diff --git a/llvm/lib/Target/AMDGPU/AMDGPUInsertDelayAlu.cpp b/llvm/lib/Target/AMDGPU/AMDGPUInsertDelayAlu.cpp
--- a/llvm/lib/Target/AMDGPU/AMDGPUInsertDelayAlu.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUInsertDelayAlu.cpp
@@ -83,6 +83,10 @@
     // an s_delay_alu instruction.
     static const unsigned TRANS_MAX = 4;
 
+    // The maximum number of SALU cycles we can encode in an s_delay_alu
+    // instruction.
+    static const unsigned SALU_CYCLES_MAX = 3;
+
     // If it was written by a (non-TRANS) VALU, remember how many clock cycles
     // are left until it completes, and how many other (non-TRANS) VALU we have
     // seen since it was issued.
@@ -120,7 +124,9 @@
         TRANSNumVALU = 0;
         break;
       case SALU:
-        SALUCycles = Cycles;
+        // Guard against pseudo-instructions like SI_CALL which are marked as
+        // SALU but with a very high latency.
+        SALUCycles = std::min(Cycles, SALU_CYCLES_MAX);
         break;
       }
     }
@@ -349,6 +355,7 @@
 
       if (instructionWaitsForVALU(MI)) {
         // Forget about all outstanding VALU delays.
+        // TODO: This is overkill since it also forgets about SALU delays.
         State = DelayState();
       } else if (Type != OTHER) {
         DelayInfo Delay;
diff --git a/llvm/test/CodeGen/AMDGPU/insert-delay-alu-bug.ll b/llvm/test/CodeGen/AMDGPU/insert-delay-alu-bug.ll
--- a/llvm/test/CodeGen/AMDGPU/insert-delay-alu-bug.ll
+++ b/llvm/test/CodeGen/AMDGPU/insert-delay-alu-bug.ll
@@ -119,7 +119,7 @@
 ; GFX11-NEXT:    s_mov_b32 s3, 0
 ; GFX11-NEXT:    v_readfirstlane_b32 s2, v0
 ; GFX11-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s25
-; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(/* invalid instid value */)
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(SALU_CYCLE_1)
 ; GFX11-NEXT:    s_or_b32 s2, s2, 1
 ; GFX11-NEXT:    s_lshr_b32 s2, s2, s30
 ; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)