[PATCH] D154980: [RISCV] Don't fold vmerge into ops if fp exception can be raised

[Luke Lau via Phabricator via llvm-commits]

luke created this revision.
luke added reviewers: craig.topper, fakepaper56, reames, frasercrmck.
Herald added subscribers: jobnoorman, asb, pmatos, VincentWu, vkmr, luismarques, apazos, sameer.abuasal, s.egerton, Jim, benna, psnobl, jocewei, PkmX, the_o, brucehoult, MartinMosbeck, rogfer01, edward-jones, zzheng, jrtc27, shiva0217, kito-cheng, niosHD, sabuasal, simoncook, johnrusso, rbar, hiraditya, arichardson.
Herald added a project: All.
luke requested review of this revision.
Herald added subscribers: llvm-commits, wangpc, eopXD, MaskRay.
Herald added a project: LLVM.

We are already checking for fp exceptions if VL changes, but I believe we
should also be checking for them if the mask changes as well, since that also
affects the set of active elements. From the spec:

> A vector floating-point exception at any active floating-point element sets
> the standard FP exception flags in the fflags register. Inactive elements do
> not set FP exception flags.

Note that we don't change the mask if IsMasked is true, i.e. True is masked
already, since in that case we keep the existing mask.


Repository:
  rG LLVM Github Monorepo

https://reviews.llvm.org/D154980

Files:
  llvm/lib/Target/RISCV/RISCVISelDAGToDAG.cpp
  llvm/test/CodeGen/RISCV/rvv/rvv-peephole-vmerge-vops.ll


Index: llvm/test/CodeGen/RISCV/rvv/rvv-peephole-vmerge-vops.ll
===================================================================
--- llvm/test/CodeGen/RISCV/rvv/rvv-peephole-vmerge-vops.ll
+++ llvm/test/CodeGen/RISCV/rvv/rvv-peephole-vmerge-vops.ll
@@ -379,12 +379,14 @@
 declare <vscale x 2 x float> @llvm.experimental.constrained.fadd(<vscale x 2 x float>, <vscale x 2 x float>, metadata, metadata)
 declare <vscale x 2 x float> @llvm.riscv.vmerge.nxv2f32.nxv2f32(<vscale x 2 x float>, <vscale x 2 x float>, <vscale x 2 x float>, <vscale x 2 x i1>, i64)
 
-; FIXME: This shouldn't be folded because the mask has been changed here
+; This shouldn't be folded because the mask has been changed here
 define <vscale x 2 x float> @vpmerge_constrained_fadd_vlmax(<vscale x 2 x float> %passthru, <vscale x 2 x float> %x, <vscale x 2 x float> %y, <vscale x 2 x i1> %m) strictfp {
 ; CHECK-LABEL: vpmerge_constrained_fadd_vlmax:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e32, m1, tu, mu
-; CHECK-NEXT:    vfadd.vv v8, v9, v10, v0.t
+; CHECK-NEXT:    vsetvli a0, zero, e32, m1, ta, ma
+; CHECK-NEXT:    vfadd.vv v9, v9, v10
+; CHECK-NEXT:    vsetvli zero, zero, e32, m1, tu, ma
+; CHECK-NEXT:    vmerge.vvm v8, v8, v9, v0
 ; CHECK-NEXT:    ret
   ; %a = fadd <vscale x 2 x float> %x, %y
   %a = call <vscale x 2 x float> @llvm.experimental.constrained.fadd(<vscale x 2 x float> %x, <vscale x 2 x float> %y, metadata !"round.dynamic", metadata !"fpexcept.strict") strictfp
Index: llvm/lib/Target/RISCV/RISCVISelDAGToDAG.cpp
===================================================================
--- llvm/lib/Target/RISCV/RISCVISelDAGToDAG.cpp
+++ llvm/lib/Target/RISCV/RISCVISelDAGToDAG.cpp
@@ -3345,17 +3345,20 @@
       True.getNumOperands() - HasVecPolicyOp - HasChainOp - HasGlueOp - 2;
   SDValue TrueVL = True.getOperand(TrueVLIndex);
 
-  auto IsNoFPExcept = [this](SDValue N) {
-    return !this->mayRaiseFPException(N.getNode()) ||
-           N->getFlags().hasNoFPExcept();
-  };
-
   // Allow the peephole for non-exception True with VLMAX vector length, since
   // all the values after VL of N are dependent on Merge. VLMAX should be
   // lowered to (XLenVT -1).
-  if (TrueVL != VL && !(IsNoFPExcept(True) && isAllOnesConstant(TrueVL)))
+  if (TrueVL != VL && !isAllOnesConstant(TrueVL))
     return false;
 
+  // If we end up changing the VL or mask of True, then we need to make sure it
+  // doesn't raise any fp exceptions, since changing the active elements will
+  // affect how fflags is set.
+  if (TrueVL != VL || !IsMasked)
+    if (mayRaiseFPException(True.getNode()) &&
+        !True->getFlags().hasNoFPExcept())
+      return false;
+
   SDLoc DL(N);
   unsigned MaskedOpc = Info->MaskedPseudo;
 #ifndef NDEBUG


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D154980.539147.patch
Type: text/x-patch
Size: 2751 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20230711/4d790071/attachment.bin>