Enabling globalMerge on external causes regression

[sushgokh]

We are able to see regression around 7% with the MultiSource/Benchmarks/VersaBench/8b10b/8b10b.test in llvm-test-suite post merging #158592

Here is the short reproducer of the same.

@davemgreen just fyi. I havent analysed this in detail.

[davemgreen]

It certainly looks better, if I have that right?

It might be worth making sure this isn't a noisy benchmark - either from a noise from the system or noise from a different binary layout.

When I ran benchmarks for #158592 it looks like this case didn't come up, probably because it was filtered out because the binary did not change. That might have been because it was ran with or without LTO.

[sushgokh]

Its definitely not noise from the machine.

Benchmark is not noisy. We have very stable performance values over a long time period.

Check this. The bad code has add x8, x9, w8, uxtw #2 that is delaying the second load. So, in the first iteration itself, we have this difference of 1 cycle.

Sorry, I forgot the mention the name of the function. Its unsigned int calc(unsigned int theWord, unsigned int k) residing in 8b10b/calc.c. The function is being called 1024 times in loop and as you can see, for 100 iterations of mca, the difference is ~200 cycles.

Instead of bad code like below:

        lsl     w8, w1, #5
        adrp    x9, .L_MergedGlobals
        add     x9, x9, :lo12:.L_MergedGlobals
        bfxil   w8, w0, #3, #5
        bfi     w0, w1, #3, #29
        add     x8, x9, w8, uxtw #2             ---------------(1)
        ldr     w10, [x9, w0, uxtw #2]
        ldr     w8, [x8, #256]                    --------------(2)
        add     w0, w8, w10

can we change the order of calculations like below:

        lsl     w8, w1, #5
        adrp    x9, .L_MergedGlobals
        add     x9, x9, :lo12:.L_MergedGlobals
        bfxil   w8, w0, #3, #5
        bfi     w0, w1, #3, #29
        add     x8, x9, #256                             ------------------(1)
        ldr     w10, [x9, w0, uxtw #2]
        ldr     w8, [x8, w8, uxtw #2]                   -------------(2)
        add     w0, w8, w10

The modified version has same no of cycles as the good version. Check here