_update.p
------------------------------------------------------------------------------ 5 Ways to Improve & Secure Unified Communications Unified Communications promises greater efficiencies for business. UC can improve internal communications as well as offer faster, more efficient ways to interact with customers and streamline customer service. Learn more! http://www.accelacomm.com/jaw/sfnl/114/51426253/
_______________________________________________ Open64-devel mailing list [email protected] https://lists.sourceforge.net/lists/listinfo/open64-devel
