; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5 ; Verifies correctness of load/store of parameters and return values. ; RUN: llc < %s -mtriple=nvptx64 -mcpu=sm_35 -verify-machineinstrs | FileCheck -allow-deprecated-dag-overlap %s ; RUN: %if ptxas %{ llc < %s -mtriple=nvptx64 -mcpu=sm_35 -verify-machineinstrs | %ptxas-verify %} %s_i8i16p = type { <{ i16, i8, i16 }>, i64 } %s_i8i32p = type { <{ i32, i8, i32 }>, i64 } %s_i8i64p = type { <{ i64, i8, i64 }>, i64 } %s_i8f16p = type { <{ half, i8, half }>, i64 } %s_i8f16x2p = type { <{ <2 x half>, i8, <2 x half> }>, i64 } %s_i8f32p = type { <{ float, i8, float }>, i64 } %s_i8f64p = type { <{ double, i8, double }>, i64 } ; -- All loads/stores from parameters aligned by one must be done one ; byte at a time. ; -- Notes: ; -- There are two fields of interest in the packed part of the struct, one ; with a proper offset and one without. The former should be loaded or ; stored as a whole, and the latter by bytes. ; -- Only loading and storing the said fields are checked in the following ; series of tests so that they are more concise. define %s_i8i16p @test_s_i8i16p(%s_i8i16p %a) { ; CHECK-LABEL: test_s_i8i16p( ; CHECK: { ; CHECK-NEXT: .reg .b16 %rs<9>; ; CHECK-NEXT: .reg .b32 %r<2>; ; CHECK-NEXT: .reg .b64 %rd<3>; ; CHECK-EMPTY: ; CHECK-NEXT: // %bb.0: ; CHECK-NEXT: ld.param.b32 %r1, [test_s_i8i16p_param_0]; ; CHECK-NEXT: ld.param.b64 %rd1, [test_s_i8i16p_param_0+8]; ; CHECK-NEXT: ld.param.b8 %rs1, [test_s_i8i16p_param_0+4]; ; CHECK-NEXT: { // callseq 0, 0 ; CHECK-NEXT: .param .align 8 .b8 param0[16]; ; CHECK-NEXT: .param .align 8 .b8 retval0[16]; ; CHECK-NEXT: st.param.b8 [param0+4], %rs1; ; CHECK-NEXT: st.param.b64 [param0+8], %rd1; ; CHECK-NEXT: st.param.b32 [param0], %r1; ; CHECK-NEXT: call.uni (retval0), test_s_i8i16p, (param0); ; CHECK-NEXT: ld.param.b64 %rd2, [retval0+8]; ; CHECK-NEXT: ld.param.b8 %rs2, [retval0+2]; ; CHECK-NEXT: ld.param.b16 %rs3, [retval0]; ; CHECK-NEXT: ld.param.b8 %rs4, [retval0+4]; ; CHECK-NEXT: ld.param.b8 %rs5, [retval0+3]; ; CHECK-NEXT: } // callseq 0 ; CHECK-NEXT: shl.b16 %rs6, %rs4, 8; ; CHECK-NEXT: or.b16 %rs7, %rs6, %rs5; ; CHECK-NEXT: st.param.b8 [func_retval0+3], %rs5; ; CHECK-NEXT: st.param.b64 [func_retval0+8], %rd2; ; CHECK-NEXT: st.param.b8 [func_retval0+2], %rs2; ; CHECK-NEXT: st.param.b16 [func_retval0], %rs3; ; CHECK-NEXT: shr.u16 %rs8, %rs7, 8; ; CHECK-NEXT: st.param.b8 [func_retval0+4], %rs8; ; CHECK-NEXT: ret; %r = tail call %s_i8i16p @test_s_i8i16p(%s_i8i16p %a) ret %s_i8i16p %r } define %s_i8i32p @test_s_i8i32p(%s_i8i32p %a) { ; CHECK-LABEL: test_s_i8i32p( ; CHECK: { ; CHECK-NEXT: .reg .b16 %rs<3>; ; CHECK-NEXT: .reg .b32 %r<19>; ; CHECK-NEXT: .reg .b64 %rd<3>; ; CHECK-EMPTY: ; CHECK-NEXT: // %bb.0: ; CHECK-NEXT: ld.param.b32 %r1, [test_s_i8i32p_param_0]; ; CHECK-NEXT: ld.param.b16 %rs1, [test_s_i8i32p_param_0+4]; ; CHECK-NEXT: ld.param.b64 %rd1, [test_s_i8i32p_param_0+16]; ; CHECK-NEXT: ld.param.b8 %r2, [test_s_i8i32p_param_0+6]; ; CHECK-NEXT: ld.param.b8 %r3, [test_s_i8i32p_param_0+7]; ; CHECK-NEXT: ld.param.b8 %r4, [test_s_i8i32p_param_0+8]; ; CHECK-NEXT: { // callseq 1, 0 ; CHECK-NEXT: .param .align 8 .b8 param0[24]; ; CHECK-NEXT: .param .align 8 .b8 retval0[24]; ; CHECK-NEXT: st.param.b8 [param0+8], %r4; ; CHECK-NEXT: st.param.b8 [param0+7], %r3; ; CHECK-NEXT: st.param.b8 [param0+6], %r2; ; CHECK-NEXT: st.param.b64 [param0+16], %rd1; ; CHECK-NEXT: st.param.b16 [param0+4], %rs1; ; CHECK-NEXT: st.param.b32 [param0], %r1; ; CHECK-NEXT: call.uni (retval0), test_s_i8i32p, (param0); ; CHECK-NEXT: ld.param.b64 %rd2, [retval0+16]; ; CHECK-NEXT: ld.param.b8 %rs2, [retval0+4]; ; CHECK-NEXT: ld.param.b32 %r5, [retval0]; ; CHECK-NEXT: ld.param.b8 %r6, [retval0+8]; ; CHECK-NEXT: ld.param.b8 %r7, [retval0+7]; ; CHECK-NEXT: ld.param.b8 %r8, [retval0+6]; ; CHECK-NEXT: ld.param.b8 %r9, [retval0+5]; ; CHECK-NEXT: } // callseq 1 ; CHECK-NEXT: shl.b32 %r10, %r8, 8; ; CHECK-NEXT: or.b32 %r11, %r10, %r9; ; CHECK-NEXT: shl.b32 %r12, %r7, 16; ; CHECK-NEXT: shl.b32 %r13, %r6, 24; ; CHECK-NEXT: or.b32 %r14, %r13, %r12; ; CHECK-NEXT: or.b32 %r15, %r14, %r11; ; CHECK-NEXT: st.param.b8 [func_retval0+5], %r9; ; CHECK-NEXT: st.param.b64 [func_retval0+16], %rd2; ; CHECK-NEXT: st.param.b8 [func_retval0+4], %rs2; ; CHECK-NEXT: st.param.b32 [func_retval0], %r5; ; CHECK-NEXT: shr.u32 %r16, %r15, 24; ; CHECK-NEXT: st.param.b8 [func_retval0+8], %r16; ; CHECK-NEXT: shr.u32 %r17, %r15, 16; ; CHECK-NEXT: st.param.b8 [func_retval0+7], %r17; ; CHECK-NEXT: shr.u32 %r18, %r15, 8; ; CHECK-NEXT: st.param.b8 [func_retval0+6], %r18; ; CHECK-NEXT: ret; %r = tail call %s_i8i32p @test_s_i8i32p(%s_i8i32p %a) ret %s_i8i32p %r } define %s_i8i64p @test_s_i8i64p(%s_i8i64p %a) { ; CHECK-LABEL: test_s_i8i64p( ; CHECK: { ; CHECK-NEXT: .reg .b16 %rs<2>; ; CHECK-NEXT: .reg .b64 %rd<36>; ; CHECK-EMPTY: ; CHECK-NEXT: // %bb.0: ; CHECK-NEXT: ld.param.b64 %rd1, [test_s_i8i64p_param_0]; ; CHECK-NEXT: ld.param.b64 %rd2, [test_s_i8i64p_param_0+8]; ; CHECK-NEXT: ld.param.b64 %rd3, [test_s_i8i64p_param_0+24]; ; CHECK-NEXT: ld.param.b8 %rd4, [test_s_i8i64p_param_0+16]; ; CHECK-NEXT: { // callseq 2, 0 ; CHECK-NEXT: .param .align 8 .b8 param0[32]; ; CHECK-NEXT: .param .align 8 .b8 retval0[32]; ; CHECK-NEXT: st.param.b8 [param0+16], %rd4; ; CHECK-NEXT: st.param.b64 [param0+24], %rd3; ; CHECK-NEXT: st.param.b64 [param0+8], %rd2; ; CHECK-NEXT: st.param.b64 [param0], %rd1; ; CHECK-NEXT: call.uni (retval0), test_s_i8i64p, (param0); ; CHECK-NEXT: ld.param.b64 %rd5, [retval0+24]; ; CHECK-NEXT: ld.param.b8 %rs1, [retval0+8]; ; CHECK-NEXT: ld.param.b64 %rd6, [retval0]; ; CHECK-NEXT: ld.param.b8 %rd7, [retval0+16]; ; CHECK-NEXT: ld.param.b8 %rd8, [retval0+15]; ; CHECK-NEXT: ld.param.b8 %rd9, [retval0+14]; ; CHECK-NEXT: ld.param.b8 %rd10, [retval0+13]; ; CHECK-NEXT: ld.param.b8 %rd11, [retval0+12]; ; CHECK-NEXT: ld.param.b8 %rd12, [retval0+11]; ; CHECK-NEXT: ld.param.b8 %rd13, [retval0+10]; ; CHECK-NEXT: ld.param.b8 %rd14, [retval0+9]; ; CHECK-NEXT: } // callseq 2 ; CHECK-NEXT: shl.b64 %rd15, %rd13, 8; ; CHECK-NEXT: or.b64 %rd16, %rd15, %rd14; ; CHECK-NEXT: shl.b64 %rd17, %rd12, 16; ; CHECK-NEXT: shl.b64 %rd18, %rd11, 24; ; CHECK-NEXT: or.b64 %rd19, %rd18, %rd17; ; CHECK-NEXT: or.b64 %rd20, %rd19, %rd16; ; CHECK-NEXT: shl.b64 %rd21, %rd9, 8; ; CHECK-NEXT: or.b64 %rd22, %rd21, %rd10; ; CHECK-NEXT: shl.b64 %rd23, %rd8, 16; ; CHECK-NEXT: shl.b64 %rd24, %rd7, 24; ; CHECK-NEXT: or.b64 %rd25, %rd24, %rd23; ; CHECK-NEXT: or.b64 %rd26, %rd25, %rd22; ; CHECK-NEXT: shl.b64 %rd27, %rd26, 32; ; CHECK-NEXT: or.b64 %rd28, %rd27, %rd20; ; CHECK-NEXT: st.param.b8 [func_retval0+9], %rd14; ; CHECK-NEXT: st.param.b64 [func_retval0+24], %rd5; ; CHECK-NEXT: st.param.b8 [func_retval0+8], %rs1; ; CHECK-NEXT: st.param.b64 [func_retval0], %rd6; ; CHECK-NEXT: shr.u64 %rd29, %rd28, 56; ; CHECK-NEXT: st.param.b8 [func_retval0+16], %rd29; ; CHECK-NEXT: shr.u64 %rd30, %rd28, 48; ; CHECK-NEXT: st.param.b8 [func_retval0+15], %rd30; ; CHECK-NEXT: shr.u64 %rd31, %rd28, 40; ; CHECK-NEXT: st.param.b8 [func_retval0+14], %rd31; ; CHECK-NEXT: shr.u64 %rd32, %rd28, 32; ; CHECK-NEXT: st.param.b8 [func_retval0+13], %rd32; ; CHECK-NEXT: shr.u64 %rd33, %rd28, 24; ; CHECK-NEXT: st.param.b8 [func_retval0+12], %rd33; ; CHECK-NEXT: shr.u64 %rd34, %rd28, 16; ; CHECK-NEXT: st.param.b8 [func_retval0+11], %rd34; ; CHECK-NEXT: shr.u64 %rd35, %rd28, 8; ; CHECK-NEXT: st.param.b8 [func_retval0+10], %rd35; ; CHECK-NEXT: ret; %r = tail call %s_i8i64p @test_s_i8i64p(%s_i8i64p %a) ret %s_i8i64p %r } define %s_i8f16p @test_s_i8f16p(%s_i8f16p %a) { ; CHECK-LABEL: test_s_i8f16p( ; CHECK: { ; CHECK-NEXT: .reg .b16 %rs<11>; ; CHECK-NEXT: .reg .b64 %rd<3>; ; CHECK-EMPTY: ; CHECK-NEXT: // %bb.0: ; CHECK-NEXT: ld.param.b16 %rs1, [test_s_i8f16p_param_0]; ; CHECK-NEXT: ld.param.b16 %rs2, [test_s_i8f16p_param_0+2]; ; CHECK-NEXT: ld.param.b64 %rd1, [test_s_i8f16p_param_0+8]; ; CHECK-NEXT: ld.param.b8 %rs3, [test_s_i8f16p_param_0+4]; ; CHECK-NEXT: { // callseq 3, 0 ; CHECK-NEXT: .param .align 8 .b8 param0[16]; ; CHECK-NEXT: .param .align 8 .b8 retval0[16]; ; CHECK-NEXT: st.param.b8 [param0+4], %rs3; ; CHECK-NEXT: st.param.b64 [param0+8], %rd1; ; CHECK-NEXT: st.param.b16 [param0+2], %rs2; ; CHECK-NEXT: st.param.b16 [param0], %rs1; ; CHECK-NEXT: call.uni (retval0), test_s_i8f16p, (param0); ; CHECK-NEXT: ld.param.b64 %rd2, [retval0+8]; ; CHECK-NEXT: ld.param.b8 %rs4, [retval0+2]; ; CHECK-NEXT: ld.param.b16 %rs5, [retval0]; ; CHECK-NEXT: ld.param.b8 %rs6, [retval0+4]; ; CHECK-NEXT: ld.param.b8 %rs7, [retval0+3]; ; CHECK-NEXT: } // callseq 3 ; CHECK-NEXT: shl.b16 %rs8, %rs6, 8; ; CHECK-NEXT: or.b16 %rs9, %rs8, %rs7; ; CHECK-NEXT: st.param.b8 [func_retval0+3], %rs7; ; CHECK-NEXT: st.param.b64 [func_retval0+8], %rd2; ; CHECK-NEXT: st.param.b8 [func_retval0+2], %rs4; ; CHECK-NEXT: st.param.b16 [func_retval0], %rs5; ; CHECK-NEXT: shr.u16 %rs10, %rs9, 8; ; CHECK-NEXT: st.param.b8 [func_retval0+4], %rs10; ; CHECK-NEXT: ret; %r = tail call %s_i8f16p @test_s_i8f16p(%s_i8f16p %a) ret %s_i8f16p %r } define %s_i8f16x2p @test_s_i8f16x2p(%s_i8f16x2p %a) { ; CHECK-LABEL: test_s_i8f16x2p( ; CHECK: { ; CHECK-NEXT: .reg .b16 %rs<3>; ; CHECK-NEXT: .reg .b32 %r<19>; ; CHECK-NEXT: .reg .b64 %rd<3>; ; CHECK-EMPTY: ; CHECK-NEXT: // %bb.0: ; CHECK-NEXT: ld.param.b32 %r1, [test_s_i8f16x2p_param_0]; ; CHECK-NEXT: ld.param.b16 %rs1, [test_s_i8f16x2p_param_0+4]; ; CHECK-NEXT: ld.param.b64 %rd1, [test_s_i8f16x2p_param_0+16]; ; CHECK-NEXT: ld.param.b8 %r2, [test_s_i8f16x2p_param_0+6]; ; CHECK-NEXT: ld.param.b8 %r3, [test_s_i8f16x2p_param_0+7]; ; CHECK-NEXT: ld.param.b8 %r4, [test_s_i8f16x2p_param_0+8]; ; CHECK-NEXT: { // callseq 4, 0 ; CHECK-NEXT: .param .align 8 .b8 param0[24]; ; CHECK-NEXT: .param .align 8 .b8 retval0[24]; ; CHECK-NEXT: st.param.b8 [param0+8], %r4; ; CHECK-NEXT: st.param.b8 [param0+7], %r3; ; CHECK-NEXT: st.param.b8 [param0+6], %r2; ; CHECK-NEXT: st.param.b64 [param0+16], %rd1; ; CHECK-NEXT: st.param.b16 [param0+4], %rs1; ; CHECK-NEXT: st.param.b32 [param0], %r1; ; CHECK-NEXT: call.uni (retval0), test_s_i8f16x2p, (param0); ; CHECK-NEXT: ld.param.b64 %rd2, [retval0+16]; ; CHECK-NEXT: ld.param.b8 %rs2, [retval0+4]; ; CHECK-NEXT: ld.param.b32 %r5, [retval0]; ; CHECK-NEXT: ld.param.b8 %r6, [retval0+8]; ; CHECK-NEXT: ld.param.b8 %r7, [retval0+7]; ; CHECK-NEXT: ld.param.b8 %r8, [retval0+6]; ; CHECK-NEXT: ld.param.b8 %r9, [retval0+5]; ; CHECK-NEXT: } // callseq 4 ; CHECK-NEXT: shl.b32 %r10, %r8, 8; ; CHECK-NEXT: or.b32 %r11, %r10, %r9; ; CHECK-NEXT: shl.b32 %r12, %r7, 16; ; CHECK-NEXT: shl.b32 %r13, %r6, 24; ; CHECK-NEXT: or.b32 %r14, %r13, %r12; ; CHECK-NEXT: or.b32 %r15, %r14, %r11; ; CHECK-NEXT: st.param.b8 [func_retval0+5], %r9; ; CHECK-NEXT: st.param.b64 [func_retval0+16], %rd2; ; CHECK-NEXT: st.param.b8 [func_retval0+4], %rs2; ; CHECK-NEXT: st.param.b32 [func_retval0], %r5; ; CHECK-NEXT: shr.u32 %r16, %r15, 24; ; CHECK-NEXT: st.param.b8 [func_retval0+8], %r16; ; CHECK-NEXT: shr.u32 %r17, %r15, 16; ; CHECK-NEXT: st.param.b8 [func_retval0+7], %r17; ; CHECK-NEXT: shr.u32 %r18, %r15, 8; ; CHECK-NEXT: st.param.b8 [func_retval0+6], %r18; ; CHECK-NEXT: ret; %r = tail call %s_i8f16x2p @test_s_i8f16x2p(%s_i8f16x2p %a) ret %s_i8f16x2p %r } define %s_i8f32p @test_s_i8f32p(%s_i8f32p %a) { ; CHECK-LABEL: test_s_i8f32p( ; CHECK: { ; CHECK-NEXT: .reg .b16 %rs<3>; ; CHECK-NEXT: .reg .b32 %r<19>; ; CHECK-NEXT: .reg .b64 %rd<3>; ; CHECK-EMPTY: ; CHECK-NEXT: // %bb.0: ; CHECK-NEXT: ld.param.b32 %r1, [test_s_i8f32p_param_0]; ; CHECK-NEXT: ld.param.b16 %rs1, [test_s_i8f32p_param_0+4]; ; CHECK-NEXT: ld.param.b64 %rd1, [test_s_i8f32p_param_0+16]; ; CHECK-NEXT: ld.param.b8 %r2, [test_s_i8f32p_param_0+6]; ; CHECK-NEXT: ld.param.b8 %r3, [test_s_i8f32p_param_0+7]; ; CHECK-NEXT: ld.param.b8 %r4, [test_s_i8f32p_param_0+8]; ; CHECK-NEXT: { // callseq 5, 0 ; CHECK-NEXT: .param .align 8 .b8 param0[24]; ; CHECK-NEXT: .param .align 8 .b8 retval0[24]; ; CHECK-NEXT: st.param.b8 [param0+8], %r4; ; CHECK-NEXT: st.param.b8 [param0+7], %r3; ; CHECK-NEXT: st.param.b8 [param0+6], %r2; ; CHECK-NEXT: st.param.b64 [param0+16], %rd1; ; CHECK-NEXT: st.param.b16 [param0+4], %rs1; ; CHECK-NEXT: st.param.b32 [param0], %r1; ; CHECK-NEXT: call.uni (retval0), test_s_i8f32p, (param0); ; CHECK-NEXT: ld.param.b64 %rd2, [retval0+16]; ; CHECK-NEXT: ld.param.b8 %rs2, [retval0+4]; ; CHECK-NEXT: ld.param.b32 %r5, [retval0]; ; CHECK-NEXT: ld.param.b8 %r6, [retval0+8]; ; CHECK-NEXT: ld.param.b8 %r7, [retval0+7]; ; CHECK-NEXT: ld.param.b8 %r8, [retval0+6]; ; CHECK-NEXT: ld.param.b8 %r9, [retval0+5]; ; CHECK-NEXT: } // callseq 5 ; CHECK-NEXT: shl.b32 %r10, %r8, 8; ; CHECK-NEXT: or.b32 %r11, %r10, %r9; ; CHECK-NEXT: shl.b32 %r12, %r7, 16; ; CHECK-NEXT: shl.b32 %r13, %r6, 24; ; CHECK-NEXT: or.b32 %r14, %r13, %r12; ; CHECK-NEXT: or.b32 %r15, %r14, %r11; ; CHECK-NEXT: st.param.b8 [func_retval0+5], %r9; ; CHECK-NEXT: st.param.b64 [func_retval0+16], %rd2; ; CHECK-NEXT: st.param.b8 [func_retval0+4], %rs2; ; CHECK-NEXT: st.param.b32 [func_retval0], %r5; ; CHECK-NEXT: shr.u32 %r16, %r15, 24; ; CHECK-NEXT: st.param.b8 [func_retval0+8], %r16; ; CHECK-NEXT: shr.u32 %r17, %r15, 16; ; CHECK-NEXT: st.param.b8 [func_retval0+7], %r17; ; CHECK-NEXT: shr.u32 %r18, %r15, 8; ; CHECK-NEXT: st.param.b8 [func_retval0+6], %r18; ; CHECK-NEXT: ret; %r = tail call %s_i8f32p @test_s_i8f32p(%s_i8f32p %a) ret %s_i8f32p %r } define %s_i8f64p @test_s_i8f64p(%s_i8f64p %a) { ; CHECK-LABEL: test_s_i8f64p( ; CHECK: { ; CHECK-NEXT: .reg .b16 %rs<2>; ; CHECK-NEXT: .reg .b64 %rd<36>; ; CHECK-EMPTY: ; CHECK-NEXT: // %bb.0: ; CHECK-NEXT: ld.param.b64 %rd1, [test_s_i8f64p_param_0]; ; CHECK-NEXT: ld.param.b64 %rd2, [test_s_i8f64p_param_0+8]; ; CHECK-NEXT: ld.param.b64 %rd3, [test_s_i8f64p_param_0+24]; ; CHECK-NEXT: ld.param.b8 %rd4, [test_s_i8f64p_param_0+16]; ; CHECK-NEXT: { // callseq 6, 0 ; CHECK-NEXT: .param .align 8 .b8 param0[32]; ; CHECK-NEXT: .param .align 8 .b8 retval0[32]; ; CHECK-NEXT: st.param.b8 [param0+16], %rd4; ; CHECK-NEXT: st.param.b64 [param0+24], %rd3; ; CHECK-NEXT: st.param.b64 [param0+8], %rd2; ; CHECK-NEXT: st.param.b64 [param0], %rd1; ; CHECK-NEXT: call.uni (retval0), test_s_i8f64p, (param0); ; CHECK-NEXT: ld.param.b64 %rd5, [retval0+24]; ; CHECK-NEXT: ld.param.b8 %rs1, [retval0+8]; ; CHECK-NEXT: ld.param.b64 %rd6, [retval0]; ; CHECK-NEXT: ld.param.b8 %rd7, [retval0+16]; ; CHECK-NEXT: ld.param.b8 %rd8, [retval0+15]; ; CHECK-NEXT: ld.param.b8 %rd9, [retval0+14]; ; CHECK-NEXT: ld.param.b8 %rd10, [retval0+13]; ; CHECK-NEXT: ld.param.b8 %rd11, [retval0+12]; ; CHECK-NEXT: ld.param.b8 %rd12, [retval0+11]; ; CHECK-NEXT: ld.param.b8 %rd13, [retval0+10]; ; CHECK-NEXT: ld.param.b8 %rd14, [retval0+9]; ; CHECK-NEXT: } // callseq 6 ; CHECK-NEXT: shl.b64 %rd15, %rd13, 8; ; CHECK-NEXT: or.b64 %rd16, %rd15, %rd14; ; CHECK-NEXT: shl.b64 %rd17, %rd12, 16; ; CHECK-NEXT: shl.b64 %rd18, %rd11, 24; ; CHECK-NEXT: or.b64 %rd19, %rd18, %rd17; ; CHECK-NEXT: or.b64 %rd20, %rd19, %rd16; ; CHECK-NEXT: shl.b64 %rd21, %rd9, 8; ; CHECK-NEXT: or.b64 %rd22, %rd21, %rd10; ; CHECK-NEXT: shl.b64 %rd23, %rd8, 16; ; CHECK-NEXT: shl.b64 %rd24, %rd7, 24; ; CHECK-NEXT: or.b64 %rd25, %rd24, %rd23; ; CHECK-NEXT: or.b64 %rd26, %rd25, %rd22; ; CHECK-NEXT: shl.b64 %rd27, %rd26, 32; ; CHECK-NEXT: or.b64 %rd28, %rd27, %rd20; ; CHECK-NEXT: st.param.b8 [func_retval0+9], %rd14; ; CHECK-NEXT: st.param.b64 [func_retval0+24], %rd5; ; CHECK-NEXT: st.param.b8 [func_retval0+8], %rs1; ; CHECK-NEXT: st.param.b64 [func_retval0], %rd6; ; CHECK-NEXT: shr.u64 %rd29, %rd28, 56; ; CHECK-NEXT: st.param.b8 [func_retval0+16], %rd29; ; CHECK-NEXT: shr.u64 %rd30, %rd28, 48; ; CHECK-NEXT: st.param.b8 [func_retval0+15], %rd30; ; CHECK-NEXT: shr.u64 %rd31, %rd28, 40; ; CHECK-NEXT: st.param.b8 [func_retval0+14], %rd31; ; CHECK-NEXT: shr.u64 %rd32, %rd28, 32; ; CHECK-NEXT: st.param.b8 [func_retval0+13], %rd32; ; CHECK-NEXT: shr.u64 %rd33, %rd28, 24; ; CHECK-NEXT: st.param.b8 [func_retval0+12], %rd33; ; CHECK-NEXT: shr.u64 %rd34, %rd28, 16; ; CHECK-NEXT: st.param.b8 [func_retval0+11], %rd34; ; CHECK-NEXT: shr.u64 %rd35, %rd28, 8; ; CHECK-NEXT: st.param.b8 [func_retval0+10], %rd35; ; CHECK-NEXT: ret; %r = tail call %s_i8f64p @test_s_i8f64p(%s_i8f64p %a) ret %s_i8f64p %r }