#13629: sqrt should use machine instruction on x86_64 -------------------------------------+------------------------------------- Reporter: bgamari | Owner: (none) Type: bug | Status: new Priority: normal | Milestone: 8.4.1 Component: Compiler (NCG) | Version: 8.0.1 Resolution: | Keywords: Operating System: Unknown/Multiple | Architecture: | Unknown/Multiple Type of failure: Runtime | Test Case: performance bug | numeric/num009 Blocked By: | Blocking: Related Tickets: #13570 | Differential Rev(s): Phab:D3508 Wiki Page: | -------------------------------------+------------------------------------- Comment (by bgamari): I've reproduced the regression locally. While it's hard to measure the runtime difference with the default test configuration, if you increase the argument from 3000 to 10000 it becomes quite clear. Before the patch the runtime is 1.14 seconds, after it's 1.30 seconds. This is quite surprising since there are a few bits of the generated code that get rather significantly shorter. Namely, {{{#!asm ... subq $8,%rsp movsd %xmm0,%xmm4 movss %xmm6,%xmm0 mulss %xmm6,%xmm0 mulss %xmm6,%xmm0 movq %rax,%rbx movl $1,%eax movq %rcx,%r14 movq %rdx,72(%rsp) movq %rsi,80(%rsp) movq %rdi,88(%rsp) movsd %xmm4,96(%rsp) movsd %xmm1,104(%rsp) movsd %xmm2,112(%rsp) movsd %xmm3,120(%rsp) call sqrtf addq $8,%rsp movss _n8uc(%rip),%xmm1 divss %xmm0,%xmm1 movss %xmm1,%xmm0 movsd 112(%rsp),%xmm2 mulss %xmm2,%xmm0 movss %xmm1,%xmm2 movsd 104(%rsp),%xmm3 mulss %xmm3,%xmm2 movsd 96(%rsp),%xmm3 mulss %xmm3,%xmm1 movq 64(%rsp),%rax leaq 1(%rax),%rcx }}} Whereas after we get, {{{#!asm ... movss %xmm6,%xmm4 mulss %xmm6,%xmm4 mulss %xmm6,%xmm4 sqrtss %xmm4,%xmm4 movss _n8ud(%rip),%xmm5 divss %xmm4,%xmm5 movss %xmm5,%xmm4 mulss %xmm3,%xmm4 movss %xmm5,%xmm3 mulss %xmm2,%xmm3 mulss %xmm1,%xmm5 leaq 1(%rdx),%rbx }}} -- Ticket URL: http://ghc.haskell.org/trac/ghc/ticket/13629#comment:16 GHC http://www.haskell.org/ghc/ The Glasgow Haskell Compiler