[arm64] Add the full neon regs to the context

The full regs are still not handled in the trampolines. Only the lower half is saved.

diff --git a/mono/mini/exceptions-arm64.c b/mono/mini/exceptions-arm64.c
index fb8d1736c10ac7b69b803ea9d0174677b902b9e7..9e19a8d25e720d597476dad7277367f752da16d4 100644 (file)
--- a/mono/mini/exceptions-arm64.c
+++ b/mono/mini/exceptions-arm64.c
@@ -43,7 +43,7 @@ mono_arch_get_restore_context (MonoTrampInfo **info, gboolean aot)
        labels [0] = code;
        arm_cbzx (code, ARMREG_IP1, 0);
        for (i = 0; i < 32; ++i)
-               arm_ldrfpx (code, i, ctx_reg, MONO_STRUCT_OFFSET (MonoContext, fregs) + (i * 8));
+               arm_ldrfpx (code, i, ctx_reg, MONO_STRUCT_OFFSET (MonoContext, fregs) + (i * sizeof (MonoContextSimdReg)));
        mono_arm_patch (labels [0], code, MONO_R_ARM64_CBZ);
        /* Restore gregs */
        // FIXME: Restore less registers
@@ -124,7 +124,7 @@ mono_arch_get_call_filter (MonoTrampInfo **info, gboolean aot)
        labels [0] = code;
        arm_cbzx (code, ARMREG_IP0, 0);
        for (i = 0; i < num_fregs; ++i)
-               arm_ldrfpx (code, ARMREG_D8 + i, ARMREG_R0, MONO_STRUCT_OFFSET (MonoContext, fregs) + ((i + 8) * 8));
+               arm_ldrfpx (code, ARMREG_D8 + i, ARMREG_R0, MONO_STRUCT_OFFSET (MonoContext, fregs) + ((i + 8) * sizeof (MonoContextSimdReg)));
        mono_arm_patch (labels [0], code, MONO_R_ARM64_CBZ);
        /* Load fp */
        arm_ldrx (code, ARMREG_FP, ARMREG_R0, MONO_STRUCT_OFFSET (MonoContext, regs) + (ARMREG_FP * 8));
@@ -393,7 +393,8 @@ mono_arm_throw_exception (gpointer arg, mgreg_t pc, mgreg_t *int_regs, gdouble *
        /* Initialize a ctx based on the arguments */
        memset (&ctx, 0, sizeof (MonoContext));
        memcpy (&(ctx.regs [0]), int_regs, sizeof (mgreg_t) * 32);
-       memcpy (&(ctx.fregs [ARMREG_D8]), fp_regs, sizeof (double) * 8);
+       for (int i = 0; i < 8; i++)
+               *((gdouble*)&ctx.fregs [ARMREG_D8 + i]) = fp_regs [i];
        ctx.has_fregs = 1;
        ctx.pc = pc;
 
@@ -422,7 +423,8 @@ mono_arm_resume_unwind (gpointer arg, mgreg_t pc, mgreg_t *int_regs, gdouble *fp
        /* Initialize a ctx based on the arguments */
        memset (&ctx, 0, sizeof (MonoContext));
        memcpy (&(ctx.regs [0]), int_regs, sizeof (mgreg_t) * 32);
-       memcpy (&(ctx.fregs [ARMREG_D8]), fp_regs, sizeof (double) * 8);
+       for (int i = 0; i < 8; i++)
+               *((gdouble*)&ctx.fregs [ARMREG_D8 + i]) = fp_regs [i];
        ctx.has_fregs = 1;
        ctx.pc = pc;
 
@@ -460,7 +462,8 @@ mono_arch_unwind_frame (MonoDomain *domain, MonoJitTlsData *jit_tls,
 
                memcpy (regs, &new_ctx->regs, sizeof (mgreg_t) * 32);
                /* v8..v15 are callee saved */
-               memcpy (regs + MONO_MAX_IREGS, &(new_ctx->fregs [8]), sizeof (mgreg_t) * 8);
+               for (int i = 0; i < 8; i++)
+                       (regs + MONO_MAX_IREGS) [i] = *((mgreg_t*)&new_ctx->fregs [8 + i]);
 
                mono_unwind_frame (unwind_info, unwind_info_len, ji->code_start, 
                                                   (guint8*)ji->code_start + ji->code_size,
@@ -468,7 +471,8 @@ mono_arch_unwind_frame (MonoDomain *domain, MonoJitTlsData *jit_tls,
                                                   save_locations, MONO_MAX_IREGS, &cfa);
 
                memcpy (&new_ctx->regs, regs, sizeof (mgreg_t) * 32);
-               memcpy (&(new_ctx->fregs [8]), regs + MONO_MAX_IREGS, sizeof (mgreg_t) * 8);
+               for (int i = 0; i < 8; i++)
+                       *((mgreg_t*)&new_ctx->fregs [8 + i]) = (regs + MONO_MAX_IREGS) [i];
 
                new_ctx->pc = regs [ARMREG_LR];
                new_ctx->regs [ARMREG_SP] = (mgreg_t)cfa;

diff --git a/mono/utils/mono-context.c b/mono/utils/mono-context.c
index 3861e54f5e6a04b75c661be7308939c904008859..59ef74b215c1b642bc2967402a7f42d941be5ae5 100644 (file)
--- a/mono/utils/mono-context.c
+++ b/mono/utils/mono-context.c
@@ -401,8 +401,7 @@ mono_sigctx_to_monoctx (void *sigctx, MonoContext *mctx)
 
        g_assert (fpctx->head.magic == FPSIMD_MAGIC);
        for (i = 0; i < 32; ++i)
-               /* Only store the bottom 8 bytes for now */
-               *(guint64*)&(mctx->fregs [i]) = fpctx->vregs [i];
+               mctx->fregs [i] = fpctx->vregs [i];
 #endif
        /* FIXME: apple */
 #endif

diff --git a/mono/utils/mono-context.h b/mono/utils/mono-context.h
index bac2fb331a1633ee4e143999f7b101870529e748..290f24f115cb8ef7a412716205b0515e4d846779 100644 (file)
--- a/mono/utils/mono-context.h
+++ b/mono/utils/mono-context.h
@@ -32,6 +32,8 @@ typedef struct __darwin_xmm_reg MonoContextSimdReg;
 #elif defined(__linux__)
 typedef struct _libc_xmmreg MonoContextSimdReg;
 #endif
+#elif defined(TARGET_ARM64)
+typedef __uint128_t MonoContextSimdReg;
 #endif
 
 /*
@@ -425,7 +427,8 @@ typedef struct {
 
 typedef struct {
        mgreg_t regs [32];
-       double fregs [32];
+       /* FIXME not fully saved in trampolines */
+       MonoContextSimdReg fregs [32];
        mgreg_t pc;
        /*
         * fregs might not be initialized if this context was created from a
@@ -478,22 +481,22 @@ typedef struct {
                "stp x30, xzr, [x16], #8\n"     \
                "mov x30, sp\n"                         \
                "str x30, [x16], #8\n"          \
-               "stp d0, d1, [x16], #16\n"      \
-               "stp d2, d3, [x16], #16\n"      \
-               "stp d4, d5, [x16], #16\n"      \
-               "stp d6, d7, [x16], #16\n"      \
-               "stp d8, d9, [x16], #16\n"      \
-               "stp d10, d11, [x16], #16\n"    \
-               "stp d12, d13, [x16], #16\n"    \
-               "stp d14, d15, [x16], #16\n"    \
-               "stp d16, d17, [x16], #16\n"    \
-               "stp d18, d19, [x16], #16\n"    \
-               "stp d20, d21, [x16], #16\n"    \
-               "stp d22, d23, [x16], #16\n"    \
-               "stp d24, d25, [x16], #16\n"    \
-               "stp d26, d27, [x16], #16\n"    \
-               "stp d28, d29, [x16], #16\n"    \
-               "stp d30, d31, [x16], #16\n"    \
+               "stp q0, q1, [x16], #32\n"      \
+               "stp q2, q3, [x16], #32\n"      \
+               "stp q4, q5, [x16], #32\n"      \
+               "stp q6, q7, [x16], #32\n"      \
+               "stp q8, q9, [x16], #32\n"      \
+               "stp q10, q11, [x16], #32\n"    \
+               "stp q12, q13, [x16], #32\n"    \
+               "stp q14, q15, [x16], #32\n"    \
+               "stp q16, q17, [x16], #32\n"    \
+               "stp q18, q19, [x16], #32\n"    \
+               "stp q20, q21, [x16], #32\n"    \
+               "stp q22, q23, [x16], #32\n"    \
+               "stp q24, q25, [x16], #32\n"    \
+               "stp q26, q27, [x16], #32\n"    \
+               "stp q28, q29, [x16], #32\n"    \
+               "stp q30, q31, [x16], #32\n"            \
                :                                                       \
                : "r" (&ctx.regs)                       \
                : "x16", "x30", "memory"                \