[PATCH] D26897: AMDGPU: Refactor immediate folding logic

[Matt Arsenault via llvm-commits]

arsenm created this revision.
arsenm added a subscriber: llvm-commits.
Herald added a reviewer: tstellarAMD.
Herald added subscribers: tony-tye, yaxunl, nhaehnle, wdng, kzhuravl.

Change the logic for when to fold immediates to
consider the destination operand rather than the
source of the materializing mov instruction.

      

No change yet, but this will allow for correctly handling
i16/f16 operands. Since 32-bit moves are used to materialize
constants for these, the same bitvalue will not be in the
register.


https://reviews.llvm.org/D26897

Files:
  lib/Target/AMDGPU/SIFoldOperands.cpp


Index: lib/Target/AMDGPU/SIFoldOperands.cpp
===================================================================
--- lib/Target/AMDGPU/SIFoldOperands.cpp
+++ lib/Target/AMDGPU/SIFoldOperands.cpp
@@ -524,28 +524,63 @@
           !TargetRegisterInfo::isVirtualRegister(Dst.getReg()))
         continue;
 
-      // Folding immediates with more than one use will increase program size.
-      // FIXME: This will also reduce register usage, which may be better
-      // in some cases.  A better heuristic is needed.
-      if (FoldingImm && !TII->isInlineConstant(MI, OpToFold, OpSize) &&
-          !MRI.hasOneUse(MI.getOperand(0).getReg()))
-        continue;
-
-
       // We need mutate the operands of new mov instructions to add implicit
       // uses of EXEC, but adding them invalidates the use_iterator, so defer
       // this.
       SmallVector<MachineInstr *, 4> CopiesToReplace;
 
       std::vector<FoldCandidate> FoldList;
-      for (MachineRegisterInfo::use_iterator
-           Use = MRI.use_begin(MI.getOperand(0).getReg()), E = MRI.use_end();
-           Use != E; ++Use) {
-
-        MachineInstr *UseMI = Use->getParent();
+      if (FoldingImm) {
+        unsigned NumLiteralUses = 0;
+        MachineOperand *NonInlineUse = nullptr;
+        int NonInlineUseOpNo = -1;
+
+        // Try to fold any inline immediate uses, and then only fold other
+        // constants if they have one use.
+        //
+        // The legality of the inline immediate must be checked based on the use
+        // operand, not the defining instruction, because 32-bit instructions
+        // with 32-bit inline immediate sources may be used to materialize
+        // constants used in 16-bit operands.
+        //
+        // e.g. it is unsafe to fold:
+        //  s_mov_b32 s0, 1.0    // materializes 0x3f800000
+        //  v_add_f16 v0, v1, s0 // 1.0 f16 inline immediate sees 0x00003c00
+
+        // Folding immediates with more than one use will increase program size.
+        // FIXME: This will also reduce register usage, which may be better
+        // in some cases. A better heuristic is needed.
+        for (MachineRegisterInfo::use_iterator
+               Use = MRI.use_begin(Dst.getReg()), E = MRI.use_end();
+             Use != E; ++Use) {
+          MachineInstr *UseMI = Use->getParent();
+
+          if (TII->isInlineConstant(OpToFold, OpSize)) {
+            foldOperand(OpToFold, UseMI, Use.getOperandNo(), FoldList,
+                        CopiesToReplace, TII, TRI, MRI);
+          } else {
+            if (++NumLiteralUses == 1) {
+              NonInlineUse = &*Use;
+              NonInlineUseOpNo = Use.getOperandNo();
+            }
+          }
+        }
 
-        foldOperand(OpToFold, UseMI, Use.getOperandNo(), FoldList,
-                    CopiesToReplace, TII, TRI, MRI);
+        if (NumLiteralUses == 1) {
+          MachineInstr *UseMI = NonInlineUse->getParent();
+          foldOperand(OpToFold, UseMI, NonInlineUseOpNo, FoldList,
+                      CopiesToReplace, TII, TRI, MRI);
+        }
+      } else {
+        // Folding register.
+        for (MachineRegisterInfo::use_iterator
+               Use = MRI.use_begin(Dst.getReg()), E = MRI.use_end();
+             Use != E; ++Use) {
+          MachineInstr *UseMI = Use->getParent();
+
+          foldOperand(OpToFold, UseMI, Use.getOperandNo(), FoldList,
+                      CopiesToReplace, TII, TRI, MRI);
+        }
       }
 
       // Make sure we add EXEC uses to any new v_mov instructions created.


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D26897.78668.patch
Type: text/x-patch
Size: 3531 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20161120/6aa70870/attachment.bin>