[Bug target/14798] New: In case of SH target with -O2 option #pragma interrupt doesn't get resetted.

[Jim Wilson]

Anil Paranjape wrote:
> 	* sh/sh.c (sh_insert_attributes) : Flag pragma_interrupt is set to zero.

This patch has a number of problems.

It is only a partial fix.  If pragma interrupt is broken, then the other similar 
pragmas are all broken also.  We shouldn't fix just one of them while leaving 
the rest of them broken.  Anything using a global variable in sh.c is suspect. 
trap_exit, sp_switch, pragma_trapa, pragma_nosave_low_regs are all broken the 
exact same way.

The reason why they are all broken is because they assume parsing, tree 
generation, rtl generation, assembly generation all occur in order for one 
function at a time.  Note that the variables are cleared in 
sh_output_function_epilogue, which occurs at the end of assembly generation. 
This ordering is not true when -funit-at-a-time, which is now enabled by default 
at -O2.  This is also not true in the presence of nested functions, so this has 
always been broken.  It is just more obviously broken now.

current_function_interrupt is OK because it is set in expand_prologue and used 
in expand_epilogue, and these will always occur in sequence, without an 
intervening function.

current_function_anonymous_args is OK because it isn't used anywhere.  It should 
be deleted.

I think you are clearing pragma_interrupt in the wrong place in 
sh_insert_attributes.  I think it should be set only if an attribute is generated.

Since there are a number of places that use pragma_interrupt (the variable), 
clearing it in sh_insert_attributes will break that code.  Specifically, this 
breaks the trapa and sp_switch attributes.  This is trivial to show by adding an 
attribute to your testcase like so.
void isr() __attribute__ ((trap_exit (4)))
{
}
This works (sort of) with current sources, but with your patch I get
tmp.c:3: warning: `trap_exit' attribute only applies to interrupt functions
tmp.c:3: error: parse error before '{' token

These problems can all be solved by getting rid of the global variables, and 
using function calls to check for the presence of an attribute.  See for 
instance how the renesas attribute works.

Another alternative is to use variables in cfun, the current function context, 
instead of global variables.  Associating the variables with a function ensures 
that they get saved/restored along with the rest of the function context.  See 
for example any target that defines a struct machine_function type, along with a 
function to initialize it.
-- 
Jim Wilson, GNU Tools Support, http://www.SpecifixInc.com