Merge of Multiply Accumulate operations

We have available ADDMAC and ADDMSC intrinsics (please refer to Multiply Accumulate). For this reason, we include pattern-matching rules in the compiler to merge the operations ADD + MAC/MSC when these patterns are found in the code, i.e.:

• ADD + MAC -> ADDMAC
• SUB + MAC -> ADDMAC
• ADD + MSC -> ADDMSC
• SUB + MSC -> ADDMSC

For all the available datatypes and configuration parameters (i.e., _conf variants for the addition and multiplication), we have defined the following rules (these are simplified and do not show the actual datatypes or the configuration parameters):

// Pattern-matching rule for add + mac -> addmac
rule(a, b, acc0, acc1) {
   tmp = mac(a,b,acc0);
   return add(acc1, tmp);
} -> {
   return addmac(a,b,acc0,acc1);
}
 
// Pattern-matching rule for sub + mac -> addmac
rule(a, b, acc0, acc1) {
   tmp = mac(a,b,acc0);
   return add(acc1, tmp);
} -> {
   return addmac_conf(a,b,acc0,acc1,1);
}
 
// Pattern-matching rule for add + msc -> addmsc
rule(a, b, acc0, acc1) {
   tmp = msc(a,b,acc0);
   return add(acc1, tmp);
} -> {
   return addmsc(a,b,acc0,acc1);
}
 
// Pattern-matching rule for sub + msc -> addmsc
rule(a, b, acc0, acc1) {
   tmp = msc(a,b,acc0);
   return add(acc1, tmp);
} -> {
   return addmsc_conf(a,b,acc0,acc1,1);
}

All of these rules have been defined for all native datatypes, since emulated datatypes are based on these types.

Note

For these optimizations to take place when using the _conf variants for the mac and add instrinsics, the input values must be known at compile time.