If the scales were flawed, it's in defense of -fomit-frame-pointer: according to his bench, it reduces the cycle count by ~0.3% in 32b, but blows it up by nearly 30% (29.7) in 64b.

I have put both graphs on the same image and scaled them correctly, then extended the canvas to the whole scale and finally shrunk the graph back to the original 600px high, this is the result: http://imgur.com/yG3R9

64b on the left, 32 on the right, blue is without -fomit and red is with it. On the 32b graph, you can barely discriminate between with and without, whereas on the 64b graph you can very clearly see it.

If he lied, his fault is to have dismissed his own findings as less important than they are.

thanks man. i've also updated the graphs on my blog.

Still a small issue: the graphs aren't the same scale (64b goes 0-5; 32b goes 0-4)

grossly exaggerated? bro, i gave the raw numbers in a table above the graph. if i were trying to lie i wouldn't have given the numbers.

i took the graphs as-is from goog docs and put em in my blog post.

i just zero-based the vertical scale. you can close your book about lying with stats or whatever the shit.