fix PR10605 / rdar://9930964 by adding a pretty scary missed check.

It's somewhat surprising anything works without this.  Before we would
compile the testcase into:

test:                                   # @test
	movl	$4, 8(%rdi)
	movl	8(%rdi), %eax
	orl	%esi, %eax
	cmpl	$32, %edx
	movl	%eax, -4(%rsp)          # 4-byte Spill
	je	.LBB0_2

now we produce:

test:                                   # @test
	movl	8(%rdi), %eax
	movl	$4, 8(%rdi)
	orl	%esi, %eax
	cmpl	$32, %edx
	movl	%eax, -4(%rsp)          # 4-byte Spill
	je	.LBB0_2
llvm-svn: 137303

llvm/lib/CodeGen/SelectionDAG/SelectionDAGISel.cpp

@@ -754,6 +754,11 @@ bool SelectionDAGISel::TryToFoldFastISelLoad(const LoadInst *LI,
     TheUser = TheUser->use_back();
   }
   
+  // If we didn't find the fold instruction, then we failed to collapse the
+  // sequence.
+  if (TheUser != FoldInst)
+    return false;
+  
   // Don't try to fold volatile loads.  Target has to deal with alignment
   // constraints.
   if (LI->isVolatile()) return false;

llvm/test/CodeGen/X86/fast-isel-gep.ll

@@ -107,3 +107,30 @@ lpad:                                             ; preds = %if.end19, %if.then1
   unreachable
 }
 declare i8* @_ZNK18G__FastAllocString4dataEv() nounwind
+
+
+; PR10605 / rdar://9930964 - Don't fold loads incorrectly.  The load should
+; happen before the store.  
+define i32 @test7({i32,i32,i32}* %tmp1, i32 %tmp71, i32 %tmp63) nounwind  {
+; X64: test7:
+; X64:    movl	8({{%rdi|%rcx}}), %eax
+; X64     movl	$4, 8({{%rdi|%rcx}})
+
+
+  %tmp29 = getelementptr inbounds {i32,i32,i32}* %tmp1, i32 0, i32 2
+  %tmp30 = load i32* %tmp29, align 4
+
+  %p2 = getelementptr inbounds {i32,i32,i32}* %tmp1, i32 0, i32 2
+  store i32 4, i32* %p2
+  
+  %tmp72 = or i32 %tmp71, %tmp30
+  %tmp73 = icmp ne i32 %tmp63, 32
+  br i1 %tmp73, label %T, label %F
+
+T:
+  ret i32 %tmp72
+
+F:
+  ret i32 4
+}
+