[PATCH] D11945: [X86] Allow x86 call frame optimization to fold more loads into pushes

Michael Kuperstein via llvm-commits llvm-commits at lists.llvm.org
Tue Aug 11 07:42:54 PDT 2015


mkuper created this revision.
mkuper added a reviewer: rnk.
mkuper added a subscriber: llvm-commits.

This abstracts away the test for "when can we fold across a MachineInstruction" as a part of the MI interface, and uses the same test in the peephole optimizer and in callframe optimization.

http://reviews.llvm.org/D11945

Files:
  include/llvm/CodeGen/MachineInstr.h
  lib/CodeGen/MachineInstr.cpp
  lib/CodeGen/PeepholeOptimizer.cpp
  lib/Target/X86/X86CallFrameOptimization.cpp
  test/CodeGen/X86/movtopush.ll

Index: test/CodeGen/X86/movtopush.ll
===================================================================
--- test/CodeGen/X86/movtopush.ll
+++ test/CodeGen/X86/movtopush.ll
@@ -357,3 +357,26 @@
   call void @good(i32 9, i32 10, i32 11, i32 12)
   ret void
 }
+
+; Make sure the add does not prevent folding loads into pushes.
+; val1 and val2 will not be folded into pushes since they have
+; an additional use, but val3 should be.
+; NORMAL-LABEL: test13:
+; NORMAL: movl ([[P1:%e..]]), [[V1:%e..]]
+; NORMAL-NEXT: movl ([[P2:%e..]]), [[V2:%e..]]
+; NORMAL-NEXT: , [[ADD:%e..]]
+; NORMAL-NEXT: pushl [[ADD]]
+; NORMAL-NEXT: pushl ([[P3:%e..]])
+; NORMAL-NEXT: pushl [[V2]]
+; NORMAL-NEXT: pushl [[V1]]
+; NORMAL-NEXT: calll _good
+; NORMAL: movl [[P3]], %eax
+define i32* @test13(i32* inreg %ptr1, i32* inreg %ptr2, i32* inreg %ptr3) optsize {
+entry:
+  %val1 = load i32, i32* %ptr1
+  %val2 = load i32, i32* %ptr2
+  %val3 = load i32, i32* %ptr3
+  %add = add i32 %val1, %val2
+  call void @good(i32 %val1, i32 %val2, i32 %val3, i32 %add)
+  ret i32* %ptr3
+}
Index: lib/Target/X86/X86CallFrameOptimization.cpp
===================================================================
--- lib/Target/X86/X86CallFrameOptimization.cpp
+++ lib/Target/X86/X86CallFrameOptimization.cpp
@@ -528,13 +528,10 @@
       DefMI->getParent() != FrameSetup->getParent())
     return nullptr;
 
-  // Now, make sure everything else up until the ADJCALLSTACK is a sequence
-  // of MOVs. To be less conservative would require duplicating a lot of the
-  // logic from PeepholeOptimizer.
-  // FIXME: A possibly better approach would be to teach the PeepholeOptimizer
-  // to be smarter about folding into pushes.
+  // Make sure we don't have any instructions between DefMI and the
+  // push that make folding the load illegal.
   for (auto I = DefMI; I != FrameSetup; ++I)
-    if (I->getOpcode() != X86::MOV32rm)
+    if (I->isLoadFoldBarrier())
       return nullptr;
 
   return DefMI;
Index: lib/CodeGen/PeepholeOptimizer.cpp
===================================================================
--- lib/CodeGen/PeepholeOptimizer.cpp
+++ lib/CodeGen/PeepholeOptimizer.cpp
@@ -1236,7 +1236,7 @@
 
       // If there exists an instruction which belongs to the following
       // categories, we will discard the load candidates.
-      if (MI->mayStore() || MI->isCall() || MI->hasUnmodeledSideEffects())
+      if (MI->isLoadFoldBarrier())
         FoldAsLoadDefCandidates.clear();
 
       if (MI->isPosition() || MI->isPHI() || MI->isImplicitDef() ||
Index: lib/CodeGen/MachineInstr.cpp
===================================================================
--- lib/CodeGen/MachineInstr.cpp
+++ lib/CodeGen/MachineInstr.cpp
@@ -1501,6 +1501,10 @@
   return false;
 }
 
+bool MachineInstr::isLoadFoldBarrier() const {
+  return mayStore() || isCall() || hasUnmodeledSideEffects();
+}
+
 /// allDefsAreDead - Return true if all the defs of this instruction are dead.
 ///
 bool MachineInstr::allDefsAreDead() const {
Index: include/llvm/CodeGen/MachineInstr.h
===================================================================
--- include/llvm/CodeGen/MachineInstr.h
+++ include/llvm/CodeGen/MachineInstr.h
@@ -1100,6 +1100,9 @@
   ///
   bool hasUnmodeledSideEffects() const;
 
+  /// Returns true if it is illegal to fold a load across this instruction.
+  bool isLoadFoldBarrier() const;
+
   /// Return true if all the defs of this instruction are dead.
   bool allDefsAreDead() const;
 


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D11945.31818.patch
Type: text/x-patch
Size: 3544 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20150811/5f2b6252/attachment.bin>


More information about the llvm-commits mailing list