diff --git a/lib/Target/PTX/PTXInstrInfo.td b/lib/Target/PTX/PTXInstrInfo.td index d497283b721..d240899c6ad 100644 --- a/lib/Target/PTX/PTXInstrInfo.td +++ b/lib/Target/PTX/PTXInstrInfo.td @@ -826,31 +826,35 @@ defm STs : PTX_ST_ALL<"st.shared", store_shared>; // TODO: Do something with st.param if/when it is needed. // Conversion to pred - +// PTX does not directly support converting to a predicate type, so we fake it +// by performing a greater-than test between the value and zero. This follows +// the C convention that any non-zero value is equivalent to 'true'. def CVT_pred_u16 - : InstPTX<(outs RegPred:$d), (ins RegI16:$a), "cvt.pred.u16\t$d, $a", + : InstPTX<(outs RegPred:$d), (ins RegI16:$a), "setp.gt.b16\t$d, $a, 0", [(set RegPred:$d, (trunc RegI16:$a))]>; def CVT_pred_u32 - : InstPTX<(outs RegPred:$d), (ins RegI32:$a), "cvt.pred.u32\t$d, $a", + : InstPTX<(outs RegPred:$d), (ins RegI32:$a), "setp.gt.b32\t$d, $a, 0", [(set RegPred:$d, (trunc RegI32:$a))]>; def CVT_pred_u64 - : InstPTX<(outs RegPred:$d), (ins RegI64:$a), "cvt.pred.u64\t$d, $a", + : InstPTX<(outs RegPred:$d), (ins RegI64:$a), "setp.gt.b64\t$d, $a, 0", [(set RegPred:$d, (trunc RegI64:$a))]>; def CVT_pred_f32 - : InstPTX<(outs RegPred:$d), (ins RegF32:$a), "cvt.rzi.pred.f32\t$d, $a", + : InstPTX<(outs RegPred:$d), (ins RegF32:$a), "setp.gt.b32\t$d, $a, 0", [(set RegPred:$d, (fp_to_uint RegF32:$a))]>; def CVT_pred_f64 - : InstPTX<(outs RegPred:$d), (ins RegF64:$a), "cvt.rzi.pred.f64\t$d, $a", + : InstPTX<(outs RegPred:$d), (ins RegF64:$a), "setp.gt.b64\t$d, $a, 0", [(set RegPred:$d, (fp_to_uint RegF64:$a))]>; // Conversion to u16 - +// PTX does not directly support converting a predicate to a value, so we +// use a select instruction to select either 0 or 1 (integer or fp) based +// on the truth value of the predicate. def CVT_u16_pred - : InstPTX<(outs RegI16:$d), (ins RegPred:$a), "cvt.u16.pred\t$d, $a", + : InstPTX<(outs RegI16:$d), (ins RegPred:$a), "selp.u16\t$d, 1, 0, $a", [(set RegI16:$d, (zext RegPred:$a))]>; def CVT_u16_u32 @@ -872,7 +876,7 @@ def CVT_u16_f64 // Conversion to u32 def CVT_u32_pred - : InstPTX<(outs RegI32:$d), (ins RegPred:$a), "cvt.u32.pred\t$d, $a", + : InstPTX<(outs RegI32:$d), (ins RegPred:$a), "selp.u32\t$d, 1, 0, $a", [(set RegI32:$d, (zext RegPred:$a))]>; def CVT_u32_u16 @@ -894,7 +898,7 @@ def CVT_u32_f64 // Conversion to u64 def CVT_u64_pred - : InstPTX<(outs RegI64:$d), (ins RegPred:$a), "cvt.u64.pred\t$d, $a", + : InstPTX<(outs RegI64:$d), (ins RegPred:$a), "selp.u64\t$d, 1, 0, $a", [(set RegI64:$d, (zext RegPred:$a))]>; def CVT_u64_u16 @@ -916,7 +920,8 @@ def CVT_u64_f64 // Conversion to f32 def CVT_f32_pred - : InstPTX<(outs RegF32:$d), (ins RegPred:$a), "cvt.rn.f32.pred\t$d, $a", + : InstPTX<(outs RegF32:$d), (ins RegPred:$a), + "selp.f32\t$d, 0F3F800000, 0F00000000, $a", // 1.0 [(set RegF32:$d, (uint_to_fp RegPred:$a))]>; def CVT_f32_u16 @@ -938,7 +943,8 @@ def CVT_f32_f64 // Conversion to f64 def CVT_f64_pred - : InstPTX<(outs RegF64:$d), (ins RegPred:$a), "cvt.rn.f64.pred\t$d, $a", + : InstPTX<(outs RegF64:$d), (ins RegPred:$a), + "selp.f64\t$d, 0D3F80000000000000, 0D0000000000000000, $a", // 1.0 [(set RegF64:$d, (uint_to_fp RegPred:$a))]>; def CVT_f64_u16 diff --git a/test/CodeGen/PTX/cvt.ll b/test/CodeGen/PTX/cvt.ll index bf18bd79211..f7233697c04 100644 --- a/test/CodeGen/PTX/cvt.ll +++ b/test/CodeGen/PTX/cvt.ll @@ -4,8 +4,10 @@ ; (note: we convert back to i32 to return) define ptx_device i32 @cvt_pred_i16(i16 %x, i1 %y) { -; CHECK: cvt.pred.u16 p0, rh1; -; CHECK: ret; +; CHECK: setp.gt.b16 p0, rh1, 0 +; CHECK-NEXT: and.pred p0, p0, p1; +; CHECK-NEXT: selp.u32 r0, 1, 0, p0; +; CHECK-NEXT: ret; %a = trunc i16 %x to i1 %b = and i1 %a, %y %c = zext i1 %b to i32 @@ -13,8 +15,10 @@ define ptx_device i32 @cvt_pred_i16(i16 %x, i1 %y) { } define ptx_device i32 @cvt_pred_i32(i32 %x, i1 %y) { -; CHECK: cvt.pred.u32 p0, r1; -; CHECK: ret; +; CHECK: setp.gt.b32 p0, r1, 0 +; CHECK-NEXT: and.pred p0, p0, p1; +; CHECK-NEXT: selp.u32 r0, 1, 0, p0; +; CHECK-NEXT: ret; %a = trunc i32 %x to i1 %b = and i1 %a, %y %c = zext i1 %b to i32 @@ -22,8 +26,10 @@ define ptx_device i32 @cvt_pred_i32(i32 %x, i1 %y) { } define ptx_device i32 @cvt_pred_i64(i64 %x, i1 %y) { -; CHECK: cvt.pred.u64 p0, rd1; -; CHECK: ret; +; CHECK: setp.gt.b64 p0, rd1, 0 +; CHECK-NEXT: and.pred p0, p0, p1; +; CHECK-NEXT: selp.u32 r0, 1, 0, p0; +; CHECK-NEXT: ret; %a = trunc i64 %x to i1 %b = and i1 %a, %y %c = zext i1 %b to i32 @@ -31,8 +37,10 @@ define ptx_device i32 @cvt_pred_i64(i64 %x, i1 %y) { } define ptx_device i32 @cvt_pred_f32(float %x, i1 %y) { -; CHECK: cvt.rzi.pred.f32 p0, r1; -; CHECK: ret; +; CHECK: setp.gt.b32 p0, r1, 0 +; CHECK-NEXT: and.pred p0, p0, p1; +; CHECK-NEXT: selp.u32 r0, 1, 0, p0; +; CHECK-NEXT: ret; %a = fptoui float %x to i1 %b = and i1 %a, %y %c = zext i1 %b to i32 @@ -40,8 +48,10 @@ define ptx_device i32 @cvt_pred_f32(float %x, i1 %y) { } define ptx_device i32 @cvt_pred_f64(double %x, i1 %y) { -; CHECK: cvt.rzi.pred.f64 p0, rd1; -; CHECK: ret; +; CHECK: setp.gt.b64 p0, rd1, 0 +; CHECK-NEXT: and.pred p0, p0, p1; +; CHECK-NEXT: selp.u32 r0, 1, 0, p0; +; CHECK-NEXT: ret; %a = fptoui double %x to i1 %b = and i1 %a, %y %c = zext i1 %b to i32 @@ -51,36 +61,36 @@ define ptx_device i32 @cvt_pred_f64(double %x, i1 %y) { ; i16 define ptx_device i16 @cvt_i16_preds(i1 %x) { -; CHECK: cvt.u16.pred rh0, p1; -; CHECK: ret; +; CHECK: selp.u16 rh0, 1, 0, p1; +; CHECK-NEXT: ret; %a = zext i1 %x to i16 ret i16 %a } define ptx_device i16 @cvt_i16_i32(i32 %x) { ; CHECK: cvt.u16.u32 rh0, r1; -; CHECK: ret; +; CHECK-NEXT: ret; %a = trunc i32 %x to i16 ret i16 %a } define ptx_device i16 @cvt_i16_i64(i64 %x) { ; CHECK: cvt.u16.u64 rh0, rd1; -; CHECK: ret; +; CHECK-NEXT: ret; %a = trunc i64 %x to i16 ret i16 %a } define ptx_device i16 @cvt_i16_f32(float %x) { ; CHECK: cvt.rzi.u16.f32 rh0, r1; -; CHECK: ret; +; CHECK-NEXT: ret; %a = fptoui float %x to i16 ret i16 %a } define ptx_device i16 @cvt_i16_f64(double %x) { ; CHECK: cvt.rzi.u16.f64 rh0, rd1; -; CHECK: ret; +; CHECK-NEXT: ret; %a = fptoui double %x to i16 ret i16 %a } @@ -88,36 +98,36 @@ define ptx_device i16 @cvt_i16_f64(double %x) { ; i32 define ptx_device i32 @cvt_i32_preds(i1 %x) { -; CHECK: cvt.u32.pred r0, p1; -; CHECK: ret; +; CHECK: selp.u32 r0, 1, 0, p1; +; CHECK-NEXT: ret; %a = zext i1 %x to i32 ret i32 %a } define ptx_device i32 @cvt_i32_i16(i16 %x) { ; CHECK: cvt.u32.u16 r0, rh1; -; CHECK: ret; +; CHECK-NEXT: ret; %a = zext i16 %x to i32 ret i32 %a } define ptx_device i32 @cvt_i32_i64(i64 %x) { ; CHECK: cvt.u32.u64 r0, rd1; -; CHECK: ret; +; CHECK-NEXT: ret; %a = trunc i64 %x to i32 ret i32 %a } define ptx_device i32 @cvt_i32_f32(float %x) { ; CHECK: cvt.rzi.u32.f32 r0, r1; -; CHECK: ret; +; CHECK-NEXT: ret; %a = fptoui float %x to i32 ret i32 %a } define ptx_device i32 @cvt_i32_f64(double %x) { ; CHECK: cvt.rzi.u32.f64 r0, rd1; -; CHECK: ret; +; CHECK-NEXT: ret; %a = fptoui double %x to i32 ret i32 %a } @@ -125,29 +135,29 @@ define ptx_device i32 @cvt_i32_f64(double %x) { ; i64 define ptx_device i64 @cvt_i64_preds(i1 %x) { -; CHECK: cvt.u64.pred rd0, p1; -; CHECK: ret; +; CHECK: selp.u64 rd0, 1, 0, p1; +; CHECK-NEXT: ret; %a = zext i1 %x to i64 ret i64 %a } define ptx_device i64 @cvt_i64_i16(i16 %x) { ; CHECK: cvt.u64.u16 rd0, rh1; -; CHECK: ret; +; CHECK-NEXT: ret; %a = zext i16 %x to i64 ret i64 %a } define ptx_device i64 @cvt_i64_i32(i32 %x) { ; CHECK: cvt.u64.u32 rd0, r1; -; CHECK: ret; +; CHECK-NEXT: ret; %a = zext i32 %x to i64 ret i64 %a } define ptx_device i64 @cvt_i64_f32(float %x) { ; CHECK: cvt.rzi.u64.f32 rd0, r1; -; CHECK: ret; +; CHECK-NEXT: ret; %a = fptoui float %x to i64 ret i64 %a } @@ -162,36 +172,36 @@ define ptx_device i64 @cvt_i64_f64(double %x) { ; f32 define ptx_device float @cvt_f32_preds(i1 %x) { -; CHECK: cvt.rn.f32.pred r0, p1; -; CHECK: ret; +; CHECK: selp.f32 r0, 0F3F800000, 0F00000000, p1; +; CHECK-NEXT: ret; %a = uitofp i1 %x to float ret float %a } define ptx_device float @cvt_f32_i16(i16 %x) { ; CHECK: cvt.rn.f32.u16 r0, rh1; -; CHECK: ret; +; CHECK-NEXT: ret; %a = uitofp i16 %x to float ret float %a } define ptx_device float @cvt_f32_i32(i32 %x) { ; CHECK: cvt.rn.f32.u32 r0, r1; -; CHECK: ret; +; CHECK-NEXT: ret; %a = uitofp i32 %x to float ret float %a } define ptx_device float @cvt_f32_i64(i64 %x) { ; CHECK: cvt.rn.f32.u64 r0, rd1; -; CHECK: ret; +; CHECK-NEXT: ret; %a = uitofp i64 %x to float ret float %a } define ptx_device float @cvt_f32_f64(double %x) { ; CHECK: cvt.rn.f32.f64 r0, rd1; -; CHECK: ret; +; CHECK-NEXT: ret; %a = fptrunc double %x to float ret float %a } @@ -199,36 +209,36 @@ define ptx_device float @cvt_f32_f64(double %x) { ; f64 define ptx_device double @cvt_f64_preds(i1 %x) { -; CHECK: cvt.rn.f64.pred rd0, p1; -; CHECK: ret; +; CHECK: selp.f64 rd0, 0D3F80000000000000, 0D0000000000000000, p1; +; CHECK-NEXT: ret; %a = uitofp i1 %x to double ret double %a } define ptx_device double @cvt_f64_i16(i16 %x) { ; CHECK: cvt.rn.f64.u16 rd0, rh1; -; CHECK: ret; +; CHECK-NEXT: ret; %a = uitofp i16 %x to double ret double %a } define ptx_device double @cvt_f64_i32(i32 %x) { ; CHECK: cvt.rn.f64.u32 rd0, r1; -; CHECK: ret; +; CHECK-NEXT: ret; %a = uitofp i32 %x to double ret double %a } define ptx_device double @cvt_f64_i64(i64 %x) { ; CHECK: cvt.rn.f64.u64 rd0, rd1; -; CHECK: ret; +; CHECK-NEXT: ret; %a = uitofp i64 %x to double ret double %a } define ptx_device double @cvt_f64_f32(float %x) { ; CHECK: cvt.f64.f32 rd0, r1; -; CHECK: ret; +; CHECK-NEXT: ret; %a = fpext float %x to double ret double %a } diff --git a/test/CodeGen/PTX/setp.ll b/test/CodeGen/PTX/setp.ll index 5836122049e..a5d7484268c 100644 --- a/test/CodeGen/PTX/setp.ll +++ b/test/CodeGen/PTX/setp.ll @@ -2,7 +2,7 @@ define ptx_device i32 @test_setp_eq_u32_rr(i32 %x, i32 %y) { ; CHECK: setp.eq.u32 p0, r1, r2; -; CHECK-NEXT: cvt.u32.pred r0, p0; +; CHECK-NEXT: selp.u32 r0, 1, 0, p0; ; CHECK-NEXT: ret; %p = icmp eq i32 %x, %y %z = zext i1 %p to i32 @@ -11,7 +11,7 @@ define ptx_device i32 @test_setp_eq_u32_rr(i32 %x, i32 %y) { define ptx_device i32 @test_setp_ne_u32_rr(i32 %x, i32 %y) { ; CHECK: setp.ne.u32 p0, r1, r2; -; CHECK-NEXT: cvt.u32.pred r0, p0; +; CHECK-NEXT: selp.u32 r0, 1, 0, p0; ; CHECK-NEXT: ret; %p = icmp ne i32 %x, %y %z = zext i1 %p to i32 @@ -20,7 +20,7 @@ define ptx_device i32 @test_setp_ne_u32_rr(i32 %x, i32 %y) { define ptx_device i32 @test_setp_lt_u32_rr(i32 %x, i32 %y) { ; CHECK: setp.lt.u32 p0, r1, r2; -; CHECK-NEXT: cvt.u32.pred r0, p0; +; CHECK-NEXT: selp.u32 r0, 1, 0, p0; ; CHECK-NEXT: ret; %p = icmp ult i32 %x, %y %z = zext i1 %p to i32 @@ -29,7 +29,7 @@ define ptx_device i32 @test_setp_lt_u32_rr(i32 %x, i32 %y) { define ptx_device i32 @test_setp_le_u32_rr(i32 %x, i32 %y) { ; CHECK: setp.le.u32 p0, r1, r2; -; CHECK-NEXT: cvt.u32.pred r0, p0; +; CHECK-NEXT: selp.u32 r0, 1, 0, p0; ; CHECK-NEXT: ret; %p = icmp ule i32 %x, %y %z = zext i1 %p to i32 @@ -38,7 +38,7 @@ define ptx_device i32 @test_setp_le_u32_rr(i32 %x, i32 %y) { define ptx_device i32 @test_setp_gt_u32_rr(i32 %x, i32 %y) { ; CHECK: setp.gt.u32 p0, r1, r2; -; CHECK-NEXT: cvt.u32.pred r0, p0; +; CHECK-NEXT: selp.u32 r0, 1, 0, p0; ; CHECK-NEXT: ret; %p = icmp ugt i32 %x, %y %z = zext i1 %p to i32 @@ -47,7 +47,7 @@ define ptx_device i32 @test_setp_gt_u32_rr(i32 %x, i32 %y) { define ptx_device i32 @test_setp_ge_u32_rr(i32 %x, i32 %y) { ; CHECK: setp.ge.u32 p0, r1, r2; -; CHECK-NEXT: cvt.u32.pred r0, p0; +; CHECK-NEXT: selp.u32 r0, 1, 0, p0; ; CHECK-NEXT: ret; %p = icmp uge i32 %x, %y %z = zext i1 %p to i32 @@ -56,7 +56,7 @@ define ptx_device i32 @test_setp_ge_u32_rr(i32 %x, i32 %y) { define ptx_device i32 @test_setp_eq_u32_ri(i32 %x) { ; CHECK: setp.eq.u32 p0, r1, 1; -; CHECK-NEXT: cvt.u32.pred r0, p0; +; CHECK-NEXT: selp.u32 r0, 1, 0, p0; ; CHECK-NEXT: ret; %p = icmp eq i32 %x, 1 %z = zext i1 %p to i32 @@ -65,7 +65,7 @@ define ptx_device i32 @test_setp_eq_u32_ri(i32 %x) { define ptx_device i32 @test_setp_ne_u32_ri(i32 %x) { ; CHECK: setp.ne.u32 p0, r1, 1; -; CHECK-NEXT: cvt.u32.pred r0, p0; +; CHECK-NEXT: selp.u32 r0, 1, 0, p0; ; CHECK-NEXT: ret; %p = icmp ne i32 %x, 1 %z = zext i1 %p to i32 @@ -74,7 +74,7 @@ define ptx_device i32 @test_setp_ne_u32_ri(i32 %x) { define ptx_device i32 @test_setp_lt_u32_ri(i32 %x) { ; CHECK: setp.eq.u32 p0, r1, 0; -; CHECK-NEXT: cvt.u32.pred r0, p0; +; CHECK-NEXT: selp.u32 r0, 1, 0, p0; ; CHECK-NEXT: ret; %p = icmp ult i32 %x, 1 %z = zext i1 %p to i32 @@ -83,7 +83,7 @@ define ptx_device i32 @test_setp_lt_u32_ri(i32 %x) { define ptx_device i32 @test_setp_le_u32_ri(i32 %x) { ; CHECK: setp.lt.u32 p0, r1, 2; -; CHECK-NEXT: cvt.u32.pred r0, p0; +; CHECK-NEXT: selp.u32 r0, 1, 0, p0; ; CHECK-NEXT: ret; %p = icmp ule i32 %x, 1 %z = zext i1 %p to i32 @@ -92,7 +92,7 @@ define ptx_device i32 @test_setp_le_u32_ri(i32 %x) { define ptx_device i32 @test_setp_gt_u32_ri(i32 %x) { ; CHECK: setp.gt.u32 p0, r1, 1; -; CHECK-NEXT: cvt.u32.pred r0, p0; +; CHECK-NEXT: selp.u32 r0, 1, 0, p0; ; CHECK-NEXT: ret; %p = icmp ugt i32 %x, 1 %z = zext i1 %p to i32 @@ -101,7 +101,7 @@ define ptx_device i32 @test_setp_gt_u32_ri(i32 %x) { define ptx_device i32 @test_setp_ge_u32_ri(i32 %x) { ; CHECK: setp.ne.u32 p0, r1, 0; -; CHECK-NEXT: cvt.u32.pred r0, p0; +; CHECK-NEXT: selp.u32 r0, 1, 0, p0; ; CHECK-NEXT: ret; %p = icmp uge i32 %x, 1 %z = zext i1 %p to i32 @@ -111,7 +111,7 @@ define ptx_device i32 @test_setp_ge_u32_ri(i32 %x) { define ptx_device i32 @test_setp_4_op_format_1(i32 %x, i32 %y, i32 %u, i32 %v) { ; CHECK: setp.gt.u32 p0, r3, r4; ; CHECK-NEXT: setp.eq.and.u32 p0, r1, r2, p0; -; CHECK-NEXT: cvt.u32.pred r0, p0; +; CHECK-NEXT: selp.u32 r0, 1, 0, p0; ; CHECK-NEXT: ret; %c = icmp eq i32 %x, %y %d = icmp ugt i32 %u, %v @@ -121,9 +121,9 @@ define ptx_device i32 @test_setp_4_op_format_1(i32 %x, i32 %y, i32 %u, i32 %v) { } define ptx_device i32 @test_setp_4_op_format_2(i32 %x, i32 %y, i32 %w) { -; CHECK: cvt.pred.u32 p0, r3; +; CHECK: setp.gt.b32 p0, r3, 0; ; CHECK-NEXT: setp.eq.and.u32 p0, r1, r2, !p0; -; CHECK-NEXT: cvt.u32.pred r0, p0; +; CHECK-NEXT: selp.u32 r0, 1, 0, p0; ; CHECK-NEXT: ret; %c = trunc i32 %w to i1 %d = icmp eq i32 %x, %y