LLDB SVE predicate variables displayed incorrectly

[yury-khrustalev]

LLDB does not display SVE predicate variables (of type svbool_t) correctly:

1. Even with DW_AT_bit_stride set correctly for svbool_t (see clang DWARF SVE predicate type does not have DW_AT_bit_stride #159285), LLDB reads 1 byte for an element of the predicate vector.
2. LLDB always reads 64 bits for a predicate vector implying vector length 512 bits.

For code

svbool_t pred16 = svwhilelt_b16_u32 (0u, 5u);

For VL = 512 bits:

(lldb) p/t *(uint64_t*)&pred16
(uint64_t) 0b0000000000000000000000000000000000000000000000000000000101010101
(lldb) p pred16
(bool __attribute__((ext_vector_type(8)))) (true, true, false, false, false, false, false, false)

For VL = 256 bits:

(lldb) p/t *(uint64_t*)&pred16
(uint64_t) 0b1010101010101010000001111011010000000000000000000000000101010101
(lldb) p pred16
(bool __attribute__((ext_vector_type(8)))) (true, true, false, false, true, true, true, true)

Notice that LLDB treats 1st byte 01010101 as true and then 2nd byte 00000001 as true while this was supposed to be several elements.

In the second example (for VL = 256 bits) LLDB reads memory beyond 32 bits of the predicate vector.

[llvmbot]

@llvm/issue-subscribers-lldb

Author: Yury Khrustalev (yury-khrustalev)

LLDB does not display SVE predicate variables (of type `svbool_t`) correctly:

1. Even with DW_AT_bit_stride set correctly for svbool_t (see #159285), LLDB reads 1 byte for an element of the predicate vector.
2. LLDB always reads 64 bits for a predicate vector implying vector length 512 bits.

For code

svbool_t pred16 = svwhilelt_b16_u32 (0u, 5u);

For VL = 512 bits:

(lldb) p/t *(uint64_t*)&pred16
(uint64_t) 0b0000000000000000000000000000000000000000000000000000000101010101
(lldb) p pred16
(bool __attribute__((ext_vector_type(8)))) (true, true, false, false, false, false, false, false)

For VL = 256 bits:

(lldb) p/t *(uint64_t*)&pred16
(uint64_t) 0b1010101010101010000001111011010000000000000000000000000101010101
(lldb) p pred16
(bool __attribute__((ext_vector_type(8)))) (true, true, false, false, true, true, true, true)

Notice that LLDB treats 1st byte 01010101 as true and then 2nd byte 00000001 as true while this was supposed to be several elements.

In the second example (for VL = 256 bits) LLDB reads memory beyond 32 bits of the predicate vector.

[llvmbot]

@llvm/issue-subscribers-backend-aarch64

Author: Yury Khrustalev (yury-khrustalev)

LLDB does not display SVE predicate variables (of type `svbool_t`) correctly:

1. Even with DW_AT_bit_stride set correctly for svbool_t (see #159285), LLDB reads 1 byte for an element of the predicate vector.
2. LLDB always reads 64 bits for a predicate vector implying vector length 512 bits.

For code

svbool_t pred16 = svwhilelt_b16_u32 (0u, 5u);

For VL = 512 bits:

(lldb) p/t *(uint64_t*)&pred16
(uint64_t) 0b0000000000000000000000000000000000000000000000000000000101010101
(lldb) p pred16
(bool __attribute__((ext_vector_type(8)))) (true, true, false, false, false, false, false, false)

For VL = 256 bits:

(lldb) p/t *(uint64_t*)&pred16
(uint64_t) 0b1010101010101010000001111011010000000000000000000000000101010101
(lldb) p pred16
(bool __attribute__((ext_vector_type(8)))) (true, true, false, false, true, true, true, true)

Notice that LLDB treats 1st byte 01010101 as true and then 2nd byte 00000001 as true while this was supposed to be several elements.

In the second example (for VL = 256 bits) LLDB reads memory beyond 32 bits of the predicate vector.

[DavidSpickett]

From the AArch64 AAPCS:
"The number of bytes in the vector is always VG×8, where VG is a runtime value determined by the execution environment."

Internal Type   arm_sve.h type      base type                                elements
__SVBool_t	    svbool_t	        single bit (fully packed into VG bytes)  VG×8

At 512 bit vector length each vector is 64 bytes. 64/8 = 8 so each svbool_t is 8 bytes. Which we should interpret as 64 1-bit boolean values (1 bit per element for the smallest possible lane size).

For 256 bits that's 32 bytes, 32/8 = 4. So we should show 32 boolean elements.

DW_AT_bit_stride

Setting this to 1 corresponds to the AAPCS saying "(fully packed into VG bytes)".

We'll need to respect the current vector length then. Right now ext_vector_type(8) is probably where we get it from.

[DavidSpickett]

I tried deleting all code that handles DW_AT_bit_stride and the similar DW_AT_byte_stride, in an AArch64 Linux build. Zero test failures.

So the first job here is to see whether fixed size types work at all, and if so add test coverage for them.