IBM POWERPC 750GL - Manual (Page 234)

User’s Manual
IBM PowerPC 750GX and 750GL RISC Microprocessor
Instruction Timing
Page 234 of 377
gx_06.fm.(1.2)
March 27, 2006
6.4.7 Integer Store Gathering
The 750GX performs store gathering for write-through operations to nonguarded space. It performs cache-
inhibited stores to nonguarded space for 4-byte, word-aligned stores. These stores are combined in the LSU
to form a double word and are sent out on the 60x bus as a single-beat operation. However, stores are gath-
ered only if the successive stores meet the criteria and are queued and pending. Store gathering occurs
regardless of the address order of the stores. Store gathering is enabled by setting HID0[SGE]. Stores can be
gathered in both endian modes.
Store gathering is not done for:
• Cacheable store operations
• Stores to guarded cache-inhibited or write-through space
• Byte-reverse store operations
• Store Word Conditional Indexed (stwcx.) instructions
• External Control Out Word Indexed (ecowx) instructions
• A store that occurs during a table-search operation
• Floating-point store operations
If store gathering is enabled and the stores do not fall under the above categories, an Enforce In-Order
Execution of I/O (eieio) or Synchronize (sync) instruction must be used to prevent two stores from being
gathered.
6.4.8 System Register Unit Execution Timing
Most instructions executed by the SRU either directly access renamed registers or access or modify
nonrenamed registers. They generally execute in a serial manner. Results from these instructions are not
available to subsequent instructions until the instruction completes and is retired. See Section 6.3.2.7,
Instruction Serialization for more information on serializing instructions executed by the SRU, and see
Table 6-5 on page 238 and Table 6-6 on page 240 for SRU instruction execution timings.
Floating-Point
8 byte
8 Optimal — — —
4 — Good Good Good
< 4 — Poor Poor Poor
4 byte
4 Optimal — — —
< 4 Poor Poor Poor Poor
Table 6-2. Performance Effects of Memory Operand Placement (Page 2 of 2)
Operand Boundary Crossing
Size Byte Alignment None 8 Byte Cache Block
Protection
Boundary
Note:
1. Optimal means one EA calculation occurs.
2. Good means multiple EA calculations occur that might cause additional bus activities with multiple bus transfers.
3. Not supported in little-endian mode; causes an alignment exception.
4. Poor means that an alignment exception occurs.