Name: cl74495 Date: 01/04/2001


Using the Sparc Hotspot Server sources from
  hotspot_server-1_3_1-pre-beta-src-b12-solsparc-14_dec_2000.zip

I am observing problems with the generated code for the
Bytecodes::_laload bytecode , for compiled methods.

The source of the problem appears to be in form3_mem_reg_long in sparc.ad ,
which is always emitting the simm13 form of an ldd instruction,
instead of also handling the ldd [r1+r2],r3 form .

Below is some debugging output, followed by the suggested changes
to sparc.ad in the "workaround" section of the bug report .
 I think the Op_StartI2C case may also need further fixes.

For a Java method like this:

private static void readObj3(long[] arr, int delta)
{
  long val = arr[1000];
  long expected = 1300 + delta;
  if (val != expected))
    System.out.println("wrong value , expected" + expected + "got" + val );
}

I observe a bad ldd instruction for the arr[1000] access , like this:

0xf9417818: ld [%i0 + 0xc], %l1 # load array length
0xf941781c: sethi %hi(0x1c00), %l0
0xf9417820: add %l0, 0x350, %l0 # 0x1c00 + 0x350 = 0x1f50 = 8016
0xf9417824: cmp %l1, 0x3e8 # 0x3e8 == 1000
0xf9417828: bleu,pn %icc,0xf94178a0
0xf941782c: nop
0xf9417830: add %i1, 0x514, %o3 # 0x514 == 1300 ; %i1 == delta
0xf9417834: ldd [%i0], %l0 # BUG, SHOULD BE: ldd [%i0 + %l0], %l0


Part of praph observed in parse2.cpp for Bytecodes::_laload, LoadLNode :
(dbx) p n->dump(3)
 55 If <== 43 54 P=0.000001, C=-1.000000 <54725> Type:{0:control, 1:control}
 66 IfFalse <== 55 #0 <54736> Type:control
 4 Start <== 4 0 #{0:control, 1:abIO, 2:memory, 3:rawptr:BotPTR, 4:return_address, 5:long[int+]:BotPTR*, 6:int} <54674> Type:{0:control, 1:abIO, 2:memory, 3:rawptr:BotPTR, 4:return_address, 5:long[int+]:BotPTR*, 6:int}
 8 Parm <== 4 Memory <54678> Type:memory
 44 CastPP <== 43 12 <54714> Type:long[int+]:NotNull*
 73 ConI <== 0 <54743> Type:int:8016
 72 AddP <== _ 44 44 73 <54742> Type:long[int+]:NotNull*[2001]
 74 LoadL <== 66 8 72 <54744> Type:long


Output from PrintOptoAssembly:

000 B1: # B8 B2 <- BLOCK HEAD IS JUNK Freq: 0.000666667 Count: 0.000666667 IDom: 0/#2 RegPressure: 2 IHRP Index: 10 FRegPressure: 0 FHRP Index: 10
 AntiDep: 0 Blocker: 0
000 --- MachPrologNode ----
00c CMP R_I0,NULL ! ptr
010 BPeq B8 # 0x0b0 P=0.000001 C=-1.000000
018
018 B2: # B7 B3 <- B1 Freq: 0.000666666 Count: -1 IDom: 1/#3 RegPressure: 4 IHRP Index: 8 FRegPressure: 0 FHRP Index: 8
 AntiDep: 0 Blocker: 0
018 + LDUW [R_I0 + #12],R_L1 !range
01c SET #8016,R_L0
024 + CMP R_L1,#1000 ! unsigned
028 BPule B7 # 0x0a0 P=0.000001 C=-1.000000
030
030 B3: # B5 B4 <- B2 Freq: 0.000666665 Count: -1 IDom: 2/#4 RegPressure: 7 IHRP Index: 8 FRegPressure: 0 FHRP Index: 8
 AntiDep: 0 Blocker: 0
030 + ADD R_I1,#1300,R_O3
034 LDD [R_I0 + R_L0],R_L0.lo ! long <<<<<< NOTE the LoadL instruction form


Output from PrintCompilation, to get method start address
 38 b testasm::readObj3 (32 bytes)
Compiled {method} methodOop=0xf7161b00 'readObj3' '([JI)V'
          in 'testasm'^M
 main code [0xf9417800,0xf94178e0] = 224 # start address = 0xf9417800

(dbx) px 0xf9417740 + 0x34
4181817152U+52 = 0xf9417774 # address of the LoadL instruction

(dbx) x 0xf9417774 / 1 i
0xf9417774: ldd [%i0], %l0 # bad value emitted by sparc.ad
(dbx) x 0xf9417774 / 1 X
0xf9417774: 0xe01e2000 # bad value as hex

(dbx) assign *(int*)0xf9417774 = 0xe01e0010 # encoding the correct instruction by hand
(dbx) x 0xf9417774 / 1 i
0xf9417774: ldd [%i0 + %l0], %l0
(Review ID: 114564)
======================================================================