i#6760 AArch64: Use smaller data types for SVE P and FFR registers #6774
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
PR #6757 fixed the way we read/write SVE register slots but unfortunately it is now broken on systems with 128-bit vector length. Both SVE vector and predicate registers use dr_simd_t slots which is a 64-byte type meant to store up to 512-bit vector registers. SVE predicate registers are always 1/8 the size of the vector register so for 512-bit vector length systems we only really need 64 / 8 = 8 bytes to store predicate registers. The ldr/str instructions we use to read and write the predicate register slots have a base+offset memory operand where the offset is a value in the range [-256, 255] scaled based by predicate register length. We read and write the registers by setting the base address to the address of the first slot, and setting the offset to n * sizeof(dr_simd_t) for each register Pn. For systems with 128-bit vector length, this means the predicate registers are 16 / 8 = 2 bytes so the maximum offset we can reach is 2 * 255 = 510 bytes. This means on 128-bit VL systems we can only reach the first 8 predicate registers (8 * sizeof(dr_simd_t) = 512). By changing the predicate register and FFR slots to use a new type dr_svep_t which is 1/8 the size of dr_simd_t we can fix this bug and save space. dr_svep_t is currently 8 bytes to correspond to 64 byte vectors, but even if we extend DynamoRIO to support the maximum SVE vector length of 2048-bits (256 bytes) dr_svep_t will only need to be increased to 256 / 8 = 32 bytes so the maximum offset (15 * 32 = 480 bytes) will always be in range. As this changes the size of the predicate register and FFR slots, this changes the size of the dr_mcontext_t structure and breaks backwards compatibility with earlier versions of DynamoRIO so the version number is increased to 10.90. Issues: #6760, #5365 Fixes: #6760
|
|
derekbruening
approved these changes
Apr 15, 2024
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
PR #6757 fixed the way we read/write SVE register slots but unfortunately it is now broken on systems with 128-bit vector length.
Both SVE vector and predicate registers use dr_simd_t slots which is a 64-byte type meant to store up to 512-bit vector registers. SVE predicate registers are always 1/8 the size of the vector register so for 512-bit vector length systems we only really need 64 / 8 = 8 bytes to store predicate registers.
The ldr/str instructions we use to read and write the predicate register slots have a base+offset memory operand where the offset is a value in the range [-256, 255] scaled based by predicate register length. We read and write the registers by setting the base address to the address of the first slot, and setting the offset to n * sizeof(dr_simd_t) for each register Pn.
For systems with 128-bit vector length, this means the predicate registers are 16 / 8 = 2 bytes so the maximum offset we can reach is 2 * 255 = 510 bytes. This means on 128-bit VL systems we can only reach the first 8 predicate registers (8 * sizeof(dr_simd_t) = 512).
By changing the predicate register and FFR slots to use a new type dr_svep_t which is 1/8 the size of dr_simd_t we can fix this bug and save space.
dr_svep_t is currently 8 bytes to correspond to 64 byte vectors, but even if we extend DynamoRIO to support the maximum SVE vector length of 2048-bits (256 bytes) dr_svep_t will only need to be increased to 256 / 8 = 32 bytes so the maximum offset (15 * 32 = 480 bytes) will always be in range.
As this changes the size of the predicate register and FFR slots, this changes the size of the dr_mcontext_t structure and breaks backwards compatibility with earlier versions of DynamoRIO so the version number is increased to 10.90.
Issues: #6760, #5365
Fixes: #6760