Merging r261360:
authorHans Wennborg <hans@hanshq.net>
Fri, 19 Feb 2016 21:35:00 +0000 (21:35 +0000)
committerHans Wennborg <hans@hanshq.net>
Fri, 19 Feb 2016 21:35:00 +0000 (21:35 +0000)
commit78e9cd40a2ea27cc9300d900a7dccc75940f9eb0
tree443c3f8163eadd198eb9d11671e74fe59b54c8de
parentfdf40bea4fc416643210790fff4345be98d97245
Merging r261360:
------------------------------------------------------------------------
r261360 | dim | 2016-02-19 12:14:11 -0800 (Fri, 19 Feb 2016) | 19 lines

Fix incorrect selection of AVX512 sqrt when OptForSize is on

Summary:
When optimizing for size, sqrt calls can be incorrectly selected as
AVX512 VSQRT instructions.  This is because X86InstrAVX512.td has a
`Requires<[OptForSize]>` in its `avx512_sqrt_scalar` multiclass
definition.  Even if the target does not support AVX512, the class can
apparently still be chosen, leading to an incorrect selection of
`vsqrtss`.

In PR26625, this lead to an assertion: Reg >= X86::FP0 && Reg <=
X86::FP6 && "Expected FP register!", because the `vsqrtss` instruction
requires an XMM register, which is not available on i686 CPUs.

Reviewers: grosbach, resistor, joker.eph

Subscribers: spatel, emaste, llvm-commits

Differential Revision: http://reviews.llvm.org/D17414
------------------------------------------------------------------------

git-svn-id: https://llvm.org/svn/llvm-project/llvm/branches/release_38@261367 91177308-0d34-0410-b5e6-96231b3b80d8
lib/Target/X86/X86InstrAVX512.td
test/CodeGen/X86/pr26625.ll [new file with mode: 0644]