[liftoff][mv] Support multi-value loops in Liftoff

R=ahaas@chromium.org

Bug: v8:10408
Change-Id: I6a4a1d82333cf31520526c766357c78d3d8dd206
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2140941
Commit-Queue: Thibaud Michaud <thibaudm@chromium.org>
Reviewed-by: Andreas Haas <ahaas@chromium.org>
Cr-Commit-Position: refs/heads/master@{#67082}

src/wasm/baseline/liftoff-assembler.cc

@@ -548,6 +548,19 @@ LiftoffRegister LiftoffAssembler::PeekToRegister(int index,
   return reg;
 }
 
+void LiftoffAssembler::PrepareLoopArgs(int num) {
+  for (int i = 0; i < num; ++i) {
+    VarState& slot = cache_state_.stack_state.end()[-1 - i];
+    if (!slot.is_const()) continue;
+    RegClass rc =
+        kNeedI64RegPair && slot.type() == kWasmI64 ? kGpRegPair : kGpReg;
+    LiftoffRegister reg = GetUnusedRegister(rc);
+    LoadConstant(reg, slot.constant());
+    slot.MakeRegister(reg);
+    cache_state_.inc_used(reg);
+  }
+}
+
 void LiftoffAssembler::MergeFullStackWith(const CacheState& target,
                                           const CacheState& source) {
   DCHECK_EQ(source.stack_height(), target.stack_height());

src/wasm/baseline/liftoff-assembler.h

@@ -297,6 +297,10 @@ class LiftoffAssembler : public TurboAssembler {
   // but discarded with {stack_state.pop_back(count)}.
   LiftoffRegister PeekToRegister(int index, LiftoffRegList pinned);
 
+  // Ensure that the loop inputs are either in a register or spilled to the
+  // stack, so that we can merge different values on the back-edge.
+  void PrepareLoopArgs(int num);
+
   int NextSpillOffset(ValueType type) {
     int offset = TopSpillOffset() + SlotSizeForType(type);
     if (NeedsAlignment(type)) {

src/wasm/baseline/liftoff-compiler.cc

@@ -752,10 +752,6 @@ class LiftoffCompiler {
   void Block(FullDecoder* decoder, Control* block) {}
 
   void Loop(FullDecoder* decoder, Control* loop) {
-    if (loop->start_merge.arity > 0 || loop->end_merge.arity > 1) {
-      return unsupported(decoder, kMultiValue, "multi-value loop");
-    }
-
     // Before entering a loop, spill all locals to the stack, in order to free
     // the cache registers, and to avoid unnecessarily reloading stack values
     // into registers at branches.
@@ -763,6 +759,8 @@ class LiftoffCompiler {
     // pre-analysis of the function.
     __ SpillLocals();
 
+    __ PrepareLoopArgs(loop->start_merge.arity);
+
     // Loop labels bind at the beginning of the block.
     __ bind(loop->label.get());