TODO notes on using sv.bc in utf8-validator

diff --git a/src/openpower/test/algorithms/svp64_utf_8_validation.py b/src/openpower/test/algorithms/svp64_utf_8_validation.py
index afcbde6809917ba138a15dad733ade4411b75a3e..fd0c39e0febd0ee40eacb05a547162781ef1d8aa 100644 (file)
--- a/src/openpower/test/algorithms/svp64_utf_8_validation.py
+++ b/src/openpower/test/algorithms/svp64_utf_8_validation.py
@@ -60,7 +60,8 @@ class UTF8FirstTwoBytesError(enum.IntFlag):
 
 # look up tables for checking for errors in the first two bytes, the final
 # error flags are generated by looking up the nibbles of the first two bytes
-# in the appropriate tables, and bitwise ANDing the results together.
+# in the appropriate tables, and bitwise ANDing the results together
+# (XXX TODO: investigate use sv.bc/all instead)
 # To figure out what to put in each entry in the LUTs, look for all cases
 # that could match the comment.
 
@@ -145,6 +146,8 @@ def svp64_utf8_validation_asm():
     # nibbles to look up in r64-r71 -- u64x8
     temp_vec2 = cur_bytes + vec_sz * 2
     temp_vec2_end = temp_vec2 + vec_sz
+    # temp consts 0xF0 0xE0 0xC0 (accessible in scalar range)
+    consts = 24
 
     def sv_set_0x80_if_ge(out_v, inp_v, temp_s, compare_rhs):
         # type: (int, int, int, int) -> list[str]
@@ -167,6 +170,10 @@ def svp64_utf8_validation_asm():
     return [
         # input addr in r3, input length in r4
         f"setvl 0, 0, {prev_bytes_sz}, 0, 1, 1",  # set VL to prev_bytes_sz
+        # create 3 consts 0xf0 0xe0 0xc0 (within scalar range, saves instrs)
+        "ori %d, 0, 0xF0" % (consts),
+        "ori %d, 0, 0xC0" % (consts+1),
+        "ori %d, 0, 0xC0" % (consts+2),
         # clear what will go into prev bytes
         f"sv.addi *{cur_bytes + vec_sz - prev_bytes_sz}, 0, 0",
         f"addis 6, 0, {FIRST_BYTE_HIGH_NIBBLE_LUT_ADDR >> 16}",
@@ -218,6 +225,7 @@ def svp64_utf8_validation_asm():
         f"sv.and *{temp_vec1}, *{temp_vec1}, *{temp_vec2}",
 
         # or-reduce error flags into temp_vec2_end
+        # XXX can likely use sv.cmpli followed by sv.bc here
         f"sv.addi {temp_vec2_end}, 0, 0",
         f"sv.ori *{temp_vec2}, *{temp_vec1}, 0",
         f"sv.or *{temp_vec2 + 1}, *{temp_vec2}, *{temp_vec2 + 1}",
@@ -242,6 +250,7 @@ def svp64_utf8_validation_asm():
         f"sv.xor *{temp_vec1}, *{temp_vec1}, *{temp_vec2}",
         # now bit 0x80 is set in temp_vec1 if there's an error
         # or-reduce into temp_vec2
+        # XXX can likely use sv.bc here, just use the Rc-vector directly
         f"sv.addi {temp_vec2}, 0, 0",
         f"sv.or *{temp_vec1 + 1}, *{temp_vec1}, *{temp_vec1 + 1}",
         # adjust count/pointer
@@ -256,23 +265,15 @@ def svp64_utf8_validation_asm():
         f"bclr 20, 0, 0 # blr",
         f"final_check:",
 
-        # need to set VL to something non-zero otherwise all our scalar
-        # instructions don't run --- I definitely don't like that ... scalar
-        # instructions should run regardless of VL.
-        f"setvl 0, 0, 1, 0, 1, 1",  # set VL to 1
+        # next loop runs on 3 consts to check 3 bytes
+        "setvl 0, 0, 3, 0, 1, 1",  # set VL to 3
 
         # check if prev input is incomplete
-        # check if byte 3 bytes from end needed 4 bytes
-        f"sv.cmpli 0, 1, {cur_bytes - 3}, {0xF0}",
-        f"bc 4, 0, fail # bge fail",
-        # check if byte 2 bytes from end needed 3 bytes
-        f"sv.cmpli 0, 1, {cur_bytes - 2}, {0xE0}",
-        f"bc 4, 0, fail # bge fail",
-        # check if byte 1 bytes from end needed 2 bytes
-        f"sv.cmpli 0, 1, {cur_bytes - 1}, {0xC0}",
-        f"bc 4, 0, fail # bge fail",
-        f"addi 3, 0, 1",
-        f"bclr 20, 0, 0 # blr",
+        # check if byte 3,2,1 bytes from end needed 4,3,2 bytes
+        "sv.cmpl 0, 1, %d, %d" % (cur_bytes - 3, consts),
+        "sv.bc 4, *0, fail # bge fail",
+        "addi 3, 0, 1",
+        "bclr 20, 0, 0 # blr",
     ]