In pr120811, we have cases where GCC is emitting an extra addi instruction
instead of using the 12-bit signed-immediate of ld.

addi t1, t1, 1
ld   t1, 0(t1)

This problem occurs when fp -> sp+offset elimination results in an
out-of-range constant and we generate an address reload in LRA using
addsi/adddi expanders.

We've already adjusted the expanders to widen the set of valid operands to
allow more constants for the 2nd input operand. These expanders, rather than
constructing the constant into a register and using an add instruction, will
generate two addi instructions (or shNadd) during initial RTL generation.

We define a new pattern for cases where we need to access the current frame
and the offsets are too large. This gets reasonable code out of LRA in a
form
fold-mem-offsets can handle, rather than having to wait for sched2 to do
the height reduction transformation and leaving in the unnecessary add
instruction in the RTL stream.

To avoid the two addi instructions being squashed back together in the
post-reload combine, we remove the adddi3_const_sum_of_two_s12 pattern.

We are seeing about 100 billion dynamic instructions saved which is about 5%
on cactuBSSN and a 2% improvement in performance on the BPI.

PR target/120811

gcc/

* config/riscv/riscv.cc (synthesize_add): Exchange constant terms when

generating addi pairs.

(synthesize_addsi): Similarly.

* config/riscv/riscv.md (addptr<mode>3): New define_expand.
(*add<mode>3_const_sum_of_two_s12): Remove pattern.


gcc/testsuite/

* gcc.target/riscv/add-synthesis-1.c: Adjust const to fit in range.

* gcc.target/riscv/pr120811.c: Add new test case.
* gcc.target/riscv/sum-of-two-s12-const-1.c: Adjust const to fit in range.

Attachment: 120811.patch
Description: Binary data

Reply via email to