RFR: 8298935: fix cyclic dependency bug in create_pack logic in SuperWord::find_adjacent_refs [v15]
Emanuel Peter
epeter at openjdk.org
Mon Mar 6 09:05:18 UTC 2023
On Mon, 6 Mar 2023 05:19:37 GMT, Jatin Bhateja <jbhateja at openjdk.org> wrote:
>> Thanks, even though newly added test now passes at all AVX and SSE level can you kindly investigate why should following be vectorized with un-aligned accesses when it carries a cross iteration true dependency with distance 4.
>>
>>
>> @Test
>> // CPU: sse4.1 to avx -> vector_width: 16 -> elements in vector: 4
>> // positive byte_offset 12 can lead to cyclic dependency
>> @IR(counts = {IRNode.LOAD_VECTOR, "> 0", IRNode.MUL_V, "> 0", IRNode.STORE_VECTOR, "> 0"},
>> applyIfAnd = {"AlignVector", "false", "MaxVectorSize", ">= 8", "MaxVectorSize", "<= 12"},
>> applyIfCPUFeatureAnd = {"sse4.1", "true", "avx2", "false"})
>> // CPU: avx2 -> vector_width: 32 -> elements in vector: 8
>> // positive byte_offset 12 can lead to cyclic dependency
>> @IR(counts = {IRNode.LOAD_VECTOR, "> 0", IRNode.MUL_V, "> 0", IRNode.STORE_VECTOR, "> 0"},
>> applyIfAnd = {"AlignVector", "false", "MaxVectorSize", ">= 8", "MaxVectorSize", "<= 12"},
>> applyIfCPUFeatureAnd = {"avx2", "true", "avx512", "false"})
>> // CPU: avx512 -> vector_width: 64 -> elements in vector: 16
>> // positive byte_offset 12 can lead to cyclic dependency
>> @IR(counts = {IRNode.LOAD_VECTOR, "> 0", IRNode.MUL_V, "> 0", IRNode.STORE_VECTOR, "> 0"},
>> applyIfAnd = {"AlignVector", "false", "MaxVectorSize", ">= 8", "MaxVectorSize", "<= 12"},
>> applyIfCPUFeature = {"avx512", "true"})
>> // CPU: asimd -> vector_width: 32 -> elements in vector: 8
>> // positive byte_offset 12 can lead to cyclic dependency
>> @IR(counts = {IRNode.LOAD_VECTOR, "> 0", IRNode.MUL_V, "> 0", IRNode.STORE_VECTOR, "> 0"},
>> applyIfAnd = {"AlignVector", "false", "MaxVectorSize", ">= 8", "MaxVectorSize", "<= 12"},
>> applyIfCPUFeature = {"asimd", "true"})
>> public static void testIntP3(int[] data) {
>> for (int j = 0; j < RANGE - 3; j++) {
>> data[j + 3] = (int)(data[j] * (int)-11);
>> }
>> }
>>
>>
>> Also SLP now operates under SuperWordMaxVectorSize so it will be good to its it instead.
>
> With +AlignVector behavior with and without Vectorize,true pragma should match.
>
>
> static void test1() {
> for (int i = 4; i < 100; i++) {
> fArr[i + 4] = fArr[i];
> }
> }
>
>
>
> CPROMPT>javad -XX:+TraceNewVectors -XX:+AlignVector -cp . bug
> WARNING: Using incubator modules: jdk.incubator.vector
> res = 0.0
> CPROMPT>
> CPROMPT>javad -XX:+TraceNewVectors -XX:+AlignVector -XX:CompileCommand=Vectorize,bug::test1,true -cp . bug
> CompileCommand: Vectorize bug.test1 bool Vectorize = true
> WARNING: Using incubator modules: jdk.incubator.vector
> new Vector node: 990 LoadVector === 373 856 824 [[ 822 802 800 798 718 706 556 196 ]] @float[int:>=0]:NotNull:exact+any *, idx=6; mismatched #vectory[8]:{float} !orig=[823],[719],[557],[199],143 !jvms: bug::test1 @ bci:18 (line 7)
> new Vector node: 991 StoreVector === 855 856 825 990 [[ 988 195 856 ]] @float[int:>=0]:NotNull:exact+any *, idx=6; mismatched Memory: @float[int:>=0]:NotNull:exact+any *, idx=6; !orig=[822],[718],[556],[196],164 !jvms: bug::test1 @ bci:19 (line 7)
> res = 0.0
@jatin-bhateja
> Thanks, even though newly added test now passes at all AVX and SSE level can you kindly investigate why should following be vectorized with un-aligned accesses when it carries a cross iteration true dependency with distance 4.
The cyclic dependency is at a distance of 3, not 4 in this example. Ints are 4 bytes. Thus, the `byte_offset` is 12 bytes. So if `MaxVectorSize <= 12`, we cannot ever have a cyclic dependency within a vector.
See my explanations at the beginning of the test file, for example:
https://github.com/openjdk/jdk/blob/fb7f6dd9fbc2a6086d2ad36e0681fbc9eff6c9a7/test/hotspot/jtreg/compiler/loopopts/superword/TestDependencyOffsets.java#L49-L57
-------------
PR: https://git.openjdk.org/jdk/pull/12350
More information about the hotspot-compiler-dev
mailing list