| buzbee | 1452bee | 2015-03-06 14:43:04 -0800 | [diff] [blame^] | 1 | /* |
| 2 | * Compare two 64-bit values. Puts 0, 1, or -1 into the destination |
| 3 | * register based on the results of the comparison. |
| 4 | * |
| 5 | * We load the full values with LDM, but in practice many values could |
| 6 | * be resolved by only looking at the high word. This could be made |
| 7 | * faster or slower by splitting the LDM into a pair of LDRs. |
| 8 | * |
| 9 | * If we just wanted to set condition flags, we could do this: |
| 10 | * subs ip, r0, r2 |
| 11 | * sbcs ip, r1, r3 |
| 12 | * subeqs ip, r0, r2 |
| 13 | * Leaving { <0, 0, >0 } in ip. However, we have to set it to a specific |
| 14 | * integer value, which we can do with 2 conditional mov/mvn instructions |
| 15 | * (set 1, set -1; if they're equal we already have 0 in ip), giving |
| 16 | * us a constant 5-cycle path plus a branch at the end to the |
| 17 | * instruction epilogue code. The multi-compare approach below needs |
| 18 | * 2 or 3 cycles + branch if the high word doesn't match, 6 + branch |
| 19 | * in the worst case (the 64-bit values are equal). |
| 20 | */ |
| 21 | /* cmp-long vAA, vBB, vCC */ |
| 22 | FETCH r0, 1 @ r0<- CCBB |
| 23 | mov r9, rINST, lsr #8 @ r9<- AA |
| 24 | and r2, r0, #255 @ r2<- BB |
| 25 | mov r3, r0, lsr #8 @ r3<- CC |
| 26 | add r2, rFP, r2, lsl #2 @ r2<- &fp[BB] |
| 27 | add r3, rFP, r3, lsl #2 @ r3<- &fp[CC] |
| 28 | ldmia r2, {r0-r1} @ r0/r1<- vBB/vBB+1 |
| 29 | ldmia r3, {r2-r3} @ r2/r3<- vCC/vCC+1 |
| 30 | cmp r1, r3 @ compare (vBB+1, vCC+1) |
| 31 | blt .L${opcode}_less @ signed compare on high part |
| 32 | bgt .L${opcode}_greater |
| 33 | subs r1, r0, r2 @ r1<- r0 - r2 |
| 34 | bhi .L${opcode}_greater @ unsigned compare on low part |
| 35 | bne .L${opcode}_less |
| 36 | b .L${opcode}_finish @ equal; r1 already holds 0 |
| 37 | %break |
| 38 | |
| 39 | .L${opcode}_less: |
| 40 | mvn r1, #0 @ r1<- -1 |
| 41 | @ Want to cond code the next mov so we can avoid branch, but don't see it; |
| 42 | @ instead, we just replicate the tail end. |
| 43 | FETCH_ADVANCE_INST 2 @ advance rPC, load rINST |
| 44 | SET_VREG r1, r9 @ vAA<- r1 |
| 45 | GET_INST_OPCODE ip @ extract opcode from rINST |
| 46 | GOTO_OPCODE ip @ jump to next instruction |
| 47 | |
| 48 | .L${opcode}_greater: |
| 49 | mov r1, #1 @ r1<- 1 |
| 50 | @ fall through to _finish |
| 51 | |
| 52 | .L${opcode}_finish: |
| 53 | FETCH_ADVANCE_INST 2 @ advance rPC, load rINST |
| 54 | SET_VREG r1, r9 @ vAA<- r1 |
| 55 | GET_INST_OPCODE ip @ extract opcode from rINST |
| 56 | GOTO_OPCODE ip @ jump to next instruction |