[all-commits] [llvm/llvm-project] dae5c4: [RISCV] Expand constant multiplication for targets...

[Iris Shi via All-commits]

  Branch: refs/heads/main
  Home:   https://github.com/llvm/llvm-project
  Commit: dae5c4e1e7deef1fddad6e83d0f109b6fc9c5cc4
      https://github.com/llvm/llvm-project/commit/dae5c4e1e7deef1fddad6e83d0f109b6fc9c5cc4
  Author: Iris Shi <0.0 at owo.li>
  Date:   2025-05-16 (Fri, 16 May 2025)

  Changed paths:
    M llvm/lib/Target/RISCV/RISCVISelLowering.cpp
    M llvm/test/CodeGen/RISCV/ctlz-cttz-ctpop.ll
    M llvm/test/CodeGen/RISCV/ctz_zero_return_test.ll
    M llvm/test/CodeGen/RISCV/mul-expand.ll
    M llvm/test/CodeGen/RISCV/mul.ll
    M llvm/test/CodeGen/RISCV/rv64xtheadbb.ll
    M llvm/test/CodeGen/RISCV/rv64zbb.ll
    M llvm/test/CodeGen/RISCV/rvv/known-never-zero.ll
    M llvm/test/CodeGen/RISCV/srem-seteq-illegal-types.ll
    M llvm/test/CodeGen/RISCV/urem-seteq-illegal-types.ll
    M llvm/test/CodeGen/RISCV/xqccmp-additional-stack.ll
    M llvm/test/CodeGen/RISCV/zcmp-additional-stack.ll

  Log Message:
  -----------
  [RISCV] Expand constant multiplication for targets without M extension (#137195)

Closes #137023

On RISC-V machines without a native multiply instruction (e.g., `rv32i`
base), multiplying a variable by a constant integer often compiles to a
call to a library routine like `__mul{s,d}i3`.

```assembly
	.globl __mulxi3
	.type  __mulxi3, @function
__mulxi3:
	mv     a2, a0
	mv     a0, zero
.L1:
	andi   a3, a1, 1
	beqz   a3, .L2
	add    a0, a0, a2
.L2:
	srli   a1, a1, 1
	slli   a2, a2, 1
	bnez   a1, .L1
	ret
```

This library function implements multiplication in software using a loop
of shifts and adds, processing the constant bit by bit. On rv32i, it
requires a minimum of 8 instructions (for multiply by `0`) and up to
about 200 instructions (by `0xffffffff`), involves heavy branching and
function call overhead.

When not optimizing for size, we could expand the constant
multiplication into a sequence of shift and add/sub instructions. For
now we use non-adjacent form for the shift and add/sub sequence, which
could save 1/2 - 2/3 instructions compared to a shl+add-only sequence.



To unsubscribe from these emails, change your notification settings at https://github.com/llvm/llvm-project/settings/notifications