Split the init.trampoline intrinsic, which currently combines GCC's

init.trampoline and adjust.trampoline intrinsics, into two intrinsics
like in GCC.  While having one combined intrinsic is tempting, it is
not natural because typically the trampoline initialization needs to
be done in one function, and the result of adjust trampoline is needed
in a different (nested) function.  To get around this llvm-gcc hacks the
nested function lowering code to insert an additional parent variable
holding the adjust.trampoline result that can be accessed from the child
function.  Dragonegg doesn't have the luxury of tweaking GCC code, so it
stored the result of adjust.trampoline in the memory GCC set aside for
the trampoline itself (this is always available in the child function),
and set up some new memory (using an alloca) to hold the trampoline.
Unfortunately this breaks Go which allocates trampoline memory on the
heap and wants to use it even after the parent has exited (!).  Rather
than doing even more hacks to get Go working, it seemed best to just use
two intrinsics like in GCC.  Patch mostly by Sanjoy Das.

llvm-svn: 139140

2 files changed, 90 insertions, 11 deletions

diff --git a/llvm/lib/Transforms/InstCombine/InstCombine.h b/llvm/lib/Transforms/InstCombine/InstCombine.h
index c6bdb089982..be4454b8789 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombine.h
+++ b/llvm/lib/Transforms/InstCombine/InstCombine.h
@@ -11,6 +11,7 @@
 #define INSTCOMBINE_INSTCOMBINE_H
 
 #include "InstCombineWorklist.h"
+#include "llvm/IntrinsicInst.h"
 #include "llvm/Operator.h"
 #include "llvm/Pass.h"
 #include "llvm/Analysis/ValueTracking.h"
@@ -214,7 +215,8 @@ private:
   Instruction *visitCallSite(CallSite CS);
   Instruction *tryOptimizeCall(CallInst *CI, const TargetData *TD);
   bool transformConstExprCastCall(CallSite CS);
-  Instruction *transformCallThroughTrampoline(CallSite CS);
+  Instruction *transformCallThroughTrampoline(CallSite CS,
+                                              IntrinsicInst *Tramp);
   Instruction *transformZExtICmp(ICmpInst *ICI, Instruction &CI,
                                  bool DoXform = true);
   Instruction *transformSExtICmp(ICmpInst *ICI, Instruction &CI);
diff --git a/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp b/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp
index 4c9cbdb4e63..7da3343fb4e 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp
@@ -12,7 +12,6 @@
 //===----------------------------------------------------------------------===//
 
 #include "InstCombine.h"
-#include "llvm/IntrinsicInst.h"
 #include "llvm/Support/CallSite.h"
 #include "llvm/Target/TargetData.h"
 #include "llvm/Analysis/MemoryBuiltins.h"
@@ -821,6 +820,83 @@ Instruction *InstCombiner::tryOptimizeCall(CallInst *CI, const TargetData *TD) {
   return Simplifier.NewInstruction;
 }
 
+static IntrinsicInst *FindInitTrampolineFromAlloca(Value *TrampMem) {
+  // Strip off at most one level of pointer casts, looking for an alloca.  This
+  // is good enough in practice and simpler than handling any number of casts.
+  Value *Underlying = TrampMem->stripPointerCasts();
+  if (Underlying != TrampMem &&
+      (!Underlying->hasOneUse() || *Underlying->use_begin() != TrampMem))
+    return 0;
+  if (!isa<AllocaInst>(Underlying))
+    return 0;
+
+  IntrinsicInst *InitTrampoline = 0;
+  for (Value::use_iterator I = TrampMem->use_begin(), E = TrampMem->use_end();
+       I != E; I++) {
+    IntrinsicInst *II = dyn_cast<IntrinsicInst>(*I);
+    if (!II)
+      return 0;
+    if (II->getIntrinsicID() == Intrinsic::init_trampoline) {
+      if (InitTrampoline)
+        // More than one init_trampoline writes to this value.  Give up.
+        return 0;
+      InitTrampoline = II;
+      continue;
+    }
+    if (II->getIntrinsicID() == Intrinsic::adjust_trampoline)
+      // Allow any number of calls to adjust.trampoline.
+      continue;
+    return 0;
+  }
+
+  // No call to init.trampoline found.
+  if (!InitTrampoline)
+    return 0;
+
+  // Check that the alloca is being used in the expected way.
+  if (InitTrampoline->getOperand(0) != TrampMem)
+    return 0;
+
+  return InitTrampoline;
+}
+
+static IntrinsicInst *FindInitTrampolineFromBB(IntrinsicInst *AdjustTramp,
+                                               Value *TrampMem) {
+  // Visit all the previous instructions in the basic block, and try to find a
+  // init.trampoline which has a direct path to the adjust.trampoline.
+  for (BasicBlock::iterator I = AdjustTramp,
+       E = AdjustTramp->getParent()->begin(); I != E; ) {
+    Instruction *Inst = --I;
+    if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
+      if (II->getIntrinsicID() == Intrinsic::init_trampoline &&
+          II->getOperand(0) == TrampMem)
+        return II;
+    if (Inst->mayWriteToMemory())
+      return 0;
+  }
+  return 0;
+}
+
+// Given a call to llvm.adjust.trampoline, find and return the corresponding
+// call to llvm.init.trampoline if the call to the trampoline can be optimized
+// to a direct call to a function.  Otherwise return NULL.
+//
+static IntrinsicInst *FindInitTrampoline(Value *Callee) {
+  Callee = Callee->stripPointerCasts();
+  IntrinsicInst *AdjustTramp = dyn_cast<IntrinsicInst>(Callee);
+  if (!AdjustTramp ||
+      AdjustTramp->getIntrinsicID() != Intrinsic::adjust_trampoline)
+    return 0;
+
+  Value *TrampMem = AdjustTramp->getOperand(0);
+
+  if (IntrinsicInst *IT = FindInitTrampolineFromAlloca(TrampMem))
+    return IT;
+  if (IntrinsicInst *IT = FindInitTrampolineFromBB(AdjustTramp, TrampMem))
+    return IT;
+  return 0;
+}
+
 // visitCallSite - Improvements for call and invoke instructions.
 //
 Instruction *InstCombiner::visitCallSite(CallSite CS) {
@@ -880,10 +956,8 @@ Instruction *InstCombiner::visitCallSite(CallSite CS) {
     return EraseInstFromFunction(*CS.getInstruction());
   }
 
-  if (BitCastInst *BC = dyn_cast<BitCastInst>(Callee))
-    if (IntrinsicInst *In = dyn_cast<IntrinsicInst>(BC->getOperand(0)))
-      if (In->getIntrinsicID() == Intrinsic::init_trampoline)
-        return transformCallThroughTrampoline(CS);
+  if (IntrinsicInst *II = FindInitTrampoline(Callee))
+    return transformCallThroughTrampoline(CS, II);
 
   PointerType *PTy = cast<PointerType>(Callee->getType());
   FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
@@ -1164,10 +1238,13 @@ bool InstCombiner::transformConstExprCastCall(CallSite CS) {
   return true;
 }
 
-// transformCallThroughTrampoline - Turn a call to a function created by the
-// init_trampoline intrinsic into a direct call to the underlying function.
+// transformCallThroughTrampoline - Turn a call to a function created by
+// init_trampoline / adjust_trampoline intrinsic pair into a direct call to the
+// underlying function.
 //
-Instruction *InstCombiner::transformCallThroughTrampoline(CallSite CS) {
+Instruction *
+InstCombiner::transformCallThroughTrampoline(CallSite CS,
+                                             IntrinsicInst *Tramp) {
   Value *Callee = CS.getCalledValue();
   PointerType *PTy = cast<PointerType>(Callee->getType());
   FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
@@ -1178,8 +1255,8 @@ Instruction *InstCombiner::transformCallThroughTrampoline(CallSite CS) {
   if (Attrs.hasAttrSomewhere(Attribute::Nest))
     return 0;
 
-  IntrinsicInst *Tramp =
-    cast<IntrinsicInst>(cast<BitCastInst>(Callee)->getOperand(0));
+  assert(Tramp &&
+         "transformCallThroughTrampoline called with incorrect CallSite.");
 
   Function *NestF =cast<Function>(Tramp->getArgOperand(1)->stripPointerCasts());
   PointerType *NestFPTy = cast<PointerType>(NestF->getType());