[InstCombine] Fold strncmp of constant arrays and variable size

[msebor]

Summary

This change extends the solution accepted in D127766 to strncmp: it simplifies strncmp(A, B, N) calls with constant A and B and variable N to the equivalent of

N <= Pos ? 0 : (A < B ? -1 : B < A ? +1 : 0)

where Pos is the offset of either the first mismatch between A and B or the terminating null character if both A and B are equal strings.

I couldn't find a good way to preserve the form of the loop in optimizeMemCmpVarSize so this change reverts it to the form originally proposed in D127766.

[msebor]

Tested by running make check-all on x86_64-linux.

[nikic]

Can you please upload the patch with full context?

[msebor]

Add more context.

[nikic]

llvm/lib/Transforms/Utils/SimplifyLibCalls.cpp

1173

Not really sure why the change in code structure is needed. As far as I can tell, the only thing that's actually changing here is that a

if (StrNCmp && (LStr[Pos] == '\0' && RStr[Pos] == '\0'))
  break;

condition gets added to the top of the loop body? Or am I misunderstanding something here?

1181

Can just return Zero here?

llvm/test/Transforms/InstCombine/strncmp-5.ll

94

Verify

[msebor]

llvm/lib/Transforms/Utils/SimplifyLibCalls.cpp

1173

The result is zero in this case, not +/-1 as the current algorithm computes. Either way it seems like six of one vs half a dozen of the other.

[nikic]

llvm/lib/Transforms/Utils/SimplifyLibCalls.cpp

1173

The result is zero in this case, not +/-1 as the current algorithm computes.

I don't really follow -- I just tested this, and it at least passes the tests.

Either way it seems like six of one vs half a dozen of the other.

Agree that in terms of final outcome both are pretty much the same and I wouldn't care either way. However, this makes the patch much simpler, because it becomes a 2-line logic change. (For reference, it looks like this is the original structure from D127766, which was then changed based on a comment from @courbet, and this restores the original again.)

[msebor]

I prefer this final version. Let me know if it's okay.

[nikic]

Fine by me, but I'll leave the approval to @courbet.

[courbet]

llvm/lib/Transforms/Utils/SimplifyLibCalls.cpp

1173

The result is zero in this case, not +/-1 as the current algorithm computes.

I don't really follow -- I just tested this, and it at least passes the tests.

If we can change the code to compute something different and it does not break the tests, can you add a test case that shows the issue ?

[msebor]

llvm/lib/Transforms/Utils/SimplifyLibCalls.cpp

1173

I'm not sure I understand the concern but my impression is that we're debating the difference between the following two equivalent forms of the loop (on trunk):

for (...) {
  if (x) {
    do-stuff;
    break;
  }
}

and (in the patch):

for (...) {
  if (x)
    break;
}
do-stuff;

The "+/-1" part of my response was to the question that

...the only thing that's actually changing here is that the if (StrNCmp && (LStr[Pos] == '\0' && RStr[Pos] == '\0')) break;condition gets added to the top of the loop body?

[msebor]

@courbet: Please let me know if you have any further questions or suggestions, otherwise I'll go ahead and commit the final revision by the end of the week. Thanks!

[courbet]

I'm not sure I understand the concern but my impression is that we're debating the difference between the following two equivalent forms of the loop

OK, I was afraid you were implying that they were not equivalent.

SGTM then.

llvm/lib/Transforms/Utils/SimplifyLibCalls.cpp

if (canTransformToMemCmp(CI, Str2P, Len1, DL))
                      ConstantInt::get(DL.getIntPtrType(CI->getContext()), Len1),
                      B, DL, TLI));
   }
 
   annotateNonNullNoUndefBasedOnAccess(CI, {0, 1});
   return nullptr;
 }
 
+// Optimize a memcmp or, when StrNCmp is true, strncmp call CI with constant
+// arrays LHS and RHS and nonconstant Size.
+static Value *optimizeMemCmpVarSize(CallInst *CI, Value *LHS, Value *RHS,
+                                    Value *Size, bool StrNCmp,
+                                    IRBuilderBase &B, const DataLayout &DL);
+
 Value *LibCallSimplifier::optimizeStrNCmp(CallInst *CI, IRBuilderBase &B) {
   Value *Str1P = CI->getArgOperand(0);
   Value *Str2P = CI->getArgOperand(1);
   Value *Size = CI->getArgOperand(2);
   if (Str1P == Str2P) // strncmp(x,x,n)  -> 0
     return ConstantInt::get(CI->getType(), 0);
 
   if (isKnownNonZero(Size, DL))
     annotateNonNullNoUndefBasedOnAccess(CI, {0, 1});
   // Get the length argument if it is constant.
   uint64_t Length;
   if (ConstantInt *LengthArg = dyn_cast<ConstantInt>(Size))
     Length = LengthArg->getZExtValue();
   else
-    return nullptr;
+    return optimizeMemCmpVarSize(CI, Str1P, Str2P, Size, true, B, DL);
 
   if (Length == 0) // strncmp(x,y,0)   -> 0
     return ConstantInt::get(CI->getType(), 0);
 
   if (Length == 1) // strncmp(x,y,1) -> memcmp(x,y,1)
     return copyFlags(*CI, emitMemCmp(Str1P, Str2P, Size, B, DL, TLI));
 
   StringRef Str1, Str2;
Value *LibCallSimplifier::optimizeMemChr(CallInst *CI, IRBuilderBase &B) {
   Value *Bits = B.CreateIsNotNull(B.CreateAnd(Shl, BitfieldC), "memchr.bits");
 
   // Finally merge both checks and cast to pointer type. The inttoptr
   // implicitly zexts the i1 to intptr type.
   return B.CreateIntToPtr(B.CreateLogicalAnd(Bounds, Bits, "memchr"),
                           CI->getType());
 }
 
-// Optimize a memcmp call CI with constant arrays LHS and RHS and nonconstant
-// Size.
+// Optimize a memcmp or, when StrNCmp is true, strncmp call CI with constant
+// arrays LHS and RHS and nonconstant Size.
 static Value *optimizeMemCmpVarSize(CallInst *CI, Value *LHS, Value *RHS,
-                                    Value *Size, IRBuilderBase &B,
-                                    const DataLayout &DL) {
+                                    Value *Size, bool StrNCmp,
+                                    IRBuilderBase &B, const DataLayout &DL) {
   if (LHS == RHS) // memcmp(s,s,x) -> 0
     return Constant::getNullValue(CI->getType());
 
   StringRef LStr, RStr;
   if (!getConstantStringInfo(LHS, LStr, 0, /*TrimAtNul=*/false) ||
       !getConstantStringInfo(RHS, RStr, 0, /*TrimAtNul=*/false))
     return nullptr;
 
   // If the contents of both constant arrays are known, fold a call to
   // memcmp(A, B, N) to
   //   N <= Pos ? 0 : (A < B ? -1 : B < A ? +1 : 0)
   // where Pos is the first mismatch between A and B, determined below.
 
-  Value *Zero = ConstantInt::get(CI->getType(), 0);
+  int IRes = 0;
+  uint64_t Pos = 0;
+  for (uint64_t MinSize = std::min(LStr.size(), RStr.size()); ; ++Pos) {
+    if (Pos == MinSize ||
+        (StrNCmp && (LStr[Pos] == '\0' && RStr[Pos] == '\0'))) {
+      // One array is a leading part of the other of equal or greater
+      // size, or for strncmp, the arrays are equal strings.
+      // Fold the result to zero.  Size is assumed to be in bounds, since
+      // otherwise the call would be undefined.
+      // Setting Pos to MAX ensures that the N <= Pos test inserted below
+      // is eventually folded away.
+      Pos = UINT64_MAX;
+      break;
+    }
 
-  uint64_t MinSize = std::min(LStr.size(), RStr.size());
-  for (uint64_t Pos = 0; Pos < MinSize; ++Pos) {
     if (LStr[Pos] != RStr[Pos]) {
-      Value *MaxSize = ConstantInt::get(Size->getType(), Pos);
-      Value *Cmp = B.CreateICmp(ICmpInst::ICMP_ULE, Size, MaxSize);
+      // Normalize the result.
       typedef unsigned char UChar;
-      int IRes = UChar(LStr[Pos]) < UChar(RStr[Pos]) ? -1 : 1;
-      Value *Res = ConstantInt::get(CI->getType(), IRes);
-      return B.CreateSelect(Cmp, Zero, Res);
+      IRes = UChar(LStr[Pos]) < UChar(RStr[Pos]) ? -1 : 1;
+      break;
     }
   }
 
-  // One array is a leading part of the other of equal or greater size.
-  // Fold the result to zero.  Size is assumed to be in bounds, since
-  // otherwise the call would be undefined.
-  return Zero;
+  Value *Zero = ConstantInt::get(CI->getType(), 0);
+  Value *MaxSize = ConstantInt::get(Size->getType(), Pos);
+  Value *Cmp = B.CreateICmp(ICmpInst::ICMP_ULE, Size, MaxSize);
+  Value *Res = ConstantInt::get(CI->getType(), IRes);
+  return B.CreateSelect(Cmp, Zero, Res);
 }
 
 // Optimize a memcmp call CI with constant size Len.
 static Value *optimizeMemCmpConstantSize(CallInst *CI, Value *LHS, Value *RHS,
                                          uint64_t Len, IRBuilderBase &B,
                                          const DataLayout &DL) {
   if (Len == 0) // memcmp(s1,s2,0) -> 0
     return Constant::getNullValue(CI->getType());

 // Most simplifications for memcmp also apply to bcmp.
 Value *LibCallSimplifier::optimizeMemCmpBCmpCommon(CallInst *CI,
                                                    IRBuilderBase &B) {
   Value *LHS = CI->getArgOperand(0), *RHS = CI->getArgOperand(1);
   Value *Size = CI->getArgOperand(2);
 
   annotateNonNullAndDereferenceable(CI, {0, 1}, Size, DL);
 
-  if (Value *Res = optimizeMemCmpVarSize(CI, LHS, RHS, Size, B, DL))
+  if (Value *Res = optimizeMemCmpVarSize(CI, LHS, RHS, Size, false, B, DL))
     return Res;
 
   // Handle constant Size.
   ConstantInt *LenC = dyn_cast<ConstantInt>(Size);
   if (!LenC)
     return nullptr;
 
   return optimizeMemCmpConstantSize(CI, LHS, RHS, LenC->getZExtValue(), B, DL);

llvm/test/Transforms/InstCombine/strncmp-5.ll

+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt < %s -passes=instcombine -S | FileCheck %s
+;
+; Exercise folding of strncmp calls with constant arrays and nonconstant
+; sizes.
+
+declare i32 @strncmp(i8*, i8*, i64)
+
+@ax = external constant [8 x i8]
+@a01230123 = constant [8 x i8] c"01230123"
+@b01230123 = constant [8 x i8] c"01230123"
+@c01230129 = constant [8 x i8] c"01230129"
+@d9123_12  = constant [7 x i8] c"9123\0012"
+@e9123_34  = constant [7 x i8] c"9123\0034"
+
+
+; Exercise strncmp(A, B, N) folding of arrays with the same bytes.
+
+define void @fold_strncmp_a_b_n(i32* %pcmp, i64 %n) {
+; CHECK-LABEL: @fold_strncmp_a_b_n(
+; CHECK-NEXT:    store i32 0, i32* [[PCMP:%.*]], align 4
+; CHECK-NEXT:    [[TMP1:%.*]] = icmp ne i64 [[N:%.*]], 0
+; CHECK-NEXT:    [[TMP2:%.*]] = sext i1 [[TMP1]] to i32
+; CHECK-NEXT:    [[S0_1:%.*]] = getelementptr i32, i32* [[PCMP]], i64 1
+; CHECK-NEXT:    store i32 [[TMP2]], i32* [[S0_1]], align 4
+; CHECK-NEXT:    [[TMP3:%.*]] = icmp ne i64 [[N]], 0
+; CHECK-NEXT:    [[TMP4:%.*]] = sext i1 [[TMP3]] to i32
+; CHECK-NEXT:    [[S0_2:%.*]] = getelementptr i32, i32* [[PCMP]], i64 2
+; CHECK-NEXT:    store i32 [[TMP4]], i32* [[S0_2]], align 4
+; CHECK-NEXT:    [[TMP5:%.*]] = icmp ne i64 [[N]], 0
+; CHECK-NEXT:    [[TMP6:%.*]] = sext i1 [[TMP5]] to i32
+; CHECK-NEXT:    [[S0_3:%.*]] = getelementptr i32, i32* [[PCMP]], i64 3
+; CHECK-NEXT:    store i32 [[TMP6]], i32* [[S0_3]], align 4
+; CHECK-NEXT:    [[S0_4:%.*]] = getelementptr i32, i32* [[PCMP]], i64 4
+; CHECK-NEXT:    store i32 0, i32* [[S0_4]], align 4
+; CHECK-NEXT:    [[TMP7:%.*]] = icmp ne i64 [[N]], 0
+; CHECK-NEXT:    [[TMP8:%.*]] = sext i1 [[TMP7]] to i32
+; CHECK-NEXT:    [[S0_5:%.*]] = getelementptr i32, i32* [[PCMP]], i64 5
+; CHECK-NEXT:    store i32 [[TMP8]], i32* [[S0_5]], align 4
+; CHECK-NEXT:    [[TMP9:%.*]] = icmp ne i64 [[N]], 0
+; CHECK-NEXT:    [[TMP10:%.*]] = zext i1 [[TMP9]] to i32
+; CHECK-NEXT:    [[S5_0:%.*]] = getelementptr i32, i32* [[PCMP]], i64 6
+; CHECK-NEXT:    store i32 [[TMP10]], i32* [[S5_0]], align 4
+; CHECK-NEXT:    ret void
+;
+
+  %p0 = getelementptr [8 x i8], [8 x i8]* @a01230123, i64 0, i64 0
+
+  %q0 = getelementptr [8 x i8], [8 x i8]* @b01230123, i64 0, i64 0
+  %q1 = getelementptr [8 x i8], [8 x i8]* @b01230123, i64 0, i64 1
+  %q2 = getelementptr [8 x i8], [8 x i8]* @b01230123, i64 0, i64 2
+  %q3 = getelementptr [8 x i8], [8 x i8]* @b01230123, i64 0, i64 3
+  %q4 = getelementptr [8 x i8], [8 x i8]* @b01230123, i64 0, i64 4
+  %q5 = getelementptr [8 x i8], [8 x i8]* @b01230123, i64 0, i64 5
+
+  ; Fold strncmp(a, b, n) to 0.
+  %c0_0 = call i32 @strncmp(i8* %p0, i8* %q0, i64 %n)
+  %s0_0 = getelementptr i32, i32* %pcmp, i64 0
+  store i32 %c0_0, i32* %s0_0
+
+  ; Fold strncmp(a, b + 1, n) to N != 0 ? -1 : 0.
+  %c0_1 = call i32 @strncmp(i8* %p0, i8* %q1, i64 %n)
+  %s0_1 = getelementptr i32, i32* %pcmp, i64 1
+  store i32 %c0_1, i32* %s0_1
+
+  ; Fold strncmp(a, b + 2, n) to N != 0 ? -1 : 0.
+  %c0_2 = call i32 @strncmp(i8* %p0, i8* %q2, i64 %n)
+  %s0_2 = getelementptr i32, i32* %pcmp, i64 2
+  store i32 %c0_2, i32* %s0_2
+
+  ; Fold strncmp(a, b + 3, n) to N != 0 ? -1 : 0.
+  %c0_3 = call i32 @strncmp(i8* %p0, i8* %q3, i64 %n)
+  %s0_3 = getelementptr i32, i32* %pcmp, i64 3
+  store i32 %c0_3, i32* %s0_3
+
+  ; Fold strncmp(a, b + 4, n) to 0.
+  %c0_4 = call i32 @strncmp(i8* %p0, i8* %q4, i64 %n)
+  %s0_4 = getelementptr i32, i32* %pcmp, i64 4
+  store i32 %c0_4, i32* %s0_4
+
+  ; Fold strncmp(a, b + 5, n) to N != 0 ? -1 : 0.
+  %c0_5 = call i32 @strncmp(i8* %p0, i8* %q5, i64 %n)
+  %s0_5 = getelementptr i32, i32* %pcmp, i64 5
+  store i32 %c0_5, i32* %s0_5
+
+  ; Fold strncmp(b + 5, a, n) to N != 0 ? +1 : 0.
+  %c5_0 = call i32 @strncmp(i8* %q5, i8* %p0, i64 %n)
+  %s5_0 = getelementptr i32, i32* %pcmp, i64 6
+  store i32 %c5_0, i32* %s5_0
+
+  ret void
+}
+
+; Vefify that a strncmp() call involving a constant array with unknown
+; contents is not folded.
+
+define void @call_strncmp_a_ax_n(i32* %pcmp, i64 %n) {
+; CHECK-LABEL: @call_strncmp_a_ax_n(
+; CHECK-NEXT:    [[C0_0:%.*]] = call i32 @strncmp(i8* getelementptr inbounds ([8 x i8], [8 x i8]* @a01230123, i64 0, i64 0), i8* getelementptr inbounds ([8 x i8], [8 x i8]* @ax, i64 0, i64 0), i64 [[N:%.*]])
+; CHECK-NEXT:    store i32 [[C0_0]], i32* [[PCMP:%.*]], align 4
+; CHECK-NEXT:    ret void
+;
+
+  %p0 = getelementptr [8 x i8], [8 x i8]* @a01230123, i64 0, i64 0
+  %q0 = getelementptr [8 x i8], [8 x i8]* @ax, i64 0, i64 0
+
+  ; Do not fold strncmp(a, ax, n).
+  %c0_0 = call i32 @strncmp(i8* %p0, i8* %q0, i64 %n)
+  %s0_0 = getelementptr i32, i32* %pcmp, i64 0
+  store i32 %c0_0, i32* %s0_0
+
+  ret void
+}
+
+
+; Exercise strncmp(A, C, N) folding of arrays with the same leading bytes
+; but a difference in the trailing byte.
+
+define void @fold_strncmp_a_c_n(i32* %pcmp, i64 %n) {
+; CHECK-LABEL: @fold_strncmp_a_c_n(
+; CHECK-NEXT:    [[TMP1:%.*]] = icmp ugt i64 [[N:%.*]], 7
+; CHECK-NEXT:    [[TMP2:%.*]] = sext i1 [[TMP1]] to i32
+; CHECK-NEXT:    store i32 [[TMP2]], i32* [[PCMP:%.*]], align 4
+; CHECK-NEXT:    [[TMP3:%.*]] = icmp ne i64 [[N]], 0
+; CHECK-NEXT:    [[TMP4:%.*]] = sext i1 [[TMP3]] to i32
+; CHECK-NEXT:    [[S0_1:%.*]] = getelementptr i32, i32* [[PCMP]], i64 1
+; CHECK-NEXT:    store i32 [[TMP4]], i32* [[S0_1]], align 4
+; CHECK-NEXT:    [[TMP5:%.*]] = icmp ne i64 [[N]], 0
+; CHECK-NEXT:    [[TMP6:%.*]] = sext i1 [[TMP5]] to i32
+; CHECK-NEXT:    [[S0_2:%.*]] = getelementptr i32, i32* [[PCMP]], i64 2
+; CHECK-NEXT:    store i32 [[TMP6]], i32* [[S0_2]], align 4
+; CHECK-NEXT:    [[TMP7:%.*]] = icmp ne i64 [[N]], 0
+; CHECK-NEXT:    [[TMP8:%.*]] = sext i1 [[TMP7]] to i32
+; CHECK-NEXT:    [[S0_3:%.*]] = getelementptr i32, i32* [[PCMP]], i64 3
+; CHECK-NEXT:    store i32 [[TMP8]], i32* [[S0_3]], align 4
+; CHECK-NEXT:    [[TMP9:%.*]] = icmp ugt i64 [[N]], 3
+; CHECK-NEXT:    [[TMP10:%.*]] = sext i1 [[TMP9]] to i32
+; CHECK-NEXT:    [[S0_4:%.*]] = getelementptr i32, i32* [[PCMP]], i64 4
+; CHECK-NEXT:    store i32 [[TMP10]], i32* [[S0_4]], align 4
+; CHECK-NEXT:    [[TMP11:%.*]] = icmp ugt i64 [[N]], 3
+; CHECK-NEXT:    [[TMP12:%.*]] = sext i1 [[TMP11]] to i32
+; CHECK-NEXT:    [[S0_5:%.*]] = getelementptr i32, i32* [[PCMP]], i64 5
+; CHECK-NEXT:    store i32 [[TMP12]], i32* [[S0_5]], align 4
+; CHECK-NEXT:    ret void
+;
+
+  %p0 = getelementptr [8 x i8], [8 x i8]* @a01230123, i64 0, i64 0
+
+  %q0 = getelementptr [8 x i8], [8 x i8]* @c01230129, i64 0, i64 0
+  %q1 = getelementptr [8 x i8], [8 x i8]* @c01230129, i64 0, i64 1
+  %q2 = getelementptr [8 x i8], [8 x i8]* @c01230129, i64 0, i64 2
+  %q3 = getelementptr [8 x i8], [8 x i8]* @c01230129, i64 0, i64 3
+  %q4 = getelementptr [8 x i8], [8 x i8]* @c01230129, i64 0, i64 4
+  %q5 = getelementptr [8 x i8], [8 x i8]* @c01230129, i64 0, i64 5
+
+  ; Fold strncmp(a, c, n) to N > 7 ? -1 : 0.
+  %c0_0 = call i32 @strncmp(i8* %p0, i8* %q0, i64 %n)
+  %s0_0 = getelementptr i32, i32* %pcmp, i64 0
+  store i32 %c0_0, i32* %s0_0
+
+  ; Fold strncmp(a, c + 1, n) to N != 0 ? -1 : 0.
+  %c0_1 = call i32 @strncmp(i8* %p0, i8* %q1, i64 %n)
+  %s0_1 = getelementptr i32, i32* %pcmp, i64 1
+  store i32 %c0_1, i32* %s0_1
+
+  ; Fold strncmp(a, c + 2, n) to N != 0 ? -1 : 0.
+  %c0_2 = call i32 @strncmp(i8* %p0, i8* %q2, i64 %n)
+  %s0_2 = getelementptr i32, i32* %pcmp, i64 2
+  store i32 %c0_2, i32* %s0_2
+
+  ; Fold strncmp(a, c + 3, n) to N != 0 ? -1 : 0.
+  %c0_3 = call i32 @strncmp(i8* %p0, i8* %q3, i64 %n)
+  %s0_3 = getelementptr i32, i32* %pcmp, i64 3
+  store i32 %c0_3, i32* %s0_3
+
+  ; Fold strncmp(a, c + 4, n) to N > 3 ? -1 : 0.
+  %c0_4 = call i32 @strncmp(i8* %p0, i8* %q4, i64 %n)
+  %s0_4 = getelementptr i32, i32* %pcmp, i64 4
+  store i32 %c0_4, i32* %s0_4
+
+  ; Fold strncmp(a, c + 5, n) to N != 0 ? -1 : 0.
+  %c0_5 = call i32 @strncmp(i8* %p0, i8* %q4, i64 %n)
+  %s0_5 = getelementptr i32, i32* %pcmp, i64 5
+  store i32 %c0_5, i32* %s0_5
+
+  ret void
+}
+
+
+; Exercise strncmp(A, D, N) folding of arrays of different sizes and
+; a difference in the leading byte.
+
+define void @fold_strncmp_a_d_n(i32* %pcmp, i64 %n) {
+; CHECK-LABEL: @fold_strncmp_a_d_n(
+; CHECK-NEXT:    [[TMP1:%.*]] = icmp ne i64 [[N:%.*]], 0
+; CHECK-NEXT:    [[TMP2:%.*]] = sext i1 [[TMP1]] to i32
+; CHECK-NEXT:    store i32 [[TMP2]], i32* [[PCMP:%.*]], align 4
+; CHECK-NEXT:    [[TMP3:%.*]] = icmp ne i64 [[N]], 0
+; CHECK-NEXT:    [[TMP4:%.*]] = sext i1 [[TMP3]] to i32
+; CHECK-NEXT:    [[S0_1:%.*]] = getelementptr i32, i32* [[PCMP]], i64 1
+; CHECK-NEXT:    store i32 [[TMP4]], i32* [[S0_1]], align 4
+; CHECK-NEXT:    [[TMP5:%.*]] = icmp ugt i64 [[N]], 3
+; CHECK-NEXT:    [[TMP6:%.*]] = zext i1 [[TMP5]] to i32
+; CHECK-NEXT:    [[S1_1:%.*]] = getelementptr i32, i32* [[PCMP]], i64 2
+; CHECK-NEXT:    store i32 [[TMP6]], i32* [[S1_1]], align 4
+; CHECK-NEXT:    [[TMP7:%.*]] = icmp ugt i64 [[N]], 2
+; CHECK-NEXT:    [[TMP8:%.*]] = zext i1 [[TMP7]] to i32
+; CHECK-NEXT:    [[S2_2:%.*]] = getelementptr i32, i32* [[PCMP]], i64 3
+; CHECK-NEXT:    store i32 [[TMP8]], i32* [[S2_2]], align 4
+; CHECK-NEXT:    [[TMP9:%.*]] = icmp ne i64 [[N]], 0
+; CHECK-NEXT:    [[TMP10:%.*]] = zext i1 [[TMP9]] to i32
+; CHECK-NEXT:    [[S4_4:%.*]] = getelementptr i32, i32* [[PCMP]], i64 4
+; CHECK-NEXT:    store i32 [[TMP10]], i32* [[S4_4]], align 4
+; CHECK-NEXT:    [[TMP11:%.*]] = icmp ne i64 [[N]], 0
+; CHECK-NEXT:    [[TMP12:%.*]] = sext i1 [[TMP11]] to i32
+; CHECK-NEXT:    [[S4_4_2:%.*]] = getelementptr i32, i32* [[PCMP]], i64 5
+; CHECK-NEXT:    store i32 [[TMP12]], i32* [[S4_4_2]], align 4
+; CHECK-NEXT:    [[S5_5:%.*]] = getelementptr i32, i32* [[PCMP]], i64 6
+; CHECK-NEXT:    store i32 0, i32* [[S5_5]], align 4
+; CHECK-NEXT:    [[S6_6:%.*]] = getelementptr i32, i32* [[PCMP]], i64 7
+; CHECK-NEXT:    store i32 0, i32* [[S6_6]], align 4
+; CHECK-NEXT:    ret void
+;
+
+  %p0 = getelementptr [8 x i8], [8 x i8]* @a01230123, i64 0, i64 0
+  %p1 = getelementptr [8 x i8], [8 x i8]* @a01230123, i64 0, i64 1
+  %p2 = getelementptr [8 x i8], [8 x i8]* @a01230123, i64 0, i64 2
+  %p3 = getelementptr [8 x i8], [8 x i8]* @a01230123, i64 0, i64 3
+  %p4 = getelementptr [8 x i8], [8 x i8]* @a01230123, i64 0, i64 4
+  %p5 = getelementptr [8 x i8], [8 x i8]* @a01230123, i64 0, i64 5
+  %p6 = getelementptr [8 x i8], [8 x i8]* @a01230123, i64 0, i64 6
+
+  %q0 = getelementptr [7 x i8], [7 x i8]* @d9123_12, i64 0, i64 0
+  %q1 = getelementptr [7 x i8], [7 x i8]* @d9123_12, i64 0, i64 1
+  %q2 = getelementptr [7 x i8], [7 x i8]* @d9123_12, i64 0, i64 2
+  %q3 = getelementptr [7 x i8], [7 x i8]* @d9123_12, i64 0, i64 3
+  %q4 = getelementptr [7 x i8], [7 x i8]* @d9123_12, i64 0, i64 4
+  %q5 = getelementptr [7 x i8], [7 x i8]* @d9123_12, i64 0, i64 5
+  %q6 = getelementptr [7 x i8], [7 x i8]* @d9123_12, i64 0, i64 6
+
+  ; Fold strncmp(a, d, n) to N != 0 ? -1 : 0.
+  %c0_0 = call i32 @strncmp(i8* %p0, i8* %q0, i64 %n)
+  %s0_0 = getelementptr i32, i32* %pcmp, i64 0
+  store i32 %c0_0, i32* %s0_0
+
+  ; Fold strncmp(a, d + 1, n) to N != 0 ? -1 : 0.
+  %c0_1 = call i32 @strncmp(i8* %p0, i8* %q1, i64 %n)
+  %s0_1 = getelementptr i32, i32* %pcmp, i64 1
+  store i32 %c0_1, i32* %s0_1
+
+  ; Fold strncmp(a + 1, d + 1, n) N > 3 ? +1 : 0.
+  %c1_1 = call i32 @strncmp(i8* %p1, i8* %q1, i64 %n)
+  %s1_1 = getelementptr i32, i32* %pcmp, i64 2
+  store i32 %c1_1, i32* %s1_1
+
+  ; Fold strncmp(a + 2, d + 2, n) N > 2 ? +1 : 0.
+  %c2_2 = call i32 @strncmp(i8* %p2, i8* %q2, i64 %n)
+  %s2_2 = getelementptr i32, i32* %pcmp, i64 3
+  store i32 %c2_2, i32* %s2_2
+
+  ; Fold strncmp(a + 3, d + 3, n) N > 1 ? +1 : 0.
+  %c3_3 = call i32 @strncmp(i8* %p3, i8* %q3, i64 %n)
+  %s3_3 = getelementptr i32, i32* %pcmp, i64 4
+  store i32 %c3_3, i32* %s3_3
+
+  ; Fold strncmp(a + 4, d + 4, n) N != 0 ? +1 : 0.
+  %c4_4 = call i32 @strncmp(i8* %p4, i8* %q4, i64 %n)
+  %s4_4 = getelementptr i32, i32* %pcmp, i64 4
+  store i32 %c4_4, i32* %s4_4
+
+  ; Fold strncmp(d + 4, a + 4, n) N != 0 ? -1 : 0 (same as above but
+  ; with the array arguments reversed).
+  %c4_4_2 = call i32 @strncmp(i8* %q4, i8* %p4, i64 %n)
+  %s4_4_2 = getelementptr i32, i32* %pcmp, i64 5
+  store i32 %c4_4_2, i32* %s4_4_2
+
+  ; Fold strncmp(a + 5, d + 5, n) to 0.
+  %c5_5 = call i32 @strncmp(i8* %p5, i8* %q5, i64 %n)
+  %s5_5 = getelementptr i32, i32* %pcmp, i64 6
+  store i32 %c5_5, i32* %s5_5
+
+  ; Fold strncmp(a + 6, d + 6, n) to 0.
+  %c6_6 = call i32 @strncmp(i8* %p6, i8* %q6, i64 %n)
+  %s6_6 = getelementptr i32, i32* %pcmp, i64 7
+  store i32 %c6_6, i32* %s6_6
+
+  ret void
+}
+
+
+; Exercise strncmp(A, D, N) folding of arrays with the same bytes and
+; a nonzero size.
+
+define void @fold_strncmp_a_d_nz(i32* %pcmp, i64 %n) {
+; CHECK-LABEL: @fold_strncmp_a_d_nz(
+; CHECK-NEXT:    store i32 -1, i32* [[PCMP:%.*]], align 4
+; CHECK-NEXT:    ret void
+;
+
+  %p0 = getelementptr [8 x i8], [8 x i8]* @a01230123, i64 0, i64 0
+  %q0 = getelementptr [7 x i8], [7 x i8]* @d9123_12, i64 0, i64 0
+  %nz = or i64 %n, 1
+
+  %c0_0 = call i32 @strncmp(i8* %p0, i8* %q0, i64 %nz)
+  %s0_0 = getelementptr i32, i32* %pcmp, i64 0
+  store i32 %c0_0, i32* %s0_0
+
+  ret void
+}
+
+
+; Exercise strncmp(D, E, N) folding of equal strings but unequal arrays.
+
+define void @fold_strncmp_d_e_n(i32* %pcmp, i64 %n) {
+; CHECK-LABEL: @fold_strncmp_d_e_n(
+; CHECK-NEXT:    store i32 0, i32* [[PCMP:%.*]], align 4
+; CHECK-NEXT:    [[TMP1:%.*]] = icmp ne i64 [[N:%.*]], 0
+; CHECK-NEXT:    [[TMP2:%.*]] = zext i1 [[TMP1]] to i32
+; CHECK-NEXT:    [[S0_1:%.*]] = getelementptr i32, i32* [[PCMP]], i64 1
+; CHECK-NEXT:    store i32 [[TMP2]], i32* [[S0_1]], align 4
+; CHECK-NEXT:    [[TMP3:%.*]] = icmp ne i64 [[N]], 0
+; CHECK-NEXT:    [[TMP4:%.*]] = sext i1 [[TMP3]] to i32
+; CHECK-NEXT:    [[S1_0:%.*]] = getelementptr i32, i32* [[PCMP]], i64 2
+; CHECK-NEXT:    store i32 [[TMP4]], i32* [[S1_0]], align 4
+; CHECK-NEXT:    [[S1_1:%.*]] = getelementptr i32, i32* [[PCMP]], i64 3
+; CHECK-NEXT:    store i32 0, i32* [[S1_1]], align 4
+; CHECK-NEXT:    ret void
+;
+
+  %p0 = getelementptr [7 x i8], [7 x i8]* @d9123_12, i64 0, i64 0
+  %p1 = getelementptr [7 x i8], [7 x i8]* @d9123_12, i64 0, i64 1
+
+  %q0 = getelementptr [7 x i8], [7 x i8]* @e9123_34, i64 0, i64 0
+  %q1 = getelementptr [7 x i8], [7 x i8]* @e9123_34, i64 0, i64 1
+
+  ; Fold to 0.
+  %c0_0 = call i32 @strncmp(i8* %p0, i8* %q0, i64 %n)
+  %s0_0 = getelementptr i32, i32* %pcmp, i64 0
+  store i32 %c0_0, i32* %s0_0
+
+  ; Fold to N ? +1 : 0.
+  %c0_1 = call i32 @strncmp(i8* %p0, i8* %q1, i64 %n)
+  %s0_1 = getelementptr i32, i32* %pcmp, i64 1
+  store i32 %c0_1, i32* %s0_1
+
+  ; Fold to N ? -1 : 0.
+  %c1_0 = call i32 @strncmp(i8* %p1, i8* %q0, i64 %n)
+  %s1_0 = getelementptr i32, i32* %pcmp, i64 2
+  store i32 %c1_0, i32* %s1_0
+
+  ; Fold to 0.
+  %c1_1 = call i32 @strncmp(i8* %p1, i8* %q1, i64 %n)
+  %s1_1 = getelementptr i32, i32* %pcmp, i64 3
+  store i32 %c1_1, i32* %s1_1
+
+  ret void
+}

llvm/test/Transforms/InstCombine/strncmp-6.ll

+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt < %s -passes=instcombine -S | FileCheck %s
+;
+; Exercise folding of strncmp calls with constant arrays including both
+; negative and positive characters and both constant and nonconstant sizes.
+
+declare i32 @strncmp(i8*, i8*, i64)
+
+@a = constant [7 x i8] c"abcdef\7f"
+@b = constant [7 x i8] c"abcdef\80"
+
+
+; Exercise strncmp(A + C, B + C, 2) folding of small arrays that differ in
+; a character with the opposite sign and a constant size.
+
+define void @fold_strncmp_cst_cst(i32* %pcmp) {
+; CHECK-LABEL: @fold_strncmp_cst_cst(
+; CHECK-NEXT:    store i32 -1, i32* [[PCMP:%.*]], align 4
+; CHECK-NEXT:    [[SB5_A5:%.*]] = getelementptr i32, i32* [[PCMP]], i64 1
+; CHECK-NEXT:    store i32 1, i32* [[SB5_A5]], align 4
+; CHECK-NEXT:    [[SA6_B6:%.*]] = getelementptr i32, i32* [[PCMP]], i64 2
+; CHECK-NEXT:    store i32 -1, i32* [[SA6_B6]], align 4
+; CHECK-NEXT:    [[SB6_A6:%.*]] = getelementptr i32, i32* [[PCMP]], i64 3
+; CHECK-NEXT:    store i32 1, i32* [[SB6_A6]], align 4
+; CHECK-NEXT:    ret void
+;
+  %p5 = getelementptr [7 x i8], [7 x i8]* @a, i64 0, i64 5
+  %p6 = getelementptr [7 x i8], [7 x i8]* @a, i64 0, i64 6
+
+  %q5 = getelementptr [7 x i8], [7 x i8]* @b, i64 0, i64 5
+  %q6 = getelementptr [7 x i8], [7 x i8]* @b, i64 0, i64 6
+
+  ; Fold strncmp(a + 5, b + 5, 2) to -1.
+  %ca5_b5 = call i32 @strncmp(i8* %p5, i8* %q5, i64 2)
+  %sa5_b5 = getelementptr i32, i32* %pcmp, i64 0
+  store i32 %ca5_b5, i32* %sa5_b5
+
+  ; Fold strncmp(b + 5, a + 5, 2) to +1.
+  %cb5_a5 = call i32 @strncmp(i8* %q5, i8* %p5, i64 2)
+  %sb5_a5 = getelementptr i32, i32* %pcmp, i64 1
+  store i32 %cb5_a5, i32* %sb5_a5
+
+  ; Fold strncmp(a + 6, b + 6, 1) to -1.
+  %ca6_b6 = call i32 @strncmp(i8* %p6, i8* %q6, i64 1)
+  %sa6_b6 = getelementptr i32, i32* %pcmp, i64 2
+  store i32 %ca6_b6, i32* %sa6_b6
+
+  ; Fold strncmp(b + 6, a + 6, 1) to  +1.
+  %cb6_a6 = call i32 @strncmp(i8* %q6, i8* %p6, i64 1)
+  %sb6_a6 = getelementptr i32, i32* %pcmp, i64 3
+  store i32 %cb6_a6, i32* %sb6_a6
+
+  ret void
+}
+
+
+; Exercise strncmp(A, B, N) folding of arrays that differ in a character
+; with the opposite sign and a variable size
+
+define void @fold_strncmp_cst_var(i32* %pcmp, i64 %n) {
+; CHECK-LABEL: @fold_strncmp_cst_var(
+; CHECK-NEXT:    [[TMP1:%.*]] = icmp ugt i64 [[N:%.*]], 6
+; CHECK-NEXT:    [[TMP2:%.*]] = sext i1 [[TMP1]] to i32
+; CHECK-NEXT:    store i32 [[TMP2]], i32* [[PCMP:%.*]], align 4
+; CHECK-NEXT:    [[TMP3:%.*]] = icmp ugt i64 [[N]], 6
+; CHECK-NEXT:    [[TMP4:%.*]] = zext i1 [[TMP3]] to i32
+; CHECK-NEXT:    [[SB0_A0:%.*]] = getelementptr i32, i32* [[PCMP]], i64 1
+; CHECK-NEXT:    store i32 [[TMP4]], i32* [[SB0_A0]], align 4
+; CHECK-NEXT:    [[TMP5:%.*]] = icmp ne i64 [[N]], 0
+; CHECK-NEXT:    [[TMP6:%.*]] = sext i1 [[TMP5]] to i32
+; CHECK-NEXT:    [[SA6_B6:%.*]] = getelementptr i32, i32* [[PCMP]], i64 2
+; CHECK-NEXT:    store i32 [[TMP6]], i32* [[SA6_B6]], align 4
+; CHECK-NEXT:    [[TMP7:%.*]] = icmp ne i64 [[N]], 0
+; CHECK-NEXT:    [[TMP8:%.*]] = zext i1 [[TMP7]] to i32
+; CHECK-NEXT:    [[SB6_A6:%.*]] = getelementptr i32, i32* [[PCMP]], i64 3
+; CHECK-NEXT:    store i32 [[TMP8]], i32* [[SB6_A6]], align 4
+; CHECK-NEXT:    ret void
+;
+  %p0 = getelementptr [7 x i8], [7 x i8]* @a, i64 0, i64 0
+  %p6 = getelementptr [7 x i8], [7 x i8]* @a, i64 0, i64 6
+
+  %q0 = getelementptr [7 x i8], [7 x i8]* @b, i64 0, i64 0
+  %q6 = getelementptr [7 x i8], [7 x i8]* @b, i64 0, i64 6
+
+  ; Fold strncmp(a, b, n) to -1.
+  %ca0_b0 = call i32 @strncmp(i8* %p0, i8* %q0, i64 %n)
+  %sa0_b0 = getelementptr i32, i32* %pcmp, i64 0
+  store i32 %ca0_b0, i32* %sa0_b0
+
+  ; Fold strncmp(b, a, n) to +1.
+  %cb0_a0 = call i32 @strncmp(i8* %q0, i8* %p0, i64 %n)
+  %sb0_a0 = getelementptr i32, i32* %pcmp, i64 1
+  store i32 %cb0_a0, i32* %sb0_a0
+
+  ; Fold strncmp(a + 6, b + 6, n) to -1.
+  %ca6_b6 = call i32 @strncmp(i8* %p6, i8* %q6, i64 %n)
+  %sa6_b6 = getelementptr i32, i32* %pcmp, i64 2
+  store i32 %ca6_b6, i32* %sa6_b6
+
+  ; Fold strncmp(b + 6, a + 6, n) to +1.
+  %cb6_a6 = call i32 @strncmp(i8* %q6, i8* %p6, i64 %n)
+  %sb6_a6 = getelementptr i32, i32* %pcmp, i64 3
+  store i32 %cb6_a6, i32* %sb6_a6
+
+  ret void
+}