Consider the following (RISC-V) instruction pair:
mul s6,a1,a2
add s6,a4,s6
Without this patch, while handling the second instruction, (a) the
existing chain for s6 will first be closed (upon the terminate_write
action for the input operand), then (b) a new one will be opened (upon
the mark_write action for the output operand). This will likely lead to
the two output operands being different physical registers, breaking the
single-output property required for some macro-op fusion pairs.
This patch, using the single_output_fused_pair_p () predicate introduced
earlier, changes the regrename behavior for such pairs to append the
input and the output operands to the existing chain (as if both actions
were mark_read), instead of breaking the current renaming chain and
starting a new one. This ensures that the output operands of both fused
instructions are kept in the same hard register, and that the
single-output property of the insn pair is preserved.
gcc/ChangeLog:
* regrename.cc (scan_rtx_reg): Handle fused insn pairs.
unsigned this_regno = REGNO (x);
int this_nregs = REG_NREGS (x);
+ /* Do not process write actions for the second instruction of
+ a macro-fused pair of two single_sets. */
+ if ((action == mark_write || action == terminate_write)
+ && single_output_fused_pair_p (insn))
+ return;
+
if (action == mark_write)
{
if (type == OP_OUT)
return;
}
- if ((type == OP_OUT) != (action == terminate_write || action == mark_access))
+ if ((type == OP_OUT) != (action == terminate_write || action == mark_access)
+ && ! (type == OP_OUT && action == mark_read
+ && single_output_fused_pair_p (insn)))
return;
for (p = &open_chains; *p;)