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

This abstracts away the test for "when can we fold across a MachineInstruction"
into the the MI interface, and changes call-frame optimization use the same test
the peephole optimizer users.

Differential Revision: http://reviews.llvm.org/D11945

llvm-svn: 244729

diff --git a/llvm/include/llvm/CodeGen/MachineInstr.h b/llvm/include/llvm/CodeGen/MachineInstr.h
index 1a4d750..8c0b567 100644 (file)
--- a/llvm/include/llvm/CodeGen/MachineInstr.h
+++ b/llvm/include/llvm/CodeGen/MachineInstr.h
@@ -1100,6 +1100,9 @@ public:
   ///
   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;
 

diff --git a/llvm/lib/CodeGen/MachineInstr.cpp b/llvm/lib/CodeGen/MachineInstr.cpp
index c3d5842..e072ee1 100644 (file)
--- a/llvm/lib/CodeGen/MachineInstr.cpp
+++ b/llvm/lib/CodeGen/MachineInstr.cpp
@@ -1503,6 +1503,10 @@ bool MachineInstr::hasUnmodeledSideEffects() const {
   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 {

diff --git a/llvm/lib/CodeGen/PeepholeOptimizer.cpp b/llvm/lib/CodeGen/PeepholeOptimizer.cpp
index 3dc73ba..4d59354 100644 (file)
--- a/llvm/lib/CodeGen/PeepholeOptimizer.cpp
+++ b/llvm/lib/CodeGen/PeepholeOptimizer.cpp
@@ -1234,9 +1234,9 @@ bool PeepholeOptimizer::runOnMachineFunction(MachineFunction &MF) {
       if (MI->isDebugValue())
           continue;
 
-      // 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 we run into an instruction we can't fold across, discard
+      // the load candidates.
+      if (MI->isLoadFoldBarrier())
         FoldAsLoadDefCandidates.clear();
 
       if (MI->isPosition() || MI->isPHI() || MI->isImplicitDef() ||

diff --git a/llvm/lib/Target/X86/X86CallFrameOptimization.cpp b/llvm/lib/Target/X86/X86CallFrameOptimization.cpp
index 4fdb896..dd33c2e 100644 (file)
--- a/llvm/lib/Target/X86/X86CallFrameOptimization.cpp
+++ b/llvm/lib/Target/X86/X86CallFrameOptimization.cpp
@@ -528,13 +528,10 @@ MachineInstr *X86CallFrameOptimization::canFoldIntoRegPush(
       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;

diff --git a/llvm/test/CodeGen/X86/movtopush.ll b/llvm/test/CodeGen/X86/movtopush.ll
index b02f9ec..cd7ec1b 100644 (file)
--- a/llvm/test/CodeGen/X86/movtopush.ll
+++ b/llvm/test/CodeGen/X86/movtopush.ll
@@ -357,3 +357,26 @@ entry:
   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
+}