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

[Thomas Symalla via Phabricator via llvm-commits]

tsymalla updated this revision to Diff 460981.
tsymalla added a comment.

Once more, change the algorithm to keep the operands in order
Use node morphing instead of updating the node operands.


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
@@ -49,7 +49,7 @@
 ; GCN-NEXT:    v_mad_f32 v10, s2, v6, v2
 ; GCN-NEXT:    s_mov_b32 s0, 0x3c23d70a
 ; GCN-NEXT:    v_fmac_f32_e32 v1, v6, v8
-; GCN-NEXT:    v_mac_f32_e32 v10, v7, v6
+; GCN-NEXT:    v_fmac_f32_e32 v10, v7, v6
 ; GCN-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN-NEXT:    v_mul_f32_e32 v9, s10, v0
 ; GCN-NEXT:    v_fma_f32 v0, -v0, s10, s14
@@ -196,10 +196,10 @@
 ; 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_mad_f32 v2, v2, v3, v6
 ; 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_mov_b32_e32 v0, v2
 ; 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
@@ -14155,26 +14155,37 @@
 
   // fadd (fma A, B, (fmul C, D)), E --> fma A, B, (fma C, D, E)
   // fadd E, (fma A, B, (fmul C, D)) --> fma A, B, (fma C, D, E)
+  // This also works with nested fma instructions:
+  // fadd (fma A, B, (fma (C, D, (fmul (E, F))))), G -->
+  // fma A, B, (fma C, D, fma (E, F, G))
+  // fadd (G, (fma A, B, (fma (C, D, (fmul (E, F)))))) -->
+  // fma A, B, (fma C, D, fma (E, F, G)).
   // This requires reassociation because it changes the order of operations.
-  SDValue FMA, E;
-  if (CanReassociate && isFusedOp(N0) &&
-      N0.getOperand(2).getOpcode() == ISD::FMUL && N0.hasOneUse() &&
-      N0.getOperand(2).hasOneUse()) {
-    FMA = N0;
-    E = N1;
-  } else if (CanReassociate && isFusedOp(N1) &&
-             N1.getOperand(2).getOpcode() == ISD::FMUL && N1.hasOneUse() &&
-             N1.getOperand(2).hasOneUse()) {
-    FMA = N1;
-    E = N0;
-  }
-  if (FMA && E) {
-    SDValue A = FMA.getOperand(0);
-    SDValue B = FMA.getOperand(1);
-    SDValue C = FMA.getOperand(2).getOperand(0);
-    SDValue D = FMA.getOperand(2).getOperand(1);
-    SDValue CDE = DAG.getNode(PreferredFusedOpcode, SL, VT, C, D, E);
-    return DAG.getNode(PreferredFusedOpcode, SL, VT, A, B, CDE);
+  if (CanReassociate) {
+    SDValue FMA, E;
+    if (isFusedOp(N0) && N0.hasOneUse()) {
+      FMA = N0;
+      E = N1;
+    } else if (isFusedOp(N1) && N1.hasOneUse()) {
+      FMA = N1;
+      E = N0;
+    }
+
+    SDValue TmpFMA = FMA;
+    while (E && TmpFMA && isFusedOp(TmpFMA)) {
+      SDValue FMul = TmpFMA->getOperand(2);
+      if (FMul.getOpcode() == ISD::FMUL && FMul.hasOneUse()) {
+        SDValue C = FMul.getOperand(0);
+        SDValue D = FMul.getOperand(1);
+
+        DAG.MorphNodeTo(FMul.getNode(), PreferredFusedOpcode, FMul->getVTList(),
+                        {C, D, E});
+
+        return FMA;
+      }
+
+      TmpFMA = TmpFMA->getOperand(2);
+    }
   }
 
   // Look through FP_EXTEND nodes to do more combining.


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D132837.460981.patch
Type: text/x-patch
Size: 3256 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20220917/860aac86/attachment.bin>