This commit also makes some slight changes to the scheduling model for AMDGPU to set the RetireOOO flag for all scheduling classes.
This flag is only used by llvm-mca and allows instructions to retire out of order.
See the differential link below for a deeper explanation of everything.
Differential Revision: https://reviews.llvm.org/D104730
// The latency values are 1 / (operations / cycle) / 4.
multiclass SICommonWriteRes {
+ let RetireOOO = 1 in { // llvm-mca specific flag
def : HWWriteRes<WriteBranch, [HWBranch], 8>;
def : HWWriteRes<WriteExport, [HWExport], 4>;
def : HWWriteRes<WriteLDS, [HWLGKM], 5>; // Can be between 2 and 64
def : HWWriteRes<Write8PassMAI, [HWXDL], 8>;
let ResourceCycles = [16] in
def : HWWriteRes<Write16PassMAI, [HWXDL], 16>;
+ } // End RetireOOO = 1
def : ReadAdvance<MIVGPRRead, -2>;
defm : SICommonWriteRes;
+let RetireOOO = 1 in { // llvm-mca specific flag
def : HWVALUWriteRes<Write64Bit, 2>;
def : HWVALUWriteRes<WriteIntMul, 4>;
def : HWVALUWriteRes<WriteFloatFMA, 1>;
def : HWVALUWriteRes<WriteDoubleAdd, 2>;
def : HWVALUWriteRes<WriteDoubleCvt, 4>;
def : HWVALUWriteRes<WriteTrans64, 4>;
+} // End RetireOOO = 1
def : InstRW<[WriteCopy], (instrs COPY)>;
defm : SICommonWriteRes;
+let RetireOOO = 1 in { // llvm-mca specific flag
def : HWVALUWriteRes<Write64Bit, 2>;
def : HWVALUWriteRes<WriteIntMul, 4>;
def : HWVALUWriteRes<WriteFloatFMA, 16>;
def : HWVALUWriteRes<WriteDoubleAdd, 8>;
def : HWVALUWriteRes<WriteDoubleCvt, 4>;
def : HWVALUWriteRes<WriteTrans64, 16>;
+} // End RetireOOO = 1
def : InstRW<[WriteCopy], (instrs COPY)>;
def : InstRW<[Write64Bit, MIReadVGPR], (instregex "^V_ACCVGPR_WRITE_B32_e64$")>;
defm : SICommonWriteRes;
+let RetireOOO = 1 in { // llvm-mca specific flag
def : HWVALUWriteRes<WriteFloatFMA, 1>;
def : HWVALUWriteRes<WriteDouble, 1>;
def : HWVALUWriteRes<WriteDoubleAdd, 1>;
def : HWVALUWriteRes<WriteTrans64, 4>;
def : HWVALUWriteRes<WriteIntMul, 1>;
def : HWVALUWriteRes<Write64Bit, 1>;
+} // End RetireOOO = 1
def : InstRW<[WriteCopy], (instrs COPY)>;
def : InstRW<[Write64Bit], (instregex "^V_ACCVGPR_WRITE_B32_e64$")>;
// The latency values are 1 / (operations / cycle).
// Add 1 stall cycle for VGPR read.
+let RetireOOO = 1 in { // llvm-mca specific flag
def : HWWriteRes<Write32Bit, [HWVALU, HWRC], 5>;
def : HWWriteRes<WriteFloatCvt, [HWVALU, HWRC], 5>;
def : HWWriteRes<Write64Bit, [HWVALU, HWRC], 6>;
def : HWWriteRes<WriteSMEM, [HWLGKM, HWRC], 20>;
def : HWWriteRes<WriteVMEM, [HWVMEM, HWRC], 320>;
def : HWWriteRes<WriteBarrier, [HWBranch], 2000>;
+} // End RetireOOO = 1
def : InstRW<[WriteCopy], (instrs COPY)>;
# CHECK: Iterations: 1
# CHECK-NEXT: Instructions: 28
-# CHECK-NEXT: Total Cycles: 224
+# CHECK-NEXT: Total Cycles: 205
# CHECK-NEXT: Total uOps: 29
# CHECK: Dispatch Width: 1
-# CHECK-NEXT: uOps Per Cycle: 0.13
-# CHECK-NEXT: IPC: 0.13
+# CHECK-NEXT: uOps Per Cycle: 0.14
+# CHECK-NEXT: IPC: 0.14
# CHECK-NEXT: Block RThroughput: 29.0
# CHECK: Instruction Info:
# CHECK-NEXT: - - - 1.00 - 1.00 1.00 - v_sqrt_f64_e32 v[4:5], v[4:5]
# CHECK: Timeline view:
-# CHECK-NEXT: 0123456789 0123456789 0123456789 0123456789 0123456789 0123456789 0123456789 0123456789 0123456789 0123456789 0123456789
-# CHECK-NEXT: Index 0123456789 0123456789 0123456789 0123456789 0123456789 0123456789 0123456789 0123456789 0123456789 0123456789 0123456789 0123
-
-# CHECK: [0,0] DeeeeeeeeeeeeeeeeeeeeeE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v_cvt_i32_f64_e32 v0, v[0:1]
-# CHECK-NEXT: [0,1] .DeeeeeeeeeeeeeeeeeeeeeE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v_cvt_f64_i32_e32 v[2:3], v2
-# CHECK-NEXT: [0,2] . DeeeeeeeeeeeeeeeeeeeeeE. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v_cvt_f32_f64_e32 v4, v[4:5]
-# CHECK-NEXT: [0,3] . DeeeeeeeeeeeeeeeeeeeeeE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v_cvt_f64_f32_e32 v[6:7], v6
-# CHECK-NEXT: [0,4] . DeeeeeeeeeeeeeeeeeeeeeE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v_cvt_u32_f64_e32 v8, v[8:9]
-# CHECK-NEXT: [0,5] . DeeeeeeeeeeeeeeeeeeeeeE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v_cvt_f64_u32_e32 v[10:11], v10
-# CHECK-NEXT: [0,6] . . . . . DeeeeeeeeeeeeeeeeeeeeeE. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v_frexp_exp_i32_f64_e32 v0, v[0:1]
-# CHECK-NEXT: [0,7] . . . . . DeeeeeeeeeeeeeeeeeeeeeE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v_frexp_mant_f64_e32 v[2:3], v[2:3]
-# CHECK-NEXT: [0,8] . . . . . DeeeeeeeeeeeeeeeeeeeeeE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v_fract_f64_e32 v[4:5], v[4:5]
-# CHECK-NEXT: [0,9] . . . . . . . . . DeeeeeeeeeeeeeeeeeeeeeE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v_trunc_f64_e32 v[0:1], v[0:1]
-# CHECK-NEXT: [0,10] . . . . . . . . . DeeeeeeeeeeeeeeeeeeeeeE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v_ceil_f64_e32 v[2:3], v[2:3]
-# CHECK-NEXT: [0,11] . . . . . . . . . .DeeeeeeeeeeeeeeeeeeeeeE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v_rndne_f64_e32 v[4:5], v[4:5]
-# CHECK-NEXT: [0,12] . . . . . . . . . . DeeeeeeeeeeeeeeeeeeeeeE. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v_floor_f64_e32 v[6:7], v[6:7]
-# CHECK-NEXT: [0,13] . . . . . . . . . . . . . .DeeeeeeeeeeeeeeeeeeeeeE . . . . . . . . . . . . . . . . . . . . . . . . . . . . v_fma_f64 v[0:1], v[0:1], v[0:1], v[0:1]
-# CHECK-NEXT: [0,14] . . . . . . . . . . . . . . DeeeeeeeeeeeeeeeeeeeeeE. . . . . . . . . . . . . . . . . . . . . . . . . . . . v_add_f64 v[2:3], v[2:3], v[2:3]
-# CHECK-NEXT: [0,15] . . . . . . . . . . . . . . DeeeeeeeeeeeeeeeeeeeeeE . . . . . . . . . . . . . . . . . . . . . . . . . . . v_mul_f64 v[4:5], v[4:5], v[4:5]
-# CHECK-NEXT: [0,16] . . . . . . . . . . . . . . DeeeeeeeeeeeeeeeeeeeeeE . . . . . . . . . . . . . . . . . . . . . . . . . . . v_min_f64 v[6:7], v[6:7], v[6:7]
-# CHECK-NEXT: [0,17] . . . . . . . . . . . . . . DeeeeeeeeeeeeeeeeeeeeeE . . . . . . . . . . . . . . . . . . . . . . . . . . . v_max_f64 v[8:9], v[8:9], v[8:9]
-# CHECK-NEXT: [0,18] . . . . . . . . . . . . . . . . . . DeeeeeeeeeeeeeeeeeeeeeE . . . . . . . . . . . . . . . . . . . . . . . v_div_fmas_f64 v[0:1], v[0:1], v[0:1], v[0:1]
-# CHECK-NEXT: [0,19] . . . . . . . . . . . . . . . . . . . . . . DeeeeeeeeeeeeeeeeeeeeeE . . . . . . . . . . . . . . . . . . . v_div_fixup_f64 v[0:1], v[0:1], v[0:1], v[0:1]
-# CHECK-NEXT: [0,20] . . . . . . . . . . . . . . . . . . . . . . . . . . . DeeeeeeeeeeeeeeeeeeeeeE. . . . . . . . . . . . . . . v_ldexp_f64 v[2:3], v[2:3], v0
-# CHECK-NEXT: [0,21] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . DeeeeeeeeeeeeeeeeeeeeeE. . . . . . . . . . . v_div_scale_f64 v[0:1], vcc_lo, v[0:1], v[0:1], v[0:1]
-# CHECK-NEXT: [0,22] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . DeeeeeeeeeeeeeeeeeeeeeE . . . . . . v_trig_preop_f64 v[2:3], v[2:3], v0
-# CHECK-NEXT: [0,23] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . DeeeeeeeeeeeeeeeeeeeeeE . . . . . . v_cmp_eq_f64_e32 vcc_lo, v[0:1], v[0:1]
-# CHECK-NEXT: [0,24] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .DeeeeeeeeeeeeeeeeeeeeeE . . v_cmp_class_f64_e64 vcc_lo, v[2:3], s0
-# CHECK-NEXT: [0,25] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . DeeeeeeeeeeeeeeeeeeeeeeeE . v_rcp_f64_e32 v[0:1], v[0:1]
-# CHECK-NEXT: [0,26] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . DeeeeeeeeeeeeeeeeeeeeeeeE. v_rsq_f64_e32 v[2:3], v[2:3]
-# CHECK-NEXT: [0,27] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . DeeeeeeeeeeeeeeeeeeeeeeeE v_sqrt_f64_e32 v[4:5], v[4:5]
+# CHECK-NEXT: 0123456789 0123456789 0123456789 0123456789 0123456789 0123456789 0123456789 0123456789 0123456789 0123456789
+# CHECK-NEXT: Index 0123456789 0123456789 0123456789 0123456789 0123456789 0123456789 0123456789 0123456789 0123456789 0123456789 01234
+
+# CHECK: [0,0] DeeeeeeeeeeeeeeeeeeeeeE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v_cvt_i32_f64_e32 v0, v[0:1]
+# CHECK-NEXT: [0,1] .DeeeeeeeeeeeeeeeeeeeeeE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v_cvt_f64_i32_e32 v[2:3], v2
+# CHECK-NEXT: [0,2] . DeeeeeeeeeeeeeeeeeeeeeE. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v_cvt_f32_f64_e32 v4, v[4:5]
+# CHECK-NEXT: [0,3] . DeeeeeeeeeeeeeeeeeeeeeE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v_cvt_f64_f32_e32 v[6:7], v6
+# CHECK-NEXT: [0,4] . DeeeeeeeeeeeeeeeeeeeeeE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v_cvt_u32_f64_e32 v8, v[8:9]
+# CHECK-NEXT: [0,5] . DeeeeeeeeeeeeeeeeeeeeeE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v_cvt_f64_u32_e32 v[10:11], v10
+# CHECK-NEXT: [0,6] . . . . . DeeeeeeeeeeeeeeeeeeeeeE. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v_frexp_exp_i32_f64_e32 v0, v[0:1]
+# CHECK-NEXT: [0,7] . . . . . DeeeeeeeeeeeeeeeeeeeeeE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v_frexp_mant_f64_e32 v[2:3], v[2:3]
+# CHECK-NEXT: [0,8] . . . . . DeeeeeeeeeeeeeeeeeeeeeE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v_fract_f64_e32 v[4:5], v[4:5]
+# CHECK-NEXT: [0,9] . . . . . . . . . DeeeeeeeeeeeeeeeeeeeeeE . . . . . . . . . . . . . . . . . . . . . . . . . . . . v_trunc_f64_e32 v[0:1], v[0:1]
+# CHECK-NEXT: [0,10] . . . . . . . . . DeeeeeeeeeeeeeeeeeeeeeE . . . . . . . . . . . . . . . . . . . . . . . . . . . . v_ceil_f64_e32 v[2:3], v[2:3]
+# CHECK-NEXT: [0,11] . . . . . . . . . .DeeeeeeeeeeeeeeeeeeeeeE . . . . . . . . . . . . . . . . . . . . . . . . . . . . v_rndne_f64_e32 v[4:5], v[4:5]
+# CHECK-NEXT: [0,12] . . . . . . . . . . DeeeeeeeeeeeeeeeeeeeeeE. . . . . . . . . . . . . . . . . . . . . . . . . . . . v_floor_f64_e32 v[6:7], v[6:7]
+# CHECK-NEXT: [0,13] . . . . . . . . . . . . . .DeeeeeeeeeeeeeeeeeeeeeE . . . . . . . . . . . . . . . . . . . . . . . . v_fma_f64 v[0:1], v[0:1], v[0:1], v[0:1]
+# CHECK-NEXT: [0,14] . . . . . . . . . . . . . . DeeeeeeeeeeeeeeeeeeeeeE. . . . . . . . . . . . . . . . . . . . . . . . v_add_f64 v[2:3], v[2:3], v[2:3]
+# CHECK-NEXT: [0,15] . . . . . . . . . . . . . . DeeeeeeeeeeeeeeeeeeeeeE . . . . . . . . . . . . . . . . . . . . . . . v_mul_f64 v[4:5], v[4:5], v[4:5]
+# CHECK-NEXT: [0,16] . . . . . . . . . . . . . . DeeeeeeeeeeeeeeeeeeeeeE . . . . . . . . . . . . . . . . . . . . . . . v_min_f64 v[6:7], v[6:7], v[6:7]
+# CHECK-NEXT: [0,17] . . . . . . . . . . . . . . DeeeeeeeeeeeeeeeeeeeeeE . . . . . . . . . . . . . . . . . . . . . . . v_max_f64 v[8:9], v[8:9], v[8:9]
+# CHECK-NEXT: [0,18] . . . . . . . . . . . . . . . . . . DeeeeeeeeeeeeeeeeeeeeeE . . . . . . . . . . . . . . . . . . . v_div_fmas_f64 v[0:1], v[0:1], v[0:1], v[0:1]
+# CHECK-NEXT: [0,19] . . . . . . . . . . . . . . . . . . . . . . DeeeeeeeeeeeeeeeeeeeeeE . . . . . . . . . . . . . . . v_div_fixup_f64 v[0:1], v[0:1], v[0:1], v[0:1]
+# CHECK-NEXT: [0,20] . . . . . . . . . . . . . . . . . . . . . . . . . . . DeeeeeeeeeeeeeeeeeeeeeE. . . . . . . . . . . v_ldexp_f64 v[2:3], v[2:3], v0
+# CHECK-NEXT: [0,21] . . . . . . . . . . . . . . . . . . . . . . . . . . . DeeeeeeeeeeeeeeeeeeeeeE . . . . . . . . . . v_div_scale_f64 v[0:1], vcc_lo, v[0:1], v[0:1], v[0:1]
+# CHECK-NEXT: [0,22] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . DeeeeeeeeeeeeeeeeeeeeeE . . . . . . v_trig_preop_f64 v[2:3], v[2:3], v0
+# CHECK-NEXT: [0,23] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .DeeeeeeeeeeeeeeeeeeeeeE . . . . . . v_cmp_eq_f64_e32 vcc_lo, v[0:1], v[0:1]
+# CHECK-NEXT: [0,24] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . DeeeeeeeeeeeeeeeeeeeeeE. . v_cmp_class_f64_e64 vcc_lo, v[2:3], s0
+# CHECK-NEXT: [0,25] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . DeeeeeeeeeeeeeeeeeeeeeeeE . v_rcp_f64_e32 v[0:1], v[0:1]
+# CHECK-NEXT: [0,26] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . DeeeeeeeeeeeeeeeeeeeeeeeE. v_rsq_f64_e32 v[2:3], v[2:3]
+# CHECK-NEXT: [0,27] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . DeeeeeeeeeeeeeeeeeeeeeeeE v_sqrt_f64_e32 v[4:5], v[4:5]
# CHECK: Average Wait times (based on the timeline view):
# CHECK-NEXT: [0]: Executions
--- /dev/null
+# NOTE: Assertions have been autogenerated by utils/update_mca_test_checks.py
+# RUN: llvm-mca -mtriple=amdgcn -mcpu=gfx900 --timeline --iterations=1 --timeline-max-cycles=0 < %s | FileCheck %s
+
+s_load_dwordx2 s[2:3], s[0:1], 0x24
+s_load_dwordx2 s[0:1], s[0:1], 0x2c
+s_waitcnt lgkmcnt(0)
+v_mov_b32_e32 v0, s2
+v_mov_b32_e32 v1, s3
+flat_load_dword v2, v[0:1]
+flat_load_dword v3, v[0:1] offset:8
+flat_load_dword v4, v[0:1] offset:16
+flat_load_dword v5, v[0:1] offset:24
+v_mov_b32_e32 v0, s0
+v_mov_b32_e32 v1, s1
+v_mov_b32_e32 v6, s6
+v_mov_b32_e32 v7, s7
+v_mov_b32_e32 v8, s8
+v_mov_b32_e32 v9, s9
+v_mov_b32_e32 v10, s10
+v_mov_b32_e32 v11, s11
+v_mov_b32_e32 v12, s12
+v_mov_b32_e32 v13, s13
+v_mov_b32_e32 v14, s14
+v_mov_b32_e32 v15, s15
+v_mov_b32_e32 v16, s16
+v_mov_b32_e32 v17, s17
+v_mov_b32_e32 v18, s18
+v_mov_b32_e32 v19, s19
+v_mov_b32_e32 v20, s20
+v_mov_b32_e32 v21, s21
+v_mov_b32_e32 v22, s22
+v_mov_b32_e32 v23, s23
+v_mov_b32_e32 v24, s24
+v_mov_b32_e32 v25, s25
+v_mov_b32_e32 v26, s26
+v_mov_b32_e32 v27, s27
+v_mov_b32_e32 v28, s28
+v_mov_b32_e32 v29, s29
+s_waitcnt vmcnt(0) lgkmcnt(0)
+
+# CHECK: Iterations: 1
+# CHECK-NEXT: Instructions: 36
+# CHECK-NEXT: Total Cycles: 94
+# CHECK-NEXT: Total uOps: 36
+
+# CHECK: Dispatch Width: 1
+# CHECK-NEXT: uOps Per Cycle: 0.38
+# CHECK-NEXT: IPC: 0.38
+# CHECK-NEXT: Block RThroughput: 36.0
+
+# CHECK: Instruction Info:
+# CHECK-NEXT: [1]: #uOps
+# CHECK-NEXT: [2]: Latency
+# CHECK-NEXT: [3]: RThroughput
+# CHECK-NEXT: [4]: MayLoad
+# CHECK-NEXT: [5]: MayStore
+# CHECK-NEXT: [6]: HasSideEffects (U)
+
+# CHECK: [1] [2] [3] [4] [5] [6] Instructions:
+# CHECK-NEXT: 1 5 1.00 * s_load_dwordx2 s[2:3], s[0:1], 0x24
+# CHECK-NEXT: 1 5 1.00 * s_load_dwordx2 s[0:1], s[0:1], 0x2c
+# CHECK-NEXT: 1 1 1.00 U s_waitcnt lgkmcnt(0)
+# CHECK-NEXT: 1 1 1.00 U v_mov_b32_e32 v0, s2
+# CHECK-NEXT: 1 1 1.00 U v_mov_b32_e32 v1, s3
+# CHECK-NEXT: 1 80 1.00 * U flat_load_dword v2, v[0:1]
+# CHECK-NEXT: 1 80 1.00 * U flat_load_dword v3, v[0:1] offset:8
+# CHECK-NEXT: 1 80 1.00 * U flat_load_dword v4, v[0:1] offset:16
+# CHECK-NEXT: 1 80 1.00 * U flat_load_dword v5, v[0:1] offset:24
+# CHECK-NEXT: 1 1 1.00 U v_mov_b32_e32 v0, s0
+# CHECK-NEXT: 1 1 1.00 U v_mov_b32_e32 v1, s1
+# CHECK-NEXT: 1 1 1.00 U v_mov_b32_e32 v6, s6
+# CHECK-NEXT: 1 1 1.00 U v_mov_b32_e32 v7, s7
+# CHECK-NEXT: 1 1 1.00 U v_mov_b32_e32 v8, s8
+# CHECK-NEXT: 1 1 1.00 U v_mov_b32_e32 v9, s9
+# CHECK-NEXT: 1 1 1.00 U v_mov_b32_e32 v10, s10
+# CHECK-NEXT: 1 1 1.00 U v_mov_b32_e32 v11, s11
+# CHECK-NEXT: 1 1 1.00 U v_mov_b32_e32 v12, s12
+# CHECK-NEXT: 1 1 1.00 U v_mov_b32_e32 v13, s13
+# CHECK-NEXT: 1 1 1.00 U v_mov_b32_e32 v14, s14
+# CHECK-NEXT: 1 1 1.00 U v_mov_b32_e32 v15, s15
+# CHECK-NEXT: 1 1 1.00 U v_mov_b32_e32 v16, s16
+# CHECK-NEXT: 1 1 1.00 U v_mov_b32_e32 v17, s17
+# CHECK-NEXT: 1 1 1.00 U v_mov_b32_e32 v18, s18
+# CHECK-NEXT: 1 1 1.00 U v_mov_b32_e32 v19, s19
+# CHECK-NEXT: 1 1 1.00 U v_mov_b32_e32 v20, s20
+# CHECK-NEXT: 1 1 1.00 U v_mov_b32_e32 v21, s21
+# CHECK-NEXT: 1 1 1.00 U v_mov_b32_e32 v22, s22
+# CHECK-NEXT: 1 1 1.00 U v_mov_b32_e32 v23, s23
+# CHECK-NEXT: 1 1 1.00 U v_mov_b32_e32 v24, s24
+# CHECK-NEXT: 1 1 1.00 U v_mov_b32_e32 v25, s25
+# CHECK-NEXT: 1 1 1.00 U v_mov_b32_e32 v26, s26
+# CHECK-NEXT: 1 1 1.00 U v_mov_b32_e32 v27, s27
+# CHECK-NEXT: 1 1 1.00 U v_mov_b32_e32 v28, s28
+# CHECK-NEXT: 1 1 1.00 U v_mov_b32_e32 v29, s29
+# CHECK-NEXT: 1 1 1.00 U s_waitcnt vmcnt(0) lgkmcnt(0)
+
+# CHECK: Resources:
+# CHECK-NEXT: [0] - HWBranch
+# CHECK-NEXT: [1] - HWExport
+# CHECK-NEXT: [2] - HWLGKM
+# CHECK-NEXT: [3] - HWSALU
+# CHECK-NEXT: [4] - HWVALU
+# CHECK-NEXT: [5] - HWVMEM
+# CHECK-NEXT: [6] - HWXDL
+
+# CHECK: Resource pressure per iteration:
+# CHECK-NEXT: [0] [1] [2] [3] [4] [5] [6]
+# CHECK-NEXT: - - 2.00 2.00 28.00 4.00 -
+
+# CHECK: Resource pressure by instruction:
+# CHECK-NEXT: [0] [1] [2] [3] [4] [5] [6] Instructions:
+# CHECK-NEXT: - - 1.00 - - - - s_load_dwordx2 s[2:3], s[0:1], 0x24
+# CHECK-NEXT: - - 1.00 - - - - s_load_dwordx2 s[0:1], s[0:1], 0x2c
+# CHECK-NEXT: - - - 1.00 - - - s_waitcnt lgkmcnt(0)
+# CHECK-NEXT: - - - - 1.00 - - v_mov_b32_e32 v0, s2
+# CHECK-NEXT: - - - - 1.00 - - v_mov_b32_e32 v1, s3
+# CHECK-NEXT: - - - - - 1.00 - flat_load_dword v2, v[0:1]
+# CHECK-NEXT: - - - - - 1.00 - flat_load_dword v3, v[0:1] offset:8
+# CHECK-NEXT: - - - - - 1.00 - flat_load_dword v4, v[0:1] offset:16
+# CHECK-NEXT: - - - - - 1.00 - flat_load_dword v5, v[0:1] offset:24
+# CHECK-NEXT: - - - - 1.00 - - v_mov_b32_e32 v0, s0
+# CHECK-NEXT: - - - - 1.00 - - v_mov_b32_e32 v1, s1
+# CHECK-NEXT: - - - - 1.00 - - v_mov_b32_e32 v6, s6
+# CHECK-NEXT: - - - - 1.00 - - v_mov_b32_e32 v7, s7
+# CHECK-NEXT: - - - - 1.00 - - v_mov_b32_e32 v8, s8
+# CHECK-NEXT: - - - - 1.00 - - v_mov_b32_e32 v9, s9
+# CHECK-NEXT: - - - - 1.00 - - v_mov_b32_e32 v10, s10
+# CHECK-NEXT: - - - - 1.00 - - v_mov_b32_e32 v11, s11
+# CHECK-NEXT: - - - - 1.00 - - v_mov_b32_e32 v12, s12
+# CHECK-NEXT: - - - - 1.00 - - v_mov_b32_e32 v13, s13
+# CHECK-NEXT: - - - - 1.00 - - v_mov_b32_e32 v14, s14
+# CHECK-NEXT: - - - - 1.00 - - v_mov_b32_e32 v15, s15
+# CHECK-NEXT: - - - - 1.00 - - v_mov_b32_e32 v16, s16
+# CHECK-NEXT: - - - - 1.00 - - v_mov_b32_e32 v17, s17
+# CHECK-NEXT: - - - - 1.00 - - v_mov_b32_e32 v18, s18
+# CHECK-NEXT: - - - - 1.00 - - v_mov_b32_e32 v19, s19
+# CHECK-NEXT: - - - - 1.00 - - v_mov_b32_e32 v20, s20
+# CHECK-NEXT: - - - - 1.00 - - v_mov_b32_e32 v21, s21
+# CHECK-NEXT: - - - - 1.00 - - v_mov_b32_e32 v22, s22
+# CHECK-NEXT: - - - - 1.00 - - v_mov_b32_e32 v23, s23
+# CHECK-NEXT: - - - - 1.00 - - v_mov_b32_e32 v24, s24
+# CHECK-NEXT: - - - - 1.00 - - v_mov_b32_e32 v25, s25
+# CHECK-NEXT: - - - - 1.00 - - v_mov_b32_e32 v26, s26
+# CHECK-NEXT: - - - - 1.00 - - v_mov_b32_e32 v27, s27
+# CHECK-NEXT: - - - - 1.00 - - v_mov_b32_e32 v28, s28
+# CHECK-NEXT: - - - - 1.00 - - v_mov_b32_e32 v29, s29
+# CHECK-NEXT: - - - 1.00 - - - s_waitcnt vmcnt(0) lgkmcnt(0)
+
+# CHECK: Timeline view:
+# CHECK-NEXT: 0123456789 0123456789 0123456789 0123456789 0123
+# CHECK-NEXT: Index 0123456789 0123456789 0123456789 0123456789 0123456789
+
+# CHECK: [0,0] DeeeeE . . . . . . . . . . . . . . . . . . s_load_dwordx2 s[2:3], s[0:1], 0x24
+# CHECK-NEXT: [0,1] .DeeeeE . . . . . . . . . . . . . . . . . . s_load_dwordx2 s[0:1], s[0:1], 0x2c
+# CHECK-NEXT: [0,2] . .DE . . . . . . . . . . . . . . . . . . s_waitcnt lgkmcnt(0)
+# CHECK-NEXT: [0,3] . . DE . . . . . . . . . . . . . . . . . . v_mov_b32_e32 v0, s2
+# CHECK-NEXT: [0,4] . . DE. . . . . . . . . . . . . . . . . . v_mov_b32_e32 v1, s3
+# CHECK-NEXT: [0,5] . . DeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeE. . flat_load_dword v2, v[0:1]
+# CHECK-NEXT: [0,6] . . DeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeE . flat_load_dword v3, v[0:1] offset:8
+# CHECK-NEXT: [0,7] . . .DeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeE . flat_load_dword v4, v[0:1] offset:16
+# CHECK-NEXT: [0,8] . . . DeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeE. flat_load_dword v5, v[0:1] offset:24
+# CHECK-NEXT: [0,9] . . . DE. . . . . . . . . . . . . . . . . v_mov_b32_e32 v0, s0
+# CHECK-NEXT: [0,10] . . . DE . . . . . . . . . . . . . . . . v_mov_b32_e32 v1, s1
+# CHECK-NEXT: [0,11] . . . DE . . . . . . . . . . . . . . . . v_mov_b32_e32 v6, s6
+# CHECK-NEXT: [0,12] . . . .DE . . . . . . . . . . . . . . . . v_mov_b32_e32 v7, s7
+# CHECK-NEXT: [0,13] . . . . DE . . . . . . . . . . . . . . . . v_mov_b32_e32 v8, s8
+# CHECK-NEXT: [0,14] . . . . DE. . . . . . . . . . . . . . . . v_mov_b32_e32 v9, s9
+# CHECK-NEXT: [0,15] . . . . DE . . . . . . . . . . . . . . . v_mov_b32_e32 v10, s10
+# CHECK-NEXT: [0,16] . . . . DE . . . . . . . . . . . . . . . v_mov_b32_e32 v11, s11
+# CHECK-NEXT: [0,17] . . . . .DE . . . . . . . . . . . . . . . v_mov_b32_e32 v12, s12
+# CHECK-NEXT: [0,18] . . . . . DE . . . . . . . . . . . . . . . v_mov_b32_e32 v13, s13
+# CHECK-NEXT: [0,19] . . . . . DE. . . . . . . . . . . . . . . v_mov_b32_e32 v14, s14
+# CHECK-NEXT: [0,20] . . . . . DE . . . . . . . . . . . . . . v_mov_b32_e32 v15, s15
+# CHECK-NEXT: [0,21] . . . . . DE . . . . . . . . . . . . . . v_mov_b32_e32 v16, s16
+# CHECK-NEXT: [0,22] . . . . . .DE . . . . . . . . . . . . . . v_mov_b32_e32 v17, s17
+# CHECK-NEXT: [0,23] . . . . . . DE . . . . . . . . . . . . . . v_mov_b32_e32 v18, s18
+# CHECK-NEXT: [0,24] . . . . . . DE. . . . . . . . . . . . . . v_mov_b32_e32 v19, s19
+# CHECK-NEXT: [0,25] . . . . . . DE . . . . . . . . . . . . . v_mov_b32_e32 v20, s20
+# CHECK-NEXT: [0,26] . . . . . . DE . . . . . . . . . . . . . v_mov_b32_e32 v21, s21
+# CHECK-NEXT: [0,27] . . . . . . .DE . . . . . . . . . . . . . v_mov_b32_e32 v22, s22
+# CHECK-NEXT: [0,28] . . . . . . . DE . . . . . . . . . . . . . v_mov_b32_e32 v23, s23
+# CHECK-NEXT: [0,29] . . . . . . . DE. . . . . . . . . . . . . v_mov_b32_e32 v24, s24
+# CHECK-NEXT: [0,30] . . . . . . . DE . . . . . . . . . . . . v_mov_b32_e32 v25, s25
+# CHECK-NEXT: [0,31] . . . . . . . DE . . . . . . . . . . . . v_mov_b32_e32 v26, s26
+# CHECK-NEXT: [0,32] . . . . . . . .DE . . . . . . . . . . . . v_mov_b32_e32 v27, s27
+# CHECK-NEXT: [0,33] . . . . . . . . DE . . . . . . . . . . . . v_mov_b32_e32 v28, s28
+# CHECK-NEXT: [0,34] . . . . . . . . DE. . . . . . . . . . . . v_mov_b32_e32 v29, s29
+# CHECK-NEXT: [0,35] . . . . . . . . . . . . . . . . . . . DE s_waitcnt vmcnt(0) lgkmcnt(0)
+
+# CHECK: Average Wait times (based on the timeline view):
+# CHECK-NEXT: [0]: Executions
+# CHECK-NEXT: [1]: Average time spent waiting in a scheduler's queue
+# CHECK-NEXT: [2]: Average time spent waiting in a scheduler's queue while ready
+# CHECK-NEXT: [3]: Average time elapsed from WB until retire stage
+
+# CHECK: [0] [1] [2] [3]
+# CHECK-NEXT: 0. 1 0.0 0.0 0.0 s_load_dwordx2 s[2:3], s[0:1], 0x24
+# CHECK-NEXT: 1. 1 0.0 0.0 0.0 s_load_dwordx2 s[0:1], s[0:1], 0x2c
+# CHECK-NEXT: 2. 1 0.0 0.0 0.0 s_waitcnt lgkmcnt(0)
+# CHECK-NEXT: 3. 1 0.0 0.0 0.0 v_mov_b32_e32 v0, s2
+# CHECK-NEXT: 4. 1 0.0 0.0 0.0 v_mov_b32_e32 v1, s3
+# CHECK-NEXT: 5. 1 0.0 0.0 0.0 flat_load_dword v2, v[0:1]
+# CHECK-NEXT: 6. 1 0.0 0.0 0.0 flat_load_dword v3, v[0:1] offset:8
+# CHECK-NEXT: 7. 1 0.0 0.0 0.0 flat_load_dword v4, v[0:1] offset:16
+# CHECK-NEXT: 8. 1 0.0 0.0 0.0 flat_load_dword v5, v[0:1] offset:24
+# CHECK-NEXT: 9. 1 0.0 0.0 0.0 v_mov_b32_e32 v0, s0
+# CHECK-NEXT: 10. 1 0.0 0.0 0.0 v_mov_b32_e32 v1, s1
+# CHECK-NEXT: 11. 1 0.0 0.0 0.0 v_mov_b32_e32 v6, s6
+# CHECK-NEXT: 12. 1 0.0 0.0 0.0 v_mov_b32_e32 v7, s7
+# CHECK-NEXT: 13. 1 0.0 0.0 0.0 v_mov_b32_e32 v8, s8
+# CHECK-NEXT: 14. 1 0.0 0.0 0.0 v_mov_b32_e32 v9, s9
+# CHECK-NEXT: 15. 1 0.0 0.0 0.0 v_mov_b32_e32 v10, s10
+# CHECK-NEXT: 16. 1 0.0 0.0 0.0 v_mov_b32_e32 v11, s11
+# CHECK-NEXT: 17. 1 0.0 0.0 0.0 v_mov_b32_e32 v12, s12
+# CHECK-NEXT: 18. 1 0.0 0.0 0.0 v_mov_b32_e32 v13, s13
+# CHECK-NEXT: 19. 1 0.0 0.0 0.0 v_mov_b32_e32 v14, s14
+# CHECK-NEXT: 20. 1 0.0 0.0 0.0 v_mov_b32_e32 v15, s15
+# CHECK-NEXT: 21. 1 0.0 0.0 0.0 v_mov_b32_e32 v16, s16
+# CHECK-NEXT: 22. 1 0.0 0.0 0.0 v_mov_b32_e32 v17, s17
+# CHECK-NEXT: 23. 1 0.0 0.0 0.0 v_mov_b32_e32 v18, s18
+# CHECK-NEXT: 24. 1 0.0 0.0 0.0 v_mov_b32_e32 v19, s19
+# CHECK-NEXT: 25. 1 0.0 0.0 0.0 v_mov_b32_e32 v20, s20
+# CHECK-NEXT: 26. 1 0.0 0.0 0.0 v_mov_b32_e32 v21, s21
+# CHECK-NEXT: 27. 1 0.0 0.0 0.0 v_mov_b32_e32 v22, s22
+# CHECK-NEXT: 28. 1 0.0 0.0 0.0 v_mov_b32_e32 v23, s23
+# CHECK-NEXT: 29. 1 0.0 0.0 0.0 v_mov_b32_e32 v24, s24
+# CHECK-NEXT: 30. 1 0.0 0.0 0.0 v_mov_b32_e32 v25, s25
+# CHECK-NEXT: 31. 1 0.0 0.0 0.0 v_mov_b32_e32 v26, s26
+# CHECK-NEXT: 32. 1 0.0 0.0 0.0 v_mov_b32_e32 v27, s27
+# CHECK-NEXT: 33. 1 0.0 0.0 0.0 v_mov_b32_e32 v28, s28
+# CHECK-NEXT: 34. 1 0.0 0.0 0.0 v_mov_b32_e32 v29, s29
+# CHECK-NEXT: 35. 1 0.0 0.0 0.0 s_waitcnt vmcnt(0) lgkmcnt(0)
+# CHECK-NEXT: 1 0.0 0.0 0.0 <total>
namespace llvm {
namespace mca {
+void AMDGPUInstrPostProcess::postProcessInstruction(
+ std::unique_ptr<Instruction> &Inst, const MCInst &MCI) {
+ switch (MCI.getOpcode()) {
+ case AMDGPU::S_WAITCNT:
+ case AMDGPU::S_WAITCNT_EXPCNT:
+ case AMDGPU::S_WAITCNT_LGKMCNT:
+ case AMDGPU::S_WAITCNT_VMCNT:
+ case AMDGPU::S_WAITCNT_VSCNT:
+ case AMDGPU::S_WAITCNT_EXPCNT_gfx10:
+ case AMDGPU::S_WAITCNT_LGKMCNT_gfx10:
+ case AMDGPU::S_WAITCNT_VMCNT_gfx10:
+ case AMDGPU::S_WAITCNT_VSCNT_gfx10:
+ case AMDGPU::S_WAITCNT_gfx10:
+ case AMDGPU::S_WAITCNT_gfx6_gfx7:
+ case AMDGPU::S_WAITCNT_vi:
+ return processWaitCnt(Inst, MCI);
+ }
+}
+
+// s_waitcnt instructions encode important information as immediate operands
+// which are lost during the MCInst -> mca::Instruction lowering.
+void AMDGPUInstrPostProcess::processWaitCnt(std::unique_ptr<Instruction> &Inst,
+ const MCInst &MCI) {
+ for (int Idx = 0, N = MCI.size(); Idx < N; Idx++) {
+ MCAOperand Op;
+ const MCOperand &MCOp = MCI.getOperand(Idx);
+ if (MCOp.isReg()) {
+ Op = MCAOperand::createReg(MCOp.getReg());
+ } else if (MCOp.isImm()) {
+ Op = MCAOperand::createImm(MCOp.getImm());
+ }
+ Op.setIndex(Idx);
+ Inst->addOperand(Op);
+ }
+}
+
AMDGPUCustomBehaviour::AMDGPUCustomBehaviour(const MCSubtargetInfo &STI,
const SourceMgr &SrcMgr,
const MCInstrInfo &MCII)
- : CustomBehaviour(STI, SrcMgr, MCII) {}
+ : CustomBehaviour(STI, SrcMgr, MCII) {
+ generateWaitCntInfo();
+}
unsigned AMDGPUCustomBehaviour::checkCustomHazard(ArrayRef<InstRef> IssuedInst,
const InstRef &IR) {
+ const Instruction &Inst = *IR.getInstruction();
+ unsigned Opcode = Inst.getOpcode();
+
+ // llvm-mca is generally run on fully compiled assembly so we wouldn't see any
+ // pseudo instructions here. However, there are plans for the future to make
+ // it possible to use mca within backend passes. As such, I have left the
+ // pseudo version of s_waitcnt within this switch statement.
+ switch (Opcode) {
+ default:
+ return 0;
+ case AMDGPU::S_WAITCNT: // This instruction
+ case AMDGPU::S_WAITCNT_EXPCNT:
+ case AMDGPU::S_WAITCNT_LGKMCNT:
+ case AMDGPU::S_WAITCNT_VMCNT:
+ case AMDGPU::S_WAITCNT_VSCNT: // to this instruction are all pseudo.
+ case AMDGPU::S_WAITCNT_EXPCNT_gfx10:
+ case AMDGPU::S_WAITCNT_LGKMCNT_gfx10:
+ case AMDGPU::S_WAITCNT_VMCNT_gfx10:
+ case AMDGPU::S_WAITCNT_VSCNT_gfx10:
+ case AMDGPU::S_WAITCNT_gfx10:
+ case AMDGPU::S_WAITCNT_gfx6_gfx7:
+ case AMDGPU::S_WAITCNT_vi:
+ // s_endpgm also behaves as if there is an implicit
+ // s_waitcnt 0, but I'm not sure if it would be appropriate
+ // to model this in llvm-mca based on how the iterations work
+ // while simulating the pipeline over and over.
+ return handleWaitCnt(IssuedInst, IR);
+ }
+
return 0;
}
+unsigned AMDGPUCustomBehaviour::handleWaitCnt(ArrayRef<InstRef> IssuedInst,
+ const InstRef &IR) {
+ // Currently, all s_waitcnt instructions are handled except s_waitcnt_depctr.
+ // I do not know how that instruction works so I did not attempt to model it.
+ // set the max values to begin
+ unsigned Vmcnt = 63;
+ unsigned Expcnt = 7;
+ unsigned Lgkmcnt = 31;
+ unsigned Vscnt = 63;
+ unsigned CurrVmcnt = 0;
+ unsigned CurrExpcnt = 0;
+ unsigned CurrLgkmcnt = 0;
+ unsigned CurrVscnt = 0;
+ unsigned CyclesToWaitVm = ~0U;
+ unsigned CyclesToWaitExp = ~0U;
+ unsigned CyclesToWaitLgkm = ~0U;
+ unsigned CyclesToWaitVs = ~0U;
+
+ computeWaitCnt(IR, Vmcnt, Expcnt, Lgkmcnt, Vscnt);
+
+ // We will now look at each of the currently executing instructions
+ // to find out if this wait instruction still needs to wait.
+ for (auto I = IssuedInst.begin(), E = IssuedInst.end(); I != E; I++) {
+ const InstRef &PrevIR = *I;
+ const Instruction &PrevInst = *PrevIR.getInstruction();
+ const unsigned PrevInstIndex = PrevIR.getSourceIndex() % SrcMgr.size();
+ const WaitCntInfo &PrevInstWaitInfo = InstrWaitCntInfo[PrevInstIndex];
+ const int CyclesLeft = PrevInst.getCyclesLeft();
+ assert(CyclesLeft != UNKNOWN_CYCLES &&
+ "We should know how many cycles are left for this instruction");
+ if (PrevInstWaitInfo.VmCnt) {
+ CurrVmcnt++;
+ if ((unsigned)CyclesLeft < CyclesToWaitVm)
+ CyclesToWaitVm = CyclesLeft;
+ }
+ if (PrevInstWaitInfo.ExpCnt) {
+ CurrExpcnt++;
+ if ((unsigned)CyclesLeft < CyclesToWaitExp)
+ CyclesToWaitExp = CyclesLeft;
+ }
+ if (PrevInstWaitInfo.LgkmCnt) {
+ CurrLgkmcnt++;
+ if ((unsigned)CyclesLeft < CyclesToWaitLgkm)
+ CyclesToWaitLgkm = CyclesLeft;
+ }
+ if (PrevInstWaitInfo.VsCnt) {
+ CurrVscnt++;
+ if ((unsigned)CyclesLeft < CyclesToWaitVs)
+ CyclesToWaitVs = CyclesLeft;
+ }
+ }
+
+ unsigned CyclesToWait = ~0U;
+ if (CurrVmcnt > Vmcnt && CyclesToWaitVm < CyclesToWait)
+ CyclesToWait = CyclesToWaitVm;
+ if (CurrExpcnt > Expcnt && CyclesToWaitExp < CyclesToWait)
+ CyclesToWait = CyclesToWaitExp;
+ if (CurrLgkmcnt > Lgkmcnt && CyclesToWaitLgkm < CyclesToWait)
+ CyclesToWait = CyclesToWaitLgkm;
+ if (CurrVscnt > Vscnt && CyclesToWaitVs < CyclesToWait)
+ CyclesToWait = CyclesToWaitVs;
+
+ // We may underestimate how many cycles we need to wait, but this
+ // isn't a big deal. Our return value is just how many cycles until
+ // this function gets run again. So as long as we don't overestimate
+ // the wait time, we'll still end up stalling at this instruction
+ // for the correct number of cycles.
+
+ if (CyclesToWait == ~0U)
+ return 0;
+ return CyclesToWait;
+}
+
+void AMDGPUCustomBehaviour::computeWaitCnt(const InstRef &IR, unsigned &Vmcnt,
+ unsigned &Expcnt, unsigned &Lgkmcnt,
+ unsigned &Vscnt) {
+ AMDGPU::IsaVersion IV = AMDGPU::getIsaVersion(STI.getCPU());
+ const Instruction &Inst = *IR.getInstruction();
+ unsigned Opcode = Inst.getOpcode();
+
+ switch (Opcode) {
+ case AMDGPU::S_WAITCNT_EXPCNT_gfx10:
+ case AMDGPU::S_WAITCNT_LGKMCNT_gfx10:
+ case AMDGPU::S_WAITCNT_VMCNT_gfx10:
+ case AMDGPU::S_WAITCNT_VSCNT_gfx10: {
+ // Should probably be checking for nullptr
+ // here, but I'm not sure how I should handle the case
+ // where we see a nullptr.
+ const MCAOperand *OpReg = Inst.getOperand(0);
+ const MCAOperand *OpImm = Inst.getOperand(1);
+ assert(OpReg && OpReg->isReg() && "First operand should be a register.");
+ assert(OpImm && OpImm->isImm() && "Second operand should be an immediate.");
+ if (OpReg->getReg() != AMDGPU::SGPR_NULL) {
+ // Instruction is using a real register.
+ // Since we can't know what value this register will have,
+ // we can't compute what the value of this wait should be.
+ WithColor::warning() << "The register component of "
+ << MCII.getName(Opcode) << " will be completely "
+ << "ignored. So the wait may not be accurate.\n";
+ }
+ switch (Opcode) {
+ // Redundant switch so I don't have to repeat the code above
+ // for each case. There are more clever ways to avoid this
+ // extra switch and anyone can feel free to implement one of them.
+ case AMDGPU::S_WAITCNT_EXPCNT_gfx10:
+ Expcnt = OpImm->getImm();
+ break;
+ case AMDGPU::S_WAITCNT_LGKMCNT_gfx10:
+ Lgkmcnt = OpImm->getImm();
+ break;
+ case AMDGPU::S_WAITCNT_VMCNT_gfx10:
+ Vmcnt = OpImm->getImm();
+ break;
+ case AMDGPU::S_WAITCNT_VSCNT_gfx10:
+ Vscnt = OpImm->getImm();
+ break;
+ }
+ return;
+ }
+ case AMDGPU::S_WAITCNT_gfx10:
+ case AMDGPU::S_WAITCNT_gfx6_gfx7:
+ case AMDGPU::S_WAITCNT_vi:
+ unsigned WaitCnt = Inst.getOperand(0)->getImm();
+ AMDGPU::decodeWaitcnt(IV, WaitCnt, Vmcnt, Expcnt, Lgkmcnt);
+ return;
+ }
+}
+
+void AMDGPUCustomBehaviour::generateWaitCntInfo() {
+ // The core logic from this function is taken from
+ // SIInsertWaitcnts::updateEventWaitcntAfter() In that pass, the instructions
+ // that are being looked at are in the MachineInstr format, whereas we have
+ // access to the MCInst format. The side effects of this are that we can't use
+ // the mayAccessVMEMThroughFlat(Inst) or mayAccessLDSThroughFlat(Inst)
+ // functions. Therefore, we conservatively assume that these functions will
+ // return true. This may cause a few instructions to be incorrectly tagged
+ // with an extra CNT. However, these are instructions that do interact with at
+ // least one CNT so giving them an extra CNT shouldn't cause issues in most
+ // scenarios.
+ AMDGPU::IsaVersion IV = AMDGPU::getIsaVersion(STI.getCPU());
+ InstrWaitCntInfo.resize(SrcMgr.size());
+
+ int Index = 0;
+ for (auto I = SrcMgr.begin(), E = SrcMgr.end(); I != E; ++I, ++Index) {
+ const std::unique_ptr<Instruction> &Inst = *I;
+ unsigned Opcode = Inst->getOpcode();
+ const MCInstrDesc &MCID = MCII.get(Opcode);
+ if ((MCID.TSFlags & SIInstrFlags::DS) &&
+ (MCID.TSFlags & SIInstrFlags::LGKM_CNT)) {
+ InstrWaitCntInfo[Index].LgkmCnt = true;
+ if (isAlwaysGDS(Opcode) || hasModifiersSet(Inst, AMDGPU::OpName::gds))
+ InstrWaitCntInfo[Index].ExpCnt = true;
+ } else if (MCID.TSFlags & SIInstrFlags::FLAT) {
+ // We conservatively assume that mayAccessVMEMThroughFlat(Inst)
+ // and mayAccessLDSThroughFlat(Inst) would both return true for this
+ // instruction. We have to do this because those functions use
+ // information about the memory operands that we don't have access to.
+ InstrWaitCntInfo[Index].LgkmCnt = true;
+ if (!STI.hasFeature(AMDGPU::FeatureVscnt))
+ InstrWaitCntInfo[Index].VmCnt = true;
+ else if (MCID.mayLoad() && !(MCID.TSFlags & SIInstrFlags::IsAtomicNoRet))
+ InstrWaitCntInfo[Index].VmCnt = true;
+ else
+ InstrWaitCntInfo[Index].VsCnt = true;
+ } else if (isVMEM(MCID) && !AMDGPU::getMUBUFIsBufferInv(Opcode)) {
+ if (!STI.hasFeature(AMDGPU::FeatureVscnt))
+ InstrWaitCntInfo[Index].VmCnt = true;
+ else if ((MCID.mayLoad() &&
+ !(MCID.TSFlags & SIInstrFlags::IsAtomicNoRet)) ||
+ ((MCID.TSFlags & SIInstrFlags::MIMG) && !MCID.mayLoad() &&
+ !MCID.mayStore()))
+ InstrWaitCntInfo[Index].VmCnt = true;
+ else if (MCID.mayStore())
+ InstrWaitCntInfo[Index].VsCnt = true;
+
+ // (IV.Major < 7) is meant to represent
+ // GCNTarget.vmemWriteNeedsExpWaitcnt()
+ // which is defined as
+ // { return getGeneration() < SEA_ISLANDS; }
+ if (IV.Major < 7 &&
+ (MCID.mayStore() || (MCID.TSFlags & SIInstrFlags::IsAtomicRet)))
+ InstrWaitCntInfo[Index].ExpCnt = true;
+ } else if (MCID.TSFlags & SIInstrFlags::SMRD) {
+ InstrWaitCntInfo[Index].LgkmCnt = true;
+ } else if (MCID.TSFlags & SIInstrFlags::EXP) {
+ InstrWaitCntInfo[Index].ExpCnt = true;
+ } else {
+ switch (Opcode) {
+ case AMDGPU::S_SENDMSG:
+ case AMDGPU::S_SENDMSGHALT:
+ case AMDGPU::S_MEMTIME:
+ case AMDGPU::S_MEMREALTIME:
+ InstrWaitCntInfo[Index].LgkmCnt = true;
+ break;
+ }
+ }
+ }
+}
+
+// taken from SIInstrInfo::isVMEM()
+bool AMDGPUCustomBehaviour::isVMEM(const MCInstrDesc &MCID) {
+ return MCID.TSFlags & SIInstrFlags::MUBUF ||
+ MCID.TSFlags & SIInstrFlags::MTBUF ||
+ MCID.TSFlags & SIInstrFlags::MIMG;
+}
+
+// taken from SIInstrInfo::hasModifiersSet()
+bool AMDGPUCustomBehaviour::hasModifiersSet(
+ const std::unique_ptr<Instruction> &Inst, unsigned OpName) const {
+ int Idx = AMDGPU::getNamedOperandIdx(Inst->getOpcode(), OpName);
+ if (Idx == -1)
+ return false;
+
+ const MCAOperand *Op = Inst->getOperand(Idx);
+ if (Op == nullptr || !Op->isImm() || !Op->getImm())
+ return false;
+
+ return true;
+}
+
+// taken from SIInstrInfo::isAlwaysGDS()
+bool AMDGPUCustomBehaviour::isAlwaysGDS(uint16_t Opcode) const {
+ return Opcode == AMDGPU::DS_ORDERED_COUNT || Opcode == AMDGPU::DS_GWS_INIT ||
+ Opcode == AMDGPU::DS_GWS_SEMA_V || Opcode == AMDGPU::DS_GWS_SEMA_BR ||
+ Opcode == AMDGPU::DS_GWS_SEMA_P ||
+ Opcode == AMDGPU::DS_GWS_SEMA_RELEASE_ALL ||
+ Opcode == AMDGPU::DS_GWS_BARRIER;
+}
+
} // namespace mca
} // namespace llvm
namespace mca {
class AMDGPUInstrPostProcess : public InstrPostProcess {
+ void processWaitCnt(std::unique_ptr<Instruction> &Inst, const MCInst &MCI);
+
public:
AMDGPUInstrPostProcess(const MCSubtargetInfo &STI, const MCInstrInfo &MCII)
: InstrPostProcess(STI, MCII) {}
~AMDGPUInstrPostProcess() {}
void postProcessInstruction(std::unique_ptr<Instruction> &Inst,
- const MCInst &MCI) override {}
+ const MCInst &MCI) override;
+};
+
+struct WaitCntInfo {
+ bool VmCnt = false;
+ bool ExpCnt = false;
+ bool LgkmCnt = false;
+ bool VsCnt = false;
};
class AMDGPUCustomBehaviour : public CustomBehaviour {
+ /// Whenever MCA would like to dispatch an s_waitcnt instructions,
+ /// we must check all the instruction that are still executing to see if
+ /// they modify the same CNT as we need to wait for. This vector
+ /// gets built in the constructor and contains 1 WaitCntInfo struct
+ /// for each instruction within the SrcManager. Each element
+ /// tells us which CNTs that instruction may interact with.
+ /// We conservatively assume some instructions interact with more
+ /// CNTs than they do in reality, so we will occasionally wait
+ /// longer than necessary, but we shouldn't ever wait for shorter.
+ std::vector<WaitCntInfo> InstrWaitCntInfo;
+
+ /// This method gets called from the constructor and is
+ /// where we setup the InstrWaitCntInfo vector.
+ /// The core logic for determining which CNTs an instruction
+ /// interacts with is taken from SIInsertWaitcnts::updateEventWaitcntAfter().
+ /// Unfortunately, some of the logic from that function is not avalable to us
+ /// in this scope so we conservatively end up assuming that some
+ /// instructions interact with more CNTs than they do in reality.
+ void generateWaitCntInfo();
+ /// Helper function used in generateWaitCntInfo()
+ bool hasModifiersSet(const std::unique_ptr<Instruction> &Inst,
+ unsigned OpName) const;
+ /// Helper function used in generateWaitCntInfo()
+ bool isAlwaysGDS(uint16_t Opcode) const;
+ /// Helper function used in generateWaitCntInfo()
+ bool isVMEM(const MCInstrDesc &MCID);
+ /// This method gets called from checkCustomHazard when mca is attempting to
+ /// dispatch an s_waitcnt instruction (or one of its variants). The method
+ /// looks at each of the instructions that are still executing in the pipeline
+ /// to determine if the waitcnt should force a wait.
+ unsigned handleWaitCnt(ArrayRef<InstRef> IssuedInst, const InstRef &IR);
+ /// Based on the type of s_waitcnt instruction we are looking at, and what its
+ /// operands are, this method will set the values for each of the cnt
+ /// references provided as arguments.
+ void computeWaitCnt(const InstRef &IR, unsigned &Vmcnt, unsigned &Expcnt,
+ unsigned &Lgkmcnt, unsigned &Vscnt);
+
public:
AMDGPUCustomBehaviour(const MCSubtargetInfo &STI, const SourceMgr &SrcMgr,
const MCInstrInfo &MCII);