AArch64 opnd_create_abs_addr() succeeds but opnd_is_memory_reference() returns false #5295
Comments
|
For 32-bit ARM: from a quick glance at the encoder, it looks like it does not think abs-addr is valid and expects only rel-addr. |
|
I had thought since abs-addr was defined on a32 it was auto-converted there, and that's why we should have a64 do the same -- but if a32 doesn't really support it, an alternative is to go and make abs-addr x86-only entirely in the docs and w/ ifdefs in the headers and code. That would then require no encoder changes. Users could still use |
|
@abhinav92003 is voting for making abs-addr x86-only. Looking at 32-bit arm implementation notes: I do not see any mention of it so I assume there was no real plan for it; so making abs-addr operands x86-only seems reasonable. |
Enables building drbbdup and its tests on AArchXX. Adds case comparison, conditional branch, and TLS access code for both AArch64 and ARM. Uses optimizations for comparing small case values; a second scratch register is needed for larger values. Ports the tests by using OPND_CREATE_ABSMEM (xref #5295). To make these pc-relative case encoding loads work on AArchXX in the presence of AArch64 encoder bugs and to avoid a more-complex interface, drbbdup auto-inserts loading the target address into its scratch register to avoid reachability problems (xref #1834 on having DR do this during mangling). Avoids storing block translations if there is no dynamic handling, improving robustness and reducing memory usage. Adds the tests to the passes-under-QEMU list for cross-compilation. Tested manually on a native ARM machine. Issue: #4134, #5295, #1834, #3995
Enables building drbbdup and its tests on AArchXX. Adds case comparison, conditional branch, and TLS access code for both AArch64 and ARM. Uses optimizations for comparing small case values; a second scratch register is needed for larger values. Ports the tests by using OPND_CREATE_ABSMEM (xref #5295). To make these pc-relative case encoding loads work on AArchXX in the presence of AArch64 encoder bugs (#5316) and to avoid a more-complex interface, drbbdup auto-inserts loading the target address into its scratch register to avoid reachability problems (xref #1834 on having DR do this during mangling). Avoids storing block translations if there is no dynamic handling, improving robustness and reducing memory usage. Adds the tests to the passes-under-QEMU list for cross-compilation. Additionally, I tested manually on native ARM and AArch64 machines. Issue: #4134, #5295, #1834, #3995, #5316
For #4134 and #3995 I'm porting drbbdup to arm and its tests call
opnd_create_abs_addr()to create theruntime_case_opnd. DR lets it create an absolute addr opnd of typeABS_ADDR_kindon AArch64; but thenopnd_is_memory_referencereturns false on AArch64 because the check foropnd_is_abs_addr()is limited toIF_X86_64. Basically, DR isn't sure it wants to support absolute address operands for AArch64, since the addressing modes there do not have that sort of thing. Maybe users are expected to useopnd_create_rel_addr()instead (AArch64 does have#define OPND_CREATE_ABSMEM(addr, size) opnd_create_rel_addr(addr, size)): but the docs foropnd_create_abs_addr()say it will auto-convert into a rel addr.The rest of the code for abs addr handling in opnd_shared.c is just gated by
X64, soopnd_is_memory_referencelooks like an anomaly. So maybe the solution is to fixopnd_is_memory_referenceto return true, and ensure the AArch64 and ARM encoders handleABS_ADDR_kindand treat it just like rel-addr (the alternative of converting to rel-addr at creation time might confuse users who would query their own just-created abs-addr opnd and it would return false foropnd_is_abs_addr()).I went to check whether the AArch64 encoder might already handle abs-addr: but it looks like it doesn't even handle rel-addr! See
#4847 (comment)
The text was updated successfully, but these errors were encountered: