Optimizes an assembly-language conversion from int to Smi in codegen.

src/codegen-ia32.cc

Index: src/codegen-ia32.cc
===================================================================
--- src/codegen-ia32.cc	(revision 1135)
+++ src/codegen-ia32.cc	(working copy)
@@ -1127,14 +1127,14 @@
         __ j(not_zero, deferred->enter(), not_taken);
         __ sar(ebx, kSmiTagSize);
         __ shl(ebx, shift_value);

[Kevin Millikin (Chromium), 2009/01/23 18:48:08]

It's late in the day so I could be wrong.

Do we really need to shift right kSmiTagSize, shift left shift_value, and then
shift left kSmiTagSize to reapply the tag?  Analogous to a multiply, can't we
just shift the tagged value if the tag is zero?

-        // This is the Smi check for the shifted result.
-        // After signed subtraction of 0xc0000000, the valid 
-        // Smis are positive.
-        __ cmp(ebx, 0xc0000000);
-        __ j(sign, deferred->enter(), not_taken); 
-        // tag result and store it in TOS (eax)
-        ASSERT(kSmiTagSize == times_2);  // adjust code if not the case
-        __ lea(eax, Operand(ebx, ebx, times_1, kSmiTag));
+        // Convert the int to a Smi, and check that it is in
+        // the range of valid Smis.
+        ASSERT(kSmiTagSize == times_2);  // Adjust code if not true.

[iposva, 2009/01/23 16:58:38]

kSmiTagSize is a number of bits, times_2 is a instruction operand encoding
field. This assert is comparing apples and pinecones. I will fix this and I also
realize that there are many other places of this misuse.

+        ASSERT(kSmiTag == 0);  // Adjust code if not true.
+        __ add(ebx, Operand(ebx));
+        __ j(overflow, deferred->enter(), not_taken);
+        __ mov(eax, Operand(ebx));
+
         __ bind(deferred->exit());
         frame_->Push(eax);
       }