• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1; RUN: llc < %s -relocation-model=pic -disable-fp-elim -mcpu=cortex-a8 -pre-RA-sched=source | FileCheck %s
2target triple = "thumbv7-apple-ios"
3; <rdar://problem/10032939>
4;
5; The vector %v2 is built like this:
6;
7;   %vreg6:ssub_1<def> = VMOVSR %vreg0<kill>, pred:14, pred:%noreg, %vreg6<imp-def>; DPR_VFP2:%vreg6 GPR:%vreg0
8;   %vreg6:ssub_0<def> = VLDRS <cp#0>, 0, pred:14, pred:%noreg; mem:LD4[ConstantPool] DPR_VFP2:%vreg6
9;
10; When %vreg6 spills, the VLDRS constant pool load cannot be rematerialized
11; since it implicitly reads the ssub_1 sub-register.
12;
13; CHECK: f1
14; CHECK: vmov    s1, r0
15; CHECK: vldr.32 s0, LCPI
16; The vector must be spilled:
17; CHECK: vstr.64 d0,
18; CHECK: asm clobber d0
19; And reloaded after the asm:
20; CHECK: vldr.64 [[D16:d[0-9]+]],
21; CHECK: vstr.64 [[D16]], [r1]
22define void @f1(float %x, <2 x float>* %p) {
23  %v1 = insertelement <2 x float> undef, float %x, i32 1
24  %v2 = insertelement <2 x float> %v1, float 0x400921FB60000000, i32 0
25  %y = call double asm sideeffect "asm clobber $0", "=w,0,~{d1},~{d2},~{d3},~{d4},~{d5},~{d6},~{d7},~{d8},~{d9},~{d10},~{d11},~{d12},~{d13},~{d14},~{d15},~{d16},~{d17},~{d18},~{d19},~{d20},~{d21},~{d22},~{d23},~{d24},~{d25},~{d26},~{d27},~{d28},~{d29},~{d30},~{d31}"(<2 x float> %v2) nounwind
26  store <2 x float> %v2, <2 x float>* %p, align 8
27  ret void
28}
29
30; On the other hand, when the partial redef doesn't read the full register
31; because the bits are undef, we should rematerialize.  The vector is now built
32; like this:
33;
34;   %vreg2:ssub_0<def> = VLDRS <cp#0>, 0, pred:14, pred:%noreg, %vreg2<imp-def>; mem:LD4[ConstantPool]
35;
36; The extra <imp-def> operand indicates that the instruction fully defines the
37; virtual register.  It doesn't read the old value.
38;
39; CHECK: f2
40; CHECK: vldr.32 s0, LCPI
41; The vector must not be spilled:
42; CHECK-NOT: vstr.64
43; CHECK: asm clobber d0
44; But instead rematerialize after the asm:
45; CHECK: vldr.32 [[S0:s[0-9]+]], LCPI
46; CHECK: vstr.64 [[D0:d[0-9]+]], [r0]
47define void @f2(<2 x float>* %p) {
48  %v2 = insertelement <2 x float> undef, float 0x400921FB60000000, i32 0
49  %y = call double asm sideeffect "asm clobber $0", "=w,0,~{d1},~{d2},~{d3},~{d4},~{d5},~{d6},~{d7},~{d8},~{d9},~{d10},~{d11},~{d12},~{d13},~{d14},~{d15},~{d16},~{d17},~{d18},~{d19},~{d20},~{d21},~{d22},~{d23},~{d24},~{d25},~{d26},~{d27},~{d28},~{d29},~{d30},~{d31}"(<2 x float> %v2) nounwind
50  store <2 x float> %v2, <2 x float>* %p, align 8
51  ret void
52}
53