[SCEV] Make SCEV or modeling more aggressive.

[efriedma]

Summary

Use haveNoCommonBitsSet to figure out whether an "or" instruction is equivalent to addition. This handles more cases than just checking for a constant on the RHS.

[sanjoy]

lgtm!

[jvesely]

this patch introduces infinite loop when compiling pyrite kernel[0] for r600 (turks) target:

Thread 1 "cl-program-test" received signal SIGSEGV, Segmentation fault.
0x00007fffdba03e36 in llvm::DenseMapBase<llvm::DenseMap<llvm::AssumptionCache::AffectedValueCallbackVH, llvm::SmallVector<llvm::WeakVH, 1u>, llvm::DenseMapInfo<llvm::Value*>, llvm::detail::DenseMapPair<llvm::AssumptionCache::AffectedValueCallbackVH, llvm::SmallVector<llvm::WeakVH, 1u> > >, llvm::AssumptionCache::AffectedValueCallbackVH, llvm::SmallVector<llvm::WeakVH, 1u>, llvm::DenseMapInfo<llvm::Value*>, llvm::detail::DenseMapPair<llvm::AssumptionCache::AffectedValueCallbackVH, llvm::SmallVector<llvm::WeakVH, 1u> > >::find_as<llvm::Value*> (
    this=0x8b09b0, Val=@0x7fffff7ff068: 0xe83e50)
    at /home/vesely/llvm/include/llvm/ADT/DenseMap.h:151
151	    if (LookupBucketFor(Val, TheBucket))
(gdb) bt
#0  0x00007fffdba03e36 in llvm::DenseMapBase<llvm::DenseMap<llvm::AssumptionCache::AffectedValueCallbackVH, llvm::SmallVector<llvm::WeakVH, 1u>, llvm::DenseMapInfo<llvm::Value*>, llvm::detail::DenseMapPair<llvm::AssumptionCache::AffectedValueCallbackVH, llvm::SmallVector<llvm::WeakVH, 1u> > >, llvm::AssumptionCache::AffectedValueCallbackVH, llvm::SmallVector<llvm::WeakVH, 1u>, llvm::DenseMapInfo<llvm::Value*>, llvm::detail::DenseMapPair<llvm::AssumptionCache::AffectedValueCallbackVH, llvm::SmallVector<llvm::WeakVH, 1u> > >::find_as<llvm::Value*> (
    this=0x8b09b0, Val=@0x7fffff7ff068: 0xe83e50)
    at /home/vesely/llvm/include/llvm/ADT/DenseMap.h:151
#1  0x00007fffdbb8195c in llvm::AssumptionCache::assumptionsFor (
    this=0x8b0930, V=0xe83e50)
    at /home/vesely/llvm/include/llvm/Analysis/AssumptionCache.h:134
#2  0x00007fffdbd3274c in computeKnownBitsFromAssume (V=0xe83e50, 
    KnownZero=..., KnownOne=..., Depth=5, Q=...)
    at /home/vesely/llvm/lib/Analysis/ValueTracking.cpp:515
#3  0x00007fffdbd3944f in computeKnownBits (V=0xe83e50, KnownZero=..., 
    KnownOne=..., Depth=5, Q=...)
    at /home/vesely/llvm/lib/Analysis/ValueTracking.cpp:1601
#4  0x00007fffdbd314c1 in computeKnownBitsAddSub (Add=true, Op0=0xe83e50, 
    Op1=0x8f5540, NSW=false, KnownZero=..., KnownOne=..., KnownZero2=..., 
    KnownOne2=..., Depth=4, Q=...)
    at /home/vesely/llvm/lib/Analysis/ValueTracking.cpp:263
#5  0x00007fffdbd372ff in computeKnownBitsFromOperator (I=0xe83f00, 
---Type <return> to continue, or q <return> to quit---
    KnownZero=..., KnownOne=..., Depth=4, Q=...)
    at /home/vesely/llvm/lib/Analysis/ValueTracking.cpp:1160
#6  0x00007fffdbd393c0 in computeKnownBits (V=0xe83f00, KnownZero=..., 
    KnownOne=..., Depth=4, Q=...)
    at /home/vesely/llvm/lib/Analysis/ValueTracking.cpp:1588
#7  0x00007fffdbd314c1 in computeKnownBitsAddSub (Add=true, Op0=0xe83f00, 
    Op1=0xe81b00, NSW=false, KnownZero=..., KnownOne=..., KnownZero2=..., 
    KnownOne2=..., Depth=3, Q=...)
    at /home/vesely/llvm/lib/Analysis/ValueTracking.cpp:263
#8  0x00007fffdbd372ff in computeKnownBitsFromOperator (I=0xe83fb0, 
    KnownZero=..., KnownOne=..., Depth=3, Q=...)
    at /home/vesely/llvm/lib/Analysis/ValueTracking.cpp:1160
#9  0x00007fffdbd393c0 in computeKnownBits (V=0xe83fb0, KnownZero=..., 
    KnownOne=..., Depth=3, Q=...)
    at /home/vesely/llvm/lib/Analysis/ValueTracking.cpp:1588
#10 0x00007fffdbd314c1 in computeKnownBitsAddSub (Add=true, Op0=0xe83fb0, 
    Op1=0xe83a60, NSW=false, KnownZero=..., KnownOne=..., KnownZero2=..., 
    KnownOne2=..., Depth=2, Q=...)
    at /home/vesely/llvm/lib/Analysis/ValueTracking.cpp:263
#11 0x00007fffdbd372ff in computeKnownBitsFromOperator (I=0xe84060, 
    KnownZero=..., KnownOne=..., Depth=2, Q=...)
    at /home/vesely/llvm/lib/Analysis/ValueTracking.cpp:1160
#12 0x00007fffdbd393c0 in computeKnownBits (V=0xe84060, KnownZero=..., 
---Type <return> to continue, or q <return> to quit---
    KnownOne=..., Depth=2, Q=...)
    at /home/vesely/llvm/lib/Analysis/ValueTracking.cpp:1588
#13 0x00007fffdbd314c1 in computeKnownBitsAddSub (Add=true, Op0=0xe84060, 
    Op1=0xe837d0, NSW=false, KnownZero=..., KnownOne=..., KnownZero2=..., 
    KnownOne2=..., Depth=1, Q=...)
    at /home/vesely/llvm/lib/Analysis/ValueTracking.cpp:263
#14 0x00007fffdbd372ff in computeKnownBitsFromOperator (I=0xe84110, 
    KnownZero=..., KnownOne=..., Depth=1, Q=...)
    at /home/vesely/llvm/lib/Analysis/ValueTracking.cpp:1160
#15 0x00007fffdbd393c0 in computeKnownBits (V=0xe84110, KnownZero=..., 
    KnownOne=..., Depth=1, Q=...)
    at /home/vesely/llvm/lib/Analysis/ValueTracking.cpp:1588
#16 0x00007fffdbd356ea in computeKnownBitsFromShiftOperator(const llvm::Operator *, llvm::APInt &, llvm::APInt &, llvm::APInt &, llvm::APInt &, unsigned int, const (anonymous namespace)::Query &, llvm::function_ref<llvm::APInt(const llvm::APInt&, unsigned int)>, llvm::function_ref<llvm::APInt(const llvm::APInt&, unsigned int)>) (I=0xe843d0, KnownZero=..., KnownOne=..., KnownZero2=..., 
    KnownOne2=..., Depth=0, Q=..., KZF=..., KOF=...)
    at /home/vesely/llvm/lib/Analysis/ValueTracking.cpp:813
#17 0x00007fffdbd37088 in computeKnownBitsFromOperator (I=0xe843d0, 
    KnownZero=..., KnownOne=..., Depth=0, Q=...)
    at /home/vesely/llvm/lib/Analysis/ValueTracking.cpp:1113
#18 0x00007fffdbd393c0 in computeKnownBits (V=0xe843d0, KnownZero=..., 
---Type <return> to continue, or q <return> to quit---
    KnownOne=..., Depth=0, Q=...)
    at /home/vesely/llvm/lib/Analysis/ValueTracking.cpp:1588
#19 0x00007fffdbd30d34 in llvm::computeKnownBits (V=0xe843d0, KnownZero=..., 
    KnownOne=..., DL=..., Depth=0, AC=0x8b0930, CxtI=0x0, DT=0x89fb80, ORE=0x0)
    at /home/vesely/llvm/lib/Analysis/ValueTracking.cpp:142


#20 0x00007fffdbd30ee1 in llvm::haveNoCommonBitsSet (LHS=0xe843d0, 
    RHS=0xe84480, DL=..., AC=0x8b0930, CxtI=0x0, DT=0x89fb80)
    at /home/vesely/llvm/lib/Analysis/ValueTracking.cpp:156
#21 0x00007fffdbc9a13e in llvm::ScalarEvolution::createSCEV (this=0x88c980, 
    V=0xe84530) at /home/vesely/llvm/lib/Analysis/ScalarEvolution.cpp:5332
#22 0x00007fffdbc913f8 in llvm::ScalarEvolution::getSCEV (this=0x88c980, 
    V=0xe84530) at /home/vesely/llvm/lib/Analysis/ScalarEvolution.cpp:3574
...
#9170 0x00007fffdbc913f8 in llvm::ScalarEvolution::getSCEV (this=0x88c980, 
    V=0xb45b70) at /home/vesely/llvm/lib/Analysis/ScalarEvolution.cpp:3574
#9171 0x00007fffdbc9a175 in llvm::ScalarEvolution::createSCEV (this=0x88c980, 
    V=0xb45c20) at /home/vesely/llvm/lib/Analysis/ScalarEvolution.cpp:5336
#9172 0x00007fffdbc913f8 in llvm::ScalarEvolution::getSCEV (this=0x88c980, 
    V=0xb45c20) at /home/vesely/llvm/lib/Analysis/ScalarEvolution.cpp:3574
#9173 0x00007fffdbc9925f in llvm::ScalarEvolution::createSCEV (this=0x88c980, 
    V=0xb45f50) at /home/vesely/llvm/lib/Analysis/ScalarEvolution.cpp:5213
#9174 0x00007fffdbc913f8 in llvm::ScalarEvolution::getSCEV (this=0x88c980, 
    V=0xb45f50) at /home/vesely/llvm/lib/Analysis/ScalarEvolution.cpp:3574
#9175 0x00007fffdbc99862 in llvm::ScalarEvolution::createSCEV (this=0x88c980, 
    V=0xb46290) at /home/vesely/llvm/lib/Analysis/ScalarEvolution.cpp:5271
#9176 0x00007fffdbc913f8 in llvm::ScalarEvolution::getSCEV (this=0x88c980, 
    V=0xb46290) at /home/vesely/llvm/lib/Analysis/ScalarEvolution.cpp:3574
#9177 0x00007fffdbc9a175 in llvm::ScalarEvolution::createSCEV (this=0x88c980, 
    V=0xb46340) at /home/vesely/llvm/lib/Analysis/ScalarEvolution.cpp:5336
#9178 0x00007fffdbc913f8 in llvm::ScalarEvolution::getSCEV (this=0x88c980, 
    V=0xb46340) at /home/vesely/llvm/lib/Analysis/ScalarEvolution.cpp:3574
#9179 0x00007fffdbc9925f in llvm::ScalarEvolution::createSCEV (this=0x88c980, 
    V=0xb46670) at /home/vesely/llvm/lib/Analysis/ScalarEvolution.cpp:5213

[0] https://cgit.freedesktop.org/piglit/tree/tests/cl/program/execute/pyrit-wpa-psk.cl

revision r300742 works OK. so does reverting this change on top of r300918

[efriedma]

Reverted in r300928.

[efriedma]

@jvesely, could you file a bug at bugs.llvm.org with a testcase for that crash? (Please include preprocessed source and command-line to reproduce.)

[jvesely]

In D32239#732987, @efriedma wrote:

@jvesely, could you file a bug at bugs.llvm.org with a testcase for that crash? (Please include preprocessed source and command-line to reproduce.)

http://bugs.llvm.org/show_bug.cgi?id=32731

[efriedma]

getSCEV is still recurses without a limit, so this still blows out the stack, and gets very slow for complex expressions.

[efriedma]

This was reverted.

llvm/lib/Analysis/ScalarEvolution.cpp

case Instruction::And:
                       IntegerType::get(getContext(), BitWidth - LZ - TZ)),
                   BO->LHS->getType()),
               MulCount);
         }
       }
       break;
 
     case Instruction::Or:
-      // If the RHS of the Or is a constant, we may have something like:
-      // X*4+1 which got turned into X*4|1.  Handle this as an Add so loop
-      // optimizations will transparently handle this case.
-      //
-      // In order for this transformation to be safe, the LHS must be of the
-      // form X*(2^n) and the Or constant must be less than 2^n.
-      if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->RHS)) {
-        const SCEV *LHS = getSCEV(BO->LHS);
-        const APInt &CIVal = CI->getValue();
-        if (GetMinTrailingZeros(LHS) >=
-            (CIVal.getBitWidth() - CIVal.countLeadingZeros())) {
-          // Build a plain add SCEV.
-          const SCEV *S = getAddExpr(LHS, getSCEV(CI));
-          // If the LHS of the add was an addrec and it has no-wrap flags,
-          // transfer the no-wrap flags, since an or won't introduce a wrap.
-          if (const SCEVAddRecExpr *NewAR = dyn_cast<SCEVAddRecExpr>(S)) {
-            const SCEVAddRecExpr *OldAR = cast<SCEVAddRecExpr>(LHS);
-            const_cast<SCEVAddRecExpr *>(NewAR)->setNoWrapFlags(
-                OldAR->getNoWrapFlags());
-          }
-          return S;
-        }
+      // Use ValueTracking to check whether this is actually an add.
+      if (haveNoCommonBitsSet(BO->LHS, BO->RHS, getDataLayout(), &AC,
+                              nullptr, &DT)) {
+        // There aren't any common bits set, so the add can't wrap.
+        auto Flags = SCEV::NoWrapFlags(SCEV::FlagNUW | SCEV::FlagNSW);
+        return getAddExpr(getSCEV(BO->LHS), getSCEV(BO->RHS), Flags);
       }
       break;
 
     case Instruction::Xor:
       if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->RHS)) {
         // If the RHS of xor is -1, then this is a not operation.
         if (CI->isAllOnesValue())
           return getNotSCEV(getSCEV(BO->LHS));

llvm/test/Analysis/ScalarEvolution/or-as-add.ll

+; RUN: opt < %s -analyze -scalar-evolution | FileCheck %s
+
+declare void @z(i32)
+declare void @z2(i64)
+
+define void @fun(i1 %bool, i32 %x) {
+entry:
+        br label %body
+body:
+        %i = phi i32 [ 0, %entry ], [ %i.next, %body ]
+        %bottom_zero = mul i32 %i, 2
+        %a = or i32 %bottom_zero, 1
+        call void @z(i32 %a)
+        %bool_ext = zext i1 %bool to i32
+        %b = or i32 %bool_ext, %bottom_zero
+        call void @z(i32 %b)
+        %shifted = lshr i32 %x, 31
+        %c = or i32 %shifted, %bottom_zero
+        call void @z(i32 %c)
+        %i_ext = zext i32 %i to i64
+        %d = or i64 %i_ext, 4294967296
+        call void @z2(i64 %d)
+        %i.next = add i32 %i, 1
+        %cond = icmp eq i32 %i.next, 10
+        br i1 %cond, label %exit, label %body
+exit:
+        ret void
+}
+
+; CHECK: %a = or i32 %bottom_zero, 1
+; CHECK-NEXT: -->  {1,+,2}<%body>
+; CHECK: %b = or i32 %bool_ext, %bottom_zero
+; CHECK-NEXT: -->  {(zext i1 %bool to i32),+,2}
+; CHECK: %c = or i32 %shifted, %bottom_zero
+; CHECK-NEXT: -->  {(%x /u -2147483648),+,2}<%body>
+; CHECK: %d = or i64 %i_ext, 4294967296
+; CHECK-NEXT: -->  {4294967296,+,1}<nuw><nsw><%body>
+