fix --ld bug introduced in previous build

2.0/include/pgenlib_misc.cc

@@ -2837,6 +2837,16 @@ void ClearGenoarrMissing1bit16Unsafe(const uintptr_t* __restrict genoarr, uint32
   }
 }
 
+double u127prod_diff_d(uint64_t plus_term0, uint64_t plus_term1, uint64_t minus_term0, uint64_t minus_term1) {
+  uint64_t plus_hi;
+  const uint64_t plus_lo = multiply64to128(plus_term0, plus_term1, &plus_hi);
+  uint64_t minus_hi;
+  const uint64_t minus_lo = multiply64to128(minus_term0, minus_term1, &minus_hi);
+  const uint64_t result_lo = plus_lo - minus_lo;
+  const uint64_t result_hi = plus_hi - minus_hi - (plus_lo < minus_lo);
+  return u127tod(result_hi, result_lo);
+}
+
 double MultiallelicDiploidMinimac3R2(const uint64_t* __restrict sums, const uint64_t* __restrict hap_ssqs_x2, uint32_t nm_sample_ct, uint32_t allele_ct, uint32_t extra_phased_het_ct) {
   // sums[k] == sum_i [left_dosage_{ik} + right_dosage_{ik}]
   // hap_ssqs_x2[k] ==

2.0/include/pgenlib_misc.h

@@ -241,11 +241,10 @@ HEADER_INLINE void SetTrailingNyps(uintptr_t nyp_ct, uintptr_t* bitarr) {
 // GetVint31 and Vint32Append moved to plink2_base.
 
 // Input must be validated, or bufp must be >= 5 characters before the end of
-// the read buffer.  Currently unused.
+// the read buffer.
 // todo: check if this has enough of a speed advantage over GetVint31() to
 // justify using this in the main loops and catching SIGSEGV.  (seems to be no
 // more than 3%?)
-/*
 HEADER_INLINE uint32_t GetVint31Unsafe(const unsigned char** buf_iterp) {
   uint32_t vint32 = *(*buf_iterp)++;
   if (vint32 <= 127) {
@@ -261,7 +260,13 @@ HEADER_INLINE uint32_t GetVint31Unsafe(const unsigned char** buf_iterp) {
   }
   return 0x80000000U;
 }
-*/
+
+HEADER_INLINE void SkipVintUnsafe(const unsigned char** buf_iterp) {
+  uint32_t cur_byte;
+  do {
+    cur_byte = *(*buf_iterp)++;
+  } while (cur_byte & 128);
+}
 
 // Does not update buf_iter.
 HEADER_INLINE uint32_t PeekVint31(const unsigned char* buf_iter, const unsigned char* buf_end) {
@@ -562,6 +567,9 @@ HEADER_INLINE double u127tod(uint64_t hi, uint64_t lo) {
   return u63tod(hi) * 18446744073709551616.0 + S_CAST(double, lo);
 }
 
+// plus_term0 * plus_term1 - minus_term0 * minus_term1
+double u127prod_diff_d(uint64_t plus_term0, uint64_t plus_term1, uint64_t minus_term0, uint64_t minus_term1);
+
 double MultiallelicDiploidMinimac3R2(const uint64_t* __restrict sums, const uint64_t* __restrict hap_ssqs_x2, uint32_t nm_sample_ct, uint32_t allele_ct, uint32_t extra_phased_het_ct);
 
 HEADER_INLINE double MultiallelicDiploidMachR2(const uint64_t* __restrict sums, const uint64_t* __restrict ssqs, uint32_t nm_sample_ct, uint32_t allele_ct) {

2.0/include/plink2_base.cc

@@ -681,6 +681,16 @@ uintptr_t FirstUnequalW(const void* arr1, const void* arr2, uintptr_t nbytes) {
 }
 #endif
 
+// TODO: obvious movemask optimization
+uintptr_t CountVints(const unsigned char* buf, const unsigned char* buf_end) {
+  const uintptr_t len = buf_end - buf;
+  uintptr_t inv_result = 0;
+  for (uintptr_t ulii = 0; ulii != len; ++ulii) {
+    inv_result += buf[ulii] >> 7;
+  }
+  return len - inv_result;
+}
+
 #ifdef __cplusplus
 }  // namespace plink2
 #endif

2.0/include/plink2_base.h

@@ -3991,6 +3991,8 @@ HEADER_INLINE uint32_t GetVint31(const unsigned char* buf_end, const unsigned ch
   return 0x80000000U;
 }
 
+uintptr_t CountVints(const unsigned char* buf, const unsigned char* buf_end);
+
 // Flagset conventions:
 // * Each 32-bit and 64-bit flagset has its own type, which is guaranteed to be
 //   the appropriate width.  (Todo: verify that bit 31 works properly in 32-bit

2.0/plink2.cc

@@ -72,7 +72,7 @@ static const char ver_str[] = "PLINK v2.00a5"
 #elif defined(USE_AOCL)
   " AMD"
 #endif
-  " (11 Sep 2023)";
+  " (13 Sep 2023)";
 static const char ver_str2[] =
   // include leading space if day < 10, so character length stays the same
   ""
@@ -4725,14 +4725,21 @@ int main(int argc, char** argv) {
             logerrputs("Error: --clump-bins does not make sense when --clump 'bins' column set has been\nexcluded.\n");
             goto main_ret_INVALID_CMDLINE_A;
           }
-          if (unlikely(EnforceParamCtRange(argvk[arg_idx], param_ct, 1, 0x7fffffff))) {
+          // may as well enforce limit of 2^26 bin-bounds.  this should never
+          // come up, and it lets us remove some bounds-checks.
+          if (unlikely(EnforceParamCtRange(argvk[arg_idx], param_ct, 1, kClumpMaxBinBounds))) {
             goto main_ret_INVALID_CMDLINE_2A;
           }
           uint32_t comma_ct = 0;
           for (uint32_t param_idx = 1; param_idx <= param_ct; ++param_idx) {
             comma_ct += CountByteInStr(argvk[arg_idx], ',');
           }
           const uint32_t bin_bound_ct = comma_ct + param_ct;
+          if (bin_bound_ct > kClumpMaxBinBounds) {
+            logerrputs("Error: --clump-bins is currently limited to 2^26 bin boundaries.\n");
+            reterr = kPglRetNotYetSupported;
+            goto main_ret_1;
+          }
           pc.clump_info.bin_bound_ct = bin_bound_ct;
           if (unlikely(pgl_malloc(sizeof(double) * bin_bound_ct, &pc.clump_info.ln_bin_boundaries))) {
             goto main_ret_NOMEM;

2.0/plink2_cmdline.cc

@@ -474,7 +474,7 @@ uintptr_t CountSortedLeqU64(const uint64_t* sorted_u64_arr, uintptr_t arr_length
 
 uintptr_t IdxToUidxW(const uintptr_t* bitvec, const uintptr_t* cumulative_popcounts, uintptr_t widx_start, uintptr_t widx_end, uintptr_t idx) {
   if (widx_end > widx_start) {
-    widx_start += CountSortedSmallerW(&(cumulative_popcounts[widx_start]), widx_end - widx_start, idx + 1);
+    widx_start += CountSortedSmallerW(&(cumulative_popcounts[widx_start]), widx_end - widx_start, idx + 1) - 1;
   }
   const uintptr_t idx_at_widx_start = cumulative_popcounts[widx_start];
   return widx_start * kBitsPerWord + WordBitIdxToUidx(bitvec[widx_start], idx - idx_at_widx_start);
@@ -483,7 +483,7 @@ uintptr_t IdxToUidxW(const uintptr_t* bitvec, const uintptr_t* cumulative_popcou
 uintptr_t ExpsearchIdxToUidxW(const uintptr_t* bitvec, const uintptr_t* cumulative_popcounts, uintptr_t widx_end, uintptr_t idx, uintptr_t* widx_startp) {
   uintptr_t widx_start = *widx_startp;
   if (widx_end > widx_start) {
-    widx_start += ExpsearchW(&(cumulative_popcounts[widx_start]), widx_end - widx_start, idx + 1);
+    widx_start += ExpsearchW(&(cumulative_popcounts[widx_start]), widx_end - widx_start, idx + 1) - 1;
     *widx_startp = widx_start;
   }
   const uintptr_t idx_at_widx_start = cumulative_popcounts[widx_start];

2.0/plink2_common.cc

@@ -3208,7 +3208,6 @@ PglErr ParseChrRanges(const char* const* argvk, const char* flagname_p, const ch
   return reterr;
 }
 
-/*
 uint32_t MultiallelicVariantPresent(const uintptr_t* variant_include, const uintptr_t* allele_idx_offsets, uint32_t variant_ct) {
   if (!allele_idx_offsets) {
     return 0;
@@ -3223,7 +3222,6 @@ uint32_t MultiallelicVariantPresent(const uintptr_t* variant_include, const uint
   }
   return 0;
 }
-*/
 
 uint32_t CountBiallelicVariants(const uintptr_t* variant_include, const uintptr_t* allele_idx_offsets, uint32_t variant_ct) {
   if (!allele_idx_offsets) {

2.0/plink2_common.h

@@ -1153,7 +1153,7 @@ void CleanupPhenoCols(uint32_t pheno_ct, PhenoCol* pheno_cols);
 PglErr ParseChrRanges(const char* const* argvk, const char* flagname_p, const char* errstr_append, uint32_t param_ct, uint32_t prohibit_extra_chrs, uint32_t xymt_subtract, char range_delim, ChrInfo* cip, uintptr_t* chr_mask);
 
 
-// uint32_t MultiallelicVariantPresent(const uintptr_t* variant_include, const uintptr_t* allele_idx_offsets, uint32_t variant_ct);
+uint32_t MultiallelicVariantPresent(const uintptr_t* variant_include, const uintptr_t* allele_idx_offsets, uint32_t variant_ct);
 
 uint32_t CountBiallelicVariants(const uintptr_t* variant_include, const uintptr_t* allele_idx_offsets, uint32_t variant_ct);