MIPS: [builtins] Fix MathMaxMin.

src/mips/builtins-mips.cc

Index: src/mips/builtins-mips.cc
diff --git a/src/mips/builtins-mips.cc b/src/mips/builtins-mips.cc
index a93e8ab4f890dc2adf6edd93d28739826c21b9be..ae3f40ce00015a8dc110bdd3bcf7be4d14a7205a 100644
--- a/src/mips/builtins-mips.cc
+++ b/src/mips/builtins-mips.cc
@@ -128,12 +128,12 @@ void Builtins::Generate_ArrayCode(MacroAssembler* masm) {
 // static
 void Builtins::Generate_MathMaxMin(MacroAssembler* masm, MathMaxMinKind kind) {
   // ----------- S t a t e -------------
-  //  -- a0                 : number of arguments
-  //  -- a1                 : function
-  //  -- cp                 : context
-  //  -- ra                 : return address
-  //  -- sp[(argc - n) * 8] : arg[n] (zero-based)
-  //  -- sp[(argc + 1) * 8] : receiver
+  //  -- a0                     : number of arguments
+  //  -- a1                     : function
+  //  -- cp                     : context
+  //  -- ra                     : return address
+  //  -- sp[(argc - n - 1) * 4] : arg[n] (zero-based)
+  //  -- sp[argc * 4]           : receiver
   // -----------------------------------
   Heap::RootListIndex const root_index =
       (kind == MathMaxMinKind::kMin) ? Heap::kInfinityValueRootIndex
@@ -143,17 +143,17 @@ void Builtins::Generate_MathMaxMin(MacroAssembler* masm, MathMaxMinKind kind) {
   // +Infinity), with the tagged value in t2 and the double value in f0.
   __ LoadRoot(t2, root_index);
   __ ldc1(f0, FieldMemOperand(t2, HeapNumber::kValueOffset));
-  __ Addu(a3, a0, Operand(1));
 
   Label done_loop, loop;
+  __ mov(a3, a0);
   __ bind(&loop);
   {
     // Check if all parameters done.
-    __ Subu(a0, a0, Operand(1));
-    __ Branch(&done_loop, lt, a0, Operand(zero_reg));
+    __ Subu(a3, a3, Operand(1));
+    __ Branch(&done_loop, lt, a3, Operand(zero_reg));

[ivica.bogosavljevic, 2016/07/01 12:40:50]

You could fill delay slot here, the next instruction writes to at, so if jump is
taken, it doesn't matter, otherwise you save one nop :)

[balazs.kilvady, 2016/07/01 13:17:25]

Since we support r6 I try to avoid delay slot usage in conditional branches as
with them we lose the chance for an r6 compact branch optimization. Especially
when the result of the previous op is the input of the branch condition.

 
     // Load the next parameter tagged value into a2.
-    __ Lsa(at, sp, a0, kPointerSizeLog2);
+    __ Lsa(at, sp, a3, kPointerSizeLog2);
     __ lw(a2, MemOperand(at));
 
     // Load the double value of the parameter into f2, maybe converting the
@@ -228,7 +228,9 @@ void Builtins::Generate_MathMaxMin(MacroAssembler* masm, MathMaxMinKind kind) {
   }
 
   __ bind(&done_loop);
-  __ Lsa(sp, sp, a3, kPointerSizeLog2);
+  // Drop all slots, including the receiver.
+  __ Addu(a0, a0, Operand(1));
+  __ Lsa(sp, sp, a0, kPointerSizeLog2);
   __ Ret(USE_DELAY_SLOT);
   __ mov(v0, t2);  // In delay slot.
 }