[pk] made PK smaller

1 files changed, 50 insertions, 40 deletions

diff --git a/pk/fp.c b/pk/fp.c
index 7cc929b..794755d 100644
--- a/pk/fp.c
+++ b/pk/fp.c
@@ -65,15 +65,22 @@ int emulate_fp(trapframe_t* tf)
 
   #define IS_INSN(x) ((tf->insn & MASK_ ## x) == MATCH_ ## x)
 
+  int do_writeback = 0;
+  int writeback_dp;
+  uint64_t writeback_value;
+  #define DO_WRITEBACK(dp, value) \
+    do { do_writeback = 1; writeback_dp = (dp); writeback_value = (value); } \
+    while(0)
+
   if(IS_INSN(FLW))
   {
     validate_address(tf, effective_address_load, 4, 0);
-    set_fp_reg(RRD, 0, *(uint32_t*)effective_address_load);
+    DO_WRITEBACK(0, *(uint32_t*)effective_address_load);
   }
   else if(IS_INSN(FLD))
   {
     validate_address(tf, effective_address_load, 8, 0);
-    set_fp_reg(RRD, 1, *(uint64_t*)effective_address_load);
+    DO_WRITEBACK(1, *(uint64_t*)effective_address_load);
   }
   else if(IS_INSN(FSW))
   {
@@ -90,89 +97,89 @@ int emulate_fp(trapframe_t* tf)
   else if(IS_INSN(MFTX_D))
     XRDR = frs2d;
   else if(IS_INSN(MXTF_S))
-    set_fp_reg(RRD, 0, XRS1);
+    DO_WRITEBACK(0, XRS1);
   else if(IS_INSN(MXTF_D))
-    set_fp_reg(RRD, 1, XRS1);
+    DO_WRITEBACK(1, XRS1);
   else if(IS_INSN(FSGNJ_S))
-    set_fp_reg(RRD, 0, (frs1s &~ (uint32_t)INT32_MIN) | (frs2s & (uint32_t)INT32_MIN));
+    DO_WRITEBACK(0, (frs1s &~ (uint32_t)INT32_MIN) | (frs2s & (uint32_t)INT32_MIN));
   else if(IS_INSN(FSGNJ_D))
-    set_fp_reg(RRD, 1, (frs1d &~ INT64_MIN) | (frs2d & INT64_MIN));
+    DO_WRITEBACK(1, (frs1d &~ INT64_MIN) | (frs2d & INT64_MIN));
   else if(IS_INSN(FSGNJN_S))
-    set_fp_reg(RRD, 0, (frs1s &~ (uint32_t)INT32_MIN) | ((~frs2s) & (uint32_t)INT32_MIN));
+    DO_WRITEBACK(0, (frs1s &~ (uint32_t)INT32_MIN) | ((~frs2s) & (uint32_t)INT32_MIN));
   else if(IS_INSN(FSGNJN_D))
-    set_fp_reg(RRD, 1, (frs1d &~ INT64_MIN) | ((~frs2d) & INT64_MIN));
+    DO_WRITEBACK(1, (frs1d &~ INT64_MIN) | ((~frs2d) & INT64_MIN));
   else if(IS_INSN(FSGNJX_S))
-    set_fp_reg(RRD, 0, frs1s ^ (frs2s & (uint32_t)INT32_MIN));
+    DO_WRITEBACK(0, frs1s ^ (frs2s & (uint32_t)INT32_MIN));
   else if(IS_INSN(FSGNJX_D))
-    set_fp_reg(RRD, 1, frs1d ^ (frs2d & INT64_MIN));
+    DO_WRITEBACK(1, frs1d ^ (frs2d & INT64_MIN));
   else if(IS_INSN(FEQ_S))
     XRDR = f32_eq(frs1s, frs2s);
   else if(IS_INSN(FEQ_D))
     XRDR = f64_eq(frs1d, frs2d);
   else if(IS_INSN(FLE_S))
-    XRDR = f32_le(frs1s, frs2s);
+    XRDR = f32_eq(frs1s, frs2s) || f32_lt(frs1s, frs2s);
   else if(IS_INSN(FLE_D))
-    XRDR = f64_le(frs1d, frs2d);
+    XRDR = f64_eq(frs1d, frs2d) || f64_lt(frs1s, frs2s);
   else if(IS_INSN(FLT_S))
     XRDR = f32_lt(frs1s, frs2s);
   else if(IS_INSN(FLT_D))
     XRDR = f64_lt(frs1d, frs2d);
   else if(IS_INSN(FCVT_S_W))
-    set_fp_reg(RRD, 0, i32_to_f32(XRS1));
+    DO_WRITEBACK(0, i64_to_f32((int64_t)(int32_t)XRS1));
   else if(IS_INSN(FCVT_S_L))
-    set_fp_reg(RRD, 0, i64_to_f32(XRS1));
+    DO_WRITEBACK(0, i64_to_f32(XRS1));
   else if(IS_INSN(FCVT_S_D))
-    set_fp_reg(RRD, 0, f64_to_f32(frs1d));
+    DO_WRITEBACK(0, f64_to_f32(frs1d));
   else if(IS_INSN(FCVT_D_W))
-    set_fp_reg(RRD, 1, i32_to_f64(XRS1));
+    DO_WRITEBACK(1, i64_to_f64((int64_t)(int32_t)XRS1));
   else if(IS_INSN(FCVT_D_L))
-    set_fp_reg(RRD, 1, i64_to_f64(XRS1));
+    DO_WRITEBACK(1, i64_to_f64(XRS1));
   else if(IS_INSN(FCVT_D_S))
-    set_fp_reg(RRD, 1, f32_to_f64(frs1s));
+    DO_WRITEBACK(1, f32_to_f64(frs1s));
   else if(IS_INSN(FCVT_S_WU))
-    set_fp_reg(RRD, 0, ui32_to_f32(XRS1));
+    DO_WRITEBACK(0, ui64_to_f32((uint64_t)(uint32_t)XRS1));
   else if(IS_INSN(FCVT_S_LU))
-    set_fp_reg(RRD, 0, ui64_to_f32(XRS1));
+    DO_WRITEBACK(0, ui64_to_f32(XRS1));
   else if(IS_INSN(FCVT_D_WU))
-    set_fp_reg(RRD, 1, ui32_to_f64(XRS1));
+    DO_WRITEBACK(1, ui64_to_f64((uint64_t)(uint32_t)XRS1));
   else if(IS_INSN(FCVT_D_LU))
-    set_fp_reg(RRD, 1, ui64_to_f64(XRS1));
+    DO_WRITEBACK(1, ui64_to_f64(XRS1));
   else if(IS_INSN(FADD_S))
-    set_fp_reg(RRD, 0, f32_add(frs1s, frs2s));
+    DO_WRITEBACK(0, f32_mulAdd(frs1s, 0x3f800000, frs2s));
   else if(IS_INSN(FADD_D))
-    set_fp_reg(RRD, 1, f64_add(frs1d, frs2d));
+    DO_WRITEBACK(1, f64_mulAdd(frs1d, 0x3ff0000000000000LL, frs2d));
   else if(IS_INSN(FSUB_S))
-    set_fp_reg(RRD, 0, f32_sub(frs1s, frs2s));
+    DO_WRITEBACK(0, f32_mulAdd(frs1s, 0x3f800000, frs2s ^ (uint32_t)INT32_MIN));
   else if(IS_INSN(FSUB_D))
-    set_fp_reg(RRD, 1, f64_sub(frs1d, frs2d));
+    DO_WRITEBACK(1, f64_mulAdd(frs1d, 0x3ff0000000000000LL, frs2d ^ INT64_MIN));
   else if(IS_INSN(FMUL_S))
-    set_fp_reg(RRD, 0, f32_mul(frs1s, frs2s));
+    DO_WRITEBACK(0, f32_mulAdd(frs1s, frs2s, 0));
   else if(IS_INSN(FMUL_D))
-    set_fp_reg(RRD, 1, f64_mul(frs1d, frs2d));
+    DO_WRITEBACK(1, f64_mulAdd(frs1d, frs2d, 0));
   else if(IS_INSN(FMADD_S))
-    set_fp_reg(RRD, 0, f32_mulAdd(frs1s, frs2s, frs3s));
+    DO_WRITEBACK(0, f32_mulAdd(frs1s, frs2s, frs3s));
   else if(IS_INSN(FMADD_D))
-    set_fp_reg(RRD, 1, f64_mulAdd(frs1d, frs2d, frs3d));
+    DO_WRITEBACK(1, f64_mulAdd(frs1d, frs2d, frs3d));
   else if(IS_INSN(FMSUB_S))
-    set_fp_reg(RRD, 0, f32_mulAdd(frs1s, frs2s, frs3s ^ (uint32_t)INT32_MIN));
+    DO_WRITEBACK(0, f32_mulAdd(frs1s, frs2s, frs3s ^ (uint32_t)INT32_MIN));
   else if(IS_INSN(FMSUB_D))
-    set_fp_reg(RRD, 1, f64_mulAdd(frs1d, frs2d, frs3d ^ INT64_MIN));
+    DO_WRITEBACK(1, f64_mulAdd(frs1d, frs2d, frs3d ^ INT64_MIN));
   else if(IS_INSN(FNMADD_S))
-    set_fp_reg(RRD, 0, f32_mulAdd(frs1s, frs2s, frs3s) ^ (uint32_t)INT32_MIN);
+    DO_WRITEBACK(0, f32_mulAdd(frs1s, frs2s, frs3s) ^ (uint32_t)INT32_MIN);
   else if(IS_INSN(FNMADD_D))
-    set_fp_reg(RRD, 1, f64_mulAdd(frs1d, frs2d, frs3d) ^ INT64_MIN);
+    DO_WRITEBACK(1, f64_mulAdd(frs1d, frs2d, frs3d) ^ INT64_MIN);
   else if(IS_INSN(FNMSUB_S))
-    set_fp_reg(RRD, 0, f32_mulAdd(frs1s, frs2s, frs3s ^ (uint32_t)INT32_MIN) ^ (uint32_t)INT32_MIN);
+    DO_WRITEBACK(0, f32_mulAdd(frs1s, frs2s, frs3s ^ (uint32_t)INT32_MIN) ^ (uint32_t)INT32_MIN);
   else if(IS_INSN(FNMSUB_D))
-    set_fp_reg(RRD, 1, f64_mulAdd(frs1d, frs2d, frs3d ^ INT64_MIN) ^ INT64_MIN);
+    DO_WRITEBACK(1, f64_mulAdd(frs1d, frs2d, frs3d ^ INT64_MIN) ^ INT64_MIN);
   else if(IS_INSN(FDIV_S))
-    set_fp_reg(RRD, 0, f32_div(frs1s, frs2s));
+    DO_WRITEBACK(0, f32_div(frs1s, frs2s));
   else if(IS_INSN(FDIV_D))
-    set_fp_reg(RRD, 1, f64_div(frs1d, frs2d));
+    DO_WRITEBACK(1, f64_div(frs1d, frs2d));
   else if(IS_INSN(FSQRT_S))
-    set_fp_reg(RRD, 0, f32_sqrt(frs1s));
+    DO_WRITEBACK(0, f32_sqrt(frs1s));
   else if(IS_INSN(FSQRT_D))
-    set_fp_reg(RRD, 1, f64_sqrt(frs1d));
+    DO_WRITEBACK(1, f64_sqrt(frs1d));
   else if(IS_INSN(FCVT_W_S))
     XRDR = f32_to_i32_r_minMag(frs1s,true);
   else if(IS_INSN(FCVT_W_D))
@@ -192,6 +199,9 @@ int emulate_fp(trapframe_t* tf)
   else
     return -1;
 
+  if(do_writeback)
+    set_fp_reg(RRD, writeback_dp, writeback_value);
+
   if(have_fp)
     put_fp_state(fp_state.fpr,fp_state.fsr);