[llvm] AMDGPU: Avoid using kernels in f16 copysign test (PR #142113)

Matt Arsenault via llvm-commits llvm-commits at lists.llvm.org
Fri May 30 02:34:04 PDT 2025


https://github.com/arsenm created https://github.com/llvm/llvm-project/pull/142113

Avoid the memory noise in tests that predate function support.

>From 8bfa0142cdc37fb55306c9802e5eaebc141e83db Mon Sep 17 00:00:00 2001
From: Matt Arsenault <Matthew.Arsenault at amd.com>
Date: Fri, 30 May 2025 09:14:55 +0200
Subject: [PATCH] AMDGPU: Avoid using kernels in f16 copysign test

Avoid the memory noise in tests that predate function support.
---
 llvm/test/CodeGen/AMDGPU/fcopysign.f16.ll | 2500 +++++++--------------
 1 file changed, 790 insertions(+), 1710 deletions(-)

diff --git a/llvm/test/CodeGen/AMDGPU/fcopysign.f16.ll b/llvm/test/CodeGen/AMDGPU/fcopysign.f16.ll
index 7ceeda810e5e6..d654537929255 100644
--- a/llvm/test/CodeGen/AMDGPU/fcopysign.f16.ll
+++ b/llvm/test/CodeGen/AMDGPU/fcopysign.f16.ll
@@ -13,684 +13,352 @@ declare <3 x half> @llvm.copysign.v3f16(<3 x half>, <3 x half>) #0
 declare <4 x half> @llvm.copysign.v4f16(<4 x half>, <4 x half>) #0
 declare i32 @llvm.amdgcn.workitem.id.x() #0
 
-define amdgpu_kernel void @s_copysign_f16(ptr addrspace(1) %arg_out, half %mag, half %sign) {
+define amdgpu_ps i16 @s_copysign_f16(half inreg %mag, half inreg %sign) {
 ; SI-LABEL: s_copysign_f16:
 ; SI:       ; %bb.0:
-; SI-NEXT:    s_load_dword s0, s[4:5], 0xb
-; SI-NEXT:    s_brev_b32 s2, -2
-; SI-NEXT:    s_mov_b32 s3, 0xf000
-; SI-NEXT:    s_waitcnt lgkmcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v0, s0
-; SI-NEXT:    s_lshr_b32 s0, s0, 16
-; SI-NEXT:    v_cvt_f32_f16_e32 v1, s0
-; SI-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x9
-; SI-NEXT:    v_bfi_b32 v0, s2, v0, v1
+; SI-NEXT:    v_cvt_f16_f32_e32 v0, s0
+; SI-NEXT:    s_brev_b32 s0, -2
+; SI-NEXT:    v_mov_b32_e32 v1, s1
+; SI-NEXT:    v_cvt_f32_f16_e32 v0, v0
+; SI-NEXT:    v_bfi_b32 v0, s0, v0, v1
 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
-; SI-NEXT:    s_mov_b32 s2, -1
-; SI-NEXT:    s_waitcnt lgkmcnt(0)
-; SI-NEXT:    buffer_store_short v0, off, s[0:3], 0
-; SI-NEXT:    s_endpgm
+; SI-NEXT:    v_readfirstlane_b32 s0, v0
+; SI-NEXT:    ; return to shader part epilog
 ;
 ; VI-LABEL: s_copysign_f16:
 ; VI:       ; %bb.0:
-; VI-NEXT:    s_load_dword s2, s[4:5], 0x2c
-; VI-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
-; VI-NEXT:    s_movk_i32 s3, 0x7fff
-; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    s_lshr_b32 s4, s2, 16
-; VI-NEXT:    v_mov_b32_e32 v0, s2
-; VI-NEXT:    v_mov_b32_e32 v1, s4
-; VI-NEXT:    v_bfi_b32 v2, s3, v0, v1
+; VI-NEXT:    s_movk_i32 s2, 0x7fff
 ; VI-NEXT:    v_mov_b32_e32 v0, s0
 ; VI-NEXT:    v_mov_b32_e32 v1, s1
-; VI-NEXT:    flat_store_short v[0:1], v2
-; VI-NEXT:    s_endpgm
+; VI-NEXT:    v_bfi_b32 v0, s2, v0, v1
+; VI-NEXT:    v_readfirstlane_b32 s0, v0
+; VI-NEXT:    ; return to shader part epilog
 ;
 ; GFX9-LABEL: s_copysign_f16:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_load_dword s2, s[4:5], 0x2c
-; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
-; GFX9-NEXT:    s_movk_i32 s3, 0x7fff
-; GFX9-NEXT:    v_mov_b32_e32 v0, 0
-; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    s_lshr_b32 s4, s2, 16
-; GFX9-NEXT:    v_mov_b32_e32 v1, s2
-; GFX9-NEXT:    v_mov_b32_e32 v2, s4
-; GFX9-NEXT:    v_bfi_b32 v1, s3, v1, v2
-; GFX9-NEXT:    global_store_short v0, v1, s[0:1]
-; GFX9-NEXT:    s_endpgm
+; GFX9-NEXT:    s_movk_i32 s2, 0x7fff
+; GFX9-NEXT:    v_mov_b32_e32 v0, s0
+; GFX9-NEXT:    v_mov_b32_e32 v1, s1
+; GFX9-NEXT:    v_bfi_b32 v0, s2, v0, v1
+; GFX9-NEXT:    v_readfirstlane_b32 s0, v0
+; GFX9-NEXT:    ; return to shader part epilog
 ;
 ; GFX11-LABEL: s_copysign_f16:
 ; GFX11:       ; %bb.0:
-; GFX11-NEXT:    s_clause 0x1
-; GFX11-NEXT:    s_load_b32 s2, s[4:5], 0x2c
-; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
-; GFX11-NEXT:    v_mov_b32_e32 v1, 0
-; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-NEXT:    s_lshr_b32 s3, s2, 16
-; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT:    v_mov_b32_e32 v0, s3
-; GFX11-NEXT:    v_bfi_b32 v0, 0x7fff, s2, v0
-; GFX11-NEXT:    global_store_b16 v1, v0, s[0:1]
-; GFX11-NEXT:    s_endpgm
+; GFX11-NEXT:    v_mov_b32_e32 v0, s1
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_bfi_b32 v0, 0x7fff, s0, v0
+; GFX11-NEXT:    v_readfirstlane_b32 s0, v0
+; GFX11-NEXT:    ; return to shader part epilog
   %out = call half @llvm.copysign.f16(half %mag, half %sign)
-  store half %out, ptr addrspace(1) %arg_out
-  ret void
+  %cast = bitcast half %out to i16
+  ret i16 %cast
 }
 
-define amdgpu_kernel void @s_test_copysign_f16_0(ptr addrspace(1) %out, half %mag) {
+define amdgpu_ps i16 @s_test_copysign_f16_0(half inreg %mag) {
 ; SI-LABEL: s_test_copysign_f16_0:
 ; SI:       ; %bb.0:
-; SI-NEXT:    s_load_dword s6, s[4:5], 0xb
-; SI-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x9
-; SI-NEXT:    s_mov_b32 s3, 0xf000
-; SI-NEXT:    s_mov_b32 s2, -1
-; SI-NEXT:    s_waitcnt lgkmcnt(0)
-; SI-NEXT:    s_and_b32 s4, s6, 0x7fff
-; SI-NEXT:    v_mov_b32_e32 v0, s4
-; SI-NEXT:    buffer_store_short v0, off, s[0:3], 0
-; SI-NEXT:    s_endpgm
+; SI-NEXT:    v_cvt_f16_f32_e32 v0, s0
+; SI-NEXT:    v_and_b32_e32 v0, 0x7fff, v0
+; SI-NEXT:    v_readfirstlane_b32 s0, v0
+; SI-NEXT:    ; return to shader part epilog
 ;
 ; VI-LABEL: s_test_copysign_f16_0:
 ; VI:       ; %bb.0:
-; VI-NEXT:    s_load_dword s2, s[4:5], 0x2c
-; VI-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
-; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    s_and_b32 s2, s2, 0x7fff
-; VI-NEXT:    v_mov_b32_e32 v0, s0
-; VI-NEXT:    v_mov_b32_e32 v1, s1
-; VI-NEXT:    v_mov_b32_e32 v2, s2
-; VI-NEXT:    flat_store_short v[0:1], v2
-; VI-NEXT:    s_endpgm
+; VI-NEXT:    s_and_b32 s0, s0, 0x7fff
+; VI-NEXT:    ; return to shader part epilog
 ;
 ; GFX9-LABEL: s_test_copysign_f16_0:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_load_dword s2, s[4:5], 0x2c
-; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
-; GFX9-NEXT:    v_mov_b32_e32 v0, 0
-; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    s_and_b32 s2, s2, 0x7fff
-; GFX9-NEXT:    v_mov_b32_e32 v1, s2
-; GFX9-NEXT:    global_store_short v0, v1, s[0:1]
-; GFX9-NEXT:    s_endpgm
+; GFX9-NEXT:    s_and_b32 s0, s0, 0x7fff
+; GFX9-NEXT:    ; return to shader part epilog
 ;
-; GFX11-TRUE16-LABEL: s_test_copysign_f16_0:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_clause 0x1
-; GFX11-TRUE16-NEXT:    s_load_b32 s2, s[4:5], 0x2c
-; GFX11-TRUE16-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v1, 0
-; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0x7fff
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, s2
-; GFX11-TRUE16-NEXT:    global_store_b16 v1, v0, s[0:1]
-; GFX11-TRUE16-NEXT:    s_endpgm
-;
-; GFX11-FAKE16-LABEL: s_test_copysign_f16_0:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_clause 0x1
-; GFX11-FAKE16-NEXT:    s_load_b32 s2, s[4:5], 0x2c
-; GFX11-FAKE16-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
-; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0x7fff
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
-; GFX11-FAKE16-NEXT:    global_store_b16 v0, v1, s[0:1]
-; GFX11-FAKE16-NEXT:    s_endpgm
+; GFX11-LABEL: s_test_copysign_f16_0:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_and_b32 s0, s0, 0x7fff
+; GFX11-NEXT:    ; return to shader part epilog
   %result = call half @llvm.copysign.f16(half %mag, half 0.0)
-  store half %result, ptr addrspace(1) %out, align 4
-  ret void
+  %cast = bitcast half %result to i16
+  ret i16 %cast
 }
 
-define amdgpu_kernel void @s_test_copysign_f16_1(ptr addrspace(1) %out, half %mag) {
+define amdgpu_ps i16 @s_test_copysign_f16_1(half inreg %mag) {
 ; SI-LABEL: s_test_copysign_f16_1:
 ; SI:       ; %bb.0:
-; SI-NEXT:    s_load_dword s6, s[4:5], 0xb
-; SI-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x9
-; SI-NEXT:    s_mov_b32 s3, 0xf000
-; SI-NEXT:    s_mov_b32 s2, -1
-; SI-NEXT:    s_waitcnt lgkmcnt(0)
-; SI-NEXT:    s_and_b32 s4, s6, 0x7fff
-; SI-NEXT:    v_mov_b32_e32 v0, s4
-; SI-NEXT:    buffer_store_short v0, off, s[0:3], 0
-; SI-NEXT:    s_endpgm
+; SI-NEXT:    v_cvt_f16_f32_e32 v0, s0
+; SI-NEXT:    v_and_b32_e32 v0, 0x7fff, v0
+; SI-NEXT:    v_readfirstlane_b32 s0, v0
+; SI-NEXT:    ; return to shader part epilog
 ;
 ; VI-LABEL: s_test_copysign_f16_1:
 ; VI:       ; %bb.0:
-; VI-NEXT:    s_load_dword s2, s[4:5], 0x2c
-; VI-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
-; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    s_and_b32 s2, s2, 0x7fff
-; VI-NEXT:    v_mov_b32_e32 v0, s0
-; VI-NEXT:    v_mov_b32_e32 v1, s1
-; VI-NEXT:    v_mov_b32_e32 v2, s2
-; VI-NEXT:    flat_store_short v[0:1], v2
-; VI-NEXT:    s_endpgm
+; VI-NEXT:    s_and_b32 s0, s0, 0x7fff
+; VI-NEXT:    ; return to shader part epilog
 ;
 ; GFX9-LABEL: s_test_copysign_f16_1:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_load_dword s2, s[4:5], 0x2c
-; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
-; GFX9-NEXT:    v_mov_b32_e32 v0, 0
-; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    s_and_b32 s2, s2, 0x7fff
-; GFX9-NEXT:    v_mov_b32_e32 v1, s2
-; GFX9-NEXT:    global_store_short v0, v1, s[0:1]
-; GFX9-NEXT:    s_endpgm
-;
-; GFX11-TRUE16-LABEL: s_test_copysign_f16_1:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_clause 0x1
-; GFX11-TRUE16-NEXT:    s_load_b32 s2, s[4:5], 0x2c
-; GFX11-TRUE16-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v1, 0
-; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0x7fff
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, s2
-; GFX11-TRUE16-NEXT:    global_store_b16 v1, v0, s[0:1]
-; GFX11-TRUE16-NEXT:    s_endpgm
+; GFX9-NEXT:    s_and_b32 s0, s0, 0x7fff
+; GFX9-NEXT:    ; return to shader part epilog
 ;
-; GFX11-FAKE16-LABEL: s_test_copysign_f16_1:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_clause 0x1
-; GFX11-FAKE16-NEXT:    s_load_b32 s2, s[4:5], 0x2c
-; GFX11-FAKE16-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
-; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0x7fff
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
-; GFX11-FAKE16-NEXT:    global_store_b16 v0, v1, s[0:1]
-; GFX11-FAKE16-NEXT:    s_endpgm
+; GFX11-LABEL: s_test_copysign_f16_1:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_and_b32 s0, s0, 0x7fff
+; GFX11-NEXT:    ; return to shader part epilog
   %result = call half @llvm.copysign.f16(half %mag, half 1.0)
-  store half %result, ptr addrspace(1) %out, align 4
-  ret void
+  %cast = bitcast half %result to i16
+  ret i16 %cast
 }
 
-define amdgpu_kernel void @s_test_copysign_f16_10.0(ptr addrspace(1) %out, half %mag) {
+define amdgpu_ps i16 @s_test_copysign_f16_10.0(half inreg %mag) {
 ; SI-LABEL: s_test_copysign_f16_10.0:
 ; SI:       ; %bb.0:
-; SI-NEXT:    s_load_dword s6, s[4:5], 0xb
-; SI-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x9
-; SI-NEXT:    s_mov_b32 s3, 0xf000
-; SI-NEXT:    s_mov_b32 s2, -1
-; SI-NEXT:    s_waitcnt lgkmcnt(0)
-; SI-NEXT:    s_and_b32 s4, s6, 0x7fff
-; SI-NEXT:    v_mov_b32_e32 v0, s4
-; SI-NEXT:    buffer_store_short v0, off, s[0:3], 0
-; SI-NEXT:    s_endpgm
+; SI-NEXT:    v_cvt_f16_f32_e32 v0, s0
+; SI-NEXT:    v_and_b32_e32 v0, 0x7fff, v0
+; SI-NEXT:    v_readfirstlane_b32 s0, v0
+; SI-NEXT:    ; return to shader part epilog
 ;
 ; VI-LABEL: s_test_copysign_f16_10.0:
 ; VI:       ; %bb.0:
-; VI-NEXT:    s_load_dword s2, s[4:5], 0x2c
-; VI-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
-; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    s_and_b32 s2, s2, 0x7fff
-; VI-NEXT:    v_mov_b32_e32 v0, s0
-; VI-NEXT:    v_mov_b32_e32 v1, s1
-; VI-NEXT:    v_mov_b32_e32 v2, s2
-; VI-NEXT:    flat_store_short v[0:1], v2
-; VI-NEXT:    s_endpgm
+; VI-NEXT:    s_and_b32 s0, s0, 0x7fff
+; VI-NEXT:    ; return to shader part epilog
 ;
 ; GFX9-LABEL: s_test_copysign_f16_10.0:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_load_dword s2, s[4:5], 0x2c
-; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
-; GFX9-NEXT:    v_mov_b32_e32 v0, 0
-; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    s_and_b32 s2, s2, 0x7fff
-; GFX9-NEXT:    v_mov_b32_e32 v1, s2
-; GFX9-NEXT:    global_store_short v0, v1, s[0:1]
-; GFX9-NEXT:    s_endpgm
+; GFX9-NEXT:    s_and_b32 s0, s0, 0x7fff
+; GFX9-NEXT:    ; return to shader part epilog
 ;
-; GFX11-TRUE16-LABEL: s_test_copysign_f16_10.0:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_clause 0x1
-; GFX11-TRUE16-NEXT:    s_load_b32 s2, s[4:5], 0x2c
-; GFX11-TRUE16-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v1, 0
-; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0x7fff
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, s2
-; GFX11-TRUE16-NEXT:    global_store_b16 v1, v0, s[0:1]
-; GFX11-TRUE16-NEXT:    s_endpgm
-;
-; GFX11-FAKE16-LABEL: s_test_copysign_f16_10.0:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_clause 0x1
-; GFX11-FAKE16-NEXT:    s_load_b32 s2, s[4:5], 0x2c
-; GFX11-FAKE16-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
-; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0x7fff
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
-; GFX11-FAKE16-NEXT:    global_store_b16 v0, v1, s[0:1]
-; GFX11-FAKE16-NEXT:    s_endpgm
+; GFX11-LABEL: s_test_copysign_f16_10.0:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_and_b32 s0, s0, 0x7fff
+; GFX11-NEXT:    ; return to shader part epilog
   %result = call half @llvm.copysign.f16(half %mag, half 10.0)
-  store half %result, ptr addrspace(1) %out, align 4
-  ret void
+  %cast = bitcast half %result to i16
+  ret i16 %cast
 }
 
-define amdgpu_kernel void @s_test_copysign_f16_neg1(ptr addrspace(1) %out, half %mag) {
+define amdgpu_ps i16 @s_test_copysign_f16_neg1(half inreg %mag) {
 ; SI-LABEL: s_test_copysign_f16_neg1:
 ; SI:       ; %bb.0:
-; SI-NEXT:    s_load_dword s6, s[4:5], 0xb
-; SI-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x9
-; SI-NEXT:    s_mov_b32 s3, 0xf000
-; SI-NEXT:    s_mov_b32 s2, -1
-; SI-NEXT:    s_waitcnt lgkmcnt(0)
-; SI-NEXT:    s_or_b32 s4, s6, 0x8000
-; SI-NEXT:    v_mov_b32_e32 v0, s4
-; SI-NEXT:    buffer_store_short v0, off, s[0:3], 0
-; SI-NEXT:    s_endpgm
+; SI-NEXT:    v_cvt_f16_f32_e32 v0, s0
+; SI-NEXT:    v_or_b32_e32 v0, 0x8000, v0
+; SI-NEXT:    v_readfirstlane_b32 s0, v0
+; SI-NEXT:    ; return to shader part epilog
 ;
 ; VI-LABEL: s_test_copysign_f16_neg1:
 ; VI:       ; %bb.0:
-; VI-NEXT:    s_load_dword s2, s[4:5], 0x2c
-; VI-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
-; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    s_bitset1_b32 s2, 15
-; VI-NEXT:    v_mov_b32_e32 v0, s0
-; VI-NEXT:    v_mov_b32_e32 v1, s1
-; VI-NEXT:    v_mov_b32_e32 v2, s2
-; VI-NEXT:    flat_store_short v[0:1], v2
-; VI-NEXT:    s_endpgm
+; VI-NEXT:    s_bitset1_b32 s0, 15
+; VI-NEXT:    ; return to shader part epilog
 ;
 ; GFX9-LABEL: s_test_copysign_f16_neg1:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_load_dword s2, s[4:5], 0x2c
-; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
-; GFX9-NEXT:    v_mov_b32_e32 v0, 0
-; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    s_bitset1_b32 s2, 15
-; GFX9-NEXT:    v_mov_b32_e32 v1, s2
-; GFX9-NEXT:    global_store_short v0, v1, s[0:1]
-; GFX9-NEXT:    s_endpgm
+; GFX9-NEXT:    s_bitset1_b32 s0, 15
+; GFX9-NEXT:    ; return to shader part epilog
 ;
-; GFX11-TRUE16-LABEL: s_test_copysign_f16_neg1:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_clause 0x1
-; GFX11-TRUE16-NEXT:    s_load_b32 s2, s[4:5], 0x2c
-; GFX11-TRUE16-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v1, 0
-; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s2, 15
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, s2
-; GFX11-TRUE16-NEXT:    global_store_b16 v1, v0, s[0:1]
-; GFX11-TRUE16-NEXT:    s_endpgm
-;
-; GFX11-FAKE16-LABEL: s_test_copysign_f16_neg1:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_clause 0x1
-; GFX11-FAKE16-NEXT:    s_load_b32 s2, s[4:5], 0x2c
-; GFX11-FAKE16-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
-; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s2, 15
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
-; GFX11-FAKE16-NEXT:    global_store_b16 v0, v1, s[0:1]
-; GFX11-FAKE16-NEXT:    s_endpgm
+; GFX11-LABEL: s_test_copysign_f16_neg1:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_bitset1_b32 s0, 15
+; GFX11-NEXT:    ; return to shader part epilog
   %result = call half @llvm.copysign.f16(half %mag, half -1.0)
-  store half %result, ptr addrspace(1) %out, align 4
-  ret void
+  %cast = bitcast half %result to i16
+  ret i16 %cast
 }
 
-define amdgpu_kernel void @s_test_copysign_f16_neg10(ptr addrspace(1) %out, half %mag) {
+define amdgpu_ps i16 @s_test_copysign_f16_neg10(half inreg %mag) {
 ; SI-LABEL: s_test_copysign_f16_neg10:
 ; SI:       ; %bb.0:
-; SI-NEXT:    s_load_dword s6, s[4:5], 0xb
-; SI-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x9
-; SI-NEXT:    s_mov_b32 s3, 0xf000
-; SI-NEXT:    s_mov_b32 s2, -1
-; SI-NEXT:    s_waitcnt lgkmcnt(0)
-; SI-NEXT:    s_or_b32 s4, s6, 0x8000
-; SI-NEXT:    v_mov_b32_e32 v0, s4
-; SI-NEXT:    buffer_store_short v0, off, s[0:3], 0
-; SI-NEXT:    s_endpgm
+; SI-NEXT:    v_cvt_f16_f32_e32 v0, s0
+; SI-NEXT:    v_or_b32_e32 v0, 0x8000, v0
+; SI-NEXT:    v_readfirstlane_b32 s0, v0
+; SI-NEXT:    ; return to shader part epilog
 ;
 ; VI-LABEL: s_test_copysign_f16_neg10:
 ; VI:       ; %bb.0:
-; VI-NEXT:    s_load_dword s2, s[4:5], 0x2c
-; VI-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
-; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    s_bitset1_b32 s2, 15
-; VI-NEXT:    v_mov_b32_e32 v0, s0
-; VI-NEXT:    v_mov_b32_e32 v1, s1
-; VI-NEXT:    v_mov_b32_e32 v2, s2
-; VI-NEXT:    flat_store_short v[0:1], v2
-; VI-NEXT:    s_endpgm
+; VI-NEXT:    s_bitset1_b32 s0, 15
+; VI-NEXT:    ; return to shader part epilog
 ;
 ; GFX9-LABEL: s_test_copysign_f16_neg10:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_load_dword s2, s[4:5], 0x2c
-; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
-; GFX9-NEXT:    v_mov_b32_e32 v0, 0
-; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    s_bitset1_b32 s2, 15
-; GFX9-NEXT:    v_mov_b32_e32 v1, s2
-; GFX9-NEXT:    global_store_short v0, v1, s[0:1]
-; GFX9-NEXT:    s_endpgm
-;
-; GFX11-TRUE16-LABEL: s_test_copysign_f16_neg10:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_clause 0x1
-; GFX11-TRUE16-NEXT:    s_load_b32 s2, s[4:5], 0x2c
-; GFX11-TRUE16-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v1, 0
-; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s2, 15
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, s2
-; GFX11-TRUE16-NEXT:    global_store_b16 v1, v0, s[0:1]
-; GFX11-TRUE16-NEXT:    s_endpgm
+; GFX9-NEXT:    s_bitset1_b32 s0, 15
+; GFX9-NEXT:    ; return to shader part epilog
 ;
-; GFX11-FAKE16-LABEL: s_test_copysign_f16_neg10:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_clause 0x1
-; GFX11-FAKE16-NEXT:    s_load_b32 s2, s[4:5], 0x2c
-; GFX11-FAKE16-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
-; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s2, 15
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
-; GFX11-FAKE16-NEXT:    global_store_b16 v0, v1, s[0:1]
-; GFX11-FAKE16-NEXT:    s_endpgm
+; GFX11-LABEL: s_test_copysign_f16_neg10:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_bitset1_b32 s0, 15
+; GFX11-NEXT:    ; return to shader part epilog
   %result = call half @llvm.copysign.f16(half %mag, half -10.0)
-  store half %result, ptr addrspace(1) %out, align 4
-  ret void
+  %cast = bitcast half %result to i16
+  ret i16 %cast
 }
 
-define amdgpu_kernel void @s_test_copysign_f16_0_mag(ptr addrspace(1) %out, half %sign) {
+define amdgpu_ps i16 @s_test_copysign_f16_0_mag(half inreg %sign) {
 ; SI-LABEL: s_test_copysign_f16_0_mag:
 ; SI:       ; %bb.0:
-; SI-NEXT:    s_load_dword s0, s[4:5], 0xb
-; SI-NEXT:    s_brev_b32 s2, -2
-; SI-NEXT:    s_mov_b32 s3, 0xf000
-; SI-NEXT:    s_waitcnt lgkmcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v0, s0
-; SI-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x9
-; SI-NEXT:    v_bfi_b32 v0, s2, 0, v0
-; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
-; SI-NEXT:    s_mov_b32 s2, -1
-; SI-NEXT:    s_waitcnt lgkmcnt(0)
-; SI-NEXT:    buffer_store_short v0, off, s[0:3], 0
-; SI-NEXT:    s_endpgm
+; SI-NEXT:    v_cvt_f16_f32_e32 v0, s0
+; SI-NEXT:    v_and_b32_e32 v0, 0x8000, v0
+; SI-NEXT:    v_readfirstlane_b32 s0, v0
+; SI-NEXT:    ; return to shader part epilog
 ;
 ; VI-LABEL: s_test_copysign_f16_0_mag:
 ; VI:       ; %bb.0:
-; VI-NEXT:    s_load_dword s2, s[4:5], 0x2c
-; VI-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
-; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    s_and_b32 s2, s2, 0x8000
-; VI-NEXT:    v_mov_b32_e32 v0, s0
-; VI-NEXT:    v_mov_b32_e32 v1, s1
-; VI-NEXT:    v_mov_b32_e32 v2, s2
-; VI-NEXT:    flat_store_short v[0:1], v2
-; VI-NEXT:    s_endpgm
+; VI-NEXT:    s_and_b32 s0, s0, 0x8000
+; VI-NEXT:    ; return to shader part epilog
 ;
 ; GFX9-LABEL: s_test_copysign_f16_0_mag:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_load_dword s2, s[4:5], 0x2c
-; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
-; GFX9-NEXT:    v_mov_b32_e32 v0, 0
-; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    s_and_b32 s2, s2, 0x8000
-; GFX9-NEXT:    v_mov_b32_e32 v1, s2
-; GFX9-NEXT:    global_store_short v0, v1, s[0:1]
-; GFX9-NEXT:    s_endpgm
+; GFX9-NEXT:    s_and_b32 s0, s0, 0x8000
+; GFX9-NEXT:    ; return to shader part epilog
 ;
 ; GFX11-LABEL: s_test_copysign_f16_0_mag:
 ; GFX11:       ; %bb.0:
-; GFX11-NEXT:    s_clause 0x1
-; GFX11-NEXT:    s_load_b32 s2, s[4:5], 0x2c
-; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
-; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-NEXT:    s_and_b32 s2, s2, 0x8000
-; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
-; GFX11-NEXT:    global_store_b16 v0, v1, s[0:1]
-; GFX11-NEXT:    s_endpgm
+; GFX11-NEXT:    s_and_b32 s0, s0, 0x8000
+; GFX11-NEXT:    ; return to shader part epilog
   %result = call half @llvm.copysign.f16(half 0.0, half %sign)
-  store half %result, ptr addrspace(1) %out, align 4
-  ret void
+  %cast = bitcast half %result to i16
+  ret i16 %cast
 }
 
 
-define amdgpu_kernel void @s_test_copysign_f16_1_mag(ptr addrspace(1) %out, half %sign) {
+define amdgpu_ps i16 @s_test_copysign_f16_1_mag(half inreg %sign) {
 ; SI-LABEL: s_test_copysign_f16_1_mag:
 ; SI:       ; %bb.0:
-; SI-NEXT:    s_load_dword s0, s[4:5], 0xb
-; SI-NEXT:    s_brev_b32 s2, -2
-; SI-NEXT:    s_mov_b32 s3, 0xf000
-; SI-NEXT:    s_waitcnt lgkmcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v0, s0
-; SI-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x9
-; SI-NEXT:    v_bfi_b32 v0, s2, 1.0, v0
-; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
-; SI-NEXT:    s_mov_b32 s2, -1
-; SI-NEXT:    s_waitcnt lgkmcnt(0)
-; SI-NEXT:    buffer_store_short v0, off, s[0:3], 0
-; SI-NEXT:    s_endpgm
+; SI-NEXT:    v_cvt_f16_f32_e32 v0, s0
+; SI-NEXT:    v_and_b32_e32 v0, 0x8000, v0
+; SI-NEXT:    v_or_b32_e32 v0, 0x3c00, v0
+; SI-NEXT:    v_readfirstlane_b32 s0, v0
+; SI-NEXT:    ; return to shader part epilog
 ;
 ; VI-LABEL: s_test_copysign_f16_1_mag:
 ; VI:       ; %bb.0:
-; VI-NEXT:    s_load_dword s2, s[4:5], 0x2c
-; VI-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
-; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    s_and_b32 s2, s2, 0x8000
-; VI-NEXT:    s_or_b32 s2, s2, 0x3c00
-; VI-NEXT:    v_mov_b32_e32 v0, s0
-; VI-NEXT:    v_mov_b32_e32 v1, s1
-; VI-NEXT:    v_mov_b32_e32 v2, s2
-; VI-NEXT:    flat_store_short v[0:1], v2
-; VI-NEXT:    s_endpgm
+; VI-NEXT:    s_and_b32 s0, s0, 0x8000
+; VI-NEXT:    s_or_b32 s0, s0, 0x3c00
+; VI-NEXT:    ; return to shader part epilog
 ;
 ; GFX9-LABEL: s_test_copysign_f16_1_mag:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_load_dword s2, s[4:5], 0x2c
-; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
-; GFX9-NEXT:    v_mov_b32_e32 v0, 0
-; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    s_and_b32 s2, s2, 0x8000
-; GFX9-NEXT:    s_or_b32 s2, s2, 0x3c00
-; GFX9-NEXT:    v_mov_b32_e32 v1, s2
-; GFX9-NEXT:    global_store_short v0, v1, s[0:1]
-; GFX9-NEXT:    s_endpgm
+; GFX9-NEXT:    s_and_b32 s0, s0, 0x8000
+; GFX9-NEXT:    s_or_b32 s0, s0, 0x3c00
+; GFX9-NEXT:    ; return to shader part epilog
 ;
 ; GFX11-LABEL: s_test_copysign_f16_1_mag:
 ; GFX11:       ; %bb.0:
-; GFX11-NEXT:    s_clause 0x1
-; GFX11-NEXT:    s_load_b32 s2, s[4:5], 0x2c
-; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
-; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-NEXT:    s_and_b32 s2, s2, 0x8000
-; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
-; GFX11-NEXT:    s_or_b32 s2, s2, 0x3c00
-; GFX11-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
-; GFX11-NEXT:    global_store_b16 v0, v1, s[0:1]
-; GFX11-NEXT:    s_endpgm
+; GFX11-NEXT:    s_and_b32 s0, s0, 0x8000
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_or_b32 s0, s0, 0x3c00
+; GFX11-NEXT:    ; return to shader part epilog
   %result = call half @llvm.copysign.f16(half 1.0, half %sign)
-  store half %result, ptr addrspace(1) %out, align 4
-  ret void
+  %cast = bitcast half %result to i16
+  ret i16 %cast
 }
 
-define amdgpu_kernel void @s_test_copysign_f16_10_mag(ptr addrspace(1) %out, half %sign) {
+define i16 @s_test_copysign_f16_10_mag(half inreg %sign) {
 ; SI-LABEL: s_test_copysign_f16_10_mag:
 ; SI:       ; %bb.0:
-; SI-NEXT:    s_load_dword s0, s[4:5], 0xb
-; SI-NEXT:    s_brev_b32 s2, -2
-; SI-NEXT:    v_mov_b32_e32 v1, 0x41200000
-; SI-NEXT:    s_mov_b32 s3, 0xf000
-; SI-NEXT:    s_waitcnt lgkmcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v0, s0
-; SI-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x9
-; SI-NEXT:    v_bfi_b32 v0, s2, v1, v0
-; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
-; SI-NEXT:    s_mov_b32 s2, -1
-; SI-NEXT:    s_waitcnt lgkmcnt(0)
-; SI-NEXT:    buffer_store_short v0, off, s[0:3], 0
-; SI-NEXT:    s_endpgm
+; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; SI-NEXT:    v_cvt_f16_f32_e32 v0, s16
+; SI-NEXT:    v_and_b32_e32 v0, 0x8000, v0
+; SI-NEXT:    v_or_b32_e32 v0, 0x4900, v0
+; SI-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; VI-LABEL: s_test_copysign_f16_10_mag:
 ; VI:       ; %bb.0:
-; VI-NEXT:    s_load_dword s2, s[4:5], 0x2c
-; VI-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
-; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    s_and_b32 s2, s2, 0x8000
-; VI-NEXT:    s_or_b32 s2, s2, 0x4900
-; VI-NEXT:    v_mov_b32_e32 v0, s0
-; VI-NEXT:    v_mov_b32_e32 v1, s1
-; VI-NEXT:    v_mov_b32_e32 v2, s2
-; VI-NEXT:    flat_store_short v[0:1], v2
-; VI-NEXT:    s_endpgm
+; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_and_b32 s4, s16, 0x8000
+; VI-NEXT:    s_or_b32 s4, s4, 0x4900
+; VI-NEXT:    v_mov_b32_e32 v0, s4
+; VI-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX9-LABEL: s_test_copysign_f16_10_mag:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_load_dword s2, s[4:5], 0x2c
-; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
-; GFX9-NEXT:    v_mov_b32_e32 v0, 0
-; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    s_and_b32 s2, s2, 0x8000
-; GFX9-NEXT:    s_or_b32 s2, s2, 0x4900
-; GFX9-NEXT:    v_mov_b32_e32 v1, s2
-; GFX9-NEXT:    global_store_short v0, v1, s[0:1]
-; GFX9-NEXT:    s_endpgm
+; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_and_b32 s4, s16, 0x8000
+; GFX9-NEXT:    s_or_b32 s4, s4, 0x4900
+; GFX9-NEXT:    v_mov_b32_e32 v0, s4
+; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX11-LABEL: s_test_copysign_f16_10_mag:
-; GFX11:       ; %bb.0:
-; GFX11-NEXT:    s_clause 0x1
-; GFX11-NEXT:    s_load_b32 s2, s[4:5], 0x2c
-; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
-; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-NEXT:    s_and_b32 s2, s2, 0x8000
-; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
-; GFX11-NEXT:    s_or_b32 s2, s2, 0x4900
-; GFX11-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
-; GFX11-NEXT:    global_store_b16 v0, v1, s[0:1]
-; GFX11-NEXT:    s_endpgm
+; GFX11-TRUE16-LABEL: s_test_copysign_f16_10_mag:
+; GFX11-TRUE16:       ; %bb.0:
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0x8000
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, 0x4900
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, s0
+; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX11-FAKE16-LABEL: s_test_copysign_f16_10_mag:
+; GFX11-FAKE16:       ; %bb.0:
+; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0x8000
+; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, 0x4900
+; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
   %result = call half @llvm.copysign.f16(half 10.0, half %sign)
-  store half %result, ptr addrspace(1) %out, align 4
-  ret void
+  %cast = bitcast half %result to i16
+  ret i16 %cast
 }
 
-define amdgpu_kernel void @s_test_copysign_f16_neg1_mag(ptr addrspace(1) %out, half %sign) {
+define amdgpu_ps i16 @s_test_copysign_f16_neg1_mag(half inreg %sign) {
 ; SI-LABEL: s_test_copysign_f16_neg1_mag:
 ; SI:       ; %bb.0:
-; SI-NEXT:    s_load_dword s0, s[4:5], 0xb
-; SI-NEXT:    s_brev_b32 s2, -2
-; SI-NEXT:    s_mov_b32 s3, 0xf000
-; SI-NEXT:    s_waitcnt lgkmcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v0, s0
-; SI-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x9
-; SI-NEXT:    v_bfi_b32 v0, s2, -1.0, v0
-; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
-; SI-NEXT:    s_mov_b32 s2, -1
-; SI-NEXT:    s_waitcnt lgkmcnt(0)
-; SI-NEXT:    buffer_store_short v0, off, s[0:3], 0
-; SI-NEXT:    s_endpgm
+; SI-NEXT:    v_cvt_f16_f32_e32 v0, s0
+; SI-NEXT:    v_and_b32_e32 v0, 0x8000, v0
+; SI-NEXT:    v_or_b32_e32 v0, 0x3c00, v0
+; SI-NEXT:    v_readfirstlane_b32 s0, v0
+; SI-NEXT:    ; return to shader part epilog
 ;
 ; VI-LABEL: s_test_copysign_f16_neg1_mag:
 ; VI:       ; %bb.0:
-; VI-NEXT:    s_load_dword s2, s[4:5], 0x2c
-; VI-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
-; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    s_and_b32 s2, s2, 0x8000
-; VI-NEXT:    s_or_b32 s2, s2, 0x3c00
-; VI-NEXT:    v_mov_b32_e32 v0, s0
-; VI-NEXT:    v_mov_b32_e32 v1, s1
-; VI-NEXT:    v_mov_b32_e32 v2, s2
-; VI-NEXT:    flat_store_short v[0:1], v2
-; VI-NEXT:    s_endpgm
+; VI-NEXT:    s_and_b32 s0, s0, 0x8000
+; VI-NEXT:    s_or_b32 s0, s0, 0x3c00
+; VI-NEXT:    ; return to shader part epilog
 ;
 ; GFX9-LABEL: s_test_copysign_f16_neg1_mag:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_load_dword s2, s[4:5], 0x2c
-; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
-; GFX9-NEXT:    v_mov_b32_e32 v0, 0
-; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    s_and_b32 s2, s2, 0x8000
-; GFX9-NEXT:    s_or_b32 s2, s2, 0x3c00
-; GFX9-NEXT:    v_mov_b32_e32 v1, s2
-; GFX9-NEXT:    global_store_short v0, v1, s[0:1]
-; GFX9-NEXT:    s_endpgm
+; GFX9-NEXT:    s_and_b32 s0, s0, 0x8000
+; GFX9-NEXT:    s_or_b32 s0, s0, 0x3c00
+; GFX9-NEXT:    ; return to shader part epilog
 ;
 ; GFX11-LABEL: s_test_copysign_f16_neg1_mag:
 ; GFX11:       ; %bb.0:
-; GFX11-NEXT:    s_clause 0x1
-; GFX11-NEXT:    s_load_b32 s2, s[4:5], 0x2c
-; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
-; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-NEXT:    s_and_b32 s2, s2, 0x8000
-; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
-; GFX11-NEXT:    s_or_b32 s2, s2, 0x3c00
-; GFX11-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
-; GFX11-NEXT:    global_store_b16 v0, v1, s[0:1]
-; GFX11-NEXT:    s_endpgm
+; GFX11-NEXT:    s_and_b32 s0, s0, 0x8000
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_or_b32 s0, s0, 0x3c00
+; GFX11-NEXT:    ; return to shader part epilog
   %result = call half @llvm.copysign.f16(half -1.0, half %sign)
-  store half %result, ptr addrspace(1) %out, align 4
-  ret void
+  %cast = bitcast half %result to i16
+  ret i16 %cast
 }
 
-define amdgpu_kernel void @s_test_copysign_f16_neg10_mag(ptr addrspace(1) %out, half %sign) {
+define amdgpu_ps i16 @s_test_copysign_f16_neg10_mag(half inreg %sign) {
 ; SI-LABEL: s_test_copysign_f16_neg10_mag:
 ; SI:       ; %bb.0:
-; SI-NEXT:    s_load_dword s0, s[4:5], 0xb
-; SI-NEXT:    s_brev_b32 s2, -2
-; SI-NEXT:    v_mov_b32_e32 v1, 0xc1200000
-; SI-NEXT:    s_mov_b32 s3, 0xf000
-; SI-NEXT:    s_waitcnt lgkmcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v0, s0
-; SI-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x9
-; SI-NEXT:    v_bfi_b32 v0, s2, v1, v0
-; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
-; SI-NEXT:    s_mov_b32 s2, -1
-; SI-NEXT:    s_waitcnt lgkmcnt(0)
-; SI-NEXT:    buffer_store_short v0, off, s[0:3], 0
-; SI-NEXT:    s_endpgm
+; SI-NEXT:    v_cvt_f16_f32_e32 v0, s0
+; SI-NEXT:    v_and_b32_e32 v0, 0x8000, v0
+; SI-NEXT:    v_or_b32_e32 v0, 0x4900, v0
+; SI-NEXT:    v_readfirstlane_b32 s0, v0
+; SI-NEXT:    ; return to shader part epilog
 ;
 ; VI-LABEL: s_test_copysign_f16_neg10_mag:
 ; VI:       ; %bb.0:
-; VI-NEXT:    s_load_dword s2, s[4:5], 0x2c
-; VI-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
-; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    s_and_b32 s2, s2, 0x8000
-; VI-NEXT:    s_or_b32 s2, s2, 0x4900
-; VI-NEXT:    v_mov_b32_e32 v0, s0
-; VI-NEXT:    v_mov_b32_e32 v1, s1
-; VI-NEXT:    v_mov_b32_e32 v2, s2
-; VI-NEXT:    flat_store_short v[0:1], v2
-; VI-NEXT:    s_endpgm
+; VI-NEXT:    s_and_b32 s0, s0, 0x8000
+; VI-NEXT:    s_or_b32 s0, s0, 0x4900
+; VI-NEXT:    ; return to shader part epilog
 ;
 ; GFX9-LABEL: s_test_copysign_f16_neg10_mag:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_load_dword s2, s[4:5], 0x2c
-; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
-; GFX9-NEXT:    v_mov_b32_e32 v0, 0
-; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    s_and_b32 s2, s2, 0x8000
-; GFX9-NEXT:    s_or_b32 s2, s2, 0x4900
-; GFX9-NEXT:    v_mov_b32_e32 v1, s2
-; GFX9-NEXT:    global_store_short v0, v1, s[0:1]
-; GFX9-NEXT:    s_endpgm
+; GFX9-NEXT:    s_and_b32 s0, s0, 0x8000
+; GFX9-NEXT:    s_or_b32 s0, s0, 0x4900
+; GFX9-NEXT:    ; return to shader part epilog
 ;
 ; GFX11-LABEL: s_test_copysign_f16_neg10_mag:
 ; GFX11:       ; %bb.0:
-; GFX11-NEXT:    s_clause 0x1
-; GFX11-NEXT:    s_load_b32 s2, s[4:5], 0x2c
-; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
-; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-NEXT:    s_and_b32 s2, s2, 0x8000
-; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
-; GFX11-NEXT:    s_or_b32 s2, s2, 0x4900
-; GFX11-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
-; GFX11-NEXT:    global_store_b16 v0, v1, s[0:1]
-; GFX11-NEXT:    s_endpgm
+; GFX11-NEXT:    s_and_b32 s0, s0, 0x8000
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_or_b32 s0, s0, 0x4900
+; GFX11-NEXT:    ; return to shader part epilog
   %result = call half @llvm.copysign.f16(half -10.0, half %sign)
-  store half %result, ptr addrspace(1) %out, align 4
-  ret void
+  %cast = bitcast half %result to i16
+  ret i16 %cast
 }
 
 define half @v_copysign_f16(half %mag, half %sign) {
@@ -901,1426 +569,838 @@ define half @v_test_copysign_f16_neg10(half %mag) {
   ret half %result
 }
 
-define amdgpu_kernel void @v_copysign_out_f32_mag_f16_sign_f32(ptr addrspace(1) %arg_out, ptr addrspace(1) %arg_mag, ptr addrspace(1) %arg_sign) {
+define float @v_copysign_out_f32_mag_f16_sign_f32(half %mag, float %sign) {
 ; SI-LABEL: v_copysign_out_f32_mag_f16_sign_f32:
 ; SI:       ; %bb.0:
-; SI-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
-; SI-NEXT:    s_load_dwordx2 s[4:5], s[4:5], 0xd
-; SI-NEXT:    s_mov_b32 s11, 0xf000
-; SI-NEXT:    s_mov_b32 s14, 0
-; SI-NEXT:    s_mov_b32 s15, s11
-; SI-NEXT:    s_waitcnt lgkmcnt(0)
-; SI-NEXT:    s_mov_b64 s[12:13], s[2:3]
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 1, v0
-; SI-NEXT:    v_mov_b32_e32 v2, 0
-; SI-NEXT:    buffer_load_ushort v3, v[1:2], s[12:15], 0 addr64
-; SI-NEXT:    s_mov_b64 s[6:7], s[14:15]
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 2, v0
-; SI-NEXT:    buffer_load_dword v0, v[1:2], s[4:7], 0 addr64
-; SI-NEXT:    s_mov_b32 s8, s0
-; SI-NEXT:    s_brev_b32 s0, -2
-; SI-NEXT:    s_mov_b32 s10, -1
-; SI-NEXT:    s_mov_b32 s9, s1
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_cvt_f32_f16_e32 v1, v3
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_bfi_b32 v0, s0, v1, v0
-; SI-NEXT:    buffer_store_dword v0, off, s[8:11], 0
-; SI-NEXT:    s_endpgm
+; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; SI-NEXT:    s_brev_b32 s4, -2
+; SI-NEXT:    v_bfi_b32 v0, s4, v0, v1
+; SI-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; VI-LABEL: v_copysign_out_f32_mag_f16_sign_f32:
 ; VI:       ; %bb.0:
-; VI-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
-; VI-NEXT:    s_load_dwordx2 s[4:5], s[4:5], 0x34
-; VI-NEXT:    v_lshlrev_b32_e32 v1, 1, v0
-; VI-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
-; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    v_mov_b32_e32 v2, s3
-; VI-NEXT:    v_add_u32_e32 v1, vcc, s2, v1
-; VI-NEXT:    v_addc_u32_e32 v2, vcc, 0, v2, vcc
-; VI-NEXT:    flat_load_ushort v2, v[1:2]
-; VI-NEXT:    v_mov_b32_e32 v1, s5
-; VI-NEXT:    v_add_u32_e32 v0, vcc, s4, v0
-; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; VI-NEXT:    flat_load_dword v3, v[0:1]
-; VI-NEXT:    v_mov_b32_e32 v0, s0
-; VI-NEXT:    s_brev_b32 s0, -2
-; VI-NEXT:    v_mov_b32_e32 v1, s1
-; VI-NEXT:    s_waitcnt vmcnt(1)
-; VI-NEXT:    v_cvt_f32_f16_e32 v2, v2
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_bfi_b32 v2, s0, v2, v3
-; VI-NEXT:    flat_store_dword v[0:1], v2
-; VI-NEXT:    s_endpgm
+; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; VI-NEXT:    v_cvt_f32_f16_e32 v0, v0
+; VI-NEXT:    s_brev_b32 s4, -2
+; VI-NEXT:    v_bfi_b32 v0, s4, v0, v1
+; VI-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX9-LABEL: v_copysign_out_f32_mag_f16_sign_f32:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
-; GFX9-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x34
-; GFX9-NEXT:    v_lshlrev_b32_e32 v1, 1, v0
-; GFX9-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
-; GFX9-NEXT:    v_mov_b32_e32 v2, 0
-; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    global_load_ushort v1, v1, s[2:3]
-; GFX9-NEXT:    s_brev_b32 s2, -2
-; GFX9-NEXT:    global_load_dword v0, v0, s[6:7]
-; GFX9-NEXT:    s_waitcnt vmcnt(1)
-; GFX9-NEXT:    v_cvt_f32_f16_e32 v1, v1
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_bfi_b32 v0, s2, v1, v0
-; GFX9-NEXT:    global_store_dword v2, v0, s[0:1]
-; GFX9-NEXT:    s_endpgm
+; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    v_cvt_f32_f16_e32 v0, v0
+; GFX9-NEXT:    s_brev_b32 s4, -2
+; GFX9-NEXT:    v_bfi_b32 v0, s4, v0, v1
+; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX11-TRUE16-LABEL: v_copysign_out_f32_mag_f16_sign_f32:
 ; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_clause 0x1
-; GFX11-TRUE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
-; GFX11-TRUE16-NEXT:    s_load_b64 s[4:5], s[4:5], 0x34
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, 0 :: v_dual_and_b32 v1, 0x3ff, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 1, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 2, v1
-; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    global_load_d16_b16 v0, v0, s[2:3]
-; GFX11-TRUE16-NEXT:    global_load_b32 v1, v1, s[4:5]
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(1)
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX11-TRUE16-NEXT:    v_cvt_f32_f16_e32 v0, v0.l
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_bfi_b32 v0, 0x7fffffff, v0, v1
-; GFX11-TRUE16-NEXT:    global_store_b32 v2, v0, s[0:1]
-; GFX11-TRUE16-NEXT:    s_endpgm
+; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX11-FAKE16-LABEL: v_copysign_out_f32_mag_f16_sign_f32:
 ; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_clause 0x1
-; GFX11-FAKE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
-; GFX11-FAKE16-NEXT:    s_load_b64 s[4:5], s[4:5], 0x34
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0x3ff, v0
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, 0 :: v_dual_lshlrev_b32 v1, 1, v0
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
-; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    global_load_u16 v1, v1, s[2:3]
-; GFX11-FAKE16-NEXT:    global_load_b32 v0, v0, s[4:5]
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(1)
-; GFX11-FAKE16-NEXT:    v_cvt_f32_f16_e32 v1, v1
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-FAKE16-NEXT:    v_cvt_f32_f16_e32 v0, v0
 ; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_bfi_b32 v0, 0x7fffffff, v1, v0
-; GFX11-FAKE16-NEXT:    global_store_b32 v2, v0, s[0:1]
-; GFX11-FAKE16-NEXT:    s_endpgm
-  %tid = call i32 @llvm.amdgcn.workitem.id.x()
-  %arg_mag_gep = getelementptr half, ptr addrspace(1) %arg_mag, i32 %tid
-  %mag = load half, ptr addrspace(1) %arg_mag_gep
+; GFX11-FAKE16-NEXT:    v_bfi_b32 v0, 0x7fffffff, v0, v1
+; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
   %mag.ext = fpext half %mag to float
-  %arg_sign_gep = getelementptr float, ptr addrspace(1) %arg_sign, i32 %tid
-  %sign = load float, ptr addrspace(1) %arg_sign_gep
   %out = call float @llvm.copysign.f32(float %mag.ext, float %sign)
-  store float %out, ptr addrspace(1) %arg_out
-  ret void
+  ret float %out
 }
 
-define amdgpu_kernel void @v_copysign_out_f64_mag_f16_sign_f64(ptr addrspace(1) %arg_out, ptr addrspace(1) %arg_mag, ptr addrspace(1) %arg_sign) {
+define double @v_copysign_out_f64_mag_f16_sign_f64(half %mag, double %sign) {
 ; SI-LABEL: v_copysign_out_f64_mag_f16_sign_f64:
 ; SI:       ; %bb.0:
-; SI-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
-; SI-NEXT:    s_load_dwordx2 s[4:5], s[4:5], 0xd
-; SI-NEXT:    s_mov_b32 s11, 0xf000
-; SI-NEXT:    s_mov_b32 s14, 0
-; SI-NEXT:    s_mov_b32 s15, s11
-; SI-NEXT:    s_waitcnt lgkmcnt(0)
-; SI-NEXT:    s_mov_b64 s[12:13], s[2:3]
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 1, v0
-; SI-NEXT:    v_mov_b32_e32 v2, 0
-; SI-NEXT:    buffer_load_ushort v3, v[1:2], s[12:15], 0 addr64
-; SI-NEXT:    s_mov_b64 s[6:7], s[14:15]
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 3, v0
-; SI-NEXT:    buffer_load_dwordx2 v[0:1], v[1:2], s[4:7], 0 addr64
-; SI-NEXT:    s_mov_b32 s8, s0
-; SI-NEXT:    s_brev_b32 s0, -2
-; SI-NEXT:    s_mov_b32 s10, -1
-; SI-NEXT:    s_mov_b32 s9, s1
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v0, v3
-; SI-NEXT:    v_cvt_f64_f32_e32 v[2:3], v0
-; SI-NEXT:    v_bfi_b32 v3, s0, v3, v1
-; SI-NEXT:    buffer_store_dwordx2 v[2:3], off, s[8:11], 0
-; SI-NEXT:    s_endpgm
+; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; SI-NEXT:    v_cvt_f64_f32_e32 v[0:1], v0
+; SI-NEXT:    s_brev_b32 s4, -2
+; SI-NEXT:    v_bfi_b32 v1, s4, v1, v2
+; SI-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; VI-LABEL: v_copysign_out_f64_mag_f16_sign_f64:
 ; VI:       ; %bb.0:
-; VI-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
-; VI-NEXT:    s_load_dwordx2 s[4:5], s[4:5], 0x34
-; VI-NEXT:    v_lshlrev_b32_e32 v1, 1, v0
-; VI-NEXT:    v_lshlrev_b32_e32 v0, 3, v0
-; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    v_mov_b32_e32 v2, s3
-; VI-NEXT:    v_add_u32_e32 v1, vcc, s2, v1
-; VI-NEXT:    v_addc_u32_e32 v2, vcc, 0, v2, vcc
-; VI-NEXT:    flat_load_ushort v2, v[1:2]
-; VI-NEXT:    v_mov_b32_e32 v1, s5
-; VI-NEXT:    v_add_u32_e32 v0, vcc, s4, v0
-; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; VI-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
-; VI-NEXT:    v_mov_b32_e32 v4, s0
-; VI-NEXT:    s_brev_b32 s0, -2
-; VI-NEXT:    v_mov_b32_e32 v5, s1
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_cvt_f32_f16_e32 v0, v2
-; VI-NEXT:    v_cvt_f64_f32_e32 v[2:3], v0
-; VI-NEXT:    v_bfi_b32 v3, s0, v3, v1
-; VI-NEXT:    flat_store_dwordx2 v[4:5], v[2:3]
-; VI-NEXT:    s_endpgm
+; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; VI-NEXT:    v_cvt_f32_f16_e32 v0, v0
+; VI-NEXT:    s_brev_b32 s4, -2
+; VI-NEXT:    v_cvt_f64_f32_e32 v[0:1], v0
+; VI-NEXT:    v_bfi_b32 v1, s4, v1, v2
+; VI-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX9-LABEL: v_copysign_out_f64_mag_f16_sign_f64:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
-; GFX9-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x34
-; GFX9-NEXT:    v_lshlrev_b32_e32 v1, 1, v0
-; GFX9-NEXT:    v_lshlrev_b32_e32 v0, 3, v0
-; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    global_load_ushort v2, v1, s[2:3]
-; GFX9-NEXT:    s_nop 0
-; GFX9-NEXT:    global_load_dwordx2 v[0:1], v0, s[6:7]
-; GFX9-NEXT:    s_brev_b32 s2, -2
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_cvt_f32_f16_e32 v0, v2
-; GFX9-NEXT:    v_cvt_f64_f32_e32 v[2:3], v0
-; GFX9-NEXT:    v_mov_b32_e32 v0, 0
-; GFX9-NEXT:    v_bfi_b32 v3, s2, v3, v1
-; GFX9-NEXT:    global_store_dwordx2 v0, v[2:3], s[0:1]
-; GFX9-NEXT:    s_endpgm
+; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    v_cvt_f32_f16_e32 v0, v0
+; GFX9-NEXT:    s_brev_b32 s4, -2
+; GFX9-NEXT:    v_cvt_f64_f32_e32 v[0:1], v0
+; GFX9-NEXT:    v_bfi_b32 v1, s4, v1, v2
+; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX11-TRUE16-LABEL: v_copysign_out_f64_mag_f16_sign_f64:
 ; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0x3ff, v0
-; GFX11-TRUE16-NEXT:    s_load_b64 s[4:5], s[4:5], 0x34
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v3, 0 :: v_dual_lshlrev_b32 v0, 1, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 3, v1
-; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    global_load_d16_b16 v0, v0, s[2:3]
-; GFX11-TRUE16-NEXT:    global_load_b64 v[1:2], v1, s[4:5]
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(1)
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX11-TRUE16-NEXT:    v_cvt_f32_f16_e32 v0, v0.l
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_cvt_f64_f32_e32 v[0:1], v0
 ; GFX11-TRUE16-NEXT:    v_bfi_b32 v1, 0x7fffffff, v1, v2
-; GFX11-TRUE16-NEXT:    global_store_b64 v3, v[0:1], s[0:1]
-; GFX11-TRUE16-NEXT:    s_endpgm
+; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX11-FAKE16-LABEL: v_copysign_out_f64_mag_f16_sign_f64:
 ; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0x3ff, v0
-; GFX11-FAKE16-NEXT:    s_load_b64 s[4:5], s[4:5], 0x34
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v1, 1, v0
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v0, 3, v0
-; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    global_load_u16 v2, v1, s[2:3]
-; GFX11-FAKE16-NEXT:    global_load_b64 v[0:1], v0, s[4:5]
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cvt_f32_f16_e32 v0, v2
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-FAKE16-NEXT:    v_cvt_f64_f32_e32 v[2:3], v0
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, 0
-; GFX11-FAKE16-NEXT:    v_bfi_b32 v3, 0x7fffffff, v3, v1
-; GFX11-FAKE16-NEXT:    global_store_b64 v0, v[2:3], s[0:1]
-; GFX11-FAKE16-NEXT:    s_endpgm
-  %tid = call i32 @llvm.amdgcn.workitem.id.x()
-  %arg_mag_gep = getelementptr half, ptr addrspace(1) %arg_mag, i32 %tid
-  %mag = load half, ptr addrspace(1) %arg_mag_gep
+; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-FAKE16-NEXT:    v_cvt_f32_f16_e32 v0, v0
+; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-FAKE16-NEXT:    v_cvt_f64_f32_e32 v[0:1], v0
+; GFX11-FAKE16-NEXT:    v_bfi_b32 v1, 0x7fffffff, v1, v2
+; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
   %mag.ext = fpext half %mag to double
-  %arg_sign_gep = getelementptr double, ptr addrspace(1) %arg_sign, i32 %tid
-  %sign = load double, ptr addrspace(1) %arg_sign_gep
   %out = call double @llvm.copysign.f64(double %mag.ext, double %sign)
-  store double %out, ptr addrspace(1) %arg_out
-  ret void
+  ret double %out
 }
 
-define amdgpu_kernel void @v_copysign_out_f32_mag_f32_sign_f16(ptr addrspace(1) %arg_out, ptr addrspace(1) %arg_mag, ptr addrspace(1) %arg_sign) {
+define float @v_copysign_out_f32_mag_f32_sign_f16(float %mag, half %sign) {
 ; SI-LABEL: v_copysign_out_f32_mag_f32_sign_f16:
 ; SI:       ; %bb.0:
-; SI-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
-; SI-NEXT:    s_load_dwordx2 s[4:5], s[4:5], 0xd
-; SI-NEXT:    s_mov_b32 s11, 0xf000
-; SI-NEXT:    s_mov_b32 s14, 0
-; SI-NEXT:    s_mov_b32 s15, s11
-; SI-NEXT:    s_waitcnt lgkmcnt(0)
-; SI-NEXT:    s_mov_b64 s[12:13], s[2:3]
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 2, v0
-; SI-NEXT:    v_mov_b32_e32 v2, 0
-; SI-NEXT:    buffer_load_dword v3, v[1:2], s[12:15], 0 addr64
-; SI-NEXT:    s_mov_b64 s[6:7], s[14:15]
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 1, v0
-; SI-NEXT:    buffer_load_ushort v0, v[1:2], s[4:7], 0 addr64
-; SI-NEXT:    s_mov_b32 s8, s0
-; SI-NEXT:    s_brev_b32 s0, -2
-; SI-NEXT:    s_mov_b32 s10, -1
-; SI-NEXT:    s_mov_b32 s9, s1
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v0, v0
-; SI-NEXT:    v_bfi_b32 v0, s0, v3, v0
-; SI-NEXT:    buffer_store_dword v0, off, s[8:11], 0
-; SI-NEXT:    s_endpgm
+; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; SI-NEXT:    s_brev_b32 s4, -2
+; SI-NEXT:    v_bfi_b32 v0, s4, v0, v1
+; SI-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; VI-LABEL: v_copysign_out_f32_mag_f32_sign_f16:
 ; VI:       ; %bb.0:
-; VI-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
-; VI-NEXT:    s_load_dwordx2 s[4:5], s[4:5], 0x34
-; VI-NEXT:    v_lshlrev_b32_e32 v2, 2, v0
-; VI-NEXT:    v_lshlrev_b32_e32 v0, 1, v0
-; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    v_mov_b32_e32 v3, s3
-; VI-NEXT:    v_mov_b32_e32 v1, s5
-; VI-NEXT:    v_add_u32_e32 v0, vcc, s4, v0
-; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; VI-NEXT:    flat_load_ushort v4, v[0:1]
-; VI-NEXT:    v_add_u32_e32 v0, vcc, s2, v2
-; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v3, vcc
-; VI-NEXT:    flat_load_dword v2, v[0:1]
-; VI-NEXT:    v_mov_b32_e32 v0, s0
-; VI-NEXT:    s_brev_b32 s0, -2
-; VI-NEXT:    v_mov_b32_e32 v1, s1
-; VI-NEXT:    s_waitcnt vmcnt(1)
-; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v4
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_bfi_b32 v2, s0, v2, v3
-; VI-NEXT:    flat_store_dword v[0:1], v2
-; VI-NEXT:    s_endpgm
+; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
+; VI-NEXT:    s_brev_b32 s4, -2
+; VI-NEXT:    v_bfi_b32 v0, s4, v0, v1
+; VI-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX9-LABEL: v_copysign_out_f32_mag_f32_sign_f16:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x34
-; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
-; GFX9-NEXT:    v_lshlrev_b32_e32 v1, 1, v0
-; GFX9-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
-; GFX9-NEXT:    v_mov_b32_e32 v2, 0
-; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    global_load_ushort v1, v1, s[6:7]
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX9-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX9-NEXT:    global_load_dword v0, v0, s[2:3]
-; GFX9-NEXT:    s_brev_b32 s2, -2
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_bfi_b32 v0, s2, v0, v1
-; GFX9-NEXT:    global_store_dword v2, v0, s[0:1]
-; GFX9-NEXT:    s_endpgm
-;
-; GFX11-TRUE16-LABEL: v_copysign_out_f32_mag_f32_sign_f16:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_clause 0x1
-; GFX11-TRUE16-NEXT:    s_load_b64 s[6:7], s[4:5], 0x34
-; GFX11-TRUE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, 0 :: v_dual_and_b32 v1, 0x3ff, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 1, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 2, v1
-; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    global_load_d16_b16 v0, v0, s[6:7]
-; GFX11-TRUE16-NEXT:    global_load_b32 v1, v1, s[2:3]
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(1)
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_bfi_b32 v0, 0x7fffffff, v1, v0
-; GFX11-TRUE16-NEXT:    global_store_b32 v2, v0, s[0:1]
-; GFX11-TRUE16-NEXT:    s_endpgm
+; GFX9-NEXT:    s_brev_b32 s4, -2
+; GFX9-NEXT:    v_bfi_b32 v0, s4, v0, v1
+; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX11-FAKE16-LABEL: v_copysign_out_f32_mag_f32_sign_f16:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_clause 0x1
-; GFX11-FAKE16-NEXT:    s_load_b64 s[6:7], s[4:5], 0x34
-; GFX11-FAKE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0x3ff, v0
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, 0 :: v_dual_lshlrev_b32 v1, 1, v0
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
-; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    global_load_u16 v1, v1, s[6:7]
-; GFX11-FAKE16-NEXT:    global_load_b32 v0, v0, s[2:3]
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(1)
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_bfi_b32 v0, 0x7fffffff, v0, v1
-; GFX11-FAKE16-NEXT:    global_store_b32 v2, v0, s[0:1]
-; GFX11-FAKE16-NEXT:    s_endpgm
-  %tid = call i32 @llvm.amdgcn.workitem.id.x()
-  %arg_mag_gep = getelementptr float, ptr addrspace(1) %arg_mag, i32 %tid
-  %mag = load float, ptr addrspace(1) %arg_mag_gep
-  %arg_sign_gep = getelementptr half, ptr addrspace(1) %arg_sign, i32 %tid
-  %sign = load half, ptr addrspace(1) %arg_sign_gep
+; GFX11-LABEL: v_copysign_out_f32_mag_f32_sign_f16:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_bfi_b32 v0, 0x7fffffff, v0, v1
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %sign.ext = fpext half %sign to float
   %out = call float @llvm.copysign.f32(float %mag, float %sign.ext)
-  store float %out, ptr addrspace(1) %arg_out
-  ret void
+  ret float %out
 }
 
-define amdgpu_kernel void @v_copysign_out_f64_mag_f64_sign_f16(ptr addrspace(1) %arg_out, ptr addrspace(1) %arg_mag, ptr addrspace(1) %arg_sign) {
+define double @v_copysign_out_f64_mag_f64_sign_f16(double %mag, half %sign) {
 ; SI-LABEL: v_copysign_out_f64_mag_f64_sign_f16:
 ; SI:       ; %bb.0:
-; SI-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
-; SI-NEXT:    s_load_dwordx2 s[4:5], s[4:5], 0xd
-; SI-NEXT:    s_mov_b32 s11, 0xf000
-; SI-NEXT:    s_mov_b32 s14, 0
-; SI-NEXT:    s_mov_b32 s15, s11
-; SI-NEXT:    v_mov_b32_e32 v1, 0
-; SI-NEXT:    s_mov_b64 s[6:7], s[14:15]
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 1, v0
-; SI-NEXT:    v_mov_b32_e32 v3, v1
-; SI-NEXT:    s_waitcnt lgkmcnt(0)
-; SI-NEXT:    buffer_load_ushort v2, v[2:3], s[4:7], 0 addr64
-; SI-NEXT:    s_mov_b64 s[12:13], s[2:3]
-; SI-NEXT:    v_lshlrev_b32_e32 v0, 3, v0
-; SI-NEXT:    buffer_load_dwordx2 v[0:1], v[0:1], s[12:15], 0 addr64
-; SI-NEXT:    s_mov_b32 s8, s0
-; SI-NEXT:    s_brev_b32 s0, -2
-; SI-NEXT:    s_mov_b32 s10, -1
-; SI-NEXT:    s_mov_b32 s9, s1
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_cvt_f32_f16_e32 v2, v2
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_bfi_b32 v1, s0, v1, v2
-; SI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[8:11], 0
-; SI-NEXT:    s_endpgm
+; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; SI-NEXT:    s_brev_b32 s4, -2
+; SI-NEXT:    v_bfi_b32 v1, s4, v1, v2
+; SI-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; VI-LABEL: v_copysign_out_f64_mag_f64_sign_f16:
 ; VI:       ; %bb.0:
-; VI-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
-; VI-NEXT:    s_load_dwordx2 s[4:5], s[4:5], 0x34
-; VI-NEXT:    v_lshlrev_b32_e32 v2, 3, v0
-; VI-NEXT:    v_lshlrev_b32_e32 v0, 1, v0
-; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    v_mov_b32_e32 v3, s3
-; VI-NEXT:    v_mov_b32_e32 v1, s5
-; VI-NEXT:    v_add_u32_e32 v0, vcc, s4, v0
-; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; VI-NEXT:    flat_load_ushort v4, v[0:1]
-; VI-NEXT:    v_add_u32_e32 v0, vcc, s2, v2
-; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v3, vcc
-; VI-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
-; VI-NEXT:    v_mov_b32_e32 v2, s0
-; VI-NEXT:    s_brev_b32 s0, -2
-; VI-NEXT:    v_mov_b32_e32 v3, s1
-; VI-NEXT:    s_waitcnt vmcnt(1)
-; VI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_bfi_b32 v1, s0, v1, v4
-; VI-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
-; VI-NEXT:    s_endpgm
+; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
+; VI-NEXT:    s_brev_b32 s4, -2
+; VI-NEXT:    v_bfi_b32 v1, s4, v1, v2
+; VI-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX9-LABEL: v_copysign_out_f64_mag_f64_sign_f16:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x34
-; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
-; GFX9-NEXT:    v_lshlrev_b32_e32 v1, 1, v0
-; GFX9-NEXT:    v_lshlrev_b32_e32 v0, 3, v0
-; GFX9-NEXT:    v_mov_b32_e32 v3, 0
-; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    global_load_ushort v2, v1, s[6:7]
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX9-NEXT:    global_load_dwordx2 v[0:1], v0, s[2:3]
-; GFX9-NEXT:    s_brev_b32 s2, -2
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_bfi_b32 v1, s2, v1, v2
-; GFX9-NEXT:    global_store_dwordx2 v3, v[0:1], s[0:1]
-; GFX9-NEXT:    s_endpgm
-;
-; GFX11-TRUE16-LABEL: v_copysign_out_f64_mag_f64_sign_f16:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_clause 0x1
-; GFX11-TRUE16-NEXT:    s_load_b64 s[6:7], s[4:5], 0x34
-; GFX11-TRUE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0x3ff, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v3, 0 :: v_dual_lshlrev_b32 v0, 1, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 3, v1
-; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    global_load_d16_b16 v0, v0, s[6:7]
-; GFX11-TRUE16-NEXT:    global_load_b64 v[1:2], v1, s[2:3]
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(1)
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_bfi_b32 v2, 0x7fffffff, v2, v0
-; GFX11-TRUE16-NEXT:    global_store_b64 v3, v[1:2], s[0:1]
-; GFX11-TRUE16-NEXT:    s_endpgm
+; GFX9-NEXT:    s_brev_b32 s4, -2
+; GFX9-NEXT:    v_bfi_b32 v1, s4, v1, v2
+; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX11-FAKE16-LABEL: v_copysign_out_f64_mag_f64_sign_f16:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_clause 0x1
-; GFX11-FAKE16-NEXT:    s_load_b64 s[6:7], s[4:5], 0x34
-; GFX11-FAKE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v3, 0 :: v_dual_and_b32 v0, 0x3ff, v0
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v1, 1, v0
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v0, 3, v0
-; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    global_load_u16 v2, v1, s[6:7]
-; GFX11-FAKE16-NEXT:    global_load_b64 v[0:1], v0, s[2:3]
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(1)
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_bfi_b32 v1, 0x7fffffff, v1, v2
-; GFX11-FAKE16-NEXT:    global_store_b64 v3, v[0:1], s[0:1]
-; GFX11-FAKE16-NEXT:    s_endpgm
-  %tid = call i32 @llvm.amdgcn.workitem.id.x()
-  %arg_mag_gep = getelementptr double, ptr addrspace(1) %arg_mag, i32 %tid
-  %mag = load double, ptr addrspace(1) %arg_mag_gep
-  %arg_sign_gep = getelementptr half, ptr addrspace(1) %arg_sign, i32 %tid
-  %sign = load half, ptr addrspace(1) %arg_sign_gep
+; GFX11-LABEL: v_copysign_out_f64_mag_f64_sign_f16:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_bfi_b32 v1, 0x7fffffff, v1, v2
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %sign.ext = fpext half %sign to double
   %out = call double @llvm.copysign.f64(double %mag, double %sign.ext)
-  store double %out, ptr addrspace(1) %arg_out
-  ret void
+  ret double %out
 }
 
-define amdgpu_kernel void @v_copysign_out_f16_mag_f16_sign_f32(ptr addrspace(1) %arg_out, ptr addrspace(1) %arg_mag, ptr addrspace(1) %arg_sign) {
+define half @v_copysign_out_f16_mag_f16_sign_f32(half %mag, float %sign) {
 ; SI-LABEL: v_copysign_out_f16_mag_f16_sign_f32:
 ; SI:       ; %bb.0:
-; SI-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
-; SI-NEXT:    s_load_dwordx2 s[4:5], s[4:5], 0xd
-; SI-NEXT:    s_mov_b32 s11, 0xf000
-; SI-NEXT:    s_mov_b32 s14, 0
-; SI-NEXT:    s_mov_b32 s15, s11
-; SI-NEXT:    s_waitcnt lgkmcnt(0)
-; SI-NEXT:    s_mov_b64 s[12:13], s[2:3]
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 1, v0
-; SI-NEXT:    v_mov_b32_e32 v2, 0
-; SI-NEXT:    buffer_load_ushort v3, v[1:2], s[12:15], 0 addr64
-; SI-NEXT:    s_mov_b64 s[6:7], s[14:15]
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 2, v0
-; SI-NEXT:    buffer_load_dword v0, v[1:2], s[4:7], 0 addr64
-; SI-NEXT:    s_brev_b32 s2, -2
-; SI-NEXT:    s_mov_b32 s10, -1
-; SI-NEXT:    s_mov_b32 s8, s0
-; SI-NEXT:    s_mov_b32 s9, s1
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_cvt_f32_f16_e32 v1, v3
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_bfi_b32 v0, s2, v1, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
-; SI-NEXT:    buffer_store_short v0, off, s[8:11], 0
-; SI-NEXT:    s_endpgm
+; SI-NEXT:    s_brev_b32 s4, -2
+; SI-NEXT:    v_cvt_f32_f16_e32 v0, v0
+; SI-NEXT:    v_bfi_b32 v0, s4, v0, v1
+; SI-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; VI-LABEL: v_copysign_out_f16_mag_f16_sign_f32:
 ; VI:       ; %bb.0:
-; VI-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
-; VI-NEXT:    s_load_dwordx2 s[4:5], s[4:5], 0x34
-; VI-NEXT:    v_lshlrev_b32_e32 v2, 1, v0
-; VI-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
-; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    v_mov_b32_e32 v3, s3
-; VI-NEXT:    v_mov_b32_e32 v1, s5
-; VI-NEXT:    v_add_u32_e32 v0, vcc, s4, v0
-; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; VI-NEXT:    flat_load_dword v4, v[0:1]
-; VI-NEXT:    v_add_u32_e32 v0, vcc, s2, v2
-; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v3, vcc
-; VI-NEXT:    flat_load_ushort v2, v[0:1]
-; VI-NEXT:    v_mov_b32_e32 v0, s0
-; VI-NEXT:    s_movk_i32 s0, 0x7fff
-; VI-NEXT:    v_mov_b32_e32 v1, s1
-; VI-NEXT:    s_waitcnt vmcnt(1)
-; VI-NEXT:    v_lshrrev_b32_e32 v3, 16, v4
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_bfi_b32 v2, s0, v2, v3
-; VI-NEXT:    flat_store_short v[0:1], v2
-; VI-NEXT:    s_endpgm
+; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; VI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
+; VI-NEXT:    s_movk_i32 s4, 0x7fff
+; VI-NEXT:    v_bfi_b32 v0, s4, v0, v1
+; VI-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX9-LABEL: v_copysign_out_f16_mag_f16_sign_f32:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x34
-; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
-; GFX9-NEXT:    v_lshlrev_b32_e32 v1, 2, v0
-; GFX9-NEXT:    v_lshlrev_b32_e32 v0, 1, v0
-; GFX9-NEXT:    v_mov_b32_e32 v2, 0
-; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    global_load_dword v1, v1, s[6:7]
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX9-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
-; GFX9-NEXT:    global_load_ushort v0, v0, s[2:3]
-; GFX9-NEXT:    s_movk_i32 s2, 0x7fff
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_bfi_b32 v0, s2, v0, v1
-; GFX9-NEXT:    global_store_short v2, v0, s[0:1]
-; GFX9-NEXT:    s_endpgm
-;
-; GFX11-TRUE16-LABEL: v_copysign_out_f16_mag_f16_sign_f32:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_clause 0x1
-; GFX11-TRUE16-NEXT:    s_load_b64 s[6:7], s[4:5], 0x34
-; GFX11-TRUE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0x3ff, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 2, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 1, v0
-; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    global_load_b32 v1, v1, s[6:7]
-; GFX11-TRUE16-NEXT:    global_load_d16_b16 v0, v0, s[2:3]
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(1)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_bfi_b32 v0, 0x7fff, v0, v1
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v1, 0
-; GFX11-TRUE16-NEXT:    global_store_b16 v1, v0, s[0:1]
-; GFX11-TRUE16-NEXT:    s_endpgm
+; GFX9-NEXT:    s_movk_i32 s4, 0x7fff
+; GFX9-NEXT:    v_bfi_b32 v0, s4, v0, v1
+; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX11-FAKE16-LABEL: v_copysign_out_f16_mag_f16_sign_f32:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_clause 0x1
-; GFX11-FAKE16-NEXT:    s_load_b64 s[6:7], s[4:5], 0x34
-; GFX11-FAKE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0x3ff, v0
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, 0 :: v_dual_lshlrev_b32 v1, 2, v0
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v0, 1, v0
-; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    global_load_b32 v1, v1, s[6:7]
-; GFX11-FAKE16-NEXT:    global_load_u16 v0, v0, s[2:3]
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(1)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_bfi_b32 v0, 0x7fff, v0, v1
-; GFX11-FAKE16-NEXT:    global_store_b16 v2, v0, s[0:1]
-; GFX11-FAKE16-NEXT:    s_endpgm
-  %tid = call i32 @llvm.amdgcn.workitem.id.x()
-  %arg_mag_gep = getelementptr half, ptr addrspace(1) %arg_mag, i32 %tid
-  %mag = load half, ptr addrspace(1) %arg_mag_gep
-  %arg_sign_gep = getelementptr float, ptr addrspace(1) %arg_sign, i32 %tid
-  %sign = load float, ptr addrspace(1) %arg_sign_gep
+; GFX11-LABEL: v_copysign_out_f16_mag_f16_sign_f32:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_bfi_b32 v0, 0x7fff, v0, v1
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %sign.trunc = fptrunc float %sign to half
   %out = call half @llvm.copysign.f16(half %mag, half %sign.trunc)
-  store half %out, ptr addrspace(1) %arg_out
-  ret void
+  ret half %out
 }
 
-define amdgpu_kernel void @v_copysign_out_f16_mag_f16_sign_f64(ptr addrspace(1) %arg_out, ptr addrspace(1) %arg_mag, ptr addrspace(1) %arg_sign) {
+define half @v_copysign_out_f16_mag_f16_sign_f64(half %mag, double %sign) {
 ; SI-LABEL: v_copysign_out_f16_mag_f16_sign_f64:
 ; SI:       ; %bb.0:
-; SI-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
-; SI-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0xd
-; SI-NEXT:    s_mov_b32 s7, 0xf000
-; SI-NEXT:    s_mov_b32 s6, -1
-; SI-NEXT:    s_mov_b32 s14, s6
-; SI-NEXT:    s_waitcnt lgkmcnt(0)
-; SI-NEXT:    s_mov_b32 s12, s2
-; SI-NEXT:    s_mov_b32 s13, s3
-; SI-NEXT:    s_mov_b32 s15, s7
-; SI-NEXT:    buffer_load_ushort v2, off, s[12:15], 0
-; SI-NEXT:    s_mov_b32 s10, 0
-; SI-NEXT:    s_mov_b32 s11, s7
-; SI-NEXT:    v_lshlrev_b32_e32 v0, 3, v0
-; SI-NEXT:    v_mov_b32_e32 v1, 0
-; SI-NEXT:    buffer_load_dwordx2 v[0:1], v[0:1], s[8:11], 0 addr64
-; SI-NEXT:    s_brev_b32 s2, -2
-; SI-NEXT:    s_mov_b32 s4, s0
-; SI-NEXT:    s_mov_b32 s5, s1
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v0, v2
-; SI-NEXT:    v_bfi_b32 v0, s2, v0, v1
+; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
-; SI-NEXT:    buffer_store_short v0, off, s[4:7], 0
-; SI-NEXT:    s_endpgm
+; SI-NEXT:    s_brev_b32 s4, -2
+; SI-NEXT:    v_cvt_f32_f16_e32 v0, v0
+; SI-NEXT:    v_bfi_b32 v0, s4, v0, v2
+; SI-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; VI-LABEL: v_copysign_out_f16_mag_f16_sign_f64:
 ; VI:       ; %bb.0:
-; VI-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
-; VI-NEXT:    s_load_dwordx2 s[4:5], s[4:5], 0x34
-; VI-NEXT:    v_lshlrev_b32_e32 v1, 3, v0
-; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    v_mov_b32_e32 v0, s2
-; VI-NEXT:    v_mov_b32_e32 v2, s5
-; VI-NEXT:    v_add_u32_e32 v1, vcc, s4, v1
-; VI-NEXT:    v_addc_u32_e32 v2, vcc, 0, v2, vcc
-; VI-NEXT:    flat_load_dwordx2 v[1:2], v[1:2]
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_mov_b32_e32 v1, s3
-; VI-NEXT:    flat_load_ushort v3, v[0:1]
-; VI-NEXT:    v_mov_b32_e32 v0, s0
-; VI-NEXT:    s_movk_i32 s0, 0x7fff
-; VI-NEXT:    v_mov_b32_e32 v1, s1
-; VI-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_bfi_b32 v2, s0, v3, v2
-; VI-NEXT:    flat_store_short v[0:1], v2
-; VI-NEXT:    s_endpgm
+; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; VI-NEXT:    v_lshrrev_b32_e32 v1, 16, v2
+; VI-NEXT:    s_movk_i32 s4, 0x7fff
+; VI-NEXT:    v_bfi_b32 v0, s4, v0, v1
+; VI-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX9-LABEL: v_copysign_out_f16_mag_f16_sign_f64:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x34
-; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
-; GFX9-NEXT:    v_lshlrev_b32_e32 v0, 3, v0
-; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    global_load_dwordx2 v[0:1], v0, s[6:7]
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_mov_b32_e32 v0, 0
-; GFX9-NEXT:    global_load_ushort v2, v0, s[2:3]
-; GFX9-NEXT:    s_movk_i32 s2, 0x7fff
-; GFX9-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_bfi_b32 v1, s2, v2, v1
-; GFX9-NEXT:    global_store_short v0, v1, s[0:1]
-; GFX9-NEXT:    s_endpgm
-;
-; GFX11-TRUE16-LABEL: v_copysign_out_f16_mag_f16_sign_f64:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_clause 0x1
-; GFX11-TRUE16-NEXT:    s_load_b64 s[6:7], s[4:5], 0x34
-; GFX11-TRUE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0x3ff, v0
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v2, 0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 3, v0
-; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    global_load_b64 v[0:1], v0, s[6:7]
-; GFX11-TRUE16-NEXT:    global_load_d16_b16 v0, v2, s[2:3]
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(1)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_bfi_b32 v0, 0x7fff, v0, v1
-; GFX11-TRUE16-NEXT:    global_store_b16 v2, v0, s[0:1]
-; GFX11-TRUE16-NEXT:    s_endpgm
+; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    v_lshrrev_b32_e32 v1, 16, v2
+; GFX9-NEXT:    s_movk_i32 s4, 0x7fff
+; GFX9-NEXT:    v_bfi_b32 v0, s4, v0, v1
+; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX11-FAKE16-LABEL: v_copysign_out_f16_mag_f16_sign_f64:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_clause 0x1
-; GFX11-FAKE16-NEXT:    s_load_b64 s[6:7], s[4:5], 0x34
-; GFX11-FAKE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0x3ff, v0
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v2, 0
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v0, 3, v0
-; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    global_load_b64 v[0:1], v0, s[6:7]
-; GFX11-FAKE16-NEXT:    global_load_u16 v0, v2, s[2:3]
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(1)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_bfi_b32 v0, 0x7fff, v0, v1
-; GFX11-FAKE16-NEXT:    global_store_b16 v2, v0, s[0:1]
-; GFX11-FAKE16-NEXT:    s_endpgm
-  %tid = call i32 @llvm.amdgcn.workitem.id.x()
-  %arg_mag_gep = getelementptr half, ptr addrspace(1) %arg_mag, i32 %tid
-  %mag = load half, ptr addrspace(1) %arg_mag
-  %arg_sign_gep = getelementptr double, ptr addrspace(1) %arg_sign, i32 %tid
-  %sign = load double, ptr addrspace(1) %arg_sign_gep
+; GFX11-LABEL: v_copysign_out_f16_mag_f16_sign_f64:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v1, 16, v2
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_bfi_b32 v0, 0x7fff, v0, v1
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %sign.trunc = fptrunc double %sign to half
   %out = call half @llvm.copysign.f16(half %mag, half %sign.trunc)
-  store half %out, ptr addrspace(1) %arg_out
-  ret void
+  ret half %out
 }
 
-define amdgpu_kernel void @v_copysign_out_f16_mag_f32_sign_f16(ptr addrspace(1) %arg_out, ptr addrspace(1) %arg_mag, ptr addrspace(1) %arg_sign) {
+define half @v_copysign_out_f16_mag_f32_sign_f16(float %mag, half %sign) {
 ; SI-LABEL: v_copysign_out_f16_mag_f32_sign_f16:
 ; SI:       ; %bb.0:
-; SI-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
-; SI-NEXT:    s_load_dwordx2 s[4:5], s[4:5], 0xd
-; SI-NEXT:    s_mov_b32 s11, 0xf000
-; SI-NEXT:    s_mov_b32 s14, 0
-; SI-NEXT:    s_mov_b32 s15, s11
-; SI-NEXT:    s_waitcnt lgkmcnt(0)
-; SI-NEXT:    s_mov_b64 s[12:13], s[2:3]
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 2, v0
-; SI-NEXT:    v_mov_b32_e32 v2, 0
-; SI-NEXT:    buffer_load_dword v3, v[1:2], s[12:15], 0 addr64
-; SI-NEXT:    s_mov_b64 s[6:7], s[14:15]
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 1, v0
-; SI-NEXT:    buffer_load_ushort v0, v[1:2], s[4:7], 0 addr64
-; SI-NEXT:    s_brev_b32 s2, -2
-; SI-NEXT:    s_mov_b32 s10, -1
-; SI-NEXT:    s_mov_b32 s8, s0
-; SI-NEXT:    s_mov_b32 s9, s1
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_cvt_f16_f32_e32 v1, v3
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v0, v0
-; SI-NEXT:    v_cvt_f32_f16_e32 v1, v1
-; SI-NEXT:    v_bfi_b32 v0, s2, v1, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
-; SI-NEXT:    buffer_store_short v0, off, s[8:11], 0
-; SI-NEXT:    s_endpgm
+; SI-NEXT:    s_brev_b32 s4, -2
+; SI-NEXT:    v_cvt_f32_f16_e32 v0, v0
+; SI-NEXT:    v_bfi_b32 v0, s4, v0, v1
+; SI-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; VI-LABEL: v_copysign_out_f16_mag_f32_sign_f16:
 ; VI:       ; %bb.0:
-; VI-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
-; VI-NEXT:    s_load_dwordx2 s[4:5], s[4:5], 0x34
-; VI-NEXT:    v_lshlrev_b32_e32 v1, 2, v0
-; VI-NEXT:    v_lshlrev_b32_e32 v0, 1, v0
-; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    v_mov_b32_e32 v2, s3
-; VI-NEXT:    v_add_u32_e32 v1, vcc, s2, v1
-; VI-NEXT:    v_addc_u32_e32 v2, vcc, 0, v2, vcc
-; VI-NEXT:    flat_load_dword v2, v[1:2]
-; VI-NEXT:    v_mov_b32_e32 v1, s5
-; VI-NEXT:    v_add_u32_e32 v0, vcc, s4, v0
-; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; VI-NEXT:    flat_load_ushort v3, v[0:1]
-; VI-NEXT:    v_mov_b32_e32 v0, s0
-; VI-NEXT:    s_movk_i32 s0, 0x7fff
-; VI-NEXT:    v_mov_b32_e32 v1, s1
-; VI-NEXT:    s_waitcnt vmcnt(1)
-; VI-NEXT:    v_cvt_f16_f32_e32 v2, v2
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_bfi_b32 v2, s0, v2, v3
-; VI-NEXT:    flat_store_short v[0:1], v2
-; VI-NEXT:    s_endpgm
+; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; VI-NEXT:    v_cvt_f16_f32_e32 v0, v0
+; VI-NEXT:    s_movk_i32 s4, 0x7fff
+; VI-NEXT:    v_bfi_b32 v0, s4, v0, v1
+; VI-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX9-LABEL: v_copysign_out_f16_mag_f32_sign_f16:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
-; GFX9-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x34
-; GFX9-NEXT:    v_lshlrev_b32_e32 v1, 2, v0
-; GFX9-NEXT:    v_lshlrev_b32_e32 v0, 1, v0
-; GFX9-NEXT:    v_mov_b32_e32 v2, 0
-; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    global_load_dword v1, v1, s[2:3]
-; GFX9-NEXT:    s_movk_i32 s2, 0x7fff
-; GFX9-NEXT:    global_load_ushort v0, v0, s[6:7]
-; GFX9-NEXT:    s_waitcnt vmcnt(1)
-; GFX9-NEXT:    v_cvt_f16_f32_e32 v1, v1
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_bfi_b32 v0, s2, v1, v0
-; GFX9-NEXT:    global_store_short v2, v0, s[0:1]
-; GFX9-NEXT:    s_endpgm
+; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    v_cvt_f16_f32_e32 v0, v0
+; GFX9-NEXT:    s_movk_i32 s4, 0x7fff
+; GFX9-NEXT:    v_bfi_b32 v0, s4, v0, v1
+; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX11-TRUE16-LABEL: v_copysign_out_f16_mag_f32_sign_f16:
 ; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_clause 0x1
-; GFX11-TRUE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
-; GFX11-TRUE16-NEXT:    s_load_b64 s[4:5], s[4:5], 0x34
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0x3ff, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 2, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 1, v0
-; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    global_load_b32 v1, v1, s[2:3]
-; GFX11-TRUE16-NEXT:    global_load_d16_b16 v0, v0, s[4:5]
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(1)
-; GFX11-TRUE16-NEXT:    v_cvt_f16_f32_e32 v1.l, v1
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    v_cvt_f16_f32_e32 v0.l, v0
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_bfi_b32 v0, 0x7fff, v1, v0
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v1, 0
-; GFX11-TRUE16-NEXT:    global_store_b16 v1, v0, s[0:1]
-; GFX11-TRUE16-NEXT:    s_endpgm
+; GFX11-TRUE16-NEXT:    v_bfi_b32 v0, 0x7fff, v0, v1
+; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX11-FAKE16-LABEL: v_copysign_out_f16_mag_f32_sign_f16:
 ; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_clause 0x1
-; GFX11-FAKE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
-; GFX11-FAKE16-NEXT:    s_load_b64 s[4:5], s[4:5], 0x34
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0x3ff, v0
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, 0 :: v_dual_lshlrev_b32 v1, 2, v0
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v0, 1, v0
-; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    global_load_b32 v1, v1, s[2:3]
-; GFX11-FAKE16-NEXT:    global_load_u16 v0, v0, s[4:5]
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(1)
-; GFX11-FAKE16-NEXT:    v_cvt_f16_f32_e32 v1, v1
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-FAKE16-NEXT:    v_cvt_f16_f32_e32 v0, v0
 ; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_bfi_b32 v0, 0x7fff, v1, v0
-; GFX11-FAKE16-NEXT:    global_store_b16 v2, v0, s[0:1]
-; GFX11-FAKE16-NEXT:    s_endpgm
-  %tid = call i32 @llvm.amdgcn.workitem.id.x()
-  %arg_mag_gep = getelementptr float, ptr addrspace(1) %arg_mag, i32 %tid
-  %mag = load float, ptr addrspace(1) %arg_mag_gep
+; GFX11-FAKE16-NEXT:    v_bfi_b32 v0, 0x7fff, v0, v1
+; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
   %mag.trunc = fptrunc float %mag to half
-  %arg_sign_gep = getelementptr half, ptr addrspace(1) %arg_sign, i32 %tid
-  %sign = load half, ptr addrspace(1) %arg_sign_gep
   %out = call half @llvm.copysign.f16(half %mag.trunc, half %sign)
-  store half %out, ptr addrspace(1) %arg_out
-  ret void
+  ret half %out
 }
 
-define amdgpu_kernel void @s_copysign_out_f16_mag_f64_sign_f16(ptr addrspace(1) %arg_out, double %mag, half %sign) {
+define amdgpu_ps i16 @s_copysign_out_f16_mag_f64_sign_f16(double inreg %mag, half inreg %sign) {
 ; SI-LABEL: s_copysign_out_f16_mag_f64_sign_f16:
 ; SI:       ; %bb.0:
-; SI-NEXT:    s_load_dword s6, s[4:5], 0xd
-; SI-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
-; SI-NEXT:    s_waitcnt lgkmcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v0, s6
-; SI-NEXT:    s_lshr_b32 s4, s3, 8
-; SI-NEXT:    s_and_b32 s5, s3, 0x1ff
-; SI-NEXT:    s_and_b32 s6, s4, 0xffe
-; SI-NEXT:    s_or_b32 s2, s5, s2
-; SI-NEXT:    s_cmp_lg_u32 s2, 0
+; SI-NEXT:    s_and_b32 s3, s1, 0x1ff
+; SI-NEXT:    s_or_b32 s0, s3, s0
+; SI-NEXT:    s_cmp_lg_u32 s0, 0
 ; SI-NEXT:    s_cselect_b64 s[4:5], -1, 0
-; SI-NEXT:    v_cndmask_b32_e64 v1, 0, 1, s[4:5]
-; SI-NEXT:    v_readfirstlane_b32 s2, v1
-; SI-NEXT:    s_bfe_u32 s5, s3, 0xb0014
-; SI-NEXT:    s_or_b32 s2, s6, s2
-; SI-NEXT:    s_sub_i32 s6, 0x3f1, s5
-; SI-NEXT:    v_med3_i32 v1, s6, 0, 13
-; SI-NEXT:    s_or_b32 s4, s2, 0x1000
-; SI-NEXT:    v_readfirstlane_b32 s6, v1
-; SI-NEXT:    s_lshr_b32 s7, s4, s6
-; SI-NEXT:    s_lshl_b32 s6, s7, s6
-; SI-NEXT:    s_cmp_lg_u32 s6, s4
-; SI-NEXT:    s_cselect_b32 s4, 1, 0
-; SI-NEXT:    s_addk_i32 s5, 0xfc10
-; SI-NEXT:    s_lshl_b32 s6, s5, 12
-; SI-NEXT:    s_or_b32 s4, s7, s4
-; SI-NEXT:    s_or_b32 s6, s2, s6
-; SI-NEXT:    s_cmp_lt_i32 s5, 1
-; SI-NEXT:    s_cselect_b32 s4, s4, s6
-; SI-NEXT:    s_and_b32 s6, s4, 7
-; SI-NEXT:    s_cmp_gt_i32 s6, 5
-; SI-NEXT:    s_cselect_b32 s7, 1, 0
-; SI-NEXT:    s_cmp_eq_u32 s6, 3
+; SI-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s[4:5]
+; SI-NEXT:    s_lshr_b32 s0, s1, 8
+; SI-NEXT:    s_bfe_u32 s4, s1, 0xb0014
+; SI-NEXT:    s_and_b32 s0, s0, 0xffe
+; SI-NEXT:    v_readfirstlane_b32 s3, v0
+; SI-NEXT:    s_sub_i32 s5, 0x3f1, s4
+; SI-NEXT:    s_or_b32 s0, s0, s3
+; SI-NEXT:    v_med3_i32 v0, s5, 0, 13
+; SI-NEXT:    s_or_b32 s3, s0, 0x1000
+; SI-NEXT:    v_readfirstlane_b32 s5, v0
+; SI-NEXT:    s_lshr_b32 s6, s3, s5
+; SI-NEXT:    s_lshl_b32 s5, s6, s5
+; SI-NEXT:    s_cmp_lg_u32 s5, s3
+; SI-NEXT:    s_cselect_b32 s3, 1, 0
+; SI-NEXT:    s_addk_i32 s4, 0xfc10
+; SI-NEXT:    s_lshl_b32 s5, s4, 12
+; SI-NEXT:    s_or_b32 s3, s6, s3
+; SI-NEXT:    s_or_b32 s5, s0, s5
+; SI-NEXT:    s_cmp_lt_i32 s4, 1
+; SI-NEXT:    s_cselect_b32 s3, s3, s5
+; SI-NEXT:    s_and_b32 s5, s3, 7
+; SI-NEXT:    s_cmp_gt_i32 s5, 5
 ; SI-NEXT:    s_cselect_b32 s6, 1, 0
-; SI-NEXT:    s_or_b32 s6, s6, s7
-; SI-NEXT:    s_lshr_b32 s4, s4, 2
-; SI-NEXT:    s_add_i32 s4, s4, s6
-; SI-NEXT:    s_cmp_lt_i32 s5, 31
-; SI-NEXT:    s_cselect_b32 s4, s4, 0x7c00
-; SI-NEXT:    s_cmp_lg_u32 s2, 0
-; SI-NEXT:    s_movk_i32 s2, 0x7e00
-; SI-NEXT:    s_cselect_b32 s2, s2, 0x7c00
-; SI-NEXT:    s_cmpk_eq_i32 s5, 0x40f
-; SI-NEXT:    s_cselect_b32 s2, s2, s4
-; SI-NEXT:    s_lshr_b32 s3, s3, 16
-; SI-NEXT:    s_and_b32 s3, s3, 0x8000
-; SI-NEXT:    s_or_b32 s2, s3, s2
-; SI-NEXT:    v_cvt_f32_f16_e32 v1, s2
-; SI-NEXT:    s_brev_b32 s2, -2
-; SI-NEXT:    s_mov_b32 s3, 0xf000
-; SI-NEXT:    v_bfi_b32 v0, s2, v1, v0
+; SI-NEXT:    s_cmp_eq_u32 s5, 3
+; SI-NEXT:    s_cselect_b32 s5, 1, 0
+; SI-NEXT:    s_or_b32 s5, s5, s6
+; SI-NEXT:    s_lshr_b32 s3, s3, 2
+; SI-NEXT:    s_add_i32 s3, s3, s5
+; SI-NEXT:    s_cmp_lt_i32 s4, 31
+; SI-NEXT:    s_cselect_b32 s3, s3, 0x7c00
+; SI-NEXT:    s_cmp_lg_u32 s0, 0
+; SI-NEXT:    s_movk_i32 s0, 0x7e00
+; SI-NEXT:    s_cselect_b32 s0, s0, 0x7c00
+; SI-NEXT:    s_cmpk_eq_i32 s4, 0x40f
+; SI-NEXT:    s_cselect_b32 s0, s0, s3
+; SI-NEXT:    s_lshr_b32 s1, s1, 16
+; SI-NEXT:    s_and_b32 s1, s1, 0x8000
+; SI-NEXT:    s_or_b32 s0, s1, s0
+; SI-NEXT:    v_cvt_f32_f16_e32 v0, s0
+; SI-NEXT:    s_brev_b32 s0, -2
+; SI-NEXT:    v_mov_b32_e32 v1, s2
+; SI-NEXT:    v_bfi_b32 v0, s0, v0, v1
 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
-; SI-NEXT:    s_mov_b32 s2, -1
-; SI-NEXT:    buffer_store_short v0, off, s[0:3], 0
-; SI-NEXT:    s_endpgm
+; SI-NEXT:    v_readfirstlane_b32 s0, v0
+; SI-NEXT:    ; return to shader part epilog
 ;
 ; VI-LABEL: s_copysign_out_f16_mag_f64_sign_f16:
 ; VI:       ; %bb.0:
-; VI-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
-; VI-NEXT:    s_load_dword s4, s[4:5], 0x34
-; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    v_mov_b32_e32 v0, s0
-; VI-NEXT:    v_mov_b32_e32 v1, s1
-; VI-NEXT:    s_lshr_b32 s0, s3, 8
-; VI-NEXT:    s_and_b32 s1, s3, 0x1ff
-; VI-NEXT:    s_and_b32 s5, s0, 0xffe
-; VI-NEXT:    s_or_b32 s0, s1, s2
+; VI-NEXT:    s_and_b32 s3, s1, 0x1ff
+; VI-NEXT:    s_or_b32 s0, s3, s0
 ; VI-NEXT:    s_cmp_lg_u32 s0, 0
-; VI-NEXT:    s_cselect_b64 s[0:1], -1, 0
-; VI-NEXT:    v_cndmask_b32_e64 v2, 0, 1, s[0:1]
-; VI-NEXT:    s_bfe_u32 s2, s3, 0xb0014
-; VI-NEXT:    v_readfirstlane_b32 s0, v2
-; VI-NEXT:    s_sub_i32 s3, 0x3f1, s2
-; VI-NEXT:    s_or_b32 s0, s5, s0
-; VI-NEXT:    v_med3_i32 v2, s3, 0, 13
-; VI-NEXT:    s_or_b32 s1, s0, 0x1000
-; VI-NEXT:    v_readfirstlane_b32 s3, v2
-; VI-NEXT:    s_lshr_b32 s5, s1, s3
-; VI-NEXT:    s_lshl_b32 s3, s5, s3
-; VI-NEXT:    s_cmp_lg_u32 s3, s1
-; VI-NEXT:    s_cselect_b32 s1, 1, 0
-; VI-NEXT:    s_addk_i32 s2, 0xfc10
-; VI-NEXT:    s_lshl_b32 s3, s2, 12
-; VI-NEXT:    s_or_b32 s1, s5, s1
-; VI-NEXT:    s_or_b32 s3, s0, s3
-; VI-NEXT:    s_cmp_lt_i32 s2, 1
-; VI-NEXT:    s_cselect_b32 s1, s1, s3
-; VI-NEXT:    s_and_b32 s3, s1, 7
-; VI-NEXT:    s_cmp_gt_i32 s3, 5
-; VI-NEXT:    s_cselect_b32 s5, 1, 0
-; VI-NEXT:    s_cmp_eq_u32 s3, 3
+; VI-NEXT:    s_cselect_b64 s[4:5], -1, 0
+; VI-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s[4:5]
+; VI-NEXT:    s_lshr_b32 s0, s1, 8
+; VI-NEXT:    s_bfe_u32 s1, s1, 0xb0014
+; VI-NEXT:    s_and_b32 s0, s0, 0xffe
+; VI-NEXT:    v_readfirstlane_b32 s3, v0
+; VI-NEXT:    s_sub_i32 s4, 0x3f1, s1
+; VI-NEXT:    s_or_b32 s0, s0, s3
+; VI-NEXT:    v_med3_i32 v0, s4, 0, 13
+; VI-NEXT:    s_or_b32 s3, s0, 0x1000
+; VI-NEXT:    v_readfirstlane_b32 s4, v0
+; VI-NEXT:    s_lshr_b32 s5, s3, s4
+; VI-NEXT:    s_lshl_b32 s4, s5, s4
+; VI-NEXT:    s_cmp_lg_u32 s4, s3
 ; VI-NEXT:    s_cselect_b32 s3, 1, 0
-; VI-NEXT:    s_or_b32 s3, s3, s5
-; VI-NEXT:    s_lshr_b32 s1, s1, 2
-; VI-NEXT:    s_add_i32 s1, s1, s3
-; VI-NEXT:    s_cmp_lt_i32 s2, 31
-; VI-NEXT:    s_cselect_b32 s1, s1, 0x7c00
+; VI-NEXT:    s_addk_i32 s1, 0xfc10
+; VI-NEXT:    s_lshl_b32 s4, s1, 12
+; VI-NEXT:    s_or_b32 s3, s5, s3
+; VI-NEXT:    s_or_b32 s4, s0, s4
+; VI-NEXT:    s_cmp_lt_i32 s1, 1
+; VI-NEXT:    s_cselect_b32 s3, s3, s4
+; VI-NEXT:    s_and_b32 s4, s3, 7
+; VI-NEXT:    s_cmp_gt_i32 s4, 5
+; VI-NEXT:    s_cselect_b32 s5, 1, 0
+; VI-NEXT:    s_cmp_eq_u32 s4, 3
+; VI-NEXT:    s_cselect_b32 s4, 1, 0
+; VI-NEXT:    s_or_b32 s4, s4, s5
+; VI-NEXT:    s_lshr_b32 s3, s3, 2
+; VI-NEXT:    s_add_i32 s3, s3, s4
+; VI-NEXT:    s_cmp_lt_i32 s1, 31
+; VI-NEXT:    s_cselect_b32 s3, s3, 0x7c00
 ; VI-NEXT:    s_cmp_lg_u32 s0, 0
 ; VI-NEXT:    s_movk_i32 s0, 0x7e00
 ; VI-NEXT:    s_cselect_b32 s0, s0, 0x7c00
-; VI-NEXT:    s_cmpk_eq_i32 s2, 0x40f
-; VI-NEXT:    s_cselect_b32 s0, s0, s1
+; VI-NEXT:    s_cmpk_eq_i32 s1, 0x40f
+; VI-NEXT:    s_cselect_b32 s0, s0, s3
 ; VI-NEXT:    s_movk_i32 s1, 0x7fff
-; VI-NEXT:    v_mov_b32_e32 v2, s0
-; VI-NEXT:    v_mov_b32_e32 v3, s4
-; VI-NEXT:    v_bfi_b32 v2, s1, v2, v3
-; VI-NEXT:    flat_store_short v[0:1], v2
-; VI-NEXT:    s_endpgm
+; VI-NEXT:    v_mov_b32_e32 v0, s0
+; VI-NEXT:    v_mov_b32_e32 v1, s2
+; VI-NEXT:    v_bfi_b32 v0, s1, v0, v1
+; VI-NEXT:    v_readfirstlane_b32 s0, v0
+; VI-NEXT:    ; return to shader part epilog
 ;
 ; GFX9-LABEL: s_copysign_out_f16_mag_f64_sign_f16:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
-; GFX9-NEXT:    s_load_dword s6, s[4:5], 0x34
-; GFX9-NEXT:    v_mov_b32_e32 v0, 0
-; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    s_lshr_b32 s4, s3, 8
-; GFX9-NEXT:    s_and_b32 s5, s3, 0x1ff
-; GFX9-NEXT:    s_and_b32 s7, s4, 0xffe
-; GFX9-NEXT:    s_or_b32 s2, s5, s2
-; GFX9-NEXT:    s_cmp_lg_u32 s2, 0
+; GFX9-NEXT:    s_and_b32 s3, s1, 0x1ff
+; GFX9-NEXT:    s_or_b32 s0, s3, s0
+; GFX9-NEXT:    s_cmp_lg_u32 s0, 0
 ; GFX9-NEXT:    s_cselect_b64 s[4:5], -1, 0
-; GFX9-NEXT:    v_cndmask_b32_e64 v1, 0, 1, s[4:5]
-; GFX9-NEXT:    s_bfe_u32 s3, s3, 0xb0014
-; GFX9-NEXT:    v_readfirstlane_b32 s2, v1
-; GFX9-NEXT:    s_sub_i32 s5, 0x3f1, s3
-; GFX9-NEXT:    s_or_b32 s2, s7, s2
-; GFX9-NEXT:    v_med3_i32 v1, s5, 0, 13
-; GFX9-NEXT:    s_or_b32 s4, s2, 0x1000
-; GFX9-NEXT:    v_readfirstlane_b32 s5, v1
-; GFX9-NEXT:    s_lshr_b32 s7, s4, s5
-; GFX9-NEXT:    s_lshl_b32 s5, s7, s5
-; GFX9-NEXT:    s_cmp_lg_u32 s5, s4
-; GFX9-NEXT:    s_cselect_b32 s4, 1, 0
-; GFX9-NEXT:    s_addk_i32 s3, 0xfc10
-; GFX9-NEXT:    s_lshl_b32 s5, s3, 12
-; GFX9-NEXT:    s_or_b32 s4, s7, s4
-; GFX9-NEXT:    s_or_b32 s5, s2, s5
-; GFX9-NEXT:    s_cmp_lt_i32 s3, 1
-; GFX9-NEXT:    s_cselect_b32 s4, s4, s5
-; GFX9-NEXT:    s_and_b32 s5, s4, 7
-; GFX9-NEXT:    s_cmp_gt_i32 s5, 5
-; GFX9-NEXT:    s_cselect_b32 s7, 1, 0
-; GFX9-NEXT:    s_cmp_eq_u32 s5, 3
+; GFX9-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s[4:5]
+; GFX9-NEXT:    s_lshr_b32 s0, s1, 8
+; GFX9-NEXT:    s_bfe_u32 s1, s1, 0xb0014
+; GFX9-NEXT:    s_and_b32 s0, s0, 0xffe
+; GFX9-NEXT:    v_readfirstlane_b32 s3, v0
+; GFX9-NEXT:    s_sub_i32 s4, 0x3f1, s1
+; GFX9-NEXT:    s_or_b32 s0, s0, s3
+; GFX9-NEXT:    v_med3_i32 v0, s4, 0, 13
+; GFX9-NEXT:    s_or_b32 s3, s0, 0x1000
+; GFX9-NEXT:    v_readfirstlane_b32 s4, v0
+; GFX9-NEXT:    s_lshr_b32 s5, s3, s4
+; GFX9-NEXT:    s_lshl_b32 s4, s5, s4
+; GFX9-NEXT:    s_cmp_lg_u32 s4, s3
+; GFX9-NEXT:    s_cselect_b32 s3, 1, 0
+; GFX9-NEXT:    s_addk_i32 s1, 0xfc10
+; GFX9-NEXT:    s_lshl_b32 s4, s1, 12
+; GFX9-NEXT:    s_or_b32 s3, s5, s3
+; GFX9-NEXT:    s_or_b32 s4, s0, s4
+; GFX9-NEXT:    s_cmp_lt_i32 s1, 1
+; GFX9-NEXT:    s_cselect_b32 s3, s3, s4
+; GFX9-NEXT:    s_and_b32 s4, s3, 7
+; GFX9-NEXT:    s_cmp_gt_i32 s4, 5
 ; GFX9-NEXT:    s_cselect_b32 s5, 1, 0
-; GFX9-NEXT:    s_or_b32 s5, s5, s7
-; GFX9-NEXT:    s_lshr_b32 s4, s4, 2
-; GFX9-NEXT:    s_add_i32 s4, s4, s5
-; GFX9-NEXT:    s_cmp_lt_i32 s3, 31
-; GFX9-NEXT:    s_cselect_b32 s4, s4, 0x7c00
-; GFX9-NEXT:    s_cmp_lg_u32 s2, 0
-; GFX9-NEXT:    s_movk_i32 s2, 0x7e00
-; GFX9-NEXT:    s_cselect_b32 s2, s2, 0x7c00
-; GFX9-NEXT:    s_cmpk_eq_i32 s3, 0x40f
-; GFX9-NEXT:    s_cselect_b32 s2, s2, s4
-; GFX9-NEXT:    s_movk_i32 s3, 0x7fff
+; GFX9-NEXT:    s_cmp_eq_u32 s4, 3
+; GFX9-NEXT:    s_cselect_b32 s4, 1, 0
+; GFX9-NEXT:    s_or_b32 s4, s4, s5
+; GFX9-NEXT:    s_lshr_b32 s3, s3, 2
+; GFX9-NEXT:    s_add_i32 s3, s3, s4
+; GFX9-NEXT:    s_cmp_lt_i32 s1, 31
+; GFX9-NEXT:    s_cselect_b32 s3, s3, 0x7c00
+; GFX9-NEXT:    s_cmp_lg_u32 s0, 0
+; GFX9-NEXT:    s_movk_i32 s0, 0x7e00
+; GFX9-NEXT:    s_cselect_b32 s0, s0, 0x7c00
+; GFX9-NEXT:    s_cmpk_eq_i32 s1, 0x40f
+; GFX9-NEXT:    s_cselect_b32 s0, s0, s3
+; GFX9-NEXT:    s_movk_i32 s1, 0x7fff
+; GFX9-NEXT:    v_mov_b32_e32 v0, s0
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s2
-; GFX9-NEXT:    v_mov_b32_e32 v2, s6
-; GFX9-NEXT:    v_bfi_b32 v1, s3, v1, v2
-; GFX9-NEXT:    global_store_short v0, v1, s[0:1]
-; GFX9-NEXT:    s_endpgm
+; GFX9-NEXT:    v_bfi_b32 v0, s1, v0, v1
+; GFX9-NEXT:    v_readfirstlane_b32 s0, v0
+; GFX9-NEXT:    ; return to shader part epilog
 ;
-; GFX11-LABEL: s_copysign_out_f16_mag_f64_sign_f16:
-; GFX11:       ; %bb.0:
-; GFX11-NEXT:    s_clause 0x1
-; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
-; GFX11-NEXT:    s_load_b32 s4, s[4:5], 0x34
-; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-NEXT:    s_and_b32 s5, s3, 0x1ff
-; GFX11-NEXT:    s_lshr_b32 s6, s3, 8
-; GFX11-NEXT:    s_or_b32 s2, s5, s2
-; GFX11-NEXT:    s_and_b32 s5, s6, 0xffe
-; GFX11-NEXT:    s_cmp_lg_u32 s2, 0
-; GFX11-NEXT:    s_cselect_b32 s2, -1, 0
-; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s2
-; GFX11-NEXT:    s_bfe_u32 s2, s3, 0xb0014
-; GFX11-NEXT:    s_sub_i32 s3, 0x3f1, s2
-; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT:    v_med3_i32 v1, s3, 0, 13
-; GFX11-NEXT:    v_readfirstlane_b32 s3, v0
-; GFX11-NEXT:    v_mov_b32_e32 v0, s4
-; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(SALU_CYCLE_1)
-; GFX11-NEXT:    v_readfirstlane_b32 s6, v1
-; GFX11-NEXT:    s_or_b32 s3, s5, s3
-; GFX11-NEXT:    v_mov_b32_e32 v1, 0
-; GFX11-NEXT:    s_or_b32 s5, s3, 0x1000
-; GFX11-NEXT:    s_lshr_b32 s7, s5, s6
-; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
-; GFX11-NEXT:    s_lshl_b32 s6, s7, s6
-; GFX11-NEXT:    s_cmp_lg_u32 s6, s5
-; GFX11-NEXT:    s_cselect_b32 s5, 1, 0
-; GFX11-NEXT:    s_addk_i32 s2, 0xfc10
-; GFX11-NEXT:    s_or_b32 s5, s7, s5
-; GFX11-NEXT:    s_lshl_b32 s6, s2, 12
-; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-NEXT:    s_or_b32 s6, s3, s6
-; GFX11-NEXT:    s_cmp_lt_i32 s2, 1
-; GFX11-NEXT:    s_cselect_b32 s5, s5, s6
-; GFX11-NEXT:    s_and_b32 s6, s5, 7
-; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT:    s_cmp_gt_i32 s6, 5
-; GFX11-NEXT:    s_cselect_b32 s7, 1, 0
-; GFX11-NEXT:    s_cmp_eq_u32 s6, 3
-; GFX11-NEXT:    s_cselect_b32 s6, 1, 0
-; GFX11-NEXT:    s_lshr_b32 s5, s5, 2
-; GFX11-NEXT:    s_or_b32 s6, s6, s7
-; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT:    s_add_i32 s5, s5, s6
-; GFX11-NEXT:    s_cmp_lt_i32 s2, 31
-; GFX11-NEXT:    s_movk_i32 s6, 0x7e00
-; GFX11-NEXT:    s_cselect_b32 s5, s5, 0x7c00
-; GFX11-NEXT:    s_cmp_lg_u32 s3, 0
-; GFX11-NEXT:    s_cselect_b32 s3, s6, 0x7c00
-; GFX11-NEXT:    s_cmpk_eq_i32 s2, 0x40f
-; GFX11-NEXT:    s_cselect_b32 s2, s3, s5
-; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT:    v_bfi_b32 v0, 0x7fff, s2, v0
-; GFX11-NEXT:    global_store_b16 v1, v0, s[0:1]
-; GFX11-NEXT:    s_endpgm
+; GFX11-TRUE16-LABEL: s_copysign_out_f16_mag_f64_sign_f16:
+; GFX11-TRUE16:       ; %bb.0:
+; GFX11-TRUE16-NEXT:    s_and_b32 s3, s1, 0x1ff
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX11-TRUE16-NEXT:    s_or_b32 s0, s3, s0
+; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s0, 0
+; GFX11-TRUE16-NEXT:    s_cselect_b32 s0, -1, 0
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; GFX11-TRUE16-NEXT:    s_bfe_u32 s0, s1, 0xb0014
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s1, s1, 8
+; GFX11-TRUE16-NEXT:    s_sub_i32 s3, 0x3f1, s0
+; GFX11-TRUE16-NEXT:    s_and_b32 s1, s1, 0xffe
+; GFX11-TRUE16-NEXT:    v_med3_i32 v1, s3, 0, 13
+; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s3, v0
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s4, v1
+; GFX11-TRUE16-NEXT:    s_or_b32 s1, s1, s3
+; GFX11-TRUE16-NEXT:    s_or_b32 s3, s1, 0x1000
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s3, s4
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s5, s4
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_4) | instid1(SALU_CYCLE_1)
+; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s4, s3
+; GFX11-TRUE16-NEXT:    s_cselect_b32 s3, 1, 0
+; GFX11-TRUE16-NEXT:    s_addk_i32 s0, 0xfc10
+; GFX11-TRUE16-NEXT:    s_or_b32 s3, s5, s3
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s0, 12
+; GFX11-TRUE16-NEXT:    s_or_b32 s4, s1, s4
+; GFX11-TRUE16-NEXT:    s_cmp_lt_i32 s0, 1
+; GFX11-TRUE16-NEXT:    s_cselect_b32 s3, s3, s4
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX11-TRUE16-NEXT:    s_and_b32 s4, s3, 7
+; GFX11-TRUE16-NEXT:    s_cmp_gt_i32 s4, 5
+; GFX11-TRUE16-NEXT:    s_cselect_b32 s5, 1, 0
+; GFX11-TRUE16-NEXT:    s_cmp_eq_u32 s4, 3
+; GFX11-TRUE16-NEXT:    s_cselect_b32 s4, 1, 0
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s3, s3, 2
+; GFX11-TRUE16-NEXT:    s_or_b32 s4, s4, s5
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-TRUE16-NEXT:    s_add_i32 s3, s3, s4
+; GFX11-TRUE16-NEXT:    s_cmp_lt_i32 s0, 31
+; GFX11-TRUE16-NEXT:    s_movk_i32 s4, 0x7e00
+; GFX11-TRUE16-NEXT:    s_cselect_b32 s3, s3, 0x7c00
+; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s1, 0
+; GFX11-TRUE16-NEXT:    s_cselect_b32 s1, s4, 0x7c00
+; GFX11-TRUE16-NEXT:    s_cmpk_eq_i32 s0, 0x40f
+; GFX11-TRUE16-NEXT:    s_cselect_b32 s0, s1, s3
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-TRUE16-NEXT:    v_bfi_b32 v0, 0x7fff, v0, s2
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s0, v0
+; GFX11-TRUE16-NEXT:    ; return to shader part epilog
+;
+; GFX11-FAKE16-LABEL: s_copysign_out_f16_mag_f64_sign_f16:
+; GFX11-FAKE16:       ; %bb.0:
+; GFX11-FAKE16-NEXT:    s_and_b32 s3, s1, 0x1ff
+; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX11-FAKE16-NEXT:    s_or_b32 s0, s3, s0
+; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s0, 0
+; GFX11-FAKE16-NEXT:    s_cselect_b32 s0, -1, 0
+; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-FAKE16-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; GFX11-FAKE16-NEXT:    s_bfe_u32 s0, s1, 0xb0014
+; GFX11-FAKE16-NEXT:    s_lshr_b32 s1, s1, 8
+; GFX11-FAKE16-NEXT:    s_sub_i32 s3, 0x3f1, s0
+; GFX11-FAKE16-NEXT:    s_and_b32 s1, s1, 0xffe
+; GFX11-FAKE16-NEXT:    v_med3_i32 v1, s3, 0, 13
+; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s3, v0
+; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s4, v1
+; GFX11-FAKE16-NEXT:    s_or_b32 s1, s1, s3
+; GFX11-FAKE16-NEXT:    s_or_b32 s3, s1, 0x1000
+; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX11-FAKE16-NEXT:    s_lshr_b32 s5, s3, s4
+; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s5, s4
+; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_4) | instid1(SALU_CYCLE_1)
+; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s4, s3
+; GFX11-FAKE16-NEXT:    s_cselect_b32 s3, 1, 0
+; GFX11-FAKE16-NEXT:    s_addk_i32 s0, 0xfc10
+; GFX11-FAKE16-NEXT:    s_or_b32 s3, s5, s3
+; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s0, 12
+; GFX11-FAKE16-NEXT:    s_or_b32 s4, s1, s4
+; GFX11-FAKE16-NEXT:    s_cmp_lt_i32 s0, 1
+; GFX11-FAKE16-NEXT:    s_cselect_b32 s3, s3, s4
+; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX11-FAKE16-NEXT:    s_and_b32 s4, s3, 7
+; GFX11-FAKE16-NEXT:    s_cmp_gt_i32 s4, 5
+; GFX11-FAKE16-NEXT:    s_cselect_b32 s5, 1, 0
+; GFX11-FAKE16-NEXT:    s_cmp_eq_u32 s4, 3
+; GFX11-FAKE16-NEXT:    s_cselect_b32 s4, 1, 0
+; GFX11-FAKE16-NEXT:    s_lshr_b32 s3, s3, 2
+; GFX11-FAKE16-NEXT:    s_or_b32 s4, s4, s5
+; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-FAKE16-NEXT:    s_add_i32 s3, s3, s4
+; GFX11-FAKE16-NEXT:    s_cmp_lt_i32 s0, 31
+; GFX11-FAKE16-NEXT:    s_movk_i32 s4, 0x7e00
+; GFX11-FAKE16-NEXT:    s_cselect_b32 s3, s3, 0x7c00
+; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s1, 0
+; GFX11-FAKE16-NEXT:    s_cselect_b32 s1, s4, 0x7c00
+; GFX11-FAKE16-NEXT:    s_cmpk_eq_i32 s0, 0x40f
+; GFX11-FAKE16-NEXT:    s_cselect_b32 s0, s1, s3
+; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-FAKE16-NEXT:    v_bfi_b32 v0, 0x7fff, s0, v0
+; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s0, v0
+; GFX11-FAKE16-NEXT:    ; return to shader part epilog
   %mag.trunc = fptrunc double %mag to half
   %result = call half @llvm.copysign.f16(half %mag.trunc, half %sign)
-  store half %result, ptr addrspace(1) %arg_out
-  ret void
+  %cast = bitcast half %result to i16
+  ret i16 %cast
 }
 
-define amdgpu_kernel void @s_copysign_v2f16(ptr addrspace(1) %arg_out, <2 x half> %arg_mag, <2 x half> %arg_sign) {
+define amdgpu_ps i32 @s_copysign_v2f16(<2 x half> inreg %arg_mag, <2 x half> inreg %arg_sign) {
 ; SI-LABEL: s_copysign_v2f16:
 ; SI:       ; %bb.0:
-; SI-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
-; SI-NEXT:    s_mov_b32 s7, 0xf000
-; SI-NEXT:    s_mov_b32 s6, -1
-; SI-NEXT:    s_waitcnt lgkmcnt(0)
-; SI-NEXT:    s_lshr_b32 s4, s2, 16
-; SI-NEXT:    s_lshr_b32 s5, s3, 16
-; SI-NEXT:    v_cvt_f32_f16_e32 v0, s4
-; SI-NEXT:    v_cvt_f32_f16_e32 v1, s5
-; SI-NEXT:    v_cvt_f32_f16_e32 v2, s2
-; SI-NEXT:    v_cvt_f32_f16_e32 v3, s3
-; SI-NEXT:    s_brev_b32 s2, -2
-; SI-NEXT:    v_bfi_b32 v0, s2, v0, v1
+; SI-NEXT:    v_cvt_f16_f32_e32 v0, s3
+; SI-NEXT:    v_cvt_f16_f32_e32 v1, s1
+; SI-NEXT:    v_cvt_f16_f32_e32 v2, s2
+; SI-NEXT:    v_cvt_f16_f32_e32 v3, s0
+; SI-NEXT:    v_cvt_f32_f16_e32 v0, v0
+; SI-NEXT:    v_cvt_f32_f16_e32 v1, v1
+; SI-NEXT:    v_cvt_f32_f16_e32 v2, v2
+; SI-NEXT:    v_cvt_f32_f16_e32 v3, v3
+; SI-NEXT:    s_brev_b32 s0, -2
+; SI-NEXT:    v_bfi_b32 v0, s0, v1, v0
 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
-; SI-NEXT:    v_bfi_b32 v1, s2, v2, v3
+; SI-NEXT:    v_bfi_b32 v1, s0, v3, v2
 ; SI-NEXT:    v_cvt_f16_f32_e32 v1, v1
-; SI-NEXT:    s_mov_b32 s4, s0
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; SI-NEXT:    s_mov_b32 s5, s1
 ; SI-NEXT:    v_or_b32_e32 v0, v1, v0
-; SI-NEXT:    buffer_store_dword v0, off, s[4:7], 0
-; SI-NEXT:    s_endpgm
+; SI-NEXT:    v_readfirstlane_b32 s0, v0
+; SI-NEXT:    ; return to shader part epilog
 ;
 ; VI-LABEL: s_copysign_v2f16:
 ; VI:       ; %bb.0:
-; VI-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
-; VI-NEXT:    s_movk_i32 s4, 0x7fff
-; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    v_mov_b32_e32 v0, s2
-; VI-NEXT:    v_mov_b32_e32 v1, s3
-; VI-NEXT:    s_lshr_b32 s3, s3, 16
-; VI-NEXT:    s_lshr_b32 s2, s2, 16
-; VI-NEXT:    v_bfi_b32 v0, s4, v0, v1
-; VI-NEXT:    v_mov_b32_e32 v1, s2
-; VI-NEXT:    v_mov_b32_e32 v2, s3
-; VI-NEXT:    v_bfi_b32 v1, s4, v1, v2
-; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; VI-NEXT:    v_or_b32_sdwa v2, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; VI-NEXT:    s_movk_i32 s2, 0x7fff
 ; VI-NEXT:    v_mov_b32_e32 v0, s0
 ; VI-NEXT:    v_mov_b32_e32 v1, s1
-; VI-NEXT:    flat_store_dword v[0:1], v2
-; VI-NEXT:    s_endpgm
+; VI-NEXT:    s_lshr_b32 s1, s1, 16
+; VI-NEXT:    s_lshr_b32 s0, s0, 16
+; VI-NEXT:    v_bfi_b32 v0, s2, v0, v1
+; VI-NEXT:    v_mov_b32_e32 v1, s0
+; VI-NEXT:    v_mov_b32_e32 v2, s1
+; VI-NEXT:    v_bfi_b32 v1, s2, v1, v2
+; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
+; VI-NEXT:    v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; VI-NEXT:    v_readfirstlane_b32 s0, v0
+; VI-NEXT:    ; return to shader part epilog
 ;
 ; GFX9-LABEL: s_copysign_v2f16:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
-; GFX9-NEXT:    s_movk_i32 s4, 0x7fff
-; GFX9-NEXT:    v_mov_b32_e32 v0, 0
-; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    v_mov_b32_e32 v1, s2
-; GFX9-NEXT:    v_mov_b32_e32 v2, s3
-; GFX9-NEXT:    s_lshr_b32 s3, s3, 16
-; GFX9-NEXT:    s_lshr_b32 s2, s2, 16
-; GFX9-NEXT:    v_bfi_b32 v1, s4, v1, v2
-; GFX9-NEXT:    v_mov_b32_e32 v2, s2
-; GFX9-NEXT:    v_mov_b32_e32 v3, s3
-; GFX9-NEXT:    v_bfi_b32 v2, s4, v2, v3
-; GFX9-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
-; GFX9-NEXT:    global_store_dword v0, v1, s[0:1]
-; GFX9-NEXT:    s_endpgm
+; GFX9-NEXT:    s_movk_i32 s2, 0x7fff
+; GFX9-NEXT:    v_mov_b32_e32 v0, s0
+; GFX9-NEXT:    v_mov_b32_e32 v1, s1
+; GFX9-NEXT:    s_lshr_b32 s1, s1, 16
+; GFX9-NEXT:    s_lshr_b32 s0, s0, 16
+; GFX9-NEXT:    v_bfi_b32 v0, s2, v0, v1
+; GFX9-NEXT:    v_mov_b32_e32 v1, s0
+; GFX9-NEXT:    v_mov_b32_e32 v2, s1
+; GFX9-NEXT:    v_bfi_b32 v1, s2, v1, v2
+; GFX9-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX9-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; GFX9-NEXT:    v_readfirstlane_b32 s0, v0
+; GFX9-NEXT:    ; return to shader part epilog
 ;
 ; GFX11-LABEL: s_copysign_v2f16:
 ; GFX11:       ; %bb.0:
-; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
-; GFX11-NEXT:    v_mov_b32_e32 v2, 0
-; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-NEXT:    v_mov_b32_e32 v0, s3
-; GFX11-NEXT:    s_lshr_b32 s3, s3, 16
+; GFX11-NEXT:    v_mov_b32_e32 v0, s1
+; GFX11-NEXT:    s_lshr_b32 s1, s1, 16
 ; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-NEXT:    v_bfi_b32 v0, 0x7fff, s2, v0
-; GFX11-NEXT:    s_lshr_b32 s2, s2, 16
+; GFX11-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-NEXT:    v_bfi_b32 v0, 0x7fff, s0, v0
+; GFX11-NEXT:    s_lshr_b32 s0, s0, 16
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instid1(SALU_CYCLE_1)
-; GFX11-NEXT:    v_bfi_b32 v1, 0x7fff, s2, v1
+; GFX11-NEXT:    v_bfi_b32 v1, 0x7fff, s0, v1
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v0
 ; GFX11-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
-; GFX11-NEXT:    global_store_b32 v2, v0, s[0:1]
-; GFX11-NEXT:    s_endpgm
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_readfirstlane_b32 s0, v0
+; GFX11-NEXT:    ; return to shader part epilog
   %out = call <2 x half> @llvm.copysign.v2f16(<2 x half> %arg_mag, <2 x half> %arg_sign)
-  store <2 x half> %out, ptr addrspace(1) %arg_out
-  ret void
+  %cast = bitcast <2 x half> %out to i32
+  ret i32 %cast
 }
 
-define amdgpu_kernel void @s_copysign_v3f16(ptr addrspace(1) %arg_out, <3 x half> %arg_mag, <3 x half> %arg_sign) {
+define amdgpu_ps <3 x i16> @s_copysign_v3f16(<3 x half> inreg %arg_mag, <3 x half> inreg %arg_sign) {
 ; SI-LABEL: s_copysign_v3f16:
 ; SI:       ; %bb.0:
-; SI-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0xb
-; SI-NEXT:    s_load_dwordx2 s[4:5], s[4:5], 0x9
-; SI-NEXT:    s_mov_b32 s7, 0xf000
-; SI-NEXT:    s_waitcnt lgkmcnt(0)
-; SI-NEXT:    s_lshr_b32 s6, s0, 16
-; SI-NEXT:    v_cvt_f32_f16_e32 v1, s0
-; SI-NEXT:    s_lshr_b32 s0, s2, 16
-; SI-NEXT:    v_cvt_f32_f16_e32 v2, s6
-; SI-NEXT:    v_cvt_f32_f16_e32 v3, s0
-; SI-NEXT:    v_cvt_f32_f16_e32 v0, s1
-; SI-NEXT:    v_cvt_f32_f16_e32 v4, s3
-; SI-NEXT:    v_cvt_f32_f16_e32 v5, s2
+; SI-NEXT:    v_cvt_f16_f32_e32 v2, s4
+; SI-NEXT:    v_cvt_f16_f32_e32 v3, s1
+; SI-NEXT:    v_cvt_f16_f32_e32 v0, s5
+; SI-NEXT:    v_cvt_f16_f32_e32 v1, s2
+; SI-NEXT:    v_cvt_f16_f32_e32 v4, s3
+; SI-NEXT:    v_cvt_f16_f32_e32 v5, s0
+; SI-NEXT:    v_cvt_f32_f16_e32 v2, v2
+; SI-NEXT:    v_cvt_f32_f16_e32 v3, v3
+; SI-NEXT:    v_cvt_f32_f16_e32 v0, v0
+; SI-NEXT:    v_cvt_f32_f16_e32 v1, v1
+; SI-NEXT:    v_cvt_f32_f16_e32 v4, v4
+; SI-NEXT:    v_cvt_f32_f16_e32 v5, v5
 ; SI-NEXT:    s_brev_b32 s0, -2
-; SI-NEXT:    v_bfi_b32 v2, s0, v2, v3
+; SI-NEXT:    v_bfi_b32 v2, s0, v3, v2
 ; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
-; SI-NEXT:    v_bfi_b32 v1, s0, v1, v5
-; SI-NEXT:    v_bfi_b32 v0, s0, v0, v4
-; SI-NEXT:    v_cvt_f16_f32_e32 v1, v1
+; SI-NEXT:    v_bfi_b32 v3, s0, v5, v4
+; SI-NEXT:    v_bfi_b32 v0, s0, v1, v0
+; SI-NEXT:    v_cvt_f16_f32_e32 v3, v3
 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
-; SI-NEXT:    s_mov_b32 s6, -1
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; SI-NEXT:    v_or_b32_e32 v1, v1, v2
-; SI-NEXT:    buffer_store_short v0, off, s[4:7], 0 offset:4
-; SI-NEXT:    buffer_store_dword v1, off, s[4:7], 0
-; SI-NEXT:    s_endpgm
+; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v2
+; SI-NEXT:    v_or_b32_e32 v2, v3, v1
+; SI-NEXT:    v_alignbit_b32 v1, v0, v1, 16
+; SI-NEXT:    v_readfirstlane_b32 s1, v1
+; SI-NEXT:    v_readfirstlane_b32 s0, v2
+; SI-NEXT:    v_readfirstlane_b32 s2, v0
+; SI-NEXT:    ; return to shader part epilog
 ;
 ; VI-LABEL: s_copysign_v3f16:
 ; VI:       ; %bb.0:
-; VI-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x2c
-; VI-NEXT:    s_load_dwordx2 s[4:5], s[4:5], 0x24
-; VI-NEXT:    s_movk_i32 s6, 0x7fff
-; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    v_mov_b32_e32 v0, s0
-; VI-NEXT:    v_mov_b32_e32 v1, s2
-; VI-NEXT:    s_lshr_b32 s2, s2, 16
-; VI-NEXT:    s_lshr_b32 s0, s0, 16
-; VI-NEXT:    v_bfi_b32 v0, s6, v0, v1
-; VI-NEXT:    v_mov_b32_e32 v1, s0
-; VI-NEXT:    v_mov_b32_e32 v2, s2
-; VI-NEXT:    v_bfi_b32 v1, s6, v1, v2
-; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; VI-NEXT:    v_or_b32_sdwa v2, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; VI-NEXT:    s_movk_i32 s4, 0x7fff
 ; VI-NEXT:    v_mov_b32_e32 v0, s1
 ; VI-NEXT:    v_mov_b32_e32 v1, s3
-; VI-NEXT:    s_add_u32 s0, s4, 4
-; VI-NEXT:    v_bfi_b32 v3, s6, v0, v1
-; VI-NEXT:    s_addc_u32 s1, s5, 0
-; VI-NEXT:    v_mov_b32_e32 v0, s0
-; VI-NEXT:    v_mov_b32_e32 v1, s1
-; VI-NEXT:    flat_store_short v[0:1], v3
-; VI-NEXT:    v_mov_b32_e32 v0, s4
-; VI-NEXT:    v_mov_b32_e32 v1, s5
-; VI-NEXT:    flat_store_dword v[0:1], v2
-; VI-NEXT:    s_endpgm
+; VI-NEXT:    s_lshr_b32 s1, s2, 16
+; VI-NEXT:    s_lshr_b32 s3, s0, 16
+; VI-NEXT:    v_bfi_b32 v0, s4, v0, v1
+; VI-NEXT:    v_mov_b32_e32 v1, s3
+; VI-NEXT:    v_mov_b32_e32 v2, s1
+; VI-NEXT:    v_bfi_b32 v1, s4, v1, v2
+; VI-NEXT:    v_mov_b32_e32 v2, s0
+; VI-NEXT:    v_mov_b32_e32 v3, s2
+; VI-NEXT:    v_bfi_b32 v2, s4, v2, v3
+; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
+; VI-NEXT:    v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; VI-NEXT:    v_readfirstlane_b32 s0, v1
+; VI-NEXT:    v_readfirstlane_b32 s1, v0
+; VI-NEXT:    ; return to shader part epilog
 ;
 ; GFX9-LABEL: s_copysign_v3f16:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x2c
-; GFX9-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x24
 ; GFX9-NEXT:    s_movk_i32 s4, 0x7fff
-; GFX9-NEXT:    v_mov_b32_e32 v0, 0
-; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s1
+; GFX9-NEXT:    v_mov_b32_e32 v1, s3
+; GFX9-NEXT:    v_bfi_b32 v0, s4, v0, v1
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s0
 ; GFX9-NEXT:    v_mov_b32_e32 v2, s2
-; GFX9-NEXT:    s_lshr_b32 s2, s2, 16
+; GFX9-NEXT:    s_lshr_b32 s1, s2, 16
 ; GFX9-NEXT:    s_lshr_b32 s0, s0, 16
 ; GFX9-NEXT:    v_bfi_b32 v1, s4, v1, v2
 ; GFX9-NEXT:    v_mov_b32_e32 v2, s0
-; GFX9-NEXT:    v_mov_b32_e32 v3, s2
+; GFX9-NEXT:    v_mov_b32_e32 v3, s1
 ; GFX9-NEXT:    v_bfi_b32 v2, s4, v2, v3
 ; GFX9-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
-; GFX9-NEXT:    v_mov_b32_e32 v2, s1
-; GFX9-NEXT:    v_mov_b32_e32 v3, s3
-; GFX9-NEXT:    v_bfi_b32 v2, s4, v2, v3
-; GFX9-NEXT:    global_store_short v0, v2, s[6:7] offset:4
-; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
-; GFX9-NEXT:    s_endpgm
+; GFX9-NEXT:    v_readfirstlane_b32 s0, v1
+; GFX9-NEXT:    v_readfirstlane_b32 s1, v0
+; GFX9-NEXT:    ; return to shader part epilog
 ;
-; GFX11-TRUE16-LABEL: s_copysign_v3f16:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_clause 0x1
-; GFX11-TRUE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x2c
-; GFX11-TRUE16-NEXT:    s_load_b64 s[4:5], s[4:5], 0x24
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v4, 0
-; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s2, s2, 16
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v1, s2
-; GFX11-TRUE16-NEXT:    v_bfi_b32 v0, 0x7fff, s0, v0
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v2, s3
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s0, s0, 16
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_bfi_b32 v1, 0x7fff, s0, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_bfi_b32 v2, 0x7fff, s1, v2
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v1, v1, 16, v3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v2.l
-; GFX11-TRUE16-NEXT:    s_clause 0x1
-; GFX11-TRUE16-NEXT:    global_store_b16 v4, v0, s[4:5] offset:4
-; GFX11-TRUE16-NEXT:    global_store_b32 v4, v1, s[4:5]
-; GFX11-TRUE16-NEXT:    s_endpgm
-;
-; GFX11-FAKE16-LABEL: s_copysign_v3f16:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_clause 0x1
-; GFX11-FAKE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x2c
-; GFX11-FAKE16-NEXT:    s_load_b64 s[4:5], s[4:5], 0x24
-; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v3, 0 :: v_dual_mov_b32 v0, s2
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s2, s2, 16
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s3 :: v_dual_mov_b32 v1, s2
-; GFX11-FAKE16-NEXT:    v_bfi_b32 v0, 0x7fff, s0, v0
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s0, s0, 16
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_bfi_b32 v2, 0x7fff, s1, v2
-; GFX11-FAKE16-NEXT:    v_bfi_b32 v1, 0x7fff, s0, v1
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
-; GFX11-FAKE16-NEXT:    s_clause 0x1
-; GFX11-FAKE16-NEXT:    global_store_b16 v3, v2, s[4:5] offset:4
-; GFX11-FAKE16-NEXT:    global_store_b32 v3, v0, s[4:5]
-; GFX11-FAKE16-NEXT:    s_endpgm
+; GFX11-LABEL: s_copysign_v3f16:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-NEXT:    s_lshr_b32 s2, s2, 16
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT:    v_dual_mov_b32 v2, s3 :: v_dual_mov_b32 v1, s2
+; GFX11-NEXT:    v_bfi_b32 v0, 0x7fff, s0, v0
+; GFX11-NEXT:    s_lshr_b32 s0, s0, 16
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_bfi_b32 v1, 0x7fff, s0, v1
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; GFX11-NEXT:    v_bfi_b32 v1, 0x7fff, s1, v2
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT:    v_readfirstlane_b32 s0, v0
+; GFX11-NEXT:    v_readfirstlane_b32 s1, v1
+; GFX11-NEXT:    ; return to shader part epilog
   %out = call <3 x half> @llvm.copysign.v3f16(<3 x half> %arg_mag, <3 x half> %arg_sign)
-  store <3 x half> %out, ptr addrspace(1) %arg_out
-  ret void
+  %cast = bitcast <3 x half> %out to <3 x i16>
+  ret <3 x i16> %cast
 }
 
-define amdgpu_kernel void @s_copysign_v4f16(ptr addrspace(1) %arg_out, <4 x half> %arg_mag, <4 x half> %arg_sign) {
+define amdgpu_ps <2 x i32> @s_copysign_v4f16(<4 x half> inreg %arg_mag, <4 x half> inreg %arg_sign) {
 ; SI-LABEL: s_copysign_v4f16:
 ; SI:       ; %bb.0:
-; SI-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0xb
-; SI-NEXT:    s_load_dwordx2 s[4:5], s[4:5], 0x9
-; SI-NEXT:    s_mov_b32 s7, 0xf000
-; SI-NEXT:    s_mov_b32 s6, -1
-; SI-NEXT:    s_waitcnt lgkmcnt(0)
-; SI-NEXT:    s_lshr_b32 s8, s0, 16
-; SI-NEXT:    v_cvt_f32_f16_e32 v2, s0
-; SI-NEXT:    s_lshr_b32 s0, s2, 16
-; SI-NEXT:    s_lshr_b32 s9, s1, 16
-; SI-NEXT:    v_cvt_f32_f16_e32 v4, s0
-; SI-NEXT:    s_lshr_b32 s0, s3, 16
-; SI-NEXT:    v_cvt_f32_f16_e32 v0, s8
-; SI-NEXT:    v_cvt_f32_f16_e32 v1, s9
-; SI-NEXT:    v_cvt_f32_f16_e32 v5, s0
-; SI-NEXT:    v_cvt_f32_f16_e32 v3, s1
-; SI-NEXT:    v_cvt_f32_f16_e32 v6, s2
-; SI-NEXT:    v_cvt_f32_f16_e32 v7, s3
+; SI-NEXT:    v_cvt_f16_f32_e32 v2, s5
+; SI-NEXT:    v_cvt_f16_f32_e32 v3, s1
+; SI-NEXT:    v_cvt_f16_f32_e32 v4, s7
+; SI-NEXT:    v_cvt_f16_f32_e32 v5, s3
+; SI-NEXT:    v_cvt_f16_f32_e32 v0, s4
+; SI-NEXT:    v_cvt_f16_f32_e32 v1, s0
+; SI-NEXT:    v_cvt_f16_f32_e32 v6, s6
+; SI-NEXT:    v_cvt_f16_f32_e32 v7, s2
+; SI-NEXT:    v_cvt_f32_f16_e32 v2, v2
+; SI-NEXT:    v_cvt_f32_f16_e32 v3, v3
+; SI-NEXT:    v_cvt_f32_f16_e32 v4, v4
+; SI-NEXT:    v_cvt_f32_f16_e32 v5, v5
+; SI-NEXT:    v_cvt_f32_f16_e32 v0, v0
+; SI-NEXT:    v_cvt_f32_f16_e32 v1, v1
+; SI-NEXT:    v_cvt_f32_f16_e32 v6, v6
+; SI-NEXT:    v_cvt_f32_f16_e32 v7, v7
 ; SI-NEXT:    s_brev_b32 s0, -2
-; SI-NEXT:    v_bfi_b32 v1, s0, v1, v5
-; SI-NEXT:    v_bfi_b32 v0, s0, v0, v4
-; SI-NEXT:    v_cvt_f16_f32_e32 v1, v1
-; SI-NEXT:    v_bfi_b32 v3, s0, v3, v7
-; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
-; SI-NEXT:    v_bfi_b32 v2, s0, v2, v6
-; SI-NEXT:    v_cvt_f16_f32_e32 v3, v3
+; SI-NEXT:    v_bfi_b32 v4, s0, v5, v4
+; SI-NEXT:    v_bfi_b32 v2, s0, v3, v2
+; SI-NEXT:    v_cvt_f16_f32_e32 v4, v4
+; SI-NEXT:    v_bfi_b32 v5, s0, v7, v6
 ; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; SI-NEXT:    v_or_b32_e32 v1, v3, v1
-; SI-NEXT:    v_or_b32_e32 v0, v2, v0
-; SI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[4:7], 0
-; SI-NEXT:    s_endpgm
+; SI-NEXT:    v_bfi_b32 v0, s0, v1, v0
+; SI-NEXT:    v_cvt_f16_f32_e32 v5, v5
+; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
+; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v4
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
+; SI-NEXT:    v_or_b32_e32 v1, v5, v1
+; SI-NEXT:    v_or_b32_e32 v0, v0, v2
+; SI-NEXT:    v_readfirstlane_b32 s0, v0
+; SI-NEXT:    v_readfirstlane_b32 s1, v1
+; SI-NEXT:    ; return to shader part epilog
 ;
 ; VI-LABEL: s_copysign_v4f16:
 ; VI:       ; %bb.0:
-; VI-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x2c
-; VI-NEXT:    s_load_dwordx2 s[4:5], s[4:5], 0x24
-; VI-NEXT:    s_movk_i32 s6, 0x7fff
-; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_movk_i32 s4, 0x7fff
 ; VI-NEXT:    v_mov_b32_e32 v0, s1
 ; VI-NEXT:    v_mov_b32_e32 v1, s3
 ; VI-NEXT:    s_lshr_b32 s3, s3, 16
 ; VI-NEXT:    s_lshr_b32 s1, s1, 16
-; VI-NEXT:    v_bfi_b32 v0, s6, v0, v1
+; VI-NEXT:    v_bfi_b32 v0, s4, v0, v1
 ; VI-NEXT:    v_mov_b32_e32 v1, s1
 ; VI-NEXT:    v_mov_b32_e32 v2, s3
-; VI-NEXT:    v_bfi_b32 v1, s6, v1, v2
+; VI-NEXT:    v_bfi_b32 v1, s4, v1, v2
 ; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; VI-NEXT:    v_or_b32_sdwa v1, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_mov_b32_e32 v0, s0
+; VI-NEXT:    v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; VI-NEXT:    v_mov_b32_e32 v1, s0
 ; VI-NEXT:    v_mov_b32_e32 v2, s2
 ; VI-NEXT:    s_lshr_b32 s1, s2, 16
 ; VI-NEXT:    s_lshr_b32 s0, s0, 16
-; VI-NEXT:    v_bfi_b32 v0, s6, v0, v2
+; VI-NEXT:    v_bfi_b32 v1, s4, v1, v2
 ; VI-NEXT:    v_mov_b32_e32 v2, s0
 ; VI-NEXT:    v_mov_b32_e32 v3, s1
-; VI-NEXT:    v_bfi_b32 v2, s6, v2, v3
+; VI-NEXT:    v_bfi_b32 v2, s4, v2, v3
 ; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; VI-NEXT:    v_or_b32_sdwa v0, v0, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_mov_b32_e32 v2, s4
-; VI-NEXT:    v_mov_b32_e32 v3, s5
-; VI-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
-; VI-NEXT:    s_endpgm
+; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; VI-NEXT:    v_readfirstlane_b32 s0, v1
+; VI-NEXT:    v_readfirstlane_b32 s1, v0
+; VI-NEXT:    ; return to shader part epilog
 ;
 ; GFX9-LABEL: s_copysign_v4f16:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x2c
-; GFX9-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x24
 ; GFX9-NEXT:    s_movk_i32 s4, 0x7fff
-; GFX9-NEXT:    v_mov_b32_e32 v2, 0
-; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    v_mov_b32_e32 v0, s1
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s3
 ; GFX9-NEXT:    s_lshr_b32 s3, s3, 16
 ; GFX9-NEXT:    s_lshr_b32 s1, s1, 16
 ; GFX9-NEXT:    v_bfi_b32 v0, s4, v0, v1
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s1
-; GFX9-NEXT:    v_mov_b32_e32 v3, s3
-; GFX9-NEXT:    v_bfi_b32 v1, s4, v1, v3
+; GFX9-NEXT:    v_mov_b32_e32 v2, s3
+; GFX9-NEXT:    v_bfi_b32 v1, s4, v1, v2
 ; GFX9-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v0
-; GFX9-NEXT:    v_mov_b32_e32 v0, s0
-; GFX9-NEXT:    v_mov_b32_e32 v3, s2
+; GFX9-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; GFX9-NEXT:    v_mov_b32_e32 v1, s0
+; GFX9-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX9-NEXT:    s_lshr_b32 s1, s2, 16
 ; GFX9-NEXT:    s_lshr_b32 s0, s0, 16
-; GFX9-NEXT:    v_bfi_b32 v0, s4, v0, v3
-; GFX9-NEXT:    v_mov_b32_e32 v3, s0
-; GFX9-NEXT:    v_mov_b32_e32 v4, s1
-; GFX9-NEXT:    v_bfi_b32 v3, s4, v3, v4
-; GFX9-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX9-NEXT:    v_lshl_or_b32 v0, v3, 16, v0
-; GFX9-NEXT:    global_store_dwordx2 v2, v[0:1], s[6:7]
-; GFX9-NEXT:    s_endpgm
+; GFX9-NEXT:    v_bfi_b32 v1, s4, v1, v2
+; GFX9-NEXT:    v_mov_b32_e32 v2, s0
+; GFX9-NEXT:    v_mov_b32_e32 v3, s1
+; GFX9-NEXT:    v_bfi_b32 v2, s4, v2, v3
+; GFX9-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
+; GFX9-NEXT:    v_readfirstlane_b32 s0, v1
+; GFX9-NEXT:    v_readfirstlane_b32 s1, v0
+; GFX9-NEXT:    ; return to shader part epilog
 ;
 ; GFX11-LABEL: s_copysign_v4f16:
 ; GFX11:       ; %bb.0:
-; GFX11-NEXT:    s_clause 0x1
-; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x2c
-; GFX11-NEXT:    s_load_b64 s[4:5], s[4:5], 0x24
-; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-NEXT:    v_dual_mov_b32 v5, 0 :: v_dual_mov_b32 v0, s3
-; GFX11-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-NEXT:    v_dual_mov_b32 v0, s3 :: v_dual_mov_b32 v1, s2
 ; GFX11-NEXT:    s_lshr_b32 s3, s3, 16
 ; GFX11-NEXT:    s_lshr_b32 s2, s2, 16
-; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_2)
 ; GFX11-NEXT:    v_dual_mov_b32 v2, s3 :: v_dual_mov_b32 v3, s2
 ; GFX11-NEXT:    v_bfi_b32 v0, 0x7fff, s1, v0
 ; GFX11-NEXT:    v_bfi_b32 v1, 0x7fff, s0, v1
-; GFX11-NEXT:    s_lshr_b32 s6, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s1, s1, 16
 ; GFX11-NEXT:    s_lshr_b32 s0, s0, 16
-; GFX11-NEXT:    v_bfi_b32 v2, 0x7fff, s6, v2
+; GFX11-NEXT:    v_bfi_b32 v2, 0x7fff, s1, v2
 ; GFX11-NEXT:    v_bfi_b32 v3, 0x7fff, s0, v3
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-NEXT:    v_and_b32_e32 v4, 0xffff, v1
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT:    v_lshl_or_b32 v1, v2, 16, v0
-; GFX11-NEXT:    v_lshl_or_b32 v0, v3, 16, v4
-; GFX11-NEXT:    global_store_b64 v5, v[0:1], s[4:5]
-; GFX11-NEXT:    s_endpgm
+; GFX11-NEXT:    v_lshl_or_b32 v1, v3, 16, v1
+; GFX11-NEXT:    v_lshl_or_b32 v0, v2, 16, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT:    v_readfirstlane_b32 s0, v1
+; GFX11-NEXT:    v_readfirstlane_b32 s1, v0
+; GFX11-NEXT:    ; return to shader part epilog
   %out = call <4 x half> @llvm.copysign.v4f16(<4 x half> %arg_mag, <4 x half> %arg_sign)
-  store <4 x half> %out, ptr addrspace(1) %arg_out
-  ret void
+  %cast = bitcast <4 x half> %out to <2 x i32>
+  ret <2 x i32> %cast
 }
 
 attributes #0 = { nocallback nofree nosync nounwind speculatable willreturn memory(none) }



More information about the llvm-commits mailing list