Fixing line numbers of inlined allocas

[Pieb, Wolfgang]

ping...

> -----Original Message-----
> From: Wolfgang Pieb [mailto:wolfgang_pieb at playstation.sony.com]
> Sent: Thursday, September 18, 2014 10:51 AM
> To: Pieb, Wolfgang
> Cc: llvm-commits at cs.uiuc.edu
> Subject: [PATCH] Fixing line numbers of inlined allocas
> 
> When a function is inlined (for example, as a result of it carrying the
> attribute alwaysinline), the cloned instructions receive the call
> site’s line number information (per r210459). After cloning, the cloned
> alloca instructions are moved to the caller’s entry block, retaining
> the call site’s line number information.
> When, subsequently, new instructions are generated at the start of the
> caller’s entry block (such as instructions generated by the stack-
> protector pass), they pick up the line number information from the
> entry block’s first instruction, which may be one of the cloned alloca
> instructions.
> 
> This can cause incorrect line number information being generated for
> the start of the calling function, with an inlined function’s call site
> being associated with the first stack-protector instruction, for
> example. Since the stack-protector instruction is likely deemed to be
> the end of the function prologue, breakpoints that are set using the
> function name will be set at the wrong location.
> 
> Example:
> 
> int __attribute__((always_inline)) foo() {
>    int arr[10];
>    arr[0] = 5;
>    int sum = 4;
>    return sum;
> }
> 
> extern void bar();
> 
> int main()
> {
>   bar();
>   int i = foo();   <== breaking at ‘main’ will stop here because main’s
> function prologue gets foo’s call site’s line number.
>   return i;
> }
> 
> 
> 
> The proposed fix is to modify the cloned alloca instructions’ debug
> information when they are moved to the caller’s entry block. Instead of
> the call site’s information they would get the same information as the
> instructions already at the start  of the caller’s entry block, i.e.
> they place where they’re being moved to.
> 
> Note to the test case: the sspstrong attribute on the 'main' function
> is not necessary to make the test case pass, but if one were to
> generate a .o file the incorrect line numbers for the function prologue
> would not appear without the stack-protector instructions being
> generated.
> 
> http://reviews.llvm.org/D5401
> 
> Files:
>   lib/Transforms/Utils/InlineFunction.cpp
>   test/DebugInfo/debug-info-always-inline.ll