ARM: prefer allocating VFP regs at stride 4 on Darwin.

This is necessary for WatchOS support, where the compact unwind format assumes
this kind of layout. For now we only want this on Swift-like CPUs though, where
it's been the Xcode behaviour for ages. Also, since it can expand the prologue
we don't want it at -Oz.

llvm-svn: 243884

2 files changed, 34 insertions, 4 deletions

diff --git a/llvm/test/CodeGen/ARM/fold-stack-adjust.ll b/llvm/test/CodeGen/ARM/fold-stack-adjust.ll
index b62d0dfee07..99174213562 100644
--- a/llvm/test/CodeGen/ARM/fold-stack-adjust.ll
+++ b/llvm/test/CodeGen/ARM/fold-stack-adjust.ll
@@ -60,8 +60,6 @@ define void @check_vfp_fold() minsize {
 ; CHECK: vpush {d6, d7, d8, d9}
 ; CHECK-NOT: sub sp,
 ; ...
-; CHECK: vldmia r[[GLOBREG]], {d8, d9}
-; ...
 ; CHECK-NOT: add sp,
 ; CHECK: vpop {d6, d7, d8, d9}
 ; CHECKL pop {r[[GLOBREG]], pc}
@@ -82,9 +80,8 @@ define void @check_vfp_fold() minsize {
 
   %var = alloca i8, i32 16
 
-  %tmp = load %bigVec, %bigVec* @var
+  call void asm "", "r,~{d8},~{d9}"(i8* %var)
   call void @bar(i8* %var)
-  store %bigVec %tmp, %bigVec* @var
 
   ret void
 }
diff --git a/llvm/test/CodeGen/ARM/vfp-reg-stride.ll b/llvm/test/CodeGen/ARM/vfp-reg-stride.ll
new file mode 100644
index 00000000000..5484cc810b0
--- /dev/null
+++ b/llvm/test/CodeGen/ARM/vfp-reg-stride.ll
@@ -0,0 +1,33 @@
+; RUN: llc -mcpu=swift -mtriple=thumbv7s-apple-ios -o - %s | FileCheck %s --check-prefix=CHECK-STRIDE4
+; RUN: llc -mcpu=cortex-a57 -mtriple=thumbv7-linux-gnueabihf -o - %s | FileCheck %s --check-prefix=CHECK-GENERIC
+
+define void @test_reg_stride(double %a, double %b) {
+; CHECK-STRIDE4-LABEL: test_reg_stride:
+; CHECK-STRIDE4-DAG: vmov d16, r
+; CHECK-STRIDE4-DAG: vmov d18, r
+
+; CHECK-GENERIC-LABEL: test_reg_stride:
+; CHECK-GENERIC-DAG: vmov.f64 d16, {{d[01]}}
+; CHECK-GENERIC-DAG: vmov.f64 d17, {{d[01]}}
+
+  call void asm "", "~{r0},~{r1},~{d0},~{d1}"()
+  call arm_aapcs_vfpcc void @eat_doubles(double %a, double %b)
+  ret void
+}
+
+define void @test_stride_minsize(float %a, float %b) minsize {
+; CHECK-STRIDE4-LABEL: test_stride_minsize:
+; CHECK-STRIDE4: vmov d2, {{r[01]}}
+; CHECK-STRIDE4: vmov d3, {{r[01]}}
+
+; CHECK-GENERIC-LABEL: test_stride_minsize:
+; CHECK-GENERIC-DAG: vmov.f32 s4, {{s[01]}}
+; CHECK-GENERIC-DAG: vmov.f32 s6, {{s[01]}}
+  call void asm "", "~{r0},~{r1},~{s0},~{s1},~{d0},~{d1}"()
+  call arm_aapcs_vfpcc void @eat_floats(float %a, float %b)
+  ret void
+}
+
+
+declare arm_aapcs_vfpcc void @eat_doubles(double, double)
+declare arm_aapcs_vfpcc void @eat_floats(float, float)