i#3544 RV64: Rebase the dcontext pointer. #7235
Conversation
The dcontext_t struct is larger than the biggest valid displacement of the load and store instructions. By rebasing the pointer by 0x800, we van access the entire struct, because the dispacement can be in the renge of -0x800 to 0x7ff.
Also worked around build failure on windows
|
I know that these patches are quite "hacky". However, I could not find a better way to work around some issues. |
FYI I am out of office today; will take a look tomorrow. |
The failure is the sourceware.org network issue. Several flaky issues were fixed in the last several days: probably worth merging the latest from master. |
| @@ -698,12 +698,13 @@ append_restore_simd_reg(dcontext_t *dcontext, instrlist_t *ilist, bool absolute) | |||
| * | |||
| * ma ta sew=8 lmul=8 */ | |||
| vtypei = (0b1 << 7) | (0b1 << 6) | (0b000 << 3) | 0b011; | |||
| memopnd = opnd_create_dcontext_field_via_reg_sz( | |||
| dcontext, DR_REG_A1, 0, reg_get_size_lmul(DR_REG_VR0, RV64_LMUL_8)); | |||
| memopnd = opnd_create_base_disp(DR_REG_A1, REG_NULL, 0, 0, | |||
There was a problem hiding this comment.
Is this temporary and you want to go back to the opnd_create_dc* routine? Added a TODO i#3544: ...? Ditto below.
There was a problem hiding this comment.
To me, it looks like opnd_create_dcontext_field_via_reg_sz() should not be used for this specific operand. We do not want to access a dcontext field by base+offset. The memory address where the the respective memory operation is expected to happen is already in register a1. This worked before, but now it tries to access a1-0x800 instead of a1. So I believe, it should not be temporary.
EDIT: s/respected/respective/
There was a problem hiding this comment.
Since this the only remaining unresolved conversation, we would like some clarification on that. I have stated my opinion that opnd_create_base_disp() is more appropriate in these cases. What are we going to do with it?
Renamed some macros, added helper macros to replace raw arithmetic operations, Added missing DCONTEXT_ACTUAL_TO_TLS_OFFSET.
For RISC-V the dcontext_t struct is larger than the biggest valid displacement of the load and store instructions.
By rebasing the pointer kept in spill_state_t's TLS slot by 0x800, we can access the entire struct, because the displacement can be in the range of -0x800 to 0x7ff. Architectures other than RV64, are unaffected by these changes.