[AVX512] add PSLLW and PSLLV Intrinsic

[oota-llvm.git] / test / CodeGen / X86 / widen_cast-4.ll

diff --git a/test/CodeGen/X86/widen_cast-4.ll b/test/CodeGen/X86/widen_cast-4.ll
index 03dffc4842ce449776198b91bb3787fbd1fef000..8ed2785ae73a89ca50f42a415ba9b476ee749ca5 100644 (file)
--- a/test/CodeGen/X86/widen_cast-4.ll
+++ b/test/CodeGen/X86/widen_cast-4.ll
@@ -18,44 +18,44 @@ entry:
        br label %forcond
 
 forcond:               ; preds = %forinc, %entry
-       %tmp = load i32* %i             ; <i32> [#uses=1]
-       %tmp1 = load i32* %n.addr               ; <i32> [#uses=1]
+       %tmp = load i32, i32* %i                ; <i32> [#uses=1]
+       %tmp1 = load i32, i32* %n.addr          ; <i32> [#uses=1]
        %cmp = icmp slt i32 %tmp, %tmp1         ; <i1> [#uses=1]
        br i1 %cmp, label %forbody, label %afterfor
 
 forbody:               ; preds = %forcond
-; CHECK: %forbody
-; CHECK-WIDE: %forbody
-       %tmp2 = load i32* %i            ; <i32> [#uses=1]
-       %tmp3 = load i64** %dst_i.addr          ; <i64*> [#uses=1]
-       %arrayidx = getelementptr i64* %tmp3, i32 %tmp2         ; <i64*> [#uses=1]
+       %tmp2 = load i32, i32* %i               ; <i32> [#uses=1]
+       %tmp3 = load i64*, i64** %dst_i.addr            ; <i64*> [#uses=1]
+       %arrayidx = getelementptr i64, i64* %tmp3, i32 %tmp2            ; <i64*> [#uses=1]
        %conv = bitcast i64* %arrayidx to <8 x i8>*             ; <<8 x i8>*> [#uses=1]
        store <8 x i8>* %conv, <8 x i8>** %dst
-       %tmp4 = load i32* %i            ; <i32> [#uses=1]
-       %tmp5 = load i64** %src_i.addr          ; <i64*> [#uses=1]
-       %arrayidx6 = getelementptr i64* %tmp5, i32 %tmp4                ; <i64*> [#uses=1]
+       %tmp4 = load i32, i32* %i               ; <i32> [#uses=1]
+       %tmp5 = load i64*, i64** %src_i.addr            ; <i64*> [#uses=1]
+       %arrayidx6 = getelementptr i64, i64* %tmp5, i32 %tmp4           ; <i64*> [#uses=1]
        %conv7 = bitcast i64* %arrayidx6 to <8 x i8>*           ; <<8 x i8>*> [#uses=1]
        store <8 x i8>* %conv7, <8 x i8>** %src
-       %tmp8 = load i32* %i            ; <i32> [#uses=1]
-       %tmp9 = load <8 x i8>** %dst            ; <<8 x i8>*> [#uses=1]
-       %arrayidx10 = getelementptr <8 x i8>* %tmp9, i32 %tmp8          ; <<8 x i8>*> [#uses=1]
-       %tmp11 = load i32* %i           ; <i32> [#uses=1]
-       %tmp12 = load <8 x i8>** %src           ; <<8 x i8>*> [#uses=1]
-       %arrayidx13 = getelementptr <8 x i8>* %tmp12, i32 %tmp11                ; <<8 x i8>*> [#uses=1]
-       %tmp14 = load <8 x i8>* %arrayidx13             ; <<8 x i8>> [#uses=1]
+       %tmp8 = load i32, i32* %i               ; <i32> [#uses=1]
+       %tmp9 = load <8 x i8>*, <8 x i8>** %dst         ; <<8 x i8>*> [#uses=1]
+       %arrayidx10 = getelementptr <8 x i8>, <8 x i8>* %tmp9, i32 %tmp8                ; <<8 x i8>*> [#uses=1]
+       %tmp11 = load i32, i32* %i              ; <i32> [#uses=1]
+       %tmp12 = load <8 x i8>*, <8 x i8>** %src                ; <<8 x i8>*> [#uses=1]
+       %arrayidx13 = getelementptr <8 x i8>, <8 x i8>* %tmp12, i32 %tmp11              ; <<8 x i8>*> [#uses=1]
+       %tmp14 = load <8 x i8>, <8 x i8>* %arrayidx13           ; <<8 x i8>> [#uses=1]
        %add = add <8 x i8> %tmp14, < i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1 >          ; <<8 x i8>> [#uses=1]
        %shr = ashr <8 x i8> %add, < i8 2, i8 2, i8 2, i8 2, i8 2, i8 2, i8 2, i8 2 >           ; <<8 x i8>> [#uses=1]
        store <8 x i8> %shr, <8 x i8>* %arrayidx10
        br label %forinc
+; CHECK: %forbody
 ; CHECK:      pmovzxbw
 ; CHECK-NEXT: paddw
 ; CHECK-NEXT: psllw $8
 ; CHECK-NEXT: psraw $8
 ; CHECK-NEXT: psraw $2
 ; CHECK-NEXT: pshufb
-; CHECK-NEXT: movlpd
+; CHECK-NEXT: movq
 ;
 ; FIXME: We shouldn't require both a movd and an insert.
+; CHECK-WIDE: %forbody
 ; CHECK-WIDE:      movd
 ; CHECK-WIDE-NEXT: pinsrd
 ; CHECK-WIDE-NEXT: paddb
@@ -67,7 +67,7 @@ forbody:              ; preds = %forcond
 ; CHECK-WIDE-NEXT: movd
 
 forinc:                ; preds = %forbody
-       %tmp15 = load i32* %i           ; <i32> [#uses=1]
+       %tmp15 = load i32, i32* %i              ; <i32> [#uses=1]
        %inc = add i32 %tmp15, 1                ; <i32> [#uses=1]
        store i32 %inc, i32* %i
        br label %forcond