ARM64: initial backend import

This adds a second implementation of the AArch64 architecture to LLVM,
accessible in parallel via the "arm64" triple. The plan over the
coming weeks & months is to merge the two into a single backend,
during which time thorough code review should naturally occur.

Everything will be easier with the target in-tree though, hence this
commit.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@205090 91177308-0d34-0410-b5e6-96231b3b80d8

1 files changed, 40 insertions, 0 deletions

diff --git a/test/CodeGen/ARM64/2012-01-11-ComparisonDAGCrash.ll b/test/CodeGen/ARM64/2012-01-11-ComparisonDAGCrash.ll
new file mode 100644
index 0000000000..d47dbb2816
--- /dev/null
+++ b/test/CodeGen/ARM64/2012-01-11-ComparisonDAGCrash.ll
@@ -0,0 +1,40 @@
+; RUN: llc < %s -march=arm64
+
+; The target lowering for integer comparisons was replacing some DAG nodes
+; during operation legalization, which resulted in dangling pointers,
+; cycles in DAGs, and eventually crashes.  This is the testcase for
+; one of those crashes. (rdar://10653656)
+
+define void @test(i1 zeroext %IsArrow) nounwind ssp align 2 {
+entry:
+  br i1 undef, label %return, label %lor.lhs.false
+
+lor.lhs.false:
+  br i1 undef, label %return, label %if.end
+
+if.end:
+  %tmp.i = load i64* undef, align 8
+  %and.i.i.i = and i64 %tmp.i, -16
+  br i1 %IsArrow, label %if.else_crit_edge, label %if.end32
+
+if.else_crit_edge:
+  br i1 undef, label %if.end32, label %return
+
+if.end32:
+  %0 = icmp ult i32 undef, 3
+  %1 = zext i64 %tmp.i to i320
+  %.pn.v = select i1 %0, i320 128, i320 64
+  %.pn = shl i320 %1, %.pn.v
+  %ins346392 = or i320 %.pn, 0
+  store i320 %ins346392, i320* undef, align 8
+  br i1 undef, label %sw.bb.i.i, label %exit
+
+sw.bb.i.i:
+  unreachable
+
+exit:
+  unreachable
+
+return:
+  ret void
+}