summaryrefslogtreecommitdiff
path: root/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.h
diff options
context:
space:
mode:
Diffstat (limited to 'src/gallium/auxiliary/rtasm/rtasm_ppc_spe.h')
-rw-r--r--src/gallium/auxiliary/rtasm/rtasm_ppc_spe.h87
1 files changed, 74 insertions, 13 deletions
diff --git a/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.h b/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.h
index d95e5aace3..47dadb343c 100644
--- a/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.h
+++ b/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.h
@@ -28,6 +28,7 @@
* For details, see /opt/cell/sdk/docs/arch/SPU_ISA_v1.2_27Jan2007_pub.pdf
*
* \author Ian Romanick <idr@us.ibm.com>
+ * \author Brian Paul
*/
#ifndef RTASM_PPC_SPE_H
@@ -39,10 +40,10 @@
/** number of general-purpose SIMD registers */
#define SPE_NUM_REGS 128
-/** Return Address register */
+/** Return Address register (aka $lr / Link Register) */
#define SPE_REG_RA 0
-/** Stack Pointer register */
+/** Stack Pointer register (aka $sp) */
#define SPE_REG_SP 1
@@ -52,25 +53,49 @@ struct spe_function
uint num_inst;
uint max_inst;
- /**
- * Mask of used / unused registers
- *
- * Each set bit corresponds to an available register. Each cleared bit
- * corresponds to an allocated register.
+ /**
+ * The "set count" reflects the number of nested register sets
+ * are allowed. In the unlikely case that we exceed the set count,
+ * register allocation will start to be confused, which is critical
+ * enough that we check for it.
+ */
+ unsigned char set_count;
+
+ /**
+ * Flags for used and unused registers. Each byte corresponds to a
+ * register; a 0 in that byte means that the register is available.
+ * A value of 1 means that the register was allocated in the current
+ * register set. Any other value N means that the register was allocated
+ * N register sets ago.
*
* \sa
* spe_allocate_register, spe_allocate_available_register,
- * spe_release_register
+ * spe_allocate_register_set, spe_release_register_set, spe_release_register,
*/
- uint64_t regs[SPE_NUM_REGS / 64];
+ unsigned char regs[SPE_NUM_REGS];
+
+ boolean print; /**< print/dump instructions as they're emitted? */
+ int indent; /**< number of spaces to indent */
};
+
extern void spe_init_func(struct spe_function *p, unsigned code_size);
extern void spe_release_func(struct spe_function *p);
+extern unsigned spe_code_size(const struct spe_function *p);
extern int spe_allocate_available_register(struct spe_function *p);
extern int spe_allocate_register(struct spe_function *p, int reg);
extern void spe_release_register(struct spe_function *p, int reg);
+extern void spe_allocate_register_set(struct spe_function *p);
+extern void spe_release_register_set(struct spe_function *p);
+
+extern unsigned
+spe_get_registers_used(const struct spe_function *p, ubyte used[]);
+
+extern void spe_print_code(struct spe_function *p, boolean enable);
+extern void spe_indent(struct spe_function *p, int spaces);
+extern void spe_comment(struct spe_function *p, int rel_indent, const char *s);
+
#endif /* RTASM_PPC_SPE_H */
@@ -106,11 +131,9 @@ extern void spe_release_register(struct spe_function *p, int reg);
/* Memory load / store instructions
*/
-EMIT_RI10(spe_lqd, 0x034);
EMIT_RR (spe_lqx, 0x1c4);
EMIT_RI16(spe_lqa, 0x061);
EMIT_RI16(spe_lqr, 0x067);
-EMIT_RI10(spe_stqd, 0x024);
EMIT_RR (spe_stqx, 0x144);
EMIT_RI16(spe_stqa, 0x041);
EMIT_RI16(spe_stqr, 0x047);
@@ -268,6 +291,12 @@ EMIT_RI16(spe_brz, 0x040);
EMIT_RI16(spe_brhnz, 0x046);
EMIT_RI16(spe_brhz, 0x044);
+extern void
+spe_lqd(struct spe_function *p, unsigned rT, unsigned rA, int offset);
+
+extern void
+spe_stqd(struct spe_function *p, unsigned rT, unsigned rA, int offset);
+
extern void spe_bi(struct spe_function *p, unsigned rA, int d, int e);
extern void spe_iret(struct spe_function *p, unsigned rA, int d, int e);
extern void spe_bisled(struct spe_function *p, unsigned rT, unsigned rA,
@@ -292,13 +321,33 @@ spe_load_float(struct spe_function *p, unsigned rT, float x);
extern void
spe_load_int(struct spe_function *p, unsigned rT, int i);
+/** Load/splat immediate unsigned int into rT. */
+extern void
+spe_load_uint(struct spe_function *p, unsigned rT, unsigned int ui);
+
+/** And immediate value into rT. */
+extern void
+spe_and_uint(struct spe_function *p, unsigned rT, unsigned rA, unsigned int ui);
+
+/** Xor immediate value into rT. */
+extern void
+spe_xor_uint(struct spe_function *p, unsigned rT, unsigned rA, unsigned int ui);
+
+/** Compare equal with immediate value. */
+extern void
+spe_compare_equal_uint(struct spe_function *p, unsigned rT, unsigned rA, unsigned int ui);
+
+/** Compare greater with immediate value. */
+extern void
+spe_compare_greater_uint(struct spe_function *p, unsigned rT, unsigned rA, unsigned int ui);
+
/** Replicate word 0 of rA across rT. */
extern void
spe_splat(struct spe_function *p, unsigned rT, unsigned rA);
-/** Complement/invert all bits in rT. */
+/** rT = complement_all_bits(rA). */
extern void
-spe_complement(struct spe_function *p, unsigned rT);
+spe_complement(struct spe_function *p, unsigned rT, unsigned rA);
/** rT = rA. */
extern void
@@ -308,6 +357,18 @@ spe_move(struct spe_function *p, unsigned rT, unsigned rA);
extern void
spe_zero(struct spe_function *p, unsigned rT);
+/** rT = splat(rA, word) */
+extern void
+spe_splat_word(struct spe_function *p, unsigned rT, unsigned rA, int word);
+
+/** rT = float min(rA, rB) */
+extern void
+spe_float_min(struct spe_function *p, unsigned rT, unsigned rA, unsigned rB);
+
+/** rT = float max(rA, rB) */
+extern void
+spe_float_max(struct spe_function *p, unsigned rT, unsigned rA, unsigned rB);
+
/* Floating-point instructions
*/