X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fgallium%2Fauxiliary%2Frtasm%2Frtasm_ppc_spe.c;h=53a0e722cffe8aa5fd407713409473680212f3d9;hb=5e8240320ac39a3e8984054bc300743725312741;hp=a6dd7ef31180715ba9d7ee575d0e2ecc94082953;hpb=73d00b9e93a9e8a5fecb0de224552741e389fc11;p=mesa.git

diff --git a/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.c b/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.c
index a6dd7ef3118..53a0e722cff 100644
--- a/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.c
+++ b/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.c
@@ -174,24 +174,54 @@ reg_name(int reg)
       return "$lr";
    default:
       {
-         static char buf[10];
-         sprintf(buf, "$%d", reg);
-         return buf;
+         /* cycle through four buffers to handle multiple calls per printf */
+         static char buf[4][10];
+         static int b = 0;
+         b = (b + 1) % 4;
+         sprintf(buf[b], "$%d", reg);
+         return buf[b];
       }
    }
 }
 
 
-static void emit_RR(struct spe_function *p, unsigned op, unsigned rT,
-		    unsigned rA, unsigned rB, const char *name)
+static void
+emit_instruction(struct spe_function *p, uint32_t inst_bits)
+{
+   if (!p->store)
+      return;  /* out of memory, drop the instruction */
+
+   if (p->num_inst == p->max_inst) {
+      /* allocate larger buffer */
+      uint32_t *newbuf;
+      p->max_inst *= 2;  /* 2x larger */
+      newbuf = align_malloc(p->max_inst * SPE_INST_SIZE, 16);
+      if (newbuf) {
+         memcpy(newbuf, p->store, p->num_inst * SPE_INST_SIZE);
+      }
+      align_free(p->store);
+      p->store = newbuf;
+      if (!p->store) {
+         /* out of memory */
+         p->num_inst = 0;
+         return;
+      }
+   }
+
+   p->store[p->num_inst++] = inst_bits;
+}
+
+
+
+static void emit_RR(struct spe_function *p, unsigned op, int rT,
+		    int rA, int rB, const char *name)
 {
     union spe_inst_RR inst;
     inst.inst.op = op;
     inst.inst.rB = rB;
     inst.inst.rA = rA;
     inst.inst.rT = rT;
-    p->store[p->num_inst++] = inst.bits;
-    assert(p->num_inst <= p->max_inst);
+    emit_instruction(p, inst.bits);
     if (p->print) {
        indent(p);
        printf("%s\t%s, %s, %s\n",
@@ -200,8 +230,8 @@ static void emit_RR(struct spe_function *p, unsigned op, unsigned rT,
 }
 
 
-static void emit_RRR(struct spe_function *p, unsigned op, unsigned rT,
-                     unsigned rA, unsigned rB, unsigned rC, const char *name)
+static void emit_RRR(struct spe_function *p, unsigned op, int rT,
+                     int rA, int rB, int rC, const char *name)
 {
     union spe_inst_RRR inst;
     inst.inst.op = op;
@@ -209,8 +239,7 @@ static void emit_RRR(struct spe_function *p, unsigned op, unsigned rT,
     inst.inst.rB = rB;
     inst.inst.rA = rA;
     inst.inst.rC = rC;
-    p->store[p->num_inst++] = inst.bits;
-    assert(p->num_inst <= p->max_inst);
+    emit_instruction(p, inst.bits);
     if (p->print) {
        indent(p);
        printf("%s\t%s, %s, %s, %s\n", rem_prefix(name), reg_name(rT),
@@ -219,16 +248,15 @@ static void emit_RRR(struct spe_function *p, unsigned op, unsigned rT,
 }
 
 
-static void emit_RI7(struct spe_function *p, unsigned op, unsigned rT,
-		     unsigned rA, int imm, const char *name)
+static void emit_RI7(struct spe_function *p, unsigned op, int rT,
+		     int rA, int imm, const char *name)
 {
     union spe_inst_RI7 inst;
     inst.inst.op = op;
     inst.inst.i7 = imm;
     inst.inst.rA = rA;
     inst.inst.rT = rT;
-    p->store[p->num_inst++] = inst.bits;
-    assert(p->num_inst <= p->max_inst);
+    emit_instruction(p, inst.bits);
     if (p->print) {
        indent(p);
        printf("%s\t%s, %s, 0x%x\n",
@@ -238,16 +266,15 @@ static void emit_RI7(struct spe_function *p, unsigned op, unsigned rT,
 
 
 
-static void emit_RI8(struct spe_function *p, unsigned op, unsigned rT,
-		     unsigned rA, int imm, const char *name)
+static void emit_RI8(struct spe_function *p, unsigned op, int rT,
+		     int rA, int imm, const char *name)
 {
     union spe_inst_RI8 inst;
     inst.inst.op = op;
     inst.inst.i8 = imm;
     inst.inst.rA = rA;
     inst.inst.rT = rT;
-    p->store[p->num_inst++] = inst.bits;
-    assert(p->num_inst <= p->max_inst);
+    emit_instruction(p, inst.bits);
     if (p->print) {
        indent(p);
        printf("%s\t%s, %s, 0x%x\n",
@@ -257,40 +284,41 @@ static void emit_RI8(struct spe_function *p, unsigned op, unsigned rT,
 
 
 
-static void emit_RI10(struct spe_function *p, unsigned op, unsigned rT,
-		      unsigned rA, int imm, const char *name)
+static void emit_RI10(struct spe_function *p, unsigned op, int rT,
+		      int rA, int imm, const char *name)
 {
     union spe_inst_RI10 inst;
     inst.inst.op = op;
     inst.inst.i10 = imm;
     inst.inst.rA = rA;
     inst.inst.rT = rT;
-    p->store[p->num_inst++] = inst.bits;
-    assert(p->num_inst <= p->max_inst);
+    emit_instruction(p, inst.bits);
     if (p->print) {
        indent(p);
-       if (strcmp(name, "spe_lqd") == 0 ||
-           strcmp(name, "spe_stqd") == 0) {
-          printf("%s\t%s, %d(%s)\n",
-                 rem_prefix(name), reg_name(rT), imm, reg_name(rA));
-       }
-       else {
-          printf("%s\t%s, %s, 0x%x\n",
-                 rem_prefix(name), reg_name(rT), reg_name(rA), imm);
-       }
+       printf("%s\t%s, %s, 0x%x\n",
+              rem_prefix(name), reg_name(rT), reg_name(rA), imm);
     }
 }
 
 
-static void emit_RI16(struct spe_function *p, unsigned op, unsigned rT,
+/** As above, but do range checking on signed immediate value */
+static void emit_RI10s(struct spe_function *p, unsigned op, int rT,
+                       int rA, int imm, const char *name)
+{
+    assert(imm <= 511);
+    assert(imm >= -512);
+    emit_RI10(p, op, rT, rA, imm, name);
+}
+
+
+static void emit_RI16(struct spe_function *p, unsigned op, int rT,
 		      int imm, const char *name)
 {
     union spe_inst_RI16 inst;
     inst.inst.op = op;
     inst.inst.i16 = imm;
     inst.inst.rT = rT;
-    p->store[p->num_inst++] = inst.bits;
-    assert(p->num_inst <= p->max_inst);
+    emit_instruction(p, inst.bits);
     if (p->print) {
        indent(p);
        printf("%s\t%s, 0x%x\n", rem_prefix(name), reg_name(rT), imm);
@@ -298,15 +326,14 @@ static void emit_RI16(struct spe_function *p, unsigned op, unsigned rT,
 }
 
 
-static void emit_RI18(struct spe_function *p, unsigned op, unsigned rT,
+static void emit_RI18(struct spe_function *p, unsigned op, int rT,
 		      int imm, const char *name)
 {
     union spe_inst_RI18 inst;
     inst.inst.op = op;
     inst.inst.i18 = imm;
     inst.inst.rT = rT;
-    p->store[p->num_inst++] = inst.bits;
-    assert(p->num_inst <= p->max_inst);
+    emit_instruction(p, inst.bits);
     if (p->print) {
        indent(p);
        printf("%s\t%s, 0x%x\n", rem_prefix(name), reg_name(rT), imm);
@@ -314,58 +341,68 @@ static void emit_RI18(struct spe_function *p, unsigned op, unsigned rT,
 }
 
 
-
+#define EMIT(_name, _op) \
+void _name (struct spe_function *p) \
+{ \
+   emit_RR(p, _op, 0, 0, 0, __FUNCTION__); \
+}
 
 #define EMIT_(_name, _op) \
-void _name (struct spe_function *p, unsigned rT) \
+void _name (struct spe_function *p, int rT) \
 { \
    emit_RR(p, _op, rT, 0, 0, __FUNCTION__); \
 }
 
 #define EMIT_R(_name, _op) \
-void _name (struct spe_function *p, unsigned rT, unsigned rA) \
+void _name (struct spe_function *p, int rT, int rA) \
 { \
    emit_RR(p, _op, rT, rA, 0, __FUNCTION__);                 \
 }
 
 #define EMIT_RR(_name, _op) \
-void _name (struct spe_function *p, unsigned rT, unsigned rA, unsigned rB) \
+void _name (struct spe_function *p, int rT, int rA, int rB) \
 { \
    emit_RR(p, _op, rT, rA, rB, __FUNCTION__);                \
 }
 
 #define EMIT_RRR(_name, _op) \
-void _name (struct spe_function *p, unsigned rT, unsigned rA, unsigned rB, unsigned rC) \
+void _name (struct spe_function *p, int rT, int rA, int rB, int rC) \
 { \
    emit_RRR(p, _op, rT, rA, rB, rC, __FUNCTION__);           \
 }
 
 #define EMIT_RI7(_name, _op) \
-void _name (struct spe_function *p, unsigned rT, unsigned rA, int imm) \
+void _name (struct spe_function *p, int rT, int rA, int imm) \
 { \
    emit_RI7(p, _op, rT, rA, imm, __FUNCTION__);              \
 }
 
 #define EMIT_RI8(_name, _op, bias) \
-void _name (struct spe_function *p, unsigned rT, unsigned rA, int imm) \
+void _name (struct spe_function *p, int rT, int rA, int imm) \
 { \
    emit_RI8(p, _op, rT, rA, bias - imm, __FUNCTION__);       \
 }
 
 #define EMIT_RI10(_name, _op) \
-void _name (struct spe_function *p, unsigned rT, unsigned rA, int imm) \
+void _name (struct spe_function *p, int rT, int rA, int imm) \
 { \
    emit_RI10(p, _op, rT, rA, imm, __FUNCTION__);             \
 }
 
+#define EMIT_RI10s(_name, _op) \
+void _name (struct spe_function *p, int rT, int rA, int imm) \
+{ \
+   emit_RI10s(p, _op, rT, rA, imm, __FUNCTION__);             \
+}
+
 #define EMIT_RI16(_name, _op) \
-void _name (struct spe_function *p, unsigned rT, int imm) \
+void _name (struct spe_function *p, int rT, int imm) \
 { \
    emit_RI16(p, _op, rT, imm, __FUNCTION__);                 \
 }
 
 #define EMIT_RI18(_name, _op) \
-void _name (struct spe_function *p, unsigned rT, int imm) \
+void _name (struct spe_function *p, int rT, int imm) \
 { \
    emit_RI18(p, _op, rT, imm, __FUNCTION__);                 \
 }
@@ -379,17 +416,22 @@ void _name (struct spe_function *p, int imm) \
 #include "rtasm_ppc_spe.h"
 
 
+
 /**
  * Initialize an spe_function.
- * \param code_size  size of instruction buffer to allocate, in bytes.
+ * \param code_size  initial size of instruction buffer to allocate, in bytes.
+ *                   If zero, use a default.
  */
 void spe_init_func(struct spe_function *p, unsigned code_size)
 {
-    register unsigned int i;
+    uint i;
+
+    if (!code_size)
+       code_size = 64;
 
-    p->store = align_malloc(code_size, 16);
     p->num_inst = 0;
     p->max_inst = code_size / SPE_INST_SIZE;
+    p->store = align_malloc(code_size, 16);
 
     p->set_count = 0;
     memset(p->regs, 0, SPE_NUM_REGS * sizeof(p->regs[0]));
@@ -401,7 +443,7 @@ void spe_init_func(struct spe_function *p, unsigned code_size)
       p->regs[i] = 1;
     }
 
-    p->print = false;
+    p->print = FALSE;
     p->indent = 0;
 }
 
@@ -461,6 +503,7 @@ int spe_allocate_register(struct spe_function *p, int reg)
  */
 void spe_release_register(struct spe_function *p, int reg)
 {
+   assert(reg >= 0);
    assert(reg < SPE_NUM_REGS);
    assert(p->regs[reg] == 1);
 
@@ -475,7 +518,7 @@ void spe_release_register(struct spe_function *p, int reg)
  */
 void spe_allocate_register_set(struct spe_function *p)
 {
-   register unsigned int i;
+   uint i;
 
    /* Keep track of the set count.  If it ever wraps around to 0, 
     * we're in trouble.
@@ -489,13 +532,14 @@ void spe_allocate_register_set(struct spe_function *p)
     * when the register set is released.
     */
    for (i = 0; i < SPE_NUM_REGS; i++) {
-      if (p->regs[i] > 0) p->regs[i]++;
+      if (p->regs[i] > 0)
+         p->regs[i]++;
    }
 }
 
 void spe_release_register_set(struct spe_function *p)
 {
-   unsigned int i;
+   uint i;
 
    /* If the set count drops below zero, we're in trouble. */
    assert(p->set_count > 0);
@@ -506,8 +550,23 @@ void spe_release_register_set(struct spe_function *p)
     * available.
     */
    for (i = 0; i < SPE_NUM_REGS; i++) {
-      if (p->regs[i] > 0) p->regs[i]--;
+      if (p->regs[i] > 0)
+         p->regs[i]--;
+   }
+}
+
+
+unsigned
+spe_get_registers_used(const struct spe_function *p, ubyte used[])
+{
+   unsigned i, num = 0;
+   /* only count registers in the range available to callers */
+   for (i = 2; i < 80; i++) {
+      if (p->regs[i]) {
+         used[num++] = i;
+      }
    }
+   return num;
 }
 
 
@@ -525,7 +584,7 @@ spe_indent(struct spe_function *p, int spaces)
 }
 
 
-extern void
+void
 spe_comment(struct spe_function *p, int rel_indent, const char *s)
 {
    if (p->print) {
@@ -537,6 +596,56 @@ spe_comment(struct spe_function *p, int rel_indent, const char *s)
 }
 
 
+/**
+ * Load quad word.
+ * NOTE: offset is in bytes and the least significant 4 bits must be zero!
+ */
+void spe_lqd(struct spe_function *p, int rT, int rA, int offset)
+{
+   const boolean pSave = p->print;
+
+   /* offset must be a multiple of 16 */
+   assert(offset % 16 == 0);
+   /* offset must fit in 10-bit signed int field, after shifting */
+   assert((offset >> 4) <= 511);
+   assert((offset >> 4) >= -512);
+
+   p->print = FALSE;
+   emit_RI10(p, 0x034, rT, rA, offset >> 4, "spe_lqd");
+   p->print = pSave;
+
+   if (p->print) {
+      indent(p);
+      printf("lqd\t%s, %d(%s)\n", reg_name(rT), offset, reg_name(rA));
+   }
+}
+
+
+/**
+ * Store quad word.
+ * NOTE: offset is in bytes and the least significant 4 bits must be zero!
+ */
+void spe_stqd(struct spe_function *p, int rT, int rA, int offset)
+{
+   const boolean pSave = p->print;
+
+   /* offset must be a multiple of 16 */
+   assert(offset % 16 == 0);
+   /* offset must fit in 10-bit signed int field, after shifting */
+   assert((offset >> 4) <= 511);
+   assert((offset >> 4) >= -512);
+
+   p->print = FALSE;
+   emit_RI10(p, 0x024, rT, rA, offset >> 4, "spe_stqd");
+   p->print = pSave;
+
+   if (p->print) {
+      indent(p);
+      printf("stqd\t%s, %d(%s)\n", reg_name(rT), offset, reg_name(rA));
+   }
+}
+
+
 /**
  * For branch instructions:
  * \param d  if 1, disable interupts if branch is taken
@@ -545,51 +654,51 @@ spe_comment(struct spe_function *p, int rel_indent, const char *s)
  */
 
 /** Branch Indirect to address in rA */
-void spe_bi(struct spe_function *p, unsigned rA, int d, int e)
+void spe_bi(struct spe_function *p, int rA, int d, int e)
 {
    emit_RI7(p, 0x1a8, 0, rA, (d << 5) | (e << 4), __FUNCTION__);
 }
 
 /** Interupt Return */
-void spe_iret(struct spe_function *p, unsigned rA, int d, int e)
+void spe_iret(struct spe_function *p, int rA, int d, int e)
 {
    emit_RI7(p, 0x1aa, 0, rA, (d << 5) | (e << 4), __FUNCTION__);
 }
 
 /** Branch indirect and set link on external data */
-void spe_bisled(struct spe_function *p, unsigned rT, unsigned rA, int d,
+void spe_bisled(struct spe_function *p, int rT, int rA, int d,
 		int e)
 {
    emit_RI7(p, 0x1ab, rT, rA, (d << 5) | (e << 4), __FUNCTION__);
 }
 
 /** Branch indirect and set link.  Save PC in rT, jump to rA. */
-void spe_bisl(struct spe_function *p, unsigned rT, unsigned rA, int d,
+void spe_bisl(struct spe_function *p, int rT, int rA, int d,
 		int e)
 {
    emit_RI7(p, 0x1a9, rT, rA, (d << 5) | (e << 4), __FUNCTION__);
 }
 
 /** Branch indirect if zero word.  If rT.word[0]==0, jump to rA. */
-void spe_biz(struct spe_function *p, unsigned rT, unsigned rA, int d, int e)
+void spe_biz(struct spe_function *p, int rT, int rA, int d, int e)
 {
    emit_RI7(p, 0x128, rT, rA, (d << 5) | (e << 4), __FUNCTION__);
 }
 
 /** Branch indirect if non-zero word.  If rT.word[0]!=0, jump to rA. */
-void spe_binz(struct spe_function *p, unsigned rT, unsigned rA, int d, int e)
+void spe_binz(struct spe_function *p, int rT, int rA, int d, int e)
 {
    emit_RI7(p, 0x129, rT, rA, (d << 5) | (e << 4), __FUNCTION__);
 }
 
 /** Branch indirect if zero halfword.  If rT.halfword[1]==0, jump to rA. */
-void spe_bihz(struct spe_function *p, unsigned rT, unsigned rA, int d, int e)
+void spe_bihz(struct spe_function *p, int rT, int rA, int d, int e)
 {
    emit_RI7(p, 0x12a, rT, rA, (d << 5) | (e << 4), __FUNCTION__);
 }
 
 /** Branch indirect if non-zero halfword.  If rT.halfword[1]!=0, jump to rA. */
-void spe_bihnz(struct spe_function *p, unsigned rT, unsigned rA, int d, int e)
+void spe_bihnz(struct spe_function *p, int rT, int rA, int d, int e)
 {
    emit_RI7(p, 0x12b, rT, rA, (d << 5) | (e << 4), __FUNCTION__);
 }
@@ -609,7 +718,6 @@ hbrr;
 #if 0
 stop;
 EMIT_RR  (spe_stopd, 0x140);
-EMIT_    (spe_lnop,  0x001);
 EMIT_    (spe_nop,   0x201);
 sync;
 EMIT_    (spe_dsync, 0x003);
@@ -626,7 +734,7 @@ EMIT_R   (spe_mtspr, 0x10c);
 
 
 void
-spe_load_float(struct spe_function *p, unsigned rT, float x)
+spe_load_float(struct spe_function *p, int rT, float x)
 {
    if (x == 0.0f) {
       spe_il(p, rT, 0x0);
@@ -653,7 +761,7 @@ spe_load_float(struct spe_function *p, unsigned rT, float x)
 
 
 void
-spe_load_int(struct spe_function *p, unsigned rT, int i)
+spe_load_int(struct spe_function *p, int rT, int i)
 {
    if (-32768 <= i && i <= 32767) {
       spe_il(p, rT, i);
@@ -665,7 +773,7 @@ spe_load_int(struct spe_function *p, unsigned rT, int i)
    }
 }
 
-void spe_load_uint(struct spe_function *p, unsigned rT, unsigned int ui)
+void spe_load_uint(struct spe_function *p, int rT, uint ui)
 {
    /* If the whole value is in the lower 18 bits, use ila, which
     * doesn't sign-extend.  Otherwise, if the two halfwords of
@@ -674,7 +782,7 @@ void spe_load_uint(struct spe_function *p, unsigned rT, unsigned int ui)
     * Bytes Immediate (fsmbi) to load the value in a single instruction.
     * Otherwise, in the general case, we have to use ilhu followed by iohl.
     */
-   if ((ui & 0xfffc0000) == ui) {
+   if ((ui & 0x0003ffff) == ui) {
       spe_ila(p, rT, ui);
    }
    else if ((ui >> 16) == (ui & 0xffff)) {
@@ -686,7 +794,7 @@ void spe_load_uint(struct spe_function *p, unsigned rT, unsigned int ui)
       ((ui & 0x00ff0000) == 0 || (ui & 0x00ff0000) == 0x00ff0000) &&
       ((ui & 0xff000000) == 0 || (ui & 0xff000000) == 0xff000000)
    ) {
-      unsigned int mask = 0;
+      uint mask = 0;
       /* fsmbi duplicates each bit in the given mask eight times,
        * using a 16-bit value to initialize a 16-byte quadword.
        * Each 4-bit nybble of the mask corresponds to a full word
@@ -710,10 +818,12 @@ void spe_load_uint(struct spe_function *p, unsigned rT, unsigned int ui)
    }
 }
 
-/* This function is constructed identically to spe_sor_uint() below.
+/**
+ * This function is constructed identically to spe_xor_uint() below.
  * Changes to one should be made in the other.
  */
-void spe_and_uint(struct spe_function *p, unsigned rT, unsigned rA, unsigned int ui)
+void
+spe_and_uint(struct spe_function *p, int rT, int rA, uint ui)
 {
    /* If we can, emit a single instruction, either And Byte Immediate
     * (which uses the same constant across each byte), And Halfword Immediate
@@ -723,7 +833,7 @@ void spe_and_uint(struct spe_function *p, unsigned rT, unsigned rA, unsigned int
     *
     * Otherwise, we'll need to use a temporary register.
     */
-   register unsigned int tmp;
+   uint tmp;
 
    /* If the upper 23 bits are all 0s or all 1s, sign extension
     * will work and we can use And Word Immediate
@@ -754,16 +864,19 @@ void spe_and_uint(struct spe_function *p, unsigned rT, unsigned rA, unsigned int
    }
 
    /* Otherwise, we'll have to use a temporary register. */
-   unsigned int tmp_reg = spe_allocate_available_register(p);
+   int tmp_reg = spe_allocate_available_register(p);
    spe_load_uint(p, tmp_reg, ui);
    spe_and(p, rT, rA, tmp_reg);
    spe_release_register(p, tmp_reg);
 }
 
-/* This function is constructed identically to spe_and_uint() above.
+
+/**
+ * This function is constructed identically to spe_and_uint() above.
  * Changes to one should be made in the other.
  */
-void spe_xor_uint(struct spe_function *p, unsigned rT, unsigned rA, unsigned int ui)
+void
+spe_xor_uint(struct spe_function *p, int rT, int rA, uint ui)
 {
    /* If we can, emit a single instruction, either Exclusive Or Byte 
     * Immediate (which uses the same constant across each byte), Exclusive 
@@ -773,7 +886,7 @@ void spe_xor_uint(struct spe_function *p, unsigned rT, unsigned rA, unsigned int
     *
     * Otherwise, we'll need to use a temporary register.
     */
-   register unsigned int tmp;
+   uint tmp;
 
    /* If the upper 23 bits are all 0s or all 1s, sign extension
     * will work and we can use Exclusive Or Word Immediate
@@ -804,14 +917,14 @@ void spe_xor_uint(struct spe_function *p, unsigned rT, unsigned rA, unsigned int
    }
 
    /* Otherwise, we'll have to use a temporary register. */
-   unsigned int tmp_reg = spe_allocate_available_register(p);
+   int tmp_reg = spe_allocate_available_register(p);
    spe_load_uint(p, tmp_reg, ui);
    spe_xor(p, rT, rA, tmp_reg);
    spe_release_register(p, tmp_reg);
 }
 
 void
-spe_compare_equal_uint(struct spe_function *p, unsigned rT, unsigned rA, unsigned int ui)
+spe_compare_equal_uint(struct spe_function *p, int rT, int rA, uint ui)
 {
    /* If the comparison value is 9 bits or less, it fits inside a
     * Compare Equal Word Immediate instruction.
@@ -821,7 +934,7 @@ spe_compare_equal_uint(struct spe_function *p, unsigned rT, unsigned rA, unsigne
    }
    /* Otherwise, we're going to have to load a word first. */
    else {
-      unsigned int tmp_reg = spe_allocate_available_register(p);
+      int tmp_reg = spe_allocate_available_register(p);
       spe_load_uint(p, tmp_reg, ui);
       spe_ceq(p, rT, rA, tmp_reg);
       spe_release_register(p, tmp_reg);
@@ -829,7 +942,7 @@ spe_compare_equal_uint(struct spe_function *p, unsigned rT, unsigned rA, unsigne
 }
 
 void
-spe_compare_greater_uint(struct spe_function *p, unsigned rT, unsigned rA, unsigned int ui)
+spe_compare_greater_uint(struct spe_function *p, int rT, int rA, uint ui)
 {
    /* If the comparison value is 10 bits or less, it fits inside a
     * Compare Logical Greater Than Word Immediate instruction.
@@ -839,7 +952,7 @@ spe_compare_greater_uint(struct spe_function *p, unsigned rT, unsigned rA, unsig
    }
    /* Otherwise, we're going to have to load a word first. */
    else {
-      unsigned int tmp_reg = spe_allocate_available_register(p);
+      int tmp_reg = spe_allocate_available_register(p);
       spe_load_uint(p, tmp_reg, ui);
       spe_clgt(p, rT, rA, tmp_reg);
       spe_release_register(p, tmp_reg);
@@ -847,23 +960,26 @@ spe_compare_greater_uint(struct spe_function *p, unsigned rT, unsigned rA, unsig
 }
 
 void
-spe_splat(struct spe_function *p, unsigned rT, unsigned rA)
+spe_splat(struct spe_function *p, int rT, int rA)
 {
+   /* Use a temporary, just in case rT == rA */
+   int tmp_reg = spe_allocate_available_register(p);
    /* Duplicate bytes 0, 1, 2, and 3 across the whole register */
-   spe_ila(p, rT, 0x00010203);
-   spe_shufb(p, rT, rA, rA, rT);
+   spe_ila(p, tmp_reg, 0x00010203);
+   spe_shufb(p, rT, rA, rA, tmp_reg);
+   spe_release_register(p, tmp_reg);
 }
 
 
 void
-spe_complement(struct spe_function *p, unsigned rT, unsigned rA)
+spe_complement(struct spe_function *p, int rT, int rA)
 {
    spe_nor(p, rT, rA, rA);
 }
 
 
 void
-spe_move(struct spe_function *p, unsigned rT, unsigned rA)
+spe_move(struct spe_function *p, int rT, int rA)
 {
    /* Use different instructions depending on the instruction address
     * to take advantage of the dual pipelines.
@@ -876,14 +992,14 @@ spe_move(struct spe_function *p, unsigned rT, unsigned rA)
 
 
 void
-spe_zero(struct spe_function *p, unsigned rT)
+spe_zero(struct spe_function *p, int rT)
 {
    spe_xor(p, rT, rT, rT);
 }
 
 
 void
-spe_splat_word(struct spe_function *p, unsigned rT, unsigned rA, int word)
+spe_splat_word(struct spe_function *p, int rT, int rA, int word)
 {
    assert(word >= 0);
    assert(word <= 3);
@@ -923,9 +1039,9 @@ spe_splat_word(struct spe_function *p, unsigned rT, unsigned rA, int word)
  * like "x = min(x, a)", we always allocate a new register to be safe.
  */
 void 
-spe_float_min(struct spe_function *p, unsigned rT, unsigned rA, unsigned rB)
+spe_float_min(struct spe_function *p, int rT, int rA, int rB)
 {
-   unsigned int compare_reg = spe_allocate_available_register(p);
+   int compare_reg = spe_allocate_available_register(p);
    spe_fcgt(p, compare_reg, rA, rB);
    spe_selb(p, rT, rA, rB, compare_reg);
    spe_release_register(p, compare_reg);
@@ -940,9 +1056,9 @@ spe_float_min(struct spe_function *p, unsigned rT, unsigned rA, unsigned rB)
  * so that the larger of the two is selected instead of the smaller.
  */
 void 
-spe_float_max(struct spe_function *p, unsigned rT, unsigned rA, unsigned rB)
+spe_float_max(struct spe_function *p, int rT, int rA, int rB)
 {
-   unsigned int compare_reg = spe_allocate_available_register(p);
+   int compare_reg = spe_allocate_available_register(p);
    spe_fcgt(p, compare_reg, rA, rB);
    spe_selb(p, rT, rB, rA, compare_reg);
    spe_release_register(p, compare_reg);