1; RUN: llc -mtriple=x86_64-linux -mcpu=nehalem < %s | FileCheck %s --check-prefix=LIN
2; RUN: llc -mtriple=x86_64-win32 -mcpu=nehalem < %s | FileCheck %s --check-prefix=WIN
3; RUN: llc -mtriple=i686-win32 -mcpu=nehalem < %s | FileCheck %s --check-prefix=LIN32
4; rdar://7398554
5
6; When doing vector gather-scatter index calculation with 32-bit indices,
7; use an efficient mov/shift sequence rather than shuffling each individual
8; element out of the index vector.
9
10; CHECK-LABEL: foo:
11; LIN: movdqa	(%rsi), %xmm0
12; LIN: pand 	(%rdx), %xmm0
13; LIN: pextrq	$1, %xmm0, %r[[REG4:.+]]
14; LIN: movd 	%xmm0, %r[[REG2:.+]]
15; LIN: movslq	%e[[REG2]], %r[[REG1:.+]]
16; LIN: sarq    $32, %r[[REG2]]
17; LIN: movslq	%e[[REG4]], %r[[REG3:.+]]
18; LIN: sarq    $32, %r[[REG4]]
19; LIN: movsd	(%rdi,%r[[REG1]],8), %xmm0
20; LIN: movhpd	(%rdi,%r[[REG2]],8), %xmm0
21; LIN: movsd	(%rdi,%r[[REG3]],8), %xmm1
22; LIN: movhpd	(%rdi,%r[[REG4]],8), %xmm1
23
24; WIN: movdqa	(%rdx), %xmm0
25; WIN: pand 	(%r8), %xmm0
26; WIN: pextrq	$1, %xmm0, %r[[REG4:.+]]
27; WIN: movd 	%xmm0, %r[[REG2:.+]]
28; WIN: movslq	%e[[REG2]], %r[[REG1:.+]]
29; WIN: sarq    $32, %r[[REG2]]
30; WIN: movslq	%e[[REG4]], %r[[REG3:.+]]
31; WIN: sarq    $32, %r[[REG4]]
32; WIN: movsd	(%rcx,%r[[REG1]],8), %xmm0
33; WIN: movhpd	(%rcx,%r[[REG2]],8), %xmm0
34; WIN: movsd	(%rcx,%r[[REG3]],8), %xmm1
35; WIN: movhpd	(%rcx,%r[[REG4]],8), %xmm1
36
37define <4 x double> @foo(double* %p, <4 x i32>* %i, <4 x i32>* %h) nounwind {
38  %a = load <4 x i32>, <4 x i32>* %i
39  %b = load <4 x i32>, <4 x i32>* %h
40  %j = and <4 x i32> %a, %b
41  %d0 = extractelement <4 x i32> %j, i32 0
42  %d1 = extractelement <4 x i32> %j, i32 1
43  %d2 = extractelement <4 x i32> %j, i32 2
44  %d3 = extractelement <4 x i32> %j, i32 3
45  %q0 = getelementptr double, double* %p, i32 %d0
46  %q1 = getelementptr double, double* %p, i32 %d1
47  %q2 = getelementptr double, double* %p, i32 %d2
48  %q3 = getelementptr double, double* %p, i32 %d3
49  %r0 = load double, double* %q0
50  %r1 = load double, double* %q1
51  %r2 = load double, double* %q2
52  %r3 = load double, double* %q3
53  %v0 = insertelement <4 x double> undef, double %r0, i32 0
54  %v1 = insertelement <4 x double> %v0, double %r1, i32 1
55  %v2 = insertelement <4 x double> %v1, double %r2, i32 2
56  %v3 = insertelement <4 x double> %v2, double %r3, i32 3
57  ret <4 x double> %v3
58}
59
60; Check that the sequence previously used above, which bounces the vector off the
61; cache works for x86-32. Note that in this case it will not be used for index
62; calculation, since indexes are 32-bit, not 64.
63; CHECK-LABEL: old:
64; LIN32: movaps	%xmm0, (%esp)
65; LIN32-DAG: {{(mov|and)}}l	(%esp),
66; LIN32-DAG: {{(mov|and)}}l	4(%esp),
67; LIN32-DAG: {{(mov|and)}}l	8(%esp),
68; LIN32-DAG: {{(mov|and)}}l	12(%esp),
69define <4 x i64> @old(double* %p, <4 x i32>* %i, <4 x i32>* %h, i64 %f) nounwind {
70  %a = load <4 x i32>, <4 x i32>* %i
71  %b = load <4 x i32>, <4 x i32>* %h
72  %j = and <4 x i32> %a, %b
73  %d0 = extractelement <4 x i32> %j, i32 0
74  %d1 = extractelement <4 x i32> %j, i32 1
75  %d2 = extractelement <4 x i32> %j, i32 2
76  %d3 = extractelement <4 x i32> %j, i32 3
77  %q0 = zext i32 %d0 to i64
78  %q1 = zext i32 %d1 to i64
79  %q2 = zext i32 %d2 to i64
80  %q3 = zext i32 %d3 to i64
81  %r0 = and i64 %q0, %f
82  %r1 = and i64 %q1, %f
83  %r2 = and i64 %q2, %f
84  %r3 = and i64 %q3, %f
85  %v0 = insertelement <4 x i64> undef, i64 %r0, i32 0
86  %v1 = insertelement <4 x i64> %v0, i64 %r1, i32 1
87  %v2 = insertelement <4 x i64> %v1, i64 %r2, i32 2
88  %v3 = insertelement <4 x i64> %v2, i64 %r3, i32 3
89  ret <4 x i64> %v3
90}
91