⚡ Bolt: [performance improvement] Replace ifelse with vectorized subsetting in llcont

.jules/bolt.md

@@ -12,3 +12,6 @@
 ## 2026-07-14 - Matrix Cross Product Optimization
 **Learning:** In R, matrix multiplication of the form `t(X) %*% Y` explicitly allocates memory for the transposed matrix. Using the optimized base function `crossprod(X, Y)` avoids this allocation.
 **Action:** Always replace `t(X) %*% Y` with `crossprod(X, Y)` for faster and more memory-efficient cross-product calculations.
+## 2026-07-16 - Optimize ifelse overhead in R
+**Learning:** In R, the `ifelse()` function evaluates both the true and false branches entirely before subsetting, which causes unnecessary calculations and memory allocations, especially for computationally expensive operations like `dnbinom()` or vector logs.
+**Action:** Replace `ifelse(cond, true_val, false_val)` with preallocation (`res <- numeric(n)`) and vectorized subsetting (`if (any(cond)) res[cond] <- ...`) to skip evaluating the expensive branch on irrelevant elements.

R/llcont.R

@@ -147,36 +147,56 @@ llcont.hurdle <- function(x, ...) {
   linkinv <- x$linkinv
 
   ## individual LL functions
+  ## Bolt: replaced ifelse() with vectorized subsetting for performance to avoid unnecessary calculations on all elements
   zeroPoisson <- function(parms) {
     mu <- as.vector(exp(Z %*% parms + offsetz))
     loglik0 <- -mu
-    Y0 * weights * loglik0 + ifelse(Y1, weights * log(1 - exp(loglik0)), 0)
+    res <- Y0 * weights * loglik0
+    if (any(Y1)) {
+      w <- if(length(weights) == 1) weights else weights[Y1]
+      res[Y1] <- res[Y1] + w * log(1 - exp(loglik0[Y1]))
+    }
+    res
   }
 
   countPoisson <- function(parms) {
     mu <- Y1 * as.vector(exp(X %*% parms + offsetx))
     loglik0 <- -mu
-    loglik1 <- Y1 * dpois(Y, lambda = mu, log = TRUE)
-    Y1 * weights * loglik1 - ifelse(Y1, weights * log(1 - exp(loglik0)), 0)
+    res <- Y * 0
+    if (any(Y1)) {
+      w <- if(length(weights) == 1) weights else weights[Y1]
+      loglik1 <- dpois(Y[Y1], lambda = mu[Y1], log = TRUE)
+      res[Y1] <- w * loglik1 - w * log(1 - exp(loglik0[Y1]))
+    }
+    res
   }
 
   zeroNegBin <- function(parms) {
     mu <- as.vector(exp(Z %*% parms[1:kz] + offsetz))
     theta <- exp(parms[kz + 1])
     loglik0 <- suppressWarnings(dnbinom(0, size = theta,
                                         mu = mu, log = TRUE))
-    Y0 * weights * loglik0 +
-        ifelse(Y1, weights * log(1 - exp(loglik0)), 0)
+    res <- Y0 * weights * loglik0
+    if (any(Y1)) {
+      w <- if(length(weights) == 1) weights else weights[Y1]
+      res[Y1] <- res[Y1] + w * log(1 - exp(loglik0[Y1]))
+    }
+    res
   }
 
   countNegBin <- function(parms) {
     mu <- as.vector(exp(X %*% parms[1:kx] + offsetx))
     theta <- exp(parms[kx + 1])
     loglik0 <- suppressWarnings(dnbinom(0, size = theta,
                                         mu = mu, log = TRUE))
-    loglik1 <- suppressWarnings(dnbinom(Y, size = theta,
-                                        mu = mu, log = TRUE))
-    ifelse(Y1, weights * loglik1 - weights * log(1 - exp(loglik0)), 0)
+    res <- Y * 0
+    if (any(Y1)) {
+      w <- if(length(weights) == 1) weights else weights[Y1]
+      loglik1 <- suppressWarnings(dnbinom(Y[Y1], size = theta,
+                                          mu = mu[Y1], log = TRUE))
+      res[Y1] <- w * loglik1 - w * log(1 - exp(loglik0[Y1]))
+    }
+    res
     }
 
   zeroGeom <- function(parms) zeroNegBin(c(parms, 0))