[AIE2P] Enable S20 Narrowing for FIFO Loads and Stores #333
Conversation
abhinay-anubola
requested review from
abnikant,
andcarminati,
F-Stuckmann,
gbossu,
katerynamuts,
khallouh,
konstantinschwarz,
martien-de-jong,
niwinanto,
SagarMaheshwari99 and
stephenneuendorffer
as code owners
abhinay-anubola
force-pushed
the
sanubola.s20narrowing.fifo.ld.st
branch
from
1e3c4ac to
54f625d
Compare
| @@ -9,7 +9,7 @@ | |||
| //===----------------------------------------------------------------------===// | |||
| include "llvm/Target/GlobalISel/Combine.td" | |||
|
|
|||
| // AIE-specific offset folding for G_GLOBAL_VALUE. | |||
| // AIE-specific offset folding for G_GLOBALllvm/lib/Target/AIE/AIECombine.td_VALUE. | |||
| case Intrinsic::aie2p_fifo_ld_pop_576_1d_bfp16: | ||
| case Intrinsic::aie2p_fifo_ld_pop_576_2d_bfp16: | ||
| case Intrinsic::aie2p_fifo_ld_pop_576_3d_bfp16: | ||
| case Intrinsic::aie2p_fifo_st_flush_2d: |
There was a problem hiding this comment.
We now have more intrinsics available, maybe double check they are all present
There was a problem hiding this comment.
I'm wondering whether all scalar operands of these instructions are accurately qualified as 'S20'. It at least seems likely that they don't consume more than 20 bits.
There was a problem hiding this comment.
See:
%4 = tail call { ptr, <32 x i32>, i32, <64 x i8>, <8 x i8> } @llvm.aie2p.fifo.ld.pop.576.1d.bfp16(ptr %0, <32 x i32> %1, i32 %2, i20 %3)
There was a problem hiding this comment.
I'm wondering whether all scalar operands of these instructions are accurately qualified as 'S20
The availability register is indeed an s32, and we should leave it as is.
| default: | ||
| case TargetOpcode::G_INTRINSIC_W_SIDE_EFFECTS: { | ||
| const unsigned IntrinsicID = cast<GIntrinsic>(Use).getIntrinsicID(); | ||
| if (!isNativeS20ConsumerIntrinsic(IntrinsicID)) { |
There was a problem hiding this comment.
Nice you factored that out!
| tracksRegLiveness: true | ||
| body: | | ||
| bb.1: | ||
| liveins: $p0, $p1, $p2 |
There was a problem hiding this comment.
nit: $p2 is not used.
| %8:_(s32) = G_PHI %2(s32), %bb.0, %9(s32), %bb.1 | ||
| %10:_(s20) = G_TRUNC %6(s32) | ||
| %11:_(s20) = G_TRUNC %8(s32) | ||
| %12:_(p0), %13:_(<32 x s32>), %14:_(s32), %15:_(s20), %16:_(s20), %17:_(<64 x s8>), %18:_(<8 x s8>) = G_INTRINSIC_W_SIDE_EFFECTS intrinsic(@llvm.aie2p.fifo.ld.pop.544.3d.bfp16), %0(p0), %5(<32 x s32>), %2(s32), %1(s20), %1(s20), %11(s20), %1(s20), %1(s20), %10(s20), %1(s20), %3(<64 x s8>), %4(<8 x s8>) |
There was a problem hiding this comment.
Can we have a test where %0(p0), %5(<32 x s32>), %2(s32), are also taken from phi nodes?
| @@ -671,16 +694,14 @@ bool canNarrowUserTreeToS20(MachineRegisterInfo &MRI, InstrNode Start, | |||
| case TargetOpcode::G_STORE: // Data operand is later modified to S20 type | |||
There was a problem hiding this comment.
This switch seems to be an extension to isNativeS20Consumer. I guess the difference is given by 'can be adapted to'. Could we delegate the trivial cases to that predicate function?
Added FIFO load and store intrinsics as S20 consumers in S20Narrowing
Added and updated the tests for the same