Merge pull request #26985 from JuliaLang/kf/simdloop

[NewOptimizer] Make simdloop marker more robust
JuliaLang · May 7, 2018 · 71f6bfe · 71f6bfe
2 parents 989de79 + 56d7ebe
commit 71f6bfe
Show file tree

Hide file tree

Showing 8 changed files with 118 additions and 41 deletions.
diff --git a/base/compiler/ssair/queries.jl b/base/compiler/ssair/queries.jl
@@ -10,7 +10,6 @@ function stmt_effect_free(@nospecialize(stmt), src, mod::Module)
     if isa(stmt, Expr)
         e = stmt::Expr
         head = e.head
-        is_meta_expr_head(head) && return true
         if head === :static_parameter
             # if we aren't certain enough about the type, it might be an UndefVarError at runtime
             return isa(e.typ, Const) || issingletontype(widenconst(e.typ))
@@ -45,9 +44,10 @@ function stmt_effect_free(@nospecialize(stmt), src, mod::Module)
                 eT ⊑ fT || return false
             end
             return true
-        elseif head === :isdefined || head === :the_exception || head === :copyast
+        elseif head === :isdefined || head === :the_exception || head === :copyast || head === :inbounds || head === :boundscheck
             return true
         else
+            # e.g. :simdloop
             return false
         end
     end

diff --git a/deps/Versions.make b/deps/Versions.make
@@ -1,5 +1,5 @@
 LLVM_VER = 6.0.0
-LLVM_BB_REL = 1
+LLVM_BB_REL = 2
 PCRE_VER = 10.30
 DSFMT_VER = 2.2.3
 LAPACK_VER = 3.5.0

diff --git a/deps/llvm.mk b/deps/llvm.mk
@@ -418,6 +418,7 @@ $(eval $(call LLVM_PATCH,llvm-D30114)) # PPC remove for 5.0
 $(eval $(call LLVM_PATCH,llvm-PR36292)) # PPC fixes #26249, remove for 6.0
 $(eval $(call LLVM_PATCH,llvm-D39297-musl-dynamiclibrary-pre5)) # Remove for 6.0
 $(eval $(call LLVM_PATCH,llvm-D28476-musl-targetlibraryinfo_3.9)) # Remove for 5.0
+$(eval $(call LLVM_PATCH,llvm-D46460))
 ifeq ($(BUILD_LLVM_CLANG),1)
 $(eval $(call LLVM_PATCH,compiler_rt-3.9-glibc_2.25.90)) # Remove for 5.0
 $(eval $(call LLVM_PATCH,clang-D28477)) # Remove for 5.0
@@ -452,6 +453,7 @@ $(eval $(call LLVM_PATCH,llvm-D30114)) # PPC remove for 5.0
 $(eval $(call LLVM_PATCH,llvm-PR36292)) # PPC fixes #26249, remove for 6.0
 $(eval $(call LLVM_PATCH,llvm-D39297-musl-dynamiclibrary-pre5)) # Remove for 6.0
 $(eval $(call LLVM_PATCH,llvm-D28476-musl-targetlibraryinfo_4.0)) # Remove for 5.0
+$(eval $(call LLVM_PATCH,llvm-D46460))
 ifeq ($(BUILD_LLVM_CLANG),1)
 $(eval $(call LLVM_PATCH,compiler_rt-3.9-glibc_2.25.90)) # Remove for 5.0
 $(eval $(call LLVM_PATCH,clang-D28477)) # Remove for 5.0
@@ -470,6 +472,7 @@ $(eval $(call LLVM_PATCH,llvm-D42262-jumpthreading-not-i1)) # remove for 7.0
 $(eval $(call LLVM_PATCH,llvm-PPC-addrspaces)) # PPC
 $(eval $(call LLVM_PATCH,llvm-PR36292-5.0)) # PPC fixes #26249, remove for 6.0
 $(eval $(call LLVM_PATCH,llvm-D39297-musl-dynamiclibrary)) # Remove for 6.0
+$(eval $(call LLVM_PATCH,llvm-D46460))
 else ifeq ($(LLVM_VER_SHORT),6.0)
 $(eval $(call LLVM_PATCH,llvm-D27629-AArch64-large_model_4.0))
 $(eval $(call LLVM_PATCH,llvm-D34078-vectorize-fdiv))
@@ -484,6 +487,7 @@ $(eval $(call LLVM_PATCH,llvm-6.0-D44650)) # mingw32 build fix
 $(eval $(call LLVM_PATCH,llvm-D45008)) # remove for 7.0
 $(eval $(call LLVM_PATCH,llvm-D45070)) # remove for 7.0
 $(eval $(call LLVM_PATCH,llvm-6.0.0-ifconv-D45819)) # remove for 7.0
+$(eval $(call LLVM_PATCH,llvm-D46460))
 endif # LLVM_VER
 
 # Remove hardcoded OS X requirements in compilter-rt cmake build

diff --git a/deps/patches/llvm-D46460.patch b/deps/patches/llvm-D46460.patch
@@ -0,0 +1,26 @@
+Index: lib/Analysis/LoopInfo.cpp
+===================================================================
+--- a/lib/Analysis/LoopInfo.cpp
++++ b/lib/Analysis/LoopInfo.cpp
+@@ -223,15 +223,14 @@
+     BasicBlock *H = getHeader();
+     for (BasicBlock *BB : this->blocks()) {
+       TerminatorInst *TI = BB->getTerminator();
+-      MDNode *MD = nullptr;
+
+       // Check if this terminator branches to the loop header.
+-      for (BasicBlock *Successor : TI->successors()) {
+-        if (Successor == H) {
+-          MD = TI->getMetadata(LLVMContext::MD_loop);
+-          break;
+-        }
+-      }
++      bool IsPredecessor = any_of(TI->successors(),
++        [=](BasicBlock *Successor) { return Successor == H; });
++      if (!IsPredecessor)
++        continue;
++
++      MDNode *MD = TI->getMetadata(LLVMContext::MD_loop);
+       if (!MD)
+         return nullptr;
+
diff --git a/src/codegen.cpp b/src/codegen.cpp
@@ -301,6 +301,7 @@ static Function *jlegal_func;
 static Function *jl_alloc_obj_func;
 static Function *jl_newbits_func;
 static Function *jl_typeof_func;
+static Function *jl_simdloop_marker_func;
 static Function *jl_write_barrier_func;
 static Function *jlisa_func;
 static Function *jlsubtype_func;
@@ -4084,7 +4085,7 @@ static jl_cgval_t emit_expr(jl_codectx_t &ctx, jl_value_t *expr)
                     maybe_decay_untracked(boxed(ctx, ast))), true, jl_expr_type);
     }
     else if (head == simdloop_sym) {
-        llvm::annotateSimdLoop(ctx.builder.GetInsertBlock());
+        ctx.builder.CreateCall(prepare_call(jl_simdloop_marker_func));
         return jl_cgval_t();
     }
     else if (head == goto_ifnot_sym) {
@@ -7437,6 +7438,13 @@ static void init_julia_llvm_env(Module *m)
     add_return_attr(jl_newbits_func, Attribute::NonNull);
     add_named_global(jl_newbits_func, (void*)jl_new_bits);
 
+    jl_simdloop_marker_func = Function::Create(FunctionType::get(T_void, {}, false),
+                                               Function::ExternalLinkage,
+                                               "julia.simdloop_marker");
+    jl_simdloop_marker_func->addFnAttr(Attribute::NoUnwind);
+    jl_simdloop_marker_func->addFnAttr(Attribute::NoRecurse);
+    jl_simdloop_marker_func->addFnAttr(Attribute::InaccessibleMemOnly);
+
     jl_typeof_func = Function::Create(FunctionType::get(T_prjlvalue, {T_prjlvalue}, false),
                                       Function::ExternalLinkage,
                                       "julia.typeof");

diff --git a/src/jitlayers.cpp b/src/jitlayers.cpp
@@ -134,6 +134,7 @@ void addOptimizationPasses(legacy::PassManagerBase *PM, int opt_level, bool dump
         PM->add(createLateLowerGCFramePass());
         PM->add(createLowerPTLSPass(dump_native));
 #endif
+        PM->add(createLowerSimdLoopPass());        // Annotate loop marked with "simdloop" as LLVM parallel loop
         if (dump_native)
             PM->add(createMultiVersioningPass());
         return;
@@ -145,8 +146,8 @@ void addOptimizationPasses(legacy::PassManagerBase *PM, int opt_level, bool dump
     }
     // list of passes from vmkit
     PM->add(createCFGSimplificationPass()); // Clean up disgusting code
-    PM->add(createDeadInstEliminationPass());
-    PM->add(createPromoteMemoryToRegisterPass()); // Kill useless allocas
+    PM->add(createDeadCodeEliminationPass());
+    PM->add(createSROAPass()); // Kill useless allocas
 
     // Due to bugs and missing features LLVM < 5.0, does not properly propagate
     // our invariants. We need to do GC rooting here. This reduces the

diff --git a/src/llvm-simdloop.cpp b/src/llvm-simdloop.cpp
@@ -55,16 +55,26 @@ bool annotateSimdLoop(BasicBlock *incr)
     return false;
 }
 
-/// Pass that lowers a loop marked by annotateSimdLoop.
 /// This pass should run after reduction variables have been converted to phi nodes,
 /// otherwise floating-point reductions might not be recognized as such and
 /// prevent SIMDization.
-struct LowerSIMDLoop: public LoopPass {
+struct LowerSIMDLoop : public ModulePass {
     static char ID;
-    LowerSIMDLoop() : LoopPass(ID) {}
+    LowerSIMDLoop() : ModulePass(ID)
+    {
+    }
+
+    protected:
+    void getAnalysisUsage(AnalysisUsage &AU) const override
+    {
+        ModulePass::getAnalysisUsage(AU);
+        AU.addRequired<LoopInfoWrapperPass>();
+        AU.addPreserved<LoopInfoWrapperPass>();
+        AU.setPreservesCFG();
+    }
 
-private:
-    bool runOnLoop(Loop *, LPPassManager &LPM) override;
+    private:
+    bool runOnModule(Module &M) override;
 
     /// Check if loop has "simd_loop" annotation.
     /// If present, the annotation is an MDNode attached to an instruction in the loop's latch.
@@ -160,41 +170,65 @@ void LowerSIMDLoop::enableUnsafeAlgebraIfReduction(PHINode *Phi, Loop *L) const
     }
 }
 
-bool LowerSIMDLoop::runOnLoop(Loop *L, LPPassManager &LPM)
+bool LowerSIMDLoop::runOnModule(Module &M)
 {
-    if (!simd_loop_mdkind) {
-        simd_loop_mdkind = L->getHeader()->getContext().getMDKindID("simd_loop");
-        simd_loop_md = MDNode::get(L->getHeader()->getContext(), ArrayRef<Metadata*>());
-    }
+    Function *simdloop_marker = M.getFunction("julia.simdloop_marker");
 
-    if (!hasSIMDLoopMetadata(L))
+    if (!simdloop_marker)
         return false;
 
-    DEBUG(dbgs() << "LSL: simd_loop found\n");
-    BasicBlock *Lh = L->getHeader();
-    DEBUG(dbgs() << "LSL: loop header: " << *Lh << "\n");
-    MDNode *n = L->getLoopID();
-    if (!n) {
-        // Loop does not have a LoopID yet, so give it one.
-        n = MDNode::get(Lh->getContext(), ArrayRef<Metadata*>(NULL));
-        n->replaceOperandWith(0,n);
-        L->setLoopID(n);
-    }
-    MDNode *m = MDNode::get(Lh->getContext(), ArrayRef<Metadata*>(n));
+    bool Changed = false;
+    std::vector<Instruction*> ToDelete;
+    for (User *U : simdloop_marker->users()) {
+        Instruction *I = cast<Instruction>(U);
+        ToDelete.push_back(I);
+        LoopInfo &LI = getAnalysis<LoopInfoWrapperPass>(*I->getParent()->getParent()).getLoopInfo();
+        Loop *L = LI.getLoopFor(I->getParent());
+        I->removeFromParent();
+        if (!L)
+            continue;
+
+        DEBUG(dbgs() << "LSL: simd_loop found\n");
+        BasicBlock *Lh = L->getHeader();
+        DEBUG(dbgs() << "LSL: loop header: " << *Lh << "\n");
+        MDNode *n = L->getLoopID();
+        if (!n) {
+            // Loop does not have a LoopID yet, so give it one.
+            n = MDNode::get(Lh->getContext(), ArrayRef<Metadata *>(NULL));
+            n->replaceOperandWith(0, n);
+            L->setLoopID(n);
+        }
+
+        assert(L->getLoopID());
 
-    // Mark memory references so that Loop::isAnnotatedParallel will return true for this loop.
-    for(Loop::block_iterator BBI = L->block_begin(), E=L->block_end(); BBI!=E; ++BBI)
-        for (BasicBlock::iterator I = (*BBI)->begin(), EE = (*BBI)->end(); I!=EE; ++I)
-            if (I->mayReadOrWriteMemory())
-                I->setMetadata("llvm.mem.parallel_loop_access", m);
-    assert(L->isAnnotatedParallel());
+        MDNode *m = MDNode::get(Lh->getContext(), ArrayRef<Metadata *>(n));
+
+        // Mark memory references so that Loop::isAnnotatedParallel will return true for this loop.
+        for (BasicBlock *BB : L->blocks()) {
+            for (Instruction &I : *BB) {
+                if (I.mayReadOrWriteMemory()) {
+                    I.setMetadata(LLVMContext::MD_mem_parallel_loop_access, m);
+                }
+            }
+        }
+        assert(L->isAnnotatedParallel());
+
+        // Mark floating-point reductions as okay to reassociate/commute.
+        for (BasicBlock::iterator I = Lh->begin(), E = Lh->end(); I != E; ++I) {
+            if (PHINode *Phi = dyn_cast<PHINode>(I))
+                enableUnsafeAlgebraIfReduction(Phi, L);
+            else
+                break;
+        }
+
+        Changed = true;
+    }
 
-    // Mark floating-point reductions as okay to reassociate/commute.
-    for (BasicBlock::iterator I = Lh->begin(), E = Lh->end(); I!=E; ++I)
-        if (PHINode *Phi = dyn_cast<PHINode>(I))
-            enableUnsafeAlgebraIfReduction(Phi,L);
+    for (Instruction *I : ToDelete)
+        I->deleteValue();
+    simdloop_marker->eraseFromParent();
 
-    return true;
+    return Changed;
 }
 
 char LowerSIMDLoop::ID = 0;

diff --git a/test/llvmpasses/simdloop.ll b/test/llvmpasses/simdloop.ll
@@ -1,5 +1,7 @@
 ; RUN: opt -load libjulia%shlibext -LowerSIMDLoop -S %s | FileCheck %s
 
+declare void @julia.simdloop_marker()
+
 define void @simd_test(double *%a, double *%b) {
 top:
   br label %loop
@@ -12,7 +14,8 @@ loop:
   %bval = load double, double *%aptr
   %cval = fadd double %aval, %bval
   store double %cval, double *%bptr
-  %nexti = add i64 %i, 1, !simd_loop !1
+  %nexti = add i64 %i, 1
+  call void @julia.simdloop_marker()
   %done = icmp sgt i64 %nexti, 500
   br i1 %done, label %loopdone, label %loop
 loopdone:
@@ -30,7 +33,8 @@ loop:
   %aval = load double, double *%aptr
   %nextv = fsub double %v, %aval
 ; CHECK: fsub fast double %v, %aval
-  %nexti = add i64 %i, 1, !simd_loop !1
+  %nexti = add i64 %i, 1
+  call void @julia.simdloop_marker()
   %done = icmp sgt i64 %nexti, 500
   br i1 %done, label %loopdone, label %loop
 loopdone: