iliekturtles / llvm-opt

llvm-opt

Given a struct struct P<D> { value: f32, d: PhantomData<D>, } that implements std::ops::Div why does llvm not optimize the struct as an f32 for a simple function to divide values?

Given a raw f32 or a struct struct V { value: f32, } the function is optimized. The f32 and struct V variants generate the same llvm-ir and asm. The struct P<D> has a return type of i32 and an extra bitcast instruction in the llvm-ir and an extra movd instruction in the asm.

Code to reproduce llvm-ir and asm is included in the repository. Run build.sh to generate .ll and .s files in the target/release/examples/ folder. Output below generated using rustc 1.13.0 (2c6933acc 2016-11-07).

f32 function implemention. struct V and struct P<D> implementations are nearly identical:

fn calc(v: f32) -> f32 {
    let a = v * 1.234_f32;
    let b = 5.0_f32;

    a / b
}

f32:

; Function Attrs: noinline norecurse nounwind readnone uwtable
define internal fastcc float @_ZN7opt_f324calc17h83aad1073897de13E(float) unnamed_addr #2 {
entry-block:
  %1 = fmul float %0, 0x3FF3BE76C0000000
  %2 = fdiv float %1, 5.000000e+00
  ret float %2

	.seh_endprologue
	mulss	__real@3f9df3b6(%rip), %xmm0
	divss	__real@40a00000(%rip), %xmm0
	retq
	.seh_handlerdata

struct V:

; Function Attrs: noinline norecurse nounwind readnone uwtable
define internal fastcc float @_ZN7opt_v324calc17hbc6edc82fc0fb0dcE(float) unnamed_addr #2 {
entry-block:
  %1 = fmul float %0, 0x3FF3BE76C0000000
  %2 = fdiv float %1, 5.000000e+00
  ret float %2

	.seh_endprologue
	mulss	__real@3f9df3b6(%rip), %xmm0
	divss	__real@40a00000(%rip), %xmm0
	retq
	.seh_handlerdata

struct P<D>:

; Function Attrs: noinline norecurse nounwind readnone uwtable
define internal fastcc i32 @_ZN7opt_p324calc17h0c0275d6390115c6E(float) unnamed_addr #2 {
entry-block:
  %1 = fmul float %0, 0x3FF3BE76C0000000
  %2 = fdiv float %1, 5.000000e+00
  %3 = bitcast float %2 to i32
  ret i32 %3

	.seh_endprologue
	mulss	__real@3f9df3b6(%rip), %xmm0
	divss	__real@40a00000(%rip), %xmm0
	movd	%xmm0, %eax
	retq
	.seh_handlerdata