fix #7197, backport a fix to SimplifyDemandedUseBits in LLVM

Make.inc

@@ -366,6 +366,7 @@ endif
 ifeq ($(USE_SYSTEM_LLVM), 1)
 LLVM_CONFIG ?= llvm-config$(EXE)
 LLVM_LLC ?= llc$(EXE)
+JCPPFLAGS+=-DSYSTEM_LLVM
 else
 LLVM_CONFIG=$(LLVMROOT)/bin/llvm-config$(EXE)
 LLVM_LLC=$(LLVMROOT)/bin/llc$(EXE)

deps/Makefile

@@ -365,6 +365,7 @@ ifeq ($(BUILD_LLDB), 1)
 endif
 ifeq ($(LLVM_VER),3.3)
 	patch -p0 < llvm-3.3.patch
+	patch -p0 < instcombine-llvm-3.3.patch
 ifeq ($(OS),WINNT)
 ifeq ($(ARCH),x86_64)
 	patch -p0 < win64-int128.llvm-3.3.patch

deps/instcombine-llvm-3.3.patch

@@ -0,0 +1,12 @@
+diff -u -r -N llvm-3.3.src/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp llvm-3.3/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp
+--- llvm-3.3.src/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp	2013-01-02 06:36:10.000000000 -0500
++++ llvm-3.3/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp	2014-06-18 23:11:49.000000000 -0400
+@@ -754,7 +754,7 @@
+       ComputeMaskedBits(I->getOperand(0), LHSKnownZero, LHSKnownOne, Depth+1);
+       // If it's known zero, our sign bit is also zero.
+       if (LHSKnownZero.isNegative())
+-        KnownZero |= LHSKnownZero;
++        KnownZero.setBit(KnownZero.getBitWidth() - 1);
+     }
+     break;
+   case Instruction::URem: {

src/codegen.cpp

@@ -3884,6 +3884,10 @@ extern "C" DLLEXPORT jl_value_t *jl_new_box(jl_value_t *v)
     return box;
 }
 
+#if LLVM_VERSION_MAJOR == 3 && LLVM_VERSION_MINOR == 3 && SYSTEM_LLVM
+#define INSTCOMBINE_BUG
+#endif
+
 static void init_julia_llvm_env(Module *m)
 {
     MDNode* tbaa_root = mbuilder->createTBAARoot("jtbaa");
@@ -4269,14 +4273,20 @@ static void init_julia_llvm_env(Module *m)
     FPM->add(createCFGSimplificationPass()); // Clean up disgusting code
     FPM->add(createPromoteMemoryToRegisterPass());// Kill useless allocas
 
+#ifndef INSTCOMBINE_BUG
     FPM->add(createInstructionCombiningPass()); // Cleanup for scalarrepl.
+#endif
     FPM->add(createScalarReplAggregatesPass()); // Break up aggregate allocas
+#ifndef INSTCOMBINE_BUG
     FPM->add(createInstructionCombiningPass()); // Cleanup for scalarrepl.
+#endif
     FPM->add(createJumpThreadingPass());        // Thread jumps.
     // NOTE: CFG simp passes after this point seem to hurt native codegen.
     // See issue #6112. Should be re-evaluated when we switch to MCJIT.
     //FPM->add(createCFGSimplificationPass());    // Merge & remove BBs
+#ifndef INSTCOMBINE_BUG
     FPM->add(createInstructionCombiningPass()); // Combine silly seq's
+#endif
 
     //FPM->add(createCFGSimplificationPass());    // Merge & remove BBs
     FPM->add(createReassociatePass());          // Reassociate expressions
@@ -4294,7 +4304,9 @@ static void init_julia_llvm_env(Module *m)
     FPM->add(createLICMPass());                 // Hoist loop invariants
     FPM->add(createLoopUnswitchPass());         // Unswitch loops.
     // Subsequent passes not stripping metadata from terminator
-    FPM->add(createInstructionCombiningPass()); 
+#ifndef INSTCOMBINE_BUG
+    FPM->add(createInstructionCombiningPass());
+#endif
     FPM->add(createIndVarSimplifyPass());       // Canonicalize indvars
     FPM->add(createLoopDeletionPass());         // Delete dead loops
     FPM->add(createLoopUnrollPass());           // Unroll small loops
@@ -4303,7 +4315,9 @@ static void init_julia_llvm_env(Module *m)
 #if LLVM_VERSION_MAJOR == 3 && LLVM_VERSION_MINOR >= 3
     FPM->add(createLoopVectorizePass());        // Vectorize loops
 #endif
+#ifndef INSTCOMBINE_BUG
     FPM->add(createInstructionCombiningPass()); // Clean up after the unroller
+#endif
     FPM->add(createGVNPass());                  // Remove redundancies
     //FPM->add(createMemCpyOptPass());            // Remove memcpy / form memset  
     FPM->add(createSCCPPass());                 // Constant prop with SCCP
@@ -4312,7 +4326,9 @@ static void init_julia_llvm_env(Module *m)
     // opened up by them.
     FPM->add(createSinkingPass()); ////////////// ****
     FPM->add(createInstructionSimplifierPass());///////// ****
+#ifndef INSTCOMBINE_BUG
     FPM->add(createInstructionCombiningPass());
+#endif
     FPM->add(createJumpThreadingPass());         // Thread jumps
     FPM->add(createDeadStoreEliminationPass());  // Delete dead stores
 

test/arrayops.jl

@@ -877,3 +877,10 @@ A = [randn(2,2) for i = 1:2, j = 1:2]
 @test issym(A.'A)
 A = [complex(randn(2,2), randn(2,2)) for i = 1:2, j = 1:2]
 @test ishermitian(A'A)
+
+# issue #7197
+function i7197()
+    S = [1 2 3; 4 5 6; 7 8 9]
+    ind2sub(size(S), 5)
+end
+@test i7197() == (2,2)