Fix printf uncommon use cases

Fix the following printf use cases:
1. `printf("%s", "");`
2. `printf("%s","foo\0foo")`

Author: aratajew

IGC/Compiler/CISACodeGen/CISABuilder.cpp

@@ -6220,9 +6220,6 @@ namespace IGC
                 else {
                     if (modMD->stringConstants.count(G))
                     {
-                        ConstantDataSequential* cds = dyn_cast<ConstantDataSequential>(G->getInitializer());
-                        IGC_ASSERT(cds && (cds->isCString() || cds->isString()));
-                        (void) cds;
                         programSyms.globalStringConst.emplace_back((vISA::GenSymType)symbolEntry.s_type, symbolEntry.s_offset, symbolEntry.s_size, G->getName().str());
                     }
                     else

IGC/Compiler/Optimizer/OpenCLPasses/OpenCLPrintf/OpenCLPrintfAnalysis.cpp

@@ -176,13 +176,31 @@ bool isPrintfOnlyStringConstantImpl(const llvm::Value *v, std::set<const llvm::U
 // path lead to a printf call.
 bool OpenCLPrintfAnalysis::isPrintfOnlyStringConstant(const llvm::GlobalVariable *GV)
 {
-    const llvm::Constant *initializer = GV->getInitializer();
-    if (!initializer)
-        return false;
-    const llvm::ConstantDataSequential* cds = llvm::dyn_cast<llvm::ConstantDataSequential>(initializer);
-    if (!cds || !cds->isCString() || !cds->isString())
+    const llvm::Constant* Initializer = GV->getInitializer();
+    if (!Initializer)
+    {
         return false;
+    }
+
+    bool IsNullTerminatedString = false;
+    if (const auto* cds = llvm::dyn_cast<llvm::ConstantDataSequential>(Initializer))
+    {
+        if (cds->isString())
+        {
+            StringRef Str = cds->getAsString();
+            IsNullTerminatedString = Str.contains(0);
+        }
+    }
+
+    bool IsZeroInitCharArray = Initializer->isZeroValue() &&
+                               isa<ArrayType>(Initializer->getType()) &&
+                               Initializer->getType()->getArrayElementType()->isIntegerTy(8);
+
+    if (IsNullTerminatedString || IsZeroInitCharArray)
+    {
+        std::set<const llvm::User*> Visited;
+        return isPrintfOnlyStringConstantImpl(GV, Visited);
+    }
 
-    std::set<const llvm::User *> visited;
-    return isPrintfOnlyStringConstantImpl(GV, visited);
+    return false;
 }

IGC/Compiler/Optimizer/OpenCLPasses/ProgramScopeConstants/ProgramScopeConstantAnalysis.cpp

@@ -129,9 +129,7 @@ bool ProgramScopeConstantAnalysis::runOnModule(Module& M)
 
         // When ZeBin is enabled, constant variables that are string literals
         // used only by printf will be stored in the second constant buffer.
-        ConstantDataSequential* cds = dyn_cast<ConstantDataSequential>(initializer);
         bool isZebinPrintfStringConst = Ctx->enableZEBinary() &&
-            cds && (cds->isCString() || cds->isString()) &&
             OpenCLPrintfAnalysis::isPrintfOnlyStringConstant(globalVar);
         // Here we follow SPV_EXT_relaxed_printf_string_address_space to relax
         // the address space requirement of printf strings and accept
@@ -155,7 +153,7 @@ bool ProgramScopeConstantAnalysis::runOnModule(Module& M)
             }
         }
 
-        if (initializer->isZeroValue())
+        if (initializer->isZeroValue() && !isZebinPrintfStringConst)
         {
             zeroInitializedGlobals.push_back(globalVar);
             continue;

IGC/Compiler/tests/ProgramScopeConstantAnalysis/printf-empty-string.ll

@@ -0,0 +1,37 @@
+;=========================== begin_copyright_notice ============================
+;
+; Copyright (C) 2025 Intel Corporation
+;
+; SPDX-License-Identifier: MIT
+;
+;============================ end_copyright_notice =============================
+
+; RUN: igc_opt -igc-programscope-constant-analysis -igc-serialize-metadata \
+; RUN:   -S < %s | FileCheck %s
+; ------------------------------------------------
+; ProgramScopeConstantAnalysis
+; ------------------------------------------------
+
+; Below LLVM IR was produced from the following OpenCL C code:
+;
+; kernel void test()
+; {
+;    printf("%s", "");
+; }
+
+; CHECK: !{!"stringConstants", ![[FORMAT_STR_MD:.*]], ![[STR_MD:.*]]}
+; CHECK: ![[FORMAT_STR_MD]] = !{!"stringConstantsSet[0]", [3 x i8] addrspace(2)* @.str}
+; CHECK: ![[STR_MD]] = !{!"stringConstantsSet[1]", [1 x i8] addrspace(2)* @.str.1}
+
+@.str = internal unnamed_addr addrspace(2) constant [3 x i8] c"%s\00", align 1
+@.str.1 = internal unnamed_addr addrspace(2) constant [1 x i8] zeroinitializer, align 1
+
+define spir_kernel void @test() {
+entry:
+  %0 = getelementptr inbounds [3 x i8], [3 x i8] addrspace(2)* @.str, i64 0, i64 0
+  %1 = getelementptr inbounds [1 x i8], [1 x i8] addrspace(2)* @.str.1, i64 0, i64 0
+  %call = call spir_func i32 (i8 addrspace(2)*, ...) @printf(i8 addrspace(2)* %0, i8 addrspace(2)* %1)
+  ret void
+}
+
+declare i32 @printf(i8 addrspace(2)*, ...)

IGC/Compiler/tests/ProgramScopeConstantAnalysis/printf-null-termination-inside-string.ll

@@ -0,0 +1,37 @@
+;=========================== begin_copyright_notice ============================
+;
+; Copyright (C) 2025 Intel Corporation
+;
+; SPDX-License-Identifier: MIT
+;
+;============================ end_copyright_notice =============================
+
+; RUN: igc_opt -igc-programscope-constant-analysis -igc-serialize-metadata \
+; RUN:   -S < %s | FileCheck %s
+; ------------------------------------------------
+; ProgramScopeConstantAnalysis
+; ------------------------------------------------
+
+; Below LLVM IR was produced from the following OpenCL C code:
+;
+; kernel void test()
+; {
+;    printf("%s","foo\0foo");
+; }
+
+; CHECK: !{!"stringConstants", ![[FORMAT_STR_MD:.*]], ![[FOO_STR_MD:.*]]}
+; CHECK: ![[FORMAT_STR_MD]] = !{!"stringConstantsSet[0]", [3 x i8] addrspace(2)* @.str}
+; CHECK: ![[FOO_STR_MD]] = !{!"stringConstantsSet[1]", [8 x i8] addrspace(2)* @.str.1}
+
+@.str = internal unnamed_addr addrspace(2) constant [3 x i8] c"%s\00", align 1
+@.str.1 = internal unnamed_addr addrspace(2) constant [8 x i8] c"foo\00foo\00", align 1
+
+define spir_kernel void @test() {
+entry:
+  %0 = getelementptr inbounds [3 x i8], [3 x i8] addrspace(2)* @.str, i64 0, i64 0
+  %1 = getelementptr inbounds [8 x i8], [8 x i8] addrspace(2)* @.str.1, i64 0, i64 0
+  %call = call spir_func i32 (i8 addrspace(2)*, ...) @printf(i8 addrspace(2)* %0, i8 addrspace(2)* %1)
+  ret void
+}
+
+declare i32 @printf(i8 addrspace(2)*, ...)