On 9/13/26 1:57 PM, Dominic P wrote:
The x & (x >> (PREC - 1)) -> MIN (x, 0) rules leave the branchless
min/max idiom
t = (a - b) & ((a - b) >> 31); a -= t; b += t;
half folded. Add
Y + MIN (X - Y, 0) -> MIN (X, Y) X - MIN (X - Y, 0) -> MAX (X, Y)
MIN (X + Y, Y) - Y -> MIN (X, 0)
and the MAX forms. Each needs the inner operation not to overflow, which
gimple_match_no_overflow_p takes from the type or, on GIMPLE, from ranges.
There is no :s because the idiom uses the masked value twice and both
uses fold.
x264's median-of-three predictor (common/base.h) goes from 20 to 11
instructions on x86_64 at -O2.
Bootstrapped and regression tested on x86_64-pc-linux-gnu.
Assisted-by: Claude Opus 5 (Anthropic)
gcc/ChangeLog:
* generic-match-head.cc (gimple_match_no_overflow_p): New.
* gimple-match-head.cc (gimple_match_no_overflow_p): New.
* match.pd (Y + MIN (X - Y, 0), X - MIN (X - Y, 0), Y + MAX (X - Y, 0),
X - MAX (X - Y, 0), MIN (X + Y, Y) - Y, MAX (X + Y, Y) - Y): New
patterns.
gcc/testsuite/ChangeLog:
* gcc.dg/torture/minmax-swar-1.c: New test.
* gcc.dg/tree-ssa/minmax-swar-1.c: New test.
* gcc.dg/tree-ssa/minmax-swar-2.c: New test.
* gcc.dg/tree-ssa/minmax-swar-3.c: New test.
* gcc.dg/tree-ssa/minmax-swar-4.c: New test.
* gcc.dg/tree-ssa/minmax-swar-5.c: New test.
* gcc.dg/tree-ssa/minmax-swar-6.c: New test.
* gcc.dg/tree-ssa/minmax-swar-7.c: New test.
Signed-off-by: Dominic P<[email protected]>
And I've pushed this to the trunk as well. Thanks again!
jeff
