[vm/compiler] "fast"-path for x64 division

Rationale:
Since 64-bit division requires twice as many cycles
and has much higher latency compared to the 32-bit
division, even for a non-speculative 64-bit path,
a 32-bit "fast path" makes sense.

Speedup:
About 2x for cases that fit 32-bits. No noticable
slowdown for the 64-bit cases.

flutter/flutter#19677
#33967

Change-Id: I0d3b44564fee2cda03fc36f089a4424084732de0
Reviewed-on: https://dart-review.googlesource.com/69200
Commit-Queue: Aart Bik <[email protected]>
Reviewed-by: Vyacheslav Egorov <[email protected]>

Author: aartbik

runtime/vm/compiler/backend/il_x64.cc

@@ -5326,9 +5326,32 @@ static void EmitInt64ModTruncDiv(FlowGraphCompiler* compiler,
   //   out = left % right
   // or
   //   out = left / right.
+  //
+  // Note that since 64-bit division requires twice as many cycles
+  // and has much higher latency compared to the 32-bit division,
+  // even for this non-speculative 64-bit path we add a "fast path".
+  // Integers are untagged at this stage, so testing if sign extending
+  // the lower half of each operand equals the full operand, effectively
+  // tests if the values fit in 32-bit operands (and the slightly
+  // dangerous division by -1 has been handled above already).
   ASSERT(left == RAX);
+  ASSERT(right != RDX);  // available at this stage
+  Label div_64;
+  Label div_merge;
+  __ movsxd(RDX, left);
+  __ cmpq(RDX, left);
+  __ j(NOT_EQUAL, &div_64, Assembler::kNearJump);
+  __ movsxd(RDX, right);
+  __ cmpq(RDX, right);
+  __ j(NOT_EQUAL, &div_64, Assembler::kNearJump);
+  __ cdq();         // sign-ext eax into edx:eax
+  __ idivl(right);  // quotient eax, remainder edx
+  __ movsxd(out, out);
+  __ jmp(&div_merge, Assembler::kNearJump);
+  __ Bind(&div_64);
   __ cqo();         // sign-ext rax into rdx:rax
   __ idivq(right);  // quotient rax, remainder rdx
+  __ Bind(&div_merge);
   if (op_kind == Token::kMOD) {
     ASSERT(out == RDX);
     ASSERT(tmp == RAX);

tests/language_2/vm/modtruncdiv_int_test.dart

@@ -176,6 +176,34 @@ main() {
     doModVars((3 << 32) - 1, (3 << 32) + 1, (3 << 32) - 1, (3 << 32) + 1);
     Expect.equals(38654705666, acc);
 
+    acc = 0;
+    doModVars(minInt32 - 4, minInt32 + 4, -11, -1);
+    Expect.equals(239, acc);
+
+    acc = 0;
+    doModVars(minInt32 - 4, minInt32 + 4, 2, 7);
+    Expect.equals(85, acc);
+
+    acc = 0;
+    doModVars(minInt32 - 4, minInt32 + 4, minInt32 - 4, minInt32 + 4);
+    Expect.equals(77309411268, acc);
+
+    acc = 0;
+    doModVars(minInt32 - 4, minInt32 + 4, maxInt32 - 4, maxInt32 + 4);
+    Expect.equals(96636763974, acc);
+
+    acc = 0;
+    doModVars(maxInt32 - 4, maxInt32 + 4, 2, 7);
+    Expect.equals(104, acc);
+
+    acc = 0;
+    doModVars(maxInt32 - 4, maxInt32 + 4, minInt32 - 4, minInt32 + 4);
+    Expect.equals(96636764139, acc);
+
+    acc = 0;
+    doModVars(maxInt32 - 4, maxInt32 + 4, maxInt32 - 4, maxInt32 + 4);
+    Expect.equals(77309411352, acc);
+
     acc = 0;
     doTruncDivVars(3, 5, 2, 6);
     Expect.equals(11, acc);
@@ -200,6 +228,38 @@ main() {
     doTruncDivVars(minInt64, minInt64, -1, -1);
     Expect.equals(minInt64, acc);
 
+    acc = 0;
+    doTruncDivVars(minInt32 - 4, minInt32 + 4, -11, -1);
+    Expect.equals(58366234918, acc);
+
+    acc = 0;
+    doTruncDivVars(minInt32 - 4, minInt32 + 4, 2, 7);
+    Expect.equals(-30785711991, acc);
+
+    acc = 0;
+    doTruncDivVars(minInt32 - 4, minInt32 + 4, minInt32 - 4, minInt32 + 4);
+    Expect.equals(45, acc);
+
+    acc = 0;
+    doTruncDivVars(minInt32 - 4, minInt32 + 4, maxInt32 - 4, maxInt32 + 4);
+    Expect.equals(-53, acc);
+
+    acc = 0;
+    doTruncDivVars(maxInt32 - 4, maxInt32 + 4, 2, 7);
+    Expect.equals(30785711975, acc);
+
+    acc = 0;
+    doTruncDivVars(maxInt32 - 4, maxInt32 + 4, minInt32 - 4, minInt32 + 4);
+    Expect.equals(-36, acc);
+
+    acc = 0;
+    doTruncDivVars(maxInt32 - 4, maxInt32 + 4, maxInt32 - 4, maxInt32 + 4);
+    Expect.equals(45, acc);
+
+    acc = 0;
+    doTruncDivVars(maxInt32 - 4, maxInt32 + 4, 1, 7);
+    Expect.equals(50113064798, acc);
+
     // Exceptions at the right time.
 
     acc = 0;