scf.if

V6 of my silly MLIR based compiler is tagged.

December 28, 2025 clang/llvm , , , , , ,

Screenshot

There’s an existing toy MLIR dialect, part of the mlir tutorial documentation, so I’ve renamed my dialect from toy to silly, and updated all the references to ‘toy calculator’ to ‘silly compiler’, or ‘silly language’. There’s no good reason to use this language, nor the compiler, so this is very appropriate. It was, however, an excellent learning tool. The toy namespace is renamed, as are various file names, and all the MLIR operators, function prefixes, and so forth.

In addition to the big rename, other changes since the V5 tag include:

  1. A GET builtin (can now to I/O, not just O)
  2. FOR loop support.
  3. Something much closer to a consistent coding style now (FooBar for structures, fooBar for functions, no more use of all of PascalCase, camelCase, and underscore separated variables).
  4. Almost all of the auto variables have been purged for clarity.
  5. I’ve removed the ‘using namespace mlir’ in lowering.cpp.  Many of my mlir:: namespace references already had the namespace tag, so removing this allowed for more consistency.  I may revert this if it proves too cumbersome, but if I do, I’ll remove all the mlir:: qualifiers consistently (unless they are needed for disambiguation).
  6. User errors in the parser/builder no longer log the internal file:line:func for the code that spots them, but just the file:line location of the code with the error.  Those errors are now reported with mlir::emitError()
  7. Declarations in scf.for and scf.if/else regions are now supported.
  8. error test script now merged into bin/testit, so there’s just one script to run the regression test.
  9. Switched to /// style doxygen markup.

GET

Here’s a sample program with a GET call:

INT32 x;
GET x;
PRINT x;

and the corresponding MLIR output:

module {
  func.func @main() -> i32 {
    "silly.scope"() ({
      "silly.declare"() <{type = i32}> {sym_name = "x"} : () -> ()
      %0 = silly.get : i32
      silly.assign @x = %0 : i32
      %1 = silly.load @x : i32
      silly.print %1 : i32
      %c0_i32 = arith.constant 0 : i32
      "silly.return"(%c0_i32) : (i32) -> ()
    }) : () -> ()
    "silly.yield"() : () -> ()
  }
}

In the generated MLIR, I’ve split the GET builtin into an SSA for the get itself. In the example above, that’s returning the %0 value, and an internal AssignOp, kind of as if the statement was:

x = GET;

with the type information for the get riding on the assignment variable. That choice doesn’t model of the language in an ideal way. However, there are plenty of other places where my generated MLIR also isn’t a great one-to-one match for the language, so I don’t feel too bad about having done that, but might make different choices, if I wanted to have a lowering pass that transformed the silly dialect into something that represented a different language.

Here’s the corresponding LLVM-IR for that MLIR (with the DI stripped out)

declare void @__silly_print_i64(i64)

declare i32 @__silly_get_i32()

define i32 @main() !dbg !4 {
  %1 = alloca i32, i64 1, align 4
  %2 = call i32 @__silly_get_i32()
  store i32 %2, ptr %1, align 4
  %3 = load i32, ptr %1, align 4
  %4 = sext i32 %3 to i64
  call void @__silly_print_i64(i64 %4)
  ret i32 0
}

The use of the store/load pair that was related to the symbol references. There’s some remnant of that left in the assembly without optimization:

   0:   push   %rax
   1:   call   6 
                        2: R_X86_64_PLT32       __silly_get_i32-0x4
   6:   mov    %eax,0x4(%rsp)
   a:   movslq %eax,%rdi
   d:   call   12 
                        e: R_X86_64_PLT32       __silly_print_i64-0x4
  12:   xor    %eax,%eax
  14:   pop    %rcx
  15:   ret

but with optimization, we are left with everything in register:

   0:   push   %rax
   1:   call   6 
                        2: R_X86_64_PLT32       __silly_get_i32-0x4
   6:   movslq %eax,%rdi
   9:   call   e 
                        a: R_X86_64_PLT32       __silly_print_i64-0x4
   e:   xor    %eax,%eax
  10:   pop    %rcx
  11:   ret

FOR

Here’s a little FOR test program:

INT32 x;

FOR ( x : (1, 11) )
{
    PRINT x;
};

FOR ( x : (1, 11, 2) )
{
    PRINT x;
};

This prints 1-10 and 1,3,5,7,9 respectively. Here’s the MLIR (with location information stripped out):

module {
  func.func @main() -> i32 {
    "silly.scope"() ({
      "silly.declare"() <{type = i32}> {sym_name = "x"} : () -> ()
      %c1_i64 = arith.constant 1 : i64
      %0 = arith.trunci %c1_i64 : i64 to i32
      %c11_i64 = arith.constant 11 : i64
      %1 = arith.trunci %c11_i64 : i64 to i32
      %c1_i64_0 = arith.constant 1 : i64
      %2 = arith.trunci %c1_i64_0 : i64 to i32
      scf.for %arg0 = %0 to %1 step %2  : i32 {
        silly.assign @x = %arg0 : i32
        %6 = silly.load @x : i32
        silly.print %6 : i32
      }
      %c1_i64_1 = arith.constant 1 : i64
      %3 = arith.trunci %c1_i64_1 : i64 to i32
      %c11_i64_2 = arith.constant 11 : i64
      %4 = arith.trunci %c11_i64_2 : i64 to i32
      %c2_i64 = arith.constant 2 : i64
      %5 = arith.trunci %c2_i64 : i64 to i32
      scf.for %arg0 = %3 to %4 step %5  : i32 {
        silly.assign @x = %arg0 : i32
        %6 = silly.load @x : i32
        silly.print %6 : i32
      }
      %c0_i32 = arith.constant 0 : i32
      "silly.return"(%c0_i32) : (i32) -> ()
    }) : () -> ()
    "silly.yield"() : () -> ()
  }
}

Observe that I did something sneaky in there: I’ve inserted a ‘silly.assign’ from the scf.for loop induction variable at the beginning of the loop, so that subsequent symbol based lookups just work. It would be cleaner to make the FOR loop variable private to the loop body (and have the builder reference the SSA induction variable directly forOp.getRegion().front().getArgument(0), instead of requiring a variable in the enclosing scope, but I did it this way to avoid the need for any additional dwarf instrumentation for that variable — basically, I was being lazy, and letting implementation guide the language “design”. Is that a hack? Absolutely!

Here’s the corresponding LLVM-IR:

declare void @__silly_print_i64(i64)

define i32 @main() { 
  %1 = alloca i32, i64 1, align 4
    #dbg_declare(ptr %1, !9, !DIExpression(), !8)
  br label %2

2:                                                ; preds = %5, %0
  %3 = phi i32 [ 1, %0 ], [ %8, %5 ]
  %4 = icmp slt i32 %3, 11
  br i1 %4, label %5, label %9

5:                                                ; preds = %2
  store i32 %3, ptr %1, align 4
  %6 = load i32, ptr %1, align 4
  %7 = sext i32 %6 to i64
  call void @__silly_print_i64(i64 %7)
  %8 = add i32 %3, 1
  br label %2

9:                                                ; preds = %2
  br label %10

10:                                               ; preds = %13, %9
  %11 = phi i32 [ 1, %9 ], [ %16, %13 ]
  %12 = icmp slt i32 %11, 11
  br i1 %12, label %13, label %17

13:                                               ; preds = %10
  store i32 %11, ptr %1, align 4
  %14 = load i32, ptr %1, align 4
  %15 = sext i32 %14 to i64
  call void @__silly_print_i64(i64 %15)
  %16 = add i32 %11, 2
  br label %10

17:                                               ; preds = %10
  ret i32 0

; uselistorder directives
  uselistorder ptr %1, { 2, 3, 0, 1 }
}

and the unoptimized codegen:

   0:   push   %rbx
   1:   sub    $0x10,%rsp
   5:   mov    $0x1,%ebx
   a:   cmp    $0xa,%ebx
   d:   jg     23 
   f:   nop
  10:   mov    %ebx,0xc(%rsp)
  14:   movslq %ebx,%rdi
  17:   call   1c 
                        18: R_X86_64_PLT32      __silly_print_i64-0x4
  1c:   inc    %ebx
  1e:   cmp    $0xa,%ebx
  21:   jle    10 
  23:   mov    $0x1,%ebx
  28:   cmp    $0xa,%ebx
  2b:   jg     44 
  2d:   nopl   (%rax)
  30:   mov    %ebx,0xc(%rsp)
  34:   movslq %ebx,%rdi
  37:   call   3c 
                        38: R_X86_64_PLT32      __silly_print_i64-0x4
  3c:   add    $0x2,%ebx
  3f:   cmp    $0xa,%ebx
  42:   jle    30 
  44:   xor    %eax,%eax
  46:   add    $0x10,%rsp
  4a:   pop    %rbx
  4b:   ret

At O2 optimization, the assembly printer chooses to unroll both loops completely, generating code like:

   0:   push   %rax
   1:   mov    $0x1,%edi
   6:   call   b 
                        7: R_X86_64_PLT32       __silly_print_i64-0x4
   b:   mov    $0x2,%edi
  10:   call   15 
                        11: R_X86_64_PLT32      __silly_print_i64-0x4
  15:   mov    $0x3,%edi
  1a:   call   1f 
                        1b: R_X86_64_PLT32      __silly_print_i64-0x4
  1f:   mov    $0x4,%edi
  24:   call   29 
                        25: R_X86_64_PLT32      __silly_print_i64-0x4
  29:   mov    $0x5,%edi
  2e:   call   33 
                        2f: R_X86_64_PLT32      __silly_print_i64-0x4
  33:   mov    $0x6,%edi
  38:   call   3d 
                        39: R_X86_64_PLT32      __silly_print_i64-0x4
...

SCF Region declarations

In the V5 tag of the compiler, a program like this wouldn’t work:

INT32 x;

x = 3;

IF ( x < 4 )
{
  INT32 y;
  y = 42;
  PRINT y;
};

PRINT "Done.";

This is because my DeclareOp needs to be in a region that has an associated symbol table (my ScopeOp). I've dealt with this by changing the insertion point for any declares to the beginning of the ScopeOp for the function (either the implicit main function, or a user defined function).

MLIR for the above program now looks like this:

module {
  func.func @main() -> i32 {
    "silly.scope"() ({
      "silly.declare"() <{type = i32}> {sym_name = "y"} : () -> ()
      "silly.declare"() <{type = i32}> {sym_name = "x"} : () -> ()
      %c3_i64 = arith.constant 3 : i64
      silly.assign @x = %c3_i64 : i64
      %0 = silly.load @x : i32
      %c4_i64 = arith.constant 4 : i64
      %1 = "silly.less"(%0, %c4_i64) : (i32, i64) -> i1
      scf.if %1 {
        %c42_i64 = arith.constant 42 : i64
        silly.assign @y = %c42_i64 : i64
        %3 = silly.load @y : i32
        silly.print %3 : i32
      }
      %2 = "silly.string_literal"() <{value = "Done."}> : () -> !llvm.ptr
      silly.print %2 : !llvm.ptr
      %c0_i32 = arith.constant 0 : i32
      "silly.return"(%c0_i32) : (i32) -> ()
    }) : () -> ()
    "silly.yield"() : () -> ()
  }
}

The declares for x, y, are no longer in the program order, but no program can observe that internal change, as I don't provide any explicit addressing operations.

Here's the generated LLVM-IR for this program:

@str_0 = private constant [5 x i8] c"Done."

declare void @__silly_print_string(i64, ptr)

declare void @__silly_print_i64(i64)

define i32 @main() !dbg !4 {
  %1 = alloca i32, i64 1, align 4
  %2 = alloca i32, i64 1, align 4
  store i32 3, ptr %2, align 4
  %3 = load i32, ptr %2, align 4
  %4 = sext i32 %3 to i64
  %5 = icmp slt i64 %4, 4
  br i1 %5, label %6, label %9

6:                                                ; preds = %0
  store i32 42, ptr %1, align 4
  %7 = load i32, ptr %1, align 4
  %8 = sext i32 %7 to i64
  call void @__silly_print_i64(i64 %8)
  br label %9

9:                                                ; preds = %6, %0
  call void @__silly_print_string(i64 5, ptr @str_0)
  ret i32 0
}

Without optimization, the codegen is:

   0:   push   %rax
   1:   movl   $0x3,(%rsp)
   8:   xor    %eax,%eax
   a:   test   %al,%al
   c:   jne    20 
   e:   movl   $0x2a,0x4(%rsp)
  16:   mov    $0x2a,%edi
  1b:   call   20 
                        1c: R_X86_64_PLT32      __silly_print_i64-0x4
  20:   mov    $0x5,%edi
  25:   mov    $0x0,%esi
                        26: R_X86_64_32 .rodata
  2a:   call   2f 
                        2b: R_X86_64_PLT32      __silly_print_string-0x4
  2f:   xor    %eax,%eax
  31:   pop    %rcx
  32:   ret

And with optimization, the branching on constant values is purged, leaving just gorp for the print calls:

   0:   push   %rax
   1:   mov    $0x2a,%edi
   6:   call   b 
                        7: R_X86_64_PLT32       __silly_print_i64-0x4
   b:   mov    $0x5,%edi
  10:   mov    $0x0,%esi
                        11: R_X86_64_32 .rodata
  15:   call   1a 
                        16: R_X86_64_PLT32      __silly_print_string-0x4
  1a:   xor    %eax,%eax
  1c:   pop    %rcx
  1d:   ret

Added IF/ELSE support to my toy MLIR/LLVM compiler.

December 16, 2025 clang/llvm , , , , , , , , , , , ,

Screenshot

I dusted off my toy compiler this weekend, and had a bit of fun implementing IF and ELSE support. The MLIR+LLVM compiler infrastructure is impressively versatile, allowing me to implement a feature like this in a handful of hours. It was, however, an invasive feature addition, requiring grammar changes, parser/builder, and lowering adjustments, as well as additional lowering for the SCF dialect ops that were used.

I have an ELIF element in the grammar too, but haven’t done that yet (and haven’t done much testing yet.)  These are the grammar elements I now have defined:

ifelifelse
  : ifStatement
    elifStatement*
    elseStatement?
  ;

ifStatement
  : IF_TOKEN BRACE_START_TOKEN booleanValue BRACE_END_TOKEN SCOPE_START_TOKEN statement* SCOPE_END_TOKEN
  ;

elifStatement
  : ELIF_TOKEN BRACE_START_TOKEN booleanValue BRACE_END_TOKEN SCOPE_START_TOKEN statement* SCOPE_END_TOKEN
  ;

elseStatement
  : ELSE_TOKEN SCOPE_START_TOKEN statement* SCOPE_END_TOKEN
  ;

Previously, I had a single monolithic ifelifelse token, but the generated parser data structure was horrendously complicated:

  class  IfelifelseContext : public antlr4::ParserRuleContext {
  public:
    IfelifelseContext(antlr4::ParserRuleContext *parent, size_t invokingState);
    virtual size_t getRuleIndex() const override;
    antlr4::tree::TerminalNode *IF_TOKEN();
    std::vector<antlr4::tree::TerminalNode *> BRACE_START_TOKEN();
    antlr4::tree::TerminalNode* BRACE_START_TOKEN(size_t i);
    std::vector<BooleanValueContext *> booleanValue();
    BooleanValueContext* booleanValue(size_t i);
    std::vector<antlr4::tree::TerminalNode *> BRACE_END_TOKEN();
    antlr4::tree::TerminalNode* BRACE_END_TOKEN(size_t i);
    std::vector<antlr4::tree::TerminalNode *> SCOPE_START_TOKEN();
    antlr4::tree::TerminalNode* SCOPE_START_TOKEN(size_t i);
    std::vector<antlr4::tree::TerminalNode *> SCOPE_END_TOKEN();
    antlr4::tree::TerminalNode* SCOPE_END_TOKEN(size_t i);
    std::vector<StatementContext *> statement();
    StatementContext* statement(size_t i);
    std::vector<antlr4::tree::TerminalNode *> ELIF_TOKEN();
    antlr4::tree::TerminalNode* ELIF_TOKEN(size_t i);
    antlr4::tree::TerminalNode *ELSE_TOKEN();

    virtual void enterRule(antlr4::tree::ParseTreeListener *listener) override;
    virtual void exitRule(antlr4::tree::ParseTreeListener *listener) override;

  };

In particular, I don’t see a way to determine if a statement should be part of the IF, the ELIF, or the ELSE body. Splitting the grammar into pieces was much better, and leave me with:

  class  IfelifelseContext : public antlr4::ParserRuleContext {
  public:
    IfelifelseContext(antlr4::ParserRuleContext *parent, size_t invokingState);
    virtual size_t getRuleIndex() const override;
    IfStatementContext *ifStatement();
    std::vector<ElifStatementContext *> elifStatement();
    ElifStatementContext* elifStatement(size_t i);
    ElseStatementContext *elseStatement();

    virtual void enterRule(antlr4::tree::ParseTreeListener *listener) override;
    virtual void exitRule(antlr4::tree::ParseTreeListener *listener) override;

  };

which allows me to drill down into each of the lower level elements, and drive the codegen from those. In particular, I can define ANTLR4 parse tree walker callbacks:

        void exitIfStatement(ToyParser::IfStatementContext *ctx) override;
        void exitElseStatement(ToyParser::ElseStatementContext *ctx) override;
        void enterIfelifelse( ToyParser::IfelifelseContext *ctx ) override;

and drive the codegen from there. The choice to use enter/exit callbacks for the various statement objects locks you into a certain mode of programming. In particular, this means that I don’t have functions for parsing a generic statement, so I am forced to emit the if/else body statements indirectly. Here’s an example:

        mlir::Value conditionPredicate = MLIRListener::parsePredicate( loc, booleanValue );

        insertionPointStack.push_back( builder.saveInsertionPoint() );

        // Create the scf.if — it will be inserted at the current IP
        auto ifOp = builder.create<mlir::scf::IfOp>( loc, conditionPredicate );

        mlir::Block &thenBlock = ifOp.getThenRegion().front();
        builder.setInsertionPointToStart( &thenBlock );

Then in the exitIf callback, if there is an else statement to process, I create a block for that, set the insertion point for that, and let the statement processing continue. When that else processing is done, I can restore the insertion point to just after the scf.if, and the rest of the function generation can proceed. Here’s an example of the MLIR dump before any lowering:

module {
  func.func @main() -> i32 {
    "toy.scope"() ({
      "toy.declare"() <{type = i32}> {sym_name = "x"} : () -> ()
      %c3_i64 = arith.constant 3 : i64
      "toy.assign"(%c3_i64) <{var_name = @x}> : (i64) -> ()
      %0 = "toy.load"() <{var_name = @x}> : () -> i32
      %c4_i64 = arith.constant 4 : i64
      %1 = "toy.less"(%0, %c4_i64) : (i32, i64) -> i1
      scf.if %1 {
        %5 = "toy.string_literal"() <{value = "x < 4"}> : () -> !llvm.ptr
        toy.print %5 : !llvm.ptr
        %6 = "toy.string_literal"() <{value = "a second statement"}> : () -> !llvm.ptr
        toy.print %6 : !llvm.ptr
      } else {
        %5 = "toy.string_literal"() <{value = "!(x < 4) -- should be dead code"}> : () -> !llvm.ptr
        toy.print %5 : !llvm.ptr
      }
      %2 = "toy.load"() <{var_name = @x}> : () -> i32
      %c5_i64 = arith.constant 5 : i64
      %3 = "toy.less"(%c5_i64, %2) : (i64, i32) -> i1
      scf.if %3 {
        %5 = "toy.string_literal"() <{value = "x > 5"}> : () -> !llvm.ptr
        toy.print %5 : !llvm.ptr
      } else {
        %5 = "toy.string_literal"() <{value = "!(x > 5) -- should see this"}> : () -> !llvm.ptr
        toy.print %5 : !llvm.ptr
      }
      %4 = "toy.string_literal"() <{value = "Done."}> : () -> !llvm.ptr
      toy.print %4 : !llvm.ptr
      %c0_i32 = arith.constant 0 : i32
      "toy.return"(%c0_i32) : (i32) -> ()
    }) : () -> ()
    "toy.yield"() : () -> ()
  }
}

This is the IR for the following “program”:

INT32 x;

x = 3;

IF ( x < 4 )
{
   PRINT "x < 4";
   PRINT "a second statement";
}
ELSE
{
   PRINT "!(x < 4) -- should be dead code";
};

IF ( x > 5 )
{
   PRINT "x > 5";
}
ELSE
{
   PRINT "!(x > 5) -- should see this";
};

PRINT "Done.";

There are existing mechanisms for lowering the SCF dialect, so it doesn’t take much work. There was one lowering quirk that I didn’t expect to have to deal with. I lower the toy.print Op in two steps, first to toy.call, and then let all the toy.call lowering kick in — however, I was doing that as part of a somewhat hacky toy.scope lowering step. That didn’t work anymore, since I can now have functions that aren’t in the scope, but part of an scf if or else block body. To fix that, I had to switch to a more conventional CallOp lowering class:

    class CallOpLowering : public ConversionPattern {
    public:
      explicit CallOpLowering(MLIRContext* context)
          : ConversionPattern(toy::CallOp::getOperationName(), 1, context) {}

      LogicalResult matchAndRewrite(Operation* op, ArrayRef<Value> operands,
                                    ConversionPatternRewriter& rewriter) const override {
        auto callOp = cast<toy::CallOp>(op);
        auto loc = callOp.getLoc();

        // Get the callee symbol reference (stored as "callee" attribute)
        auto calleeAttr = callOp->getAttrOfType<FlatSymbolRefAttr>("callee");
        if (!calleeAttr)
          return failure();

        // Get result types (empty for void, one type for scalar return)
        TypeRange resultTypes = callOp.getResultTypes();

        auto mlirCall = rewriter.create<mlir::func::CallOp>( loc, resultTypes, 
                                                             calleeAttr, callOp.getOperands() );

        // Replace uses correctly
        if (!resultTypes.empty()) {
          // Non-void: replace the single result
          rewriter.replaceOp(op, mlirCall.getResults());
        } else {
          // Void: erase the op (no result to replace)
          rewriter.eraseOp(op);
        }

        return success();
      }
    };

I lower all calls (print and any other) first to mlir::func::CallOp, and then let existing mlir func lowering kick in and do the rest.

After my “stage I” lowering, this program is mostly translated into LLVM:

module attributes {llvm.ident = "toycalculator V2"} {
  llvm.mlir.global private constant @str_5(dense<[68, 111, 110, 101, 46]> : tensor<5xi8>) {addr_space = 0 : i32} : !llvm.array<5 x i8>
  llvm.mlir.global private constant @str_4(dense<[33, 40, 120, 32, 62, 32, 53, 41, 32, 45, 45, 32, 115, 104, 111, 117, 108, 100, 32, 115, 101, 101, 32, 116, 104, 105, 115]> : tensor<27xi8>) {addr_space = 0 : i32} : !llvm.array<27 x i8>
  llvm.mlir.global private constant @str_3(dense<[120, 32, 62, 32, 53]> : tensor<5xi8>) {addr_space = 0 : i32} : !llvm.array<5 x i8>
  llvm.mlir.global private constant @str_2(dense<[33, 40, 120, 32, 60, 32, 52, 41, 32, 45, 45, 32, 115, 104, 111, 117, 108, 100, 32, 98, 101, 32, 100, 101, 97, 100, 32, 99, 111, 100, 101]> : tensor<31xi8>) {addr_space = 0 : i32} : !llvm.array<31 x i8>
  llvm.mlir.global private constant @str_1(dense<[97, 32, 115, 101, 99, 111, 110, 100, 32, 115, 116, 97, 116, 101, 109, 101, 110, 116]> : tensor<18xi8>) {addr_space = 0 : i32} : !llvm.array<18 x i8>
  func.func private @__toy_print_string(i64, !llvm.ptr)
  llvm.mlir.global private constant @str_0(dense<[120, 32, 60, 32, 52]> : tensor<5xi8>) {addr_space = 0 : i32} : !llvm.array<5 x i8>
  func.func @main() -> i32 attributes {llvm.debug.subprogram = #llvm.di_subprogram, compileUnit = , sourceLanguage = DW_LANG_C, file = <"if.toy" in ".">, producer = "toycalculator", isOptimized = false, emissionKind = Full>, scope = #llvm.di_file<"if.toy" in ".">, name = "main", linkageName = "main", file = <"if.toy" in ".">, line = 1, scopeLine = 1, subprogramFlags = Definition, type = >>} {
    "toy.scope"() ({
      %0 = llvm.mlir.constant(1 : i64) : i64
      %1 = llvm.alloca %0 x i32 {alignment = 4 : i64, bindc_name = "x"} : (i64) -> !llvm.ptr
      llvm.intr.dbg.declare #llvm.di_local_variable, compileUnit = , sourceLanguage = DW_LANG_C, file = <"if.toy" in ".">, producer = "toycalculator", isOptimized = false, emissionKind = Full>, scope = #llvm.di_file<"if.toy" in ".">, name = "main", linkageName = "main", file = <"if.toy" in ".">, line = 1, scopeLine = 1, subprogramFlags = Definition, type = >>, name = "x", file = <"if.toy" in ".">, line = 1, alignInBits = 32, type = #llvm.di_basic_type> = %1 : !llvm.ptr
      %2 = llvm.mlir.constant(3 : i64) : i64
      %3 = llvm.trunc %2 : i64 to i32
      llvm.store %3, %1 {alignment = 4 : i64} : i32, !llvm.ptr
      %4 = llvm.load %1 : !llvm.ptr -> i32
      %5 = llvm.mlir.constant(4 : i64) : i64
      %6 = llvm.sext %4 : i32 to i64
      %7 = llvm.icmp "slt" %6, %5 : i64
      scf.if %7 {
        %15 = llvm.mlir.addressof @str_0 : !llvm.ptr
        %16 = llvm.mlir.constant(5 : i64) : i64
        "toy.call"(%16, %15) <{callee = @__toy_print_string}> : (i64, !llvm.ptr) -> ()
        %17 = llvm.mlir.addressof @str_1 : !llvm.ptr
        %18 = llvm.mlir.constant(18 : i64) : i64
        "toy.call"(%18, %17) <{callee = @__toy_print_string}> : (i64, !llvm.ptr) -> ()
      } else {
        %15 = llvm.mlir.addressof @str_2 : !llvm.ptr
        %16 = llvm.mlir.constant(31 : i64) : i64
        "toy.call"(%16, %15) <{callee = @__toy_print_string}> : (i64, !llvm.ptr) -> ()
      }
      %8 = llvm.load %1 : !llvm.ptr -> i32
      %9 = llvm.mlir.constant(5 : i64) : i64
      %10 = llvm.sext %8 : i32 to i64
      %11 = llvm.icmp "slt" %9, %10 : i64
      scf.if %11 {
        %15 = llvm.mlir.addressof @str_3 : !llvm.ptr
        %16 = llvm.mlir.constant(5 : i64) : i64
        "toy.call"(%16, %15) <{callee = @__toy_print_string}> : (i64, !llvm.ptr) -> ()
      } else {
        %15 = llvm.mlir.addressof @str_4 : !llvm.ptr
        %16 = llvm.mlir.constant(27 : i64) : i64
        "toy.call"(%16, %15) <{callee = @__toy_print_string}> : (i64, !llvm.ptr) -> ()
      }
      %12 = llvm.mlir.addressof @str_5 : !llvm.ptr
      %13 = llvm.mlir.constant(5 : i64) : i64
      "toy.call"(%13, %12) <{callee = @__toy_print_string}> : (i64, !llvm.ptr) -> ()
      %14 = llvm.mlir.constant(0 : i32) : i32
      "toy.return"(%14) : (i32) -> ()
    }) : () -> ()
    "toy.yield"() : () -> ()
  }
}

The only things that are left in my MLIR toy dialect are “toy.scope”, “toy.call”, “toy.yield”, and “toy.return”. After my second lowering pass, everything is in the MLIR LLVM dialect:

llvm.mlir.global private constant @str_5(dense<[68, 111, 110, 101, 46]> : tensor<5xi8>) {addr_space = 0 : i32} : !llvm.array<5 x i8>
llvm.mlir.global private constant @str_4(dense<[33, 40, 120, 32, 62, 32, 53, 41, 32, 45, 45, 32, 115, 104, 111, 117, 108, 100, 32, 115, 101, 101, 32, 116, 104, 105, 115]> : tensor<27xi8>) {addr_space = 0 : i32} : !llvm.array<27 x i8>
llvm.mlir.global private constant @str_3(dense<[120, 32, 62, 32, 53]> : tensor<5xi8>) {addr_space = 0 : i32} : !llvm.array<5 x i8>
llvm.mlir.global private constant @str_2(dense<[33, 40, 120, 32, 60, 32, 52, 41, 32, 45, 45, 32, 115, 104, 111, 117, 108, 100, 32, 98, 101, 32, 100, 101, 97, 100, 32, 99, 111, 100, 101]> : tensor<31xi8>) {addr_space = 0 : i32} : !llvm.array<31 x i8>
llvm.mlir.global private constant @str_1(dense<[97, 32, 115, 101, 99, 111, 110, 100, 32, 115, 116, 97, 116, 101, 109, 101, 110, 116]> : tensor<18xi8>) {addr_space = 0 : i32} : !llvm.array<18 x i8>
func.func private @__toy_print_string(i64, !llvm.ptr)
llvm.mlir.global private constant @str_0(dense<[120, 32, 60, 32, 52]> : tensor<5xi8>) {addr_space = 0 : i32} : !llvm.array<5 x i8>
func.func @main() -> i32 attributes {llvm.debug.subprogram = #llvm.di_subprogram, compileUnit = , sourceLanguage = DW_LANG_C, file = <"if.toy" in ".">, producer = "toycalculator", isOptimized = false, emissionKind = Full>, scope = #llvm.di_file<"if.toy" in ".">, name = "main", linkageName = "main", file = <"if.toy" in ".">, line = 1, scopeLine = 1, subprogramFlags = Definition, type = >>} {
  %0 = llvm.mlir.constant(1 : i64) : i64
  %1 = llvm.alloca %0 x i32 {alignment = 4 : i64, bindc_name = "x"} : (i64) -> !llvm.ptr
  llvm.intr.dbg.declare #llvm.di_local_variable, compileUnit = , sourceLanguage = DW_LANG_C, file = <"if.toy" in ".">, producer = "toycalculator", isOptimized = false, emissionKind = Full>, scope = #llvm.di_file<"if.toy" in ".">, name = "main", linkageName = "main", file = <"if.toy" in ".">, line = 1, scopeLine = 1, subprogramFlags = Definition, type = >>, name = "x", file = <"if.toy" in ".">, line = 1, alignInBits = 32, type = #llvm.di_basic_type> = %1 : !llvm.ptr
  %2 = llvm.mlir.constant(3 : i64) : i64
  %3 = llvm.trunc %2 : i64 to i32
  llvm.store %3, %1 {alignment = 4 : i64} : i32, !llvm.ptr
  %4 = llvm.load %1 : !llvm.ptr -> i32
  %5 = llvm.mlir.constant(4 : i64) : i64
  %6 = llvm.sext %4 : i32 to i64
  %7 = llvm.icmp "slt" %6, %5 : i64
  llvm.cond_br %7, ^bb1, ^bb2
^bb1:  // pred: ^bb0
  %8 = llvm.mlir.addressof @str_0 : !llvm.ptr
  %9 = llvm.mlir.constant(5 : i64) : i64
  call @__toy_print_string(%9, %8) : (i64, !llvm.ptr) -> ()
  %10 = llvm.mlir.addressof @str_1 : !llvm.ptr
  %11 = llvm.mlir.constant(18 : i64) : i64
  call @__toy_print_string(%11, %10) : (i64, !llvm.ptr) -> ()
  llvm.br ^bb3
^bb2:  // pred: ^bb0
  %12 = llvm.mlir.addressof @str_2 : !llvm.ptr
  %13 = llvm.mlir.constant(31 : i64) : i64
  call @__toy_print_string(%13, %12) : (i64, !llvm.ptr) -> ()
  llvm.br ^bb3
^bb3:  // 2 preds: ^bb1, ^bb2
  %14 = llvm.load %1 : !llvm.ptr -> i32
  %15 = llvm.mlir.constant(5 : i64) : i64
  %16 = llvm.sext %14 : i32 to i64
  %17 = llvm.icmp "slt" %15, %16 : i64
  llvm.cond_br %17, ^bb4, ^bb5
^bb4:  // pred: ^bb3
  %18 = llvm.mlir.addressof @str_3 : !llvm.ptr
  %19 = llvm.mlir.constant(5 : i64) : i64
  call @__toy_print_string(%19, %18) : (i64, !llvm.ptr) -> ()
  llvm.br ^bb6
^bb5:  // pred: ^bb3
  %20 = llvm.mlir.addressof @str_4 : !llvm.ptr
  %21 = llvm.mlir.constant(27 : i64) : i64
  call @__toy_print_string(%21, %20) : (i64, !llvm.ptr) -> ()
  llvm.br ^bb6
^bb6:  // 2 preds: ^bb4, ^bb5
  %22 = llvm.mlir.addressof @str_5 : !llvm.ptr
  %23 = llvm.mlir.constant(5 : i64) : i64
  call @__toy_print_string(%23, %22) : (i64, !llvm.ptr) -> ()
  %24 = llvm.mlir.constant(0 : i32) : i32
  return %24 : i32
}

Once assembled, we are left with:

0000000000000000 
:
   0:   push   %rax
   1:   movl   $0x3,0x4(%rsp)
   9:   xor    %eax,%eax
   b:   test   %al,%al
   d:   jne    2a 
   f:   mov    $0x5,%edi
  14:   mov    $0x0,%esi
                        15: R_X86_64_32 .rodata+0x62
  19:   call   1e 
                        1a: R_X86_64_PLT32      __toy_print_string-0x4
  1e:   mov    $0x12,%edi
  23:   mov    $0x0,%esi
                        24: R_X86_64_32 .rodata+0x50
  28:   jmp    34 
  2a:   mov    $0x1f,%edi
  2f:   mov    $0x0,%esi
                        30: R_X86_64_32 .rodata+0x30
  34:   call   39 
                        35: R_X86_64_PLT32      __toy_print_string-0x4
  39:   movslq 0x4(%rsp),%rax
  3e:   cmp    $0x6,%rax
  42:   jl     50 
  44:   mov    $0x5,%edi
  49:   mov    $0x0,%esi
                        4a: R_X86_64_32 .rodata+0x2b
  4e:   jmp    5a 
  50:   mov    $0x1b,%edi
  55:   mov    $0x0,%esi
                        56: R_X86_64_32 .rodata+0x10
  5a:   call   5f 
                        5b: R_X86_64_PLT32      __toy_print_string-0x4
  5f:   mov    $0x5,%edi
  64:   mov    $0x0,%esi
                        65: R_X86_64_32 .rodata
  69:   call   6e 
                        6a: R_X86_64_PLT32      __toy_print_string-0x4
  6e:   xor    %eax,%eax
  70:   pop    %rcx
  71:   ret

Of course, we may enable optimization, and get something much nicer. With -O2, by the time we are done the LLVM lowering, all the constant propagation has kicked in, leaving just:

; ModuleID = 'if.toy'
source_filename = "if.toy"
target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-i128:128-f80:128-n8:16:32:64-S128"
target triple = "x86_64-unknown-linux-gnu"

@str_5 = private constant [5 x i8] c"Done."
@str_4 = private constant [27 x i8] c"!(x > 5) -- should see this"
@str_1 = private constant [18 x i8] c"a second statement"
@str_0 = private constant [5 x i8] c"x < 4"

declare void @__toy_print_string(i64, ptr) local_unnamed_addr

define noundef i32 @main() local_unnamed_addr !dbg !4 {
    #dbg_value(i32 3, !8, !DIExpression(), !10)
  tail call void @__toy_print_string(i64 5, ptr nonnull @str_0), !dbg !11
  tail call void @__toy_print_string(i64 18, ptr nonnull @str_1), !dbg !12
  tail call void @__toy_print_string(i64 27, ptr nonnull @str_4), !dbg !13
  tail call void @__toy_print_string(i64 5, ptr nonnull @str_5), !dbg !14
  ret i32 0, !dbg !14
}

Our program is reduced to just a couple of print statements:

0000000000400470 
:
  400470:       50                      push   %rax
  400471:       bf 05 00 00 00          mov    $0x5,%edi
  400476:       be 12 12 40 00          mov    $0x401212,%esi
  40047b:       e8 f0 fe ff ff          call   400370 <__toy_print_string@plt>
  400480:       bf 12 00 00 00          mov    $0x12,%edi
  400485:       be 00 12 40 00          mov    $0x401200,%esi
  40048a:       e8 e1 fe ff ff          call   400370 <__toy_print_string@plt>
  40048f:       bf 1b 00 00 00          mov    $0x1b,%edi
  400494:       be e0 11 40 00          mov    $0x4011e0,%esi
  400499:       e8 d2 fe ff ff          call   400370 <__toy_print_string@plt>
  40049e:       bf 05 00 00 00          mov    $0x5,%edi
  4004a3:       be d0 11 40 00          mov    $0x4011d0,%esi
  4004a8:       e8 c3 fe ff ff          call   400370 <__toy_print_string@plt>
  4004ad:       31 c0                   xor    %eax,%eax
  4004af:       59                      pop    %rcx
  4004b0:       c3                      ret

We've seen how easy it was to implement enough control flow to almost make the toy language useful. Another side effect of this MLIR+LLVM infrastructure, is that our IF/ELSE debugging support comes for free (having paid the cost earlier of having figured out how to emit the dwarf instrumentation). Here's an example:

> gdb -q out/if
Reading symbols from out/if...
(gdb) b main
Breakpoint 1 at 0x400471: file if.toy, line 3.
(gdb) run
Starting program: /home/pjoot/toycalculator/samples/out/if 
[Thread debugging using libthread_db enabled]
Using host libthread_db library "/lib64/libthread_db.so.1".

Breakpoint 1, main () at if.toy:3
3       x = 3;
(gdb) l
1       INT32 x;
2
3       x = 3;
4
5       IF ( x < 4 )
6       {
7          PRINT "x < 4";
8          PRINT "a second statement";
9       }
10      ELSE
(gdb) b 7
Breakpoint 2 at 0x40047f: file if.toy, line 7.
(gdb) c
Continuing.

Breakpoint 2, main () at if.toy:7
7          PRINT "x < 4";
(gdb) l
2
3       x = 3;
4
5       IF ( x < 4 )
6       {
7          PRINT "x < 4";
8          PRINT "a second statement";
9       }
10      ELSE
11      {
(gdb) p x
$1 = 3
(gdb) l
12         PRINT "!(x < 4) -- should be dead code";
13      };
14
15      IF ( x > 5 )
16      {
17         PRINT "x > 5";
18      }
19      ELSE
20      {
21         PRINT "!(x > 5) -- should see this";
(gdb) b 21
Breakpoint 3 at 0x4004c0: file if.toy, line 21.
(gdb) c
Continuing.
x < 4
a second statement

Breakpoint 3, main () at if.toy:21
21         PRINT "!(x > 5) -- should see this";