On Wed, Sep 10, 2025 at 5:12 PM Christophe Lyon via Gcc <gcc@gcc.gnu.org> wrote: > > Hi! > > I am looking for clarification about the permitted shift amount in RTL > (ashift, lshiftrt, ashiftrt). > > The doc in rtl.texi does not mention any constraints on the value of > 'c'. In particular, what is supposed to happen if c < 0 ? Although > it is UB at the C language level, arm targets have instructions > (e.g. asrl and lsll) which explicitly support negative shift amounts > in their "register" variant (if the shift amount is negative, the > shift direction is reversed). > > We recently fixed the asrl and lsll Arm MVE intrinsics implementation > in GCC to use the builtins instead of: > > return (value >> shift); > return (value << shift); > > respectively, because some user code was purposely using a negative > shift amount in order to shift in the opposite direction. With the > previous implementation, the compiler detected UB and did not use the > intended instruction.
I don't think RTL supports this, in the past we turned constant negative shift amounts to flip the direction early on GENERIC, but I'm not sure we're still doing this. I think rtl.texi should mention that the shift amount is interpreted unsigned. > > However, we have concerns that given > > (set (reg:SI 1) (const_int -5)) > (set (reg:DI 2) (ashift:DI (reg:DI 3) (reg:SI 1))) > > simplify_rtx could simplify this into > > (set (reg:DI 2) (ashift:DI (reg:DI 3) (const_int -5))) > > and consider this as invalid and do whatever it feels like? > > I do notice several places in simplify-rtx.cc checking that > UINTVAL (XEXP (op, 1)) < precision) > so it seems there is some assumption that the shift amount is within > the expected range. > > Does the doc need clarification on the semantics of RTL shift > operations when the shift amount is out of range? The SHIFT_COUNT_TRUNCATED target macro is also relevant in this context. On x86 which is a "mixed" ISA in this regard we make sure to recognize shifts with an (and ...) shift operand which is what the middle-end will add (the and) when !SHIFT_COUNT_TRUNCATED IIRC. Richard. > Thanks, > > Christophe
