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

[Luke Lau via Phabricator via llvm-commits]

This revision was landed with ongoing or failed builds.
This revision was automatically updated to reflect the committed changes.
luke marked 2 inline comments as done.
Closed by commit rGed15e9119b4a: [RISCV] Don't fold vmerge into ops if fp exception can be raised (authored by luke).

Changed prior to commit:
  https://reviews.llvm.org/D154980?vs=539147&id=539933#toc

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D154980/new/

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
@@ -378,13 +378,15 @@
 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 we need to preserve exceptions with
+; This shouldn't be folded because we need to preserve exceptions with
 ; "fpexcept.strict" exception behaviour, and masking may hide them.
 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 = 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
   %b = call <vscale x 2 x float> @llvm.riscv.vmerge.nxv2f32.nxv2f32(<vscale x 2 x float> %passthru, <vscale x 2 x float> %passthru, <vscale x 2 x float> %a, <vscale x 2 x i1> %m, i64 -1)
Index: llvm/lib/Target/RISCV/RISCVISelDAGToDAG.cpp
===================================================================
--- llvm/lib/Target/RISCV/RISCVISelDAGToDAG.cpp
+++ llvm/lib/Target/RISCV/RISCVISelDAGToDAG.cpp
@@ -3301,17 +3301,20 @@
       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)))
+  // We need the VLs to be the same. But if True has a VL of VLMAX then we can
+  // go ahead and use N's VL because we know it will be smaller, so any tail
+  // elements in the result will be from Merge.
+  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 observable 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: D154980.539933.patch
Type: text/x-patch
Size: 3193 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20230713/ab81ca5b/attachment.bin>