[PATCH] D120531: [SystemZ] Use VREP for storing replicated regs/immediates.

Mon Mar 7 07:58:32 PST 2022

jonpa updated this revision to Diff 413483.
jonpa added a comment.

Patch updated to use SystemZVectorConstantInfo to detect the constant splats in combineSTORE().

If we were to only handle replicated constants / registers, it seems quite straightforward to simply build a splatVector in combineSTORE() and be done.

There are however also other constants that SystemZVectorConstantInfo can handle (with VGBM / VGM), which I have experimented with. The handling of stores of constants is in this version done instead in SystemZISelDAGToDAG.cpp, where Select() detects the case and loadVectorConstant() builds the vector with a vector element extraction. The experimental option '-replicate-only' skips generation of the non-REPLICATE opcodes, which should then show the value of being able to generate any vector constant.

**trunk <> patch "replicate only":**

This is the important part of the patch I think, as it eliminates the scalar multiplies (common case seems to be load-replicate-store):

  vlrepb         :                  187                  489     +302
  llc            :                39073                38772     -301
  mhi            :                 6119                 5851     -268
  msfi           :                 7066                 6934     -132
  msgrkc         :                 7116                 7071      -45
  msgc           :                  407                  403       -4

Generally it looks like a good change:

  stef           :                 5750                 6386     +636
  vsteg          :                 5912                 6475     +563
  stg            :               373052               372504     -548
  st             :               123399               122875     -524
  vrepif         :                 7708                 8127     +419
  llihl          :                 7344                 7022     -322
  oill           :                18487                18169     -318
  vsteh          :                 2557                 2875     +318
  iilf           :                 5388                 5108     -280
  sth            :                25810                25532     -278
  lay            :                54928                55179     +251
  vrepih         :                 1131                 1374     +243
  sty            :                 3642                 3527     -115
  oilf           :                 9138                 9048      -90
  llihf          :                11328                11238      -90
  OPCDIFFS: -534

  Spill|Reload   :               613545               613472      -73
  Copies         :              1005714              1005692      -22

**patch "replicate only" <> patch "with vgbm/vgm":**

There seems to be many cases where this could be enabled:

  vsteg          :                 6475                 8912    +2437
  stg            :               372504               370085    -2419
  vstef          :                 6386                 8551    +2165
  st             :               122875               121122    -1753
  vgmf           :                13118                14454    +1336
  vgmg           :                 8171                 9496    +1325
  llilh          :                 6887                 5763    -1124
  llihh          :                 2516                 1662     -854
  llihl          :                 7022                 6564     -458
  lg             :               986019               986363     +344
  l              :               179157               179407     +250
  stfh           :                21086                20837     -249
  vgbm           :                10985                11189     +204
  oilh           :                 1533                 1341     -192
  llihf          :                11238                11063     -175
  oilf           :                 9048                 8909     -139
  iilf           :                 5108                 4974     -134
  iihf           :                 4940                 4837     -103
  cg             :                32046                32139      +93
  OPCDIFFS: 554

  Spill|Reload   :               613472               614178     +706
  Copies         :              1005692              1005667      -25

The increase in spill/reload instructions seems to be due to one file with some +816 of them. I checked and have so far learned that it was not more vector register spills due to the constants being loaded into them, and it also seems that there are many more reloads, but not spills. This file is quite big, so it may be that the change is not that much relatively greater than in other smaller files.

One advantage with handling this in Select() (or rather after legalize-types), is that i128 constants now actually get handled - see  @fun_128i. Not sure if that is of any real interest...

> I agree that this looks like the right thing to do. However, this really should use the SystemZVectorConstantInfo mechanism. (You'll probably have to add an APInt constructor there in addition to the APFloat one - currently this is only used for floating-point and vector constants, but now we also need it for integer constants. But that should be straightforward.)

Per above we could either just handle the replication of regs/constants in combineSTORE(), or if we would like to be able to use vgbm/vgm also for storing integer constants we can extend also Select(), but perhaps also try again to figure out why one file gets more reloading (f510.parest_r/build/vectors.s)...

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D120531/new/

https://reviews.llvm.org/D120531

Files:
  llvm/lib/Target/SystemZ/SystemZISelDAGToDAG.cpp
  llvm/lib/Target/SystemZ/SystemZISelLowering.cpp
  llvm/lib/Target/SystemZ/SystemZISelLowering.h
  llvm/test/CodeGen/SystemZ/store-replicated-vals.ll

-------------- next part --------------
A non-text attachment was scrubbed...
Name: D120531.413483.patch
Type: text/x-patch
Size: 21483 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20220307/9d611cb5/attachment.bin>