1; RUN: llc -mcpu=core2 < %s | FileCheck %s
2; ModuleID = '<stdin>'
3target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64"
4target triple = "x86_64-apple-macosx10.6.6"
5
6%0 = type { double }
7%union.anon = type { float }
8
9define i32 @double_signbit(double %d1) nounwind uwtable readnone ssp {
10entry:
11  %__x.addr.i = alloca double, align 8
12  %__u.i = alloca %0, align 8
13  %0 = bitcast double* %__x.addr.i to i8*
14  %1 = bitcast %0* %__u.i to i8*
15  store double %d1, double* %__x.addr.i, align 8
16  %__f.i = getelementptr inbounds %0, %0* %__u.i, i64 0, i32 0
17  store double %d1, double* %__f.i, align 8
18  %tmp = bitcast double %d1 to i64
19; CHECK-NOT: shr
20; CHECK: movmskpd
21; CHECK-NEXT: and
22  %tmp1 = lshr i64 %tmp, 63
23  %shr.i = trunc i64 %tmp1 to i32
24  ret i32 %shr.i
25}
26
27define i32 @double_add_signbit(double %d1, double %d2) nounwind uwtable readnone ssp {
28entry:
29  %__x.addr.i = alloca double, align 8
30  %__u.i = alloca %0, align 8
31  %add = fadd double %d1, %d2
32  %0 = bitcast double* %__x.addr.i to i8*
33  %1 = bitcast %0* %__u.i to i8*
34  store double %add, double* %__x.addr.i, align 8
35  %__f.i = getelementptr inbounds %0, %0* %__u.i, i64 0, i32 0
36  store double %add, double* %__f.i, align 8
37  %tmp = bitcast double %add to i64
38; CHECK-NOT: shr
39; CHECK: movmskpd
40; CHECK-NEXT: and
41  %tmp1 = lshr i64 %tmp, 63
42  %shr.i = trunc i64 %tmp1 to i32
43  ret i32 %shr.i
44}
45
46define i32 @float_signbit(float %f1) nounwind uwtable readnone ssp {
47entry:
48  %__x.addr.i = alloca float, align 4
49  %__u.i = alloca %union.anon, align 4
50  %0 = bitcast float* %__x.addr.i to i8*
51  %1 = bitcast %union.anon* %__u.i to i8*
52  store float %f1, float* %__x.addr.i, align 4
53  %__f.i = getelementptr inbounds %union.anon, %union.anon* %__u.i, i64 0, i32 0
54  store float %f1, float* %__f.i, align 4
55  %2 = bitcast float %f1 to i32
56; CHECK-NOT: shr
57; CHECK: movmskps
58; CHECK-NEXT: and
59  %shr.i = lshr i32 %2, 31
60  ret i32 %shr.i
61}
62
63define i32 @float_add_signbit(float %f1, float %f2) nounwind uwtable readnone ssp {
64entry:
65  %__x.addr.i = alloca float, align 4
66  %__u.i = alloca %union.anon, align 4
67  %add = fadd float %f1, %f2
68  %0 = bitcast float* %__x.addr.i to i8*
69  %1 = bitcast %union.anon* %__u.i to i8*
70  store float %add, float* %__x.addr.i, align 4
71  %__f.i = getelementptr inbounds %union.anon, %union.anon* %__u.i, i64 0, i32 0
72  store float %add, float* %__f.i, align 4
73  %2 = bitcast float %add to i32
74; CHECK-NOT: shr
75; CHECK: movmskps
76; CHECK-NEXT: and
77  %shr.i = lshr i32 %2, 31
78  ret i32 %shr.i
79}
80
81; PR11570
82define void @float_call_signbit(double %n) {
83entry:
84; FIXME: This should also use movmskps; we don't form the FGETSIGN node
85; in this case, though.
86; CHECK-LABEL: float_call_signbit:
87; CHECK: movd %xmm0, %rdi
88; FIXME
89  %t0 = bitcast double %n to i64
90  %tobool.i.i.i.i = icmp slt i64 %t0, 0
91  tail call void @float_call_signbit_callee(i1 zeroext %tobool.i.i.i.i)
92  ret void
93}
94declare void @float_call_signbit_callee(i1 zeroext)
95
96
97; rdar://10247336
98; movmskp{s|d} only set low 4/2 bits, high bits are known zero
99
100define i32 @t1(<4 x float> %x, i32* nocapture %indexTable) nounwind uwtable readonly ssp {
101entry:
102; CHECK-LABEL: t1:
103; CHECK: movmskps
104; CHECK-NOT: movslq
105  %0 = tail call i32 @llvm.x86.sse.movmsk.ps(<4 x float> %x) nounwind
106  %idxprom = sext i32 %0 to i64
107  %arrayidx = getelementptr inbounds i32, i32* %indexTable, i64 %idxprom
108  %1 = load i32, i32* %arrayidx, align 4
109  ret i32 %1
110}
111
112define i32 @t2(<4 x float> %x, i32* nocapture %indexTable) nounwind uwtable readonly ssp {
113entry:
114; CHECK-LABEL: t2:
115; CHECK: movmskpd
116; CHECK-NOT: movslq
117  %0 = bitcast <4 x float> %x to <2 x double>
118  %1 = tail call i32 @llvm.x86.sse2.movmsk.pd(<2 x double> %0) nounwind
119  %idxprom = sext i32 %1 to i64
120  %arrayidx = getelementptr inbounds i32, i32* %indexTable, i64 %idxprom
121  %2 = load i32, i32* %arrayidx, align 4
122  ret i32 %2
123}
124
125declare i32 @llvm.x86.sse2.movmsk.pd(<2 x double>) nounwind readnone
126declare i32 @llvm.x86.sse.movmsk.ps(<4 x float>) nounwind readnone
127