OSDN Git Service

Remove the X86 Maximal Stack Alignment Check pass as it is no longer necessary.
authorChad Rosier <mcrosier@apple.com>
Mon, 26 Nov 2012 22:55:05 +0000 (22:55 +0000)
committerChad Rosier <mcrosier@apple.com>
Mon, 26 Nov 2012 22:55:05 +0000 (22:55 +0000)
commit1243922fc1a1e3d2681ed9e78503eeabd875ba93
treee10f57435c2c026887a1f67c516df20c98a83ddb
parentbb8ddc7e4ff1f06c7289865bbf9e1fa4632d1c8e
Remove the X86 Maximal Stack Alignment Check pass as it is no longer necessary.
This pass was conservative in that it always reserved the FP to enable dynamic
stack realignment, which allowed the RA to use aligned spills for vector
registers.  This happens even when spills were not necessary.  The RA has
since been improved to use unaligned spills when necessary.

The new behavior is to realign the stack if the frame pointer was already
reserved for some other reason, but don't reserve the frame pointer just
because a function contains vector virtual registers.

Part of rdar://12719844

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@168627 91177308-0d34-0410-b5e6-96231b3b80d8
lib/Target/X86/X86.h
lib/Target/X86/X86RegisterInfo.cpp
lib/Target/X86/X86TargetMachine.cpp
test/CodeGen/X86/avx-intel-ocl.ll
test/CodeGen/X86/dynamic-allocas-VLAs.ll