[llvm-commits] [PATCH] Initial work on vector-select and modified type legalization

[Rotem, Nadav]

Hi, 

I started working on implementing vector select in LLVM.  Codegen support for vector-select is very important for vectorizing compilers.  I followed Duncan's design proposal and implemented vector-integer promotion in the type-legalizer.  This required changes all over the codegen.  The work is not complete and there are several bugs and missing features, but I am now able to compile the basic vector-select code below.  Almost all of the LIT tests pass and a good number of vectorized programs run successfully. 

This is a big patch and I would like other LLVM developers to go over it and review it.

Thanks,
Nadav

define <4 x float> @vector_code(<4 x i64> %A, <4 x i64> %B, <4 x float> %R0, <4 x float> %R1 )  {
   %C = icmp eq <4 x i64> %A, %B
   %K = xor <4 x i1> <i1 1, i1 1, i1 1, i1 1>, %C
   %D = select <4 x i1> %K, <4 x float> %R1, <4 x float> %R0   
   ret <4 x float> %D
} 


pcmpeqq     XMM1, XMM3
pcmpeqq     XMM0, XMM2
shufps          XMM0, XMM1, -120
xorps            XMM0, XMMWORD PTR [LCPI0_0]
pslld             XMM0, 31
movdqa       XMM1, XMMWORD PTR [ESP + 4]
movdqa       XMM2, XMMWORD PTR [ESP + 20]
blendvps     XMM2, XMM1, %xmm0
movdqa       XMM0, XMM2
ret 

---------------------------------------------------------------------
Intel Israel (74) Limited

This e-mail and any attachments may contain confidential material for
the sole use of the intended recipient(s). Any review or distribution
by others is strictly prohibited. If you are not the intended
recipient, please contact the sender and delete all copies.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: legalize.diff
Type: application/octet-stream
Size: 92207 bytes
Desc: legalize.diff
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20110505/b4df8e94/attachment.obj>