i#3544 RV64: Rebase the dcontext pointer. (#7235)

For RISC-V the dcontext_t struct is larger than the biggest valid
displacement of the load and store instructions.

By rebasing the pointer kept in spill_state_t's TLS slot by 0x800, we
can access the entire struct, because the displacement can be in the
range of -0x800 to 0x7ff. Architectures other than RV64 are unaffected
by these changes.

Issue: #3544

core/arch/arch_exports.h

@@ -1,6 +1,7 @@
 /* **********************************************************
  * Copyright (c) 2011-2022 Google, Inc.  All rights reserved.
  * Copyright (c) 2000-2010 VMware, Inc.  All rights reserved.
+ * Copyright (c) 2025 Foundation of Research and Technology, Hellas.
  * **********************************************************/
 
 /*
@@ -165,6 +166,10 @@ typedef struct _spill_state_t {
     reg_t reg_stolen;
 #endif
     /* XXX: move this below the tables to fit more on cache line */
+    /* In RISCV64 dcontext does not poiont to the actual dcontext. It points somewhere
+     * in the middle of that. This value is the actual pointer offseted by
+     * DCONTEXT_TLS_MIDPTR_OFFSET.
+     */
     dcontext_t *dcontext;
 #if defined(RISCV64) || defined(AARCHXX)
     /* We store addresses here so we can load pointer-sized addresses into
@@ -1819,4 +1824,29 @@ typedef struct _rseq_entry_state_t {
 } rseq_entry_state_t;
 #endif
 
+/*
+ * In riscv we cannot address the entire dcontext by using base + immediate.
+ * For the reason we add 0x800 to the saved pointer and access it by
+ * substracting 0x800 from the offset in the struct.
+ */
+#ifdef RISCV64
+#    define DCONTEXT_TLS_MIDPTR_OFFSET 0x800
+#else
+#    define DCONTEXT_TLS_MIDPTR_OFFSET 0
+#endif
+
+#if (DCONTEXT_TLS_MIDPTR_OFFSET != 0)
+#    define DCONTEXT_ACTUAL_TO_TLS_PTR(x) \
+        ((dcontext_t *)(((ptr_uint_t)x) + DCONTEXT_TLS_MIDPTR_OFFSET))
+#    define DCONTEXT_TLS_TO_ACTUAL_PTR(x) \
+        ((dcontext_t *)(((ptr_uint_t)x) - DCONTEXT_TLS_MIDPTR_OFFSET))
+#    define DCONTEXT_ACTUAL_TO_TLS_OFFSET(x) (x - DCONTEXT_TLS_MIDPTR_OFFSET)
+#    define DCONTEXT_TLS_TO_ACTUAL_OFFSET(x) (x + DCONTEXT_TLS_MIDPTR_OFFSET)
+#else
+#    define DCONTEXT_ACTUAL_TO_TLS_PTR(x) x
+#    define DCONTEXT_TLS_TO_ACTUAL_PTR(x) x
+#    define DCONTEXT_ACTUAL_TO_TLS_OFFSET(x) x
+#    define DCONTEXT_TLS_TO_ACTUAL_OFFSET(x) x
+#endif
+
 #endif /* _ARCH_EXPORTS_H_ */

core/arch/emit_utils_shared.c

@@ -1,6 +1,7 @@
 /* **********************************************************
  * Copyright (c) 2010-2022 Google, Inc.  All rights reserved.
  * Copyright (c) 2000-2010 VMware, Inc.  All rights reserved.
+ * Copyright (c) 2025 Foundation of Research and Technology, Hellas.
  * **********************************************************/
 
 /*
@@ -2180,12 +2181,16 @@ emit_fcache_enter_common(dcontext_t *dcontext, generated_code_t *code, byte *pc,
 #elif defined(RISCV64)
     APP(&ilist,
         INSTR_CREATE_ld(dcontext, opnd_create_reg(DR_REG_A0),
-                        opnd_create_base_disp(REG_DCXT, DR_REG_NULL, 0,
-                                              REG_OFFSET(DR_REG_A0), OPSZ_8)));
+                        opnd_create_base_disp(
+                            REG_DCXT, DR_REG_NULL, 0,
+                            DCONTEXT_ACTUAL_TO_TLS_OFFSET((int)REG_OFFSET(DR_REG_A0)),
+                            OPSZ_8)));
     APP(&ilist,
         INSTR_CREATE_ld(dcontext, opnd_create_reg(DR_REG_A1),
-                        opnd_create_base_disp(REG_DCXT, DR_REG_NULL, 0,
-                                              REG_OFFSET(DR_REG_A1), OPSZ_8)));
+                        opnd_create_base_disp(
+                            REG_DCXT, DR_REG_NULL, 0,
+                            DCONTEXT_ACTUAL_TO_TLS_OFFSET((int)REG_OFFSET(DR_REG_A1)),
+                            OPSZ_8)));
 
     APP(&ilist,
         INSTR_CREATE_sd(
@@ -2380,9 +2385,31 @@ append_call_dispatch(dcontext_t *dcontext, instrlist_t *ilist, bool absolute)
     /* call central d_r_dispatch routine */
     /* for x64 linux we could optimize and avoid the "mov rdi, rdi" */
     /* for ARM we use _noreturn to avoid storing to %lr */
+
+    /*
+     * REG_DCTXT holds the rebased dcontext d_r_dispatch expects the normal one
+     * so we should restore it.
+     *
+     * Currently only affects RISCV64
+     */
+#if (DCONTEXT_TLS_MIDPTR_OFFSET != 0)
+    if (absolute) {
+        dr_insert_call_noreturn((void *)dcontext, ilist, NULL /*append*/,
+                                (void *)d_r_dispatch, 1,
+                                OPND_CREATE_INTPTR((ptr_int_t)dcontext));
+    } else {
+        APP(ilist,
+            XINST_CREATE_add_2src(dcontext, opnd_create_reg(DR_REG_A0),
+                                  opnd_create_reg(REG_DCTXT),
+                                  OPND_CREATE_INT32(-DCONTEXT_TLS_MIDPTR_OFFSET)));
+        dr_insert_call_noreturn((void *)dcontext, ilist, NULL /*append*/,
+                                (void *)d_r_dispatch, 1, opnd_create_reg(DR_REG_A0));
+    }
+#else
     dr_insert_call_noreturn(
         (void *)dcontext, ilist, NULL /*append*/, (void *)d_r_dispatch, 1,
         absolute ? OPND_CREATE_INTPTR((ptr_int_t)dcontext) : opnd_create_reg(REG_DCTXT));
+#endif
 
     /* d_r_dispatch() shouldn't return! */
     insert_reachable_cti(dcontext, ilist, NULL, vmcode_get_start(),

core/arch/mangle_shared.c

@@ -2,6 +2,7 @@
  * Copyright (c) 2010-2022 Google, Inc.  All rights reserved.
  * Copyright (c) 2010 Massachusetts Institute of Technology  All rights reserved.
  * Copyright (c) 2000-2010 VMware, Inc.  All rights reserved.
+ * Copyright (c) 2025 Foundation of Research and Technology, Hellas.
  * ******************************************************************************/
 
 /*
@@ -81,6 +82,12 @@ get_clean_call_temp_stack_size(void)
 /* utility routines for inserting clean calls to an instrumentation routine
  * strategy is very similar to fcache_enter/return
  * FIXME: try to share code with fcache_enter/return?
+ * TODO i#3544: Return the correct mcontext base when DCONTEXT_TLS_MIDPTR_OFFSET is used.
+ * This will need calls like opnd_create_dcontext_field_via_reg_sz to be replaced
+ * with something else. Currently we work around that by assuming that we have the
+ * dcontext pointer (offseted) instead of the mcontext when DCONTEXT_TLS_MIDPTR_OFFSET is
+ * not zero. For that reason we substract DCONTEXT_TLS_MIDPTR_OFFSET in the offsets
+ * created at emit_fcache_enter_common().
  *
  * first swap stacks to DynamoRIO stack:
  *      SAVE_TO_UPCONTEXT %xsp,xsp_OFFSET
@@ -757,8 +764,9 @@ insert_meta_call_vargs(dcontext_t *dcontext, instrlist_t *ilist, instr_t *instr,
              * We save it to dcontext.mcontext.x0.
              */
             PRE(ilist, instr,
-                XINST_CREATE_store(dcontext, OPND_CREATE_MEMPTR(link_reg, 0),
-                                   opnd_create_reg(SCRATCH_REG0)));
+                XINST_CREATE_store(
+                    dcontext, OPND_CREATE_MEMPTR(link_reg, -DCONTEXT_TLS_MIDPTR_OFFSET),
+                    opnd_create_reg(SCRATCH_REG0)));
             instrlist_insert_mov_immed_ptrsz(dcontext, (ptr_int_t)DR_WHERE_CLEAN_CALLEE,
                                              opnd_create_reg(SCRATCH_REG0), ilist, instr,
                                              NULL, NULL);
@@ -769,8 +777,9 @@ insert_meta_call_vargs(dcontext_t *dcontext, instrlist_t *ilist, instr_t *instr,
                     WHEREAMI_OFFSET));
             /* Restore scratch_reg from dcontext.mcontext.x0. */
             PRE(ilist, instr,
-                XINST_CREATE_load(dcontext, opnd_create_reg(SCRATCH_REG0),
-                                  OPND_CREATE_MEMPTR(link_reg, 0)));
+                XINST_CREATE_load(
+                    dcontext, opnd_create_reg(SCRATCH_REG0),
+                    OPND_CREATE_MEMPTR(link_reg, -DCONTEXT_TLS_MIDPTR_OFFSET)));
 #else
             /* SCRATCH_REG0 is dead here, because clean calls only support "cdecl",
              * which specifies that the caller must save xax (and xcx and xdx).
@@ -823,8 +832,10 @@ insert_meta_call_vargs(dcontext_t *dcontext, instrlist_t *ilist, instr_t *instr,
              * We save it to dcontext.mcontext.x0.
              */
             PRE(ilist, instr,
-                XINST_CREATE_store(dcontext, OPND_CREATE_MEMPTR(SCRATCH_REG0, 0),
-                                   opnd_create_reg(SCRATCH_REG1)));
+                XINST_CREATE_store(
+                    dcontext,
+                    OPND_CREATE_MEMPTR(SCRATCH_REG0, -DCONTEXT_TLS_MIDPTR_OFFSET),
+                    opnd_create_reg(SCRATCH_REG1)));
             instrlist_insert_mov_immed_ptrsz(dcontext, (ptr_int_t)whereami,
                                              opnd_create_reg(SCRATCH_REG1), ilist, instr,
                                              NULL, NULL);
@@ -835,8 +846,9 @@ insert_meta_call_vargs(dcontext_t *dcontext, instrlist_t *ilist, instr_t *instr,
                     WHEREAMI_OFFSET));
             /* Restore scratch_reg from dcontext.mcontext.x0. */
             PRE(ilist, instr,
-                XINST_CREATE_load(dcontext, opnd_create_reg(SCRATCH_REG1),
-                                  OPND_CREATE_MEMPTR(SCRATCH_REG0, 0)));
+                XINST_CREATE_load(
+                    dcontext, opnd_create_reg(SCRATCH_REG1),
+                    OPND_CREATE_MEMPTR(SCRATCH_REG0, -DCONTEXT_TLS_MIDPTR_OFFSET)));
 #else
             PRE(ilist, instr,
                 instr_create_save_immed_to_dc_via_reg(dcontext, SCRATCH_REG0,

core/arch/riscv64/emit_utils.c

@@ -1,6 +1,6 @@
 /* **********************************************************
  * Copyright (c) 2022 Rivos, Inc.  All rights reserved.
- * Copyright (c) 2024 Foundation of Research and Technology, Hellas.
+ * Copyright (c) 2024-2025 Foundation of Research and Technology, Hellas.
  * **********************************************************/
 
 /*
@@ -691,19 +691,20 @@ append_restore_simd_reg(dcontext_t *dcontext, instrlist_t *ilist, bool absolute)
     }
 
     if (proc_has_feature(FEATURE_VECTOR)) {
+        APP(ilist,
+            INSTR_CREATE_addi(
+                dcontext, opnd_create_reg(DR_REG_A1), opnd_create_reg(REG_DCXT),
+                opnd_create_immed_int(
+                    DCONTEXT_ACTUAL_TO_TLS_OFFSET(VREG_OFFSET(DR_REG_VR0)), OPSZ_12b)));
+        memopnd = opnd_create_base_disp(DR_REG_A1, REG_NULL, 0, 0,
+                                        reg_get_size_lmul(DR_REG_VR0, RV64_LMUL_8));
         /* ma:   mask agnostic
          * ta:   tail agnostic
          * sew:  selected element width
          * lmul: vector register group multiplier
          *
          *           ma            ta         sew=8       lmul=8 */
         vtypei = (0b1 << 7) | (0b1 << 6) | (0b000 << 3) | 0b011;
-        memopnd = opnd_create_dcontext_field_via_reg_sz(
-            dcontext, DR_REG_A1, 0, reg_get_size_lmul(DR_REG_VR0, RV64_LMUL_8));
-        APP(ilist,
-            INSTR_CREATE_addi(dcontext, opnd_create_reg(DR_REG_A1),
-                              opnd_create_reg(REG_DCXT),
-                              opnd_create_immed_int(VREG_OFFSET(DR_REG_VR0), OPSZ_12b)));
         /* For the following vector instructions, set the element width to 8b, and use 8
          * registers as a group (lmul=8).
          */
@@ -828,19 +829,20 @@ append_save_simd_reg(dcontext_t *dcontext, instrlist_t *ilist, bool absolute)
     }
 
     if (proc_has_feature(FEATURE_VECTOR)) {
+        APP(ilist,
+            INSTR_CREATE_addi(
+                dcontext, opnd_create_reg(DR_REG_A1), opnd_create_reg(REG_DCXT),
+                opnd_create_immed_int(
+                    DCONTEXT_ACTUAL_TO_TLS_OFFSET(VREG_OFFSET(DR_REG_VR0)), OPSZ_12b)));
+        memopnd = opnd_create_base_disp(DR_REG_A1, REG_NULL, 0, 0,
+                                        reg_get_size_lmul(DR_REG_VR0, RV64_LMUL_8));
         /* ma:   mask agnostic
          * ta:   tail agnostic
          * sew:  selected element width
          * lmul: vector register group multiplier
          *
          *           ma            ta         sew=8       lmul=8 */
         vtypei = (0b1 << 7) | (0b1 << 6) | (0b000 << 3) | 0b011;
-        memopnd = opnd_create_dcontext_field_via_reg_sz(
-            dcontext, DR_REG_A1, 0, reg_get_size_lmul(DR_REG_VR0, RV64_LMUL_8));
-        APP(ilist,
-            INSTR_CREATE_addi(dcontext, opnd_create_reg(DR_REG_A1),
-                              opnd_create_reg(REG_DCXT),
-                              opnd_create_immed_int(VREG_OFFSET(DR_REG_VR0), OPSZ_12b)));
         /* For the following vector instructions, set the element width to 8b, and use 8
          * registers as a group (lmul=8).
          */

core/arch/riscv64/mangle.c

@@ -1,6 +1,6 @@
 /* **********************************************************
  * Copyright (c) 2022 Rivos, Inc.  All rights reserved.
- * Copyright (c) 2024 Foundation of Research and Technology, Hellas.
+ * Copyright (c) 2024-2025 Foundation of Research and Technology, Hellas.
  * **********************************************************/
 
 /*
@@ -223,19 +223,19 @@ insert_push_all_registers(dcontext_t *dcontext, clean_call_info_t *cci,
 
     /* Push vector registers. */
     if (proc_has_feature(FEATURE_VECTOR)) {
+        PRE(ilist, instr,
+            INSTR_CREATE_addi(dcontext, opnd_create_reg(DR_REG_A0),
+                              opnd_create_reg(DR_REG_SP),
+                              opnd_create_immed_int(dstack_offs, OPSZ_12b)));
+        memopnd = opnd_create_base_disp(DR_REG_A0, REG_NULL, 0, 0,
+                                        reg_get_size_lmul(DR_REG_VR0, RV64_LMUL_8));
         /* ma:   mask agnostic
          * ta:   tail agnostic
          * sew:  selected element width
          * lmul: vector register group multiplier
          *
          *           ma            ta         sew=8       lmul=8 */
         vtypei = (0b1 << 7) | (0b1 << 6) | (0b000 << 3) | 0b011;
-        memopnd = opnd_create_dcontext_field_via_reg_sz(
-            dcontext, DR_REG_A0, 0, reg_get_size_lmul(DR_REG_VR0, RV64_LMUL_8));
-        PRE(ilist, instr,
-            INSTR_CREATE_addi(dcontext, opnd_create_reg(DR_REG_A0),
-                              opnd_create_reg(DR_REG_SP),
-                              opnd_create_immed_int(dstack_offs, OPSZ_12b)));
         /* For the following vector instructions, set the element width to 8b, and use 8
          * registers as a group (lmul=8).
          */
@@ -294,19 +294,19 @@ insert_pop_all_registers(dcontext_t *dcontext, clean_call_info_t *cci, instrlist
     /* Pop vector registers. */
     current_offs -= proc_num_simd_registers() * sizeof(dr_simd_t);
     if (proc_has_feature(FEATURE_VECTOR)) {
+        PRE(ilist, instr,
+            INSTR_CREATE_addi(dcontext, opnd_create_reg(DR_REG_A0),
+                              opnd_create_reg(DR_REG_SP),
+                              opnd_create_immed_int(current_offs, OPSZ_12b)));
+        memopnd = opnd_create_base_disp(DR_REG_A0, REG_NULL, 0, 0,
+                                        reg_get_size_lmul(DR_REG_VR0, RV64_LMUL_8));
         /* ma:   mask agnostic
          * ta:   tail agnostic
          * sew:  selected element width
          * lmul: vector register group multiplier
          *
          *           ma            ta         sew=8       lmul=8 */
         vtypei = (0b1 << 7) | (0b1 << 6) | (0b000 << 3) | 0b011;
-        memopnd = opnd_create_dcontext_field_via_reg_sz(
-            dcontext, DR_REG_A0, 0, reg_get_size_lmul(DR_REG_VR0, RV64_LMUL_8));
-        PRE(ilist, instr,
-            INSTR_CREATE_addi(dcontext, opnd_create_reg(DR_REG_A0),
-                              opnd_create_reg(DR_REG_SP),
-                              opnd_create_immed_int(current_offs, OPSZ_12b)));
         /* For the following vector instructions, set the element width to 8b, and use 8
          * registers as a group (lmul=8).
          */

core/dispatch.c

@@ -1,6 +1,7 @@
 /* **********************************************************
  * Copyright (c) 2011-2023 Google, Inc.  All rights reserved.
  * Copyright (c) 2000-2010 VMware, Inc.  All rights reserved.
+ * Copyright (c) 2025 Foundation of Research and Technology, Hellas.
  * **********************************************************/
 
 /*
@@ -569,7 +570,8 @@ enter_fcache(dcontext_t *dcontext, fcache_enter_func_t entry, cache_pc pc)
      * paths were missed?
      */
     PTHREAD_JIT_READ();
-    (*entry)(dcontext);
+
+    (*entry)(DCONTEXT_ACTUAL_TO_TLS_PTR(dcontext));
     IF_WINDOWS(ASSERT_NOT_REACHED()); /* returns for signals on unix */
 }
 

core/ir/opnd_shared.c

@@ -1,6 +1,7 @@
 /* **********************************************************
  * Copyright (c) 2011-2022 Google, Inc.  All rights reserved.
  * Copyright (c) 2000-2010 VMware, Inc.  All rights reserved.
+ * Copyright (c) 2025 Foundation of Research and Technology, Hellas.
  * **********************************************************/
 
 /*
@@ -2898,14 +2899,16 @@ dcontext_opnd_common(dcontext_t *dcontext, bool absolute, reg_id_t basereg, int
             absolute ? REG_NULL : (basereg == REG_NULL ? REG_DCXT_PROT : basereg),
             REG_NULL, 0,
             ((int)(ptr_int_t)(absolute ? dcontext->upcontext.separate_upcontext : 0)) +
-                offs,
+                DCONTEXT_ACTUAL_TO_TLS_OFFSET(offs),
             size);
     } else {
         if (offs >= sizeof(unprotected_context_t))
             offs -= sizeof(unprotected_context_t);
         return opnd_create_base_disp(
             absolute ? REG_NULL : (basereg == REG_NULL ? REG_DCXT : basereg), REG_NULL, 0,
-            ((int)(ptr_int_t)(absolute ? dcontext : 0)) + offs, size);
+            ((int)(ptr_int_t)(absolute ? dcontext : 0)) +
+                DCONTEXT_ACTUAL_TO_TLS_OFFSET(offs),
+            size);
     }
 }
 
@@ -2950,7 +2953,8 @@ update_dcontext_address(opnd_t op, dcontext_t *old_dcontext, dcontext_t *new_dco
                       opnd_get_index(op) == REG_NULL,
                   "update_dcontext_address: invalid opnd");
     IF_X64(ASSERT_NOT_IMPLEMENTED(false));
-    offs = opnd_get_disp(op) - (uint)(ptr_uint_t)old_dcontext;
+    offs =
+        opnd_get_disp(op) - (uint)(ptr_uint_t)old_dcontext + DCONTEXT_TLS_MIDPTR_OFFSET;
     if (offs >= 0 && offs < sizeof(dcontext_t)) {
         /* don't pass raw offset, add in upcontext size */
         offs += sizeof(unprotected_context_t);