[AMDGPU] Fix a potential integer overflow in GCNRegPressure when true16 is enabled

Fixes SWDEV-537014.

Author: shiltian

llvm/lib/Target/AMDGPU/GCNRegPressure.cpp

@@ -66,7 +66,25 @@ void GCNRegPressure::inc(unsigned Reg,
       Value[TupleIdx] += Sign * TRI->getRegClassWeight(RC).RegWeight;
     }
     // Pressure scales with number of new registers covered by the new mask.
-    Sign *= SIRegisterInfo::getNumCoveredRegs(~PrevMask & NewMask);
+    // Note when true16 is enabled, we can no longer safely use the following
+    // approach to calculate the difference in the number of 32-bit registers
+    // between two masks:
+    //
+    // Sign *= SIRegisterInfo::getNumCoveredRegs(~PrevMask & NewMask);
+    //
+    // The issue is that the mask calculation `~PrevMask & NewMask` doesn't
+    // properly account for partial usage of a 32-bit register when dealing with
+    // 16-bit registers.
+    //
+    // Consider this example:
+    // Assume PrevMask = 0b0010 and NewMask = 0b1111. Here, the correct register
+    // usage difference should be 1, because even though PrevMask uses only half
+    // of a 32-bit register, it should still be counted as a full register use.
+    // However, the mask calculation yields `~PrevMask & NewMask = 0b1101`, and
+    // calling `getNumCoveredRegs` returns 2 instead of 1. This incorrect
+    // calculation can lead to integer overflow when Sign = -1.
+    Sign *= SIRegisterInfo::getNumCoveredRegs(NewMask) -
+            SIRegisterInfo::getNumCoveredRegs(PrevMask);
   }
   Value[RegKind] += Sign;
 }

llvm/test/CodeGen/AMDGPU/gcn-reg-pressure-true16-integer-overflow.mir

@@ -0,0 +1,70 @@
+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py UTC_ARGS: --version 5
+# RUN: llc -x mir -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1102 -run-pass=machine-scheduler %s -o - | FileCheck %s
+
+--- |
+  declare void @llvm.amdgcn.s.waitcnt(i32 immarg)
+
+  declare <2 x i32> @llvm.amdgcn.raw.buffer.load.v2i32(<4 x i32>, i32, i32, i32 immarg)
+
+  define amdgpu_kernel void @foo(ptr %p) {
+  entry:
+    %foo.kernarg.segment = call nonnull align 16 dereferenceable(264) ptr addrspace(4) @llvm.amdgcn.kernarg.segment.ptr()
+    %p.kernarg.offset1 = bitcast ptr addrspace(4) %foo.kernarg.segment to ptr addrspace(4)
+    %p.load = load ptr, ptr addrspace(4) %p.kernarg.offset1, align 16
+    %call = tail call <2 x i32> @llvm.amdgcn.raw.buffer.load.v2i32(<4 x i32> zeroinitializer, i32 0, i32 0, i32 0)
+    %cast = bitcast <2 x i32> %call to <8 x i8>
+    %shuffle = shufflevector <8 x i8> zeroinitializer, <8 x i8> %cast, <2 x i32> <i32 3, i32 11>
+    %zext = zext <2 x i8> %shuffle to <2 x i16>
+    %shl = shl <2 x i16> %zext, splat (i16 8)
+    store <2 x i16> %shl, ptr %p.load, align 4
+    tail call void @llvm.amdgcn.s.waitcnt(i32 0)
+    ret void
+  }
+
+  declare noundef align 4 ptr addrspace(4) @llvm.amdgcn.kernarg.segment.ptr()
+...
+---
+name:            foo
+tracksRegLiveness: true
+liveins:
+  - { reg: '$sgpr4_sgpr5', virtual-reg: '%3' }
+body:             |
+  bb.0.entry:
+    liveins: $sgpr4_sgpr5
+
+    ; CHECK-LABEL: name: foo
+    ; CHECK: liveins: $sgpr4_sgpr5
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: [[COPY:%[0-9]+]]:sgpr_64(p4) = COPY $sgpr4_sgpr5
+    ; CHECK-NEXT: [[S_MOV_B32_:%[0-9]+]]:sreg_32 = S_MOV_B32 0
+    ; CHECK-NEXT: undef [[COPY1:%[0-9]+]].sub0:sgpr_128 = COPY [[S_MOV_B32_]]
+    ; CHECK-NEXT: [[COPY1:%[0-9]+]].sub1:sgpr_128 = COPY [[S_MOV_B32_]]
+    ; CHECK-NEXT: [[COPY1:%[0-9]+]].sub2:sgpr_128 = COPY [[S_MOV_B32_]]
+    ; CHECK-NEXT: [[COPY1:%[0-9]+]].sub3:sgpr_128 = COPY [[S_MOV_B32_]]
+    ; CHECK-NEXT: [[BUFFER_LOAD_DWORDX2_OFFSET:%[0-9]+]]:vreg_64 = BUFFER_LOAD_DWORDX2_OFFSET [[COPY1]], 0, 0, 0, 0, implicit $exec :: (dereferenceable load (s64), align 1, addrspace 8)
+    ; CHECK-NEXT: [[S_LOAD_DWORDX2_IMM:%[0-9]+]]:sreg_64_xexec = S_LOAD_DWORDX2_IMM [[COPY]](p4), 0, 0 :: (dereferenceable invariant load (s64) from %ir.p.kernarg.offset1, align 16, addrspace 4)
+    ; CHECK-NEXT: [[V_LSHRREV_B64_e64_:%[0-9]+]]:vreg_64 = V_LSHRREV_B64_e64 24, [[BUFFER_LOAD_DWORDX2_OFFSET]], implicit $exec
+    ; CHECK-NEXT: undef [[COPY2:%[0-9]+]].lo16:vgpr_32 = COPY [[V_LSHRREV_B64_e64_]].lo16
+    ; CHECK-NEXT: [[V_LSHLREV_B32_e64_:%[0-9]+]]:vgpr_32 = V_LSHLREV_B32_e64 16, [[COPY2]], implicit $exec
+    ; CHECK-NEXT: [[COPY3:%[0-9]+]]:vreg_64 = COPY [[S_LOAD_DWORDX2_IMM]]
+    ; CHECK-NEXT: [[V_PK_LSHLREV_B16_:%[0-9]+]]:vgpr_32 = V_PK_LSHLREV_B16 0, 8, 8, [[V_LSHLREV_B32_e64_]], 0, 0, 0, 0, 0, implicit $exec
+    ; CHECK-NEXT: FLAT_STORE_DWORD [[COPY3]], [[V_PK_LSHLREV_B16_]], 0, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %ir.p.load)
+    ; CHECK-NEXT: S_WAITCNT 0
+    ; CHECK-NEXT: S_ENDPGM 0
+    %3:sgpr_64(p4) = COPY killed $sgpr4_sgpr5
+    %13:sreg_64_xexec = S_LOAD_DWORDX2_IMM killed %3(p4), 0, 0 :: (dereferenceable invariant load (s64) from %ir.p.kernarg.offset1, align 16, addrspace 4)
+    %14:sreg_32 = S_MOV_B32 0
+    undef %15.sub0:sgpr_128 = COPY %14
+    %15.sub1:sgpr_128 = COPY %14
+    %15.sub2:sgpr_128 = COPY %14
+    %15.sub3:sgpr_128 = COPY killed %14
+    %16:vreg_64 = BUFFER_LOAD_DWORDX2_OFFSET killed %15, 0, 0, 0, 0, implicit $exec :: (dereferenceable load (s64), align 1, addrspace 8)
+    %26:vreg_64 = V_LSHRREV_B64_e64 24, killed %16, implicit $exec
+    undef %28.lo16:vgpr_32 = COPY killed %26.lo16
+    %30:vgpr_32 = V_LSHLREV_B32_e64 16, killed %28, implicit $exec
+    %24:vgpr_32 = V_PK_LSHLREV_B16 0, 8, 8, killed %30, 0, 0, 0, 0, 0, implicit $exec
+    %25:vreg_64 = COPY killed %13
+    FLAT_STORE_DWORD killed %25, killed %24, 0, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %ir.p.load)
+    S_WAITCNT 0
+    S_ENDPGM 0
+...