[PATCH] [X86] Refactor PMOV[SZ]Xrm to add missing AVX2 patterns.

[Demikhovsky, Elena]

I re-checked all patterns again. I think you can commit the patch.
LGTM.

Thanks

-  Elena


-----Original Message-----
From: Ahmed Bougacha [mailto:ahmed.bougacha at gmail.com] 
Sent: Wednesday, December 03, 2014 01:37
To: ahmed.bougacha at gmail.com; craig.topper at gmail.com; qcolombet at apple.com; anemet at apple.com; Demikhovsky, Elena
Cc: llvm-commits at cs.uiuc.edu
Subject: Re: [PATCH] [X86] Refactor PMOV[SZ]Xrm to add missing AVX2 patterns.

Hi Elena!
So, any thoughts on the sse41.ll tests and associated patterns?

To recap:
- the vzmovl patterns fire on code with scalar loads:

```
	%0 = load i64* %p
	%tmp2 = insertelement <2 x i64> zeroinitializer, i64 %0, i32 0
	%1 = bitcast <2 x i64> %tmp2 to <8 x i16>
	%2 = call <4 x i32> @llvm.x86.sse41.pmovsxwd(<8 x i16> %1) ```
- the vzload patterns fire on the same code, but in 32-bit mode
- the loadf64 patterns fire on vector loads of a 64-bit type, in 32-bit mode (where load f64 is legal, but load i64 isn't):

```
 %X = load <2 x i32>* %ptr
 %Y = sext <2 x i32> %X to <2 x i64>
```

The other patterns cover the expected scalar load + extension, or vector load + extension (for intrinsics).

-Ahmed

http://reviews.llvm.org/D6125


---------------------------------------------------------------------
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.