April 2014 - aarch64-port-dev

RFR: Minor optimisation for divide by 2
by Edward Nevill 12 May '14

12 May '14

Hi, C2 currently generates mov rdst, rsrc, asr #31 mov rdst, rdst, lsr #31 add rdst, rsrc, rdst mov rdst, rdst, asr #1 for divide by 2. The following patch reduces this to add rdst, rsrc, rsrc, lsr #31 mov rdst, rdst, asr #1 I know this is very minor, but it offends me:-) OK? Ed. --- CUT HERE --- # HG changeset patch # User Edward Nevill edward.nevill(a)linaro.org # Date 1398762402 -3600 # Tue Apr 29 10:06:42 2014 +0100 # Node ID 7f9ab7b86d7a690e04ffc6331c2b9519aae2a565 # Parent d9468835bc5160b7fac6709b0afbc751b2159fbb Minor optimisation for divide by 2 diff -r d9468835bc51 -r 7f9ab7b86d7a src/cpu/aarch64/vm/aarch64.ad --- a/src/cpu/aarch64/vm/aarch64.ad Thu Apr 10 06:50:43 2014 -0400 +++ b/src/cpu/aarch64/vm/aarch64.ad Tue Apr 29 10:06:42 2014 +0100 @@ -3356,6 +3356,16 @@ interface(CONST_INTER); %} +operand immI_31() +%{ + predicate(n->get_int() == 31); + match(ConI); + + op_cost(0); + format %{ %} + interface(CONST_INTER); +%} + operand immI_8() %{ predicate(n->get_int() == 8); @@ -7274,6 +7284,30 @@ ins_pipe(pipe_class_default); %} +instruct signExtract(iRegINoSp dst, iRegI src, immI_31 div1, immI_31 div2) %{ + match(Set dst (URShiftI (RShiftI src div1) div2)); + ins_cost(INSN_COST); + format %{ "lsrw $dst, $src, $div1" %} + ins_encode %{ + __ lsrw(as_Register($dst$$reg), as_Register($src$$reg), 31); + %} + ins_pipe(pipe_class_default); +%} + +instruct div2Round(iRegINoSp dst, iRegI src, immI_31 div1, immI_31 div2) %{ + match(Set dst (AddI src (URShiftI (RShiftI src div1) div2))); + ins_cost(INSN_COST); + format %{ "addw $dst, $src, $div1" %} + + ins_encode %{ + __ addw(as_Register($dst$$reg), + as_Register($src$$reg), + as_Register($src$$reg), + Assembler::LSR, 31); + %} + ins_pipe(pipe_class_default); +%} + // Long Divide instruct divL(iRegLNoSp dst, iRegL src1, iRegL src2) %{ --- CUT HERE ---

3 3

client JTREG results for OpenJDK 8 on AArch64
by openjdk-testing＠linaro.org 30 Apr '14

30 Apr '14

This is a summary of the JTREG test results for OpenJDK 8 on AArch64. The build and test results are cycled on a weekly basis. For detailed information on the test output please refer to: http://openjdk.linaro.org/openjdk8-jtreg-nightly-tests/summary/2014/120/sum… =============================================================================== client-fastdebug/hotspot =============================================================================== Build 0: aarch64/2014/apr/15 pass: 431; fail: 5; error: 2 Build 1: aarch64/2014/apr/16 pass: 429; fail: 5; error: 4 Build 2: aarch64/2014/apr/17 pass: 429; fail: 5; error: 4 Build 3: aarch64/2014/apr/24 pass: 418; fail: 5; error: 15 Build 4: aarch64/2014/apr/25 pass: 421; fail: 5; error: 12 Build 5: aarch64/2014/apr/26 pass: 421; fail: 5; error: 12 Build 6: aarch64/2014/apr/30 pass: 422; fail: 5; error: 11 ------------------------------------------------------------------------------- =============================================================================== client-fastdebug/langtools =============================================================================== Build 0: aarch64/2014/apr/15 pass: 2,939; error: 33 Build 1: aarch64/2014/apr/16 pass: 2,936; error: 36 Build 2: aarch64/2014/apr/17 pass: 2,937; error: 35 Build 3: aarch64/2014/apr/24 pass: 2,916; error: 56 Build 4: aarch64/2014/apr/25 pass: 2,917; error: 55 Build 5: aarch64/2014/apr/26 pass: 2,917; error: 55 Build 6: aarch64/2014/apr/30 pass: 2,917; error: 55 ------------------------------------------------------------------------------- =============================================================================== client-release/jdk =============================================================================== Build 0: aarch64/2014/apr/15 pass: 5,236; fail: 155; error: 59 Build 1: aarch64/2014/apr/16 pass: 5,231; fail: 156; error: 63 Build 2: aarch64/2014/apr/17 pass: 5,231; fail: 157; error: 62 Build 3: aarch64/2014/apr/24 pass: 4,886; fail: 474; error: 90 Build 4: aarch64/2014/apr/25 pass: 4,905; fail: 474; error: 71 Build 5: aarch64/2014/apr/26 pass: 4,905; fail: 472; error: 72 Build 6: aarch64/2014/apr/30 pass: 4,904; fail: 473; error: 72 ------------------------------------------------------------------------------- Previous results can be found here: http://openjdk.linaro.org/openjdk8-jtreg-nightly-tests/index.html

1 0

server JTREG results for OpenJDK 8 on AArch64
by openjdk-testing＠linaro.org 30 Apr '14

30 Apr '14

This is a summary of the JTREG test results for OpenJDK 8 on AArch64. The build and test results are cycled on a weekly basis. For detailed information on the test output please refer to: http://openjdk.linaro.org/openjdk8-jtreg-nightly-tests/summary/2014/120/sum… =============================================================================== server-fastdebug/hotspot =============================================================================== Build 0: aarch64/2014/apr/15 pass: 432; fail: 2; error: 4 Build 1: aarch64/2014/apr/16 pass: 433; fail: 1; error: 4 Build 2: aarch64/2014/apr/17 pass: 433; fail: 1; error: 4 Build 3: aarch64/2014/apr/24 pass: 416; fail: 1; error: 21 Build 4: aarch64/2014/apr/25 pass: 425; fail: 1; error: 12 Build 5: aarch64/2014/apr/26 pass: 423; fail: 1; error: 14 Build 6: aarch64/2014/apr/30 pass: 421; fail: 2; error: 15 ------------------------------------------------------------------------------- =============================================================================== server-fastdebug/langtools =============================================================================== Build 0: aarch64/2014/apr/15 pass: 2,936; error: 36 Build 1: aarch64/2014/apr/16 pass: 2,934; error: 38 Build 2: aarch64/2014/apr/17 pass: 2,933; error: 39 Build 3: aarch64/2014/apr/24 pass: 2,894; error: 78 Build 4: aarch64/2014/apr/25 pass: 2,911; error: 61 Build 5: aarch64/2014/apr/26 pass: 2,912; error: 60 Build 6: aarch64/2014/apr/30 pass: 2,915; error: 57 ------------------------------------------------------------------------------- =============================================================================== server-release/jdk =============================================================================== Build 0: aarch64/2014/apr/15 pass: 5,230; fail: 149; error: 71 Build 1: aarch64/2014/apr/16 pass: 5,242; fail: 134; error: 74 Build 2: aarch64/2014/apr/17 pass: 5,230; fail: 146; error: 74 Build 3: aarch64/2014/apr/24 pass: 4,699; fail: 472; error: 279 Build 4: aarch64/2014/apr/25 pass: 4,725; fail: 472; error: 253 Build 5: aarch64/2014/apr/26 pass: 4,723; fail: 475; error: 251 Build 6: aarch64/2014/apr/30 pass: 4,725; fail: 470; error: 254 ------------------------------------------------------------------------------- Previous results can be found here: http://openjdk.linaro.org/openjdk8-jtreg-nightly-tests/index.html

1 0

JDK9
by Andrew Haley 29 Apr '14

29 Apr '14

I'm building a JDK 9 tree as the first stage of upstreaming. I'll write a JEP proposal for this project next. Andrew.

2 2

hg: aarch64-port/jdk8/hotspot: Minor optimisation for divide by 2
by ed＠camswl.com 29 Apr '14

29 Apr '14

Changeset: 9d641fdeea4d Author: Edward Nevill edward.nevill(a)linaro.org Date: 2014-04-29 14:58 +0100 URL: http://hg.openjdk.java.net/aarch64-port/jdk8/hotspot/rev/9d641fdeea4d Minor optimisation for divide by 2 ! src/cpu/aarch64/vm/aarch64.ad

1 0

simulator: multiplication/divsion for LONG/INT_MIN and -1
by D.Sturm 27 Apr '14

27 Apr '14

The simulator does not correctly handle multiplications/divisions in the case where we have the smallest negative value and -1. This is undefined in C/C++ anyhow, but works correctly for me in the case of multiplication (INT_MIN * -1 == INT_MIN), and is actually guaranteed to work for the int division due to the way the code is written but causes an arithmetic exception for LONG_MIN / -1 under Linux Mint 16 gcc 4.8.1. That exception is caught by the signal handler in the JVM and then correctly handled there in most situations - I ran into some throwing assertion when it was simulating graal code but haven't looked further into it. The following patch should take care of this. The smulh instruction uses inline assembly but I think x86 does the right thing(TM) there. -- Daniel # HG changeset patch # User Daniel Sturm <d.sturm42(a)gmail.com> # Date 1398525661 -7200 # Sat Apr 26 17:21:01 2014 +0200 # Node ID 31460830078ecdb590d602ccfa12f93c8e022ba9 # Parent d70a2050eb423046283bc8f1280fd9d9e1fabc21 fixed signed multiply and division in case of MIN_VAL and -1 as operands diff -r d70a2050eb42 -r 31460830078e simulator.cpp --- a/simulator.cpp Fri Jan 03 13:42:36 2014 +0000 +++ b/simulator.cpp Sat Apr 26 17:21:01 2014 +0200 @@ -44,6 +44,7 @@ #include <syscall.h> #include <signal.h> #include <assert.h> +#include <limits> #define DEBUG // #include <shared_ptr.h> @@ -7073,6 +7074,16 @@ // multiply +template <class T> +static T mul(T a, T b) { + if (a == -1 || b == -1) { + if (a == std::numeric_limits<T>::min() || b == std::numeric_limits<T>::min()) { + return std::numeric_limits<T>::min(); + } + } + return a * b; +} + // 32 bit multiply and add void AArch64Simulator::madd32() { @@ -7080,7 +7091,7 @@ // instr[14,10] = ra : may not be SP // instr[9,5] = rn : may not be SP // instr[4,0] = rd : may not be SP - xreg(0, NO_SP) = wreg(10, NO_SP) + wreg(5, NO_SP) * wreg(16, NO_SP); + xreg(0, NO_SP) = wreg(10, NO_SP) + mul(wreg(5, NO_SP), wreg(16, NO_SP)); } // 64 bit multiply and add @@ -7090,7 +7101,7 @@ // instr[14,10] = ra : may not be SP // instr[9,5] = rn : may not be SP // instr[4,0] = rd : may not be SP - xreg(0, NO_SP) = xreg(10, NO_SP) + xreg(5, NO_SP) * xreg(16, NO_SP); + xreg(0, NO_SP) = xreg(10, NO_SP) + mul(xreg(5, NO_SP), xreg(16, NO_SP)); } // 32 bit multiply and sub @@ -7100,7 +7111,7 @@ // instr[14,10] = ra : may not be SP // instr[9,5] = rn : may not be SP // instr[4,0] = rd : may not be SP - xreg(0, NO_SP) = wreg(10, NO_SP) - wreg(5, NO_SP) * wreg(16, NO_SP); + xreg(0, NO_SP) = wreg(10, NO_SP) - mul(wreg(5, NO_SP), wreg(16, NO_SP)); } // 64 bit multiply and sub @@ -7110,7 +7121,7 @@ // instr[14,10] = ra : may not be SP // instr[9,5] = rn : may not be SP // instr[4,0] = rd : may not be SP - xreg(0, NO_SP) = xreg(10, NO_SP) - xreg(5, NO_SP) * xreg(16, NO_SP); + xreg(0, NO_SP) = xreg(10, NO_SP) - mul(xreg(5, NO_SP), xreg(16, NO_SP)); } // signed multiply add long -- source, source2 : 32 bit, source3 : 64 bit @@ -7122,7 +7133,7 @@ // instr[4,0] = rd : may not be SP // n.b. we need to multiply the signed 32 bit values in rn, rm to // obtain a 64 bit product - xregs(0, NO_SP) = xregs(10, NO_SP) + ((int64_t)wregs(5, NO_SP)) * ((int64_t)wregs(16, NO_SP)); + xregs(0, NO_SP) = xregs(10, NO_SP) + mul((int64_t)wregs(5, NO_SP), (int64_t)wregs(16, NO_SP)); } // signed multiply sub long -- source, source2 : 32 bit, source3 : 64 bit @@ -7134,7 +7145,7 @@ // instr[4,0] = rd : may not be SP // n.b. we need to multiply the signed 32 bit values in rn, rm to // obtain a 64 bit product - xregs(0, NO_SP) = xregs(10, NO_SP) - ((int64_t)wregs(5, NO_SP)) * ((int64_t)wregs(16, NO_SP)); + xregs(0, NO_SP) = xregs(10, NO_SP) - mul((int64_t)wregs(5, NO_SP), (int64_t)wregs(16, NO_SP)); } // signed multiply high, source, source2 : 64 bit, dest <-- high 64-bit of result @@ -7225,8 +7236,13 @@ // instr[9,5] = rn : may not be SP // instr[4,0] = rd : may not be SP // TODO : check that this rounds towards zero as required + int64_t dividend = xregs(5, NO_SP); int64_t divisor = xregs(16, NO_SP); - xregs(0, NO_SP) = (divisor ? (xregs(5, NO_SP) / divisor) : 0); + if (divisor == -1 && dividend == std::numeric_limits<int64_t>::min()) { + xregs(0, NO_SP) = std::numeric_limits<int64_t>::min(); + } else { + xregs(0, NO_SP) = (divisor ? (dividend / divisor) : 0); + } } // 32 bit unsigned divide

2 2

frintz instruction in simulator
by D.Sturm 26 Apr '14

26 Apr '14

Hi, Since I need a frintz instruction when generating the floating point remainder operation* I had to implement it in the simulator. The implementation should be correct except in case of a signaling NaN since that should throw an exception (could make a difference if someone generates a signaling NaN manually from the integer pattern, but that seems like an extreme edge case). Maybe useful for the general branch too. --Daniel

2 2

hg: aarch64-port/jdk8: Add test to JTreg exclude which also fails on x86
by ed＠camswl.com 25 Apr '14

25 Apr '14

Changeset: e78570806ec8 Author: Edward Nevill edward.nevill(a)linaro.org Date: 2014-04-25 14:28 +0100 URL: http://hg.openjdk.java.net/aarch64-port/jdk8/rev/e78570806ec8 Add test to JTreg exclude which also fails on x86 ! test/exclude_aarch64.txt

1 0

Add test to JTreg exclude which also fails on x86
by Edward Nevill 25 Apr '14

25 Apr '14

Hi, The following test fails on x86 also. I have filed a bug report against it on x86. Patch to exclude from JTreg tests. Regards, Ed. --- CUT HERE --- # HG changeset patch # User Edward Nevill edward.nevill(a)linaro.org # Date 1398432523 -3600 # Fri Apr 25 14:28:43 2014 +0100 # Node ID e78570806ec82432313f6543bb7601cc3197aae3 # Parent e03c59f25d2b8d12d3a4906d72887bfe14494c85 Add test to JTreg exclude which also fails on x86 diff -r e03c59f25d2b -r e78570806ec8 test/exclude_aarch64.txt --- a/test/exclude_aarch64.txt Wed Mar 26 17:19:44 2014 +0000 +++ b/test/exclude_aarch64.txt Fri Apr 25 14:28:43 2014 +0100 @@ -15,3 +15,4 @@ runtime/SharedArchiveFile/CdsSameObjectAlignment.java generic-all runtime/SharedArchiveFile/CdsDifferentObjectAlignment.java generic-all compiler/whitebox/IsMethodCompilableTest.java generic-all +sun/java2d/OpenGL/DrawBufImgOp.java generic-all --- CUT HERE ---

1 0

client JTREG results for OpenJDK 8 on AArch64
by openjdk-testing＠linaro.org 25 Apr '14

25 Apr '14

This is a summary of the JTREG test results for OpenJDK 8 on AArch64. The build and test results are cycled on a weekly basis. For detailed information on the test output please refer to: http://openjdk.linaro.org/openjdk8-jtreg-nightly-tests/summary/2014/115/sum… =============================================================================== client-fastdebug/hotspot =============================================================================== Build 0: aarch64/2014/apr/13 pass: 431; fail: 5; error: 2 Build 1: aarch64/2014/apr/14 pass: 431; fail: 5; error: 2 Build 2: aarch64/2014/apr/15 pass: 431; fail: 5; error: 2 Build 3: aarch64/2014/apr/16 pass: 429; fail: 5; error: 4 Build 4: aarch64/2014/apr/17 pass: 429; fail: 5; error: 4 Build 5: aarch64/2014/apr/24 pass: 418; fail: 5; error: 15 Build 6: aarch64/2014/apr/25 pass: 421; fail: 5; error: 12 ------------------------------------------------------------------------------- =============================================================================== client-fastdebug/langtools =============================================================================== Build 0: aarch64/2014/apr/13 pass: 2,937; error: 35 Build 1: aarch64/2014/apr/14 pass: 2,938; error: 34 Build 2: aarch64/2014/apr/15 pass: 2,939; error: 33 Build 3: aarch64/2014/apr/16 pass: 2,936; error: 36 Build 4: aarch64/2014/apr/17 pass: 2,937; error: 35 Build 5: aarch64/2014/apr/24 pass: 2,916; error: 56 Build 6: aarch64/2014/apr/25 pass: 2,917; error: 55 ------------------------------------------------------------------------------- =============================================================================== client-release/jdk =============================================================================== Build 0: aarch64/2014/apr/13 pass: 5,238; fail: 152; error: 60 Build 1: aarch64/2014/apr/14 pass: 5,236; fail: 157; error: 57 Build 2: aarch64/2014/apr/15 pass: 5,236; fail: 155; error: 59 Build 3: aarch64/2014/apr/16 pass: 5,231; fail: 156; error: 63 Build 4: aarch64/2014/apr/17 pass: 5,231; fail: 157; error: 62 Build 5: aarch64/2014/apr/24 pass: 4,886; fail: 474; error: 90 Build 6: aarch64/2014/apr/25 pass: 4,905; fail: 474; error: 71 ------------------------------------------------------------------------------- Previous results can be found here: http://openjdk.linaro.org/openjdk8-jtreg-nightly-tests/index.html

1 0