On Wed, 25 May 2011, Andrew Stubbs wrote: > I know that, potentially, not all shifted operands are going to be widening > multiplies, but I *think* this should be safe because other random shift > values are unlikely to match a real widening mult instruction (and if they do > then the code would already be broken). If somebody knows a reason why this > isn't safe then I think I'm going to need some help figuring out what > conditions to use.
Random supposition like that is not a sensible basis for modifying GCC. I haven't managed to produce an example of code demonstrating the problem, but that's probably because I'm not sufficiently familiar with all the RTL optimizers. Where is the guarantee that the inputs to these functions must represent real instructions, or that the outputs will only be used if they represent real instructions? Where are the assertions to ensure that wrong code is not quietly generated if this is not the case? Where is the documentation of what instruction patterns it is not permitted to put in .md files because they would violate the assumptions about what instructions you are permitted to represent in RTL? How have you checked there are no existing problematic instruction patterns? RTL has defined abstract semantics and RTL transformations should be ones that are valid in accordance with those semantics, with proper assertions if there are additional constraints on the input passed to a function. This means actually counting the numbers of variable bits in the operands to determine whether the multiplication could overflow. -- Joseph S. Myers jos...@codesourcery.com
