[PATCH RFC 2/2] RISC-V: add T-Head vector errata handling

[Stefan O'Rear]

On Tue, Feb 28, 2023, at 4:54 PM, Heiko Stuebner wrote:
> @@ -29,6 +78,7 @@ static __always_inline bool has_vector(void)
>  static inline void __vstate_clean(struct pt_regs *regs)
>  {
>  	regs->status = (regs->status & ~(SR_VS)) | SR_VS_CLEAN;
> +
>  }
> 
>  static inline void vstate_off(struct pt_regs *regs)
> @@ -58,30 +108,75 @@ static __always_inline void rvv_disable(void)
> 
>  static __always_inline void __vstate_csr_save(struct __riscv_v_state *dest)
>  {
> -	asm volatile (
> +	register u32 t1 asm("t1") = (SR_FS);
> +
> +	/*
> +	 * CSR_VCSR is defined as
> +	 * [2:1] - vxrm[1:0]
> +	 * [0] - vxsat
> +	 * The earlier vector spec implemented by T-Head uses separate
> +	 * registers for the same bit-elements, so just combine those
> +	 * into the existing output field.
> +	 *
> +	 * Additionally T-Head cores need FS to be enabled when accessing
> +	 * the VXRM and VXSAT CSRs, otherwise ending in illegal instructions.
> +	 */
> +	asm volatile (ALTERNATIVE(
>  		"csrr	%0, " CSR_STR(CSR_VSTART) "\n\t"
>  		"csrr	%1, " CSR_STR(CSR_VTYPE) "\n\t"
>  		"csrr	%2, " CSR_STR(CSR_VL) "\n\t"
>  		"csrr	%3, " CSR_STR(CSR_VCSR) "\n\t"
> +		__nops(5),
> +		"csrs	sstatus, t1\n\t"
> +		"csrr	%0, " CSR_STR(CSR_VSTART) "\n\t"
> +		"csrr	%1, " CSR_STR(CSR_VTYPE) "\n\t"
> +		"csrr	%2, " CSR_STR(CSR_VL) "\n\t"
> +		"csrr	%3, " CSR_STR(THEAD_C9XX_CSR_VXRM) "\n\t"
> +		"slliw	%3, %3, " CSR_STR(VCSR_VXRM_SHIFT) "\n\t"
> +		"csrr	t4, " CSR_STR(THEAD_C9XX_CSR_VXSAT) "\n\t"
> +		"or	%3, %3, t4\n\t"
> +		"csrc	sstatus, t1\n\t",
> +		THEAD_VENDOR_ID,
> +		ERRATA_THEAD_VECTOR, CONFIG_ERRATA_THEAD_VECTOR)
>  		: "=r" (dest->vstart), "=r" (dest->vtype), "=r" (dest->vl),
> -		  "=r" (dest->vcsr) : :);
> +		  "=r" (dest->vcsr) : "r"(t1) : "t4");
>  }
> 
>  static __always_inline void __vstate_csr_restore(struct __riscv_v_state *src)
>  {
> -	asm volatile (
> +	register u32 t1 asm("t1") = (SR_FS);
> +
> +	/*
> +	 * Similar to __vstate_csr_save above, restore values for the
> +	 * separate VXRM and VXSAT CSRs from the vcsr variable.
> +	 */
> +	asm volatile (ALTERNATIVE(
>  		"vsetvl	 x0, %2, %1\n\t"
>  		"csrw	" CSR_STR(CSR_VSTART) ", %0\n\t"
>  		"csrw	" CSR_STR(CSR_VCSR) ", %3\n\t"
> +		__nops(6),
> +		"csrs	sstatus, t1\n\t"
> +		"vsetvl	 x0, %2, %1\n\t"
> +		"csrw	" CSR_STR(CSR_VSTART) ", %0\n\t"
> +		"srliw	t4, %3, " CSR_STR(VCSR_VXRM_SHIFT) "\n\t"
> +		"andi	t4, t4, " CSR_STR(VCSR_VXRM_MASK) "\n\t"
> +		"csrw	" CSR_STR(THEAD_C9XX_CSR_VXRM) ", t4\n\t"
> +		"andi	%3, %3, " CSR_STR(VCSR_VXSAT_MASK) "\n\t"
> +		"csrw	" CSR_STR(THEAD_C9XX_CSR_VXSAT) ", %3\n\t"
> +		"csrc	sstatus, t1\n\t",
> +		THEAD_VENDOR_ID,
> +		ERRATA_THEAD_VECTOR, CONFIG_ERRATA_THEAD_VECTOR)
>  		: : "r" (src->vstart), "r" (src->vtype), "r" (src->vl),
> -		    "r" (src->vcsr) :);
> +		    "r" (src->vcsr), "r"(t1): "t4");
>  }

vxrm and vxsat are part of fcsr in 0.7.1, so they should already have been
handled by __fstate_save and __fstate_restore, and this code is likely to
misbehave (saving the new process's vxrm/vxsat in the old process's save area
because float state is swapped before vector state in __switch_to).

-s