Add fallback for BMI or BMI2 instructions

.github/workflows/ci-mac

@@ -0,0 +1,13 @@
+name: CI-Mac
+on: [push]
+jobs:
+  build-and-test:
+    runs-on: macos-latest
+    steps:
+      - name: install compilers etc
+        run: sudo apt install g++
+      - name: Check out repository code
+        uses: actions/checkout@v3
+      - run: make -j$(nproc) test no-BMI2=1
+      - run: make -j$(nproc) release no-BMI2=1
+      - run: make -j$(nproc) perft no-BMI2=1

.github/workflows/ci.yml → .github/workflows/ci-ubuntu.yml

@@ -1,4 +1,4 @@
-name: CI
+name: CI-Ubuntu
 on: [push]
 jobs:
   build-and-test:

Makefile

@@ -23,6 +23,14 @@ ifneq ($(LIBS),)
 	LINK_FLAGS += $(shell pkg-config --libs $(LIBS))
 endif
 
+# BMI and BMI2 instruction set
+ifeq ($(no-BMI),)
+	COMPILE_FLAGS += -mbmi
+ifeq ($(no-BMI2),)
+	COMPILE_FLAGS += -mbmi2
+endif
+endif
+
 # Verbose option, to output compile and link commands
 export V := false
 export CMD_PREFIX := @

config.mk

@@ -10,7 +10,7 @@ MAINS_PATH = mains
 # Space-separated pkg-config libraries used by this project
 LIBS =
 # General compiler flags
-COMPILE_FLAGS = -std=c++17 -Wall -Wextra -g -mbmi -mbmi2 -O3
+COMPILE_FLAGS = -std=c++17 -Wall -Wextra -g -O3
 # Additional release-specific flags
 RCOMPILE_FLAGS = -D NDEBUG -DDOCTEST_CONFIG_DISABLE
 # Additional debug-specific flags

includes/attacks.hpp

@@ -2,6 +2,7 @@
 
 # include "bitboard.hpp"
 # include "lookup.hpp"
+# include "bmi2_fallback.hpp"
 
 # include <tuple>
 
@@ -150,14 +151,14 @@ constexpr BitMask compute_file_blocker(const size_t square){
 const auto file_blocker_lookup = lookup_table<BitMask, 64>(compute_file_blocker);
 
 inline BitMask rook_seen_rank(const Square square, const BitMask occ){
-	const BitMask ext_block_mask = _pext_u64(occ, rank_blocker_lookup[square]);
+	const BitMask ext_block_mask = PEXT(occ, rank_blocker_lookup[square]);
 	const BitMask ext_seen_mask = slider_attack_table[square % 8][ext_block_mask];
-	return _pdep_u64(ext_seen_mask, rank_lookup[square]);
+	return PDEP(ext_seen_mask, rank_lookup[square]);
 }
 inline BitMask rook_seen_file(const Square square, const BitMask occ){
-	const BitMask ext_block_mask = _pext_u64(occ, file_blocker_lookup[square]);
+	const BitMask ext_block_mask = PEXT(occ, file_blocker_lookup[square]);
 	const BitMask ext_seen_mask = slider_attack_table[square / 8][ext_block_mask];
-	return _pdep_u64(ext_seen_mask, file_lookup[square]);
+	return PDEP(ext_seen_mask, file_lookup[square]);
 }
 inline BitMask rook_seen(const Square square, const BitMask occ){
 	return rook_seen_rank(square, occ) | rook_seen_file(square, occ);
@@ -198,15 +199,15 @@ constexpr auto index_on_lower(Square square){
 }
 
 inline BitMask bishop_seen_upper(const Square square, const BitMask occ){
-	const BitMask ext_block_mask = _pext_u64(occ, upper_blocker_lookup[square]);
+	const BitMask ext_block_mask = PEXT(occ, upper_blocker_lookup[square]);
 	const BitMask ext_seen_mask = slider_attack_table[index_on_upper(square)][ext_block_mask];
-	return _pdep_u64(ext_seen_mask, upper_diag_lookup[square]);
+	return PDEP(ext_seen_mask, upper_diag_lookup[square]);
 }
 
 inline BitMask bishop_seen_lower(const Square square, const BitMask occ){
-	const BitMask ext_block_mask = _pext_u64(occ, lower_blocker_lookup[square]);
+	const BitMask ext_block_mask = PEXT(occ, lower_blocker_lookup[square]);
 	const BitMask ext_seen_mask = slider_attack_table[index_on_lower(square)][ext_block_mask];
-	return _pdep_u64(ext_seen_mask, lower_diag_lookup[square]);
+	return PDEP(ext_seen_mask, lower_diag_lookup[square]);
 }
 
 inline BitMask bishop_seen(const Square square, const BitMask occ){

includes/bitboard.hpp

@@ -2,10 +2,11 @@
 
 # include <cstdint>
 # include <x86intrin.h>
+# include "bmi_fallback.hpp"
 
 #define SquareOf(X) _tzcnt_u64(X)
 #define ToMask(X) (1ull << (X))
-#define Bitloop(X, var) for(auto var = X; var; var = __blsr_u64(var))
+#define Bitloop(X, var) for(auto var = X; var; var = BLSR(var))
 #define Flip(X) ((X) ^ 56)
 #define FlipIf(cond, X) ((cond) ? ((X) ^ 56) : (X))
 

includes/bmi2_fallback.hpp

@@ -0,0 +1,33 @@
+# pragma once
+
+# ifdef __BMI2__
+
+# include <x86intrin.h>
+# define PEXT _pext_u64
+# define PDEP _pdep_u64
+
+# else
+
+# include <cstdint>
+
+constexpr uint64_t PEXT(const uint64_t val, uint64_t mask) {
+  uint64_t res = 0;
+  for (uint64_t bb = 1; mask; bb += bb) {
+    if ( val & mask & -mask )
+      res |= bb;
+    mask &= mask - 1;
+  }
+  return res;
+}
+
+constexpr uint64_t PDEP(const uint64_t val, uint64_t mask) {
+  uint64_t res = 0;
+  for (uint64_t bb = 1; mask; bb += bb) {
+    if (val & bb)
+      res |= mask & -mask;
+    mask &= mask - 1;
+  }
+  return res;
+}
+
+# endif

includes/bmi_fallback.hpp

@@ -0,0 +1,13 @@
+# pragma once
+
+# ifdef __BMI__
+
+# include <x86intrin.h>
+# define BLSR __blsr_u64
+
+# else
+
+# include <cstdint>
+constexpr uint64_t BLSR(const uint64_t x){ return x & (x-1); }
+
+# endif

src/movegen.cpp

@@ -4,6 +4,7 @@
 # include "movegen.hpp"
 # include "parse_format.hpp"
 # include "attacks.hpp"
+# include "bmi2_fallback.hpp"
 
 /* KNIGHTS */
 
@@ -85,21 +86,21 @@ std::tuple<BitMask, BitMask> rook_checks_and_pins(
 	const BitMask rank = rank_lookup[king];
 	Bitloop(rank & rooks, loop_var){
 		const Square rook = SquareOf(loop_var);
-		const BitMask ext_block_mask = _pext_u64(occ, rank_blocker_lookup[king]);
+		const BitMask ext_block_mask = PEXT(occ, rank_blocker_lookup[king]);
 		BitMask check, pin;
 		std::tie(check, pin) = slider_cnp_table[rook % 8][king % 8][ext_block_mask];
-		h_check_mask |= _pdep_u64(check, rank);
-		hv_pin |= _pdep_u64(pin, rank);
+		h_check_mask |= PDEP(check, rank);
+		hv_pin |= PDEP(pin, rank);
 	}
 
 	const BitMask file = file_lookup[king];
 	Bitloop(file & rooks, loop_var){
 		const Square rook = SquareOf(loop_var);
-		const BitMask ext_block_mask = _pext_u64(occ, file_blocker_lookup[king]);
+		const BitMask ext_block_mask = PEXT(occ, file_blocker_lookup[king]);
 		BitMask check, pin;
 		std::tie(check, pin) = slider_cnp_table[rook / 8][king / 8][ext_block_mask];
-		v_check_mask |= _pdep_u64(check, file);
-		hv_pin |= _pdep_u64(pin, file);
+		v_check_mask |= PDEP(check, file);
+		hv_pin |= PDEP(pin, file);
 	}
 
 	BitMask check_mask = (h_check_mask ? h_check_mask : FULL_BOARD) &
@@ -142,25 +143,25 @@ std::tuple<BitMask, BitMask> bishop_checks_and_pins(
 	const BitMask upper_diag = upper_diag_lookup[king];
 	Bitloop(upper_diag & bishops, loop_var){
 		const Square bishop = SquareOf(loop_var);
-		const BitMask ext_block_mask = _pext_u64(occ, upper_blocker_lookup[king]);
+		const BitMask ext_block_mask = PEXT(occ, upper_blocker_lookup[king]);
 		const auto k_index = index_on_upper(king);
 		const auto b_index = index_on_upper(bishop);
 		BitMask check, pin;
 		std::tie(check, pin) = slider_cnp_table[b_index][k_index][ext_block_mask];
-		check_mask |= _pdep_u64(check, upper_diag);
-		diag_pin |= _pdep_u64(pin, upper_diag);
+		check_mask |= PDEP(check, upper_diag);
+		diag_pin |= PDEP(pin, upper_diag);
 	}
 
 	const BitMask lower_diag = lower_diag_lookup[king];
 	Bitloop(lower_diag & bishops, loop_var){
 		const Square bishop = SquareOf(loop_var);
-		const BitMask ext_block_mask = _pext_u64(occ, lower_blocker_lookup[king]);
+		const BitMask ext_block_mask = PEXT(occ, lower_blocker_lookup[king]);
 		const auto k_index = index_on_lower(king);
 		const auto b_index = index_on_lower(bishop);
 		BitMask check, pin;
 		std::tie(check, pin) = slider_cnp_table[b_index][k_index][ext_block_mask];
-		check_mask |= _pdep_u64(check, lower_diag);
-		diag_pin |= _pdep_u64(pin, lower_diag);
+		check_mask |= PDEP(check, lower_diag);
+		diag_pin |= PDEP(pin, lower_diag);
 	}
 
 	return std::make_tuple(check_mask ? check_mask : FULL_BOARD, diag_pin);