diff --git a/src/xenia/base/math.h b/src/xenia/base/math.h
index 824a1d746..662ceced7 100644
--- a/src/xenia/base/math.h
+++ b/src/xenia/base/math.h
@@ -64,6 +64,22 @@ constexpr uint32_t select_bits(uint32_t value, uint32_t a, uint32_t b) {
   return (value & make_bitmask(a, b)) >> a;
 }
 
+inline uint32_t bit_count(uint32_t v) {
+  v = v - ((v >> 1) & 0x55555555);
+  v = (v & 0x33333333) + ((v >> 2) & 0x33333333);
+  return ((v + (v >> 4) & 0xF0F0F0F) * 0x1010101) >> 24;
+}
+
+inline uint32_t bit_count(uint64_t v) {
+  v = (v & 0x5555555555555555LU) + (v >> 1 & 0x5555555555555555LU);
+  v = (v & 0x3333333333333333LU) + (v >> 2 & 0x3333333333333333LU);
+  v = v + (v >> 4) & 0x0F0F0F0F0F0F0F0FLU;
+  v = v + (v >> 8);
+  v = v + (v >> 16);
+  v = v + (v >> 32) & 0x0000007F;
+  return static_cast<uint32_t>(v);
+}
+
 // lzcnt instruction, typed for integers of all sizes.
 // The number of leading zero bits in the value parameter. If value is zero, the
 // return value is the size of the input operand (8, 16, 32, or 64). If the most
diff --git a/src/xenia/gpu/shader.h b/src/xenia/gpu/shader.h
index e3e4d7072..476369e53 100644
--- a/src/xenia/gpu/shader.h
+++ b/src/xenia/gpu/shader.h
@@ -504,8 +504,11 @@ class Shader {
     // Each bit corresponds to a storage index [0-31].
     uint32_t int_bitmap;
     // Bitmap of all kConstantBool registers read by the shader.
-    // Each bit corresponds to a storage index [0-31].
-    uint32_t bool_bitmap;
+    // Each bit corresponds to a storage index [0-255].
+    uint32_t bool_bitmap[256 / 32];
+
+    // Computed byte count of all registers required when packed.
+    uint32_t packed_byte_length;
   };
 
   Shader(ShaderType shader_type, uint64_t ucode_data_hash,
diff --git a/src/xenia/gpu/shader_translator.cc b/src/xenia/gpu/shader_translator.cc
index 68a70d5fb..1c9f31962 100644
--- a/src/xenia/gpu/shader_translator.cc
+++ b/src/xenia/gpu/shader_translator.cc
@@ -125,6 +125,24 @@ bool ShaderTranslator::Translate(Shader* shader) {
 
   TranslateBlocks();
 
+  // Compute total bytes used by the register map.
+  // This saves us work later when we need to pack them.
+  constant_register_map_.packed_byte_length = 0;
+  for (int i = 0; i < 4; ++i) {
+    // Each bit indicates a vec4 (4 floats).
+    constant_register_map_.packed_byte_length +=
+        4 * 4 * xe::bit_count(constant_register_map_.float_bitmap[i]);
+  }
+  // Each bit indicates a single word.
+  constant_register_map_.packed_byte_length +=
+      4 * xe::bit_count(constant_register_map_.int_bitmap);
+  // Direct map between words and words we upload.
+  for (int i = 0; i < 4; ++i) {
+    if (constant_register_map_.bool_bitmap[i]) {
+      constant_register_map_.packed_byte_length += 4;
+    }
+  }
+
   shader->errors_ = std::move(errors_);
   shader->translated_binary_ = CompleteTranslation();
   shader->ucode_disassembly_ = ucode_disasm_buffer_.to_string();
@@ -490,7 +508,8 @@ void ShaderTranslator::TranslateControlFlowCondExec(
   i.instruction_count = cf.count();
   i.type = ParsedExecInstruction::Type::kConditional;
   i.bool_constant_index = cf.bool_address();
-  constant_register_map_.bool_bitmap |= 1 << i.bool_constant_index;
+  constant_register_map_.bool_bitmap[i.bool_constant_index / 32] |=
+      1 << (i.bool_constant_index % 32);
   i.condition = cf.condition();
   switch (cf.opcode()) {
     case ControlFlowOpcode::kCondExec:
@@ -567,7 +586,8 @@ void ShaderTranslator::TranslateControlFlowCondCall(
   } else {
     i.type = ParsedCallInstruction::Type::kConditional;
     i.bool_constant_index = cf.bool_address();
-    constant_register_map_.bool_bitmap |= 1 << i.bool_constant_index;
+    constant_register_map_.bool_bitmap[i.bool_constant_index / 32] |=
+        1 << (i.bool_constant_index % 32);
     i.condition = cf.condition();
   }
 
@@ -599,7 +619,8 @@ void ShaderTranslator::TranslateControlFlowCondJmp(
   } else {
     i.type = ParsedJumpInstruction::Type::kConditional;
     i.bool_constant_index = cf.bool_address();
-    constant_register_map_.bool_bitmap |= 1 << i.bool_constant_index;
+    constant_register_map_.bool_bitmap[i.bool_constant_index / 32] |=
+        1 << (i.bool_constant_index % 32);
     i.condition = cf.condition();
   }