ARM cost model: Add test case to make sure we would notice a change in CodeGen

In r176898 I updated the cost model to reflect the fact that sext/zext/cast on
v8i32 <-> v8i8 and v16i32 <-> v16i8 are expensive.

This test case is so that we make sure to update the cost model once we fix
CodeGen.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@176955 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/test/CodeGen/ARM/vcvt.ll b/test/CodeGen/ARM/vcvt.ll
index c078f493094bea3bbb08e804b0fd2f2b32f9932e..72d36456fd25abe60fa2120bff70125d1bca3ba9 100644 (file)
--- a/test/CodeGen/ARM/vcvt.ll
+++ b/test/CodeGen/ARM/vcvt.ll
@@ -156,3 +156,156 @@ define <4 x i16> @vcvt_f32tof16(<4 x float>* %A) nounwind {
 
 declare <4 x float> @llvm.arm.neon.vcvthf2fp(<4 x i16>) nounwind readnone
 declare <4 x i16> @llvm.arm.neon.vcvtfp2hf(<4 x float>) nounwind readnone
+
+; We currently estimate the cost of sext/zext/trunc v8(v16)i32 <-> v8(v16)i8
+; instructions as expensive. If lowering is improved the cost model needs to
+; change.
+; RUN: opt < %s  -cost-model -analyze -mtriple=thumbv7-apple-ios6.0.0 -march=arm -mcpu=cortex-a8 | FileCheck %s --check-prefix=COST
+%T0_5 = type <8 x i8>
+%T1_5 = type <8 x i32>
+; CHECK: func_cvt5:
+define void @func_cvt5(%T0_5* %loadaddr, %T1_5* %storeaddr) {
+; CHECK: strh
+; CHECK: strh
+; CHECK: strh
+; CHECK: strh
+; CHECK: strh
+; CHECK: strh
+; CHECK: strh
+; CHECK: strh
+  %v0 = load %T0_5* %loadaddr
+; COST: func_cvt5
+; COST: cost of 24 {{.*}} sext
+  %r = sext %T0_5 %v0 to %T1_5
+  store %T1_5 %r, %T1_5* %storeaddr
+  ret void
+}
+;; We currently estimate the cost of this instruction as expensive. If lowering
+;; is improved the cost needs to change.
+%TA0_5 = type <8 x i8>
+%TA1_5 = type <8 x i32>
+; CHECK: func_cvt1:
+define void @func_cvt1(%TA0_5* %loadaddr, %TA1_5* %storeaddr) {
+; CHECK: strh
+; CHECK: strh
+; CHECK: strh
+; CHECK: strh
+; CHECK: strh
+; CHECK: strh
+; CHECK: strh
+; CHECK: strh
+  %v0 = load %TA0_5* %loadaddr
+; COST: func_cvt1
+; COST: cost of 22 {{.*}} zext
+  %r = zext %TA0_5 %v0 to %TA1_5
+  store %TA1_5 %r, %TA1_5* %storeaddr
+  ret void
+}
+;; We currently estimate the cost of this instruction as expensive. If lowering
+;; is improved the cost needs to change.
+%T0_51 = type <8 x i32>
+%T1_51 = type <8 x i8>
+; CHECK: func_cvt51:
+define void @func_cvt51(%T0_51* %loadaddr, %T1_51* %storeaddr) {
+; CHECK: strb
+; CHECK: strb
+; CHECK: strb
+; CHECK: strb
+; CHECK: strb
+; CHECK: strb
+; CHECK: strb
+; CHECK: strb
+  %v0 = load %T0_51* %loadaddr
+; COST: func_cvt51
+; COST: cost of 19 {{.*}} trunc
+  %r = trunc %T0_51 %v0 to %T1_51
+  store %T1_51 %r, %T1_51* %storeaddr
+  ret void
+}
+;; We currently estimate the cost of this instruction as expensive. If lowering
+;; is improved the cost needs to change.
+%TT0_5 = type <16 x i8>
+%TT1_5 = type <16 x i32>
+; CHECK: func_cvt52:
+define void @func_cvt52(%TT0_5* %loadaddr, %TT1_5* %storeaddr) {
+; CHECK: strh
+; CHECK: strh
+; CHECK: strh
+; CHECK: strh
+; CHECK: strh
+; CHECK: strh
+; CHECK: strh
+; CHECK: strh
+; CHECK: strh
+; CHECK: strh
+; CHECK: strh
+; CHECK: strh
+; CHECK: strh
+; CHECK: strh
+; CHECK: strh
+; CHECK: strh
+  %v0 = load %TT0_5* %loadaddr
+; COST: func_cvt52
+; COST: cost of 48 {{.*}} sext
+  %r = sext %TT0_5 %v0 to %TT1_5
+  store %TT1_5 %r, %TT1_5* %storeaddr
+  ret void
+}
+;; We currently estimate the cost of this instruction as expensive. If lowering
+;; is improved the cost needs to change.
+%TTA0_5 = type <16 x i8>
+%TTA1_5 = type <16 x i32>
+; CHECK: func_cvt12:
+define void @func_cvt12(%TTA0_5* %loadaddr, %TTA1_5* %storeaddr) {
+; CHECK: strh
+; CHECK: strh
+; CHECK: strh
+; CHECK: strh
+; CHECK: strh
+; CHECK: strh
+; CHECK: strh
+; CHECK: strh
+; CHECK: strh
+; CHECK: strh
+; CHECK: strh
+; CHECK: strh
+; CHECK: strh
+; CHECK: strh
+; CHECK: strh
+; CHECK: strh
+  %v0 = load %TTA0_5* %loadaddr
+; COST: func_cvt12
+; COST: cost of 44 {{.*}} zext
+  %r = zext %TTA0_5 %v0 to %TTA1_5
+  store %TTA1_5 %r, %TTA1_5* %storeaddr
+  ret void
+}
+;; We currently estimate the cost of this instruction as expensive. If lowering
+;; is improved the cost needs to change.
+%TT0_51 = type <16 x i32>
+%TT1_51 = type <16 x i8>
+; CHECK: func_cvt512:
+define void @func_cvt512(%TT0_51* %loadaddr, %TT1_51* %storeaddr) {
+; CHECK: strb
+; CHECK: strb
+; CHECK: strb
+; CHECK: strb
+; CHECK: strb
+; CHECK: strb
+; CHECK: strb
+; CHECK: strb
+; CHECK: strb
+; CHECK: strb
+; CHECK: strb
+; CHECK: strb
+; CHECK: strb
+; CHECK: strb
+; CHECK: strb
+; CHECK: strb
+  %v0 = load %TT0_51* %loadaddr
+; COST: func_cvt512
+; COST: cost of 38 {{.*}} trunc
+  %r = trunc %TT0_51 %v0 to %TT1_51
+  store %TT1_51 %r, %TT1_51* %storeaddr
+  ret void
+}