[pk] various PK cleanups/speedups

10 files changed, 177 insertions, 97 deletions

diff --git a/pk/fp.c b/pk/fp.c
index 7ed43cb..a034c5a 100644
--- a/pk/fp.c
+++ b/pk/fp.c
@@ -2,6 +2,7 @@
 #include "softfloat.h"
 #include "riscv-opc.h"
 #include "pk.h"
+#include "fp.h"
 #include <stdint.h>
 
 #define noisy 0
@@ -10,8 +11,6 @@ static void set_fp_reg(unsigned int which, unsigned int dp, uint64_t val);
 static uint64_t get_fp_reg(unsigned int which, unsigned int dp);
 
 static fp_state_t fp_state;
-static void get_fp_state();
-static void put_fp_state();
 
 static inline void
 validate_address(trapframe_t* tf, long addr, int size, int store)
@@ -24,9 +23,8 @@ validate_address(trapframe_t* tf, long addr, int size, int store)
 
 int emulate_fp(trapframe_t* tf)
 {
-  fp_state.fsr = mfcr(CR_FSR);
   if(have_fp)
-    get_fp_state();
+    fp_state.fsr = get_fp_state(fp_state.fpr);
 
   if(noisy)
     printk("FPU emulation at pc %lx, insn %x\n",tf->epc,(uint32_t)tf->insn);
@@ -45,9 +43,9 @@ int emulate_fp(trapframe_t* tf)
   #define XRS2 (tf->gpr[RRS2])
   #define XRDR (tf->gpr[RRD])
 
-  uint64_t frs1d = get_fp_reg(RRS1, 1);
-  uint64_t frs2d = get_fp_reg(RRS2, 1);
-  uint64_t frs3d = get_fp_reg(RRS3, 1);
+  uint64_t frs1d = fp_state.fpr[RRS1];
+  uint64_t frs2d = fp_state.fpr[RRS2];
+  uint64_t frs3d = fp_state.fpr[RRS3];
   uint32_t frs1s = get_fp_reg(RRS1, 0);
   uint32_t frs2s = get_fp_reg(RRS2, 0);
   uint32_t frs3s = get_fp_reg(RRS3, 0);
@@ -217,9 +215,8 @@ int emulate_fp(trapframe_t* tf)
   else
     return -1;
 
-  mtcr(fp_state.fsr, CR_FSR);
   if(have_fp)
-    put_fp_state();
+    put_fp_state(fp_state.fpr,fp_state.fsr);
 
   advance_pc(tf);
 
@@ -231,6 +228,8 @@ int emulate_fp(trapframe_t* tf)
 
 #define PUT_FP_REG(which, type, val) asm("mtf." STR(type) " $f" STR(which) ",%0" : : "r"(val))
 #define GET_FP_REG(which, type, val) asm("mff." STR(type) " %0,$f" STR(which) : "=r"(val))
+#define LOAD_FP_REG(which, type, val) asm("l." STR(type) " $f" STR(which) ",%0" : : "m"(val))
+#define STORE_FP_REG(which, type, val)  asm("s." STR(type) " $f" STR(which) ",%0" : "=m"(val) : : "memory")
 
 static void __attribute__((noinline))
 set_fp_reg(unsigned int which, unsigned int dp, uint64_t val)
@@ -248,11 +247,8 @@ set_fp_reg(unsigned int which, unsigned int dp, uint64_t val)
     // to set an SP value, move the SP value into the FPU
     // then move it back out as a DP value.  OK to clobber $f0
     // because we'll restore it later.
-    uint64_t tmp;
-    GET_FP_REG(0,d,tmp);
     PUT_FP_REG(0,s,val);
     GET_FP_REG(0,d,fp_state.fpr[which]);
-    PUT_FP_REG(0,d,tmp);
   }
 }
 
@@ -267,11 +263,8 @@ get_fp_reg(unsigned int which, unsigned int dp)
     // to get an SP value, move the DP value into the FPU
     // then move it back out as an SP value.  OK to clobber $f0
     // because we'll restore it later.
-    uint64_t tmp;
-    GET_FP_REG(0,d,tmp);
     PUT_FP_REG(0,d,fp_state.fpr[which]);
     GET_FP_REG(0,s,val);
-    PUT_FP_REG(0,d,tmp);
   }
 
   if(noisy)
@@ -283,83 +276,11 @@ get_fp_reg(unsigned int which, unsigned int dp)
   return val;
 }
 
-static void __attribute__((noinline)) get_fp_state()
-{
-  GET_FP_REG(0, d, fp_state.fpr[0]);
-  GET_FP_REG(1, d, fp_state.fpr[1]);
-  GET_FP_REG(2, d, fp_state.fpr[2]);
-  GET_FP_REG(3, d, fp_state.fpr[3]);
-  GET_FP_REG(4, d, fp_state.fpr[4]);
-  GET_FP_REG(5, d, fp_state.fpr[5]);
-  GET_FP_REG(6, d, fp_state.fpr[6]);
-  GET_FP_REG(7, d, fp_state.fpr[7]);
-  GET_FP_REG(8, d, fp_state.fpr[8]);
-  GET_FP_REG(9, d, fp_state.fpr[9]);
-  GET_FP_REG(10, d, fp_state.fpr[10]);
-  GET_FP_REG(11, d, fp_state.fpr[11]);
-  GET_FP_REG(12, d, fp_state.fpr[12]);
-  GET_FP_REG(13, d, fp_state.fpr[13]);
-  GET_FP_REG(14, d, fp_state.fpr[14]);
-  GET_FP_REG(15, d, fp_state.fpr[15]);
-  GET_FP_REG(16, d, fp_state.fpr[16]);
-  GET_FP_REG(17, d, fp_state.fpr[17]);
-  GET_FP_REG(18, d, fp_state.fpr[18]);
-  GET_FP_REG(19, d, fp_state.fpr[19]);
-  GET_FP_REG(20, d, fp_state.fpr[20]);
-  GET_FP_REG(21, d, fp_state.fpr[21]);
-  GET_FP_REG(22, d, fp_state.fpr[22]);
-  GET_FP_REG(23, d, fp_state.fpr[23]);
-  GET_FP_REG(24, d, fp_state.fpr[24]);
-  GET_FP_REG(25, d, fp_state.fpr[25]);
-  GET_FP_REG(26, d, fp_state.fpr[26]);
-  GET_FP_REG(27, d, fp_state.fpr[27]);
-  GET_FP_REG(28, d, fp_state.fpr[28]);
-  GET_FP_REG(29, d, fp_state.fpr[29]);
-  GET_FP_REG(30, d, fp_state.fpr[30]);
-  GET_FP_REG(31, d, fp_state.fpr[31]);
-}
-
-static void __attribute__((noinline)) put_fp_state()
-{
-  PUT_FP_REG(0, d, fp_state.fpr[0]);
-  PUT_FP_REG(1, d, fp_state.fpr[1]);
-  PUT_FP_REG(2, d, fp_state.fpr[2]);
-  PUT_FP_REG(3, d, fp_state.fpr[3]);
-  PUT_FP_REG(4, d, fp_state.fpr[4]);
-  PUT_FP_REG(5, d, fp_state.fpr[5]);
-  PUT_FP_REG(6, d, fp_state.fpr[6]);
-  PUT_FP_REG(7, d, fp_state.fpr[7]);
-  PUT_FP_REG(8, d, fp_state.fpr[8]);
-  PUT_FP_REG(9, d, fp_state.fpr[9]);
-  PUT_FP_REG(10, d, fp_state.fpr[10]);
-  PUT_FP_REG(11, d, fp_state.fpr[11]);
-  PUT_FP_REG(12, d, fp_state.fpr[12]);
-  PUT_FP_REG(13, d, fp_state.fpr[13]);
-  PUT_FP_REG(14, d, fp_state.fpr[14]);
-  PUT_FP_REG(15, d, fp_state.fpr[15]);
-  PUT_FP_REG(16, d, fp_state.fpr[16]);
-  PUT_FP_REG(17, d, fp_state.fpr[17]);
-  PUT_FP_REG(18, d, fp_state.fpr[18]);
-  PUT_FP_REG(19, d, fp_state.fpr[19]);
-  PUT_FP_REG(20, d, fp_state.fpr[20]);
-  PUT_FP_REG(21, d, fp_state.fpr[21]);
-  PUT_FP_REG(22, d, fp_state.fpr[22]);
-  PUT_FP_REG(23, d, fp_state.fpr[23]);
-  PUT_FP_REG(24, d, fp_state.fpr[24]);
-  PUT_FP_REG(25, d, fp_state.fpr[25]);
-  PUT_FP_REG(26, d, fp_state.fpr[26]);
-  PUT_FP_REG(27, d, fp_state.fpr[27]);
-  PUT_FP_REG(28, d, fp_state.fpr[28]);
-  PUT_FP_REG(29, d, fp_state.fpr[29]);
-  PUT_FP_REG(30, d, fp_state.fpr[30]);
-  PUT_FP_REG(31, d, fp_state.fpr[31]);
-}
-
 void init_fp_regs()
 {
   long sr = mfpcr(PCR_SR);
   mtpcr(sr | SR_EF, PCR_SR);
-  put_fp_state();
+  put_fp_state(fp_state.fpr,fp_state.fsr);
   mtpcr(sr, PCR_SR);
 }
 
diff --git a/pk/fp.h b/pk/fp.h
new file mode 100644
index 0000000..6cdae46
--- /dev/null
+++ b/pk/fp.h
@@ -0,0 +1,13 @@
+#ifndef _FP_H
+#define _FP_H
+
+typedef struct
+{
+  uint64_t fpr[32];
+  uint32_t fsr;
+} fp_state_t;
+
+void put_fp_state(const void* fp_regs, long fsr);
+long get_fp_state(void* fp_regs);
+
+#endif
diff --git a/pk/fp_asm.S b/pk/fp_asm.S
new file mode 100644
index 0000000..4788e84
--- /dev/null
+++ b/pk/fp_asm.S
@@ -0,0 +1,84 @@
+#include "pcr.h"
+
+	.text
+	.globl	get_fp_state
+	.ent	get_fp_state
+get_fp_state:
+
+	mfcr	$v0, ASM_CR(CR_FSR)
+
+	s.d		$f0 ,  0($a0)
+	s.d		$f1 ,  8($a0)
+	s.d		$f2 , 16($a0)
+	s.d		$f3 , 24($a0)
+	s.d		$f4 , 32($a0)
+	s.d		$f5 , 40($a0)
+	s.d		$f6 , 48($a0)
+	s.d		$f7 , 56($a0)
+	s.d		$f8 , 64($a0)
+	s.d		$f9 , 72($a0)
+	s.d		$f10, 80($a0)
+	s.d		$f11, 88($a0)
+	s.d		$f12, 96($a0)
+	s.d		$f13,104($a0)
+	s.d		$f14,112($a0)
+	s.d		$f15,120($a0)
+	s.d		$f16,128($a0)
+	s.d		$f17,136($a0)
+	s.d		$f18,144($a0)
+	s.d		$f19,152($a0)
+	s.d		$f20,160($a0)
+	s.d		$f21,168($a0)
+	s.d		$f22,176($a0)
+	s.d		$f23,184($a0)
+	s.d		$f24,192($a0)
+	s.d		$f25,200($a0)
+	s.d		$f26,208($a0)
+	s.d		$f27,216($a0)
+	s.d		$f28,224($a0)
+	s.d		$f29,232($a0)
+	s.d		$f30,240($a0)
+	s.d		$f31,248($a0)
+
+	.end	get_fp_state
+
+	.globl	put_fp_state
+	.ent	put_fp_state
+put_fp_state:
+
+	l.d		$f0 ,  0($a0)
+	l.d		$f1 ,  8($a0)
+	l.d		$f2 , 16($a0)
+	l.d		$f3 , 24($a0)
+	l.d		$f4 , 32($a0)
+	l.d		$f5 , 40($a0)
+	l.d		$f6 , 48($a0)
+	l.d		$f7 , 56($a0)
+	l.d		$f8 , 64($a0)
+	l.d		$f9 , 72($a0)
+	l.d		$f10, 80($a0)
+	l.d		$f11, 88($a0)
+	l.d		$f12, 96($a0)
+	l.d		$f13,104($a0)
+	l.d		$f14,112($a0)
+	l.d		$f15,120($a0)
+	l.d		$f16,128($a0)
+	l.d		$f17,136($a0)
+	l.d		$f18,144($a0)
+	l.d		$f19,152($a0)
+	l.d		$f20,160($a0)
+	l.d		$f21,168($a0)
+	l.d		$f22,176($a0)
+	l.d		$f23,184($a0)
+	l.d		$f24,192($a0)
+	l.d		$f25,200($a0)
+	l.d		$f26,208($a0)
+	l.d		$f27,216($a0)
+	l.d		$f28,224($a0)
+	l.d		$f29,232($a0)
+	l.d		$f30,240($a0)
+	l.d		$f31,248($a0)
+
+	mtcr	$a1, ASM_CR(CR_FSR)
+
+	.end	put_fp_state
diff --git a/pk/memset.c b/pk/memset.c
new file mode 100644
index 0000000..f062ad4
--- /dev/null
+++ b/pk/memset.c
@@ -0,0 +1,39 @@
+#include <stdlib.h>
+#include <limits.h>
+#include <string.h>
+
+void* memset(void* m, int ch, size_t s)
+{
+  char* mem = (char*)m;
+  while(((long)m & (sizeof(long)-1)) && s)
+  {
+    *mem++ = ch;
+    s--;
+  }
+
+  long l = ch & 0xFF;
+  l = l | (l << 8);
+  l = l | (l << 16);
+  if(sizeof(long) == 8)
+    l = l | (l << 32);
+  else if(sizeof(long) != 4)
+    abort();
+
+  long* lmem = (long*)mem;
+  for(size_t i = 0; i < (s+sizeof(long)-1)/sizeof(long)*sizeof(long); i += 8)
+  {
+    lmem[i+0] = l;
+    lmem[i+1] = l;
+    lmem[i+2] = l;
+    lmem[i+3] = l;
+    lmem[i+4] = l;
+    lmem[i+5] = l;
+    lmem[i+6] = l;
+    lmem[i+7] = l;
+  }
+
+  for(size_t i = (s+sizeof(long)-1)/sizeof(long)*sizeof(long); i < s; i++)
+    mem[i] = ch;
+
+  return m;
+}
diff --git a/pk/pk.c b/pk/pk.c
index e9e3e71..f4e0424 100644
--- a/pk/pk.c
+++ b/pk/pk.c
@@ -145,6 +145,7 @@ static void mainvars_init()
 
 static void jump_usrstart()
 {
+  printk("strlen(\"\") = %d\n",strlen(""));
   trapframe_t tf;
   init_tf(&tf, USER_START, USER_MEM_SIZE-USER_MAINVARS_SIZE);
   pop_tf(&tf);
diff --git a/pk/pk.h b/pk/pk.h
index 0e56fcc..a1fd0da 100644
--- a/pk/pk.h
+++ b/pk/pk.h
@@ -13,12 +13,6 @@ typedef struct
   long insn;
 } trapframe_t;
 
-typedef struct
-{
-  uint64_t fpr[32];
-  uint32_t fsr;
-} fp_state_t;
-
 #define USER_MEM_SIZE 0x70000000
 #define USER_MAINVARS_SIZE 0x1000
 #define USER_START 0x10000
diff --git a/pk/pk.ld b/pk/pk.ld
index 0ccf7ef..577de0f 100644
--- a/pk/pk.ld
+++ b/pk/pk.ld
@@ -2,8 +2,6 @@ OUTPUT_ARCH( "mips:riscv" )
 
 ENTRY( __start )
 
-GROUP( -lc -lgcc -lgloss )
-
 SECTIONS
 {
 
diff --git a/pk/pk.mk.in b/pk/pk.mk.in
index 1de66e8..8d1a06a 100644
--- a/pk/pk.mk.in
+++ b/pk/pk.mk.in
@@ -5,6 +5,7 @@ pk_subproject_deps = \
 pk_hdrs = \
 	pk.h \
 	pcr.h \
+	fp.h \
 	atomic.h \
 	file.h \
 	frontend.h \
@@ -17,9 +18,12 @@ pk_c_srcs = \
 	handlers.c \
 	frontend.c \
 	fp.c \
+	memset.c \
+	strlen.c \
 
 pk_asm_srcs = \
 	entry.S \
+	fp_asm.S \
 
 pk_test_srcs =
 
diff --git a/pk/riscv-pk.c b/pk/riscv-pk.c
index ea5d4d4..771982c 100644
--- a/pk/riscv-pk.c
+++ b/pk/riscv-pk.c
@@ -22,7 +22,7 @@ void __attribute__((section(".boottext"))) __start()
   #endif
 
   mtpcr(sr0 | SR_EF, PCR_SR);
-  have_fp = 0;//mfpcr(PCR_SR) & SR_EF;
+  have_fp = mfpcr(PCR_SR) & SR_EF;
   mtpcr(sr0, PCR_SR);
 
   extern void boot();
diff --git a/pk/strlen.c b/pk/strlen.c
new file mode 100644
index 0000000..f600d47
--- /dev/null
+++ b/pk/strlen.c
@@ -0,0 +1,26 @@
+#include <string.h>
+#include <stdlib.h>
+
+// from http://www-graphics.stanford.edu/~seander/bithacks.html
+static inline long hasZeroByte(long l)
+{
+  if(sizeof(long) == 4)
+    return (l - 0x01010101UL) & ~l & 0x80808080UL;
+  else if(sizeof(long) == 8)
+    return (l - 0x0101010101010101UL) & ~l & 0x8080808080808080UL;
+}
+
+size_t strlen(const char* s)
+{
+  size_t i = 0;
+
+  // use optimized version if string starts on a long boundary
+  if(((long)s & (sizeof(long)-1)) == 0)
+    while(!hasZeroByte(*(long*)(s+i)))
+      i += sizeof(long);
+
+  while(s[i])
+    i++;
+
+  return i;
+}