[PATCH] D132837: [ISel] Enable generating more fma instructions.

[Thomas Symalla via Phabricator via llvm-commits]

tsymalla updated this revision to Diff 456632.
tsymalla added a comment.

Rebase.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D132837/new/

https://reviews.llvm.org/D132837

Files:
  llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
  llvm/test/CodeGen/AMDGPU/dagcombine-fma-fmad.ll


Index: llvm/test/CodeGen/AMDGPU/dagcombine-fma-fmad.ll
===================================================================
--- llvm/test/CodeGen/AMDGPU/dagcombine-fma-fmad.ll
+++ llvm/test/CodeGen/AMDGPU/dagcombine-fma-fmad.ll
@@ -196,10 +196,9 @@
 ; GCN:       ; %bb.0:
 ; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GCN-NEXT:    s_waitcnt_vscnt null, 0x0
-; GCN-NEXT:    v_mul_f32_e32 v2, v2, v3
-; GCN-NEXT:    v_fmac_f32_e32 v2, v0, v1
-; GCN-NEXT:    v_fmac_f32_e32 v2, v4, v5
-; GCN-NEXT:    v_add_f32_e32 v0, v2, v6
+; GCN-NEXT:    v_fma_f32 v0, v0, v1, v6
+; GCN-NEXT:    v_fmac_f32_e32 v0, v4, v5
+; GCN-NEXT:    v_mac_f32_e32 v0, v2, v3
 ; GCN-NEXT:    s_setpc_b64 s[30:31]
   %t0 = fmul fast float %a, %b
   %t1 = fmul fast float %c, %d
Index: llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
===================================================================
--- llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -14171,6 +14171,59 @@
     return DAG.getNode(PreferredFusedOpcode, SL, VT, A, B, CDE);
   }
 
+  // Transform
+  // (fadd (fma (a, b, ... (fma (c, d, (fmul (e, f)))))), g)
+  // into
+  // (fadd (fma (a, b, ... (fma (c, d, g)))), (fmul (e, f))
+  // to allow for further FMA combines.
+  if (CanReassociate) {
+    auto IsCopyOrExtract = [](SDValue Val) {
+      return Val.getOpcode() == ISD::CopyFromReg ||
+             Val.getOpcode() == ISD::EXTRACT_VECTOR_ELT;
+    };
+    auto TryOperandPair = [&](SDValue N0, SDValue N1) {
+      return N0.getOpcode() == ISD::FMA && N1.hasOneUse() &&
+             IsCopyOrExtract(N1);
+    };
+
+    SDValue Fma;
+    SDValue G;
+    bool TryCombine = false;
+
+    if (TryOperandPair(N0, N1)) {
+      Fma = N0;
+      G = N1;
+      TryCombine = true;
+    } else if (TryOperandPair(N1, N0)) {
+      Fma = N1;
+      G = N0;
+      TryCombine = true;
+    }
+
+    if (TryCombine) {
+      SDValue Tmp = Fma->getOperand(2);
+      while (Tmp.getOpcode() == ISD::FMA) {
+        if (Tmp.getOperand(2).getOpcode() == ISD::FMUL) {
+          SDValue FMul = Tmp.getOperand(2);
+          SDValue FMul0 = FMul.getOperand(0);
+          SDValue FMul1 = FMul.getOperand(1);
+          if (IsCopyOrExtract(FMul0) && IsCopyOrExtract(FMul1)) {
+            SDValue UpdatedFAdd = DAG.getNode(ISD::FADD, SL, VT, Fma, FMul);
+            DAG.UpdateNodeOperands(Tmp.getNode(), Tmp.getOperand(0),
+                                   Tmp.getOperand(1), G);
+
+            return UpdatedFAdd;
+          }
+
+          // Give up
+          break;
+        }
+
+        Tmp = Tmp.getOperand(2);
+      }
+    }
+  }
+
   // Look through FP_EXTEND nodes to do more combining.
 
   // fold (fadd (fpext (fmul x, y)), z) -> (fma (fpext x), (fpext y), z)


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D132837.456632.patch
Type: text/x-patch
Size: 2766 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20220830/1c80644c/attachment.bin>