Changeset 279343 in webkit for trunk/Source/JavaScriptCore/offlineasm/arm.rb

Timestamp:

Jun 28, 2021, 10:55:12 AM (4 years ago)

Author:

[email protected]

Message:

Add LLInt fast path for less, lesseq, greater and greatereq
​https://bugs.webkit.org/show_bug.cgi?id=226266

Patch by Mikhail R. Gadelha <Mikhail R. Gadelha> on 2021-06-28
Reviewed by Tadeu Zagallo.

The motivation is to add fast path for integers and doubles
in LLInt, so we don't need to go to slow path for those cases.

This patch implements the less, lesseq, greater, greatereq
instruction for ARMv7, MIPS and CLoop.

Microbenchmarking results:

• x86_64: number-comparison-inline definitely 1.3520x faster
• ARMv7: number-comparison-inline definitely 1.3520x faster

JetStream2 results:

• x86_64 jit: 1.015 times better
• x86_64 no-jit: 1.018 times better
• ARMv7 no-jit: 1.004 times worse

• llint/LowLevelInterpreter.asm:
• llint/LowLevelInterpreter32_64.asm:
• llint/LowLevelInterpreter64.asm:
• offlineasm/arm.rb:
• offlineasm/cloop.rb:
• offlineasm/mips.rb:

File:

• trunk/Source/JavaScriptCore/offlineasm/arm.rb (modified) (2 diffs)

trunk/Source/JavaScriptCore/offlineasm/arm.rb

@@ -373 +373 @@
         $asm.puts "tst #{value.armOperand}, #{mask.armOperand}"
     end
+end
+
+def emitArmDoubleCompare(operands, code)
+    $asm.puts "mov #{operands[2].armOperand}, \#0"
+    $asm.puts "vcmpe.f64 #{armOperands(operands[0..1])}"
+    $asm.puts "vmrs APSR_nzcv, FPSCR"
+    $asm.puts "it #{code}"
+    $asm.puts "mov#{code} #{operands[2].armOperand}, \#1"
 end
 
@@ -642 +650 @@
         when "cilteq", "cplteq", "cblteq"
             emitArmCompare(operands, "le")
+        when "cdgt"
+            emitArmDoubleCompare(operands, "gt")
+        when "cdgteq"
+            emitArmDoubleCompare(operands, "ge")
+        when "cdlt"
+            emitArmDoubleCompare(operands, "mi")
+        when "cdlteq"
+            emitArmDoubleCompare(operands, "ls")
         when "tis", "tbs", "tps"
             emitArmTestSet(operands, "mi")