ldrd/strd fix - we definitely need to know that in ldrd r0, r1, [pc], we can't use r1 as scratch

5 files changed, 38 insertions, 20 deletions

diff --git a/lib/dis-arm.inc.h b/lib/dis-arm.inc.h
index 0779205..2f06234 100644
--- a/lib/dis-arm.inc.h
+++ b/lib/dis-arm.inc.h
@@ -65,16 +65,20 @@ static INLINE void P(GPR_Rt_addr_offset_none_addr_am2offset_reg_offset_S_4_STRBT
     data(r(addr), rs(offset, 0, 4), r(Rt));
 }
 static INLINE void P(GPR_Rt_addr_offset_none_addr_am3offset_offset_S_2_STRD_POST)(tdis_ctx ctx, struct bitslice offset, struct bitslice Rt, struct bitslice addr) {
-    data(r(addr), rs(offset, 0, 4), r(Rt));
+    data_flags(IS_LDRD_STRD, r(Rt), r(addr), rs(offset, 0, 4));
 }
 static INLINE void P(GPR_Rt_addr_offset_none_addr_postidx_imm8_offset_S_1_STRHTi)(tdis_ctx ctx, UNUSED struct bitslice offset, struct bitslice Rt, struct bitslice addr) {
     data(r(addr), r(Rt));
 }
 static INLINE void P(GPR_Rt_addrmode3_addr_S_2_STRD)(tdis_ctx ctx, struct bitslice addr, struct bitslice Rt) {
-    data(rs(addr, 9, 4), rs(addr, 0, 4), r(Rt));
+    unsigned addr_val = bs_get(addr, ctx->op);
+    if (addr_val & 1 << 13)
+        data_flags(IS_LDRD_STRD, r(Rt), rs(addr, 9, 4));
+    else
+        data_flags(IS_LDRD_STRD, r(Rt), rs(addr, 9, 4), rs(addr, 0, 4));
 }
 static INLINE void P(GPR_Rt_addrmode3_pre_addr_S_2_STRD_PRE)(tdis_ctx ctx, struct bitslice addr, struct bitslice Rt) {
-    data(rs(addr, 9, 4), rs(addr, 0, 4), r(Rt));
+    return P(GPR_Rt_addrmode3_addr_S_2_STRD)(ctx, addr, Rt);
 }
 static INLINE void P(GPR_Rt_addrmode_imm12_addr_S_1_STRi12)(tdis_ctx ctx, struct bitslice addr, struct bitslice Rt) {
     data(rs(addr, 13, 4), r(Rt));
@@ -125,12 +129,16 @@ static INLINE void P(addr_offset_none_addr_postidx_imm8s4_offset_S_4_STC2L_POST)
 static INLINE void P(addr_offset_none_addr_unk_Rt_13_LDA)(tdis_ctx ctx, struct bitslice Rt, struct bitslice addr) {
     data(rout(Rt), r(addr));
 }
-static INLINE void P(addrmode3_addr_unk_Rt_4_LDRD)(tdis_ctx ctx, struct bitslice addr, UNUSED struct bitslice Rt) {
-    /* ignoring Rt2 = Rt + 1, but it isn't supposed to load PC anyway */
-    data(rs(addr, 9, 4), rs(addr, 0, 4));
+static INLINE void P(addrmode3_addr_unk_Rt_4_LDRD)(tdis_ctx ctx, struct bitslice addr, struct bitslice Rt) {
+    /* ignoring Rt2 = Rt + 1, but LDRD itself isn't supposed to load PC anyway */
+    unsigned addr_val = bs_get(addr, ctx->op);
+    if (addr_val & 1 << 13)
+        data_flags(IS_LDRD_STRD, rout(Rt), rs(addr, 9, 4));
+    else
+        data_flags(IS_LDRD_STRD, rout(Rt), rs(addr, 9, 4), rs(addr, 0, 4));
 }
-static INLINE void P(addrmode3_pre_addr_unk_Rt_4_LDRD_PRE)(tdis_ctx ctx, struct bitslice addr, UNUSED struct bitslice Rt) {
-    data(rs(addr, 9, 4), rs(addr, 0, 4));
+static INLINE void P(addrmode3_pre_addr_unk_Rt_4_LDRD_PRE)(tdis_ctx ctx, struct bitslice addr, struct bitslice Rt) {
+    return P(addrmode3_addr_unk_Rt_4_LDRD)(ctx, addr, Rt);
 }
 static INLINE void P(addrmode5_addr_8_LDC2L_OFFSET)(tdis_ctx ctx, struct bitslice addr) {
     data(rs(addr, 9, 4));
diff --git a/lib/dis-thumb2.inc.h b/lib/dis-thumb2.inc.h
index fa744bc..a9d7f9d 100644
--- a/lib/dis-thumb2.inc.h
+++ b/lib/dis-thumb2.inc.h
@@ -100,10 +100,10 @@ static INLINE void P(rGPR_Rt_t2addrmode_imm8_pre_addr_S_2_t2STRB_PRE)(tdis_ctx c
     data(r(Rt), rs(addr, 9, 4));
 }
 static INLINE void P(rGPR_Rt_t2addrmode_imm8s4_addr_S_1_t2STRDi8)(tdis_ctx ctx, struct bitslice addr, struct bitslice Rt) {
-    data(r(Rt), rs(addr, 9, 4));
+    data_flags(IS_LDRD_STRD, r(Rt), rs(addr, 9, 4));
 }
 static INLINE void P(rGPR_Rt_t2addrmode_imm8s4_pre_addr_S_1_t2STRD_PRE)(tdis_ctx ctx, struct bitslice addr, struct bitslice Rt) {
-    data(r(Rt), rs(addr, 9, 4));
+    data_flags(IS_LDRD_STRD, r(Rt), rs(addr, 9, 4));
 }
 static INLINE void P(rGPR_Rt_t2addrmode_negimm8_addr_S_2_t2STRBi8)(tdis_ctx ctx, struct bitslice addr, struct bitslice Rt) {
     data(r(Rt), rs(addr, 9, 4));
@@ -129,11 +129,11 @@ static INLINE void P(t2addrmode_imm8_pre_addr_unk_Rt_5_t2LDRB_PRE)(tdis_ctx ctx,
 static INLINE void P(addr_offset_none_Rn_t2am_imm8_offset_offset_unk_Rt_5_t2LDRB_POST)(tdis_ctx ctx, UNUSED struct bitslice offset, struct bitslice Rt, struct bitslice Rn) {
     data(rout(Rt), r(Rn));
 }
-static INLINE void P(t2addrmode_imm8s4_addr_unk_Rt_1_t2LDRDi8)(tdis_ctx ctx, struct bitslice addr, UNUSED struct bitslice Rt) {
-    data(rs(addr, 9, 4));
+static INLINE void P(t2addrmode_imm8s4_addr_unk_Rt_1_t2LDRDi8)(tdis_ctx ctx, struct bitslice addr, struct bitslice Rt) {
+    data_flags(IS_LDRD_STRD, rout(Rt), rs(addr, 9, 4));
 }
-static INLINE void P(t2addrmode_imm8s4_pre_addr_unk_Rt_1_t2LDRD_PRE)(tdis_ctx ctx, struct bitslice addr, UNUSED struct bitslice Rt) {
-    data(rs(addr, 9, 4));
+static INLINE void P(t2addrmode_imm8s4_pre_addr_unk_Rt_1_t2LDRD_PRE)(tdis_ctx ctx, struct bitslice addr, struct bitslice Rt) {
+    data_flags(IS_LDRD_STRD, rout(Rt), rs(addr, 9, 4));
 }
 static INLINE void P(t2addrmode_negimm8_addr_unk_Rt_5_t2LDRBi8)(tdis_ctx ctx, struct bitslice addr, struct bitslice Rt) {
     data(rout(Rt), rs(addr, 9, 4));
diff --git a/lib/dis.h b/lib/dis.h
index e0cbc4f..7a53f27 100644
--- a/lib/dis.h
+++ b/lib/dis.h
@@ -81,9 +81,11 @@ static const unsigned null_op = -0x100;
 #define rout(nn)        nn,                 true,  true
 #define rsout(nn, l, s) bs_slice(nn, l, s), true,  true
 #define rnull           null_bs,             false, false
-#define data(...) data_(__VA_ARGS__, rnull, rnull, rnull, rnull)
+
+#define data(...) data_flags(0, __VA_ARGS__)
+#define data_flags(...) data_(__VA_ARGS__, rnull, rnull, rnull, rnull)
 #define data_(...) data__(__VA_ARGS__)
-#define data__(b1, o1, v1, b2, o2, v2, b3, o3, v3, b4, o4, v4, ...) do { \
+#define data__(fl, b1, o1, v1, b2, o2, v2, b3, o3, v3, b4, o4, v4, ...) do { \
     P(data)(ctx, \
         v1 ? bs_get(b1, ctx->op) : null_op, \
         v2 ? bs_get(b2, ctx->op) : null_op, \
@@ -92,7 +94,8 @@ static const unsigned null_op = -0x100;
         (o1 << 0) | \
         (o2 << 1) | \
         (o3 << 2) | \
-        (o4 << 3)); \
+        (o4 << 3) | \
+        fl); \
     if (TDIS_CTX_MODIFY(ctx)) { \
         unsigned new = ctx->op; \
         new = bs_set(b1, TDIS_CTX_NEWVAL(ctx, 0), new); \
@@ -116,6 +119,7 @@ static const unsigned null_op = -0x100;
     #define MIN_INSN_SIZE 2
     #define TARGET_DIS_HEADER "dis-arm-multi.inc.h"
     struct arch_dis_ctx { unsigned thumb_it_length; };
+    enum { IS_LDRD_STRD = 1 << 16 };
 #elif defined(TARGET_arm64)
     #define MIN_INSN_SIZE 4
     #define TARGET_DIS_HEADER "dis-arm64.inc.h"
diff --git a/lib/transform-dis-arm-multi.inc.h b/lib/transform-dis-arm-multi.inc.h
index dc066ff..662b501 100644
--- a/lib/transform-dis-arm-multi.inc.h
+++ b/lib/transform-dis-arm-multi.inc.h
@@ -113,9 +113,11 @@ static NOINLINE UNUSED void transform_dis_data(struct transform_dis_ctx *ctx,
     for (int i = 0; i < 4; i++) {
         if (out_mask & 1 << i)
             out_reg = newval[i];
-        else
+        else if (newval[i] != null_op)
             in_regs |= 1 << newval[i];
     }
+    if (out_mask & IS_LDRD_STRD)
+        in_regs |= 1 << (newval[0] + 1);
     uint32_t pc = ctx->pc + (ctx->pc_low_bit ? 4 : 8);
     int scratch = __builtin_ctz(~(in_regs | (1 << out_reg)));
 
@@ -149,7 +151,7 @@ static NOINLINE UNUSED void transform_dis_data(struct transform_dis_ctx *ctx,
             /* case 4 */
             PUSHone(ctx, scratch);
             MOVW_MOVT(ctx, scratch, pc);
-            for (int i = 1; i < 4; i++)
+            for (int i = 0; i < 4; i++)
                 if (newval[i] == 15)
                     newval[i] = scratch;
             ctx->write_newop_here = *rpp; *rpp += ctx->op_size;
@@ -157,6 +159,10 @@ static NOINLINE UNUSED void transform_dis_data(struct transform_dis_ctx *ctx,
         }
     }
     ctx->modify = true;
+#ifdef TRANSFORM_DIS_VERBOSE
+    printf("transform_dis_data: => %x %x %x %x\n",
+           newval[0], newval[1], newval[2], newval[3]);
+#endif
 }
 
 static NOINLINE UNUSED void transform_dis_pcrel(struct transform_dis_ctx *ctx,
diff --git a/test/insns-arm.S b/test/insns-arm.S
index b5a2cce..2470c35 100644
--- a/test/insns-arm.S
+++ b/test/insns-arm.S
@@ -39,7 +39,7 @@ ldrd  r0, r1, [pc]
 str   r1, [pc]
 strd  r0, r1, [pc]
 
-push {r0-r3, pc}
+/* push {r0-r3, pc} */
 push {r0-r3, lr}
 /* pop {r0-r3, pc} */