[wasm] Preserve R2R variable debug information

docs/design/datacontracts/DebugInfo.md

@@ -225,6 +225,50 @@ Each variable entry in the Vars section is nibble-encoded as follows:
 
 Signed integers are encoded using the same unsigned scheme, with the sign bit stored in bit 0 (`value = unsigned >> 1`, negate if `unsigned & 1`). On x86, stack offsets are DWORD-aligned and stored divided by `sizeof(DWORD)`.
 
+### WebAssembly Variable Register Encoding
+
+WASM has no physical registers. Every register field in the Vars stream holds a WebAssembly local
+index biased past the reserved register numbers:
+
+```text
+register = localIndex + WASM_LOCAL_REGNUM_BASE
+```
+
+`WASM_LOCAL_REGNUM_BASE` is `3`, one past `REGNUM_AMBIENT_SP`. Values `0` through `2` remain
+`REGNUM_PC`, `REGNUM_COUNT`, and `REGNUM_AMBIENT_SP`. As on other targets, the value type is not
+encoded: it is implied by the variable's type and by the local's declaration in the WebAssembly
+function.
+
+`WASM_LOCAL_REGNUM_BASE` (`3`) and the stack base register (`2`, described below) are fixed
+constants of DebugInfo contract version `c1`; readers hardcode them. Changing either requires a new
+DebugInfo contract version. Because ReadyToRun images persist this encoding, it also requires a
+ReadyToRun format version change so static readers such as R2RDump do not misdecode older images.
+
+### WebAssembly Stack Base Encoding
+
+WASM `VLT_STK` and `VLT_STK2` records currently encode base register `2`.
+`REG_FPBASE`, `REG_SPBASE`, and `REGNUM_AMBIENT_SP` all have that value on this target, so the
+debug record identifies a logical frame-relative stack home; it does not identify a particular
+WebAssembly engine local.
+
+The absolute logical frame address is reconstructed by the runtime stack-walk and unwind
+protocol from shadow-stack linear memory. The engine's current per-function SP/FP local allocation
+is a separate code-generation detail:
+
+* Frame access allocates an FP value when a method has frame locals, uses `localloc`, or has
+  funclets.
+* Without `localloc`, the root function's FP aliases its SP, including methods that make calls.
+* `localloc` gives the root a distinct FP so later SP movement does not change frame-relative
+  addresses.
+* Funclets receive a distinct parent establishing FP; with `localloc`, that remains the root's
+  pre-adjustment frame base.
+
+The numeric WebAssembly local indices holding those values can vary with function parameters and
+compiler-created locals and are not part of this debug-info format. Readers must not infer the
+logical frame address from a hardcoded `$varN`, and the producer does not advertise SP/FP engine
+local indices. A future producer that changes stack records away from base `2` requires a
+coordinated, fail-loud reader update.
+
 ### Async Suspension Point APIs
 
 We also support decoding async suspension points (and their captured continuation-object locals) from the `AsyncInfo` chunk of the debug info blob. The chunk is present only for methods that the JIT compiled with runtime-async suspension points; for all other methods, `AsyncInfoSize` is `0` in the FAT header and the API returns an empty list.

src/coreclr/inc/cordebuginfo.h

@@ -299,6 +299,12 @@ class ICorDebugInfo
 
     };
 
+#ifdef TARGET_WASM
+    // WASM has no physical registers. A register field in a variable location holds a wasm
+    // local index biased past the reserved RegNum values above.
+    static constexpr uint32_t WASM_LOCAL_REGNUM_BASE = REGNUM_AMBIENT_SP + 1;
+#endif // TARGET_WASM
+
     // VarLoc describes the location of a native variable.  Note that currently, VLT_REG_BYREF and VLT_STK_BYREF
     // are only used for value types on X64.
     // [cDAC]: Mirrored in managed code (IDacDbiInterface.cs).

src/coreclr/jit/codegen.h

@@ -741,6 +741,12 @@ class CodeGen final : public CodeGenInterface
 
     IL_OFFSET siLastEndOffs; // IL offset of the (exclusive) end of the last block processed
 
+#if defined(TARGET_WASM)
+    // Wasm opens its method-wide scopes in the first emitted block rather than relying on the
+    // monotonic enter/exit cursors used by other targets.
+    bool siWasmScopesOpened;
+#endif // defined(TARGET_WASM)
+
     /*
     XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
     XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX

src/coreclr/jit/compiler.cpp

@@ -2901,13 +2901,6 @@ void Compiler::compInitOptions(JitFlags* jitFlags)
 
     opts.compScopeInfo = opts.compDbgInfo;
 
-#ifdef TARGET_WASM
-    // Wasm uses virtual registers that cannot be encoded in the
-    // ICorDebugInfo register scheme, and there is no native debugger
-    // to consume scope info, so disable it entirely.
-    opts.compScopeInfo = false;
-#endif
-
 #ifdef LATE_DISASM
     codeGen->getDisAssembler().disOpenForLateDisAsm(info.compMethodName, info.compClassName,
                                                     info.compMethodInfo->args.pSig);

src/coreclr/jit/ee_il_dll.cpp

@@ -866,7 +866,11 @@ void Compiler::eeDispVar(ICorDebugInfo::NativeVarInfo* var)
     {
         case CodeGenInterface::VLT_REG:
         case CodeGenInterface::VLT_REG_BYREF:
+#if defined(TARGET_WASM)
+            printf("$%u", (unsigned)var->loc.vlReg.vlrReg - ICorDebugInfo::WASM_LOCAL_REGNUM_BASE);
+#else
             printf("%s", getRegName(var->loc.vlReg.vlrReg));
+#endif
             if (var->loc.vlType == (ICorDebugInfo::VarLocType)CodeGenInterface::VLT_REG_BYREF)
             {
                 printf(" byref");

src/coreclr/jit/lclvars.cpp

@@ -1281,7 +1281,7 @@ unsigned Compiler::compMap2ILvarNum(unsigned varNum) const
     }
 
 #if defined(TARGET_WASM)
-    if (varNum == lvaWasmSpArg)
+    if ((varNum == lvaWasmSpArg) || (varNum == lvaWasmPortableEntryPointArg))
     {
         return (unsigned)ICorDebugInfo::UNKNOWN_ILNUM;
     }
@@ -1319,6 +1319,11 @@ unsigned Compiler::compMap2ILvarNum(unsigned varNum) const
     {
         varNum--;
     }
+
+    if ((lvaWasmPortableEntryPointArg != BAD_VAR_NUM) && (originalVarNum > lvaWasmPortableEntryPointArg))
+    {
+        varNum--;
+    }
 #endif
 
     if (varNum >= info.compLocalsCount)

src/coreclr/jit/scopeinfo.cpp

@@ -57,6 +57,28 @@ XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
 #include "emit.h"
 #include "codegen.h"
 
+#if defined(TARGET_WASM)
+// Stack VarLoc records use the reserved ambient SP register number, not a wasm local.
+static_assert(REG_FPBASE == REG_SPBASE);
+static_assert(REG_FPBASE == REG_NA);
+static_assert(static_cast<int>(REG_NA) == static_cast<int>(ICorDebugInfo::REGNUM_AMBIENT_SP));
+static_assert(static_cast<int>(ICorDebugInfo::REGNUM_AMBIENT_SP) == 2);
+
+//------------------------------------------------------------------------
+// siWasmDebugRegNum: Encode a wasm local as a debug-info register number.
+//
+// Arguments:
+//    reg - the JIT register representing a wasm local.
+//
+// Return Value:
+//    The wasm local index biased past the reserved debug-info register numbers.
+//
+static regNumber siWasmDebugRegNum(regNumber reg)
+{
+    return static_cast<regNumber>(ICorDebugInfo::WASM_LOCAL_REGNUM_BASE + WasmRegToIndex(reg));
+}
+#endif // defined(TARGET_WASM)
+
 //============================================================================
 //           siVarLoc functions
 //============================================================================
@@ -208,6 +230,30 @@ void CodeGenInterface::siVarLoc::storeVariableInRegisters(regNumber reg, regNumb
     // Note: mask registers (K0-K7) and XMM16+ are accepted but will produce
     // VLT_INVALID since they can't be encoded in debug info.
 
+#if defined(TARGET_WASM)
+    if (reg == REG_NA)
+    {
+        vlType = VLT_INVALID;
+        return;
+    }
+
+    assert(genIsValidReg(reg));
+
+    if (otherReg == REG_NA)
+    {
+        vlType       = VLT_REG;
+        vlReg.vlrReg = siWasmDebugRegNum(reg);
+    }
+    else
+    {
+        assert(genIsValidReg(otherReg));
+        vlType            = VLT_REG_REG;
+        vlRegReg.vlrrReg1 = siWasmDebugRegNum(reg);
+        vlRegReg.vlrrReg2 = siWasmDebugRegNum(otherReg);
+    }
+    return;
+#endif // defined(TARGET_WASM)
+
     if (otherReg == REG_NA)
     {
         if (genIsValidFloatReg(reg))
@@ -462,6 +508,12 @@ void CodeGenInterface::siVarLoc::siFillStackVarLoc(
 void CodeGenInterface::siVarLoc::siFillRegisterVarLoc(
     const LclVarDsc* varDsc, var_types type, regNumber baseReg, int offset, bool isFramePointerUsed)
 {
+#if defined(TARGET_WASM)
+    this->vlType       = VLT_REG;
+    this->vlReg.vlrReg = siWasmDebugRegNum(varDsc->GetRegNum());
+    return;
+#endif // defined(TARGET_WASM)
+
     switch (type)
     {
         case TYP_INT:
@@ -623,7 +675,11 @@ void CodeGenInterface::dumpSiVarLoc(const siVarLoc* varLoc) const
     {
         case VLT_REG:
         case VLT_REG_BYREF:
+#if defined(TARGET_WASM)
+            printf("$%u", (unsigned)varLoc->vlReg.vlrReg - ICorDebugInfo::WASM_LOCAL_REGNUM_BASE);
+#else
             printf("%s", getRegName(varLoc->vlReg.vlrReg));
+#endif
             if (varLoc->vlType == VLT_REG_BYREF)
             {
                 printf(" byref");
@@ -1589,6 +1645,10 @@ void CodeGen::siInit()
     siLastEndOffs = 0;
 
     m_compiler->compResetScopeLists();
+
+#if defined(TARGET_WASM)
+    siWasmScopesOpened = false;
+#endif // defined(TARGET_WASM)
 }
 
 /*****************************************************************************
@@ -1674,11 +1734,35 @@ void CodeGen::siBeginBlock(BasicBlock* block)
 void CodeGen::siOpenScopesForNonTrackedVars(const BasicBlock* block, unsigned int lastBlockILEndOffset)
 {
 #if defined(TARGET_WASM)
-    // TODO-WASM: Wasm structured control flow
-    // requirements are incompatible with debug codegen's
-    // desire to keep blocks in increasing IL offset
-    // order. Figure out the proper scope manipulations.
-    //
+    // Wasm is compiled only ahead of time, where getVars reports no explicit ranges, so every
+    // scope covers the whole method. Our elimination of irreducible flow does not emit blocks in
+    // increasing IL offset order, so open all scopes in the first emitted block.
+    if (m_compiler->opts.OptimizationDisabled() && !siWasmScopesOpened)
+    {
+        siWasmScopesOpened = true;
+
+        for (unsigned i = 0; i < m_compiler->info.compVarScopesCount; i++)
+        {
+            VarScopeDsc* varScope = &m_compiler->info.compVarScopes[i];
+            assert(varScope->vsdLifeBeg == 0);
+
+            LclVarDsc* lclVarDsc = m_compiler->lvaGetDesc(varScope->vsdVarNum);
+
+            // Only report locals that were referenced, if we're not doing debug codegen
+            if (m_compiler->opts.compDbgCode || (lclVarDsc->lvRefCnt() > 0))
+            {
+                JITDUMP("Scope info: opening scope, LVnum=%u [%03X..%03X)\n", varScope->vsdLVnum, varScope->vsdLifeBeg,
+                        varScope->vsdLifeEnd);
+
+                varLiveKeeper->siStartVariableLiveRange(lclVarDsc, varScope->vsdVarNum);
+            }
+            else
+            {
+                JITDUMP("Skipping open scope for V%02u, unreferenced\n", varScope->vsdVarNum);
+            }
+        }
+    }
+
     return;
 #endif // defined(TARGET_WASM)