caio/clang-p2996
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
Files
f2e7285b03fbfc263c8a2eaa2b5e2e1cfafc6abd
clang-p2996/llvm/test/CodeGen/AMDGPU/rsq.f32.ll
Matt Arsenault 8287f3af9d AMDGPU: Overhaul and improve rcp and rsq f32 formation 
The highlight change is a new denormal safe 1ulp lowering which uses
rcp after using frexp to perform input scaling. This saves 2
instructions compared to other implementations which performed an
explicit denormal range change. This improves the OpenCL default, and
requires a flag for HIP. I don't believe there's any flag wired up for
OpenMP to emit the necessary fpmath metadata.

This provides several improvements and changes that were hard to
separate without regressing one case or another. Disturbingly the
OpenCL conformance test seems to have the reciprocal test commented
out. I locally hacked it back in to test this.

Starts introducing f32 rsq intrinsics in AMDGPUCodeGenPrepare. Like
the rcp case, we could do this in codegen if !fpmath were preserved
(although we would lose some computeKnownFPClass tricks). Start
requiring contract flags to form rsq. The rsq fusion actually improves
the result from ~2ulp to ~1ulp. We have some older fusion in codegen
which only keys off unsafe math which should be refined.

Expand rsq patterns by checking for denormal inputs and pre/post
multiplying like the current library code does. We also take advantage
of computeKnownFPClass to avoid the scaling when we can statically
prove the input cannot be a denormal. We could do the same for the rcp
case, but unlike rsq a large input can underflow to denormal. We need
additional upper bound exponent checks on the input in order to do the
same for rcp.

This rsq handling also now starts handling the negated case. We
introduce rsq with an fneg. In the case the fneg doesn't fold into its
user, it's a neutral change but provides improvement if it is foldable
as a source modifier.

Also starts respecting the arcp attribute properly, and more strictly
interprets afn. We were previously interpreting afn as implying you
could do the reciprocal expansion of an fdiv. The codegen handling of
these also needs to be revisited.

This also effectively introduces the optimization
combineRepeatedFPDivisors enables, just done in the IR instead (and
only for f32).

This is almost across the board better. The one minor regression is
for gfx6/buggy frexp case where for multiple reciprocals, we could
previously reuse rematerialized constants per instance (it's neutral
for a single rcp).

The fdiv.fast and sqrt handling need to be revisited next.

https://reviews.llvm.org/D155593
2023-07-21 16:35:53 -04:00

1420 lines
63 KiB
LLVM
Raw Blame History

; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 3
; RUN: llc -amdgpu-scalarize-global-loads=false -march=amdgcn -mcpu=tahiti -denormal-fp-math-f32=preserve-sign -enable-unsafe-fp-math < %s | FileCheck -check-prefixes=GCN-DAZ,GCN-DAZ-UNSAFE,SI-DAZ-UNSAFE %s
; RUN: llc -amdgpu-scalarize-global-loads=false -march=amdgcn -mcpu=tahiti -denormal-fp-math-f32=ieee -enable-unsafe-fp-math < %s | FileCheck -check-prefixes=GCN-IEEE,GCN-IEEE-UNSAFE,SI-IEEE-UNSAFE %s
; RUN: llc -amdgpu-scalarize-global-loads=false -march=amdgcn -mcpu=tahiti -denormal-fp-math-f32=preserve-sign < %s | FileCheck -check-prefixes=GCN-DAZ,GCN-DAZ-SAFE,SI-DAZ-SAFE %s
; RUN: llc -amdgpu-scalarize-global-loads=false -march=amdgcn -mcpu=tahiti -denormal-fp-math-f32=ieee < %s | FileCheck -check-prefixes=GCN-IEEE,GCN-IEEE-SAFE,SI-IEEE-SAFE %s
; RUN: llc -amdgpu-scalarize-global-loads=false -march=amdgcn -mcpu=hawaii -denormal-fp-math-f32=preserve-sign -enable-unsafe-fp-math < %s | FileCheck -check-prefixes=GCN-DAZ,GCN-DAZ-UNSAFE,CI-DAZ-UNSAFE %s
; RUN: llc -amdgpu-scalarize-global-loads=false -march=amdgcn -mcpu=hawaii -denormal-fp-math-f32=ieee -enable-unsafe-fp-math < %s | FileCheck -check-prefixes=GCN-IEEE,GCN-IEEE-UNSAFE,CI-IEEE-UNSAFE %s
; RUN: llc -amdgpu-scalarize-global-loads=false -march=amdgcn -mcpu=hawaii -denormal-fp-math-f32=preserve-sign < %s | FileCheck -check-prefixes=GCN-DAZ,GCN-DAZ-SAFE,CI-DAZ-SAFE %s
; RUN: llc -amdgpu-scalarize-global-loads=false -march=amdgcn -mcpu=hawaii -denormal-fp-math-f32=ieee < %s | FileCheck -check-prefixes=GCN-IEEE,GCN-IEEE-SAFE,CI-IEEE-SAFE %s
declare i32 @llvm.amdgcn.workitem.id.x() nounwind readnone
declare float @llvm.sqrt.f32(float) nounwind readnone
declare <2 x float> @llvm.sqrt.v2f32(<2 x float>) nounwind readnone
define amdgpu_kernel void @rsq_f32(ptr addrspace(1) noalias %out, ptr addrspace(1) noalias %in) {
; GCN-DAZ-LABEL: rsq_f32:
; GCN-DAZ: ; %bb.0:
; GCN-DAZ-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9
; GCN-DAZ-NEXT: s_mov_b32 s7, 0xf000
; GCN-DAZ-NEXT: s_mov_b32 s6, -1
; GCN-DAZ-NEXT: s_mov_b32 s10, s6
; GCN-DAZ-NEXT: s_mov_b32 s11, s7
; GCN-DAZ-NEXT: s_waitcnt lgkmcnt(0)
; GCN-DAZ-NEXT: s_mov_b32 s8, s2
; GCN-DAZ-NEXT: s_mov_b32 s9, s3
; GCN-DAZ-NEXT: buffer_load_dword v0, off, s[8:11], 0
; GCN-DAZ-NEXT: s_mov_b32 s4, s0
; GCN-DAZ-NEXT: s_mov_b32 s5, s1
; GCN-DAZ-NEXT: s_waitcnt vmcnt(0)
; GCN-DAZ-NEXT: v_rsq_f32_e32 v0, v0
; GCN-DAZ-NEXT: buffer_store_dword v0, off, s[4:7], 0
; GCN-DAZ-NEXT: s_endpgm
;
; GCN-IEEE-UNSAFE-LABEL: rsq_f32:
; GCN-IEEE-UNSAFE: ; %bb.0:
; GCN-IEEE-UNSAFE-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9
; GCN-IEEE-UNSAFE-NEXT: s_mov_b32 s7, 0xf000
; GCN-IEEE-UNSAFE-NEXT: s_mov_b32 s6, -1
; GCN-IEEE-UNSAFE-NEXT: s_mov_b32 s10, s6
; GCN-IEEE-UNSAFE-NEXT: s_mov_b32 s11, s7
; GCN-IEEE-UNSAFE-NEXT: s_waitcnt lgkmcnt(0)
; GCN-IEEE-UNSAFE-NEXT: s_mov_b32 s8, s2
; GCN-IEEE-UNSAFE-NEXT: s_mov_b32 s9, s3
; GCN-IEEE-UNSAFE-NEXT: buffer_load_dword v0, off, s[8:11], 0
; GCN-IEEE-UNSAFE-NEXT: s_mov_b32 s4, s0
; GCN-IEEE-UNSAFE-NEXT: s_mov_b32 s5, s1
; GCN-IEEE-UNSAFE-NEXT: s_waitcnt vmcnt(0)
; GCN-IEEE-UNSAFE-NEXT: v_rsq_f32_e32 v0, v0
; GCN-IEEE-UNSAFE-NEXT: buffer_store_dword v0, off, s[4:7], 0
; GCN-IEEE-UNSAFE-NEXT: s_endpgm
;
; SI-IEEE-SAFE-LABEL: rsq_f32:
; SI-IEEE-SAFE: ; %bb.0:
; SI-IEEE-SAFE-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9
; SI-IEEE-SAFE-NEXT: s_mov_b32 s7, 0xf000
; SI-IEEE-SAFE-NEXT: s_mov_b32 s6, -1
; SI-IEEE-SAFE-NEXT: s_mov_b32 s10, s6
; SI-IEEE-SAFE-NEXT: s_mov_b32 s11, s7
; SI-IEEE-SAFE-NEXT: s_waitcnt lgkmcnt(0)
; SI-IEEE-SAFE-NEXT: s_mov_b32 s8, s2
; SI-IEEE-SAFE-NEXT: s_mov_b32 s9, s3
; SI-IEEE-SAFE-NEXT: buffer_load_dword v0, off, s[8:11], 0
; SI-IEEE-SAFE-NEXT: s_mov_b32 s2, 0x7f800000
; SI-IEEE-SAFE-NEXT: s_mov_b32 s4, s0
; SI-IEEE-SAFE-NEXT: s_mov_b32 s5, s1
; SI-IEEE-SAFE-NEXT: s_waitcnt vmcnt(0)
; SI-IEEE-SAFE-NEXT: v_sqrt_f32_e32 v0, v0
; SI-IEEE-SAFE-NEXT: v_frexp_mant_f32_e32 v1, v0
; SI-IEEE-SAFE-NEXT: v_cmp_lt_f32_e64 vcc, |v0|, s2
; SI-IEEE-SAFE-NEXT: v_cndmask_b32_e32 v1, v0, v1, vcc
; SI-IEEE-SAFE-NEXT: v_rcp_f32_e32 v1, v1
; SI-IEEE-SAFE-NEXT: v_frexp_exp_i32_f32_e32 v0, v0
; SI-IEEE-SAFE-NEXT: v_sub_i32_e32 v0, vcc, 0, v0
; SI-IEEE-SAFE-NEXT: v_ldexp_f32_e32 v0, v1, v0
; SI-IEEE-SAFE-NEXT: buffer_store_dword v0, off, s[4:7], 0
; SI-IEEE-SAFE-NEXT: s_endpgm
;
; CI-IEEE-SAFE-LABEL: rsq_f32:
; CI-IEEE-SAFE: ; %bb.0:
; CI-IEEE-SAFE-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9
; CI-IEEE-SAFE-NEXT: s_mov_b32 s7, 0xf000
; CI-IEEE-SAFE-NEXT: s_mov_b32 s6, -1
; CI-IEEE-SAFE-NEXT: s_mov_b32 s10, s6
; CI-IEEE-SAFE-NEXT: s_mov_b32 s11, s7
; CI-IEEE-SAFE-NEXT: s_waitcnt lgkmcnt(0)
; CI-IEEE-SAFE-NEXT: s_mov_b32 s8, s2
; CI-IEEE-SAFE-NEXT: s_mov_b32 s9, s3
; CI-IEEE-SAFE-NEXT: buffer_load_dword v0, off, s[8:11], 0
; CI-IEEE-SAFE-NEXT: s_mov_b32 s4, s0
; CI-IEEE-SAFE-NEXT: s_mov_b32 s5, s1
; CI-IEEE-SAFE-NEXT: s_waitcnt vmcnt(0)
; CI-IEEE-SAFE-NEXT: v_sqrt_f32_e32 v0, v0
; CI-IEEE-SAFE-NEXT: v_frexp_mant_f32_e32 v1, v0
; CI-IEEE-SAFE-NEXT: v_rcp_f32_e32 v1, v1
; CI-IEEE-SAFE-NEXT: v_frexp_exp_i32_f32_e32 v0, v0
; CI-IEEE-SAFE-NEXT: v_sub_i32_e32 v0, vcc, 0, v0
; CI-IEEE-SAFE-NEXT: v_ldexp_f32_e32 v0, v1, v0
; CI-IEEE-SAFE-NEXT: buffer_store_dword v0, off, s[4:7], 0
; CI-IEEE-SAFE-NEXT: s_endpgm
; GCN-UNSAFE-LABEL: rsq_f32:
; GCN-UNSAFE: ; %bb.0:
; GCN-UNSAFE-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9
; GCN-UNSAFE-NEXT: s_mov_b32 s7, 0xf000
; GCN-UNSAFE-NEXT: s_mov_b32 s6, -1
; GCN-UNSAFE-NEXT: s_mov_b32 s10, s6
; GCN-UNSAFE-NEXT: s_mov_b32 s11, s7
; GCN-UNSAFE-NEXT: s_waitcnt lgkmcnt(0)
; GCN-UNSAFE-NEXT: s_mov_b32 s8, s2
; GCN-UNSAFE-NEXT: s_mov_b32 s9, s3
; GCN-UNSAFE-NEXT: buffer_load_dword v0, off, s[8:11], 0
; GCN-UNSAFE-NEXT: s_mov_b32 s4, s0
; GCN-UNSAFE-NEXT: s_mov_b32 s5, s1
; GCN-UNSAFE-NEXT: s_waitcnt vmcnt(0)
; GCN-UNSAFE-NEXT: v_rsq_f32_e32 v0, v0
; GCN-UNSAFE-NEXT: buffer_store_dword v0, off, s[4:7], 0
; GCN-UNSAFE-NEXT: s_endpgm
%val = load float, ptr addrspace(1) %in, align 4
%sqrt = call contract float @llvm.sqrt.f32(float %val) nounwind readnone
%div = fdiv contract float 1.0, %sqrt, !fpmath !0
store float %div, ptr addrspace(1) %out, align 4
ret void
}
define amdgpu_kernel void @rsq_f32_sgpr(ptr addrspace(1) noalias %out, float %val) {
; GCN-DAZ-LABEL: rsq_f32_sgpr:
; GCN-DAZ: ; %bb.0:
; GCN-DAZ-NEXT: s_load_dword s2, s[0:1], 0xb
; GCN-DAZ-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x9
; GCN-DAZ-NEXT: s_mov_b32 s3, 0xf000
; GCN-DAZ-NEXT: s_waitcnt lgkmcnt(0)
; GCN-DAZ-NEXT: v_rsq_f32_e32 v0, s2
; GCN-DAZ-NEXT: s_mov_b32 s2, -1
; GCN-DAZ-NEXT: buffer_store_dword v0, off, s[0:3], 0
; GCN-DAZ-NEXT: s_endpgm
;
; GCN-IEEE-UNSAFE-LABEL: rsq_f32_sgpr:
; GCN-IEEE-UNSAFE: ; %bb.0:
; GCN-IEEE-UNSAFE-NEXT: s_load_dword s2, s[0:1], 0xb
; GCN-IEEE-UNSAFE-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x9
; GCN-IEEE-UNSAFE-NEXT: s_mov_b32 s3, 0xf000
; GCN-IEEE-UNSAFE-NEXT: s_waitcnt lgkmcnt(0)
; GCN-IEEE-UNSAFE-NEXT: v_rsq_f32_e32 v0, s2
; GCN-IEEE-UNSAFE-NEXT: s_mov_b32 s2, -1
; GCN-IEEE-UNSAFE-NEXT: buffer_store_dword v0, off, s[0:3], 0
; GCN-IEEE-UNSAFE-NEXT: s_endpgm
;
; SI-IEEE-SAFE-LABEL: rsq_f32_sgpr:
; SI-IEEE-SAFE: ; %bb.0:
; SI-IEEE-SAFE-NEXT: s_load_dword s2, s[0:1], 0xb
; SI-IEEE-SAFE-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x9
; SI-IEEE-SAFE-NEXT: s_mov_b32 s4, 0x7f800000
; SI-IEEE-SAFE-NEXT: s_mov_b32 s3, 0xf000
; SI-IEEE-SAFE-NEXT: s_waitcnt lgkmcnt(0)
; SI-IEEE-SAFE-NEXT: v_sqrt_f32_e32 v0, s2
; SI-IEEE-SAFE-NEXT: s_mov_b32 s2, -1
; SI-IEEE-SAFE-NEXT: v_frexp_mant_f32_e32 v1, v0
; SI-IEEE-SAFE-NEXT: v_cmp_lt_f32_e64 vcc, |v0|, s4
; SI-IEEE-SAFE-NEXT: v_cndmask_b32_e32 v1, v0, v1, vcc
; SI-IEEE-SAFE-NEXT: v_rcp_f32_e32 v1, v1
; SI-IEEE-SAFE-NEXT: v_frexp_exp_i32_f32_e32 v0, v0
; SI-IEEE-SAFE-NEXT: v_sub_i32_e32 v0, vcc, 0, v0
; SI-IEEE-SAFE-NEXT: v_ldexp_f32_e32 v0, v1, v0
; SI-IEEE-SAFE-NEXT: buffer_store_dword v0, off, s[0:3], 0
; SI-IEEE-SAFE-NEXT: s_endpgm
;
; CI-IEEE-SAFE-LABEL: rsq_f32_sgpr:
; CI-IEEE-SAFE: ; %bb.0:
; CI-IEEE-SAFE-NEXT: s_load_dword s2, s[0:1], 0xb
; CI-IEEE-SAFE-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x9
; CI-IEEE-SAFE-NEXT: s_mov_b32 s3, 0xf000
; CI-IEEE-SAFE-NEXT: s_waitcnt lgkmcnt(0)
; CI-IEEE-SAFE-NEXT: v_sqrt_f32_e32 v0, s2
; CI-IEEE-SAFE-NEXT: s_mov_b32 s2, -1
; CI-IEEE-SAFE-NEXT: v_frexp_mant_f32_e32 v1, v0
; CI-IEEE-SAFE-NEXT: v_rcp_f32_e32 v1, v1
; CI-IEEE-SAFE-NEXT: v_frexp_exp_i32_f32_e32 v0, v0
; CI-IEEE-SAFE-NEXT: v_sub_i32_e32 v0, vcc, 0, v0
; CI-IEEE-SAFE-NEXT: v_ldexp_f32_e32 v0, v1, v0
; CI-IEEE-SAFE-NEXT: buffer_store_dword v0, off, s[0:3], 0
; CI-IEEE-SAFE-NEXT: s_endpgm
; GCN-UNSAFE-LABEL: rsq_f32_sgpr:
; GCN-UNSAFE: ; %bb.0:
; GCN-UNSAFE-NEXT: s_load_dword s2, s[0:1], 0xb
; GCN-UNSAFE-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x9
; GCN-UNSAFE-NEXT: s_mov_b32 s3, 0xf000
; GCN-UNSAFE-NEXT: s_waitcnt lgkmcnt(0)
; GCN-UNSAFE-NEXT: v_rsq_f32_e32 v0, s2
; GCN-UNSAFE-NEXT: s_mov_b32 s2, -1
; GCN-UNSAFE-NEXT: buffer_store_dword v0, off, s[0:3], 0
; GCN-UNSAFE-NEXT: s_endpgm
%sqrt = call contract float @llvm.sqrt.f32(float %val) nounwind readnone
%div = fdiv contract float 1.0, %sqrt, !fpmath !0
store float %div, ptr addrspace(1) %out, align 4
ret void
}
; Recognize that this is rsqrt(a) * rcp(b) * c,
; not 1 / ( 1 / sqrt(a)) * rcp(b) * c.
; NOTE: c * rcp( sqrt(a) * b ) is generated when we move rcp generation to AMGGPUCogenPrepare.
define amdgpu_kernel void @rsqrt_fmul(ptr addrspace(1) %out, ptr addrspace(1) %in) {
; GCN-UNSAFE-LABEL: rsqrt_fmul:
; GCN-UNSAFE: ; %bb.0:
; GCN-UNSAFE-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9
; GCN-UNSAFE-NEXT: s_mov_b32 s7, 0xf000
; GCN-UNSAFE-NEXT: s_mov_b32 s6, 0
; GCN-UNSAFE-NEXT: v_lshlrev_b32_e32 v0, 2, v0
; GCN-UNSAFE-NEXT: v_mov_b32_e32 v1, 0
; GCN-UNSAFE-NEXT: s_waitcnt lgkmcnt(0)
; GCN-UNSAFE-NEXT: s_mov_b64 s[8:9], s[2:3]
; GCN-UNSAFE-NEXT: s_mov_b64 s[10:11], s[6:7]
; GCN-UNSAFE-NEXT: buffer_load_dword v2, v[0:1], s[8:11], 0 addr64 glc
; GCN-UNSAFE-NEXT: s_waitcnt vmcnt(0)
; GCN-UNSAFE-NEXT: buffer_load_dword v3, v[0:1], s[8:11], 0 addr64 offset:4 glc
; GCN-UNSAFE-NEXT: s_waitcnt vmcnt(0)
; GCN-UNSAFE-NEXT: buffer_load_dword v4, v[0:1], s[8:11], 0 addr64 offset:8 glc
; GCN-UNSAFE-NEXT: s_waitcnt vmcnt(0)
; GCN-UNSAFE-NEXT: s_mov_b64 s[4:5], s[0:1]
; GCN-UNSAFE-NEXT: v_sqrt_f32_e32 v2, v2
; GCN-UNSAFE-NEXT: v_mul_f32_e32 v2, v2, v3
; GCN-UNSAFE-NEXT: v_rcp_f32_e32 v2, v2
; GCN-UNSAFE-NEXT: v_mul_f32_e32 v2, v4, v2
; GCN-UNSAFE-NEXT: buffer_store_dword v2, v[0:1], s[4:7], 0 addr64
; GCN-UNSAFE-NEXT: s_endpgm
; GCN-DAZ-UNSAFE-LABEL: rsqrt_fmul:
; GCN-DAZ-UNSAFE: ; %bb.0:
; GCN-DAZ-UNSAFE-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9
; GCN-DAZ-UNSAFE-NEXT: s_mov_b32 s7, 0xf000
; GCN-DAZ-UNSAFE-NEXT: s_mov_b32 s6, 0
; GCN-DAZ-UNSAFE-NEXT: v_lshlrev_b32_e32 v0, 2, v0
; GCN-DAZ-UNSAFE-NEXT: v_mov_b32_e32 v1, 0
; GCN-DAZ-UNSAFE-NEXT: s_waitcnt lgkmcnt(0)
; GCN-DAZ-UNSAFE-NEXT: s_mov_b64 s[8:9], s[2:3]
; GCN-DAZ-UNSAFE-NEXT: s_mov_b64 s[10:11], s[6:7]
; GCN-DAZ-UNSAFE-NEXT: buffer_load_dword v2, v[0:1], s[8:11], 0 addr64 glc
; GCN-DAZ-UNSAFE-NEXT: s_waitcnt vmcnt(0)
; GCN-DAZ-UNSAFE-NEXT: buffer_load_dword v3, v[0:1], s[8:11], 0 addr64 offset:4 glc
; GCN-DAZ-UNSAFE-NEXT: s_waitcnt vmcnt(0)
; GCN-DAZ-UNSAFE-NEXT: buffer_load_dword v4, v[0:1], s[8:11], 0 addr64 offset:8 glc
; GCN-DAZ-UNSAFE-NEXT: s_waitcnt vmcnt(0)
; GCN-DAZ-UNSAFE-NEXT: s_mov_b64 s[4:5], s[0:1]
; GCN-DAZ-UNSAFE-NEXT: v_rsq_f32_e32 v2, v2
; GCN-DAZ-UNSAFE-NEXT: v_rcp_f32_e32 v3, v3
; GCN-DAZ-UNSAFE-NEXT: v_mul_f32_e32 v2, v2, v3
; GCN-DAZ-UNSAFE-NEXT: v_mul_f32_e32 v2, v4, v2
; GCN-DAZ-UNSAFE-NEXT: buffer_store_dword v2, v[0:1], s[4:7], 0 addr64
; GCN-DAZ-UNSAFE-NEXT: s_endpgm
;
; GCN-IEEE-UNSAFE-LABEL: rsqrt_fmul:
; GCN-IEEE-UNSAFE: ; %bb.0:
; GCN-IEEE-UNSAFE-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9
; GCN-IEEE-UNSAFE-NEXT: s_mov_b32 s7, 0xf000
; GCN-IEEE-UNSAFE-NEXT: s_mov_b32 s6, 0
; GCN-IEEE-UNSAFE-NEXT: v_lshlrev_b32_e32 v0, 2, v0
; GCN-IEEE-UNSAFE-NEXT: v_mov_b32_e32 v1, 0
; GCN-IEEE-UNSAFE-NEXT: s_waitcnt lgkmcnt(0)
; GCN-IEEE-UNSAFE-NEXT: s_mov_b64 s[8:9], s[2:3]
; GCN-IEEE-UNSAFE-NEXT: s_mov_b64 s[10:11], s[6:7]
; GCN-IEEE-UNSAFE-NEXT: buffer_load_dword v2, v[0:1], s[8:11], 0 addr64 glc
; GCN-IEEE-UNSAFE-NEXT: s_waitcnt vmcnt(0)
; GCN-IEEE-UNSAFE-NEXT: buffer_load_dword v3, v[0:1], s[8:11], 0 addr64 offset:4 glc
; GCN-IEEE-UNSAFE-NEXT: s_waitcnt vmcnt(0)
; GCN-IEEE-UNSAFE-NEXT: buffer_load_dword v4, v[0:1], s[8:11], 0 addr64 offset:8 glc
; GCN-IEEE-UNSAFE-NEXT: s_waitcnt vmcnt(0)
; GCN-IEEE-UNSAFE-NEXT: s_mov_b64 s[4:5], s[0:1]
; GCN-IEEE-UNSAFE-NEXT: v_rsq_f32_e32 v2, v2
; GCN-IEEE-UNSAFE-NEXT: v_rcp_f32_e32 v3, v3
; GCN-IEEE-UNSAFE-NEXT: v_mul_f32_e32 v2, v2, v3
; GCN-IEEE-UNSAFE-NEXT: v_mul_f32_e32 v2, v4, v2
; GCN-IEEE-UNSAFE-NEXT: buffer_store_dword v2, v[0:1], s[4:7], 0 addr64
; GCN-IEEE-UNSAFE-NEXT: s_endpgm
;
; GCN-DAZ-SAFE-LABEL: rsqrt_fmul:
; GCN-DAZ-SAFE: ; %bb.0:
; GCN-DAZ-SAFE-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9
; GCN-DAZ-SAFE-NEXT: s_mov_b32 s7, 0xf000
; GCN-DAZ-SAFE-NEXT: s_mov_b32 s6, 0
; GCN-DAZ-SAFE-NEXT: v_lshlrev_b32_e32 v0, 2, v0
; GCN-DAZ-SAFE-NEXT: v_mov_b32_e32 v1, 0
; GCN-DAZ-SAFE-NEXT: s_waitcnt lgkmcnt(0)
; GCN-DAZ-SAFE-NEXT: s_mov_b64 s[8:9], s[2:3]
; GCN-DAZ-SAFE-NEXT: s_mov_b64 s[10:11], s[6:7]
; GCN-DAZ-SAFE-NEXT: buffer_load_dword v2, v[0:1], s[8:11], 0 addr64 glc
; GCN-DAZ-SAFE-NEXT: s_waitcnt vmcnt(0)
; GCN-DAZ-SAFE-NEXT: buffer_load_dword v3, v[0:1], s[8:11], 0 addr64 offset:4 glc
; GCN-DAZ-SAFE-NEXT: s_waitcnt vmcnt(0)
; GCN-DAZ-SAFE-NEXT: buffer_load_dword v4, v[0:1], s[8:11], 0 addr64 offset:8 glc
; GCN-DAZ-SAFE-NEXT: s_waitcnt vmcnt(0)
; GCN-DAZ-SAFE-NEXT: s_mov_b64 s[4:5], s[0:1]
; GCN-DAZ-SAFE-NEXT: v_sqrt_f32_e32 v2, v2
; GCN-DAZ-SAFE-NEXT: v_mul_f32_e32 v2, v2, v3
; GCN-DAZ-SAFE-NEXT: v_div_scale_f32 v3, s[0:1], v2, v2, v4
; GCN-DAZ-SAFE-NEXT: v_rcp_f32_e32 v5, v3
; GCN-DAZ-SAFE-NEXT: v_div_scale_f32 v6, vcc, v4, v2, v4
; GCN-DAZ-SAFE-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 3
; GCN-DAZ-SAFE-NEXT: v_fma_f32 v7, -v3, v5, 1.0
; GCN-DAZ-SAFE-NEXT: v_fma_f32 v5, v7, v5, v5
; GCN-DAZ-SAFE-NEXT: v_mul_f32_e32 v7, v6, v5
; GCN-DAZ-SAFE-NEXT: v_fma_f32 v8, -v3, v7, v6
; GCN-DAZ-SAFE-NEXT: v_fma_f32 v7, v8, v5, v7
; GCN-DAZ-SAFE-NEXT: v_fma_f32 v3, -v3, v7, v6
; GCN-DAZ-SAFE-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 0
; GCN-DAZ-SAFE-NEXT: v_div_fmas_f32 v3, v3, v5, v7
; GCN-DAZ-SAFE-NEXT: v_div_fixup_f32 v2, v3, v2, v4
; GCN-DAZ-SAFE-NEXT: buffer_store_dword v2, v[0:1], s[4:7], 0 addr64
; GCN-DAZ-SAFE-NEXT: s_endpgm
;
; GCN-IEEE-SAFE-LABEL: rsqrt_fmul:
; GCN-IEEE-SAFE: ; %bb.0:
; GCN-IEEE-SAFE-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9
; GCN-IEEE-SAFE-NEXT: s_mov_b32 s7, 0xf000
; GCN-IEEE-SAFE-NEXT: s_mov_b32 s6, 0
; GCN-IEEE-SAFE-NEXT: v_lshlrev_b32_e32 v0, 2, v0
; GCN-IEEE-SAFE-NEXT: v_mov_b32_e32 v1, 0
; GCN-IEEE-SAFE-NEXT: s_waitcnt lgkmcnt(0)
; GCN-IEEE-SAFE-NEXT: s_mov_b64 s[8:9], s[2:3]
; GCN-IEEE-SAFE-NEXT: s_mov_b64 s[10:11], s[6:7]
; GCN-IEEE-SAFE-NEXT: buffer_load_dword v2, v[0:1], s[8:11], 0 addr64 glc
; GCN-IEEE-SAFE-NEXT: s_waitcnt vmcnt(0)
; GCN-IEEE-SAFE-NEXT: buffer_load_dword v3, v[0:1], s[8:11], 0 addr64 offset:4 glc
; GCN-IEEE-SAFE-NEXT: s_waitcnt vmcnt(0)
; GCN-IEEE-SAFE-NEXT: buffer_load_dword v4, v[0:1], s[8:11], 0 addr64 offset:8 glc
; GCN-IEEE-SAFE-NEXT: s_waitcnt vmcnt(0)
; GCN-IEEE-SAFE-NEXT: s_mov_b64 s[4:5], s[0:1]
; GCN-IEEE-SAFE-NEXT: v_sqrt_f32_e32 v2, v2
; GCN-IEEE-SAFE-NEXT: v_mul_f32_e32 v2, v2, v3
; GCN-IEEE-SAFE-NEXT: v_div_scale_f32 v3, s[2:3], v2, v2, v4
; GCN-IEEE-SAFE-NEXT: v_rcp_f32_e32 v5, v3
; GCN-IEEE-SAFE-NEXT: v_div_scale_f32 v6, vcc, v4, v2, v4
; GCN-IEEE-SAFE-NEXT: v_fma_f32 v7, -v3, v5, 1.0
; GCN-IEEE-SAFE-NEXT: v_fma_f32 v5, v7, v5, v5
; GCN-IEEE-SAFE-NEXT: v_mul_f32_e32 v7, v6, v5
; GCN-IEEE-SAFE-NEXT: v_fma_f32 v8, -v3, v7, v6
; GCN-IEEE-SAFE-NEXT: v_fma_f32 v7, v8, v5, v7
; GCN-IEEE-SAFE-NEXT: v_fma_f32 v3, -v3, v7, v6
; GCN-IEEE-SAFE-NEXT: v_div_fmas_f32 v3, v3, v5, v7
; GCN-IEEE-SAFE-NEXT: v_div_fixup_f32 v2, v3, v2, v4
; GCN-IEEE-SAFE-NEXT: buffer_store_dword v2, v[0:1], s[4:7], 0 addr64
; GCN-IEEE-SAFE-NEXT: s_endpgm
%tid = call i32 @llvm.amdgcn.workitem.id.x() nounwind readnone
%out.gep = getelementptr float, ptr addrspace(1) %out, i32 %tid
%gep.0 = getelementptr float, ptr addrspace(1) %in, i32 %tid
%gep.1 = getelementptr float, ptr addrspace(1) %gep.0, i32 1
%gep.2 = getelementptr float, ptr addrspace(1) %gep.0, i32 2
%a = load volatile float, ptr addrspace(1) %gep.0
%b = load volatile float, ptr addrspace(1) %gep.1
%c = load volatile float, ptr addrspace(1) %gep.2
%x = call contract float @llvm.sqrt.f32(float %a)
%y = fmul contract float %x, %b
%z = fdiv contract float %c, %y
store float %z, ptr addrspace(1) %out.gep
ret void
}
define amdgpu_kernel void @neg_rsq_f32(ptr addrspace(1) noalias %out, ptr addrspace(1) noalias %in) {
; GCN-DAZ-UNSAFE-LABEL: neg_rsq_f32:
; GCN-DAZ-UNSAFE: ; %bb.0:
; GCN-DAZ-UNSAFE-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9
; GCN-DAZ-UNSAFE-NEXT: s_mov_b32 s7, 0xf000
; GCN-DAZ-UNSAFE-NEXT: s_mov_b32 s6, -1
; GCN-DAZ-UNSAFE-NEXT: s_mov_b32 s10, s6
; GCN-DAZ-UNSAFE-NEXT: s_mov_b32 s11, s7
; GCN-DAZ-UNSAFE-NEXT: s_waitcnt lgkmcnt(0)
; GCN-DAZ-UNSAFE-NEXT: s_mov_b32 s8, s2
; GCN-DAZ-UNSAFE-NEXT: s_mov_b32 s9, s3
; GCN-DAZ-UNSAFE-NEXT: buffer_load_dword v0, off, s[8:11], 0
; GCN-DAZ-UNSAFE-NEXT: s_mov_b32 s4, s0
; GCN-DAZ-UNSAFE-NEXT: s_mov_b32 s5, s1
; GCN-DAZ-UNSAFE-NEXT: s_waitcnt vmcnt(0)
; GCN-DAZ-UNSAFE-NEXT: v_rsq_f32_e32 v0, v0
; GCN-DAZ-UNSAFE-NEXT: v_sub_f32_e32 v0, 0x80000000, v0
; GCN-DAZ-UNSAFE-NEXT: buffer_store_dword v0, off, s[4:7], 0
; GCN-DAZ-UNSAFE-NEXT: s_endpgm
;
; GCN-IEEE-UNSAFE-LABEL: neg_rsq_f32:
; GCN-IEEE-UNSAFE: ; %bb.0:
; GCN-IEEE-UNSAFE-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9
; GCN-IEEE-UNSAFE-NEXT: s_mov_b32 s7, 0xf000
; GCN-IEEE-UNSAFE-NEXT: s_mov_b32 s6, -1
; GCN-IEEE-UNSAFE-NEXT: s_mov_b32 s10, s6
; GCN-IEEE-UNSAFE-NEXT: s_mov_b32 s11, s7
; GCN-IEEE-UNSAFE-NEXT: s_waitcnt lgkmcnt(0)
; GCN-IEEE-UNSAFE-NEXT: s_mov_b32 s8, s2
; GCN-IEEE-UNSAFE-NEXT: s_mov_b32 s9, s3
; GCN-IEEE-UNSAFE-NEXT: buffer_load_dword v0, off, s[8:11], 0
; GCN-IEEE-UNSAFE-NEXT: s_mov_b32 s4, s0
; GCN-IEEE-UNSAFE-NEXT: s_mov_b32 s5, s1
; GCN-IEEE-UNSAFE-NEXT: s_waitcnt vmcnt(0)
; GCN-IEEE-UNSAFE-NEXT: v_rsq_f32_e32 v0, v0
; GCN-IEEE-UNSAFE-NEXT: v_xor_b32_e32 v0, 0x80000000, v0
; GCN-IEEE-UNSAFE-NEXT: buffer_store_dword v0, off, s[4:7], 0
; GCN-IEEE-UNSAFE-NEXT: s_endpgm
;
; GCN-DAZ-SAFE-LABEL: neg_rsq_f32:
; GCN-DAZ-SAFE: ; %bb.0:
; GCN-DAZ-SAFE-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9
; GCN-DAZ-SAFE-NEXT: s_mov_b32 s7, 0xf000
; GCN-DAZ-SAFE-NEXT: s_mov_b32 s6, -1
; GCN-DAZ-SAFE-NEXT: s_mov_b32 s10, s6
; GCN-DAZ-SAFE-NEXT: s_mov_b32 s11, s7
; GCN-DAZ-SAFE-NEXT: s_waitcnt lgkmcnt(0)
; GCN-DAZ-SAFE-NEXT: s_mov_b32 s8, s2
; GCN-DAZ-SAFE-NEXT: s_mov_b32 s9, s3
; GCN-DAZ-SAFE-NEXT: buffer_load_dword v0, off, s[8:11], 0
; GCN-DAZ-SAFE-NEXT: s_mov_b32 s4, s0
; GCN-DAZ-SAFE-NEXT: s_mov_b32 s5, s1
; GCN-DAZ-SAFE-NEXT: s_waitcnt vmcnt(0)
; GCN-DAZ-SAFE-NEXT: v_sqrt_f32_e32 v0, v0
; GCN-DAZ-SAFE-NEXT: v_rcp_f32_e64 v0, -v0
; GCN-DAZ-SAFE-NEXT: buffer_store_dword v0, off, s[4:7], 0
; GCN-DAZ-SAFE-NEXT: s_endpgm
;
; SI-IEEE-SAFE-LABEL: neg_rsq_f32:
; SI-IEEE-SAFE: ; %bb.0:
; SI-IEEE-SAFE-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9
; SI-IEEE-SAFE-NEXT: s_mov_b32 s7, 0xf000
; SI-IEEE-SAFE-NEXT: s_mov_b32 s6, -1
; SI-IEEE-SAFE-NEXT: s_mov_b32 s10, s6
; SI-IEEE-SAFE-NEXT: s_mov_b32 s11, s7
; SI-IEEE-SAFE-NEXT: s_waitcnt lgkmcnt(0)
; SI-IEEE-SAFE-NEXT: s_mov_b32 s8, s2
; SI-IEEE-SAFE-NEXT: s_mov_b32 s9, s3
; SI-IEEE-SAFE-NEXT: buffer_load_dword v0, off, s[8:11], 0
; SI-IEEE-SAFE-NEXT: s_mov_b32 s2, 0x7f800000
; SI-IEEE-SAFE-NEXT: s_mov_b32 s4, s0
; SI-IEEE-SAFE-NEXT: s_mov_b32 s5, s1
; SI-IEEE-SAFE-NEXT: s_waitcnt vmcnt(0)
; SI-IEEE-SAFE-NEXT: v_sqrt_f32_e32 v0, v0
; SI-IEEE-SAFE-NEXT: v_frexp_mant_f32_e64 v1, -v0
; SI-IEEE-SAFE-NEXT: v_cmp_lt_f32_e64 s[0:1], |v0|, s2
; SI-IEEE-SAFE-NEXT: v_cndmask_b32_e64 v1, -v0, v1, s[0:1]
; SI-IEEE-SAFE-NEXT: v_rcp_f32_e32 v1, v1
; SI-IEEE-SAFE-NEXT: v_frexp_exp_i32_f32_e32 v0, v0
; SI-IEEE-SAFE-NEXT: v_sub_i32_e32 v0, vcc, 0, v0
; SI-IEEE-SAFE-NEXT: v_ldexp_f32_e32 v0, v1, v0
; SI-IEEE-SAFE-NEXT: buffer_store_dword v0, off, s[4:7], 0
; SI-IEEE-SAFE-NEXT: s_endpgm
;
; CI-IEEE-SAFE-LABEL: neg_rsq_f32:
; CI-IEEE-SAFE: ; %bb.0:
; CI-IEEE-SAFE-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9
; CI-IEEE-SAFE-NEXT: s_mov_b32 s7, 0xf000
; CI-IEEE-SAFE-NEXT: s_mov_b32 s6, -1
; CI-IEEE-SAFE-NEXT: s_mov_b32 s10, s6
; CI-IEEE-SAFE-NEXT: s_mov_b32 s11, s7
; CI-IEEE-SAFE-NEXT: s_waitcnt lgkmcnt(0)
; CI-IEEE-SAFE-NEXT: s_mov_b32 s8, s2
; CI-IEEE-SAFE-NEXT: s_mov_b32 s9, s3
; CI-IEEE-SAFE-NEXT: buffer_load_dword v0, off, s[8:11], 0
; CI-IEEE-SAFE-NEXT: s_mov_b32 s4, s0
; CI-IEEE-SAFE-NEXT: s_mov_b32 s5, s1
; CI-IEEE-SAFE-NEXT: s_waitcnt vmcnt(0)
; CI-IEEE-SAFE-NEXT: v_sqrt_f32_e32 v0, v0
; CI-IEEE-SAFE-NEXT: v_frexp_mant_f32_e64 v1, -v0
; CI-IEEE-SAFE-NEXT: v_rcp_f32_e32 v1, v1
; CI-IEEE-SAFE-NEXT: v_frexp_exp_i32_f32_e32 v0, v0
; CI-IEEE-SAFE-NEXT: v_sub_i32_e32 v0, vcc, 0, v0
; CI-IEEE-SAFE-NEXT: v_ldexp_f32_e32 v0, v1, v0
; CI-IEEE-SAFE-NEXT: buffer_store_dword v0, off, s[4:7], 0
; CI-IEEE-SAFE-NEXT: s_endpgm
; GCN-UNSAFE-LABEL: neg_rsq_f32:
; GCN-UNSAFE: ; %bb.0:
; GCN-UNSAFE-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9
; GCN-UNSAFE-NEXT: s_mov_b32 s7, 0xf000
; GCN-UNSAFE-NEXT: s_mov_b32 s6, -1
; GCN-UNSAFE-NEXT: s_mov_b32 s10, s6
; GCN-UNSAFE-NEXT: s_mov_b32 s11, s7
; GCN-UNSAFE-NEXT: s_waitcnt lgkmcnt(0)
; GCN-UNSAFE-NEXT: s_mov_b32 s8, s2
; GCN-UNSAFE-NEXT: s_mov_b32 s9, s3
; GCN-UNSAFE-NEXT: buffer_load_dword v0, off, s[8:11], 0
; GCN-UNSAFE-NEXT: s_mov_b32 s4, s0
; GCN-UNSAFE-NEXT: s_mov_b32 s5, s1
; GCN-UNSAFE-NEXT: s_waitcnt vmcnt(0)
; GCN-UNSAFE-NEXT: v_sqrt_f32_e32 v0, v0
; GCN-UNSAFE-NEXT: v_rcp_f32_e64 v0, -v0
; GCN-UNSAFE-NEXT: buffer_store_dword v0, off, s[4:7], 0
; GCN-UNSAFE-NEXT: s_endpgm
%val = load float, ptr addrspace(1) %in, align 4
%sqrt = call contract float @llvm.sqrt.f32(float %val)
%div = fdiv contract float -1.0, %sqrt, !fpmath !0
store float %div, ptr addrspace(1) %out, align 4
ret void
}
define amdgpu_kernel void @neg_rsq_neg_f32(ptr addrspace(1) noalias %out, ptr addrspace(1) noalias %in) {
; GCN-DAZ-UNSAFE-LABEL: neg_rsq_neg_f32:
; GCN-DAZ-UNSAFE: ; %bb.0:
; GCN-DAZ-UNSAFE-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9
; GCN-DAZ-UNSAFE-NEXT: s_mov_b32 s7, 0xf000
; GCN-DAZ-UNSAFE-NEXT: s_mov_b32 s6, -1
; GCN-DAZ-UNSAFE-NEXT: s_mov_b32 s10, s6
; GCN-DAZ-UNSAFE-NEXT: s_mov_b32 s11, s7
; GCN-DAZ-UNSAFE-NEXT: s_waitcnt lgkmcnt(0)
; GCN-DAZ-UNSAFE-NEXT: s_mov_b32 s8, s2
; GCN-DAZ-UNSAFE-NEXT: s_mov_b32 s9, s3
; GCN-DAZ-UNSAFE-NEXT: buffer_load_dword v0, off, s[8:11], 0
; GCN-DAZ-UNSAFE-NEXT: s_mov_b32 s4, s0
; GCN-DAZ-UNSAFE-NEXT: s_mov_b32 s5, s1
; GCN-DAZ-UNSAFE-NEXT: s_waitcnt vmcnt(0)
; GCN-DAZ-UNSAFE-NEXT: v_rsq_f32_e64 v0, -v0
; GCN-DAZ-UNSAFE-NEXT: v_sub_f32_e32 v0, 0x80000000, v0
; GCN-DAZ-UNSAFE-NEXT: buffer_store_dword v0, off, s[4:7], 0
; GCN-DAZ-UNSAFE-NEXT: s_endpgm
;
; GCN-IEEE-UNSAFE-LABEL: neg_rsq_neg_f32:
; GCN-IEEE-UNSAFE: ; %bb.0:
; GCN-IEEE-UNSAFE-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9
; GCN-IEEE-UNSAFE-NEXT: s_mov_b32 s7, 0xf000
; GCN-IEEE-UNSAFE-NEXT: s_mov_b32 s6, -1
; GCN-IEEE-UNSAFE-NEXT: s_mov_b32 s10, s6
; GCN-IEEE-UNSAFE-NEXT: s_mov_b32 s11, s7
; GCN-IEEE-UNSAFE-NEXT: s_waitcnt lgkmcnt(0)
; GCN-IEEE-UNSAFE-NEXT: s_mov_b32 s8, s2
; GCN-IEEE-UNSAFE-NEXT: s_mov_b32 s9, s3
; GCN-IEEE-UNSAFE-NEXT: buffer_load_dword v0, off, s[8:11], 0
; GCN-IEEE-UNSAFE-NEXT: s_mov_b32 s4, s0
; GCN-IEEE-UNSAFE-NEXT: s_mov_b32 s5, s1
; GCN-IEEE-UNSAFE-NEXT: s_waitcnt vmcnt(0)
; GCN-IEEE-UNSAFE-NEXT: v_rsq_f32_e64 v0, -v0
; GCN-IEEE-UNSAFE-NEXT: v_xor_b32_e32 v0, 0x80000000, v0
; GCN-IEEE-UNSAFE-NEXT: buffer_store_dword v0, off, s[4:7], 0
; GCN-IEEE-UNSAFE-NEXT: s_endpgm
;
; GCN-DAZ-SAFE-LABEL: neg_rsq_neg_f32:
; GCN-DAZ-SAFE: ; %bb.0:
; GCN-DAZ-SAFE-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9
; GCN-DAZ-SAFE-NEXT: s_mov_b32 s7, 0xf000
; GCN-DAZ-SAFE-NEXT: s_mov_b32 s6, -1
; GCN-DAZ-SAFE-NEXT: s_mov_b32 s10, s6
; GCN-DAZ-SAFE-NEXT: s_mov_b32 s11, s7
; GCN-DAZ-SAFE-NEXT: s_waitcnt lgkmcnt(0)
; GCN-DAZ-SAFE-NEXT: s_mov_b32 s8, s2
; GCN-DAZ-SAFE-NEXT: s_mov_b32 s9, s3
; GCN-DAZ-SAFE-NEXT: buffer_load_dword v0, off, s[8:11], 0
; GCN-DAZ-SAFE-NEXT: s_mov_b32 s4, s0
; GCN-DAZ-SAFE-NEXT: s_mov_b32 s5, s1
; GCN-DAZ-SAFE-NEXT: s_waitcnt vmcnt(0)
; GCN-DAZ-SAFE-NEXT: v_sqrt_f32_e64 v0, -v0
; GCN-DAZ-SAFE-NEXT: v_rcp_f32_e64 v0, -v0
; GCN-DAZ-SAFE-NEXT: buffer_store_dword v0, off, s[4:7], 0
; GCN-DAZ-SAFE-NEXT: s_endpgm
;
; SI-IEEE-SAFE-LABEL: neg_rsq_neg_f32:
; SI-IEEE-SAFE: ; %bb.0:
; SI-IEEE-SAFE-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9
; SI-IEEE-SAFE-NEXT: s_mov_b32 s7, 0xf000
; SI-IEEE-SAFE-NEXT: s_mov_b32 s6, -1
; SI-IEEE-SAFE-NEXT: s_mov_b32 s10, s6
; SI-IEEE-SAFE-NEXT: s_mov_b32 s11, s7
; SI-IEEE-SAFE-NEXT: s_waitcnt lgkmcnt(0)
; SI-IEEE-SAFE-NEXT: s_mov_b32 s8, s2
; SI-IEEE-SAFE-NEXT: s_mov_b32 s9, s3
; SI-IEEE-SAFE-NEXT: buffer_load_dword v0, off, s[8:11], 0
; SI-IEEE-SAFE-NEXT: s_mov_b32 s2, 0x7f800000
; SI-IEEE-SAFE-NEXT: s_mov_b32 s4, s0
; SI-IEEE-SAFE-NEXT: s_mov_b32 s5, s1
; SI-IEEE-SAFE-NEXT: s_waitcnt vmcnt(0)
; SI-IEEE-SAFE-NEXT: v_sqrt_f32_e64 v0, -v0
; SI-IEEE-SAFE-NEXT: v_frexp_mant_f32_e64 v1, -v0
; SI-IEEE-SAFE-NEXT: v_cmp_lt_f32_e64 s[0:1], |v0|, s2
; SI-IEEE-SAFE-NEXT: v_cndmask_b32_e64 v1, -v0, v1, s[0:1]
; SI-IEEE-SAFE-NEXT: v_rcp_f32_e32 v1, v1
; SI-IEEE-SAFE-NEXT: v_frexp_exp_i32_f32_e32 v0, v0
; SI-IEEE-SAFE-NEXT: v_sub_i32_e32 v0, vcc, 0, v0
; SI-IEEE-SAFE-NEXT: v_ldexp_f32_e32 v0, v1, v0
; SI-IEEE-SAFE-NEXT: buffer_store_dword v0, off, s[4:7], 0
; SI-IEEE-SAFE-NEXT: s_endpgm
;
; CI-IEEE-SAFE-LABEL: neg_rsq_neg_f32:
; CI-IEEE-SAFE: ; %bb.0:
; CI-IEEE-SAFE-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9
; CI-IEEE-SAFE-NEXT: s_mov_b32 s7, 0xf000
; CI-IEEE-SAFE-NEXT: s_mov_b32 s6, -1
; CI-IEEE-SAFE-NEXT: s_mov_b32 s10, s6
; CI-IEEE-SAFE-NEXT: s_mov_b32 s11, s7
; CI-IEEE-SAFE-NEXT: s_waitcnt lgkmcnt(0)
; CI-IEEE-SAFE-NEXT: s_mov_b32 s8, s2
; CI-IEEE-SAFE-NEXT: s_mov_b32 s9, s3
; CI-IEEE-SAFE-NEXT: buffer_load_dword v0, off, s[8:11], 0
; CI-IEEE-SAFE-NEXT: s_mov_b32 s4, s0
; CI-IEEE-SAFE-NEXT: s_mov_b32 s5, s1
; CI-IEEE-SAFE-NEXT: s_waitcnt vmcnt(0)
; CI-IEEE-SAFE-NEXT: v_sqrt_f32_e64 v0, -v0
; CI-IEEE-SAFE-NEXT: v_frexp_mant_f32_e64 v1, -v0
; CI-IEEE-SAFE-NEXT: v_rcp_f32_e32 v1, v1
; CI-IEEE-SAFE-NEXT: v_frexp_exp_i32_f32_e32 v0, v0
; CI-IEEE-SAFE-NEXT: v_sub_i32_e32 v0, vcc, 0, v0
; CI-IEEE-SAFE-NEXT: v_ldexp_f32_e32 v0, v1, v0
; CI-IEEE-SAFE-NEXT: buffer_store_dword v0, off, s[4:7], 0
; CI-IEEE-SAFE-NEXT: s_endpgm
; GCN-UNSAFE-LABEL: neg_rsq_neg_f32:
; GCN-UNSAFE: ; %bb.0:
; GCN-UNSAFE-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9
; GCN-UNSAFE-NEXT: s_mov_b32 s7, 0xf000
; GCN-UNSAFE-NEXT: s_mov_b32 s6, -1
; GCN-UNSAFE-NEXT: s_mov_b32 s10, s6
; GCN-UNSAFE-NEXT: s_mov_b32 s11, s7
; GCN-UNSAFE-NEXT: s_waitcnt lgkmcnt(0)
; GCN-UNSAFE-NEXT: s_mov_b32 s8, s2
; GCN-UNSAFE-NEXT: s_mov_b32 s9, s3
; GCN-UNSAFE-NEXT: buffer_load_dword v0, off, s[8:11], 0
; GCN-UNSAFE-NEXT: s_mov_b32 s4, s0
; GCN-UNSAFE-NEXT: s_mov_b32 s5, s1
; GCN-UNSAFE-NEXT: s_waitcnt vmcnt(0)
; GCN-UNSAFE-NEXT: v_sqrt_f32_e64 v0, -v0
; GCN-UNSAFE-NEXT: v_rcp_f32_e64 v0, -v0
; GCN-UNSAFE-NEXT: buffer_store_dword v0, off, s[4:7], 0
; GCN-UNSAFE-NEXT: s_endpgm
%val = load float, ptr addrspace(1) %in, align 4
%val.fneg = fneg float %val
%sqrt = call contract float @llvm.sqrt.f32(float %val.fneg)
%div = fdiv contract float -1.0, %sqrt, !fpmath !0
store float %div, ptr addrspace(1) %out, align 4
ret void
}
define float @v_neg_rsq_neg_f32(float %val) {
; GCN-DAZ-UNSAFE-LABEL: v_neg_rsq_neg_f32:
; GCN-DAZ-UNSAFE: ; %bb.0:
; GCN-DAZ-UNSAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-DAZ-UNSAFE-NEXT: v_rsq_f32_e64 v0, -v0
; GCN-DAZ-UNSAFE-NEXT: v_sub_f32_e32 v0, 0x80000000, v0
; GCN-DAZ-UNSAFE-NEXT: s_setpc_b64 s[30:31]
;
; GCN-IEEE-UNSAFE-LABEL: v_neg_rsq_neg_f32:
; GCN-IEEE-UNSAFE: ; %bb.0:
; GCN-IEEE-UNSAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-IEEE-UNSAFE-NEXT: v_rsq_f32_e64 v0, -v0
; GCN-IEEE-UNSAFE-NEXT: v_xor_b32_e32 v0, 0x80000000, v0
; GCN-IEEE-UNSAFE-NEXT: s_setpc_b64 s[30:31]
;
; GCN-DAZ-SAFE-LABEL: v_neg_rsq_neg_f32:
; GCN-DAZ-SAFE: ; %bb.0:
; GCN-DAZ-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-DAZ-SAFE-NEXT: v_sqrt_f32_e64 v0, -v0
; GCN-DAZ-SAFE-NEXT: v_rcp_f32_e64 v0, -v0
; GCN-DAZ-SAFE-NEXT: s_setpc_b64 s[30:31]
;
; SI-IEEE-SAFE-LABEL: v_neg_rsq_neg_f32:
; SI-IEEE-SAFE: ; %bb.0:
; SI-IEEE-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-IEEE-SAFE-NEXT: v_sqrt_f32_e64 v0, -v0
; SI-IEEE-SAFE-NEXT: s_mov_b32 s4, 0x7f800000
; SI-IEEE-SAFE-NEXT: v_frexp_mant_f32_e64 v1, -v0
; SI-IEEE-SAFE-NEXT: v_cmp_lt_f32_e64 s[4:5], |v0|, s4
; SI-IEEE-SAFE-NEXT: v_cndmask_b32_e64 v1, -v0, v1, s[4:5]
; SI-IEEE-SAFE-NEXT: v_rcp_f32_e32 v1, v1
; SI-IEEE-SAFE-NEXT: v_frexp_exp_i32_f32_e32 v0, v0
; SI-IEEE-SAFE-NEXT: v_sub_i32_e32 v0, vcc, 0, v0
; SI-IEEE-SAFE-NEXT: v_ldexp_f32_e32 v0, v1, v0
; SI-IEEE-SAFE-NEXT: s_setpc_b64 s[30:31]
;
; CI-IEEE-SAFE-LABEL: v_neg_rsq_neg_f32:
; CI-IEEE-SAFE: ; %bb.0:
; CI-IEEE-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; CI-IEEE-SAFE-NEXT: v_sqrt_f32_e64 v0, -v0
; CI-IEEE-SAFE-NEXT: v_frexp_mant_f32_e64 v1, -v0
; CI-IEEE-SAFE-NEXT: v_rcp_f32_e32 v1, v1
; CI-IEEE-SAFE-NEXT: v_frexp_exp_i32_f32_e32 v0, v0
; CI-IEEE-SAFE-NEXT: v_sub_i32_e32 v0, vcc, 0, v0
; CI-IEEE-SAFE-NEXT: v_ldexp_f32_e32 v0, v1, v0
; CI-IEEE-SAFE-NEXT: s_setpc_b64 s[30:31]
%val.fneg = fneg float %val
%sqrt = call contract float @llvm.sqrt.f32(float %val.fneg)
%div = fdiv contract float -1.0, %sqrt, !fpmath !0
ret float %div
}
define <2 x float> @v_neg_rsq_neg_v2f32(<2 x float> %val) {
; GCN-DAZ-UNSAFE-LABEL: v_neg_rsq_neg_v2f32:
; GCN-DAZ-UNSAFE: ; %bb.0:
; GCN-DAZ-UNSAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-DAZ-UNSAFE-NEXT: v_rsq_f32_e64 v0, -v0
; GCN-DAZ-UNSAFE-NEXT: v_rsq_f32_e64 v1, -v1
; GCN-DAZ-UNSAFE-NEXT: v_sub_f32_e32 v0, 0x80000000, v0
; GCN-DAZ-UNSAFE-NEXT: v_sub_f32_e32 v1, 0x80000000, v1
; GCN-DAZ-UNSAFE-NEXT: s_setpc_b64 s[30:31]
;
; GCN-IEEE-UNSAFE-LABEL: v_neg_rsq_neg_v2f32:
; GCN-IEEE-UNSAFE: ; %bb.0:
; GCN-IEEE-UNSAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-IEEE-UNSAFE-NEXT: v_rsq_f32_e64 v0, -v0
; GCN-IEEE-UNSAFE-NEXT: v_rsq_f32_e64 v1, -v1
; GCN-IEEE-UNSAFE-NEXT: v_xor_b32_e32 v0, 0x80000000, v0
; GCN-IEEE-UNSAFE-NEXT: v_xor_b32_e32 v1, 0x80000000, v1
; GCN-IEEE-UNSAFE-NEXT: s_setpc_b64 s[30:31]
;
; GCN-DAZ-SAFE-LABEL: v_neg_rsq_neg_v2f32:
; GCN-DAZ-SAFE: ; %bb.0:
; GCN-DAZ-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-DAZ-SAFE-NEXT: v_sqrt_f32_e64 v0, -v0
; GCN-DAZ-SAFE-NEXT: v_sqrt_f32_e64 v1, -v1
; GCN-DAZ-SAFE-NEXT: v_rcp_f32_e64 v0, -v0
; GCN-DAZ-SAFE-NEXT: v_rcp_f32_e64 v1, -v1
; GCN-DAZ-SAFE-NEXT: s_setpc_b64 s[30:31]
;
; SI-IEEE-SAFE-LABEL: v_neg_rsq_neg_v2f32:
; SI-IEEE-SAFE: ; %bb.0:
; SI-IEEE-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-IEEE-SAFE-NEXT: v_sqrt_f32_e64 v0, -v0
; SI-IEEE-SAFE-NEXT: s_mov_b32 s6, 0x7f800000
; SI-IEEE-SAFE-NEXT: v_sqrt_f32_e64 v1, -v1
; SI-IEEE-SAFE-NEXT: v_frexp_mant_f32_e64 v2, -v0
; SI-IEEE-SAFE-NEXT: v_cmp_lt_f32_e64 s[4:5], |v0|, s6
; SI-IEEE-SAFE-NEXT: v_cndmask_b32_e64 v2, -v0, v2, s[4:5]
; SI-IEEE-SAFE-NEXT: v_rcp_f32_e32 v2, v2
; SI-IEEE-SAFE-NEXT: v_frexp_exp_i32_f32_e32 v0, v0
; SI-IEEE-SAFE-NEXT: v_sub_i32_e32 v0, vcc, 0, v0
; SI-IEEE-SAFE-NEXT: v_ldexp_f32_e32 v0, v2, v0
; SI-IEEE-SAFE-NEXT: v_frexp_mant_f32_e64 v2, -v1
; SI-IEEE-SAFE-NEXT: v_cmp_lt_f32_e64 s[4:5], |v1|, s6
; SI-IEEE-SAFE-NEXT: v_cndmask_b32_e64 v2, -v1, v2, s[4:5]
; SI-IEEE-SAFE-NEXT: v_rcp_f32_e32 v2, v2
; SI-IEEE-SAFE-NEXT: v_frexp_exp_i32_f32_e32 v1, v1
; SI-IEEE-SAFE-NEXT: v_sub_i32_e32 v1, vcc, 0, v1
; SI-IEEE-SAFE-NEXT: v_ldexp_f32_e32 v1, v2, v1
; SI-IEEE-SAFE-NEXT: s_setpc_b64 s[30:31]
;
; CI-IEEE-SAFE-LABEL: v_neg_rsq_neg_v2f32:
; CI-IEEE-SAFE: ; %bb.0:
; CI-IEEE-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; CI-IEEE-SAFE-NEXT: v_sqrt_f32_e64 v0, -v0
; CI-IEEE-SAFE-NEXT: v_sqrt_f32_e64 v1, -v1
; CI-IEEE-SAFE-NEXT: v_frexp_mant_f32_e64 v2, -v0
; CI-IEEE-SAFE-NEXT: v_rcp_f32_e32 v2, v2
; CI-IEEE-SAFE-NEXT: v_frexp_exp_i32_f32_e32 v0, v0
; CI-IEEE-SAFE-NEXT: v_frexp_mant_f32_e64 v3, -v1
; CI-IEEE-SAFE-NEXT: v_sub_i32_e32 v0, vcc, 0, v0
; CI-IEEE-SAFE-NEXT: v_ldexp_f32_e32 v0, v2, v0
; CI-IEEE-SAFE-NEXT: v_rcp_f32_e32 v2, v3
; CI-IEEE-SAFE-NEXT: v_frexp_exp_i32_f32_e32 v1, v1
; CI-IEEE-SAFE-NEXT: v_sub_i32_e32 v1, vcc, 0, v1
; CI-IEEE-SAFE-NEXT: v_ldexp_f32_e32 v1, v2, v1
; CI-IEEE-SAFE-NEXT: s_setpc_b64 s[30:31]
%val.fneg = fneg <2 x float> %val
%sqrt = call contract <2 x float> @llvm.sqrt.v2f32(<2 x float> %val.fneg)
%div = fdiv contract <2 x float> <float -1.0, float -1.0>, %sqrt, !fpmath !0
ret <2 x float> %div
}
define float @v_neg_rsq_neg_f32_foldable_user(float %val0, float %val1) {
; GCN-DAZ-UNSAFE-LABEL: v_neg_rsq_neg_f32_foldable_user:
; GCN-DAZ-UNSAFE: ; %bb.0:
; GCN-DAZ-UNSAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-DAZ-UNSAFE-NEXT: v_rsq_f32_e64 v0, -v0
; GCN-DAZ-UNSAFE-NEXT: v_mul_f32_e64 v0, -v0, v1
; GCN-DAZ-UNSAFE-NEXT: s_setpc_b64 s[30:31]
;
; GCN-IEEE-UNSAFE-LABEL: v_neg_rsq_neg_f32_foldable_user:
; GCN-IEEE-UNSAFE: ; %bb.0:
; GCN-IEEE-UNSAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-IEEE-UNSAFE-NEXT: v_rsq_f32_e64 v0, -v0
; GCN-IEEE-UNSAFE-NEXT: v_mul_f32_e64 v0, -v0, v1
; GCN-IEEE-UNSAFE-NEXT: s_setpc_b64 s[30:31]
;
; GCN-DAZ-SAFE-LABEL: v_neg_rsq_neg_f32_foldable_user:
; GCN-DAZ-SAFE: ; %bb.0:
; GCN-DAZ-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-DAZ-SAFE-NEXT: v_sqrt_f32_e64 v0, -v0
; GCN-DAZ-SAFE-NEXT: v_rcp_f32_e64 v0, -v0
; GCN-DAZ-SAFE-NEXT: v_mul_f32_e32 v0, v0, v1
; GCN-DAZ-SAFE-NEXT: s_setpc_b64 s[30:31]
;
; SI-IEEE-SAFE-LABEL: v_neg_rsq_neg_f32_foldable_user:
; SI-IEEE-SAFE: ; %bb.0:
; SI-IEEE-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-IEEE-SAFE-NEXT: v_sqrt_f32_e64 v0, -v0
; SI-IEEE-SAFE-NEXT: s_mov_b32 s4, 0x7f800000
; SI-IEEE-SAFE-NEXT: v_frexp_mant_f32_e64 v2, -v0
; SI-IEEE-SAFE-NEXT: v_cmp_lt_f32_e64 s[4:5], |v0|, s4
; SI-IEEE-SAFE-NEXT: v_cndmask_b32_e64 v2, -v0, v2, s[4:5]
; SI-IEEE-SAFE-NEXT: v_rcp_f32_e32 v2, v2
; SI-IEEE-SAFE-NEXT: v_frexp_exp_i32_f32_e32 v0, v0
; SI-IEEE-SAFE-NEXT: v_sub_i32_e32 v0, vcc, 0, v0
; SI-IEEE-SAFE-NEXT: v_ldexp_f32_e32 v0, v2, v0
; SI-IEEE-SAFE-NEXT: v_mul_f32_e32 v0, v0, v1
; SI-IEEE-SAFE-NEXT: s_setpc_b64 s[30:31]
;
; CI-IEEE-SAFE-LABEL: v_neg_rsq_neg_f32_foldable_user:
; CI-IEEE-SAFE: ; %bb.0:
; CI-IEEE-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; CI-IEEE-SAFE-NEXT: v_sqrt_f32_e64 v0, -v0
; CI-IEEE-SAFE-NEXT: v_frexp_mant_f32_e64 v2, -v0
; CI-IEEE-SAFE-NEXT: v_rcp_f32_e32 v2, v2
; CI-IEEE-SAFE-NEXT: v_frexp_exp_i32_f32_e32 v0, v0
; CI-IEEE-SAFE-NEXT: v_sub_i32_e32 v0, vcc, 0, v0
; CI-IEEE-SAFE-NEXT: v_ldexp_f32_e32 v0, v2, v0
; CI-IEEE-SAFE-NEXT: v_mul_f32_e32 v0, v0, v1
; CI-IEEE-SAFE-NEXT: s_setpc_b64 s[30:31]
%val0.neg = fneg float %val0
%sqrt = call contract float @llvm.sqrt.f32(float %val0.neg)
%div = fdiv contract float -1.0, %sqrt, !fpmath !0
%user = fmul contract float %div, %val1
ret float %user
}
define <2 x float> @v_neg_rsq_neg_v2f32_foldable_user(<2 x float> %val0, <2 x float> %val1) {
; GCN-DAZ-UNSAFE-LABEL: v_neg_rsq_neg_v2f32_foldable_user:
; GCN-DAZ-UNSAFE: ; %bb.0:
; GCN-DAZ-UNSAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-DAZ-UNSAFE-NEXT: v_rsq_f32_e64 v0, -v0
; GCN-DAZ-UNSAFE-NEXT: v_rsq_f32_e64 v1, -v1
; GCN-DAZ-UNSAFE-NEXT: v_mul_f32_e64 v0, -v0, v2
; GCN-DAZ-UNSAFE-NEXT: v_mul_f32_e64 v1, -v1, v3
; GCN-DAZ-UNSAFE-NEXT: s_setpc_b64 s[30:31]
;
; GCN-IEEE-UNSAFE-LABEL: v_neg_rsq_neg_v2f32_foldable_user:
; GCN-IEEE-UNSAFE: ; %bb.0:
; GCN-IEEE-UNSAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-IEEE-UNSAFE-NEXT: v_rsq_f32_e64 v0, -v0
; GCN-IEEE-UNSAFE-NEXT: v_rsq_f32_e64 v1, -v1
; GCN-IEEE-UNSAFE-NEXT: v_mul_f32_e64 v0, -v0, v2
; GCN-IEEE-UNSAFE-NEXT: v_mul_f32_e64 v1, -v1, v3
; GCN-IEEE-UNSAFE-NEXT: s_setpc_b64 s[30:31]
;
; GCN-DAZ-SAFE-LABEL: v_neg_rsq_neg_v2f32_foldable_user:
; GCN-DAZ-SAFE: ; %bb.0:
; GCN-DAZ-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-DAZ-SAFE-NEXT: v_sqrt_f32_e64 v0, -v0
; GCN-DAZ-SAFE-NEXT: v_sqrt_f32_e64 v1, -v1
; GCN-DAZ-SAFE-NEXT: v_rcp_f32_e64 v0, -v0
; GCN-DAZ-SAFE-NEXT: v_rcp_f32_e64 v1, -v1
; GCN-DAZ-SAFE-NEXT: v_mul_f32_e32 v0, v0, v2
; GCN-DAZ-SAFE-NEXT: v_mul_f32_e32 v1, v1, v3
; GCN-DAZ-SAFE-NEXT: s_setpc_b64 s[30:31]
;
; SI-IEEE-SAFE-LABEL: v_neg_rsq_neg_v2f32_foldable_user:
; SI-IEEE-SAFE: ; %bb.0:
; SI-IEEE-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-IEEE-SAFE-NEXT: v_sqrt_f32_e64 v0, -v0
; SI-IEEE-SAFE-NEXT: s_mov_b32 s6, 0x7f800000
; SI-IEEE-SAFE-NEXT: v_sqrt_f32_e64 v1, -v1
; SI-IEEE-SAFE-NEXT: v_frexp_mant_f32_e64 v4, -v0
; SI-IEEE-SAFE-NEXT: v_cmp_lt_f32_e64 s[4:5], |v0|, s6
; SI-IEEE-SAFE-NEXT: v_cndmask_b32_e64 v4, -v0, v4, s[4:5]
; SI-IEEE-SAFE-NEXT: v_rcp_f32_e32 v4, v4
; SI-IEEE-SAFE-NEXT: v_frexp_exp_i32_f32_e32 v0, v0
; SI-IEEE-SAFE-NEXT: v_sub_i32_e32 v0, vcc, 0, v0
; SI-IEEE-SAFE-NEXT: v_ldexp_f32_e32 v0, v4, v0
; SI-IEEE-SAFE-NEXT: v_frexp_mant_f32_e64 v4, -v1
; SI-IEEE-SAFE-NEXT: v_cmp_lt_f32_e64 s[4:5], |v1|, s6
; SI-IEEE-SAFE-NEXT: v_cndmask_b32_e64 v4, -v1, v4, s[4:5]
; SI-IEEE-SAFE-NEXT: v_rcp_f32_e32 v4, v4
; SI-IEEE-SAFE-NEXT: v_frexp_exp_i32_f32_e32 v1, v1
; SI-IEEE-SAFE-NEXT: v_sub_i32_e32 v1, vcc, 0, v1
; SI-IEEE-SAFE-NEXT: v_ldexp_f32_e32 v1, v4, v1
; SI-IEEE-SAFE-NEXT: v_mul_f32_e32 v0, v0, v2
; SI-IEEE-SAFE-NEXT: v_mul_f32_e32 v1, v1, v3
; SI-IEEE-SAFE-NEXT: s_setpc_b64 s[30:31]
;
; CI-IEEE-SAFE-LABEL: v_neg_rsq_neg_v2f32_foldable_user:
; CI-IEEE-SAFE: ; %bb.0:
; CI-IEEE-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; CI-IEEE-SAFE-NEXT: v_sqrt_f32_e64 v0, -v0
; CI-IEEE-SAFE-NEXT: v_sqrt_f32_e64 v1, -v1
; CI-IEEE-SAFE-NEXT: v_frexp_mant_f32_e64 v4, -v0
; CI-IEEE-SAFE-NEXT: v_rcp_f32_e32 v4, v4
; CI-IEEE-SAFE-NEXT: v_frexp_exp_i32_f32_e32 v0, v0
; CI-IEEE-SAFE-NEXT: v_sub_i32_e32 v0, vcc, 0, v0
; CI-IEEE-SAFE-NEXT: v_ldexp_f32_e32 v0, v4, v0
; CI-IEEE-SAFE-NEXT: v_frexp_mant_f32_e64 v4, -v1
; CI-IEEE-SAFE-NEXT: v_rcp_f32_e32 v4, v4
; CI-IEEE-SAFE-NEXT: v_frexp_exp_i32_f32_e32 v1, v1
; CI-IEEE-SAFE-NEXT: v_sub_i32_e32 v1, vcc, 0, v1
; CI-IEEE-SAFE-NEXT: v_ldexp_f32_e32 v1, v4, v1
; CI-IEEE-SAFE-NEXT: v_mul_f32_e32 v0, v0, v2
; CI-IEEE-SAFE-NEXT: v_mul_f32_e32 v1, v1, v3
; CI-IEEE-SAFE-NEXT: s_setpc_b64 s[30:31]
%val0.fneg = fneg <2 x float> %val0
%sqrt = call contract <2 x float> @llvm.sqrt.v2f32(<2 x float> %val0.fneg)
%div = fdiv contract <2 x float> <float -1.0, float -1.0>, %sqrt, !fpmath !0
%user = fmul contract <2 x float> %div, %val1
ret <2 x float> %user
}
define float @v_neg_rsq_f32(float %val) {
; GCN-DAZ-UNSAFE-LABEL: v_neg_rsq_f32:
; GCN-DAZ-UNSAFE: ; %bb.0:
; GCN-DAZ-UNSAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-DAZ-UNSAFE-NEXT: v_rsq_f32_e32 v0, v0
; GCN-DAZ-UNSAFE-NEXT: v_sub_f32_e32 v0, 0x80000000, v0
; GCN-DAZ-UNSAFE-NEXT: s_setpc_b64 s[30:31]
;
; GCN-IEEE-UNSAFE-LABEL: v_neg_rsq_f32:
; GCN-IEEE-UNSAFE: ; %bb.0:
; GCN-IEEE-UNSAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-IEEE-UNSAFE-NEXT: v_rsq_f32_e32 v0, v0
; GCN-IEEE-UNSAFE-NEXT: v_xor_b32_e32 v0, 0x80000000, v0
; GCN-IEEE-UNSAFE-NEXT: s_setpc_b64 s[30:31]
;
; GCN-DAZ-SAFE-LABEL: v_neg_rsq_f32:
; GCN-DAZ-SAFE: ; %bb.0:
; GCN-DAZ-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-DAZ-SAFE-NEXT: v_sqrt_f32_e32 v0, v0
; GCN-DAZ-SAFE-NEXT: v_rcp_f32_e64 v0, -v0
; GCN-DAZ-SAFE-NEXT: s_setpc_b64 s[30:31]
;
; SI-IEEE-SAFE-LABEL: v_neg_rsq_f32:
; SI-IEEE-SAFE: ; %bb.0:
; SI-IEEE-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-IEEE-SAFE-NEXT: v_sqrt_f32_e32 v0, v0
; SI-IEEE-SAFE-NEXT: s_mov_b32 s4, 0x7f800000
; SI-IEEE-SAFE-NEXT: v_frexp_mant_f32_e64 v1, -v0
; SI-IEEE-SAFE-NEXT: v_cmp_lt_f32_e64 s[4:5], |v0|, s4
; SI-IEEE-SAFE-NEXT: v_cndmask_b32_e64 v1, -v0, v1, s[4:5]
; SI-IEEE-SAFE-NEXT: v_rcp_f32_e32 v1, v1
; SI-IEEE-SAFE-NEXT: v_frexp_exp_i32_f32_e32 v0, v0
; SI-IEEE-SAFE-NEXT: v_sub_i32_e32 v0, vcc, 0, v0
; SI-IEEE-SAFE-NEXT: v_ldexp_f32_e32 v0, v1, v0
; SI-IEEE-SAFE-NEXT: s_setpc_b64 s[30:31]
;
; CI-IEEE-SAFE-LABEL: v_neg_rsq_f32:
; CI-IEEE-SAFE: ; %bb.0:
; CI-IEEE-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; CI-IEEE-SAFE-NEXT: v_sqrt_f32_e32 v0, v0
; CI-IEEE-SAFE-NEXT: v_frexp_mant_f32_e64 v1, -v0
; CI-IEEE-SAFE-NEXT: v_rcp_f32_e32 v1, v1
; CI-IEEE-SAFE-NEXT: v_frexp_exp_i32_f32_e32 v0, v0
; CI-IEEE-SAFE-NEXT: v_sub_i32_e32 v0, vcc, 0, v0
; CI-IEEE-SAFE-NEXT: v_ldexp_f32_e32 v0, v1, v0
; CI-IEEE-SAFE-NEXT: s_setpc_b64 s[30:31]
%sqrt = call contract float @llvm.sqrt.f32(float %val)
%div = fdiv contract float -1.0, %sqrt, !fpmath !0
ret float %div
}
define <2 x float> @v_neg_rsq_v2f32(<2 x float> %val) {
; GCN-DAZ-UNSAFE-LABEL: v_neg_rsq_v2f32:
; GCN-DAZ-UNSAFE: ; %bb.0:
; GCN-DAZ-UNSAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-DAZ-UNSAFE-NEXT: v_rsq_f32_e32 v0, v0
; GCN-DAZ-UNSAFE-NEXT: v_rsq_f32_e32 v1, v1
; GCN-DAZ-UNSAFE-NEXT: v_sub_f32_e32 v0, 0x80000000, v0
; GCN-DAZ-UNSAFE-NEXT: v_sub_f32_e32 v1, 0x80000000, v1
; GCN-DAZ-UNSAFE-NEXT: s_setpc_b64 s[30:31]
;
; GCN-IEEE-UNSAFE-LABEL: v_neg_rsq_v2f32:
; GCN-IEEE-UNSAFE: ; %bb.0:
; GCN-IEEE-UNSAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-IEEE-UNSAFE-NEXT: v_rsq_f32_e32 v0, v0
; GCN-IEEE-UNSAFE-NEXT: v_rsq_f32_e32 v1, v1
; GCN-IEEE-UNSAFE-NEXT: v_xor_b32_e32 v0, 0x80000000, v0
; GCN-IEEE-UNSAFE-NEXT: v_xor_b32_e32 v1, 0x80000000, v1
; GCN-IEEE-UNSAFE-NEXT: s_setpc_b64 s[30:31]
;
; GCN-DAZ-SAFE-LABEL: v_neg_rsq_v2f32:
; GCN-DAZ-SAFE: ; %bb.0:
; GCN-DAZ-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-DAZ-SAFE-NEXT: v_sqrt_f32_e32 v0, v0
; GCN-DAZ-SAFE-NEXT: v_sqrt_f32_e32 v1, v1
; GCN-DAZ-SAFE-NEXT: v_rcp_f32_e64 v0, -v0
; GCN-DAZ-SAFE-NEXT: v_rcp_f32_e64 v1, -v1
; GCN-DAZ-SAFE-NEXT: s_setpc_b64 s[30:31]
;
; SI-IEEE-SAFE-LABEL: v_neg_rsq_v2f32:
; SI-IEEE-SAFE: ; %bb.0:
; SI-IEEE-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-IEEE-SAFE-NEXT: v_sqrt_f32_e32 v0, v0
; SI-IEEE-SAFE-NEXT: s_mov_b32 s6, 0x7f800000
; SI-IEEE-SAFE-NEXT: v_sqrt_f32_e32 v1, v1
; SI-IEEE-SAFE-NEXT: v_frexp_mant_f32_e64 v2, -v0
; SI-IEEE-SAFE-NEXT: v_cmp_lt_f32_e64 s[4:5], |v0|, s6
; SI-IEEE-SAFE-NEXT: v_cndmask_b32_e64 v2, -v0, v2, s[4:5]
; SI-IEEE-SAFE-NEXT: v_rcp_f32_e32 v2, v2
; SI-IEEE-SAFE-NEXT: v_frexp_exp_i32_f32_e32 v0, v0
; SI-IEEE-SAFE-NEXT: v_sub_i32_e32 v0, vcc, 0, v0
; SI-IEEE-SAFE-NEXT: v_ldexp_f32_e32 v0, v2, v0
; SI-IEEE-SAFE-NEXT: v_frexp_mant_f32_e64 v2, -v1
; SI-IEEE-SAFE-NEXT: v_cmp_lt_f32_e64 s[4:5], |v1|, s6
; SI-IEEE-SAFE-NEXT: v_cndmask_b32_e64 v2, -v1, v2, s[4:5]
; SI-IEEE-SAFE-NEXT: v_rcp_f32_e32 v2, v2
; SI-IEEE-SAFE-NEXT: v_frexp_exp_i32_f32_e32 v1, v1
; SI-IEEE-SAFE-NEXT: v_sub_i32_e32 v1, vcc, 0, v1
; SI-IEEE-SAFE-NEXT: v_ldexp_f32_e32 v1, v2, v1
; SI-IEEE-SAFE-NEXT: s_setpc_b64 s[30:31]
;
; CI-IEEE-SAFE-LABEL: v_neg_rsq_v2f32:
; CI-IEEE-SAFE: ; %bb.0:
; CI-IEEE-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; CI-IEEE-SAFE-NEXT: v_sqrt_f32_e32 v0, v0
; CI-IEEE-SAFE-NEXT: v_sqrt_f32_e32 v1, v1
; CI-IEEE-SAFE-NEXT: v_frexp_mant_f32_e64 v2, -v0
; CI-IEEE-SAFE-NEXT: v_rcp_f32_e32 v2, v2
; CI-IEEE-SAFE-NEXT: v_frexp_exp_i32_f32_e32 v0, v0
; CI-IEEE-SAFE-NEXT: v_frexp_mant_f32_e64 v3, -v1
; CI-IEEE-SAFE-NEXT: v_sub_i32_e32 v0, vcc, 0, v0
; CI-IEEE-SAFE-NEXT: v_ldexp_f32_e32 v0, v2, v0
; CI-IEEE-SAFE-NEXT: v_rcp_f32_e32 v2, v3
; CI-IEEE-SAFE-NEXT: v_frexp_exp_i32_f32_e32 v1, v1
; CI-IEEE-SAFE-NEXT: v_sub_i32_e32 v1, vcc, 0, v1
; CI-IEEE-SAFE-NEXT: v_ldexp_f32_e32 v1, v2, v1
; CI-IEEE-SAFE-NEXT: s_setpc_b64 s[30:31]
%sqrt = call contract <2 x float> @llvm.sqrt.v2f32(<2 x float> %val)
%div = fdiv contract <2 x float> <float -1.0, float -1.0>, %sqrt, !fpmath !0
ret <2 x float> %div
}
define float @v_neg_rsq_f32_foldable_user(float %val0, float %val1) {
; GCN-DAZ-UNSAFE-LABEL: v_neg_rsq_f32_foldable_user:
; GCN-DAZ-UNSAFE: ; %bb.0:
; GCN-DAZ-UNSAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-DAZ-UNSAFE-NEXT: v_rsq_f32_e32 v0, v0
; GCN-DAZ-UNSAFE-NEXT: v_mul_f32_e64 v0, -v0, v1
; GCN-DAZ-UNSAFE-NEXT: s_setpc_b64 s[30:31]
;
; GCN-IEEE-UNSAFE-LABEL: v_neg_rsq_f32_foldable_user:
; GCN-IEEE-UNSAFE: ; %bb.0:
; GCN-IEEE-UNSAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-IEEE-UNSAFE-NEXT: v_rsq_f32_e32 v0, v0
; GCN-IEEE-UNSAFE-NEXT: v_mul_f32_e64 v0, -v0, v1
; GCN-IEEE-UNSAFE-NEXT: s_setpc_b64 s[30:31]
;
; GCN-DAZ-SAFE-LABEL: v_neg_rsq_f32_foldable_user:
; GCN-DAZ-SAFE: ; %bb.0:
; GCN-DAZ-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-DAZ-SAFE-NEXT: v_sqrt_f32_e32 v0, v0
; GCN-DAZ-SAFE-NEXT: v_rcp_f32_e64 v0, -v0
; GCN-DAZ-SAFE-NEXT: v_mul_f32_e32 v0, v0, v1
; GCN-DAZ-SAFE-NEXT: s_setpc_b64 s[30:31]
;
; SI-IEEE-SAFE-LABEL: v_neg_rsq_f32_foldable_user:
; SI-IEEE-SAFE: ; %bb.0:
; SI-IEEE-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-IEEE-SAFE-NEXT: v_sqrt_f32_e32 v0, v0
; SI-IEEE-SAFE-NEXT: s_mov_b32 s4, 0x7f800000
; SI-IEEE-SAFE-NEXT: v_frexp_mant_f32_e64 v2, -v0
; SI-IEEE-SAFE-NEXT: v_cmp_lt_f32_e64 s[4:5], |v0|, s4
; SI-IEEE-SAFE-NEXT: v_cndmask_b32_e64 v2, -v0, v2, s[4:5]
; SI-IEEE-SAFE-NEXT: v_rcp_f32_e32 v2, v2
; SI-IEEE-SAFE-NEXT: v_frexp_exp_i32_f32_e32 v0, v0
; SI-IEEE-SAFE-NEXT: v_sub_i32_e32 v0, vcc, 0, v0
; SI-IEEE-SAFE-NEXT: v_ldexp_f32_e32 v0, v2, v0
; SI-IEEE-SAFE-NEXT: v_mul_f32_e32 v0, v0, v1
; SI-IEEE-SAFE-NEXT: s_setpc_b64 s[30:31]
;
; CI-IEEE-SAFE-LABEL: v_neg_rsq_f32_foldable_user:
; CI-IEEE-SAFE: ; %bb.0:
; CI-IEEE-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; CI-IEEE-SAFE-NEXT: v_sqrt_f32_e32 v0, v0
; CI-IEEE-SAFE-NEXT: v_frexp_mant_f32_e64 v2, -v0
; CI-IEEE-SAFE-NEXT: v_rcp_f32_e32 v2, v2
; CI-IEEE-SAFE-NEXT: v_frexp_exp_i32_f32_e32 v0, v0
; CI-IEEE-SAFE-NEXT: v_sub_i32_e32 v0, vcc, 0, v0
; CI-IEEE-SAFE-NEXT: v_ldexp_f32_e32 v0, v2, v0
; CI-IEEE-SAFE-NEXT: v_mul_f32_e32 v0, v0, v1
; CI-IEEE-SAFE-NEXT: s_setpc_b64 s[30:31]
%sqrt = call contract float @llvm.sqrt.f32(float %val0)
%div = fdiv contract float -1.0, %sqrt, !fpmath !0
%user = fmul contract float %div, %val1
ret float %user
}
define <2 x float> @v_neg_rsq_v2f32_foldable_user(<2 x float> %val0, <2 x float> %val1) {
; GCN-DAZ-UNSAFE-LABEL: v_neg_rsq_v2f32_foldable_user:
; GCN-DAZ-UNSAFE: ; %bb.0:
; GCN-DAZ-UNSAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-DAZ-UNSAFE-NEXT: v_rsq_f32_e32 v0, v0
; GCN-DAZ-UNSAFE-NEXT: v_rsq_f32_e32 v1, v1
; GCN-DAZ-UNSAFE-NEXT: v_mul_f32_e64 v0, -v0, v2
; GCN-DAZ-UNSAFE-NEXT: v_mul_f32_e64 v1, -v1, v3
; GCN-DAZ-UNSAFE-NEXT: s_setpc_b64 s[30:31]
;
; GCN-IEEE-UNSAFE-LABEL: v_neg_rsq_v2f32_foldable_user:
; GCN-IEEE-UNSAFE: ; %bb.0:
; GCN-IEEE-UNSAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-IEEE-UNSAFE-NEXT: v_rsq_f32_e32 v0, v0
; GCN-IEEE-UNSAFE-NEXT: v_rsq_f32_e32 v1, v1
; GCN-IEEE-UNSAFE-NEXT: v_mul_f32_e64 v0, -v0, v2
; GCN-IEEE-UNSAFE-NEXT: v_mul_f32_e64 v1, -v1, v3
; GCN-IEEE-UNSAFE-NEXT: s_setpc_b64 s[30:31]
;
; GCN-DAZ-SAFE-LABEL: v_neg_rsq_v2f32_foldable_user:
; GCN-DAZ-SAFE: ; %bb.0:
; GCN-DAZ-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-DAZ-SAFE-NEXT: v_sqrt_f32_e32 v0, v0
; GCN-DAZ-SAFE-NEXT: v_sqrt_f32_e32 v1, v1
; GCN-DAZ-SAFE-NEXT: v_rcp_f32_e64 v0, -v0
; GCN-DAZ-SAFE-NEXT: v_rcp_f32_e64 v1, -v1
; GCN-DAZ-SAFE-NEXT: v_mul_f32_e32 v0, v0, v2
; GCN-DAZ-SAFE-NEXT: v_mul_f32_e32 v1, v1, v3
; GCN-DAZ-SAFE-NEXT: s_setpc_b64 s[30:31]
;
; SI-IEEE-SAFE-LABEL: v_neg_rsq_v2f32_foldable_user:
; SI-IEEE-SAFE: ; %bb.0:
; SI-IEEE-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-IEEE-SAFE-NEXT: v_sqrt_f32_e32 v0, v0
; SI-IEEE-SAFE-NEXT: s_mov_b32 s6, 0x7f800000
; SI-IEEE-SAFE-NEXT: v_sqrt_f32_e32 v1, v1
; SI-IEEE-SAFE-NEXT: v_frexp_mant_f32_e64 v4, -v0
; SI-IEEE-SAFE-NEXT: v_cmp_lt_f32_e64 s[4:5], |v0|, s6
; SI-IEEE-SAFE-NEXT: v_cndmask_b32_e64 v4, -v0, v4, s[4:5]
; SI-IEEE-SAFE-NEXT: v_rcp_f32_e32 v4, v4
; SI-IEEE-SAFE-NEXT: v_frexp_exp_i32_f32_e32 v0, v0
; SI-IEEE-SAFE-NEXT: v_sub_i32_e32 v0, vcc, 0, v0
; SI-IEEE-SAFE-NEXT: v_ldexp_f32_e32 v0, v4, v0
; SI-IEEE-SAFE-NEXT: v_frexp_mant_f32_e64 v4, -v1
; SI-IEEE-SAFE-NEXT: v_cmp_lt_f32_e64 s[4:5], |v1|, s6
; SI-IEEE-SAFE-NEXT: v_cndmask_b32_e64 v4, -v1, v4, s[4:5]
; SI-IEEE-SAFE-NEXT: v_rcp_f32_e32 v4, v4
; SI-IEEE-SAFE-NEXT: v_frexp_exp_i32_f32_e32 v1, v1
; SI-IEEE-SAFE-NEXT: v_sub_i32_e32 v1, vcc, 0, v1
; SI-IEEE-SAFE-NEXT: v_ldexp_f32_e32 v1, v4, v1
; SI-IEEE-SAFE-NEXT: v_mul_f32_e32 v0, v0, v2
; SI-IEEE-SAFE-NEXT: v_mul_f32_e32 v1, v1, v3
; SI-IEEE-SAFE-NEXT: s_setpc_b64 s[30:31]
;
; CI-IEEE-SAFE-LABEL: v_neg_rsq_v2f32_foldable_user:
; CI-IEEE-SAFE: ; %bb.0:
; CI-IEEE-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; CI-IEEE-SAFE-NEXT: v_sqrt_f32_e32 v0, v0
; CI-IEEE-SAFE-NEXT: v_sqrt_f32_e32 v1, v1
; CI-IEEE-SAFE-NEXT: v_frexp_mant_f32_e64 v4, -v0
; CI-IEEE-SAFE-NEXT: v_rcp_f32_e32 v4, v4
; CI-IEEE-SAFE-NEXT: v_frexp_exp_i32_f32_e32 v0, v0
; CI-IEEE-SAFE-NEXT: v_sub_i32_e32 v0, vcc, 0, v0
; CI-IEEE-SAFE-NEXT: v_ldexp_f32_e32 v0, v4, v0
; CI-IEEE-SAFE-NEXT: v_frexp_mant_f32_e64 v4, -v1
; CI-IEEE-SAFE-NEXT: v_rcp_f32_e32 v4, v4
; CI-IEEE-SAFE-NEXT: v_frexp_exp_i32_f32_e32 v1, v1
; CI-IEEE-SAFE-NEXT: v_sub_i32_e32 v1, vcc, 0, v1
; CI-IEEE-SAFE-NEXT: v_ldexp_f32_e32 v1, v4, v1
; CI-IEEE-SAFE-NEXT: v_mul_f32_e32 v0, v0, v2
; CI-IEEE-SAFE-NEXT: v_mul_f32_e32 v1, v1, v3
; CI-IEEE-SAFE-NEXT: s_setpc_b64 s[30:31]
%sqrt = call contract <2 x float> @llvm.sqrt.v2f32(<2 x float> %val0)
%div = fdiv contract <2 x float> <float -1.0, float -1.0>, %sqrt, !fpmath !0
%user = fmul contract <2 x float> %div, %val1
ret <2 x float> %user
}
define float @v_rsq_f32(float %val) {
; GCN-DAZ-LABEL: v_rsq_f32:
; GCN-DAZ: ; %bb.0:
; GCN-DAZ-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-DAZ-NEXT: v_rsq_f32_e32 v0, v0
; GCN-DAZ-NEXT: s_setpc_b64 s[30:31]
;
; GCN-IEEE-UNSAFE-LABEL: v_rsq_f32:
; GCN-IEEE-UNSAFE: ; %bb.0:
; GCN-IEEE-UNSAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-IEEE-UNSAFE-NEXT: v_rsq_f32_e32 v0, v0
; GCN-IEEE-UNSAFE-NEXT: s_setpc_b64 s[30:31]
;
; GCN-IEEE-SAFE-LABEL: v_rsq_f32:
; GCN-IEEE-SAFE: ; %bb.0:
; GCN-IEEE-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-IEEE-SAFE-NEXT: s_mov_b32 s4, 0x800000
; GCN-IEEE-SAFE-NEXT: v_mov_b32_e32 v1, 0x4b800000
; GCN-IEEE-SAFE-NEXT: v_cmp_gt_f32_e32 vcc, s4, v0
; GCN-IEEE-SAFE-NEXT: v_cndmask_b32_e32 v1, 1.0, v1, vcc
; GCN-IEEE-SAFE-NEXT: v_mul_f32_e32 v0, v0, v1
; GCN-IEEE-SAFE-NEXT: v_rsq_f32_e32 v0, v0
; GCN-IEEE-SAFE-NEXT: v_mov_b32_e32 v1, 0x45800000
; GCN-IEEE-SAFE-NEXT: v_cndmask_b32_e32 v1, 1.0, v1, vcc
; GCN-IEEE-SAFE-NEXT: v_mul_f32_e32 v0, v0, v1
; GCN-IEEE-SAFE-NEXT: s_setpc_b64 s[30:31]
%sqrt = call contract float @llvm.sqrt.f32(float %val), !fpmath !1
%div = fdiv contract float 1.0, %sqrt, !fpmath !1
ret float %div
}
define float @v_rsq_f32_missing_contract0(float %val) {
; GCN-DAZ-LABEL: v_rsq_f32_missing_contract0:
; GCN-DAZ: ; %bb.0:
; GCN-DAZ-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-DAZ-NEXT: v_rsq_f32_e32 v0, v0
; GCN-DAZ-NEXT: s_setpc_b64 s[30:31]
;
; GCN-IEEE-UNSAFE-LABEL: v_rsq_f32_missing_contract0:
; GCN-IEEE-UNSAFE: ; %bb.0:
; GCN-IEEE-UNSAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-IEEE-UNSAFE-NEXT: v_rsq_f32_e32 v0, v0
; GCN-IEEE-UNSAFE-NEXT: s_setpc_b64 s[30:31]
;
; SI-IEEE-SAFE-LABEL: v_rsq_f32_missing_contract0:
; SI-IEEE-SAFE: ; %bb.0:
; SI-IEEE-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-IEEE-SAFE-NEXT: v_sqrt_f32_e32 v0, v0
; SI-IEEE-SAFE-NEXT: s_mov_b32 s4, 0x7f800000
; SI-IEEE-SAFE-NEXT: v_frexp_mant_f32_e32 v1, v0
; SI-IEEE-SAFE-NEXT: v_cmp_lt_f32_e64 vcc, |v0|, s4
; SI-IEEE-SAFE-NEXT: v_cndmask_b32_e32 v1, v0, v1, vcc
; SI-IEEE-SAFE-NEXT: v_rcp_f32_e32 v1, v1
; SI-IEEE-SAFE-NEXT: v_frexp_exp_i32_f32_e32 v0, v0
; SI-IEEE-SAFE-NEXT: v_sub_i32_e32 v0, vcc, 0, v0
; SI-IEEE-SAFE-NEXT: v_ldexp_f32_e32 v0, v1, v0
; SI-IEEE-SAFE-NEXT: s_setpc_b64 s[30:31]
;
; CI-IEEE-SAFE-LABEL: v_rsq_f32_missing_contract0:
; CI-IEEE-SAFE: ; %bb.0:
; CI-IEEE-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; CI-IEEE-SAFE-NEXT: v_sqrt_f32_e32 v0, v0
; CI-IEEE-SAFE-NEXT: v_frexp_mant_f32_e32 v1, v0
; CI-IEEE-SAFE-NEXT: v_rcp_f32_e32 v1, v1
; CI-IEEE-SAFE-NEXT: v_frexp_exp_i32_f32_e32 v0, v0
; CI-IEEE-SAFE-NEXT: v_sub_i32_e32 v0, vcc, 0, v0
; CI-IEEE-SAFE-NEXT: v_ldexp_f32_e32 v0, v1, v0
; CI-IEEE-SAFE-NEXT: s_setpc_b64 s[30:31]
%sqrt = call float @llvm.sqrt.f32(float %val), !fpmath !1
%div = fdiv contract float 1.0, %sqrt, !fpmath !1
ret float %div
}
define float @v_rsq_f32_missing_contract1(float %val) {
; GCN-DAZ-LABEL: v_rsq_f32_missing_contract1:
; GCN-DAZ: ; %bb.0:
; GCN-DAZ-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-DAZ-NEXT: v_rsq_f32_e32 v0, v0
; GCN-DAZ-NEXT: s_setpc_b64 s[30:31]
;
; GCN-IEEE-UNSAFE-LABEL: v_rsq_f32_missing_contract1:
; GCN-IEEE-UNSAFE: ; %bb.0:
; GCN-IEEE-UNSAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-IEEE-UNSAFE-NEXT: v_rsq_f32_e32 v0, v0
; GCN-IEEE-UNSAFE-NEXT: s_setpc_b64 s[30:31]
;
; SI-IEEE-SAFE-LABEL: v_rsq_f32_missing_contract1:
; SI-IEEE-SAFE: ; %bb.0:
; SI-IEEE-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-IEEE-SAFE-NEXT: v_sqrt_f32_e32 v0, v0
; SI-IEEE-SAFE-NEXT: s_mov_b32 s4, 0x7f800000
; SI-IEEE-SAFE-NEXT: v_frexp_mant_f32_e32 v1, v0
; SI-IEEE-SAFE-NEXT: v_cmp_lt_f32_e64 vcc, |v0|, s4
; SI-IEEE-SAFE-NEXT: v_cndmask_b32_e32 v1, v0, v1, vcc
; SI-IEEE-SAFE-NEXT: v_rcp_f32_e32 v1, v1
; SI-IEEE-SAFE-NEXT: v_frexp_exp_i32_f32_e32 v0, v0
; SI-IEEE-SAFE-NEXT: v_sub_i32_e32 v0, vcc, 0, v0
; SI-IEEE-SAFE-NEXT: v_ldexp_f32_e32 v0, v1, v0
; SI-IEEE-SAFE-NEXT: s_setpc_b64 s[30:31]
;
; CI-IEEE-SAFE-LABEL: v_rsq_f32_missing_contract1:
; CI-IEEE-SAFE: ; %bb.0:
; CI-IEEE-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; CI-IEEE-SAFE-NEXT: v_sqrt_f32_e32 v0, v0
; CI-IEEE-SAFE-NEXT: v_frexp_mant_f32_e32 v1, v0
; CI-IEEE-SAFE-NEXT: v_rcp_f32_e32 v1, v1
; CI-IEEE-SAFE-NEXT: v_frexp_exp_i32_f32_e32 v0, v0
; CI-IEEE-SAFE-NEXT: v_sub_i32_e32 v0, vcc, 0, v0
; CI-IEEE-SAFE-NEXT: v_ldexp_f32_e32 v0, v1, v0
; CI-IEEE-SAFE-NEXT: s_setpc_b64 s[30:31]
%sqrt = call contract float @llvm.sqrt.f32(float %val), !fpmath !1
%div = fdiv float 1.0, %sqrt, !fpmath !1
ret float %div
}
; Test that we contract into FMA for an fadd user after introducing
; the fmul.
define float @v_rsq_f32_contractable_user(float %val0, float %val1) {
; GCN-DAZ-LABEL: v_rsq_f32_contractable_user:
; GCN-DAZ: ; %bb.0:
; GCN-DAZ-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-DAZ-NEXT: v_rsq_f32_e32 v0, v0
; GCN-DAZ-NEXT: v_add_f32_e32 v0, v0, v1
; GCN-DAZ-NEXT: s_setpc_b64 s[30:31]
;
; GCN-IEEE-UNSAFE-LABEL: v_rsq_f32_contractable_user:
; GCN-IEEE-UNSAFE: ; %bb.0:
; GCN-IEEE-UNSAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-IEEE-UNSAFE-NEXT: v_rsq_f32_e32 v0, v0
; GCN-IEEE-UNSAFE-NEXT: v_add_f32_e32 v0, v0, v1
; GCN-IEEE-UNSAFE-NEXT: s_setpc_b64 s[30:31]
;
; GCN-IEEE-SAFE-LABEL: v_rsq_f32_contractable_user:
; GCN-IEEE-SAFE: ; %bb.0:
; GCN-IEEE-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-IEEE-SAFE-NEXT: s_mov_b32 s4, 0x800000
; GCN-IEEE-SAFE-NEXT: v_mov_b32_e32 v2, 0x4b800000
; GCN-IEEE-SAFE-NEXT: v_cmp_gt_f32_e32 vcc, s4, v0
; GCN-IEEE-SAFE-NEXT: v_cndmask_b32_e32 v2, 1.0, v2, vcc
; GCN-IEEE-SAFE-NEXT: v_mul_f32_e32 v0, v0, v2
; GCN-IEEE-SAFE-NEXT: v_rsq_f32_e32 v0, v0
; GCN-IEEE-SAFE-NEXT: v_mov_b32_e32 v2, 0x45800000
; GCN-IEEE-SAFE-NEXT: v_cndmask_b32_e32 v2, 1.0, v2, vcc
; GCN-IEEE-SAFE-NEXT: v_fma_f32 v0, v0, v2, v1
; GCN-IEEE-SAFE-NEXT: s_setpc_b64 s[30:31]
%sqrt = call contract float @llvm.sqrt.f32(float %val0), !fpmath !1
%div = fdiv contract float 1.0, %sqrt, !fpmath !1
%add = fadd contract float %div, %val1
ret float %add
}
; Missing contract on the fdiv
define float @v_rsq_f32_contractable_user_missing_contract0(float %val0, float %val1) {
; GCN-DAZ-LABEL: v_rsq_f32_contractable_user_missing_contract0:
; GCN-DAZ: ; %bb.0:
; GCN-DAZ-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-DAZ-NEXT: v_rsq_f32_e32 v0, v0
; GCN-DAZ-NEXT: v_add_f32_e32 v0, v0, v1
; GCN-DAZ-NEXT: s_setpc_b64 s[30:31]
;
; GCN-IEEE-UNSAFE-LABEL: v_rsq_f32_contractable_user_missing_contract0:
; GCN-IEEE-UNSAFE: ; %bb.0:
; GCN-IEEE-UNSAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-IEEE-UNSAFE-NEXT: v_rsq_f32_e32 v0, v0
; GCN-IEEE-UNSAFE-NEXT: v_add_f32_e32 v0, v0, v1
; GCN-IEEE-UNSAFE-NEXT: s_setpc_b64 s[30:31]
;
; GCN-IEEE-SAFE-LABEL: v_rsq_f32_contractable_user_missing_contract0:
; GCN-IEEE-SAFE: ; %bb.0:
; GCN-IEEE-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-IEEE-SAFE-NEXT: s_mov_b32 s4, 0x800000
; GCN-IEEE-SAFE-NEXT: v_mov_b32_e32 v2, 0x4b800000
; GCN-IEEE-SAFE-NEXT: v_cmp_gt_f32_e32 vcc, s4, v0
; GCN-IEEE-SAFE-NEXT: v_cndmask_b32_e32 v2, 1.0, v2, vcc
; GCN-IEEE-SAFE-NEXT: v_mul_f32_e32 v0, v0, v2
; GCN-IEEE-SAFE-NEXT: v_rsq_f32_e32 v0, v0
; GCN-IEEE-SAFE-NEXT: v_mov_b32_e32 v2, 0x45800000
; GCN-IEEE-SAFE-NEXT: v_cndmask_b32_e32 v2, 1.0, v2, vcc
; GCN-IEEE-SAFE-NEXT: v_fma_f32 v0, v0, v2, v1
; GCN-IEEE-SAFE-NEXT: s_setpc_b64 s[30:31]
%sqrt = call contract float @llvm.sqrt.f32(float %val0), !fpmath !1
%div = fdiv contract float 1.0, %sqrt, !fpmath !1
%add = fadd contract float %div, %val1
ret float %add
}
; Missing contract on the fadd
define float @v_rsq_f32_contractable_user_missing_contract1(float %val0, float %val1) {
; GCN-DAZ-LABEL: v_rsq_f32_contractable_user_missing_contract1:
; GCN-DAZ: ; %bb.0:
; GCN-DAZ-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-DAZ-NEXT: v_rsq_f32_e32 v0, v0
; GCN-DAZ-NEXT: v_add_f32_e32 v0, v0, v1
; GCN-DAZ-NEXT: s_setpc_b64 s[30:31]
;
; GCN-IEEE-UNSAFE-LABEL: v_rsq_f32_contractable_user_missing_contract1:
; GCN-IEEE-UNSAFE: ; %bb.0:
; GCN-IEEE-UNSAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-IEEE-UNSAFE-NEXT: v_rsq_f32_e32 v0, v0
; GCN-IEEE-UNSAFE-NEXT: v_add_f32_e32 v0, v0, v1
; GCN-IEEE-UNSAFE-NEXT: s_setpc_b64 s[30:31]
;
; GCN-IEEE-SAFE-LABEL: v_rsq_f32_contractable_user_missing_contract1:
; GCN-IEEE-SAFE: ; %bb.0:
; GCN-IEEE-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-IEEE-SAFE-NEXT: s_mov_b32 s4, 0x800000
; GCN-IEEE-SAFE-NEXT: v_mov_b32_e32 v2, 0x4b800000
; GCN-IEEE-SAFE-NEXT: v_cmp_gt_f32_e32 vcc, s4, v0
; GCN-IEEE-SAFE-NEXT: v_cndmask_b32_e32 v2, 1.0, v2, vcc
; GCN-IEEE-SAFE-NEXT: v_mul_f32_e32 v0, v0, v2
; GCN-IEEE-SAFE-NEXT: v_rsq_f32_e32 v0, v0
; GCN-IEEE-SAFE-NEXT: v_mov_b32_e32 v2, 0x45800000
; GCN-IEEE-SAFE-NEXT: v_cndmask_b32_e32 v2, 1.0, v2, vcc
; GCN-IEEE-SAFE-NEXT: v_mul_f32_e32 v0, v0, v2
; GCN-IEEE-SAFE-NEXT: v_add_f32_e32 v0, v0, v1
; GCN-IEEE-SAFE-NEXT: s_setpc_b64 s[30:31]
%sqrt = call contract float @llvm.sqrt.f32(float %val0), !fpmath !1
%div = fdiv contract float 1.0, %sqrt, !fpmath !1
%add = fadd float %div, %val1
ret float %add
}
define float @v_rsq_f32_known_never_denormal(float nofpclass(sub) %val) {
; GCN-DAZ-LABEL: v_rsq_f32_known_never_denormal:
; GCN-DAZ: ; %bb.0:
; GCN-DAZ-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-DAZ-NEXT: v_rsq_f32_e32 v0, v0
; GCN-DAZ-NEXT: s_setpc_b64 s[30:31]
;
; GCN-IEEE-LABEL: v_rsq_f32_known_never_denormal:
; GCN-IEEE: ; %bb.0:
; GCN-IEEE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-IEEE-NEXT: v_rsq_f32_e32 v0, v0
; GCN-IEEE-NEXT: s_setpc_b64 s[30:31]
%sqrt = call contract float @llvm.sqrt.f32(float %val), !fpmath !1
%div = fdiv contract float 1.0, %sqrt, !fpmath !1
ret float %div
}
define float @v_rsq_f32_known_never_posdenormal(float nofpclass(psub) %val) {
; GCN-DAZ-LABEL: v_rsq_f32_known_never_posdenormal:
; GCN-DAZ: ; %bb.0:
; GCN-DAZ-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-DAZ-NEXT: v_rsq_f32_e32 v0, v0
; GCN-DAZ-NEXT: s_setpc_b64 s[30:31]
;
; GCN-IEEE-UNSAFE-LABEL: v_rsq_f32_known_never_posdenormal:
; GCN-IEEE-UNSAFE: ; %bb.0:
; GCN-IEEE-UNSAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-IEEE-UNSAFE-NEXT: v_rsq_f32_e32 v0, v0
; GCN-IEEE-UNSAFE-NEXT: s_setpc_b64 s[30:31]
;
; GCN-IEEE-SAFE-LABEL: v_rsq_f32_known_never_posdenormal:
; GCN-IEEE-SAFE: ; %bb.0:
; GCN-IEEE-SAFE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-IEEE-SAFE-NEXT: s_mov_b32 s4, 0x800000
; GCN-IEEE-SAFE-NEXT: v_mov_b32_e32 v1, 0x4b800000
; GCN-IEEE-SAFE-NEXT: v_cmp_gt_f32_e32 vcc, s4, v0
; GCN-IEEE-SAFE-NEXT: v_cndmask_b32_e32 v1, 1.0, v1, vcc
; GCN-IEEE-SAFE-NEXT: v_mul_f32_e32 v0, v0, v1
; GCN-IEEE-SAFE-NEXT: v_rsq_f32_e32 v0, v0
; GCN-IEEE-SAFE-NEXT: v_mov_b32_e32 v1, 0x45800000
; GCN-IEEE-SAFE-NEXT: v_cndmask_b32_e32 v1, 1.0, v1, vcc
; GCN-IEEE-SAFE-NEXT: v_mul_f32_e32 v0, v0, v1
; GCN-IEEE-SAFE-NEXT: s_setpc_b64 s[30:31]
%sqrt = call contract float @llvm.sqrt.f32(float %val), !fpmath !1
%div = fdiv contract float 1.0, %sqrt, !fpmath !1
ret float %div
}
!0 = !{float 2.500000e+00}
!1 = !{float 1.000000e+00}
attributes #0 = { nounwind "denormal-fp-math-f32"="preserve-sign,preserve-sign" }
;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
; CI-DAZ-SAFE: {{.*}}
; CI-DAZ-UNSAFE: {{.*}}
; CI-IEEE-UNSAFE: {{.*}}
; SI-DAZ-SAFE: {{.*}}
; SI-DAZ-UNSAFE: {{.*}}
; SI-IEEE-UNSAFE: {{.*}}
Reference in New Issue View Git Blame Copy Permalink