[llvm] [AMDGPU] Restore the default rounding mode before calls and returns (PR #222247)

[Matt Arsenault via llvm-commits]

================
@@ -0,0 +1,120 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 6
+; RUN: llc -mtriple=amdgcn -mcpu=gfx900 < %s | FileCheck %s
+
+; A mode SIModeRegister set for llvm.fptrunc.round is restored at a call or a
+; return. A mode the program asked for is not.
+
+declare void @sink(float)
+declare float @tail_target(float)
+
+define float @ret_restores_default(double %a) {
+; CHECK-LABEL: ret_restores_default:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    s_setreg_imm32_b32 hwreg(HW_REG_MODE, 2, 1), 1
+; CHECK-NEXT:    v_cvt_f32_f64_e32 v0, v[0:1]
+; CHECK-NEXT:    s_nop 0
+; CHECK-NEXT:    s_setreg_imm32_b32 hwreg(HW_REG_MODE, 2, 2), 0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %t = call float @llvm.fptrunc.round.f32.f64(double %a, metadata !"round.upward")
+  ret float %t
+}
+
+define float @tail_call_restores_default(double %a) {
+; CHECK-LABEL: tail_call_restores_default:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    s_getpc_b64 s[16:17]
+; CHECK-NEXT:    s_add_u32 s16, s16, tail_target at gotpcrel32@lo+4
+; CHECK-NEXT:    s_addc_u32 s17, s17, tail_target at gotpcrel32@hi+12
+; CHECK-NEXT:    s_load_dwordx2 s[16:17], s[16:17], 0x0
+; CHECK-NEXT:    s_setreg_imm32_b32 hwreg(HW_REG_MODE, 2, 1), 1
+; CHECK-NEXT:    v_cvt_f32_f64_e32 v0, v[0:1]
+; CHECK-NEXT:    s_waitcnt lgkmcnt(0)
+; CHECK-NEXT:    s_setreg_imm32_b32 hwreg(HW_REG_MODE, 2, 2), 0
+; CHECK-NEXT:    s_setpc_b64 s[16:17]
+  %t = call float @llvm.fptrunc.round.f32.f64(double %a, metadata !"round.upward")
+  %u = tail call float @tail_target(float %t)
+  ret float %u
+}
+
+define void @call_restores_default(double %a) {
+; CHECK-LABEL: call_restores_default:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    s_mov_b32 s16, s33
+; CHECK-NEXT:    s_mov_b32 s33, s32
+; CHECK-NEXT:    s_or_saveexec_b64 s[18:19], -1
+; CHECK-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; CHECK-NEXT:    s_mov_b64 exec, s[18:19]
+; CHECK-NEXT:    v_writelane_b32 v40, s16, 2
+; CHECK-NEXT:    v_writelane_b32 v40, s30, 0
+; CHECK-NEXT:    s_addk_i32 s32, 0x400
+; CHECK-NEXT:    v_writelane_b32 v40, s31, 1
+; CHECK-NEXT:    s_getpc_b64 s[16:17]
+; CHECK-NEXT:    s_add_u32 s16, s16, sink at gotpcrel32@lo+4
+; CHECK-NEXT:    s_addc_u32 s17, s17, sink at gotpcrel32@hi+12
+; CHECK-NEXT:    s_load_dwordx2 s[16:17], s[16:17], 0x0
+; CHECK-NEXT:    s_setreg_imm32_b32 hwreg(HW_REG_MODE, 2, 1), 1
+; CHECK-NEXT:    v_cvt_f32_f64_e32 v0, v[0:1]
+; CHECK-NEXT:    s_waitcnt lgkmcnt(0)
+; CHECK-NEXT:    s_setreg_imm32_b32 hwreg(HW_REG_MODE, 2, 2), 0
+; CHECK-NEXT:    s_swappc_b64 s[30:31], s[16:17]
+; CHECK-NEXT:    v_readlane_b32 s30, v40, 0
+; CHECK-NEXT:    v_readlane_b32 s31, v40, 1
+; CHECK-NEXT:    s_mov_b32 s32, s33
+; CHECK-NEXT:    v_readlane_b32 s4, v40, 2
+; CHECK-NEXT:    s_or_saveexec_b64 s[6:7], -1
+; CHECK-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; CHECK-NEXT:    s_mov_b64 exec, s[6:7]
+; CHECK-NEXT:    s_mov_b32 s33, s4
+; CHECK-NEXT:    s_waitcnt vmcnt(0)
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %t = call float @llvm.fptrunc.round.f32.f64(double %a, metadata !"round.upward")
+  call void @sink(float %t)
+  ret void
+}
+
+; s_endpgm ends the wave, so there is no caller to restore for.
+define amdgpu_kernel void @kernel_no_restore(ptr addrspace(1) %out, double %a) {
+; CHECK-LABEL: kernel_no_restore:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
+; CHECK-NEXT:    v_mov_b32_e32 v2, 0
+; CHECK-NEXT:    s_waitcnt lgkmcnt(0)
+; CHECK-NEXT:    v_mov_b32_e32 v0, s2
+; CHECK-NEXT:    v_mov_b32_e32 v1, s3
+; CHECK-NEXT:    s_setreg_imm32_b32 hwreg(HW_REG_MODE, 2, 1), 1
+; CHECK-NEXT:    v_cvt_f32_f64_e32 v0, v[0:1]
+; CHECK-NEXT:    global_store_dword v2, v0, s[0:1]
+; CHECK-NEXT:    s_endpgm
+  %t = call float @llvm.fptrunc.round.f32.f64(double %a, metadata !"round.upward")
+  store float %t, ptr addrspace(1) %out
+  ret void
+}
+
+; A return to the shader epilog is not a call boundary either.
----------------
arsenm wrote:

it's still a return to external code, should probably restore here 

https://github.com/llvm/llvm-project/pull/222247