Weird C optimization killer at the matmult benchmark

[gabrielsferre]

At matmul's benchmark main function, m.matmul, moving the CheckGC() intermediate representation command may cause a drop in performance of 36%. It seems like moving this command kills some important C optimization, but it's not clear what optimization or why this happens.

This caught my eye while implementing a compiler optimization that would reduce a basic block's number of CheckGCs by grouping them all at the end of the block.

Currently, the first basic block of m.matmul looks like this:

function m.matmul(x1, x2): x3 {
  1:
    x4 <- NewArr(0)
    CheckGC()
    x5 <- #x1
    x6 <- #x2
    x8 <- x2[1]
    x7 <- #x8
    x9, x12, x11, x10 <- ForPrep(1, x5, 1)
    jmpIf x12, 2, 5

If we move CheckGC() to the end of the block, just before jmpIf, the runtime of the benchmark increases considerably.

function m.matmul(x1, x2): x3 {
  1:
    x4 <- NewArr(0)
    x5 <- #x1
    x6 <- #x2
    RenormArr(x2, 1)
    x8 <- x2[1]
    x7 <- #x8
    x9, x12, x11, x10 <- ForPrep(1, x5, 1)
    CheckGC()
    jmpIf x12, 2, 5

Right after this block, the function enters it's first loop. I seems like moving CheckGC() is disturbing the C compiler's optimizer (I'm using GCC v.15.1.1 with the -O3 flag). Furthermore, if we change m.matmul to use a while--loop instead of a for--loop as the first loop, moving CheckGC no longer causes this problem.