summaryrefslogtreecommitdiff
path: root/src/gallium/auxiliary/rtasm
diff options
context:
space:
mode:
Diffstat (limited to 'src/gallium/auxiliary/rtasm')
-rw-r--r--src/gallium/auxiliary/rtasm/rtasm_execmem.c19
-rw-r--r--src/gallium/auxiliary/rtasm/rtasm_ppc.c79
-rw-r--r--src/gallium/auxiliary/rtasm/rtasm_ppc.h9
-rw-r--r--src/gallium/auxiliary/rtasm/rtasm_ppc_spe.c57
4 files changed, 115 insertions, 49 deletions
diff --git a/src/gallium/auxiliary/rtasm/rtasm_execmem.c b/src/gallium/auxiliary/rtasm/rtasm_execmem.c
index 19087589a8..be7433baf8 100644
--- a/src/gallium/auxiliary/rtasm/rtasm_execmem.c
+++ b/src/gallium/auxiliary/rtasm/rtasm_execmem.c
@@ -38,12 +38,13 @@
#include "rtasm_execmem.h"
-#if defined(__linux__)
+#if defined(PIPE_OS_LINUX)
+
/*
* Allocate a large block of memory which can hold code then dole it out
* in pieces by means of the generic memory manager code.
-*/
+ */
#include <unistd.h>
#include <sys/mman.h>
@@ -62,7 +63,7 @@ static void
init_heap(void)
{
if (!exec_heap)
- exec_heap = mmInit( 0, EXEC_HEAP_SIZE );
+ exec_heap = u_mmInit( 0, EXEC_HEAP_SIZE );
if (!exec_mem)
exec_mem = (unsigned char *) mmap(0, EXEC_HEAP_SIZE,
@@ -82,8 +83,8 @@ rtasm_exec_malloc(size_t size)
init_heap();
if (exec_heap) {
- size = (size + 31) & ~31;
- block = mmAllocMem( exec_heap, size, 32, 0 );
+ size = (size + 31) & ~31; /* next multiple of 32 bytes */
+ block = u_mmAllocMem( exec_heap, size, 5, 0 ); /* 5 -> 32-byte alignment */
}
if (block)
@@ -103,17 +104,17 @@ rtasm_exec_free(void *addr)
pipe_mutex_lock(exec_mutex);
if (exec_heap) {
- struct mem_block *block = mmFindBlock(exec_heap, (unsigned char *)addr - exec_mem);
+ struct mem_block *block = u_mmFindBlock(exec_heap, (unsigned char *)addr - exec_mem);
if (block)
- mmFreeMem(block);
+ u_mmFreeMem(block);
}
pipe_mutex_unlock(exec_mutex);
}
-#else
+#else /* PIPE_OS_LINUX */
/*
* Just use regular memory.
@@ -133,4 +134,4 @@ rtasm_exec_free(void *addr)
}
-#endif
+#endif /* PIPE_OS_LINUX */
diff --git a/src/gallium/auxiliary/rtasm/rtasm_ppc.c b/src/gallium/auxiliary/rtasm/rtasm_ppc.c
index 7dd8263749..6d11263be8 100644
--- a/src/gallium/auxiliary/rtasm/rtasm_ppc.c
+++ b/src/gallium/auxiliary/rtasm/rtasm_ppc.c
@@ -38,17 +38,18 @@
#include <stdio.h>
#include "util/u_memory.h"
#include "pipe/p_debug.h"
+#include "rtasm_execmem.h"
#include "rtasm_ppc.h"
void
-ppc_init_func(struct ppc_function *p, unsigned max_inst)
+ppc_init_func(struct ppc_function *p)
{
uint i;
- p->store = align_malloc(max_inst * PPC_INST_SIZE, 16);
p->num_inst = 0;
- p->max_inst = max_inst;
+ p->max_inst = 100; /* first guess at buffer size */
+ p->store = rtasm_exec_malloc(p->max_inst * PPC_INST_SIZE);
p->reg_used = 0x0;
p->fp_used = 0x0;
p->vec_used = 0x0;
@@ -66,12 +67,19 @@ ppc_release_func(struct ppc_function *p)
{
assert(p->num_inst <= p->max_inst);
if (p->store != NULL) {
- align_free(p->store);
+ rtasm_exec_free(p->store);
}
p->store = NULL;
}
+uint
+ppc_num_instructions(const struct ppc_function *p)
+{
+ return p->num_inst;
+}
+
+
void (*ppc_get_func(struct ppc_function *p))(void)
{
#if 0
@@ -202,6 +210,35 @@ ppc_release_vec_register(struct ppc_function *p, int reg)
}
+/**
+ * Append instruction to instruction buffer. Grow buffer if out of room.
+ */
+static void
+emit_instruction(struct ppc_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 = rtasm_exec_malloc(p->max_inst * PPC_INST_SIZE);
+ if (newbuf) {
+ memcpy(newbuf, p->store, p->num_inst * PPC_INST_SIZE);
+ }
+ rtasm_exec_free(p->store);
+ p->store = newbuf;
+ if (!p->store) {
+ /* out of memory */
+ p->num_inst = 0;
+ return;
+ }
+ }
+
+ p->store[p->num_inst++] = inst_bits;
+}
+
union vx_inst {
uint32_t bits;
@@ -223,8 +260,7 @@ emit_vx(struct ppc_function *p, uint op2, uint vD, uint vA, uint vB)
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);
+ emit_instruction(p, inst.bits);
};
@@ -250,8 +286,7 @@ emit_vxr(struct ppc_function *p, uint op2, uint vD, uint vA, uint vB)
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);
+ emit_instruction(p, inst.bits);
};
@@ -277,8 +312,7 @@ emit_va(struct ppc_function *p, uint op2, uint vD, uint vA, uint vB, uint vC)
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);
+ emit_instruction(p, inst.bits);
};
@@ -300,8 +334,7 @@ emit_i(struct ppc_function *p, uint op, uint li, uint aa, uint lk)
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);
+ emit_instruction(p, inst.bits);
}
@@ -330,8 +363,7 @@ emit_xl(struct ppc_function *p, uint op, uint bo, uint bi, uint bh,
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);
+ emit_instruction(p, inst.bits);
}
static INLINE void
@@ -373,8 +405,7 @@ emit_x(struct ppc_function *p, uint op, uint vrs, uint ra, uint rb, uint op2)
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);
+ emit_instruction(p, inst.bits);
}
@@ -398,8 +429,7 @@ emit_d(struct ppc_function *p, uint op, uint rt, uint ra, int si)
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);
+ emit_instruction(p, inst.bits);
};
@@ -428,8 +458,7 @@ emit_a(struct ppc_function *p, uint op, uint frt, uint fra, uint frb, uint op2,
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);
+ emit_instruction(p, inst.bits);
};
@@ -458,8 +487,7 @@ emit_xo(struct ppc_function *p, uint op, uint rt, uint ra, uint rb, uint oe,
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);
+ emit_instruction(p, inst.bits);
}
@@ -505,6 +533,13 @@ ppc_vmaddfp(struct ppc_function *p, uint vD, uint vA, uint vB, uint vC)
emit_va(p, 46, vD, vA, vC, vB); /* note arg order */
}
+/** vector float negative mult subtract: vD = vA - vB * vC */
+void
+ppc_vnmsubfp(struct ppc_function *p, uint vD, uint vA, uint vB, uint vC)
+{
+ emit_va(p, 47, vD, vB, vA, vC); /* note arg order */
+}
+
/** vector float compare greater than */
void
ppc_vcmpgtfpx(struct ppc_function *p, uint vD, uint vA, uint vB)
diff --git a/src/gallium/auxiliary/rtasm/rtasm_ppc.h b/src/gallium/auxiliary/rtasm/rtasm_ppc.h
index f938d8d759..afb4704c39 100644
--- a/src/gallium/auxiliary/rtasm/rtasm_ppc.h
+++ b/src/gallium/auxiliary/rtasm/rtasm_ppc.h
@@ -62,8 +62,9 @@ struct ppc_function
-extern void ppc_init_func(struct ppc_function *p, unsigned max_inst);
+extern void ppc_init_func(struct ppc_function *p);
extern void ppc_release_func(struct ppc_function *p);
+extern uint ppc_num_instructions(const struct ppc_function *p);
extern void (*ppc_get_func( struct ppc_function *p ))( void );
extern void ppc_dump_func(const struct ppc_function *p);
@@ -97,10 +98,14 @@ ppc_vminfp(struct ppc_function *p, uint vD, uint vA, uint vB);
extern void
ppc_vmaxfp(struct ppc_function *p, uint vD, uint vA, uint vB);
-/** vector float mult add */
+/** vector float mult add: vD = vA * vB + vC */
extern void
ppc_vmaddfp(struct ppc_function *p, uint vD, uint vA, uint vB, uint vC);
+/** vector float negative mult subtract: vD = vA - vB * vC */
+extern void
+ppc_vnmsubfp(struct ppc_function *p, uint vD, uint vA, uint vB, uint vC);
+
/** vector float compare greater than */
extern void
ppc_vcmpgtfpx(struct ppc_function *p, uint vD, uint vA, uint vB);
diff --git a/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.c b/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.c
index dea1aed032..f8568f690b 100644
--- a/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.c
+++ b/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.c
@@ -185,6 +185,34 @@ reg_name(int reg)
}
+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, unsigned rT,
unsigned rA, unsigned rB, const char *name)
{
@@ -193,8 +221,7 @@ 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->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",
@@ -212,8 +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),
@@ -230,8 +256,7 @@ 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->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",
@@ -249,8 +274,7 @@ 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->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",
@@ -268,8 +292,7 @@ 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->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",
@@ -295,8 +318,7 @@ 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->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);
@@ -311,8 +333,7 @@ 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->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);
@@ -394,15 +415,19 @@ void _name (struct spe_function *p, int imm) \
/**
* 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)
{
unsigned int i;
- p->store = align_malloc(code_size, 16);
+ if (!code_size)
+ code_size = 64;
+
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]));
generated by cgit v1.2.3 (git 2.39.1) at 2025-04-05 15:43:19 +0000