[Tarantool-patches] [PATCH luajit 1/2] Optimize table.new() with constant args to (sinkable) IR_TNEW.

[Sergey Kaplun]

From: Mike Pall <mike>

Thanks to Peter Cawley.

(cherry picked from commit d1236a4caa999b29e774ef5103df3b424d821d9b)

This patch adds optimization for calls of `table.new()` with constant
argument refs when asize is in the proper range (i.e., is less than IR
operand width). The call is replaced with IR TNEW. This opens up
opportunities for other optimizations in the pipeline.

Sergey Kaplun:
* added the description and the test for the feature

Part of tarantool/tarantool#9595
---
 src/lj_ffrecord.c                             |   9 ++
 .../lj-1128-table-new-opt-tnew.test.lua       | 112 ++++++++++++++++++
 2 files changed, 121 insertions(+)
 create mode 100644 test/tarantool-tests/lj-1128-table-new-opt-tnew.test.lua

diff --git a/src/lj_ffrecord.c b/src/lj_ffrecord.c
index 99a6b918..7c6e3590 100644
--- a/src/lj_ffrecord.c
+++ b/src/lj_ffrecord.c
@@ -1065,6 +1065,15 @@ static void LJ_FASTCALL recff_table_new(jit_State *J, RecordFFData *rd)
 {
   TRef tra = lj_opt_narrow_toint(J, J->base[0]);
   TRef trh = lj_opt_narrow_toint(J, J->base[1]);
+  if (tref_isk(tra) && tref_isk(trh)) {
+    int32_t a = IR(tref_ref(tra))->i;
+    if (a < 0x7fff) {
+      uint32_t hbits = hsize2hbits(IR(tref_ref(trh))->i);
+      a = a > 0 ? a+1 : 0;
+      J->base[0] = emitir(IRTG(IR_TNEW, IRT_TAB), (uint32_t)a, hbits);
+      return;
+    }
+  }
   J->base[0] = lj_ir_call(J, IRCALL_lj_tab_new_ah, tra, trh);
   UNUSED(rd);
 }
diff --git a/test/tarantool-tests/lj-1128-table-new-opt-tnew.test.lua b/test/tarantool-tests/lj-1128-table-new-opt-tnew.test.lua
new file mode 100644
index 00000000..805e6de6
--- /dev/null
+++ b/test/tarantool-tests/lj-1128-table-new-opt-tnew.test.lua
@@ -0,0 +1,112 @@
+local tap = require('tap')
+local test = tap.test('lj-1128-table-new-opt-tnew'):skipcond({
+  ['Test requires JIT enabled'] = not jit.status(),
+  ['Disabled on *BSD due to #4819'] = jit.os == 'BSD',
+})
+
+-- Test LuaJIT optimization when a call to the `lj_tab_new_ah()`
+-- is replaced with the corresponding TNEW IR.
+-- See also https://github.com/LuaJIT/LuaJIT/issues/1128.
+
+local jparse = require('utils').jit.parse
+
+-- API follows the semantics of `lua_createtable()`.
+local table_new = require('table.new')
+
+-- `hbits` for different `hsizes`, see <lj_tab.h> for details.
+local HBITS = {
+  [1] = 1,
+  [3] = 2,
+}
+
+-- XXX: Avoid any other traces compilation due to hotcount
+-- collisions for predictable results.
+jit.off()
+jit.flush()
+
+test:plan(10)
+
+jit.on()
+jit.opt.start('hotloop=1')
+jparse.start('i')
+
+local anchor
+
+for _ = 1, 4 do
+  anchor = table_new(1, 1)
+end
+
+local traces = jparse.finish()
+jit.off()
+
+test:ok(type(anchor) == 'table', 'base result')
+test:ok(traces[1]:has_ir(('TNEW.*#2.*#%d'):format(HBITS[1])), 'base IR value')
+
+jit.flush()
+jit.on()
+-- XXX: Reset hotcounters.
+jit.opt.start('hotloop=1')
+jparse.start('i')
+
+for _ = 1, 4 do
+  anchor = table_new(0, 0)
+end
+
+traces = jparse.finish()
+jit.off()
+
+test:ok(type(anchor) == 'table', 'base result')
+test:ok(traces[1]:has_ir('TNEW.*#0.*#0'), '0 asize, 0 hsize')
+
+jit.flush()
+jit.on()
+-- XXX: Reset hotcounters.
+jit.opt.start('hotloop=1')
+jparse.start('i')
+
+for _ = 1, 4 do
+  anchor = table_new(0, 3)
+end
+
+traces = jparse.finish()
+jit.off()
+
+test:ok(type(anchor) == 'table', 'base result')
+test:ok(traces[1]:has_ir(('TNEW.*#0.*#%d'):format(HBITS[3])),
+        '3 hsize -> 2 hbits')
+
+jit.flush()
+jit.on()
+-- XXX: Reset hotcounters.
+jit.opt.start('hotloop=1')
+jparse.start('i')
+
+for _ = 1, 4 do
+  anchor = table_new(-1, 0)
+end
+
+traces = jparse.finish()
+jit.off()
+
+test:ok(type(anchor) == 'table', 'base result')
+test:ok(traces[1]:has_ir('TNEW.*#0.*#0'), 'negative asize')
+
+jit.flush()
+jit.on()
+-- XXX: Reset hotcounters.
+jit.opt.start('hotloop=1')
+jparse.start('i')
+
+for _ = 1, 4 do
+  anchor = table_new(0xffff, 0)
+end
+
+traces = jparse.finish()
+jit.off()
+
+test:ok(type(anchor) == 'table', 'base result')
+-- Test that TNEW isn't emitted for `asize` bigger than the IR
+-- operand width (>0x8000).
+test:ok(not traces[1]:has_ir('TNEW'), 'asize out of range')
+
+test:done(true)
-- 
2.43.0