PPC: Add string.rep/reverse/lower/upper() fast functions.

author: Mike Pall <mike> 2010-10-06 16:17:07 +0200
committer: Mike Pall <mike> 2010-10-06 16:17:07 +0200
commit: b4f3d4525e35939865f4f0e61940a022f19b78ae (patch)
tree: e409699db9f3f9d5d08d4ad0e198cfd07ee1e612 /src
parent: 82b2f6156a0fc15029ae9c65a7751c5b4f7ff29f (diff)
download: luajit-b4f3d4525e35939865f4f0e61940a022f19b78ae.tar.gz
luajit-b4f3d4525e35939865f4f0e61940a022f19b78ae.tar.bz2
luajit-b4f3d4525e35939865f4f0e61940a022f19b78ae.zip
1 files changed, 84 insertions, 12 deletions
diff --git a/src/buildvm_ppc.dasc b/src/buildvm_ppc.dasc
index e7ca7fc4..09bb1d36 100644
--- a/src/buildvm_ppc.dasc
+++ b/src/buildvm_ppc.dasc
@@ -1431,22 +1431,94 @@ static void build_subroutines(BuildCtx *ctx)
  |.ffunc string_sub
  |  NYI
  |
-  |->fff_emptystr:  // Range underflow.
+  |.ffunc string_rep                    // Only handle the 1-char case inline.
-  |  NYI
+  |  ffgccheck
-  |
+  |  cmplwi NARGS8:RC, 16
-  |.ffunc_2 string_rep                  // Only handle the 1-char case inline.
+  |   evldd CARG1, 0(BASE)
-  |  NYI
+  |   evldd CARG2, 8(BASE)
+  |  blt ->fff_fallback
+  |  checknum CARG2
+  |  checkfail ->fff_fallback
+  |  checkstr STR:CARG1
+  |   efdctsiz CARG3, CARG2
+  |  checkfail ->fff_fallback
+  |   lwz TMP0, STR:CARG1->len
+  |  cmpwi CARG3, 0
+  |   lwz TMP1, DISPATCH_GL(tmpbuf.sz)(DISPATCH)
+  |  ble >2                             // Count <= 0? (or non-int)
+  |   cmplwi TMP0, 1
+  |  subi TMP2, CARG3, 1
+  |   blt >2                            // Zero length string?
+  |  cmplw cr1, TMP1, CARG3
+  |   bne ->fff_fallback                // Fallback for > 1-char strings.
+  |   lbz TMP0, STR:CARG1[1]
+  |   lwz CARG2, DISPATCH_GL(tmpbuf.buf)(DISPATCH)
+  |  blt cr1, ->fff_fallback
+  |1:  // Fill buffer with char. Yes, this is suboptimal code (do you care?).
+  |  cmplwi TMP2, 0
+  |   stbx TMP0, CARG2, TMP2
+  |   subi TMP2, TMP2, 1
+  |  bne <1
+  |  b ->fff_newstr
+  |2:  // Return empty string.
+  |  la STR:CRET1, DISPATCH_GL(strempty)(DISPATCH)
+  |  evmergelo CRET1, TISSTR, STR:CRET1
+  |  b ->fff_restv
  |
-  |.ffunc_1 string_reverse
+  |.ffunc string_reverse
-  |  NYI
+  |  ffgccheck
+  |  cmplwi NARGS8:RC, 8
+  |   evldd CARG1, 0(BASE)
+  |  blt ->fff_fallback
+  |  checkstr STR:CARG1
+  |   lwz TMP1, DISPATCH_GL(tmpbuf.sz)(DISPATCH)
+  |  checkfail ->fff_fallback
+  |  lwz CARG3, STR:CARG1->len
+  |   la CARG1, #STR(STR:CARG1)
+  |   lwz CARG2, DISPATCH_GL(tmpbuf.buf)(DISPATCH)
+  |   li TMP2, 0
+  |  cmplw TMP1, CARG3
+  |   subi TMP3, CARG3, 1
+  |  blt ->fff_fallback
+  |1:  // Reverse string copy.
+  |  cmpwi TMP3, 0
+  |   lbzx TMP1, CARG1, TMP2
+  |  blt ->fff_newstr
+  |   stbx TMP1, CARG2, TMP3
+  |  subi TMP3, TMP3, 1
+  |  addi TMP2, TMP2, 1
+  |  b <1
  |
-  |.macro ffstring_case, name, lo, hi
+  |.macro ffstring_case, name, lo
-  |  .ffunc_1 name
+  |  .ffunc name
-  |  NYI
+  |  ffgccheck
+  |  cmplwi NARGS8:RC, 8
+  |   evldd CARG1, 0(BASE)
+  |  blt ->fff_fallback
+  |  checkstr STR:CARG1
+  |   lwz TMP1, DISPATCH_GL(tmpbuf.sz)(DISPATCH)
+  |  checkfail ->fff_fallback
+  |  lwz CARG3, STR:CARG1->len
+  |   la CARG1, #STR(STR:CARG1)
+  |   lwz CARG2, DISPATCH_GL(tmpbuf.buf)(DISPATCH)
+  |  cmplw TMP1, CARG3
+  |   li TMP2, 0
+  |  blt ->fff_fallback
+  |1:  // ASCII case conversion.
+  |  cmplw TMP2, CARG3
+  |   lbzx TMP1, CARG1, TMP2
+  |  bge ->fff_newstr
+  |   subi TMP0, TMP1, lo
+  |    xori TMP3, TMP1, 0x20
+  |   cmplwi TMP0, 26
+  |   isellt TMP1, TMP3, TMP1
+  |   stbx TMP1, CARG2, TMP2
+  |  addi TMP2, TMP2, 1
+  |  b <1
  |.endmacro
  |
-  |ffstring_case string_lower, 0x41, 0x5a
+  |ffstring_case string_lower, 0x41
-  |ffstring_case string_upper, 0x61, 0x7a
+  |ffstring_case string_upper, 0x61
  |
  |//-- Table library ------------------------------------------------------
  |
author	Mike Pall <mike>	2010-10-06 16:17:07 +0200
committer	Mike Pall <mike>	2010-10-06 16:17:07 +0200
commit	b4f3d4525e35939865f4f0e61940a022f19b78ae (patch)
tree	e409699db9f3f9d5d08d4ad0e198cfd07ee1e612 /src
parent	82b2f6156a0fc15029ae9c65a7751c5b4f7ff29f (diff)
download	luajit-b4f3d4525e35939865f4f0e61940a022f19b78ae.tar.gz luajit-b4f3d4525e35939865f4f0e61940a022f19b78ae.tar.bz2 luajit-b4f3d4525e35939865f4f0e61940a022f19b78ae.zip

diff --git a/src/buildvm_ppc.dasc b/src/buildvm_ppc.dasc index e7ca7fc4..09bb1d36 100644 --- a/src/buildvm_ppc.dasc +++ b/src/buildvm_ppc.dasc
@@ -1431,22 +1431,94 @@ static void build_subroutines(BuildCtx *ctx)
1431	\|.ffunc string_sub	1431	\|.ffunc string_sub
1432	\| NYI	1432	\| NYI
1433	\|	1433	\|
1434	\|->fff_emptystr: // Range underflow.	1434	\|.ffunc string_rep // Only handle the 1-char case inline.
1435	\| NYI	1435	\| ffgccheck
1436	\|	1436	\| cmplwi NARGS8:RC, 16
1437	\|.ffunc_2 string_rep // Only handle the 1-char case inline.	1437	\| evldd CARG1, 0(BASE)
1438	\| NYI	1438	\| evldd CARG2, 8(BASE)
		1439	\| blt ->fff_fallback
		1440	\| checknum CARG2
		1441	\| checkfail ->fff_fallback
		1442	\| checkstr STR:CARG1
		1443	\| efdctsiz CARG3, CARG2
		1444	\| checkfail ->fff_fallback
		1445	\| lwz TMP0, STR:CARG1->len
		1446	\| cmpwi CARG3, 0
		1447	\| lwz TMP1, DISPATCH_GL(tmpbuf.sz)(DISPATCH)
		1448	\| ble >2 // Count <= 0? (or non-int)
		1449	\| cmplwi TMP0, 1
		1450	\| subi TMP2, CARG3, 1
		1451	\| blt >2 // Zero length string?
		1452	\| cmplw cr1, TMP1, CARG3
		1453	\| bne ->fff_fallback // Fallback for > 1-char strings.
		1454	\| lbz TMP0, STR:CARG1[1]
		1455	\| lwz CARG2, DISPATCH_GL(tmpbuf.buf)(DISPATCH)
		1456	\| blt cr1, ->fff_fallback
		1457	\|1: // Fill buffer with char. Yes, this is suboptimal code (do you care?).
		1458	\| cmplwi TMP2, 0
		1459	\| stbx TMP0, CARG2, TMP2
		1460	\| subi TMP2, TMP2, 1
		1461	\| bne <1
		1462	\| b ->fff_newstr
		1463	\|2: // Return empty string.
		1464	\| la STR:CRET1, DISPATCH_GL(strempty)(DISPATCH)
		1465	\| evmergelo CRET1, TISSTR, STR:CRET1
		1466	\| b ->fff_restv
1439	\|	1467	\|
1440	\|.ffunc_1 string_reverse	1468	\|.ffunc string_reverse
1441	\| NYI	1469	\| ffgccheck
		1470	\| cmplwi NARGS8:RC, 8
		1471	\| evldd CARG1, 0(BASE)
		1472	\| blt ->fff_fallback
		1473	\| checkstr STR:CARG1
		1474	\| lwz TMP1, DISPATCH_GL(tmpbuf.sz)(DISPATCH)
		1475	\| checkfail ->fff_fallback
		1476	\| lwz CARG3, STR:CARG1->len
		1477	\| la CARG1, #STR(STR:CARG1)
		1478	\| lwz CARG2, DISPATCH_GL(tmpbuf.buf)(DISPATCH)
		1479	\| li TMP2, 0
		1480	\| cmplw TMP1, CARG3
		1481	\| subi TMP3, CARG3, 1
		1482	\| blt ->fff_fallback
		1483	\|1: // Reverse string copy.
		1484	\| cmpwi TMP3, 0
		1485	\| lbzx TMP1, CARG1, TMP2
		1486	\| blt ->fff_newstr
		1487	\| stbx TMP1, CARG2, TMP3
		1488	\| subi TMP3, TMP3, 1
		1489	\| addi TMP2, TMP2, 1
		1490	\| b <1
1442	\|	1491	\|
1443	\|.macro ffstring_case, name, lo, hi	1492	\|.macro ffstring_case, name, lo
1444	\| .ffunc_1 name	1493	\| .ffunc name
1445	\| NYI	1494	\| ffgccheck
		1495	\| cmplwi NARGS8:RC, 8
		1496	\| evldd CARG1, 0(BASE)
		1497	\| blt ->fff_fallback
		1498	\| checkstr STR:CARG1
		1499	\| lwz TMP1, DISPATCH_GL(tmpbuf.sz)(DISPATCH)
		1500	\| checkfail ->fff_fallback
		1501	\| lwz CARG3, STR:CARG1->len
		1502	\| la CARG1, #STR(STR:CARG1)
		1503	\| lwz CARG2, DISPATCH_GL(tmpbuf.buf)(DISPATCH)
		1504	\| cmplw TMP1, CARG3
		1505	\| li TMP2, 0
		1506	\| blt ->fff_fallback
		1507	\|1: // ASCII case conversion.
		1508	\| cmplw TMP2, CARG3
		1509	\| lbzx TMP1, CARG1, TMP2
		1510	\| bge ->fff_newstr
		1511	\| subi TMP0, TMP1, lo
		1512	\| xori TMP3, TMP1, 0x20
		1513	\| cmplwi TMP0, 26
		1514	\| isellt TMP1, TMP3, TMP1
		1515	\| stbx TMP1, CARG2, TMP2
		1516	\| addi TMP2, TMP2, 1
		1517	\| b <1
1446	\|.endmacro	1518	\|.endmacro
1447	\|	1519	\|
1448	\|ffstring_case string_lower, 0x41, 0x5a	1520	\|ffstring_case string_lower, 0x41
1449	\|ffstring_case string_upper, 0x61, 0x7a	1521	\|ffstring_case string_upper, 0x61
1450	\|	1522	\|
1451	\|//-- Table library ------------------------------------------------------	1523	\|//-- Table library ------------------------------------------------------
1452	\|	1524	\|