[llvm] [X86] Elect to tail call when `sret` ptr is passed to the callee (PR #146575)

[Reid Kleckner via llvm-commits]

================
@@ -725,27 +662,17 @@ define fastcc void @t21_sret_to_sret_more_args2(ptr noalias sret(%struct.foo) %a
 ;
 ; X64-LABEL: t21_sret_to_sret_more_args2:
 ; X64:       # %bb.0:
-; X64-NEXT:    pushq %rbx
 ; X64-NEXT:    movl %esi, %eax
-; X64-NEXT:    movq %rdi, %rbx
 ; X64-NEXT:    movl %edx, %esi
 ; X64-NEXT:    movl %eax, %edx
-; X64-NEXT:    callq f_sret at PLT
-; X64-NEXT:    movq %rbx, %rax
-; X64-NEXT:    popq %rbx
-; X64-NEXT:    retq
+; X64-NEXT:    jmp f_sret at PLT # TAILCALL
 ;
 ; X32-LABEL: t21_sret_to_sret_more_args2:
 ; X32:       # %bb.0:
-; X32-NEXT:    pushq %rbx
 ; X32-NEXT:    movl %esi, %eax
-; X32-NEXT:    movq %rdi, %rbx
 ; X32-NEXT:    movl %edx, %esi
 ; X32-NEXT:    movl %eax, %edx
-; X32-NEXT:    callq f_sret at PLT
-; X32-NEXT:    movl %ebx, %eax
-; X32-NEXT:    popq %rbx
-; X32-NEXT:    retq
+; X32-NEXT:    jmp f_sret at PLT # TAILCALL
----------------
rnk wrote:

Please add new whitebox test cases that try to fool your heuristic. The C code that comes to mind for me looks like:
```
typedef struct SRet {
  long x, y, z;
} SRet;
SRet foo();
SRet bar(SRet *p) {
  SRet r = {0, 0, 0};
  *p = foo();
  return r;
}
```

The call to `foo` could end up in the tail position, but the implicit sret arg is not forwarded to `foo` here.

https://github.com/llvm/llvm-project/pull/146575