[llvm9 regression] i586-unknown-linux-gnu does not generate movmskps anymore

[gnzlbg]

This does looks like a regression with the LLVM9 update @nikic . The _mm_movemask_ps intrinsic used to generate a single movmskps instruction with LLVM8 on the i586-unknown-linux-gnu target, and now it generates:

---- core_arch::x86::sse::assert__mm_movemask_ps_movmskps stdout ----
disassembly for stdarch_test_shim__mm_movemask_ps_movmskps: 
	 0: sub $0x1c,%esp
	 1: movaps %xmm0,(%esp)
	 2: mov $0x811698e,%eax
	 3: mov %eax,0x81cd008
	 4: cmpl $0x0,(%esp)
	 5: sets %al
	 6: cmpl $0x0,0x4(%esp)
	 7: sets %cl
	 8: add %cl,%cl
	 9: or %al,%cl
	10: cmpl $0x0,0x8(%esp)
	11: sets %al
	12: cmpl $0x0,0xc(%esp)
	13: sets %dl
	14: add %dl,%dl
	15: or %al,%dl
	16: shl $0x2,%dl
	17: or %cl,%dl
	18: movzbl %dl,%eax
	19: add $0x1c,%esp
	20: ret
	21: xchg %ax,%ax
	22: xchg %ax,%ax
	23: xchg %ax,%ax
	24: xchg %ax,%ax
	25: xchg %ax,%ax
	26: xchg %ax,%ax
thread 'core_arch::x86::sse::assert__mm_movemask_ps_movmskps' panicked at 'failed to find instruction `movmskps` in the disassembly', crates/stdarch-test/src/lib.rs:152:9

[gnzlbg]

Rust nightly MWE: https://godbolt.org/z/fA5UNl

Using the generated LLVM-IR: https://godbolt.org/z/LU7i9q

source_filename = "example.3a1fbbbh-cgu.0"
target datalayout = "e-m:e-p:32:32-f64:32:64-f80:32-n8:16:32-S128"
target triple = "i586-unknown-linux-gnu"

define i32 @_ZN7example15_mm_movemask_ps17h7f545271133b7928E(<4 x float>* noalias nocapture readonly dereferenceable(16) %a) unnamed_addr #0 {
start:
  %0 = bitcast <4 x float>* %a to <4 x i32>*
  %1 = load <4 x i32>, <4 x i32>* %0, align 16
  %2 = icmp slt <4 x i32> %1, zeroinitializer
  %3 = bitcast <4 x i1> %2 to i4
  %4 = zext i4 %3 to i32
  ret i32 %4
}

attributes #0 = { norecurse nounwind nonlazybind readonly "probe-stack"="__rust_probestack" "target-cpu"="pentium" "target-features"="+sse" }

!llvm.module.flags = !{!0}

!0 = !{i32 2, !"RtLibUseGOT", i32 1}

I cannot generate the old machine code using any llc version. So maybe this is unrelated to the LLVM8 upgrade, and we changed something somewhere about the i586-unknown-linux-gnu target ?

[gnzlbg]

Using Rust 1.36.0: https://godbolt.org/z/k22p47

does generate

example::_mm_movemask_ps:
        mov     eax, dword ptr [esp + 4]
        movaps  xmm0, xmmword ptr [eax]
        movmskps        eax, xmm0
        ret
.Lfunc_end0:

and the LLVM-IR generated: https://godbolt.org/z/oMW4OK

is

source_filename = "example.3a1fbbbh-cgu.0"
target datalayout = "e-m:e-p:32:32-f64:32:64-f80:32-n8:16:32-S128"
target triple = "i586-unknown-linux-gnu"

define i32 @_ZN7example15_mm_movemask_ps17h9d7ca884d8f840c4E(<4 x float>* noalias nocapture readonly dereferenceable(16) %a) unnamed_addr #0 {
start:
  %0 = load <4 x float>, <4 x float>* %a, align 16
  %1 = tail call i32 @llvm.x86.sse.movmsk.ps(<4 x float> %0) #2
  ret i32 %1
}

declare i32 @llvm.x86.sse.movmsk.ps(<4 x float>) unnamed_addr #1

attributes #0 = { nounwind nonlazybind readonly "probe-stack"="__rust_probestack" "target-cpu"="pentium" "target-features"="+sse" }
attributes #1 = { nounwind readnone }
attributes #2 = { nounwind }

!llvm.module.flags = !{!0}

!0 = !{i32 2, !"RtLibUseGOT", i32 1}

notice how the @llvm.x86.sse.movmsk.ps is not transformed like it is transformed with LLVM 9.

[gnzlbg]

Found the bug. The bug is in opt: https://godbolt.org/z/vVVDFd

[gnzlbg]

Opened LLVM bug: https://bugs.llvm.org/show_bug.cgi?id=42870

[Nugine]

The LLVM bug was resolved on 2020-03-24.

I find this issue from:

stdarch/crates/core_arch/src/x86/sse.rs

Lines 1081 to 1090 in b168a4c

	#[inline]
	#[target_feature(enable = "sse")]
	// FIXME: LLVM9 trunk has the following bug:
	// https://github.com/rust-lang/stdarch/issues/794
	// so we only temporarily test this on i686 and x86_64 but not on i586:
	#[cfg_attr(all(test, target_feature = "sse2"), assert_instr(movmskps))]
	#[stable(feature = "simd_x86", since = "1.27.0")]
	pub unsafe fn _mm_movemask_ps(a: __m128) -> i32 {
	movmskps(a)
	}