From 85fc716732bc6e85a634335847999f411269f282 Mon Sep 17 00:00:00 2001 From: Richard Henderson Date: Fri, 9 Mar 2018 17:09:43 +0000 Subject: linux-user: Implement aarch64 PR_SVE_SET/GET_VL MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit As an implementation choice, widening VL has zeroed the previously inaccessible portion of the sve registers. Reviewed-by: Peter Maydell Signed-off-by: Richard Henderson Acked-by: Alex Bennée Message-id: 20180303143823.27055-2-richard.henderson@linaro.org Signed-off-by: Peter Maydell --- linux-user/aarch64/target_syscall.h | 3 +++ linux-user/syscall.c | 27 ++++++++++++++++++++++++ target/arm/cpu.h | 1 + target/arm/cpu64.c | 41 +++++++++++++++++++++++++++++++++++++ 4 files changed, 72 insertions(+) diff --git a/linux-user/aarch64/target_syscall.h b/linux-user/aarch64/target_syscall.h index 604ab99..205265e 100644 --- a/linux-user/aarch64/target_syscall.h +++ b/linux-user/aarch64/target_syscall.h @@ -19,4 +19,7 @@ struct target_pt_regs { #define TARGET_MLOCKALL_MCL_CURRENT 1 #define TARGET_MLOCKALL_MCL_FUTURE 2 +#define TARGET_PR_SVE_SET_VL 50 +#define TARGET_PR_SVE_GET_VL 51 + #endif /* AARCH64_TARGET_SYSCALL_H */ diff --git a/linux-user/syscall.c b/linux-user/syscall.c index a8abfd4..b4f7b14 100644 --- a/linux-user/syscall.c +++ b/linux-user/syscall.c @@ -10672,6 +10672,33 @@ abi_long do_syscall(void *cpu_env, int num, abi_long arg1, break; } #endif +#ifdef TARGET_AARCH64 + case TARGET_PR_SVE_SET_VL: + /* We cannot support either PR_SVE_SET_VL_ONEXEC + or PR_SVE_VL_INHERIT. Therefore, anything above + ARM_MAX_VQ results in EINVAL. */ + ret = -TARGET_EINVAL; + if (arm_feature(cpu_env, ARM_FEATURE_SVE) + && arg2 >= 0 && arg2 <= ARM_MAX_VQ * 16 && !(arg2 & 15)) { + CPUARMState *env = cpu_env; + int old_vq = (env->vfp.zcr_el[1] & 0xf) + 1; + int vq = MAX(arg2 / 16, 1); + + if (vq < old_vq) { + aarch64_sve_narrow_vq(env, vq); + } + env->vfp.zcr_el[1] = vq - 1; + ret = vq * 16; + } + break; + case TARGET_PR_SVE_GET_VL: + ret = -TARGET_EINVAL; + if (arm_feature(cpu_env, ARM_FEATURE_SVE)) { + CPUARMState *env = cpu_env; + ret = ((env->vfp.zcr_el[1] & 0xf) + 1) * 16; + } + break; +#endif /* AARCH64 */ case PR_GET_SECCOMP: case PR_SET_SECCOMP: /* Disable seccomp to prevent the target disabling syscalls we diff --git a/target/arm/cpu.h b/target/arm/cpu.h index 3fa8fda..36711cd 100644 --- a/target/arm/cpu.h +++ b/target/arm/cpu.h @@ -866,6 +866,7 @@ int arm_cpu_write_elf32_note(WriteCoreDumpFunction f, CPUState *cs, #ifdef TARGET_AARCH64 int aarch64_cpu_gdb_read_register(CPUState *cpu, uint8_t *buf, int reg); int aarch64_cpu_gdb_write_register(CPUState *cpu, uint8_t *buf, int reg); +void aarch64_sve_narrow_vq(CPUARMState *env, unsigned vq); #endif target_ulong do_arm_semihosting(CPUARMState *env); diff --git a/target/arm/cpu64.c b/target/arm/cpu64.c index dd9ba97..f12a485 100644 --- a/target/arm/cpu64.c +++ b/target/arm/cpu64.c @@ -368,3 +368,44 @@ static void aarch64_cpu_register_types(void) } type_init(aarch64_cpu_register_types) + +/* The manual says that when SVE is enabled and VQ is widened the + * implementation is allowed to zero the previously inaccessible + * portion of the registers. The corollary to that is that when + * SVE is enabled and VQ is narrowed we are also allowed to zero + * the now inaccessible portion of the registers. + * + * The intent of this is that no predicate bit beyond VQ is ever set. + * Which means that some operations on predicate registers themselves + * may operate on full uint64_t or even unrolled across the maximum + * uint64_t[4]. Performing 4 bits of host arithmetic unconditionally + * may well be cheaper than conditionals to restrict the operation + * to the relevant portion of a uint16_t[16]. + * + * TODO: Need to call this for changes to the real system registers + * and EL state changes. + */ +void aarch64_sve_narrow_vq(CPUARMState *env, unsigned vq) +{ + int i, j; + uint64_t pmask; + + assert(vq >= 1 && vq <= ARM_MAX_VQ); + + /* Zap the high bits of the zregs. */ + for (i = 0; i < 32; i++) { + memset(&env->vfp.zregs[i].d[2 * vq], 0, 16 * (ARM_MAX_VQ - vq)); + } + + /* Zap the high bits of the pregs and ffr. */ + pmask = 0; + if (vq & 3) { + pmask = ~(-1ULL << (16 * (vq & 3))); + } + for (j = vq / 4; j < ARM_MAX_VQ / 4; j++) { + for (i = 0; i < 17; ++i) { + env->vfp.pregs[i].p[j] &= pmask; + } + pmask = 0; + } +} -- cgit v1.1