It's not necessary to do rounding for alloca operations when the requested
alignment is equal to the stack alignment.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40004 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/test/CodeGen/Generic/badFoldGEP.ll b/test/CodeGen/Generic/badFoldGEP.ll
new file mode 100644
index 0000000..63acd46
--- /dev/null
+++ b/test/CodeGen/Generic/badFoldGEP.ll
@@ -0,0 +1,29 @@
+; RUN: llvm-upgrade %s | llvm-as | llc
+
+;; GetMemInstArgs() folded the two getElementPtr instructions together,
+;; producing an illegal getElementPtr. That's because the type generated
+;; by the last index for the first one is a structure field, not an array
+;; element, and the second one indexes off that structure field.
+;; The code is legal but not type-safe and the two GEPs should not be folded.
+;;
+;; This code fragment is from Spec/CINT2000/197.parser/197.parser.bc,
+;; file post_process.c, function build_domain().
+;; (Modified to replace store with load and return load value.)
+;;
+
+%Domain = type { sbyte*, int, int*, int, int, int*, %Domain* }
+%domain_array = uninitialized global [497 x %Domain]
+
+implementation; Functions:
+
+declare void %opaque([497 x %Domain]*)
+
+int %main(int %argc, sbyte** %argv) {
+bb0: ;[#uses=0]
+ call void %opaque([497 x %Domain]* %domain_array)
+ %cann-indvar-idxcast = cast int %argc to long
+ %reg841 = getelementptr [497 x %Domain]* %domain_array, long 0, long %cann-indvar-idxcast, uint 3
+ %reg846 = getelementptr int* %reg841, long 1
+ %reg820 = load int* %reg846
+ ret int %reg820
+}