On 2/28/22 04:48, Weiwei Li wrote:
+#define GEN_SHA512H_RV32(NAME, OP, NUM1, NUM2, NUM3) \
+static void gen_##NAME(TCGv dest, TCGv src1, TCGv src2) \
+{ \
+ TCGv_i64 t0 = tcg_temp_new_i64(); \
+ TCGv_i64 t1 = tcg_temp_new_i64(); \
+ TCGv_i64 t2 = tcg_temp_new_i64(); \
+ \
+ tcg_gen_concat_tl_i64(t0, src1, src2); \
+ tcg_gen_##OP##_i64(t1, t0, NUM1); \
+ tcg_gen_concat_tl_i64(t2, src1, tcg_const_tl(0)); \
The bug here is tcg_const_tl instead of tcg_constant_tl, which leaks a
temporary.
It's not the best option for zero-extension, though, as we don't optimize a deposit of
zero like this (we probably should, but, hey).
Better would be
tcg_gen_extu_tl_i64(t2, src1);
tcg_gen_ext32u_i64(t2, t2);
Note that the first operation will *not* extend if TARGET_RISCV64, since it doesn't
actually change type. The second operation will be optimized away if TARGET_RISCV32,
since the zero-extend has already happened.
BTW, it would be better to not use a large macro for this, and in the previous patch.
Passing in parameters to a helper function would be easier to read and debug.
r~
