cell: implement many more PPC instructions for code gen

3 files changed, 704 insertions, 41 deletions

diff --git a/src/gallium/auxiliary/rtasm/Makefile b/src/gallium/auxiliary/rtasm/Makefile
index 39b8a4dbd7a..252dc5274ab 100644
--- a/src/gallium/auxiliary/rtasm/Makefile
+++ b/src/gallium/auxiliary/rtasm/Makefile
@@ -7,6 +7,7 @@ C_SOURCES = \
 	rtasm_cpu.c \
 	rtasm_execmem.c \
 	rtasm_x86sse.c \
+	rtasm_ppc.c \
 	rtasm_ppc_spe.c
 
 include ../../Makefile.template
diff --git a/src/gallium/auxiliary/rtasm/rtasm_ppc.c b/src/gallium/auxiliary/rtasm/rtasm_ppc.c
index 534a23568d5..4a94ed0460d 100644
--- a/src/gallium/auxiliary/rtasm/rtasm_ppc.c
+++ b/src/gallium/auxiliary/rtasm/rtasm_ppc.c
@@ -23,10 +23,19 @@
 
 /**
  * PPC code generation.
+ * For reference, see http://www.power.org/resources/reading/PowerISA_V2.05.pdf
+ * ABI info: http://www.cs.utsa.edu/~whaley/teach/cs6463FHPO/LEC/lec12_ho.pdf
+ *
+ * Other PPC refs:
+ * http://www-01.ibm.com/chips/techlib/techlib.nsf/techdocs/852569B20050FF778525699600719DF2
+ * http://www.ibm.com/developerworks/eserver/library/es-archguide-v2.html
+ * http://www.freescale.com/files/product/doc/MPCFPE32B.pdf
+ *
  * \author Brian Paul
  */
 
 
+#include <stdio.h>
 #include "util/u_memory.h"
 #include "pipe/p_debug.h"
 #include "rtasm_ppc.h"
@@ -35,30 +44,125 @@
 void
 ppc_init_func(struct ppc_function *p, unsigned max_inst)
 {
-    p->store = align_malloc(max_inst * PPC_INST_SIZE, 16);
-    p->num_inst = 0;
-    p->max_inst = max_inst;
-    p->vec_used = ~0;
+   uint i;
+
+   p->store = align_malloc(max_inst * PPC_INST_SIZE, 16);
+   p->num_inst = 0;
+   p->max_inst = max_inst;
+   p->fp_used = ~0x0;
+   p->vec_used = ~0x0;
+
+   /* only allow using gp registers 7..12 for now */
+   p->reg_used = 0x0;
+   for (i = 7; i < 13; i++)
+      p->reg_used |= (1 << i);
 }
 
 
 void
 ppc_release_func(struct ppc_function *p)
 {
-    assert(p->num_inst <= p->max_inst);
-    if (p->store != NULL) {
-        align_free(p->store);
-    }
-    p->store = NULL;
+   assert(p->num_inst <= p->max_inst);
+   if (p->store != NULL) {
+      align_free(p->store);
+   }
+   p->store = NULL;
+}
+
+
+void (*ppc_get_func(struct ppc_function *p))(void)
+{
+#if 0
+   DUMP_END();
+   if (DISASSEM && p->store)
+      debug_printf("disassemble %p %p\n", p->store, p->csr);
+
+   if (p->store == p->error_overflow)
+      return (void (*)(void)) NULL;
+   else
+#endif
+      return (void (*)(void)) p->store;
+}
+
+
+void
+ppc_dump_func(const struct ppc_function *p)
+{
+   uint i;
+   for (i = 0; i < p->num_inst; i++) {
+      debug_printf("%3u: 0x%08x\n", i, p->store[i]);
+   }
+}
+
+
+/**
+ * Allocate a general purpose register.
+ * \return register index or -1 if none left.
+ */
+int
+ppc_allocate_register(struct ppc_function *p)
+{
+   unsigned i;
+   for (i = 0; i < PPC_NUM_REGS; i++) {
+      const uint64_t mask = 1 << i;
+      if ((p->reg_used & mask) != 0) {
+         p->reg_used &= ~mask;
+         return i;
+      }
+   }
+   return -1;
 }
 
 
 /**
- * Alloate a vector register.
+ * Mark the given general purpose register as "unallocated".
+ */
+void
+ppc_release_register(struct ppc_function *p, int reg)
+{
+   assert(reg < PPC_NUM_REGS);
+   assert((p->reg_used & (1 << reg)) == 0);
+   p->reg_used |= (1 << reg);
+}
+
+
+/**
+ * Allocate a floating point register.
  * \return register index or -1 if none left.
  */
 int
-ppc_allocate_vec_register(struct ppc_function *p, int reg)
+ppc_allocate_fp_register(struct ppc_function *p)
+{
+   unsigned i;
+   for (i = 0; i < PPC_NUM_FP_REGS; i++) {
+      const uint64_t mask = 1 << i;
+      if ((p->fp_used & mask) != 0) {
+         p->fp_used &= ~mask;
+         return i;
+      }
+   }
+   return -1;
+}
+
+
+/**
+ * Mark the given floating point register as "unallocated".
+ */
+void
+ppc_release_fp_register(struct ppc_function *p, int reg)
+{
+   assert(reg < PPC_NUM_FP_REGS);
+   assert((p->fp_used & (1 << reg)) == 0);
+   p->fp_used |= (1 << reg);
+}
+
+
+/**
+ * Allocate a vector register.
+ * \return register index or -1 if none left.
+ */
+int
+ppc_allocate_vec_register(struct ppc_function *p)
 {
    unsigned i;
    for (i = 0; i < PPC_NUM_VEC_REGS; i++) {
@@ -68,7 +172,6 @@ ppc_allocate_vec_register(struct ppc_function *p, int reg)
          return i;
       }
    }
-
    return -1;
 }
 
@@ -81,7 +184,6 @@ ppc_release_vec_register(struct ppc_function *p, int reg)
 {
    assert(reg < PPC_NUM_VEC_REGS);
    assert((p->vec_used & (1 << reg)) == 0);
-
    p->vec_used |= (1 << reg);
 }
 
@@ -98,6 +200,20 @@ union vx_inst {
    } inst;
 };
 
+static inline void
+emit_vx(struct ppc_function *p, uint op2, uint vD, uint vA, uint vB)
+{
+   union vx_inst inst;
+   inst.inst.op = 4;
+   inst.inst.vD = vD;
+   inst.inst.vA = vA;
+   inst.inst.vB = vB;
+   inst.inst.op2 = op2;
+   p->store[p->num_inst++] = inst.bits;
+   assert(p->num_inst <= p->max_inst);
+};
+
+
 union vxr_inst {
    uint32_t bits;
    struct {
@@ -110,6 +226,21 @@ union vxr_inst {
    } inst;
 };
 
+static inline void
+emit_vxr(struct ppc_function *p, uint op2, uint vD, uint vA, uint vB)
+{
+   union vxr_inst inst;
+   inst.inst.op = 4;
+   inst.inst.vD = vD;
+   inst.inst.vA = vA;
+   inst.inst.vB = vB;
+   inst.inst.rC = 0;
+   inst.inst.op2 = op2;
+   p->store[p->num_inst++] = inst.bits;
+   assert(p->num_inst <= p->max_inst);
+};
+
+
 union va_inst {
    uint32_t bits;
    struct {
@@ -122,49 +253,204 @@ union va_inst {
    } inst;
 };
 
-
 static inline void
-emit_vx(struct ppc_function *p, uint op2, uint vD, uint vA, uint vB)
+emit_va(struct ppc_function *p, uint op2, uint vD, uint vA, uint vB, uint vC)
 {
-   union vx_inst inst;
+   union va_inst inst;
    inst.inst.op = 4;
    inst.inst.vD = vD;
    inst.inst.vA = vA;
    inst.inst.vB = vB;
+   inst.inst.vC = vC;
    inst.inst.op2 = op2;
    p->store[p->num_inst++] = inst.bits;
    assert(p->num_inst <= p->max_inst);
 };
 
-static inline void
-emit_vxr(struct ppc_function *p, uint op2, uint vD, uint vA, uint vB)
+
+union i_inst {
+   uint32_t bits;
+   struct {
+      unsigned op:6;
+      unsigned li:24;
+      unsigned aa:1;
+      unsigned lk:1;
+   } inst;
+};
+
+static INLINE void
+emit_i(struct ppc_function *p, uint op, uint li, uint aa, uint lk)
 {
-   union vxr_inst inst;
-   inst.inst.op = 4;
-   inst.inst.vD = vD;
-   inst.inst.vA = vA;
-   inst.inst.vB = vB;
-   inst.inst.rC = 0;
+   union i_inst inst;
+   inst.inst.op = op;
+   inst.inst.li = li;
+   inst.inst.aa = aa;
+   inst.inst.lk = lk;
+   p->store[p->num_inst++] = inst.bits;
+   assert(p->num_inst <= p->max_inst);
+}
+
+
+union xl_inst {
+   uint32_t bits;
+   struct {
+      unsigned op:6;
+      unsigned bo:5;
+      unsigned bi:5;
+      unsigned unused:3;
+      unsigned bh:2;
+      unsigned op2:10;
+      unsigned lk:1;
+   } inst;
+};
+
+static INLINE void
+emit_xl(struct ppc_function *p, uint op, uint bo, uint bi, uint bh,
+        uint op2, uint lk)
+{
+   union xl_inst inst;
+   inst.inst.op = op;
+   inst.inst.bo = bo;
+   inst.inst.bi = bi;
+   inst.inst.unused = 0x0;
+   inst.inst.bh = bh;
    inst.inst.op2 = op2;
+   inst.inst.lk = lk;
    p->store[p->num_inst++] = inst.bits;
    assert(p->num_inst <= p->max_inst);
+}
+
+static INLINE void
+dump_xl(const char *name, uint inst)
+{
+   union xl_inst i;
+
+   i.bits = inst;
+   debug_printf("%s = 0x%08x\n", name, inst);
+   debug_printf(" op: %d 0x%x\n", i.inst.op, i.inst.op);
+   debug_printf(" bo: %d 0x%x\n", i.inst.bo, i.inst.bo);
+   debug_printf(" bi: %d 0x%x\n", i.inst.bi, i.inst.bi);
+   debug_printf(" unused: %d 0x%x\n", i.inst.unused, i.inst.unused);
+   debug_printf(" bh: %d 0x%x\n", i.inst.bh, i.inst.bh);
+   debug_printf(" op2: %d 0x%x\n", i.inst.op2, i.inst.op2);
+   debug_printf(" lk: %d 0x%x\n", i.inst.lk, i.inst.lk);
+}
+
+
+union x_inst {
+   uint32_t bits;
+   struct {
+      unsigned op:6;
+      unsigned vrs:5;
+      unsigned ra:5;
+      unsigned rb:5;
+      unsigned op2:10;
+      unsigned unused:1;
+   } inst;
 };
 
-static inline void
-emit_va(struct ppc_function *p, uint op2, uint vD, uint vA, uint vB, uint vC)
+static INLINE void
+emit_x(struct ppc_function *p, uint op, uint vrs, uint ra, uint rb, uint op2)
 {
-   union va_inst inst;
-   inst.inst.op = 4;
-   inst.inst.vD = vD;
-   inst.inst.vA = vA;
-   inst.inst.vB = vB;
-   inst.inst.vC = vC;
+   union x_inst inst;
+   inst.inst.op = op;
+   inst.inst.vrs = vrs;
+   inst.inst.ra = ra;
+   inst.inst.rb = rb;
    inst.inst.op2 = op2;
+   inst.inst.unused = 0x0;
+   p->store[p->num_inst++] = inst.bits;
+   assert(p->num_inst <= p->max_inst);
+}
+
+
+union d_inst {
+   uint32_t bits;
+   struct {
+      unsigned op:6;
+      unsigned rt:5;
+      unsigned ra:5;
+      unsigned si:16;
+   } inst;
+};
+
+static inline void
+emit_d(struct ppc_function *p, uint op, uint rt, uint ra, int si)
+{
+   union d_inst inst;
+   assert(si >= -32768);
+   assert(si <= 32767);
+   inst.inst.op = op;
+   inst.inst.rt = rt;
+   inst.inst.ra = ra;
+   inst.inst.si = (unsigned) (si & 0xffff);
    p->store[p->num_inst++] = inst.bits;
    assert(p->num_inst <= p->max_inst);
 };
 
 
+union a_inst {
+   uint32_t bits;
+   struct {
+      unsigned op:6;
+      unsigned frt:5;
+      unsigned fra:5;
+      unsigned frb:5;
+      unsigned unused:5;
+      unsigned op2:5;
+      unsigned rc:1;
+   } inst;
+};
+
+static inline void
+emit_a(struct ppc_function *p, uint op, uint frt, uint fra, uint frb, uint op2,
+       uint rc)
+{
+   union a_inst inst;
+   inst.inst.op = op;
+   inst.inst.frt = frt;
+   inst.inst.fra = fra;
+   inst.inst.frb = frb;
+   inst.inst.unused = 0x0;
+   inst.inst.op2 = op2;
+   inst.inst.rc = rc;
+   p->store[p->num_inst++] = inst.bits;
+   assert(p->num_inst <= p->max_inst);
+};
+
+
+union xo_inst {
+   uint32_t bits;
+   struct {
+      unsigned op:6;
+      unsigned rt:5;
+      unsigned ra:5;
+      unsigned rb:5;
+      unsigned oe:1;
+      unsigned op2:9;
+      unsigned rc:1;
+   } inst;
+};
+
+static INLINE void
+emit_xo(struct ppc_function *p, uint op, uint rt, uint ra, uint rb, uint oe,
+        uint op2, uint rc)
+{
+   union xo_inst inst;
+   inst.inst.op = op;
+   inst.inst.rt = rt;
+   inst.inst.ra = ra;
+   inst.inst.rb = rb;
+   inst.inst.oe = oe;
+   inst.inst.op2 = op2;
+   inst.inst.rc = rc;
+   p->store[p->num_inst++] = inst.bits;
+   assert(p->num_inst <= p->max_inst);
+}
+
+
+
+
 
 /**
  ** float vector arithmetic
@@ -172,7 +458,7 @@ emit_va(struct ppc_function *p, uint op2, uint vD, uint vA, uint vB, uint vC)
 
 /** vector float add */
 void
-ppc_vaddfp(struct ppc_function *p,uint vD, uint vA, uint vB)
+ppc_vaddfp(struct ppc_function *p, uint vD, uint vA, uint vB)
 {
    emit_vx(p, 10, vD, vA, vB);
 }
@@ -198,11 +484,11 @@ ppc_vmaxfp(struct ppc_function *p, uint vD, uint vA, uint vB)
    emit_vx(p, 1034, vD, vA, vB);
 }
 
-/** vector float mult add */
+/** vector float mult add: vD = vA * vB + vC */
 void
 ppc_vmaddfp(struct ppc_function *p, uint vD, uint vA, uint vB, uint vC)
 {
-   emit_va(p, 46, vD, vA, vB, vC);
+   emit_va(p, 46, vD, vA, vC, vB); /* note arg order */
 }
 
 /** vector float compare greater than */
@@ -282,13 +568,26 @@ ppc_vrfiz(struct ppc_function *p, uint vD, uint vB)
    emit_vx(p, 586, vD, 0, vB);
 }
 
+/** vector store: store vR at mem[vA+vB] */
+void
+ppc_stvx(struct ppc_function *p, uint vR, uint vA, uint vB)
+{
+   emit_x(p, 31, vR, vA, vB, 231);
+}
+
+/** vector load: vR = mem[vA+vB] */
+void
+ppc_lvx(struct ppc_function *p, uint vR, uint vA, uint vB)
+{
+   emit_x(p, 31, vR, vA, vB, 103);
+}
+
 
 
 /**
- ** bitwise operations
+ ** vector bitwise operations
  **/
 
-
 /** vector and */
 void
 ppc_vand(struct ppc_function *p, uint vD, uint vA, uint vB)
@@ -324,6 +623,14 @@ ppc_vxor(struct ppc_function *p, uint vD, uint vA, uint vB)
    emit_vx(p, 1220, vD, vA, vB);
 }
 
+/** Pseudo-instruction: vector move */
+void
+ppc_vecmove(struct ppc_function *p, uint vD, uint vA)
+{
+   ppc_vor(p, vD, vA, vA);
+}
+
+
 
 /**
  ** Vector shuffle / select / splat / etc
@@ -363,3 +670,225 @@ ppc_vspltw(struct ppc_function *p, uint vD, uint vB, uint imm)
 {
    emit_vx(p, 652, vD, imm, vB);
 }
+
+/** vector splat signed immediate word */
+void
+ppc_vspltisw(struct ppc_function *p, uint vD, int imm)
+{
+   assert(imm >= -16);
+   assert(imm < 15);
+   emit_vx(p, 908, vD, imm, 0);
+}
+
+/** vector shift left word: vD[word] = vA[word] << (vB[word] & 0x1f) */
+void
+ppc_vslw(struct ppc_function *p, uint vD, uint vA, uint vB)
+{
+   emit_vx(p, 388, vD, vA, vB);
+}
+
+
+
+
+/**
+ ** integer arithmetic
+ **/
+
+/** rt = ra + imm */
+void
+ppc_addi(struct ppc_function *p, uint rt, uint ra, int imm)
+{
+   emit_d(p, 14, rt, ra, imm);
+}
+
+/** rt = ra + (imm << 16) */
+void
+ppc_addis(struct ppc_function *p, uint rt, uint ra, int imm)
+{
+   emit_d(p, 15, rt, ra, imm);
+}
+
+/** rt = ra + rb */
+void
+ppc_add(struct ppc_function *p, uint rt, uint ra, uint rb)
+{
+   emit_xo(p, 31, rt, ra, rb, 0, 266, 0);
+}
+
+/** rt = ra AND ra */
+void
+ppc_and(struct ppc_function *p, uint rt, uint ra, uint rb)
+{
+   emit_x(p, 31, ra, rt, rb, 28);  /* note argument order */
+}
+
+/** rt = ra AND imm */
+void
+ppc_andi(struct ppc_function *p, uint rt, uint ra, int imm)
+{
+   emit_d(p, 28, ra, rt, imm);  /* note argument order */
+}
+
+/** rt = ra OR ra */
+void
+ppc_or(struct ppc_function *p, uint rt, uint ra, uint rb)
+{
+   emit_x(p, 31, ra, rt, rb, 444);  /* note argument order */
+}
+
+/** rt = ra OR imm */
+void
+ppc_ori(struct ppc_function *p, uint rt, uint ra, int imm)
+{
+   emit_d(p, 24, ra, rt, imm);  /* note argument order */
+}
+
+/** rt = ra XOR ra */
+void
+ppc_xor(struct ppc_function *p, uint rt, uint ra, uint rb)
+{
+   emit_x(p, 31, ra, rt, rb, 316);  /* note argument order */
+}
+
+/** rt = ra XOR imm */
+void
+ppc_xori(struct ppc_function *p, uint rt, uint ra, int imm)
+{
+   emit_d(p, 26, ra, rt, imm);  /* note argument order */
+}
+
+/** pseudo instruction: move: rt = ra */
+void
+ppc_mr(struct ppc_function *p, uint rt, uint ra)
+{
+   ppc_or(p, rt, ra, ra);
+}
+
+/** pseudo instruction: load immediate: rt = imm */
+void
+ppc_li(struct ppc_function *p, uint rt, int imm)
+{
+   ppc_addi(p, rt, 0, imm);
+}
+
+/** rt = imm << 16 */
+void
+ppc_lis(struct ppc_function *p, uint rt, int imm)
+{
+   ppc_addis(p, rt, 0, imm);
+}
+
+/** rt = imm */
+void
+ppc_load_int(struct ppc_function *p, uint rt, int imm)
+{
+   ppc_lis(p, rt, (imm >> 16));          /* rt = imm >> 16 */
+   ppc_ori(p, rt, rt, (imm & 0xffff));   /* rt = rt | (imm & 0xffff) */
+}
+
+
+
+
+/**
+ ** integer load/store
+ **/
+
+/** store rs at memory[(ra)+d],
+ * then update ra = (ra)+d
+ */
+void
+ppc_stwu(struct ppc_function *p, uint rs, uint ra, int d)
+{
+   emit_d(p, 37, rs, ra, d);
+}
+
+/** store rs at memory[(ra)+d] */
+void
+ppc_stw(struct ppc_function *p, uint rs, uint ra, int d)
+{
+   emit_d(p, 36, rs, ra, d);
+}
+
+/** Load rt = mem[(ra)+d];  then zero set high 32 bits to zero. */
+void
+ppc_lwz(struct ppc_function *p, uint rt, uint ra, int d)
+{
+   emit_d(p, 32, rt, ra, d);
+}
+
+
+
+/**
+ ** Float (non-vector) arithmetic
+ **/
+
+/** add: frt = fra + frb */
+void
+ppc_fadd(struct ppc_function *p, uint frt, uint fra, uint frb)
+{
+   emit_a(p, 63, frt, fra, frb, 21, 0);
+}
+
+/** sub: frt = fra - frb */
+void
+ppc_fsub(struct ppc_function *p, uint frt, uint fra, uint frb)
+{
+   emit_a(p, 63, frt, fra, frb, 20, 0);
+}
+
+/** convert to int: rt = (int) ra */
+void
+ppc_fctiwz(struct ppc_function *p, uint rt, uint fra)
+{
+   emit_x(p, 63, rt, 0, fra, 15);
+}
+
+/** store frs at mem[(ra)+offset] */
+void
+ppc_stfs(struct ppc_function *p, uint frs, uint ra, int offset)
+{
+   emit_d(p, 52, frs, ra, offset);
+}
+
+/** store frs at mem[(ra)+(rb)] */
+void
+ppc_stfiwx(struct ppc_function *p, uint frs, uint ra, uint rb)
+{
+   emit_x(p, 31, frs, ra, rb, 983);
+}
+
+/** load frt = mem[(ra)+offset] */
+void
+ppc_lfs(struct ppc_function *p, uint frt, uint ra, int offset)
+{
+   emit_d(p, 48, frt, ra, offset);
+}
+
+
+
+
+
+/**
+ ** branch instructions
+ **/
+
+/** BLR: Branch to link register (p. 35) */
+void
+ppc_blr(struct ppc_function *p)
+{
+   emit_i(p, 18, 0, 0, 1);
+}
+
+/** Branch Conditional to Link Register (p. 36) */
+void
+ppc_bclr(struct ppc_function *p, uint condOp, uint branchHint, uint condReg)
+{
+   emit_xl(p, 19, condOp, condReg, branchHint, 16, 0);
+}
+
+/** Pseudo instruction: return from subroutine */
+void
+ppc_return(struct ppc_function *p)
+{
+   ppc_bclr(p, BRANCH_COND_ALWAYS, BRANCH_HINT_SUB_RETURN, 0);
+}
diff --git a/src/gallium/auxiliary/rtasm/rtasm_ppc.h b/src/gallium/auxiliary/rtasm/rtasm_ppc.h
index ed14e943df6..6370b604944 100644
--- a/src/gallium/auxiliary/rtasm/rtasm_ppc.h
+++ b/src/gallium/auxiliary/rtasm/rtasm_ppc.h
@@ -36,27 +36,46 @@
 
 #define PPC_INST_SIZE 4  /**< 4 bytes / instruction */
 
+#define PPC_NUM_REGS 32
+#define PPC_NUM_FP_REGS 32
 #define PPC_NUM_VEC_REGS 32
 
+/** Stack pointer register */
+#define PPC_REG_SP 1
+
+/** Branch conditions */
+#define BRANCH_COND_ALWAYS       0x14  /* binary 1z1zz (z=ignored) */
+
+/** Branch hints */
+#define BRANCH_HINT_SUB_RETURN   0x0   /* binary 00 */
+
 
 struct ppc_function
 {
    uint32_t *store;  /**< instruction buffer */
    uint num_inst;
    uint max_inst;
-   uint32_t vec_used;   /** used/free vector registers bitmask */
    uint32_t reg_used;   /** used/free general-purpose registers bitmask */
+   uint32_t fp_used;   /** used/free floating point registers bitmask */
+   uint32_t vec_used;   /** used/free vector registers bitmask */
 };
 
 
 
 extern void ppc_init_func(struct ppc_function *p, unsigned max_inst);
 extern void ppc_release_func(struct ppc_function *p);
-
-extern int ppc_allocate_vec_register(struct ppc_function *p, int reg);
+extern void (*ppc_get_func( struct ppc_function *p ))( void );
+extern void ppc_dump_func(const struct ppc_function *p);
+
+extern int ppc_allocate_register(struct ppc_function *p);
+extern void ppc_release_register(struct ppc_function *p, int reg);
+extern int ppc_allocate_fp_register(struct ppc_function *p);
+extern void ppc_release_fp_register(struct ppc_function *p, int reg);
+extern int ppc_allocate_vec_register(struct ppc_function *p);
 extern void ppc_release_vec_register(struct ppc_function *p, int reg);
 
 
+
 /**
  ** float vector arithmetic
  **/
@@ -126,9 +145,18 @@ extern void
 ppc_vrfiz(struct ppc_function *p, uint vD, uint vB);
 
 
+/** vector store: store vR at mem[vA+vB] */
+extern void
+ppc_stvx(struct ppc_function *p, uint vR, uint vA, uint vB);
+
+/** vector load: vR = mem[vA+vB] */
+extern void
+ppc_lvx(struct ppc_function *p, uint vR, uint vA, uint vB);
+
+
 
 /**
- ** bitwise operations
+ ** vector bitwise operations
  **/
 
 
@@ -152,6 +180,10 @@ ppc_vnor(struct ppc_function *p, uint vD, uint vA, uint vB);
 extern void
 ppc_vxor(struct ppc_function *p, uint vD, uint vA, uint vB);
 
+/** Pseudo-instruction: vector move */
+extern void
+ppc_vecmove(struct ppc_function *p, uint vD, uint vA);
+
 
 /**
  ** Vector shuffle / select / splat / etc
@@ -177,5 +209,106 @@ ppc_vsplthw(struct ppc_function *p, uint vD, uint vB, uint imm);
 extern void
 ppc_vspltw(struct ppc_function *p, uint vD, uint vB, uint imm);
 
+/** vector splat signed immediate word */
+extern void
+ppc_vspltisw(struct ppc_function *p, uint vD, int imm);
+
+/** vector shift left word: vD[word] = vA[word] << (vB[word] & 0x1f) */
+extern void
+ppc_vslw(struct ppc_function *p, uint vD, uint vA, uint vB);
+
+
+
+/**
+ ** scalar arithmetic
+ **/
+
+extern void
+ppc_add(struct ppc_function *p, uint rt, uint ra, uint rb);
+
+extern void
+ppc_addi(struct ppc_function *p, uint rt, uint ra, int imm);
+
+extern void
+ppc_and(struct ppc_function *p, uint rt, uint ra, uint rb);
+
+extern void
+ppc_andi(struct ppc_function *p, uint rt, uint ra, int imm);
+
+extern void
+ppc_or(struct ppc_function *p, uint rt, uint ra, uint rb);
+
+extern void
+ppc_ori(struct ppc_function *p, uint rt, uint ra, int imm);
+
+extern void
+ppc_xor(struct ppc_function *p, uint rt, uint ra, uint rb);
+
+extern void
+ppc_xori(struct ppc_function *p, uint rt, uint ra, int imm);
+
+extern void
+ppc_mr(struct ppc_function *p, uint rt, uint ra);
+
+extern void
+ppc_li(struct ppc_function *p, uint rt, int imm);
+
+extern void
+ppc_lis(struct ppc_function *p, uint rt, int imm);
+
+extern void
+ppc_load_int(struct ppc_function *p, uint rt, int imm);
+
+
+
+/**
+ ** scalar load/store
+ **/
+
+extern void
+ppc_stwu(struct ppc_function *p, uint rs, uint ra, int d);
+
+extern void
+ppc_stw(struct ppc_function *p, uint rs, uint ra, int d);
+
+extern void
+ppc_lwz(struct ppc_function *p, uint rs, uint ra, int d);
+
+
+
+/**
+ ** Float (non-vector) arithmetic
+ **/
+
+extern void
+ppc_fadd(struct ppc_function *p, uint frt, uint fra, uint frb);
+
+extern void
+ppc_fsub(struct ppc_function *p, uint frt, uint fra, uint frb);
+
+extern void
+ppc_fctiwz(struct ppc_function *p, uint rt, uint ra);
+
+extern void
+ppc_stfs(struct ppc_function *p, uint frs, uint ra, int offset);
+
+extern void
+ppc_stfiwx(struct ppc_function *p, uint frs, uint ra, uint rb);
+
+
+
+/**
+ ** branch instructions
+ **/
+
+extern void
+ppc_blr(struct ppc_function *p);
+
+void
+ppc_bclr(struct ppc_function *p, uint condOp, uint branchHint, uint condReg);
+
+extern void
+ppc_return(struct ppc_function *p);
+
 
 #endif /* RTASM_PPC_H */