gallium: assorted additions and fixes to Cell SPE rtasm code

Fix incorrect opcode for fsmbi.
Added "macro" functions for loading floats/ints, register complement, zero, move.
Added #defines for return address and stack pointer registers.
Added assertions to check that the instruction buffer doesn't overflow.

diff --git a/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.c b/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.c
index fe5beba45689fadc0bac1d828274fe6cb7d02fe8..61010e4333934bb97850da5589568160d2b3299b 100644 (file)
--- a/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.c
+++ b/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.c
@@ -151,8 +151,8 @@ static void emit_RR(struct spe_function *p, unsigned op, unsigned rT,
     inst.inst.rB = rB;
     inst.inst.rA = rA;
     inst.inst.rT = rT;
-    *p->csr = inst.bits;
-    p->csr++;
+    p->store[p->num_inst++] = inst.bits;
+    assert(p->num_inst <= p->max_inst);
 }
 
 
@@ -165,8 +165,8 @@ 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->csr = inst.bits;
-    p->csr++;
+    p->store[p->num_inst++] = inst.bits;
+    assert(p->num_inst <= p->max_inst);
 }
 
 
@@ -178,8 +178,8 @@ static void emit_RI7(struct spe_function *p, unsigned op, unsigned rT,
     inst.inst.i7 = imm;
     inst.inst.rA = rA;
     inst.inst.rT = rT;
-    *p->csr = inst.bits;
-    p->csr++;
+    p->store[p->num_inst++] = inst.bits;
+    assert(p->num_inst <= p->max_inst);
 }
 
 
@@ -192,8 +192,8 @@ static void emit_RI8(struct spe_function *p, unsigned op, unsigned rT,
     inst.inst.i8 = imm;
     inst.inst.rA = rA;
     inst.inst.rT = rT;
-    *p->csr = inst.bits;
-    p->csr++;
+    p->store[p->num_inst++] = inst.bits;
+    assert(p->num_inst <= p->max_inst);
 }
 
 
@@ -206,8 +206,8 @@ static void emit_RI10(struct spe_function *p, unsigned op, unsigned rT,
     inst.inst.i10 = imm;
     inst.inst.rA = rA;
     inst.inst.rT = rT;
-    *p->csr = inst.bits;
-    p->csr++;
+    p->store[p->num_inst++] = inst.bits;
+    assert(p->num_inst <= p->max_inst);
 }
 
 
@@ -218,8 +218,8 @@ static void emit_RI16(struct spe_function *p, unsigned op, unsigned rT,
     inst.inst.op = op;
     inst.inst.i16 = imm;
     inst.inst.rT = rT;
-    *p->csr = inst.bits;
-    p->csr++;
+    p->store[p->num_inst++] = inst.bits;
+    assert(p->num_inst <= p->max_inst);
 }
 
 
@@ -230,8 +230,8 @@ static void emit_RI18(struct spe_function *p, unsigned op, unsigned rT,
     inst.inst.op = op;
     inst.inst.i18 = imm;
     inst.inst.rT = rT;
-    *p->csr = inst.bits;
-    p->csr++;
+    p->store[p->num_inst++] = inst.bits;
+    assert(p->num_inst <= p->max_inst);
 }
 
 
@@ -307,8 +307,9 @@ void _name (struct spe_function *p, int imm) \
 void spe_init_func(struct spe_function *p, unsigned code_size)
 {
     p->store = align_malloc(code_size, 16);
-    p->csr = p->store;
-    
+    p->num_inst = 0;
+    p->max_inst = code_size / SPE_INST_SIZE;
+
     /* Conservatively treat R0 - R2 and R80 - R127 as non-volatile.
      */
     p->regs[0] = ~7;
@@ -318,11 +319,11 @@ void spe_init_func(struct spe_function *p, unsigned code_size)
 
 void spe_release_func(struct spe_function *p)
 {
+    assert(p->num_inst <= p->max_inst);
     if (p->store != NULL) {
         align_free(p->store);
     }
     p->store = NULL;
-    p->csr = NULL;
 }
 
 
@@ -337,6 +338,7 @@ int spe_allocate_available_register(struct spe_function *p)
       const uint64_t mask = (1ULL << (i % 64));
       const unsigned idx = i / 64;
 
+      assert(idx < 2);
       if ((p->regs[idx] & mask) != 0) {
          p->regs[idx] &= ~mask;
          return i;
@@ -371,6 +373,8 @@ void spe_release_register(struct spe_function *p, int reg)
    const unsigned idx = reg / 64;
    const unsigned bit = reg % 64;
 
+   assert(idx < 2);
+
    assert(reg < SPE_NUM_REGS);
    assert((p->regs[idx] & (1ULL << bit)) == 0);
 
@@ -458,4 +462,54 @@ EMIT_R   (spe_mfspr, 0x00c);
 EMIT_R   (spe_mtspr, 0x10c);
 #endif
 
+
+/**
+ ** Helper / "macro" instructions.
+ ** Use somewhat verbose names as a reminder that these aren't native
+ ** SPE instructions.
+ **/
+
+
+void
+spe_load_float(struct spe_function *p, unsigned rT, float x)
+{
+   union {
+      float f;
+      unsigned u;
+   } bits;
+   bits.f = x;
+   spe_ilhu(p, rT, bits.u >> 16);
+   spe_iohl(p, rT, bits.u & 0xffff);
+}
+
+
+void
+spe_load_int(struct spe_function *p, unsigned rT, int i)
+{
+   spe_ilhu(p, rT, i >> 16);
+   spe_iohl(p, rT, i & 0xffff);
+}
+
+
+void
+spe_complement(struct spe_function *p, unsigned rT)
+{
+   spe_nor(p, rT, rT, rT);
+}
+
+
+void
+spe_move(struct spe_function *p, unsigned rT, unsigned rA)
+{
+   spe_ori(p, rT, rA, 0);
+}
+
+
+void
+spe_zero(struct spe_function *p, unsigned rT)
+{
+   spe_xor(p, rT, rT, rT);
+}
+
+
 #endif /* GALLIUM_CELL */

diff --git a/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.h b/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.h
index 7dd754ba773db8144c91d2af31e5d8e60c6ae829..dee8c55c4a99dff228a135ebc1cc5155318fd89d 100644 (file)
--- a/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.h
+++ b/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.h
@@ -25,6 +25,7 @@
 /**
  * \file
  * Real-time assembly generation interface for Cell B.E. SPEs.
+ * For details, see /opt/cell/sdk/docs/arch/SPU_ISA_v1.2_27Jan2007_pub.pdf
  *
  * \author Ian Romanick <idr@us.ibm.com>
  */
@@ -38,11 +39,18 @@
 /** number of general-purpose SIMD registers */
 #define SPE_NUM_REGS  128
 
+/** Return Address register */
+#define SPE_REG_RA  0
+
+/** Stack Pointer register */
+#define SPE_REG_SP  1
+
+
 struct spe_function
 {
-    uint32_t *store;  /**< instruction buffer */
-    uint32_t *csr;    /**< next free pos in instruction buffer */
-    const char *fn;   /**< unused */
+   uint32_t *store;  /**< instruction buffer */
+   uint num_inst;
+   uint max_inst;
 
     /**
      * Mask of used / unused registers
@@ -123,7 +131,8 @@ EMIT_RI16(spe_ilhu,  0x082);
 EMIT_RI16(spe_il,    0x081);
 EMIT_RI18(spe_ila,   0x021);
 EMIT_RI16(spe_iohl,  0x0c1);
-EMIT_RI16(spe_fsmbi, 0x0c5);
+EMIT_RI16(spe_fsmbi, 0x065);
+
 
 
 /* Integer and logical instructions
@@ -275,6 +284,27 @@ extern void spe_bihnz(struct spe_function *p, unsigned rT, unsigned rA,
     int d, int e);
 
 
+/** Load/splat immediate float into rT. */
+extern void
+spe_load_float(struct spe_function *p, unsigned rT, float x);
+
+/** Load/splat immediate int into rT. */
+extern void
+spe_load_int(struct spe_function *p, unsigned rT, int i);
+
+/** Complement/invert all bits in rT. */
+extern void
+spe_complement(struct spe_function *p, unsigned rT);
+
+/** rT = rA. */
+extern void
+spe_move(struct spe_function *p, unsigned rT, unsigned rA);
+
+/** rT = {0,0,0,0}. */
+extern void
+spe_zero(struct spe_function *p, unsigned rT);
+
+
 /* Floating-point instructions
  */
 EMIT_RR  (spe_fa,         0x2c4);