Skip to content

[AMDGPU] Combine or remove redundant waitcnts at the end of each MBB #87539

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom
Merged
merged 2 commits into from
Apr 4, 2024
Merged

[AMDGPU] Combine or remove redundant waitcnts at the end of each MBB #87539

merged 2 commits into from
Apr 4, 2024

Conversation

jayfoad
Copy link
Contributor

@jayfoad jayfoad commented Apr 3, 2024

Call generateWaitcnt unconditionally at the end of
SIInsertWaitcnts::insertWaitcntInBlock. Even if we don't need to
generate a new waitcnt instruction it has the effect of combining or
removing redundant waitcnts that were already present. Tests show
various small improvements in waitcnt placement.

@jayfoad
[AMDGPU] Combine or remove redundant waitcnts at the end of each MBB
5dbe01c 
Call generateWaitcnt unconditionally at the end of
SIInsertWaitcnts::insertWaitcntInBlock. Even if we don't need to
generate a new waitcnt instruction it has the effect of combining or
removing redundant waitcnts that were already present. Tests show
various small improvements in waitcnt placement.
@llvmbot
Copy link
Member

llvmbot commented Apr 3, 2024

@llvm/pr-subscribers-backend-amdgpu

Author: Jay Foad (jayfoad)

Changes

Call generateWaitcnt unconditionally at the end of
SIInsertWaitcnts::insertWaitcntInBlock. Even if we don't need to
generate a new waitcnt instruction it has the effect of combining or
removing redundant waitcnts that were already present. Tests show
various small improvements in waitcnt placement.

Patch is 49.35 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/87539.diff

13 Files Affected:

  • (modified) llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp (+15-30)
  • (modified) llvm/test/CodeGen/AMDGPU/GlobalISel/divergent-control-flow.ll (-4)
  • (modified) llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.is.private.ll (-2)
  • (modified) llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.is.shared.ll (-2)
  • (modified) llvm/test/CodeGen/AMDGPU/atomic_optimizations_local_pointer.ll (+40-46)
  • (modified) llvm/test/CodeGen/AMDGPU/atomicrmw-expand.ll (+1-1)
  • (modified) llvm/test/CodeGen/AMDGPU/cgp-addressing-modes-gfx908.ll (+1-1)
  • (modified) llvm/test/CodeGen/AMDGPU/extract-subvector-16bit.ll (-11)
  • (modified) llvm/test/CodeGen/AMDGPU/function-args.ll (-1)
  • (modified) llvm/test/CodeGen/AMDGPU/lds-global-non-entry-func.ll (+12-16)
  • (modified) llvm/test/CodeGen/AMDGPU/local-atomics-fp.ll (+2-6)
  • (modified) llvm/test/CodeGen/AMDGPU/skip-if-dead.ll (-4)
  • (modified) llvm/test/CodeGen/AMDGPU/transform-block-with-return-to-epilog.ll (+2-2)
diff --git a/llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp b/llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp
index 27621906e4c5fd..1f44ed8c628d94 100644
--- a/llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp
+++ b/llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp
@@ -708,9 +708,6 @@ class SIInsertWaitcnts : public MachineFunctionPass {
                                  WaitcntBrackets &ScoreBrackets,
                                  MachineInstr *OldWaitcntInstr,
                                  bool FlushVmCnt);
-  bool generateWaitcntBlockEnd(MachineBasicBlock &Block,
-                               WaitcntBrackets &ScoreBrackets,
-                               MachineInstr *OldWaitcntInstr);
   bool generateWaitcnt(AMDGPU::Waitcnt Wait,
                        MachineBasicBlock::instr_iterator It,
                        MachineBasicBlock &Block, WaitcntBrackets &ScoreBrackets,
@@ -1902,31 +1899,6 @@ bool SIInsertWaitcnts::generateWaitcntInstBefore(MachineInstr &MI,
                          OldWaitcntInstr);
 }
 
-// Add a waitcnt to flush the LOADcnt, SAMPLEcnt and BVHcnt counters at the
-// end of the given block if needed.
-bool SIInsertWaitcnts::generateWaitcntBlockEnd(MachineBasicBlock &Block,
-                                               WaitcntBrackets &ScoreBrackets,
-                                               MachineInstr *OldWaitcntInstr) {
-  AMDGPU::Waitcnt Wait;
-
-  unsigned LoadCntPending = ScoreBrackets.hasPendingEvent(LOAD_CNT);
-  unsigned SampleCntPending = ScoreBrackets.hasPendingEvent(SAMPLE_CNT);
-  unsigned BvhCntPending = ScoreBrackets.hasPendingEvent(BVH_CNT);
-
-  if (LoadCntPending == 0 && SampleCntPending == 0 && BvhCntPending == 0)
-    return false;
-
-  if (LoadCntPending != 0)
-    Wait.LoadCnt = 0;
-  if (SampleCntPending != 0)
-    Wait.SampleCnt = 0;
-  if (BvhCntPending != 0)
-    Wait.BvhCnt = 0;
-
-  return generateWaitcnt(Wait, Block.instr_end(), Block, ScoreBrackets,
-                         OldWaitcntInstr);
-}
-
 bool SIInsertWaitcnts::generateWaitcnt(AMDGPU::Waitcnt Wait,
                                        MachineBasicBlock::instr_iterator It,
                                        MachineBasicBlock &Block,
@@ -2355,9 +2327,22 @@ bool SIInsertWaitcnts::insertWaitcntInBlock(MachineFunction &MF,
     ++Iter;
   }
 
+  // Flush the LOADcnt, SAMPLEcnt and BVHcnt counters at the end of the block if
+  // needed.
+  AMDGPU::Waitcnt Wait;
   if (Block.getFirstTerminator() == Block.end() &&
-      isPreheaderToFlush(Block, ScoreBrackets))
-    Modified |= generateWaitcntBlockEnd(Block, ScoreBrackets, OldWaitcntInstr);
+      isPreheaderToFlush(Block, ScoreBrackets)) {
+    if (ScoreBrackets.hasPendingEvent(LOAD_CNT))
+      Wait.LoadCnt = 0;
+    if (ScoreBrackets.hasPendingEvent(SAMPLE_CNT))
+      Wait.SampleCnt = 0;
+    if (ScoreBrackets.hasPendingEvent(BVH_CNT))
+      Wait.BvhCnt = 0;
+  }
+
+  // Combine or remove any redundant waitcnts ar the end of the block.
+  Modified |= generateWaitcnt(Wait, Block.instr_end(), Block, ScoreBrackets,
+                              OldWaitcntInstr);
 
   return Modified;
 }
diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/divergent-control-flow.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/divergent-control-flow.ll
index c25b0f21282663..78d908455e019b 100644
--- a/llvm/test/CodeGen/AMDGPU/GlobalISel/divergent-control-flow.ll
+++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/divergent-control-flow.ll
@@ -16,7 +16,6 @@ define i32 @divergent_if_swap_brtarget_order0(i32 %value) {
 ; CHECK-NEXT:    s_waitcnt vmcnt(0)
 ; CHECK-NEXT:  .LBB0_2: ; %endif
 ; CHECK-NEXT:    s_or_b64 exec, exec, s[4:5]
-; CHECK-NEXT:    s_waitcnt vmcnt(0)
 ; CHECK-NEXT:    s_setpc_b64 s[30:31]
 entry:
   %c = icmp ne i32 %value, 0
@@ -44,7 +43,6 @@ define i32 @divergent_if_swap_brtarget_order1(i32 %value) {
 ; CHECK-NEXT:    s_waitcnt vmcnt(0)
 ; CHECK-NEXT:  .LBB1_2: ; %endif
 ; CHECK-NEXT:    s_or_b64 exec, exec, s[4:5]
-; CHECK-NEXT:    s_waitcnt vmcnt(0)
 ; CHECK-NEXT:    s_setpc_b64 s[30:31]
 entry:
   %c = icmp ne i32 %value, 0
@@ -74,7 +72,6 @@ define i32 @divergent_if_nonboolean_condition0(i32 %value) {
 ; CHECK-NEXT:    s_waitcnt vmcnt(0)
 ; CHECK-NEXT:  .LBB2_2: ; %endif
 ; CHECK-NEXT:    s_or_b64 exec, exec, s[4:5]
-; CHECK-NEXT:    s_waitcnt vmcnt(0)
 ; CHECK-NEXT:    s_setpc_b64 s[30:31]
 entry:
   %c = trunc i32 %value to i1
@@ -106,7 +103,6 @@ define i32 @divergent_if_nonboolean_condition1(ptr addrspace(1) %ptr) {
 ; CHECK-NEXT:    s_waitcnt vmcnt(0)
 ; CHECK-NEXT:  .LBB3_2: ; %endif
 ; CHECK-NEXT:    s_or_b64 exec, exec, s[4:5]
-; CHECK-NEXT:    s_waitcnt vmcnt(0)
 ; CHECK-NEXT:    s_setpc_b64 s[30:31]
 entry:
   %value = load i32, ptr addrspace(1) %ptr
diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.is.private.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.is.private.ll
index 303dc46e2c8845..5c22d5bdcf7449 100644
--- a/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.is.private.ll
+++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.is.private.ll
@@ -131,8 +131,6 @@ define amdgpu_kernel void @is_private_sgpr(ptr %ptr) {
 ; GFX11-NEXT:    global_store_b32 v[0:1], v0, off dlc
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-NEXT:  .LBB1_2: ; %bb1
-; GFX11-NEXT:    s_nop 0
-; GFX11-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-NEXT:    s_endpgm
   %val = call i1 @llvm.amdgcn.is.private(ptr %ptr)
   br i1 %val, label %bb0, label %bb1
diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.is.shared.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.is.shared.ll
index 63702d2587574b..e005c38355a3ce 100644
--- a/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.is.shared.ll
+++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.is.shared.ll
@@ -131,8 +131,6 @@ define amdgpu_kernel void @is_local_sgpr(ptr %ptr) {
 ; GFX11-NEXT:    global_store_b32 v[0:1], v0, off dlc
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-NEXT:  .LBB1_2: ; %bb1
-; GFX11-NEXT:    s_nop 0
-; GFX11-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-NEXT:    s_endpgm
   %val = call i1 @llvm.amdgcn.is.shared(ptr %ptr)
   br i1 %val, label %bb0, label %bb1
diff --git a/llvm/test/CodeGen/AMDGPU/atomic_optimizations_local_pointer.ll b/llvm/test/CodeGen/AMDGPU/atomic_optimizations_local_pointer.ll
index 352adacbda3e4b..af6f69130910d0 100644
--- a/llvm/test/CodeGen/AMDGPU/atomic_optimizations_local_pointer.ll
+++ b/llvm/test/CodeGen/AMDGPU/atomic_optimizations_local_pointer.ll
@@ -39,9 +39,9 @@ define amdgpu_kernel void @add_i32_constant(ptr addrspace(1) %out) {
 ; GFX7LESS-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
 ; GFX7LESS-NEXT:    s_mov_b32 s3, 0xf000
 ; GFX7LESS-NEXT:    s_mov_b32 s2, -1
-; GFX7LESS-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX7LESS-NEXT:    v_readfirstlane_b32 s4, v1
 ; GFX7LESS-NEXT:    v_mad_u32_u24 v0, v0, 5, s4
+; GFX7LESS-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX7LESS-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 ; GFX7LESS-NEXT:    s_endpgm
 ;
@@ -65,11 +65,11 @@ define amdgpu_kernel void @add_i32_constant(ptr addrspace(1) %out) {
 ; GFX8-NEXT:  .LBB0_2:
 ; GFX8-NEXT:    s_or_b64 exec, exec, s[2:3]
 ; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX8-NEXT:    v_readfirstlane_b32 s4, v1
 ; GFX8-NEXT:    s_mov_b32 s3, 0xf000
 ; GFX8-NEXT:    s_mov_b32 s2, -1
 ; GFX8-NEXT:    v_mad_u32_u24 v0, v0, 5, s4
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX8-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 ; GFX8-NEXT:    s_endpgm
 ;
@@ -92,11 +92,11 @@ define amdgpu_kernel void @add_i32_constant(ptr addrspace(1) %out) {
 ; GFX9-NEXT:  .LBB0_2:
 ; GFX9-NEXT:    s_or_b64 exec, exec, s[2:3]
 ; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    v_readfirstlane_b32 s4, v1
 ; GFX9-NEXT:    s_mov_b32 s3, 0xf000
 ; GFX9-NEXT:    s_mov_b32 s2, -1
 ; GFX9-NEXT:    v_mad_u32_u24 v0, v0, 5, s4
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 ; GFX9-NEXT:    s_endpgm
 ;
@@ -253,8 +253,8 @@ define amdgpu_kernel void @add_i32_uniform(ptr addrspace(1) %out, i32 %additive)
 ; GFX7LESS-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
 ; GFX7LESS-NEXT:    s_mov_b32 s3, 0xf000
 ; GFX7LESS-NEXT:    s_mov_b32 s2, -1
-; GFX7LESS-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX7LESS-NEXT:    v_readfirstlane_b32 s4, v1
+; GFX7LESS-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX7LESS-NEXT:    v_mul_lo_u32 v0, s6, v0
 ; GFX7LESS-NEXT:    v_add_i32_e32 v0, vcc, s4, v0
 ; GFX7LESS-NEXT:    buffer_store_dword v0, off, s[0:3], 0
@@ -504,11 +504,11 @@ define amdgpu_kernel void @add_i32_varying(ptr addrspace(1) %out) {
 ; GFX8-NEXT:  .LBB2_4:
 ; GFX8-NEXT:    s_or_b64 exec, exec, s[2:3]
 ; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX8-NEXT:    v_readfirstlane_b32 s4, v0
 ; GFX8-NEXT:    s_mov_b32 s3, 0xf000
 ; GFX8-NEXT:    s_mov_b32 s2, -1
 ; GFX8-NEXT:    v_add_u32_e32 v0, vcc, s4, v1
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX8-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 ; GFX8-NEXT:    s_endpgm
 ;
@@ -544,11 +544,11 @@ define amdgpu_kernel void @add_i32_varying(ptr addrspace(1) %out) {
 ; GFX9-NEXT:  .LBB2_4:
 ; GFX9-NEXT:    s_or_b64 exec, exec, s[2:3]
 ; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    v_readfirstlane_b32 s4, v0
 ; GFX9-NEXT:    s_mov_b32 s3, 0xf000
 ; GFX9-NEXT:    s_mov_b32 s2, -1
 ; GFX9-NEXT:    v_add_u32_e32 v0, s4, v1
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 ; GFX9-NEXT:    s_endpgm
 ;
@@ -944,7 +944,6 @@ define amdgpu_kernel void @add_i64_constant(ptr addrspace(1) %out) {
 ; GFX7LESS-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
 ; GFX7LESS-NEXT:    s_mov_b32 s3, 0xf000
 ; GFX7LESS-NEXT:    s_mov_b32 s2, -1
-; GFX7LESS-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX7LESS-NEXT:    v_readfirstlane_b32 s4, v0
 ; GFX7LESS-NEXT:    v_readfirstlane_b32 s5, v1
 ; GFX7LESS-NEXT:    v_mul_hi_u32_u24_e32 v1, 5, v2
@@ -952,6 +951,7 @@ define amdgpu_kernel void @add_i64_constant(ptr addrspace(1) %out) {
 ; GFX7LESS-NEXT:    v_mov_b32_e32 v2, s5
 ; GFX7LESS-NEXT:    v_add_i32_e32 v0, vcc, s4, v0
 ; GFX7LESS-NEXT:    v_addc_u32_e32 v1, vcc, v2, v1, vcc
+; GFX7LESS-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX7LESS-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
 ; GFX7LESS-NEXT:    s_endpgm
 ;
@@ -974,7 +974,6 @@ define amdgpu_kernel void @add_i64_constant(ptr addrspace(1) %out) {
 ; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX8-NEXT:  .LBB4_2:
 ; GFX8-NEXT:    s_or_b64 exec, exec, s[2:3]
-; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX8-NEXT:    v_readfirstlane_b32 s2, v0
 ; GFX8-NEXT:    v_readfirstlane_b32 s3, v1
 ; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
@@ -1006,7 +1005,6 @@ define amdgpu_kernel void @add_i64_constant(ptr addrspace(1) %out) {
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:  .LBB4_2:
 ; GFX9-NEXT:    s_or_b64 exec, exec, s[2:3]
-; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    v_readfirstlane_b32 s2, v0
 ; GFX9-NEXT:    v_readfirstlane_b32 s3, v1
 ; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
@@ -1219,11 +1217,11 @@ define amdgpu_kernel void @add_i64_uniform(ptr addrspace(1) %out, i64 %additive)
 ; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX8-NEXT:  .LBB5_2:
 ; GFX8-NEXT:    s_or_b64 exec, exec, s[4:5]
-; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX8-NEXT:    v_readfirstlane_b32 s4, v0
 ; GFX8-NEXT:    v_readfirstlane_b32 s5, v1
 ; GFX8-NEXT:    v_mov_b32_e32 v0, s4
 ; GFX8-NEXT:    v_mov_b32_e32 v1, s5
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX8-NEXT:    v_mul_lo_u32 v3, s3, v2
 ; GFX8-NEXT:    v_mad_u64_u32 v[0:1], s[2:3], s2, v2, v[0:1]
 ; GFX8-NEXT:    s_mov_b32 s7, 0xf000
@@ -1258,11 +1256,11 @@ define amdgpu_kernel void @add_i64_uniform(ptr addrspace(1) %out, i64 %additive)
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:  .LBB5_2:
 ; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    v_readfirstlane_b32 s4, v0
 ; GFX9-NEXT:    v_readfirstlane_b32 s5, v1
 ; GFX9-NEXT:    v_mov_b32_e32 v0, s4
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s5
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    v_mad_u64_u32 v[0:1], s[4:5], s2, v2, v[0:1]
 ; GFX9-NEXT:    s_mov_b32 s7, 0xf000
 ; GFX9-NEXT:    s_mov_b32 s6, -1
@@ -1530,10 +1528,10 @@ define amdgpu_kernel void @sub_i32_constant(ptr addrspace(1) %out) {
 ; GFX7LESS-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
 ; GFX7LESS-NEXT:    s_mov_b32 s3, 0xf000
 ; GFX7LESS-NEXT:    s_mov_b32 s2, -1
-; GFX7LESS-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX7LESS-NEXT:    v_readfirstlane_b32 s4, v1
 ; GFX7LESS-NEXT:    v_mul_u32_u24_e32 v0, 5, v0
 ; GFX7LESS-NEXT:    v_sub_i32_e32 v0, vcc, s4, v0
+; GFX7LESS-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX7LESS-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 ; GFX7LESS-NEXT:    s_endpgm
 ;
@@ -1557,12 +1555,12 @@ define amdgpu_kernel void @sub_i32_constant(ptr addrspace(1) %out) {
 ; GFX8-NEXT:  .LBB7_2:
 ; GFX8-NEXT:    s_or_b64 exec, exec, s[2:3]
 ; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX8-NEXT:    v_readfirstlane_b32 s4, v1
 ; GFX8-NEXT:    v_mul_u32_u24_e32 v0, 5, v0
 ; GFX8-NEXT:    s_mov_b32 s3, 0xf000
 ; GFX8-NEXT:    s_mov_b32 s2, -1
 ; GFX8-NEXT:    v_sub_u32_e32 v0, vcc, s4, v0
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX8-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 ; GFX8-NEXT:    s_endpgm
 ;
@@ -1585,12 +1583,12 @@ define amdgpu_kernel void @sub_i32_constant(ptr addrspace(1) %out) {
 ; GFX9-NEXT:  .LBB7_2:
 ; GFX9-NEXT:    s_or_b64 exec, exec, s[2:3]
 ; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    v_readfirstlane_b32 s4, v1
 ; GFX9-NEXT:    v_mul_u32_u24_e32 v0, 5, v0
 ; GFX9-NEXT:    s_mov_b32 s3, 0xf000
 ; GFX9-NEXT:    s_mov_b32 s2, -1
 ; GFX9-NEXT:    v_sub_u32_e32 v0, s4, v0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 ; GFX9-NEXT:    s_endpgm
 ;
@@ -1751,8 +1749,8 @@ define amdgpu_kernel void @sub_i32_uniform(ptr addrspace(1) %out, i32 %subitive)
 ; GFX7LESS-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
 ; GFX7LESS-NEXT:    s_mov_b32 s3, 0xf000
 ; GFX7LESS-NEXT:    s_mov_b32 s2, -1
-; GFX7LESS-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX7LESS-NEXT:    v_readfirstlane_b32 s4, v1
+; GFX7LESS-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX7LESS-NEXT:    v_mul_lo_u32 v0, s6, v0
 ; GFX7LESS-NEXT:    v_sub_i32_e32 v0, vcc, s4, v0
 ; GFX7LESS-NEXT:    buffer_store_dword v0, off, s[0:3], 0
@@ -2006,11 +2004,11 @@ define amdgpu_kernel void @sub_i32_varying(ptr addrspace(1) %out) {
 ; GFX8-NEXT:  .LBB9_4:
 ; GFX8-NEXT:    s_or_b64 exec, exec, s[2:3]
 ; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX8-NEXT:    v_readfirstlane_b32 s4, v0
 ; GFX8-NEXT:    s_mov_b32 s3, 0xf000
 ; GFX8-NEXT:    s_mov_b32 s2, -1
 ; GFX8-NEXT:    v_sub_u32_e32 v0, vcc, s4, v1
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX8-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 ; GFX8-NEXT:    s_endpgm
 ;
@@ -2046,11 +2044,11 @@ define amdgpu_kernel void @sub_i32_varying(ptr addrspace(1) %out) {
 ; GFX9-NEXT:  .LBB9_4:
 ; GFX9-NEXT:    s_or_b64 exec, exec, s[2:3]
 ; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    v_readfirstlane_b32 s4, v0
 ; GFX9-NEXT:    s_mov_b32 s3, 0xf000
 ; GFX9-NEXT:    s_mov_b32 s2, -1
 ; GFX9-NEXT:    v_sub_u32_e32 v0, s4, v1
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 ; GFX9-NEXT:    s_endpgm
 ;
@@ -2446,7 +2444,6 @@ define amdgpu_kernel void @sub_i64_constant(ptr addrspace(1) %out) {
 ; GFX7LESS-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
 ; GFX7LESS-NEXT:    s_mov_b32 s3, 0xf000
 ; GFX7LESS-NEXT:    s_mov_b32 s2, -1
-; GFX7LESS-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX7LESS-NEXT:    v_readfirstlane_b32 s4, v0
 ; GFX7LESS-NEXT:    v_readfirstlane_b32 s5, v1
 ; GFX7LESS-NEXT:    v_mul_hi_u32_u24_e32 v1, 5, v2
@@ -2454,6 +2451,7 @@ define amdgpu_kernel void @sub_i64_constant(ptr addrspace(1) %out) {
 ; GFX7LESS-NEXT:    v_mov_b32_e32 v2, s5
 ; GFX7LESS-NEXT:    v_sub_i32_e32 v0, vcc, s4, v0
 ; GFX7LESS-NEXT:    v_subb_u32_e32 v1, vcc, v2, v1, vcc
+; GFX7LESS-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX7LESS-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
 ; GFX7LESS-NEXT:    s_endpgm
 ;
@@ -2477,7 +2475,6 @@ define amdgpu_kernel void @sub_i64_constant(ptr addrspace(1) %out) {
 ; GFX8-NEXT:  .LBB11_2:
 ; GFX8-NEXT:    s_or_b64 exec, exec, s[2:3]
 ; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX8-NEXT:    v_readfirstlane_b32 s4, v0
 ; GFX8-NEXT:    v_readfirstlane_b32 s5, v1
 ; GFX8-NEXT:    v_mul_u32_u24_e32 v0, 5, v2
@@ -2487,6 +2484,7 @@ define amdgpu_kernel void @sub_i64_constant(ptr addrspace(1) %out) {
 ; GFX8-NEXT:    s_mov_b32 s3, 0xf000
 ; GFX8-NEXT:    s_mov_b32 s2, -1
 ; GFX8-NEXT:    v_subb_u32_e32 v1, vcc, v2, v1, vcc
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX8-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
 ; GFX8-NEXT:    s_endpgm
 ;
@@ -2509,7 +2507,6 @@ define amdgpu_kernel void @sub_i64_constant(ptr addrspace(1) %out) {
 ; GFX9-NEXT:  .LBB11_2:
 ; GFX9-NEXT:    s_or_b64 exec, exec, s[2:3]
 ; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    v_readfirstlane_b32 s4, v0
 ; GFX9-NEXT:    v_readfirstlane_b32 s5, v1
 ; GFX9-NEXT:    v_mul_u32_u24_e32 v0, 5, v2
@@ -2519,6 +2516,7 @@ define amdgpu_kernel void @sub_i64_constant(ptr addrspace(1) %out) {
 ; GFX9-NEXT:    s_mov_b32 s3, 0xf000
 ; GFX9-NEXT:    s_mov_b32 s2, -1
 ; GFX9-NEXT:    v_subb_co_u32_e32 v1, vcc, v2, v1, vcc
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
 ; GFX9-NEXT:    s_endpgm
 ;
@@ -3081,11 +3079,11 @@ define amdgpu_kernel void @and_i32_varying(ptr addrspace(1) %out) {
 ; GFX8-NEXT:  .LBB14_4:
 ; GFX8-NEXT:    s_or_b64 exec, exec, s[2:3]
 ; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX8-NEXT:    v_readfirstlane_b32 s4, v0
 ; GFX8-NEXT:    s_mov_b32 s3, 0xf000
 ; GFX8-NEXT:    s_mov_b32 s2, -1
 ; GFX8-NEXT:    v_and_b32_e32 v0, s4, v1
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX8-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 ; GFX8-NEXT:    s_endpgm
 ;
@@ -3121,11 +3119,11 @@ define amdgpu_kernel void @and_i32_varying(ptr addrspace(1) %out) {
 ; GFX9-NEXT:  .LBB14_4:
 ; GFX9-NEXT:    s_or_b64 exec, exec, s[2:3]
 ; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    v_readfirstlane_b32 s4, v0
 ; GFX9-NEXT:    s_mov_b32 s3, 0xf000
 ; GFX9-NEXT:    s_mov_b32 s2, -1
 ; GFX9-NEXT:    v_and_b32_e32 v0, s4, v1
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 ; GFX9-NEXT:    s_endpgm
 ;
@@ -3355,11 +3353,11 @@ define amdgpu_kernel void @or_i32_varying(ptr addrspace(1) %out) {
 ; GFX8-NEXT:  .LBB15_4:
 ; GFX8-NEXT:    s_or_b64 exec, exec, s[2:3]
 ; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX8-NEXT:    v_readfirstlane_b32 s4, v0
 ; GFX8-NEXT:    s_mov_b32 s3, 0xf000
 ; GFX8-NEXT:    s_mov_b32 s2, -1
 ; GFX8-NEXT:    v_or_b32_e32 v0, s4, v1
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX8-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 ; GFX8-NEXT:    s_endpgm
 ;
@@ -3395,11 +3393,11 @@ define amdgpu_kernel void @or_i32_varying(ptr addrspace(1) %out) {
 ; GFX9-NEXT:  .LBB15_4:
 ; GFX9-NEXT:    s_or_b64 exec, exec, s[2:3]
 ; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    v_readfirstlane_b32 s4, v0
 ; GFX9-NEXT:    s_mov_b32 s3, 0xf000
 ; GFX9-NEXT:    s_mov_b32 s2, -1
 ; GFX9-NEXT:    v_or_b32_e32 v0, s4, v1
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 ; GFX9-NEXT:    s_endpgm
 ;
@@ -3629,11 +3627,11 @@ define amdgpu_kernel void @xor_i32_varying(ptr addrspace(1) %out) {
 ; GFX8-NEXT:  .LBB16_4:
 ; GFX8-NEXT:    s_or_b64 exec, exec, s[2:3]
 ; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX8-NEXT:    v_readfirstlane_b32 s4, v0
 ; GFX8-NEXT:    s_mov_b32 s3, 0xf000
 ; GFX8-NEXT:    s_mov_b32 s2, -1
 ; GFX8-NEXT:    v_xor_b32_e32 v0, s4, v1
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
 ; ...
[truncated]

@llvmbot
Copy link
Member

llvmbot commented Apr 3, 2024

@llvm/pr-subscribers-llvm-globalisel

Author: Jay Foad (jayfoad)

Changes

Call generateWaitcnt unconditionally at the end of
SIInsertWaitcnts::insertWaitcntInBlock. Even if we don't need to
generate a new waitcnt instruction it has the effect of combining or
removing redundant waitcnts that were already present. Tests show
various small improvements in waitcnt placement.

Patch is 49.35 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/87539.diff

13 Files Affected:

  • (modified) llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp (+15-30)
  • (modified) llvm/test/CodeGen/AMDGPU/GlobalISel/divergent-control-flow.ll (-4)
  • (modified) llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.is.private.ll (-2)
  • (modified) llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.is.shared.ll (-2)
  • (modified) llvm/test/CodeGen/AMDGPU/atomic_optimizations_local_pointer.ll (+40-46)
  • (modified) llvm/test/CodeGen/AMDGPU/atomicrmw-expand.ll (+1-1)
  • (modified) llvm/test/CodeGen/AMDGPU/cgp-addressing-modes-gfx908.ll (+1-1)
  • (modified) llvm/test/CodeGen/AMDGPU/extract-subvector-16bit.ll (-11)
  • (modified) llvm/test/CodeGen/AMDGPU/function-args.ll (-1)
  • (modified) llvm/test/CodeGen/AMDGPU/lds-global-non-entry-func.ll (+12-16)
  • (modified) llvm/test/CodeGen/AMDGPU/local-atomics-fp.ll (+2-6)
  • (modified) llvm/test/CodeGen/AMDGPU/skip-if-dead.ll (-4)
  • (modified) llvm/test/CodeGen/AMDGPU/transform-block-with-return-to-epilog.ll (+2-2)
diff --git a/llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp b/llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp
index 27621906e4c5fd..1f44ed8c628d94 100644
--- a/llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp
+++ b/llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp
@@ -708,9 +708,6 @@ class SIInsertWaitcnts : public MachineFunctionPass {
                                  WaitcntBrackets &ScoreBrackets,
                                  MachineInstr *OldWaitcntInstr,
                                  bool FlushVmCnt);
-  bool generateWaitcntBlockEnd(MachineBasicBlock &Block,
-                               WaitcntBrackets &ScoreBrackets,
-                               MachineInstr *OldWaitcntInstr);
   bool generateWaitcnt(AMDGPU::Waitcnt Wait,
                        MachineBasicBlock::instr_iterator It,
                        MachineBasicBlock &Block, WaitcntBrackets &ScoreBrackets,
@@ -1902,31 +1899,6 @@ bool SIInsertWaitcnts::generateWaitcntInstBefore(MachineInstr &MI,
                          OldWaitcntInstr);
 }
 
-// Add a waitcnt to flush the LOADcnt, SAMPLEcnt and BVHcnt counters at the
-// end of the given block if needed.
-bool SIInsertWaitcnts::generateWaitcntBlockEnd(MachineBasicBlock &Block,
-                                               WaitcntBrackets &ScoreBrackets,
-                                               MachineInstr *OldWaitcntInstr) {
-  AMDGPU::Waitcnt Wait;
-
-  unsigned LoadCntPending = ScoreBrackets.hasPendingEvent(LOAD_CNT);
-  unsigned SampleCntPending = ScoreBrackets.hasPendingEvent(SAMPLE_CNT);
-  unsigned BvhCntPending = ScoreBrackets.hasPendingEvent(BVH_CNT);
-
-  if (LoadCntPending == 0 && SampleCntPending == 0 && BvhCntPending == 0)
-    return false;
-
-  if (LoadCntPending != 0)
-    Wait.LoadCnt = 0;
-  if (SampleCntPending != 0)
-    Wait.SampleCnt = 0;
-  if (BvhCntPending != 0)
-    Wait.BvhCnt = 0;
-
-  return generateWaitcnt(Wait, Block.instr_end(), Block, ScoreBrackets,
-                         OldWaitcntInstr);
-}
-
 bool SIInsertWaitcnts::generateWaitcnt(AMDGPU::Waitcnt Wait,
                                        MachineBasicBlock::instr_iterator It,
                                        MachineBasicBlock &Block,
@@ -2355,9 +2327,22 @@ bool SIInsertWaitcnts::insertWaitcntInBlock(MachineFunction &MF,
     ++Iter;
   }
 
+  // Flush the LOADcnt, SAMPLEcnt and BVHcnt counters at the end of the block if
+  // needed.
+  AMDGPU::Waitcnt Wait;
   if (Block.getFirstTerminator() == Block.end() &&
-      isPreheaderToFlush(Block, ScoreBrackets))
-    Modified |= generateWaitcntBlockEnd(Block, ScoreBrackets, OldWaitcntInstr);
+      isPreheaderToFlush(Block, ScoreBrackets)) {
+    if (ScoreBrackets.hasPendingEvent(LOAD_CNT))
+      Wait.LoadCnt = 0;
+    if (ScoreBrackets.hasPendingEvent(SAMPLE_CNT))
+      Wait.SampleCnt = 0;
+    if (ScoreBrackets.hasPendingEvent(BVH_CNT))
+      Wait.BvhCnt = 0;
+  }
+
+  // Combine or remove any redundant waitcnts ar the end of the block.
+  Modified |= generateWaitcnt(Wait, Block.instr_end(), Block, ScoreBrackets,
+                              OldWaitcntInstr);
 
   return Modified;
 }
diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/divergent-control-flow.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/divergent-control-flow.ll
index c25b0f21282663..78d908455e019b 100644
--- a/llvm/test/CodeGen/AMDGPU/GlobalISel/divergent-control-flow.ll
+++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/divergent-control-flow.ll
@@ -16,7 +16,6 @@ define i32 @divergent_if_swap_brtarget_order0(i32 %value) {
 ; CHECK-NEXT:    s_waitcnt vmcnt(0)
 ; CHECK-NEXT:  .LBB0_2: ; %endif
 ; CHECK-NEXT:    s_or_b64 exec, exec, s[4:5]
-; CHECK-NEXT:    s_waitcnt vmcnt(0)
 ; CHECK-NEXT:    s_setpc_b64 s[30:31]
 entry:
   %c = icmp ne i32 %value, 0
@@ -44,7 +43,6 @@ define i32 @divergent_if_swap_brtarget_order1(i32 %value) {
 ; CHECK-NEXT:    s_waitcnt vmcnt(0)
 ; CHECK-NEXT:  .LBB1_2: ; %endif
 ; CHECK-NEXT:    s_or_b64 exec, exec, s[4:5]
-; CHECK-NEXT:    s_waitcnt vmcnt(0)
 ; CHECK-NEXT:    s_setpc_b64 s[30:31]
 entry:
   %c = icmp ne i32 %value, 0
@@ -74,7 +72,6 @@ define i32 @divergent_if_nonboolean_condition0(i32 %value) {
 ; CHECK-NEXT:    s_waitcnt vmcnt(0)
 ; CHECK-NEXT:  .LBB2_2: ; %endif
 ; CHECK-NEXT:    s_or_b64 exec, exec, s[4:5]
-; CHECK-NEXT:    s_waitcnt vmcnt(0)
 ; CHECK-NEXT:    s_setpc_b64 s[30:31]
 entry:
   %c = trunc i32 %value to i1
@@ -106,7 +103,6 @@ define i32 @divergent_if_nonboolean_condition1(ptr addrspace(1) %ptr) {
 ; CHECK-NEXT:    s_waitcnt vmcnt(0)
 ; CHECK-NEXT:  .LBB3_2: ; %endif
 ; CHECK-NEXT:    s_or_b64 exec, exec, s[4:5]
-; CHECK-NEXT:    s_waitcnt vmcnt(0)
 ; CHECK-NEXT:    s_setpc_b64 s[30:31]
 entry:
   %value = load i32, ptr addrspace(1) %ptr
diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.is.private.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.is.private.ll
index 303dc46e2c8845..5c22d5bdcf7449 100644
--- a/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.is.private.ll
+++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.is.private.ll
@@ -131,8 +131,6 @@ define amdgpu_kernel void @is_private_sgpr(ptr %ptr) {
 ; GFX11-NEXT:    global_store_b32 v[0:1], v0, off dlc
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-NEXT:  .LBB1_2: ; %bb1
-; GFX11-NEXT:    s_nop 0
-; GFX11-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-NEXT:    s_endpgm
   %val = call i1 @llvm.amdgcn.is.private(ptr %ptr)
   br i1 %val, label %bb0, label %bb1
diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.is.shared.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.is.shared.ll
index 63702d2587574b..e005c38355a3ce 100644
--- a/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.is.shared.ll
+++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.is.shared.ll
@@ -131,8 +131,6 @@ define amdgpu_kernel void @is_local_sgpr(ptr %ptr) {
 ; GFX11-NEXT:    global_store_b32 v[0:1], v0, off dlc
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-NEXT:  .LBB1_2: ; %bb1
-; GFX11-NEXT:    s_nop 0
-; GFX11-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-NEXT:    s_endpgm
   %val = call i1 @llvm.amdgcn.is.shared(ptr %ptr)
   br i1 %val, label %bb0, label %bb1
diff --git a/llvm/test/CodeGen/AMDGPU/atomic_optimizations_local_pointer.ll b/llvm/test/CodeGen/AMDGPU/atomic_optimizations_local_pointer.ll
index 352adacbda3e4b..af6f69130910d0 100644
--- a/llvm/test/CodeGen/AMDGPU/atomic_optimizations_local_pointer.ll
+++ b/llvm/test/CodeGen/AMDGPU/atomic_optimizations_local_pointer.ll
@@ -39,9 +39,9 @@ define amdgpu_kernel void @add_i32_constant(ptr addrspace(1) %out) {
 ; GFX7LESS-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
 ; GFX7LESS-NEXT:    s_mov_b32 s3, 0xf000
 ; GFX7LESS-NEXT:    s_mov_b32 s2, -1
-; GFX7LESS-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX7LESS-NEXT:    v_readfirstlane_b32 s4, v1
 ; GFX7LESS-NEXT:    v_mad_u32_u24 v0, v0, 5, s4
+; GFX7LESS-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX7LESS-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 ; GFX7LESS-NEXT:    s_endpgm
 ;
@@ -65,11 +65,11 @@ define amdgpu_kernel void @add_i32_constant(ptr addrspace(1) %out) {
 ; GFX8-NEXT:  .LBB0_2:
 ; GFX8-NEXT:    s_or_b64 exec, exec, s[2:3]
 ; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX8-NEXT:    v_readfirstlane_b32 s4, v1
 ; GFX8-NEXT:    s_mov_b32 s3, 0xf000
 ; GFX8-NEXT:    s_mov_b32 s2, -1
 ; GFX8-NEXT:    v_mad_u32_u24 v0, v0, 5, s4
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX8-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 ; GFX8-NEXT:    s_endpgm
 ;
@@ -92,11 +92,11 @@ define amdgpu_kernel void @add_i32_constant(ptr addrspace(1) %out) {
 ; GFX9-NEXT:  .LBB0_2:
 ; GFX9-NEXT:    s_or_b64 exec, exec, s[2:3]
 ; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    v_readfirstlane_b32 s4, v1
 ; GFX9-NEXT:    s_mov_b32 s3, 0xf000
 ; GFX9-NEXT:    s_mov_b32 s2, -1
 ; GFX9-NEXT:    v_mad_u32_u24 v0, v0, 5, s4
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 ; GFX9-NEXT:    s_endpgm
 ;
@@ -253,8 +253,8 @@ define amdgpu_kernel void @add_i32_uniform(ptr addrspace(1) %out, i32 %additive)
 ; GFX7LESS-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
 ; GFX7LESS-NEXT:    s_mov_b32 s3, 0xf000
 ; GFX7LESS-NEXT:    s_mov_b32 s2, -1
-; GFX7LESS-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX7LESS-NEXT:    v_readfirstlane_b32 s4, v1
+; GFX7LESS-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX7LESS-NEXT:    v_mul_lo_u32 v0, s6, v0
 ; GFX7LESS-NEXT:    v_add_i32_e32 v0, vcc, s4, v0
 ; GFX7LESS-NEXT:    buffer_store_dword v0, off, s[0:3], 0
@@ -504,11 +504,11 @@ define amdgpu_kernel void @add_i32_varying(ptr addrspace(1) %out) {
 ; GFX8-NEXT:  .LBB2_4:
 ; GFX8-NEXT:    s_or_b64 exec, exec, s[2:3]
 ; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX8-NEXT:    v_readfirstlane_b32 s4, v0
 ; GFX8-NEXT:    s_mov_b32 s3, 0xf000
 ; GFX8-NEXT:    s_mov_b32 s2, -1
 ; GFX8-NEXT:    v_add_u32_e32 v0, vcc, s4, v1
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX8-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 ; GFX8-NEXT:    s_endpgm
 ;
@@ -544,11 +544,11 @@ define amdgpu_kernel void @add_i32_varying(ptr addrspace(1) %out) {
 ; GFX9-NEXT:  .LBB2_4:
 ; GFX9-NEXT:    s_or_b64 exec, exec, s[2:3]
 ; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    v_readfirstlane_b32 s4, v0
 ; GFX9-NEXT:    s_mov_b32 s3, 0xf000
 ; GFX9-NEXT:    s_mov_b32 s2, -1
 ; GFX9-NEXT:    v_add_u32_e32 v0, s4, v1
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 ; GFX9-NEXT:    s_endpgm
 ;
@@ -944,7 +944,6 @@ define amdgpu_kernel void @add_i64_constant(ptr addrspace(1) %out) {
 ; GFX7LESS-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
 ; GFX7LESS-NEXT:    s_mov_b32 s3, 0xf000
 ; GFX7LESS-NEXT:    s_mov_b32 s2, -1
-; GFX7LESS-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX7LESS-NEXT:    v_readfirstlane_b32 s4, v0
 ; GFX7LESS-NEXT:    v_readfirstlane_b32 s5, v1
 ; GFX7LESS-NEXT:    v_mul_hi_u32_u24_e32 v1, 5, v2
@@ -952,6 +951,7 @@ define amdgpu_kernel void @add_i64_constant(ptr addrspace(1) %out) {
 ; GFX7LESS-NEXT:    v_mov_b32_e32 v2, s5
 ; GFX7LESS-NEXT:    v_add_i32_e32 v0, vcc, s4, v0
 ; GFX7LESS-NEXT:    v_addc_u32_e32 v1, vcc, v2, v1, vcc
+; GFX7LESS-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX7LESS-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
 ; GFX7LESS-NEXT:    s_endpgm
 ;
@@ -974,7 +974,6 @@ define amdgpu_kernel void @add_i64_constant(ptr addrspace(1) %out) {
 ; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX8-NEXT:  .LBB4_2:
 ; GFX8-NEXT:    s_or_b64 exec, exec, s[2:3]
-; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX8-NEXT:    v_readfirstlane_b32 s2, v0
 ; GFX8-NEXT:    v_readfirstlane_b32 s3, v1
 ; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
@@ -1006,7 +1005,6 @@ define amdgpu_kernel void @add_i64_constant(ptr addrspace(1) %out) {
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:  .LBB4_2:
 ; GFX9-NEXT:    s_or_b64 exec, exec, s[2:3]
-; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    v_readfirstlane_b32 s2, v0
 ; GFX9-NEXT:    v_readfirstlane_b32 s3, v1
 ; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
@@ -1219,11 +1217,11 @@ define amdgpu_kernel void @add_i64_uniform(ptr addrspace(1) %out, i64 %additive)
 ; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX8-NEXT:  .LBB5_2:
 ; GFX8-NEXT:    s_or_b64 exec, exec, s[4:5]
-; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX8-NEXT:    v_readfirstlane_b32 s4, v0
 ; GFX8-NEXT:    v_readfirstlane_b32 s5, v1
 ; GFX8-NEXT:    v_mov_b32_e32 v0, s4
 ; GFX8-NEXT:    v_mov_b32_e32 v1, s5
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX8-NEXT:    v_mul_lo_u32 v3, s3, v2
 ; GFX8-NEXT:    v_mad_u64_u32 v[0:1], s[2:3], s2, v2, v[0:1]
 ; GFX8-NEXT:    s_mov_b32 s7, 0xf000
@@ -1258,11 +1256,11 @@ define amdgpu_kernel void @add_i64_uniform(ptr addrspace(1) %out, i64 %additive)
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:  .LBB5_2:
 ; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]
-; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    v_readfirstlane_b32 s4, v0
 ; GFX9-NEXT:    v_readfirstlane_b32 s5, v1
 ; GFX9-NEXT:    v_mov_b32_e32 v0, s4
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s5
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    v_mad_u64_u32 v[0:1], s[4:5], s2, v2, v[0:1]
 ; GFX9-NEXT:    s_mov_b32 s7, 0xf000
 ; GFX9-NEXT:    s_mov_b32 s6, -1
@@ -1530,10 +1528,10 @@ define amdgpu_kernel void @sub_i32_constant(ptr addrspace(1) %out) {
 ; GFX7LESS-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
 ; GFX7LESS-NEXT:    s_mov_b32 s3, 0xf000
 ; GFX7LESS-NEXT:    s_mov_b32 s2, -1
-; GFX7LESS-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX7LESS-NEXT:    v_readfirstlane_b32 s4, v1
 ; GFX7LESS-NEXT:    v_mul_u32_u24_e32 v0, 5, v0
 ; GFX7LESS-NEXT:    v_sub_i32_e32 v0, vcc, s4, v0
+; GFX7LESS-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX7LESS-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 ; GFX7LESS-NEXT:    s_endpgm
 ;
@@ -1557,12 +1555,12 @@ define amdgpu_kernel void @sub_i32_constant(ptr addrspace(1) %out) {
 ; GFX8-NEXT:  .LBB7_2:
 ; GFX8-NEXT:    s_or_b64 exec, exec, s[2:3]
 ; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX8-NEXT:    v_readfirstlane_b32 s4, v1
 ; GFX8-NEXT:    v_mul_u32_u24_e32 v0, 5, v0
 ; GFX8-NEXT:    s_mov_b32 s3, 0xf000
 ; GFX8-NEXT:    s_mov_b32 s2, -1
 ; GFX8-NEXT:    v_sub_u32_e32 v0, vcc, s4, v0
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX8-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 ; GFX8-NEXT:    s_endpgm
 ;
@@ -1585,12 +1583,12 @@ define amdgpu_kernel void @sub_i32_constant(ptr addrspace(1) %out) {
 ; GFX9-NEXT:  .LBB7_2:
 ; GFX9-NEXT:    s_or_b64 exec, exec, s[2:3]
 ; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    v_readfirstlane_b32 s4, v1
 ; GFX9-NEXT:    v_mul_u32_u24_e32 v0, 5, v0
 ; GFX9-NEXT:    s_mov_b32 s3, 0xf000
 ; GFX9-NEXT:    s_mov_b32 s2, -1
 ; GFX9-NEXT:    v_sub_u32_e32 v0, s4, v0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 ; GFX9-NEXT:    s_endpgm
 ;
@@ -1751,8 +1749,8 @@ define amdgpu_kernel void @sub_i32_uniform(ptr addrspace(1) %out, i32 %subitive)
 ; GFX7LESS-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
 ; GFX7LESS-NEXT:    s_mov_b32 s3, 0xf000
 ; GFX7LESS-NEXT:    s_mov_b32 s2, -1
-; GFX7LESS-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX7LESS-NEXT:    v_readfirstlane_b32 s4, v1
+; GFX7LESS-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX7LESS-NEXT:    v_mul_lo_u32 v0, s6, v0
 ; GFX7LESS-NEXT:    v_sub_i32_e32 v0, vcc, s4, v0
 ; GFX7LESS-NEXT:    buffer_store_dword v0, off, s[0:3], 0
@@ -2006,11 +2004,11 @@ define amdgpu_kernel void @sub_i32_varying(ptr addrspace(1) %out) {
 ; GFX8-NEXT:  .LBB9_4:
 ; GFX8-NEXT:    s_or_b64 exec, exec, s[2:3]
 ; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX8-NEXT:    v_readfirstlane_b32 s4, v0
 ; GFX8-NEXT:    s_mov_b32 s3, 0xf000
 ; GFX8-NEXT:    s_mov_b32 s2, -1
 ; GFX8-NEXT:    v_sub_u32_e32 v0, vcc, s4, v1
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX8-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 ; GFX8-NEXT:    s_endpgm
 ;
@@ -2046,11 +2044,11 @@ define amdgpu_kernel void @sub_i32_varying(ptr addrspace(1) %out) {
 ; GFX9-NEXT:  .LBB9_4:
 ; GFX9-NEXT:    s_or_b64 exec, exec, s[2:3]
 ; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    v_readfirstlane_b32 s4, v0
 ; GFX9-NEXT:    s_mov_b32 s3, 0xf000
 ; GFX9-NEXT:    s_mov_b32 s2, -1
 ; GFX9-NEXT:    v_sub_u32_e32 v0, s4, v1
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 ; GFX9-NEXT:    s_endpgm
 ;
@@ -2446,7 +2444,6 @@ define amdgpu_kernel void @sub_i64_constant(ptr addrspace(1) %out) {
 ; GFX7LESS-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
 ; GFX7LESS-NEXT:    s_mov_b32 s3, 0xf000
 ; GFX7LESS-NEXT:    s_mov_b32 s2, -1
-; GFX7LESS-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX7LESS-NEXT:    v_readfirstlane_b32 s4, v0
 ; GFX7LESS-NEXT:    v_readfirstlane_b32 s5, v1
 ; GFX7LESS-NEXT:    v_mul_hi_u32_u24_e32 v1, 5, v2
@@ -2454,6 +2451,7 @@ define amdgpu_kernel void @sub_i64_constant(ptr addrspace(1) %out) {
 ; GFX7LESS-NEXT:    v_mov_b32_e32 v2, s5
 ; GFX7LESS-NEXT:    v_sub_i32_e32 v0, vcc, s4, v0
 ; GFX7LESS-NEXT:    v_subb_u32_e32 v1, vcc, v2, v1, vcc
+; GFX7LESS-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX7LESS-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
 ; GFX7LESS-NEXT:    s_endpgm
 ;
@@ -2477,7 +2475,6 @@ define amdgpu_kernel void @sub_i64_constant(ptr addrspace(1) %out) {
 ; GFX8-NEXT:  .LBB11_2:
 ; GFX8-NEXT:    s_or_b64 exec, exec, s[2:3]
 ; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX8-NEXT:    v_readfirstlane_b32 s4, v0
 ; GFX8-NEXT:    v_readfirstlane_b32 s5, v1
 ; GFX8-NEXT:    v_mul_u32_u24_e32 v0, 5, v2
@@ -2487,6 +2484,7 @@ define amdgpu_kernel void @sub_i64_constant(ptr addrspace(1) %out) {
 ; GFX8-NEXT:    s_mov_b32 s3, 0xf000
 ; GFX8-NEXT:    s_mov_b32 s2, -1
 ; GFX8-NEXT:    v_subb_u32_e32 v1, vcc, v2, v1, vcc
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX8-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
 ; GFX8-NEXT:    s_endpgm
 ;
@@ -2509,7 +2507,6 @@ define amdgpu_kernel void @sub_i64_constant(ptr addrspace(1) %out) {
 ; GFX9-NEXT:  .LBB11_2:
 ; GFX9-NEXT:    s_or_b64 exec, exec, s[2:3]
 ; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    v_readfirstlane_b32 s4, v0
 ; GFX9-NEXT:    v_readfirstlane_b32 s5, v1
 ; GFX9-NEXT:    v_mul_u32_u24_e32 v0, 5, v2
@@ -2519,6 +2516,7 @@ define amdgpu_kernel void @sub_i64_constant(ptr addrspace(1) %out) {
 ; GFX9-NEXT:    s_mov_b32 s3, 0xf000
 ; GFX9-NEXT:    s_mov_b32 s2, -1
 ; GFX9-NEXT:    v_subb_co_u32_e32 v1, vcc, v2, v1, vcc
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
 ; GFX9-NEXT:    s_endpgm
 ;
@@ -3081,11 +3079,11 @@ define amdgpu_kernel void @and_i32_varying(ptr addrspace(1) %out) {
 ; GFX8-NEXT:  .LBB14_4:
 ; GFX8-NEXT:    s_or_b64 exec, exec, s[2:3]
 ; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX8-NEXT:    v_readfirstlane_b32 s4, v0
 ; GFX8-NEXT:    s_mov_b32 s3, 0xf000
 ; GFX8-NEXT:    s_mov_b32 s2, -1
 ; GFX8-NEXT:    v_and_b32_e32 v0, s4, v1
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX8-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 ; GFX8-NEXT:    s_endpgm
 ;
@@ -3121,11 +3119,11 @@ define amdgpu_kernel void @and_i32_varying(ptr addrspace(1) %out) {
 ; GFX9-NEXT:  .LBB14_4:
 ; GFX9-NEXT:    s_or_b64 exec, exec, s[2:3]
 ; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    v_readfirstlane_b32 s4, v0
 ; GFX9-NEXT:    s_mov_b32 s3, 0xf000
 ; GFX9-NEXT:    s_mov_b32 s2, -1
 ; GFX9-NEXT:    v_and_b32_e32 v0, s4, v1
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 ; GFX9-NEXT:    s_endpgm
 ;
@@ -3355,11 +3353,11 @@ define amdgpu_kernel void @or_i32_varying(ptr addrspace(1) %out) {
 ; GFX8-NEXT:  .LBB15_4:
 ; GFX8-NEXT:    s_or_b64 exec, exec, s[2:3]
 ; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX8-NEXT:    v_readfirstlane_b32 s4, v0
 ; GFX8-NEXT:    s_mov_b32 s3, 0xf000
 ; GFX8-NEXT:    s_mov_b32 s2, -1
 ; GFX8-NEXT:    v_or_b32_e32 v0, s4, v1
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX8-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 ; GFX8-NEXT:    s_endpgm
 ;
@@ -3395,11 +3393,11 @@ define amdgpu_kernel void @or_i32_varying(ptr addrspace(1) %out) {
 ; GFX9-NEXT:  .LBB15_4:
 ; GFX9-NEXT:    s_or_b64 exec, exec, s[2:3]
 ; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    v_readfirstlane_b32 s4, v0
 ; GFX9-NEXT:    s_mov_b32 s3, 0xf000
 ; GFX9-NEXT:    s_mov_b32 s2, -1
 ; GFX9-NEXT:    v_or_b32_e32 v0, s4, v1
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 ; GFX9-NEXT:    s_endpgm
 ;
@@ -3629,11 +3627,11 @@ define amdgpu_kernel void @xor_i32_varying(ptr addrspace(1) %out) {
 ; GFX8-NEXT:  .LBB16_4:
 ; GFX8-NEXT:    s_or_b64 exec, exec, s[2:3]
 ; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX8-NEXT:    v_readfirstlane_b32 s4, v0
 ; GFX8-NEXT:    s_mov_b32 s3, 0xf000
 ; GFX8-NEXT:    s_mov_b32 s2, -1
 ; GFX8-NEXT:    v_xor_b32_e32 v0, s4, v1
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
 ; ...
[truncated]

arsenm
arsenm reviewed Apr 3, 2024
View reviewed changes
llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.is.private.ll
@@ -131,8 +131,6 @@ define amdgpu_kernel void @is_private_sgpr(ptr %ptr) {
; GFX11-NEXT: global_store_b32 v[0:1], v0, off dlc
; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-NEXT: .LBB1_2: ; %bb1
; GFX11-NEXT: s_nop 0
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Dont' understand this change, no waits changed?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

SIInsertWaitcnts only inserts that s_nop/s_sendmsg sequence if it thinks that vscnt might be non-zero at the s_endpgm. I think what happened here is that the pass now takes notice of the "s_waitcnt_vscnt 0" at the end of the previous MBB. You're right that it didn't get changed, but the pass's bookkeeping got updated with the knowledge that vscnt must be zero after that.

@jayfoad jayfoad mentioned this pull request Apr 3, 2024
Merged
jwanggit86
jwanggit86 reviewed Apr 3, 2024
View reviewed changes
llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp Outdated
Wait.BvhCnt = 0;
}

// Combine or remove any redundant waitcnts ar the end of the block.
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

"ar the end" => "at the end"

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Done

jwanggit86
jwanggit86 reviewed Apr 3, 2024
View reviewed changes
llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp
@@ -2355,9 +2327,22 @@ bool SIInsertWaitcnts::insertWaitcntInBlock(MachineFunction &MF,
++Iter;
}

// Flush the LOADcnt, SAMPLEcnt and BVHcnt counters at the end of the block if
// needed.
AMDGPU::Waitcnt Wait;
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Move the variable "Wait" inside the if?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

No. It's used outside the if.

@jayfoad jayfoad merged commit 3cf539f into llvm:main Apr 4, 2024
@jayfoad jayfoad deleted the combine-redundant-waitcnts branch April 4, 2024 09:14
jwanggit86 pushed a commit to jwanggit86/llvm-project that referenced this pull request Apr 4, 2024
With llvm#87539, previous commit that checks for the instruction imme…
49cad2d 
…diately

after a load/store is not necessary.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Reviewers

@jwanggit86 jwanggit86 jwanggit86 left review comments

@arsenm arsenm arsenm approved these changes

@Pierre-vh Pierre-vh Awaiting requested review from Pierre-vh

@stepthomas stepthomas Awaiting requested review from stepthomas

Assignees
No one assigned
Labels
backend:AMDGPU llvm:globalisel
Projects
None yet
Milestone
No milestone
Development

Successfully merging this pull request may close these issues.
4 participants
@jayfoad @llvmbot @arsenm @jwanggit86