[all-commits] [llvm/llvm-project] f59096: [RISCV] Implement RISCVTTIImpl::getPreferredAddres...

[Philipp van Kempen via All-commits]

  Branch: refs/heads/main
  Home:   https://github.com/llvm/llvm-project
  Commit: f590963db836ccbf7c547a3dea9dc719f24444d1
      https://github.com/llvm/llvm-project/commit/f590963db836ccbf7c547a3dea9dc719f24444d1
  Author: Philipp van Kempen <philipp.van-kempen at tum.de>
  Date:   2024-12-31 (Tue, 31 Dec 2024)

  Changed paths:
    M llvm/lib/Target/RISCV/RISCVTargetTransformInfo.cpp
    M llvm/lib/Target/RISCV/RISCVTargetTransformInfo.h
    A llvm/test/CodeGen/RISCV/xcvmem-heuristic.ll

  Log Message:
  -----------
  [RISCV] Implement RISCVTTIImpl::getPreferredAddressingMode for HasVendorXCVmem (#120533)

For a simple matmult kernel this heuristic reduces the length of the
critical basic block from 15 to 20 instructions, resulting in a 20%
speedup.

**Without heuristic:**

```
       13688: 001b838b      cv.lb   t2, (s7), 0x1
       1368c: 09cdbcab      cv.lb   s9, t3(s11)
       13690: 089db62b      cv.lb   a2, s1(s11)
       13694: 092dbdab      cv.lb   s11, s2(s11)
       13698: 001d028b      cv.lb   t0, (s10), 0x1
       1369c: 00f282b3      add     t0, t0, a5
       136a0: 9072b52b      cv.mac  a0, t0, t2
       136a4: 9192bfab      cv.mac  t6, t0, s9
       136a8: 90c2beab      cv.mac  t4, t0, a2
       136ac: 91b2bf2b      cv.mac  t5, t0, s11
       136b0: fffc0c13      addi    s8, s8, -0x1
       136b4: 018e0633      add     a2, t3, s8
       136b8: 91b2b0ab      cv.mac  ra, t0, s11
       136bc: 000b8d93      mv      s11, s7
       136c0: fc0614e3      bnez    a2, 0x13688 <muriscv_nn_vec_mat_mult_t_s8+0x2f0>

       #instrs = 15
```

**With heuristic:**

```
        7bc0: 001c860b      cv.lb   a2, (s9), 0x1
        7bc4: 001e0d0b      cv.lb   s10, (t3), 0x1
        7bc8: 001e808b      cv.lb   ra, (t4), 0x1
        7bcc: 0015038b      cv.lb   t2, (a0), 0x1
        7bd0: 001c028b      cv.lb   t0, (s8), 0x1
        7bd4: 00f282b3      add     t0, t0, a5
        7bd8: 90c2bfab      cv.mac  t6, t0, a2
        7bdc: 91a2b92b      cv.mac  s2, t0, s10
        7be0: 9012b5ab      cv.mac  a1, t0, ra
        7be4: 9072b9ab      cv.mac  s3, t0, t2
        7be8: 9072b72b      cv.mac  a4, t0, t2
        7bec: fc851ae3      bne     a0, s0, 0x7bc0 <muriscv_nn_vec_mat_mult_t_s8+0x338>

        #instrs = 12

        improvement = 1 - 12/15 = 0.2 = 20%
```



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