[ARM64] Fix the cost model for cheap large constants.

Originally the cost model would give up for large constants and just return the
maximum cost. This is not what we want for constant hoisting, because some of
these constants are large in bitwidth, but are still cheap to materialize.

This commit fixes the cost model to either return TCC_Free if the cost cannot be
determined, or accurately calculate the cost even for large constants
(bitwidth > 128).

This fixes <rdar://problem/16591573>.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@206100 91177308-0d34-0410-b5e6-96231b3b80d8

1 files changed, 8 insertions, 0 deletions

diff --git a/test/Transforms/ConstantHoisting/ARM64/large-immediate.ll b/test/Transforms/ConstantHoisting/ARM64/large-immediate.ll
index 0965a45ea6..92ad6ee70d 100644
--- a/test/Transforms/ConstantHoisting/ARM64/large-immediate.ll
+++ b/test/Transforms/ConstantHoisting/ARM64/large-immediate.ll
@@ -8,3 +8,11 @@ define i128 @test1(i128 %a) nounwind {
   ret i128 %2
 }
 
+; Check that we don't hoist large, but cheap constants
+define i512 @test2(i512 %a) nounwind {
+; CHECK-LABEL: test2
+; CHECK-NOT: %const = bitcast i512 7 to i512
+  %1 = and i512 %a, 7
+  %2 = or i512 %1, 7
+  ret i512 %2
+}