fix some ARM stuff including calls, and test it

wow this code sucks

11 files changed, 82 insertions, 29 deletions

diff --git a/configure b/configure
index 37b54ef..480d027 100755
--- a/configure
+++ b/configure
@@ -156,8 +156,13 @@ if settings.enable_tests:
         ('transform-dis-cases-arm64.o', 'transform-dis-cases-arm64.S', [], machs[3]),
         ('transform-dis-cases-i386.o', 'transform-dis-cases-i386.S', [], machs[1]),
         ('transform-dis-cases-x86_64.o', 'transform-dis-cases-x86_64.S', [], machs[0]),
+        ('transform-dis-cases-arm.o', 'transform-dis-cases-arm.S', [], machs[2]),
+        ('transform-dis-cases-thumb.o', 'transform-dis-cases-arm.S', ['-DTHUMB'], machs[2]),
     ]:
-        mconfig.build_c_objs(emitter, mach, settings.specialize(override_obj_fn='(out)/'+ofile), ['(src)/test/'+sfile])
+        mconfig.build_c_objs(emitter, mach, settings.specialize(
+            override_obj_fn='(out)/'+ofile,
+            override_cflags=cflags+settings.host.cflags
+        ), ['(src)/test/'+sfile])
         o_to_bin('(out)/'+ofile)
 
     tests = [
diff --git a/lib/arm/arch-transform-dis.inc.h b/lib/arm/arch-transform-dis.inc.h
index 4c44913..d74a06f 100644
--- a/lib/arm/arch-transform-dis.inc.h
+++ b/lib/arm/arch-transform-dis.inc.h
@@ -12,7 +12,8 @@ static struct assemble_ctx tdctx_to_actx(const struct transform_dis_ctx *ctx) {
     }
     return (struct assemble_ctx) {
         ctx->rewritten_ptr_ptr,
-        (uint_tptr) (uintptr_t) ctx->rewritten_ptr_ptr,
+        *ctx->rewritten_ptr_ptr,
+        (uint_tptr) (uintptr_t) *ctx->rewritten_ptr_ptr,
         ctx->arch.pc_low_bit,
         cond
     };
@@ -20,8 +21,7 @@ static struct assemble_ctx tdctx_to_actx(const struct transform_dis_ctx *ctx) {
 }
 
 static int invert_arm_cond(int cc) {
-    if (cc >= 0xe)
-        __builtin_abort();
+    substitute_assert(cc < 0xe);
     return cc ^ 1;
 }
 
@@ -152,9 +152,10 @@ void transform_dis_pcrel(struct transform_dis_ctx *ctx, uint_tptr dpc,
 static NOINLINE UNUSED
 void transform_dis_branch(struct transform_dis_ctx *ctx, uint_tptr dpc, int cc) {
 #ifdef TRANSFORM_DIS_VERBOSE
-    printf("transform_dis (0x%llx): branch => 0x%llx\n",
+    printf("transform_dis (0x%llx): branch => 0x%llx cc=%x\n",
            (unsigned long long) ctx->base.pc,
-           (unsigned long long) dpc);
+           (unsigned long long) dpc,
+           cc);
 #endif
     /* The check in transform_dis_branch_top is correct under the simplifying
      * assumption here that functions will not try to branch into the middle of
@@ -164,18 +165,23 @@ void transform_dis_branch(struct transform_dis_ctx *ctx, uint_tptr dpc, int cc)
     transform_dis_branch_top(ctx, dpc, cc);
     struct assemble_ctx actx = tdctx_to_actx(ctx);
     ctx->write_newop_here = NULL;
+    int replacement_size = 8 + (actx.thumb ? 2 : 4);
     if ((cc & CC_ARMCC) == CC_ARMCC) {
+        replacement_size += actx.thumb ? 2 : 4;
         actx.cond = invert_arm_cond(cc & 0xf);
-        Bccrel(actx, 2+8);
+        Bccrel(actx, replacement_size);
     } else if ((cc & CC_CBXZ) == CC_CBXZ) {
+        replacement_size += 2;
         ctx->base.modify = true;
-        ctx->base.newval[0] = ctx->base.pc + 2+8;
+        ctx->base.newval[0] = actx.pc_of_code_base + replacement_size;
         ctx->base.newval[1] = 1; /* do invert */
         void **codep = ctx->rewritten_ptr_ptr;
         ctx->write_newop_here = *codep; *codep += 2;
     }
     actx.cond = 0xe;
-    LDR_PC(actx, dpc | ctx->arch.pc_low_bit);
+    MOVW_MOVT(actx, 14, dpc | ctx->arch.pc_low_bit);
+    BLXr(actx, 14);
+    substitute_assert(*actx.codep - actx.code_base == replacement_size);
 }
 
 static void transform_dis_pre_dis(struct transform_dis_ctx *ctx) {
@@ -193,6 +199,7 @@ static void transform_dis_pre_dis(struct transform_dis_ctx *ctx) {
 static void transform_dis_post_dis(struct transform_dis_ctx *ctx) {
     if (ctx->arch.bccrel_p) {
         struct assemble_ctx actx = {&ctx->arch.bccrel_p,
+                                    ctx->arch.bccrel_p,
                                     (uint_tptr) (uintptr_t) ctx->arch.bccrel_p,
                                     /*thumb*/ true,
                                     ctx->arch.bccrel_bits};
diff --git a/lib/arm/assemble.h b/lib/arm/assemble.h
index 2303a7a..1d4250b 100644
--- a/lib/arm/assemble.h
+++ b/lib/arm/assemble.h
@@ -3,11 +3,16 @@
 
 struct assemble_ctx {
     void **codep;
-    uint_tptr pc;
+    void *code_base;
+    uint_tptr pc_of_code_base;
     bool thumb;
     int cond;
 };
 
+static inline uint_tptr actx_pc(struct assemble_ctx ctx) {
+    return ctx.pc_of_code_base + (*ctx.codep - ctx.code_base);
+}
+
 static inline void PUSHone(struct assemble_ctx ctx, int Rt) {
     if (ctx.thumb)
         op32(ctx.codep, 0x0d04f84d | Rt << 28);
@@ -63,7 +68,7 @@ static inline void LDRxi(struct assemble_ctx ctx, int Rt, int Rn, uint32_t off,
             case PLM_U16: subop = 1; sign = 0; break;
             case PLM_S16: subop = 1; sign = 1; break;
             case PLM_U32: subop = 2; sign = 0; break;
-            default: __builtin_abort();
+            default: substitute_assert(false);
         }
         op32(ctx.codep, 0x0000f890 | Rn | Rt << 28 | subop << 5 | sign << 8 |
                         off << 16);
@@ -85,11 +90,18 @@ static inline void LDRxi(struct assemble_ctx ctx, int Rt, int Rn, uint32_t off,
                                 ctx.cond << 28);
                 break;
             default:
-                __builtin_abort();
+                substitute_assert(false);
         }
     }
 }
 
+static inline void BLXr(struct assemble_ctx ctx, int Rm) {
+    if (ctx.thumb)
+        op16(ctx.codep, 0x4780 | Rm << 3);
+    else
+        op32(ctx.codep, 0xe12fff30 | Rm | ctx.cond << 28);
+}
+
 static inline void Bccrel(struct assemble_ctx ctx, int offset) {
     if (ctx.thumb) {
         offset = (offset - 4) / 2;
@@ -101,7 +113,7 @@ static inline void Bccrel(struct assemble_ctx ctx, int offset) {
 }
 
 static inline void LDR_PC(struct assemble_ctx ctx, uint32_t dpc) {
-    if (ctx.pc & 2)
+    if (actx_pc(ctx) & 2)
         op16(ctx.codep, 0xbf00);
     if (ctx.thumb)
         op32(ctx.codep, 0xf000f8df);
diff --git a/lib/arm/dis-arm.inc.h b/lib/arm/dis-arm.inc.h
index 18285cd..021227c 100644
--- a/lib/arm/dis-arm.inc.h
+++ b/lib/arm/dis-arm.inc.h
@@ -164,7 +164,7 @@ static INLINE void P(adrlabel_label_unk_Rd_1_ADR)(tdis_ctx ctx, struct bitslice
 }
 static INLINE void P(br_target_target_pred_p_B_1_Bcc)(tdis_ctx ctx, struct bitslice target, struct bitslice p) {
     unsigned p_val = bs_get(p, ctx->base.op);
-    return P(branch)(ctx, ctx->base.pc + 8 + sext(bs_get(target, ctx->base.op), 24),
+    return P(branch)(ctx, ctx->base.pc + 8 + 4 * sext(bs_get(target, ctx->base.op), 24),
                      p_val == 0xe ? 0 : (CC_ARMCC | p_val));
 }
 static INLINE void P(ldst_so_reg_addr_unk_Rt_2_LDRB_PRE_REG)(tdis_ctx ctx, struct bitslice addr, struct bitslice Rt) {
@@ -210,8 +210,9 @@ static INLINE void P(GPR_func_3_BLX)(tdis_ctx ctx, UNUSED struct bitslice func)
     return P(indirect_call)(ctx);
 }
 static INLINE void P(bl_target_func_2_BL)(tdis_ctx ctx, struct bitslice func) {
-    return P(branch)(ctx, ctx->base.pc + 8 + sext(bs_get(func, ctx->base.op), 24),
-                     CC_CALL);
+    unsigned p_val = ctx->base.op >> 28; // XXX fix this to actually be an op
+    return P(branch)(ctx, ctx->base.pc + 8 + 4 * sext(bs_get(func, ctx->base.op), 24),
+                     CC_CALL | (p_val == 0xe ? 0 : (CC_ARMCC | p_val)));
 }
 
 static INLINE void P(dis_arm)(tdis_ctx ctx) {
diff --git a/lib/substitute-internal.h b/lib/substitute-internal.h
index c9b78f0..2d26e66 100644
--- a/lib/substitute-internal.h
+++ b/lib/substitute-internal.h
@@ -109,3 +109,7 @@ static const char *xbasename(const char *path) {
     return slash ? slash + 1 : path;
 }
 
+#define substitute_assert(x) do { \
+    if (!(x)) { __builtin_abort(); } \
+} while(0)
+
diff --git a/script/test-transform-dis.sh b/script/test-transform-dis.sh
index 3a9adf3..0b45fd2 100755
--- a/script/test-transform-dis.sh
+++ b/script/test-transform-dis.sh
@@ -1,6 +1,12 @@
 #!/bin/sh
 set -xe
-make -j8 out/transform-dis-cases-$1.bin out/test-transform-dis-$1
-out/test-transform-dis-$1 auto < out/transform-dis-cases-$1.bin
+barch="$1"
+is_thumb=0
+if [ "$1" = "thumb" ]; then
+    barch=arm
+    is_thumb=1
+fi
+make -j8 out/transform-dis-cases-$1.bin out/test-transform-dis-"$barch"
+out/test-transform-dis-"$barch" auto "$is_thumb" < out/transform-dis-cases-$1.bin
 
 
diff --git a/test/test-transform-dis.c b/test/test-transform-dis.c
index 98c98e3..d147a49 100644
--- a/test/test-transform-dis.c
+++ b/test/test-transform-dis.c
@@ -127,7 +127,7 @@ static void do_auto(uint8_t *in, size_t in_size, struct arch_dis_ctx arch) {
             pc_trampoline,
             &arch,
             offsets,
-            TRANSFORM_DIS_BAN_CALLS);
+            0);//TRANSFORM_DIS_BAN_CALLS);
         if (ret) {
             if (expect_err) {
                 printf("OK\n");
diff --git a/test/transform-dis-cases-arm.S b/test/transform-dis-cases-arm.S
new file mode 100644
index 0000000..a9e01b1
--- /dev/null
+++ b/test/transform-dis-cases-arm.S
@@ -0,0 +1,15 @@
+#include "transform-dis-cases.h"
+#ifdef THUMB
+.thumb
+#endif
+
+
+#ifndef THUMB
+GIVEN blne 0f; nop; nop; 0:
+EXPECT beq 1f; movw lr, #0x000c; movt lr, #0xdead; blx lr; 1: nop; nop
+GIVEN bl 0f; 0:
+EXPECT movw lr, #0x0004; movt lr, #0xdead; blx lr
+#else
+GIVEN bl 0f; 0:
+EXPECT movw lr, #0x0005; movt lr, #0xdead; blx lr
+#endif
diff --git a/test/transform-dis-cases-arm64.S b/test/transform-dis-cases-arm64.S
index cc3bd4c..99a353b 100644
--- a/test/transform-dis-cases-arm64.S
+++ b/test/transform-dis-cases-arm64.S
@@ -1,6 +1,4 @@
-#define GIVEN .ascii "GIVEN";
-#define EXPECT .ascii "EXPECT";
-#define EXPECT_ERR .ascii "EXPECT_ERR";
+#include "transform-dis-cases.h"
 
 /* yay clang, no semicolons allowed */
 
@@ -12,7 +10,10 @@ EXPECT
 GIVEN
     blr x5
     nop
-EXPECT_ERR
+//EXPECT_ERR (with ban_calls)
+EXPECT
+    blr x5
+
 
 GIVEN
     cbnz x8, .+0x100
diff --git a/test/transform-dis-cases-i386.S b/test/transform-dis-cases-i386.S
index c02a044..a39ee02 100644
--- a/test/transform-dis-cases-i386.S
+++ b/test/transform-dis-cases-i386.S
@@ -1,6 +1,4 @@
-#define GIVEN .ascii "GIVEN";
-#define EXPECT .ascii "EXPECT";
-#define EXPECT_ERR .ascii "EXPECT_ERR";
+#include "transform-dis-cases.h"
 
 GIVEN call 0f; 0: pop %edx
 /* XXX the extra push isn't necessary in 32-bit mode */
@@ -10,9 +8,9 @@ GIVEN jmp 0f; 0: nop
 EXPECT_ERR
 
 GIVEN jne .+0x1000
-/* we expect to generate an unnecessarily long jump, so hardcode it the 0x10000
- * is because we pretend our trampoline is 0x10000 bytes before the original
- * function */
+/* we expect to generate an unnecessarily long jump, so hardcode it
+ * the 0x10000 is because we pretend our trampoline is 0x10000 bytes before the
+ * original function */
 EXPECT 2: .byte 0x0f, 0x85; .long 2; jmp 1f; 0: jmp .+0x10000+0x1000-8; 1:
 
 GIVEN loopne .+0x80
diff --git a/test/transform-dis-cases.h b/test/transform-dis-cases.h
new file mode 100644
index 0000000..94f39bd
--- /dev/null
+++ b/test/transform-dis-cases.h
@@ -0,0 +1,4 @@
+#pragma once
+#define GIVEN .ascii "GIVEN";
+#define EXPECT .ascii "EXPECT";
+#define EXPECT_ERR .ascii "EXPECT_ERR";