projects / platform / upstream / llvm.git / commit
? search:
summary | shortlog | log | commit | commitdiff | tree
(parent: ff8bb49) | patch
AMDGPU: Materialize frame index before add
authorMatt Arsenault <Matthew.Arsenault@amd.com>
Tue, 29 Nov 2016 19:20:48 +0000 (19:20 +0000)
committerMatt Arsenault <Matthew.Arsenault@amd.com>
Tue, 29 Nov 2016 19:20:48 +0000 (19:20 +0000)
commitf96eeec0057c8535be809d324be22ccb690345bb
tree96426ec82b5b3e059eb0529018d30c783c28bc28tree | snapshot
parentff8bb49bf45aa0de0908841720a7c592a76ebc0ccommit | diff
AMDGPU: Materialize frame index before add

It isn't generally safe to fold the frame index
directly into the operand since it will possibly
not be an inline immediate after it is expanded.

This surprisingly seems to produce better code, since
the FI doesn't prevent folding other immediate operands.

llvm-svn: 288185
llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp diff | blob | history
llvm/test/CodeGen/AMDGPU/local-stack-slot-bug.ll diff | blob | history
Domain: System / Toolchain;