MIPS[64]: Fix deoptimizer generate table for large deoptimization tables

Fix 233b6926

With the introduction of upfront deopt table generation, case where
deoptimization tables have more than 8000 entries become much more common
and bugs started to appear.

This CL fixes the issue, but it is far from perfect. We had to
increase number of instruction per table entry from 2 to 3 in
order to accommodate larger tables.

TEST=mjsunit/array-sort,mjsunit/json,mjsunit/md5
BUG=

Review-Url: https://codereview.chromium.org/2901393003
Cr-Commit-Position: refs/heads/master@{#45531}

src/mips/deoptimizer-mips.cc

@@ -323,14 +323,14 @@ void Deoptimizer::TableEntryGenerator::Generate() {
 
 
 // Maximum size of a table entry generated below.
-const int Deoptimizer::table_entry_size_ = 2 * Assembler::kInstrSize;
+const int Deoptimizer::table_entry_size_ = 3 * Assembler::kInstrSize;
 
 void Deoptimizer::TableEntryGenerator::GeneratePrologue() {
   Assembler::BlockTrampolinePoolScope block_trampoline_pool(masm());
 
   // Create a sequence of deoptimization entries.
   // Note that registers are still live when jumping to an entry.
-  Label table_start, done, done_special, trampoline_jump;
+  Label table_start, done, trampoline_jump;
   __ bind(&table_start);
   int kMaxEntriesBranchReach = (1 << (kImm16Bits - 2))/
      (table_entry_size_ /  Assembler::kInstrSize);
@@ -343,6 +343,7 @@ void Deoptimizer::TableEntryGenerator::GeneratePrologue() {
       DCHECK(is_int16(i));
       __ BranchShort(USE_DELAY_SLOT, &done);  // Expose delay slot.
       __ li(at, i);  // In the delay slot.
+      __ nop();
 
       DCHECK_EQ(table_entry_size_, masm()->SizeOfCodeGeneratedSince(&start));
     }
@@ -353,34 +354,29 @@ void Deoptimizer::TableEntryGenerator::GeneratePrologue() {
     __ Push(at);
   } else {
     // Uncommon case, the branch cannot reach.
-    // Create mini trampoline and adjust id constants to get proper value at
-    // the end of table.
-    for (int i = kMaxEntriesBranchReach; i > 1; i--) {
+    // Create mini trampoline to reach the end of the table
+    for (int i = 0, j = 0; i < count(); i++, j++) {
       Label start;
       __ bind(&start);
       DCHECK(is_int16(i));
-      __ BranchShort(USE_DELAY_SLOT, &trampoline_jump);  // Expose delay slot.
-      __ li(at, - i);  // In the delay slot.
+      if (j >= kMaxEntriesBranchReach) {
+        j = 0;
+        __ li(at, i);
+        __ bind(&trampoline_jump);
+        trampoline_jump = Label();
+        __ BranchShort(USE_DELAY_SLOT, &trampoline_jump);
+        __ nop();
+      } else {
+        __ BranchShort(USE_DELAY_SLOT, &trampoline_jump);  // Expose delay slot.
+        __ li(at, i);                                      // In the delay slot.
+        __ nop();
+      }
       DCHECK_EQ(table_entry_size_, masm()->SizeOfCodeGeneratedSince(&start));
     }
-    // Entry with id == kMaxEntriesBranchReach - 1.
-    __ bind(&trampoline_jump);
-    __ BranchShort(USE_DELAY_SLOT, &done_special);
-    __ li(at, -1);
-
-    for (int i = kMaxEntriesBranchReach ; i < count(); i++) {
-      Label start;
-      __ bind(&start);
-      DCHECK(is_int16(i));
-      __ BranchShort(USE_DELAY_SLOT, &done);  // Expose delay slot.
-      __ li(at, i);  // In the delay slot.
-    }
 
     DCHECK_EQ(masm()->SizeOfCodeGeneratedSince(&table_start),
         count() * table_entry_size_);
-    __ bind(&done_special);
-    __ addiu(at, at, kMaxEntriesBranchReach);
-    __ bind(&done);
+    __ bind(&trampoline_jump);
     __ Push(at);
   }
 }

src/mips64/deoptimizer-mips64.cc

@@ -323,14 +323,14 @@ void Deoptimizer::TableEntryGenerator::Generate() {
 
 
 // Maximum size of a table entry generated below.
-const int Deoptimizer::table_entry_size_ = 2 * Assembler::kInstrSize;
+const int Deoptimizer::table_entry_size_ = 3 * Assembler::kInstrSize;
 
 void Deoptimizer::TableEntryGenerator::GeneratePrologue() {
   Assembler::BlockTrampolinePoolScope block_trampoline_pool(masm());
 
   // Create a sequence of deoptimization entries.
   // Note that registers are still live when jumping to an entry.
-  Label table_start, done, done_special, trampoline_jump;
+  Label table_start, done, trampoline_jump;
   __ bind(&table_start);
   int kMaxEntriesBranchReach =
       (1 << (kImm16Bits - 2)) / (table_entry_size_ / Assembler::kInstrSize);
@@ -343,6 +343,7 @@ void Deoptimizer::TableEntryGenerator::GeneratePrologue() {
       DCHECK(is_int16(i));
       __ BranchShort(USE_DELAY_SLOT, &done);  // Expose delay slot.
       __ li(at, i);                      // In the delay slot.
+      __ nop();
 
       DCHECK_EQ(table_entry_size_, masm()->SizeOfCodeGeneratedSince(&start));
     }
@@ -353,34 +354,29 @@ void Deoptimizer::TableEntryGenerator::GeneratePrologue() {
     __ Push(at);
   } else {
     // Uncommon case, the branch cannot reach.
-    // Create mini trampoline and adjust id constants to get proper value at
-    // the end of table.
-    for (int i = kMaxEntriesBranchReach; i > 1; i--) {
+    // Create mini trampoline to reach the end of the table
+    for (int i = 0, j = 0; i < count(); i++, j++) {
       Label start;
       __ bind(&start);
       DCHECK(is_int16(i));
-      __ BranchShort(USE_DELAY_SLOT, &trampoline_jump);  // Expose delay slot.
-      __ li(at, -i);                                // In the delay slot.
+      if (j >= kMaxEntriesBranchReach) {
+        j = 0;
+        __ li(at, i);
+        __ bind(&trampoline_jump);
+        trampoline_jump = Label();
+        __ BranchShort(USE_DELAY_SLOT, &trampoline_jump);
+        __ nop();
+      } else {
+        __ BranchShort(USE_DELAY_SLOT, &trampoline_jump);  // Expose delay slot.
+        __ li(at, i);                                      // In the delay slot.
+        __ nop();
+      }
       DCHECK_EQ(table_entry_size_, masm()->SizeOfCodeGeneratedSince(&start));
     }
-    // Entry with id == kMaxEntriesBranchReach - 1.
-    __ bind(&trampoline_jump);
-    __ BranchShort(USE_DELAY_SLOT, &done_special);
-    __ li(at, -1);
-
-    for (int i = kMaxEntriesBranchReach; i < count(); i++) {
-      Label start;
-      __ bind(&start);
-      DCHECK(is_int16(i));
-      __ Branch(USE_DELAY_SLOT, &done);  // Expose delay slot.
-      __ li(at, i);                      // In the delay slot.
-    }
 
     DCHECK_EQ(masm()->SizeOfCodeGeneratedSince(&table_start),
               count() * table_entry_size_);
-    __ bind(&done_special);
-    __ daddiu(at, at, kMaxEntriesBranchReach);
-    __ bind(&done);
+    __ bind(&trampoline_jump);
     __ Push(at);
   }
 }