GCC's middle-end provides a default cost model for RTL expressions, for
backends that don't specify their own instruction timings, that can be
summarized as multiplications are COSTS_N_INSNS(4), divisions are
COSTS_N_INSNS(7) and all other operations are COSTS_N_INSNS(1).
This patch tweaks the above definition so that fused-multiply-add
(FMA) and high-part multiplications cost the same as regular multiplications,
or more importantly aren't (by default) considered less expensive. Likewise
the saturating forms of multiplication and division cost the same as the
regular variants. These values can always be changed by the target, but
the goal is to avoid RTL expansion substituting a suitable operation with
its saturating equivalent because it (accidentally) looks much cheaper.
For example, PR 89845 is about implementing division/modulus via highpart
multiply, which may accidentally look extremely cheap.
2022-05-30 Roger Sayle <roger@nextmovesoftware.com>
gcc/ChangeLog
* rtlanal.cc (rtx_cost) <MULT>: Treat FMA, SS_MULT, US_MULT,
SMUL_HIGHPART and UMUL_HIGHPART as having the same cost as MULT.
<DIV>: Likewise, SS_DIV and US_DIV have the same default as DIV.
switch (code)
{
case MULT:
+ case FMA:
+ case SS_MULT:
+ case US_MULT:
+ case SMUL_HIGHPART:
+ case UMUL_HIGHPART:
/* Multiplication has time-complexity O(N*N), where N is the
number of units (translated from digits) when using
schoolbook long multiplication. */
case UDIV:
case MOD:
case UMOD:
+ case SS_DIV:
+ case US_DIV:
/* Similarly, complexity for schoolbook long division. */
total = factor * factor * COSTS_N_INSNS (7);
break;