[PATCH] D129639: [RISCV] Add a test showing a miscompilation with subreg liveness

Wed Jul 13 06:42:44 PDT 2022

frasercrmck created this revision.
frasercrmck added reviewers: craig.topper, reames, rogfer01, kito-cheng.
Herald added subscribers: sunshaoce, VincentWu, luke957, StephenFan, vkmr, evandro, luismarques, apazos, sameer.abuasal, s.egerton, Jim, benna, psnobl, jocewei, PkmX, the_o, brucehoult, MartinMosbeck, edward-jones, zzheng, jrtc27, shiva0217, niosHD, sabuasal, simoncook, johnrusso, rbar, asb, arichardson.
Herald added a project: All.
frasercrmck requested review of this revision.
Herald added subscribers: llvm-commits, pcwang-thead, eopXD, MaskRay.
Herald added a project: LLVM.

This patch adds a test which shows that we may incorrectly register
allocate for RVV instructions which have no-overlap constraints on
source/dest registers of different LMUL groups.

The particular case shows that a vrgatherei16 instruction writes to a
LMUL=1 register group v11 and reads from an EMUL=2 register group
v10/v11. This breaks the overlap constraints of the vrgatherei16
instruction.

The test also shows that disabling subregister liveness fixes the test.

We use `early-clobber` on the `VR` dest and the `VRM2` source to enforce
the constraint but with subregister liveness this constraint is not met.

It's unclear to me at this point whether this is per-design of
early-clobber in conjunction with subregisters (meaning we should find
another way of expressing this constraint) or whether it's a bug in the
register allocator somewhere.


Repository:
  rG LLVM Github Monorepo

https://reviews.llvm.org/D129639

Files:
  llvm/test/CodeGen/RISCV/rvv/vrgatherei16-subreg-liveness.ll


Index: llvm/test/CodeGen/RISCV/rvv/vrgatherei16-subreg-liveness.ll
===================================================================

--- /dev/null
+++ llvm/test/CodeGen/RISCV/rvv/vrgatherei16-subreg-liveness.ll
@@ -0,0 +1,73 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=riscv64 -mattr=+v -verify-machineinstrs < %s | FileCheck %s
+; RUN: llc -mtriple=riscv64 -mattr=+v -verify-machineinstrs < %s -riscv-enable-subreg-liveness=false | FileCheck %s --check-prefix NOSUBREG
+
+; This test checks that vrgatherei16 instructions are correctly
+; register-allocated. The LMUL=1 destination register groups may not overlap
+; with the EMUL=2 source vector register groups.
+
+; FIXME: enabling subregister liveness results in incorrect register
+; allocation!
+
+define internal void @foo(<vscale x 1 x i16> %v15, <vscale x 1 x i16> %0, <vscale x 1 x i16> %vs12.i.i.i, <vscale x 1 x i16> %1, <vscale x 8 x i8> %v37) {
+; CHECK-LABEL: foo:
+; CHECK:       # %bb.0: # %loopIR.preheader.i.i
+; CHECK-NEXT:    vsetvli a0, zero, e16, m2, ta, mu
+; CHECK-NEXT:    vmv.v.i v14, 0
+; CHECK-NEXT:    vsetvli zero, zero, e8, m1, ta, mu
+; CHECK-NEXT:    vmv.v.i v9, 0
+; CHECK-NEXT:    vsetivli zero, 4, e8, m1, tu, mu
+; CHECK-NEXT:    vmv1r.v v8, v9
+; CHECK-NEXT:    vrgatherei16.vv v8, v9, v14
+; CHECK-NEXT:  .LBB0_1: # %loopIR3.i.i
+; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
+; CHECK-NEXT:    vl1r.v v9, (zero)
+; CHECK-NEXT:    vsetivli zero, 4, e8, m1, tu, mu
+; CHECK-NEXT:    vmv1r.v v11, v12
+; CHECK-NEXT:    vrgatherei16.vv v11, v9, v10
+; CHECK-NEXT:    vsetvli a0, zero, e8, m1, ta, mu
+; CHECK-NEXT:    vand.vv v9, v8, v11
+; CHECK-NEXT:    vs1r.v v9, (zero)
+; CHECK-NEXT:    j .LBB0_1
+;
+; NOSUBREG-LABEL: foo:
+; NOSUBREG:       # %bb.0: # %loopIR.preheader.i.i
+; NOSUBREG-NEXT:    # kill: def $v10 killed $v10 def $v10m2
+; NOSUBREG-NEXT:    vsetvli a0, zero, e16, m2, ta, mu
+; NOSUBREG-NEXT:    vmv.v.i v14, 0
+; NOSUBREG-NEXT:    vsetvli zero, zero, e8, m1, ta, mu
+; NOSUBREG-NEXT:    vmv.v.i v9, 0
+; NOSUBREG-NEXT:    vsetivli zero, 4, e8, m1, tu, mu
+; NOSUBREG-NEXT:    vmv1r.v v8, v9
+; NOSUBREG-NEXT:    vrgatherei16.vv v8, v9, v14
+; NOSUBREG-NEXT:  .LBB0_1: # %loopIR3.i.i
+; NOSUBREG-NEXT:    # =>This Inner Loop Header: Depth=1
+; NOSUBREG-NEXT:    vl1r.v v9, (zero)
+; NOSUBREG-NEXT:    vsetivli zero, 4, e8, m1, tu, mu
+; NOSUBREG-NEXT:    vmv1r.v v13, v12
+; NOSUBREG-NEXT:    vrgatherei16.vv v13, v9, v10
+; NOSUBREG-NEXT:    vsetvli a0, zero, e8, m1, ta, mu
+; NOSUBREG-NEXT:    vand.vv v9, v8, v13
+; NOSUBREG-NEXT:    vs1r.v v9, (zero)
+; NOSUBREG-NEXT:    j .LBB0_1
+loopIR.preheader.i.i:
+  %v18 = tail call <vscale x 8 x i16> @llvm.vector.insert.nxv8i16.nxv1i16(<vscale x 8 x i16> poison, <vscale x 1 x i16> %vs12.i.i.i, i64 0)
+  br label %loopIR3.i.i
+
+loopIR3.i.i:                                      ; preds = %loopIR3.i.i, %loopIR.preheader.i.i
+  %v376 = load <vscale x 8 x i8>, ptr addrspace(1) null, align 8
+  %v38 = tail call <vscale x 8 x i8> @llvm.riscv.vrgatherei16.vv.nxv8i8.i64(<vscale x 8 x i8> zeroinitializer, <vscale x 8 x i8> zeroinitializer, <vscale x 8 x i16> zeroinitializer, i64 4)
+  %v40 = tail call <vscale x 8 x i8> @llvm.riscv.vrgatherei16.vv.nxv8i8.i64(<vscale x 8 x i8> %v37, <vscale x 8 x i8> %v376, <vscale x 8 x i16> %v18, i64 4)
+  %v42 = and <vscale x 8 x i8> %v38, %v40
+  store <vscale x 8 x i8> %v42, ptr addrspace(1) null, align 4
+  br label %loopIR3.i.i
+}
+
+; Function Attrs: nocallback nofree nosync nounwind readnone willreturn
+declare <vscale x 8 x i16> @llvm.vector.insert.nxv8i16.nxv1i16(<vscale x 8 x i16>, <vscale x 1 x i16>, i64 immarg) #0
+
+; Function Attrs: nounwind readnone
+declare <vscale x 8 x i8> @llvm.riscv.vrgatherei16.vv.nxv8i8.i64(<vscale x 8 x i8>, <vscale x 8 x i8>, <vscale x 8 x i16>, i64) #1
+
+attributes #0 = { nocallback nofree nosync nounwind readnone willreturn }
+attributes #1 = { nounwind readnone }


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D129639.444246.patch
Type: text/x-patch
Size: 3982 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20220713/78538460/attachment.bin>
