1; RUN: llc -O2 < %s -march=nvptx -mcpu=sm_35 | FileCheck %s --check-prefix=O2 --check-prefix=CHECK 2; RUN: llc -O0 < %s -march=nvptx -mcpu=sm_35 | FileCheck %s --check-prefix=O0 --check-prefix=CHECK 3 4; The following IR 5; 6; quot = n / d 7; rem = n % d 8; 9; should be transformed into 10; 11; quot = n / d 12; rem = n - (n / d) * d 13; 14; during NVPTX isel, at -O2. At -O0, we should leave it alone. 15 16; CHECK-LABEL: sdiv32( 17define void @sdiv32(i32 %n, i32 %d, i32* %quot_ret, i32* %rem_ret) { 18 ; CHECK: div.s32 [[quot:%r[0-9]+]], [[num:%r[0-9]+]], [[den:%r[0-9]+]]; 19 %quot = sdiv i32 %n, %d 20 21 ; O0: rem.s32 22 ; (This is unfortunately order-sensitive, even though mul is commutative.) 23 ; O2: mul.lo.s32 [[mul:%r[0-9]+]], [[quot]], [[den]]; 24 ; O2: sub.s32 [[rem:%r[0-9]+]], [[num]], [[mul]] 25 %rem = srem i32 %n, %d 26 27 ; O2: st{{.*}}[[quot]] 28 store i32 %quot, i32* %quot_ret 29 ; O2: st{{.*}}[[rem]] 30 store i32 %rem, i32* %rem_ret 31 ret void 32} 33 34; CHECK-LABEL: udiv32( 35define void @udiv32(i32 %n, i32 %d, i32* %quot_ret, i32* %rem_ret) { 36 ; CHECK: div.u32 [[quot:%r[0-9]+]], [[num:%r[0-9]+]], [[den:%r[0-9]+]]; 37 %quot = udiv i32 %n, %d 38 39 ; O0: rem.u32 40 41 ; Selection DAG doesn't know whether this is signed or unsigned 42 ; multiplication and subtraction, but it doesn't make a difference either 43 ; way. 44 ; O2: mul.lo.{{u|s}}32 [[mul:%r[0-9]+]], [[quot]], [[den]]; 45 ; O2: sub.{{u|s}}32 [[rem:%r[0-9]+]], [[num]], [[mul]] 46 %rem = urem i32 %n, %d 47 48 ; O2: st{{.*}}[[quot]] 49 store i32 %quot, i32* %quot_ret 50 ; O2: st{{.*}}[[rem]] 51 store i32 %rem, i32* %rem_ret 52 ret void 53} 54 55; Check that we don't perform this optimization if one operation is signed and 56; the other isn't. 57; CHECK-LABEL: mismatched_types1( 58define void @mismatched_types1(i32 %n, i32 %d, i32* %quot_ret, i32* %rem_ret) { 59 ; CHECK: div.u32 60 ; CHECK: rem.s32 61 %quot = udiv i32 %n, %d 62 %rem = srem i32 %n, %d 63 store i32 %quot, i32* %quot_ret 64 store i32 %rem, i32* %rem_ret 65 ret void 66} 67 68; CHECK-LABEL: mismatched_types2( 69define void @mismatched_types2(i32 %n, i32 %d, i32* %quot_ret, i32* %rem_ret) { 70 ; CHECK: div.s32 71 ; CHECK: rem.u32 72 %quot = sdiv i32 %n, %d 73 %rem = urem i32 %n, %d 74 store i32 %quot, i32* %quot_ret 75 store i32 %rem, i32* %rem_ret 76 ret void 77} 78 79; Check that we don't perform this optimization if the inputs to the div don't 80; match the inputs to the rem. 81; CHECK-LABEL: mismatched_inputs1( 82define void @mismatched_inputs1(i32 %n, i32 %d, i32* %quot_ret, i32* %rem_ret) { 83 ; CHECK: div.s32 84 ; CHECK: rem.s32 85 %quot = sdiv i32 %n, %d 86 %rem = srem i32 %d, %n 87 store i32 %quot, i32* %quot_ret 88 store i32 %rem, i32* %rem_ret 89 ret void 90} 91 92; CHECK-LABEL: mismatched_inputs2( 93define void @mismatched_inputs2(i32 %n1, i32 %n2, i32 %d, i32* %quot_ret, i32* %rem_ret) { 94 ; CHECK: div.s32 95 ; CHECK: rem.s32 96 %quot = sdiv i32 %n1, %d 97 %rem = srem i32 %n2, %d 98 store i32 %quot, i32* %quot_ret 99 store i32 %rem, i32* %rem_ret 100 ret void 101} 102 103; CHECK-LABEL: mismatched_inputs3( 104define void @mismatched_inputs3(i32 %n, i32 %d1, i32 %d2, i32* %quot_ret, i32* %rem_ret) { 105 ; CHECK: div.s32 106 ; CHECK: rem.s32 107 %quot = sdiv i32 %n, %d1 108 %rem = srem i32 %n, %d2 109 store i32 %quot, i32* %quot_ret 110 store i32 %rem, i32* %rem_ret 111 ret void 112} 113