1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
|
; RUN: llc < %s -mcpu=cortex-a8 -verify-machineinstrs -disable-block-placement | FileCheck %s
target triple = "thumbv7-apple-ios"
; The 1.0e+10 constant is loaded from the constant pool and kept in a register.
; CHECK: %entry
; CHECK: vldr s
; The float loop variable is initialized with a vmovs from the constant register.
; The vmovs is first widened to a vmovd, and then converted to a vorr because of the v2f32 vadd.f32.
; CHECK: vorr [[DL:d[0-9]+]], [[DN:d[0-9]+]]
; CHECK: , [[DN]]
; CHECK: %for.body.i
; CHECK: vadd.f32 [[DL]], [[DL]], [[DN]]
; CHECK: %rInnerproduct.exit
;
; This test is verifying:
; - The VMOVS widening is happening.
; - Register liveness is verified.
; - The execution domain switch to vorr works across basic blocks.
define void @Mm(i32 %in, ptr %addr) nounwind {
entry:
br label %for.body4
for.body4:
br label %for.body.i
for.body.i:
%tmp3.i = phi float [ 1.000000e+10, %for.body4 ], [ %add.i, %for.body.i ]
%add.i = fadd float %tmp3.i, 1.000000e+10
%exitcond.i = icmp eq i32 %in, 41
br i1 %exitcond.i, label %rInnerproduct.exit, label %for.body.i
rInnerproduct.exit:
store float %add.i, ptr %addr, align 4
br label %for.body4
}
|
