From: Chris Lattner <sabre@nondot.org>
Date: Fri, 13 Apr 2007 20:42:26 +0000 (+0000)
Subject: Now that codegen prepare isn't defeating me, I can finally fix what I set
X-Git-Url: http://plrg.eecs.uci.edu/git/?p=oota-llvm.git;a=commitdiff_plain;h=c5494af8a90f398046c45bc2b7549ab9004c01d9

Now that codegen prepare isn't defeating me, I can finally fix what I set
out to do! :)

This fixes a problem where LSR would insert a bunch of code into each MBB
that uses a particular subexpression (e.g. IV+base+C).  The problem is that
this code cannot be CSE'd back together if inserted into different blocks.

This patch changes LSR to attempt to insert a single copy of this code and
share it, allowing codegenprepare to duplicate the code if it can be sunk
into various addressing modes.  On CodeGen/ARM/lsr-code-insertion.ll,
for example, this gives us code like:

        add r8, r0, r5
        str r6, [r8, #+4]
..
        ble LBB1_4      @cond_next
LBB1_3: @cond_true
        str r10, [r8, #+4]
LBB1_4: @cond_next
...
LBB1_5: @cond_true55
        ldr r6, LCPI1_1
        str r6, [r8, #+4]

instead of:

        add r10, r0, r6
        str r8, [r10, #+4]
...
        ble LBB1_4      @cond_next
LBB1_3: @cond_true
        add r8, r0, r6
        str r10, [r8, #+4]
LBB1_4: @cond_next
...
LBB1_5: @cond_true55
        add r8, r0, r6
        ldr r10, LCPI1_1
        str r10, [r8, #+4]

Besides being smaller and more efficient, this makes it immediately
obvious that it is profitable to predicate LBB1_3 now :)


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35972 91177308-0d34-0410-b5e6-96231b3b80d8
---

diff --git a/lib/Transforms/Scalar/LoopStrengthReduce.cpp b/lib/Transforms/Scalar/LoopStrengthReduce.cpp
index 7540f44e4e7..332ddfa488e 100644
--- a/lib/Transforms/Scalar/LoopStrengthReduce.cpp
+++ b/lib/Transforms/Scalar/LoopStrengthReduce.cpp
@@ -568,7 +568,25 @@ void BasedUser::RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
                                                SCEVExpander &Rewriter,
                                                Loop *L, Pass *P) {
   if (!isa<PHINode>(Inst)) {
-    Value *NewVal = InsertCodeForBaseAtPosition(NewBase, Rewriter, Inst, L);
+    // By default, insert code at the user instruction.
+    BasicBlock::iterator InsertPt = Inst;
+    
+    // However, if the Operand is itself an instruction, the (potentially
+    // complex) inserted code may be shared by many users.  Because of this, we
+    // want to emit code for the computation of the operand right before its old
+    // computation.  This is usually safe, because we obviously used to use the
+    // computation when it was computed in its current block.  However, in some
+    // cases (e.g. use of a post-incremented induction variable) the NewBase
+    // value will be pinned to live somewhere after the original computation.
+    // In this case, we have to back off.
+    if (!isUseOfPostIncrementedValue) {
+      if (Instruction *OpInst = dyn_cast<Instruction>(OperandValToReplace)) { 
+        InsertPt = OpInst;
+        while (isa<PHINode>(InsertPt)) ++InsertPt;
+      }
+    }
+    
+    Value *NewVal = InsertCodeForBaseAtPosition(NewBase, Rewriter, InsertPt, L);
     // Replace the use of the operand Value with the new Phi we just created.
     Inst->replaceUsesOfWith(OperandValToReplace, NewVal);
     DOUT << "    CHANGED: IMM =" << *Imm << "  Inst = " << *Inst;