[ESIMD] Fix an issue with copy_to that incorrectly copies char buffers in some circumstances #6298
Conversation
|
The test that is supposed to verify this fix has failed here: http://llvm-ci-test2.intel.com:8080/blue/organizations/jenkins/SYCL_CI%2Fintel%2FLLVM_Test_Suite_Associate_CI/detail/LLVM_Test_Suite_Associate_CI/175/pipeline |
| Vals.template select<4, 1>() = | ||
| Tmp.template bit_cast_view<int32_t>().template select<4, 1>( | ||
| NumChunks * ChunkSize); | ||
| block_store<int32_t, 8>(reinterpret_cast<int32_t *>(Addr) + |
There was a problem hiding this comment.
How does this work? Vals has 8 int elements where the low 4 hold some useful values, while the upper 4 keep garbage, which is block_store'd to memory.
Looks like it writes to memory that is not supposed to be written, right?
There was a problem hiding this comment.
Fixed this and accessor version, added tests to test both versions
There was a problem hiding this comment.
It uses mask to exclude data that don't need to be copied.
|
/verify with intel/llvm-test-suite#1057 |
Seems to work now |
Remove special handling for accessor version as it is correctly handled.
|
/verify with intel/llvm-test-suite#1057 |
|
/verify with intel/llvm-test-suite#1057 |
1 similar comment
|
/verify with intel/llvm-test-suite#1057 |
2 participants
This fix addresses an issue of incorrect handling of char buffers by copy_to function.
Complementary test PR intel/llvm-test-suite#1057