llvm-6502/test/CodeGen/X86/avx2-pmovx-256-old-shuffle.ll
Ahmed Bougacha 34288d885e [X86] Declare SSE4.1/AVX2 vector extloads covered by PMOV[SZ]X legal.
Now that we can fully specify extload legality, we can declare them
legal for the PMOVSX/PMOVZX instructions.  This for instance enables
a DAGCombine to fire on code such as
  (and (<zextload-equivalent> ...), <redundant mask>)
to turn it into:
  (zextload ...)
as seen in the testcase changes.

There is one regression, in widen_load-2.ll: we're no longer able
to do store-to-load forwarding with illegal extload memory types.
This will be addressed separately.

Differential Revision: http://reviews.llvm.org/D6533


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@226676 91177308-0d34-0410-b5e6-96231b3b80d8
2015-01-21 17:07:06 +00:00

28 lines
1.0 KiB
LLVM

; RUN: llc < %s -x86-experimental-vector-shuffle-lowering=false -mattr=+avx2 | FileCheck %s
target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
target triple = "x86_64-apple-darwin"
; PR21876
; The old shuffle lowering sometimes generates VZEXT nodes with both input
; and output same-sized types, here 256-bits. For instance, a v8i8 to v8i32
; zero-extend would become a (v8i32 (VZEXT v32i8)) node, which can't happen
; otherwise. The companion commit r223996 added those patterns temporarily.
; This test, along with the VR256 for AVX2 PMOVXrr instructions, should be
; removed once the old vector shuffle lowering goes away.
define void @test_avx2_pmovx_256(<8 x i8>* %tmp64, <8 x float>* %tmp75) {
; CHECK-LABEL: test_avx2_pmovx_256
; We really don't care about the generated code.
; CHECK: vpmovzxbd
; CHECK: vcvtdq2ps
; CHECK: vmovups
; CHECK: vzeroupper
; CHECK: retq
%wide.load458 = load <8 x i8>* %tmp64, align 1
%tmp68 = uitofp <8 x i8> %wide.load458 to <8 x float>
store <8 x float> %tmp68, <8 x float>* %tmp75, align 4
ret void
}