This patch fixes the FAIL of gcc.target/i386/pr78794.c on ia32, which is caused by minor STV rtx_cost differences with -march=silvermont. It turns out that generic tuning results in pandn, but the lack of accurate parameterization for COMPARE in compute_convert_gain combined with small differences in scalar<->SSE costs on silvermont results in this DImode chain not being converted.
The solution is to provide more accurate costs/gains for converting (DImode and SImode) comparisons. I'd been holding off of doing this as I'd thought it would be possible to turn pandn;ptestz into ptestc (for an even bigger scalar-to-vector win) but I've recently realized that these optimizations (as I've implemented them) occur in the wrong order (stv2 occurs after combine), so it isn't easy for STV to convert CCZmode into CCCmode. Doh! Perhaps something can be done in peephole2... This patch has been tested on x86_64-pc-linux-gnu with make bootstrap and make -k check, both with and without --target_board=unix{-m32} with no new failures. Ok for mainline? 2023-06-27 Roger Sayle <ro...@nextmovesoftware.com> gcc/ChangeLog PR target/78794 * config/i386/i386-features.cc (compute_convert_gain): Provide more accurate gains for conversion of scalar comparisons to PTEST. Thanks for your patience. Roger --
diff --git a/gcc/config/i386/i386-features.cc b/gcc/config/i386/i386-features.cc index 4a3b07a..53bec08 100644 --- a/gcc/config/i386/i386-features.cc +++ b/gcc/config/i386/i386-features.cc @@ -631,7 +631,31 @@ general_scalar_chain::compute_convert_gain () break; case COMPARE: - /* Assume comparison cost is the same. */ + if (XEXP (src, 1) != const0_rtx) + { + /* cmp vs. pxor;pshufd;ptest. */ + igain += COSTS_N_INSNS (m - 3); + } + else if (GET_CODE (XEXP (src, 0)) != AND) + { + /* test vs. pshufd;ptest. */ + igain += COSTS_N_INSNS (m - 2); + } + else if (GET_CODE (XEXP (XEXP (src, 0), 0)) != NOT) + { + /* and;test vs. pshufd;ptest. */ + igain += COSTS_N_INSNS (2 * m - 2); + } + else if (TARGET_BMI) + { + /* andn;test vs. pandn;pshufd;ptest. */ + igain += COSTS_N_INSNS (2 * m - 3); + } + else + { + /* not;and;test vs. pandn;pshufd;ptest. */ + igain += COSTS_N_INSNS (3 * m - 3); + } break; case CONST_INT: