TTI: Improve default costs for addrspacecast

For some reason multiple places need to do this, and the variant the
loop unroller and inliner use was not handling it.

Also, introduce a new wrapper to be slightly more precise, since on
AMDGPU some addrspacecasts are free, but not no-ops.

llvm-svn: 362436

1 files changed, 27 insertions, 6 deletions

diff --git a/llvm/test/Analysis/CostModel/AMDGPU/addrspacecast.ll b/llvm/test/Analysis/CostModel/AMDGPU/addrspacecast.ll
index ddb31482adc..f15ab500a9e 100644
--- a/llvm/test/Analysis/CostModel/AMDGPU/addrspacecast.ll
+++ b/llvm/test/Analysis/CostModel/AMDGPU/addrspacecast.ll
@@ -1,45 +1,66 @@
 ; RUN: opt -cost-model -analyze -mtriple=amdgcn-unknown-amdhsa -mcpu=kaveri < %s | FileCheck %s
 
-; CHECK: 'addrspacecast_global_to_flat'
+; CHECK-LABEL: 'addrspacecast_global_to_flat'
 ; CHECK: estimated cost of 0 for {{.*}} addrspacecast i8 addrspace(1)* %ptr to i8*
 define i8* @addrspacecast_global_to_flat(i8 addrspace(1)* %ptr) #0 {
   %cast = addrspacecast i8 addrspace(1)* %ptr to i8*
   ret i8* %cast
 }
 
-; CHECK: 'addrspacecast_global_to_flat_v2'
+; CHECK-LABEL: 'addrspacecast_global_to_flat_v2'
 ; CHECK: estimated cost of 0 for {{.*}} addrspacecast <2 x i8 addrspace(1)*> %ptr to <2 x i8*>
 define <2 x i8*> @addrspacecast_global_to_flat_v2(<2 x i8 addrspace(1)*> %ptr) #0 {
   %cast = addrspacecast <2 x i8 addrspace(1)*> %ptr to <2 x i8*>
   ret <2 x i8*> %cast
 }
 
-; CHECK: 'addrspacecast_global_to_flat_v32'
+; CHECK-LABEL: 'addrspacecast_global_to_flat_v32'
 ; CHECK: estimated cost of 0 for {{.*}} addrspacecast <32 x i8 addrspace(1)*> %ptr to <32 x i8*>
 define <32 x i8*> @addrspacecast_global_to_flat_v32(<32 x i8 addrspace(1)*> %ptr) #0 {
   %cast = addrspacecast <32 x i8 addrspace(1)*> %ptr to <32 x i8*>
   ret <32 x i8*> %cast
 }
 
-; CHECK: 'addrspacecast_local_to_flat'
+; CHECK-LABEL: 'addrspacecast_local_to_flat'
 ; CHECK: estimated cost of 1 for {{.*}} addrspacecast i8 addrspace(3)* %ptr to i8*
 define i8* @addrspacecast_local_to_flat(i8 addrspace(3)* %ptr) #0 {
   %cast = addrspacecast i8 addrspace(3)* %ptr to i8*
   ret i8* %cast
 }
 
-; CHECK: 'addrspacecast_local_to_flat_v2'
+; CHECK-LABEL: 'addrspacecast_local_to_flat_v2'
 ; CHECK: estimated cost of 2 for {{.*}} addrspacecast <2 x i8 addrspace(3)*> %ptr to <2 x i8*>
 define <2 x i8*> @addrspacecast_local_to_flat_v2(<2 x i8 addrspace(3)*> %ptr) #0 {
   %cast = addrspacecast <2 x i8 addrspace(3)*> %ptr to <2 x i8*>
   ret <2 x i8*> %cast
 }
 
-; CHECK: 'addrspacecast_local_to_flat_v32'
+; CHECK-LABEL: 'addrspacecast_local_to_flat_v32'
 ; CHECK: estimated cost of 32 for {{.*}} addrspacecast <32 x i8 addrspace(3)*> %ptr to <32 x i8*>
 define <32 x i8*> @addrspacecast_local_to_flat_v32(<32 x i8 addrspace(3)*> %ptr) #0 {
   %cast = addrspacecast <32 x i8 addrspace(3)*> %ptr to <32 x i8*>
   ret <32 x i8*> %cast
 }
 
+; CHECK-LABEL: 'addrspacecast_flat_to_local'
+; CHECK: estimated cost of 0 for {{.*}} addrspacecast i8* %ptr to i8 addrspace(3)*
+define i8 addrspace(3)* @addrspacecast_flat_to_local(i8* %ptr) #0 {
+  %cast = addrspacecast i8* %ptr to i8 addrspace(3)*
+  ret i8 addrspace(3)* %cast
+}
+
+; CHECK-LABEL: 'addrspacecast_flat_to_local_v2'
+; CHECK: estimated cost of 0 for {{.*}} addrspacecast <2 x i8*> %ptr to <2 x i8 addrspace(3)*>
+define <2 x i8 addrspace(3)*> @addrspacecast_flat_to_local_v2(<2 x i8*> %ptr) #0 {
+  %cast = addrspacecast <2 x i8*> %ptr to <2 x i8 addrspace(3)*>
+  ret <2 x i8 addrspace(3)*> %cast
+}
+
+; CHECK-LABEL: 'addrspacecast_flat_to_local_v32'
+; CHECK: estimated cost of 0 for {{.*}} addrspacecast <32 x i8*> %ptr to <32 x i8 addrspace(3)*>
+define <32 x i8 addrspace(3)*> @addrspacecast_flat_to_local_v32(<32 x i8*> %ptr) #0 {
+  %cast = addrspacecast <32 x i8*> %ptr to <32 x i8 addrspace(3)*>
+  ret <32 x i8 addrspace(3)*> %cast
+}
+
 attributes #0 = { nounwind readnone }