Clang's call lowering often coerces pointers to ints, which blocks alias analysis

[rnk]

Bugzilla Link	31724
Version	unspecified
OS	Windows NT

Extended Description

Consider this C code:

struct Handle { void *p; }; extern void g(struct Handle); void f(struct Handle o) { g(o); }

Here are three examples of clang coercing the 'p' field to an int:

$ clang -S -x c --target=ppc64-linux t.cpp -o - -emit-llvm | grep define define void @​f(i64 %o.coerce) #​0 {

$ clang -S -x c --target=x86_64-windows t.cpp -o - -emit-llvm | grep define define void @​f(i64 %o.coerce) #​0 {

$ clang -S -x c --target=mips-linux t.cpp -o - -emit-llvm | grep define define void @​f(i32 inreg %o.coerce) #​0 {

According to dannyb, this is bad, because the ptrtoint calls block alias analysis.

[llvmbot]

(and by special, i mean "only escape/etc as much as a pointer would, and not 'complete lost to the void'", which is usually how ptrtoint is treated)

[llvmbot]

In general, we can't do anything with pointer to int'd pointers. We can't tell where they point to. In an IPA analysis, we can't easily join them up, etc.

We also assume they escape.

BasicAA only handles bitcasts, it doesn't go back through ptrtoint/inttoptr to see if they are the same thing, unmodified (and doing so is expensive), so even inlining won't help until the inttoptr/ptrtoint pair is eliminated.

There are a couple ways to solve this.

Truthfully, the best way is to make them pointers again if possible.

If that really can't be solved, we should add metadata saying they are special, and really pointers, and include the real pointer address expression somewhere, but this seems .. hard.