1. [RAGreedy] Enable -consider-local-interval-cost for AArch64 (details)
Commit f649f24d388c745d20fab5573d27b822b92818ed by Sanne.Wouda
[RAGreedy] Enable -consider-local-interval-cost for AArch64
Summary: The greedy register allocator occasionally decides to insert a
large number of unnecessary copies, see below for an example.  The
-consider-local-interval-cost option (which X86 already enables by
default) fixes this.  We enable this option for AArch64 only after
receiving feedback that this change is not beneficial for PowerPC.
We evaluated the impact of this change on compile time, code size and
performance benchmarks.
This option has a small impact on compile time, measured on CTMark. A
0.1% geomean regression on -O1 and -O2, and 0.2% geomean for -O3, with
at most 0.5% on individual benchmarks.
The effect on both code size and performance on AArch64 for the LLVM
test suite is nil on the geomean with individual outliers (ignoring
short exec_times) between:
                 best     worst
size..text     -3.3%    +0.0%
exec_time      -5.8%    +2.3%
On SPEC CPU® 2017 (compiled for AArch64) there is a minor reduction
(-0.2% at most) in code size on some benchmarks, with a tiny movement
(-0.01%) on the geomean.  Neither intrate nor fprate show any change in
This patch makes the following changes.
- For the AArch64 target, enableAdvancedRASplitCost() now returns true.
- Ensures that -consider-local-interval-cost=false can disable the new
behaviour if necessary.
This matrix multiply example:
   $ cat test.c
  long A[8][8];
  long B[8][8];
  long C[8][8];
   void run_test() {
    for (int k = 0; k < 8; k++) {
      for (int i = 0; i < 8; i++) {
for (int j = 0; j < 8; j++) {
   C[i][j] += A[i][k] * B[k][j];
results in the following generated code on AArch64:
  $ clang --target=aarch64-arm-none-eabi -O3 -S test.c -o -
                                       // %for.cond1.preheader
                                       // =>This Inner Loop Header:
       add     x14, x11, x9
       str     q0, [sp, #16]           // 16-byte Folded Spill
       ldr     q0, [x14]
       mov     v2.16b, v15.16b
       mov     v15.16b, v14.16b
       mov     v14.16b, v13.16b
       mov     v13.16b, v12.16b
       mov     v12.16b, v11.16b
       mov     v11.16b, v10.16b
       mov     v10.16b, v9.16b
       mov     v9.16b, v8.16b
       mov     v8.16b, v31.16b
       mov     v31.16b, v30.16b
       mov     v30.16b, v29.16b
       mov     v29.16b, v28.16b
       mov     v28.16b, v27.16b
       mov     v27.16b, v26.16b
       mov     v26.16b, v25.16b
       mov     v25.16b, v24.16b
       mov     v24.16b, v23.16b
       mov     v23.16b, v22.16b
       mov     v22.16b, v21.16b
       mov     v21.16b, v20.16b
       mov     v20.16b, v19.16b
       mov     v19.16b, v18.16b
       mov     v18.16b, v17.16b
       mov     v17.16b, v16.16b
       mov     v16.16b, v7.16b
       mov     v7.16b, v6.16b
       mov     v6.16b, v5.16b
       mov     v5.16b, v4.16b
       mov     v4.16b, v3.16b
       mov     v3.16b, v1.16b
       mov     x12, v0.d[1]
       fmov    x15, d0
       ldp     q1, q0, [x14, #16]
       ldur    x1, [x10, #-256]
       ldur    x2, [x10, #-192]
       add     x9, x9, #64             // =64
       mov     x13, v1.d[1]
       fmov    x16, d1
       ldr     q1, [x14, #48]
       mul     x3, x15, x1
       mov     x14, v0.d[1]
       fmov    x17, d0
       mov     x18, v1.d[1]
       fmov    x0, d1
       mov     v1.16b, v3.16b
       mov     v3.16b, v4.16b
       mov     v4.16b, v5.16b
       mov     v5.16b, v6.16b
       mov     v6.16b, v7.16b
       mov     v7.16b, v16.16b
       mov     v16.16b, v17.16b
       mov     v17.16b, v18.16b
       mov     v18.16b, v19.16b
       mov     v19.16b, v20.16b
       mov     v20.16b, v21.16b
       mov     v21.16b, v22.16b
       mov     v22.16b, v23.16b
       mov     v23.16b, v24.16b
       mov     v24.16b, v25.16b
       mov     v25.16b, v26.16b
       mov     v26.16b, v27.16b
       mov     v27.16b, v28.16b
       mov     v28.16b, v29.16b
       mov     v29.16b, v30.16b
       mov     v30.16b, v31.16b
       mov     v31.16b, v8.16b
       mov     v8.16b, v9.16b
       mov     v9.16b, v10.16b
       mov     v10.16b, v11.16b
       mov     v11.16b, v12.16b
       mov     v12.16b, v13.16b
       mov     v13.16b, v14.16b
       mov     v14.16b, v15.16b
       mov     v15.16b, v2.16b
       ldr     q2, [sp]                // 16-byte Folded Reload
       fmov    d0, x3
       mul     x3, x12, x1
With -consider-local-interval-cost the same section of code results in
the following:
  $ clang --target=aarch64-arm-none-eabi -mllvm
-consider-local-interval-cost -O3 -S test.c -o -
.LBB0_1:                              // %for.cond1.preheader
                                       // =>This Inner Loop Header:
       add     x14, x11, x9
       ldp     q0, q1, [x14]
       ldur    x1, [x10, #-256]
       ldur    x2, [x10, #-192]
       add     x9, x9, #64             // =64
       mov     x12, v0.d[1]
       fmov    x15, d0
       mov     x13, v1.d[1]
       fmov    x16, d1
       ldp     q0, q1, [x14, #32]
       mul     x3, x15, x1
       cmp     x9, #512                // =512
       mov     x14, v0.d[1]
       fmov    x17, d0
       fmov    d0, x3
       mul     x3, x12, x1
Reviewers: SjoerdMeijer, samparker, dmgreen, qcolombet
Reviewed By: dmgreen
Subscribers: ZhangKang, jsji, wuzish, ppc-slack, lkail, steven.zhang,
MatzeB, qcolombet, kristof.beyls, hiraditya, llvm-commits
Tags: #llvm
Differential Revision:
The file was modifiedllvm/lib/CodeGen/RegAllocGreedy.cpp
The file was modifiedllvm/lib/Target/AArch64/AArch64Subtarget.h
The file was modifiedllvm/test/CodeGen/AArch64/ragreedy-local-interval-cost.ll