As with SVE rather than impose ambitious toolchain requirements for SME we manually encode the few instructions which we require in order to perform the work the kernel needs to do. That is currently:
- Vector store and load for the ZA array. - Zeroing of the whole ZA array.
We do not use SMSTART and SMSTOP, instead saving and restoring SVCR via system register access.
Signed-off-by: Mark Brown broonie@kernel.org --- arch/arm64/include/asm/fpsimdmacros.h | 44 +++++++++++++++++++++++++++ 1 file changed, 44 insertions(+)
diff --git a/arch/arm64/include/asm/fpsimdmacros.h b/arch/arm64/include/asm/fpsimdmacros.h index b22538a6137e..bc45bb984c49 100644 --- a/arch/arm64/include/asm/fpsimdmacros.h +++ b/arch/arm64/include/asm/fpsimdmacros.h @@ -93,6 +93,12 @@ .endif .endm
+.macro _sme_check_wv v + .if (\v) < 12 || (\v) > 15 + .error "Bad vector select register \v." + .endif +.endm + /* SVE instruction encodings for non-SVE-capable assemblers */ /* (pre binutils 2.28, all kernel capable clang versions support SVE) */
@@ -174,6 +180,44 @@ | (\np) .endm
+/* SME instruction encodings for non-SME-capable assemblers */ + +/* + * STR (vector from ZA array): + * STR ZA[\nw, #\offset], [X\nxbase, #\offset, MUL VL] + */ +.macro _sme_str_zav nw, nxbase, offset=0 + _sme_check_wv \nw + _check_general_reg \nxbase + _check_num (\offset), -0x100, 0xff + .inst 0xe1200000 \ + | (((\nw) & 3) << 13) \ + | ((\nxbase) << 5) \ + | ((\offset) & 7) +.endm + +/* + * LDR (vector to ZA array): + * LDR ZA[\nw, #\offset], [X\nxbase, #\offset, MUL VL] + */ +.macro _sme_ldr_zav nw, nxbase, offset=0 + _sme_check_wv \nw + _check_general_reg \nxbase + _check_num (\offset), -0x100, 0xff + .inst 0xe1000000 \ + | (((\nw) & 3) << 13) \ + | ((\nxbase) << 5) \ + | ((\offset) & 7) +.endm + +/* + * Zero the entire ZA array + * ZERO ZA + */ +.macro zero_za + .inst 0xc00800ff +.endm + .macro __for from:req, to:req .if (\from) == (\to) _for__body %\from