[clang] Reapply "[analyzer] Handle [[assume(cond)]] as __builtin_assume(cond)" (PR #129234)

[Balazs Benics via cfe-commits]

================
@@ -180,3 +183,58 @@ int test_reference_that_might_be_after_the_end(int idx) {
   return ref;
 }
 
+// From: https://github.com/llvm/llvm-project/issues/100762
+extern int arrOf10[10];
+void using_builtin(int x) {
+  __builtin_assume(x > 101); // CallExpr
+  arrOf10[x] = 404; // expected-warning {{Out of bound access to memory}}
+}
+
+void using_assume_attr(int ax) {
+  [[assume(ax > 100)]]; // NullStmt with an "assume" attribute.
+  arrOf10[ax] = 405; // expected-warning {{Out of bound access to memory}}
+}
+
+void using_many_assume_attr(int yx) {
+  [[assume(yx > 104), assume(yx > 200), assume(yx < 300)]]; // NullStmt with an attribute
+  arrOf10[yx] = 406; // expected-warning{{Out of bound access to memory}}
+}
+
+
+int using_builtin_assume_has_no_sideeffects(int y) {
+  // We should not apply sideeffects of the argument of [[assume(...)]].
+  // "y" should not get incremented;
+  __builtin_assume(++y == 43); // expected-warning {{assumption is ignored because it contains (potential) side-effects}}
+  clang_analyzer_eval(y == 42); // expected-warning {{FALSE}}
+  return y;
+}
+
+
+
+int using_assume_attr_has_no_sideeffects(int y) {
+
+  // We should not apply sideeffects of the argument of [[assume(...)]].
+  // "y" should not get incremented;
+  [[assume(++y == 43)]]; // expected-warning {{assumption is ignored because it contains (potential) side-effects}}
+ 
+  clang_analyzer_eval(y == 42); // expected-warning {{TRUE}} expected-warning {{FALSE}} FIXME: This should be only TRUE.
+
+  clang_analyzer_eval(y == 43); // expected-warning {{FALSE}} expected-warning {{TRUE}} FIXME: This should be only FALSE.
+
+  return y;
+}
+
+
+int using_builtinassume_has_no_sideeffects(int u) {
+  // We should not apply sideeffects of the argument of __builtin_assume(...)
+  // "u" should not get incremented;
+  __builtin_assume(++u == 43); // expected-warning {{assumption is ignored because it contains (potential) side-effects}}
----------------
steakhal wrote:

Actually, in the test above we had the right outcome: `y` remained opaque, with the same value before-and-after the assume attribute. That is good, and the eval just split the path.
I refined the test to make it more direct.

Speaking of the test with the builtin assume, it was slightly different.
Actually, in the CFG when we visited the CallExpr, we didn't check if the callee is a builtin assume, and we should have omitted the argument if it had sideeffects. Thanks for spotting this.
Now it should be fixed.

https://github.com/llvm/llvm-project/pull/129234