On Thu, Aug 12, 2021 at 11:52 AM Roger Sayle <ro...@nextmovesoftware.com> wrote: > > > This patch improves the bit bounds for MINUS_EXPR during tree-ssa's > conditional constant propagation (CCP) pass (and as an added bonus > adds support for POINTER_DIFF_EXPR). > > The pessimistic assumptions made by the current algorithm are > demonstrated by considering 1 - (x&1). Intuitively this should > have possible values 0 and 1, and therefore an unknown mask of 1. > Alas by treating subtraction as a negation followed by addition, > the second operand first becomes 0 or -1, with an unknown mask > of all ones, which results in the addition containing no known bits. > > Improved bounds are achieved by using the same approach used for > PLUS_EXPR, determining the result with the minimum number of borrows, > the result from the maximum number of borrows, and examining the bits > they have in common. One additional benefit of this approach > is that it is applicable to POINTER_DIFF_EXPR, where previously the > negation of a pointer didn't/doesn't make sense. > > A more convincing example, where a transformation missed by .032t.cpp > isn't caught a few passes later by .038t.evrp, is the expression > (7 - (x&5)) & 2, which (in the new test case) currently survives the > tree-level optimizers but with this patch is now simplified to the > constant value 2. > > This patch has been tested on x86_64-pc-linux-gnu with "make bootstrap" > and "make -k check" with no new failures. > > Ok for mainline?
OK. Thanks, Richard. > 2021-08-12 Roger Sayle <ro...@nextmovesoftware.com> > > gcc/ChangeLog > * tree-ssa-ccp.c (bit_value_binop) [MINUS_EXPR]: Use same > algorithm as PLUS_EXPR to improve subtraction bit bounds. > [POINTER_DIFF_EXPR]: Treat as synonymous with MINUS_EXPR. > > gcc/testsuite/ChangeLog > * gcc.dg/tree-ssa/ssa-ccp-40.c: New test case. > > > Roger > -- > Roger Sayle > NextMove Software > Cambridge, UK >
