diff --git a/wally-pipelined/src/ifu/icache.sv b/wally-pipelined/src/ifu/icache.sv
index b07e64056..d8083bb6c 100644
--- a/wally-pipelined/src/ifu/icache.sv
+++ b/wally-pipelined/src/ifu/icache.sv
@@ -40,18 +40,26 @@ module icache(
   output logic [31:0]     InstrRawD
 );
 
-    logic        DelayF, DelaySideF, FlushDLastCycle;
-    logic  [1:0] InstrDMuxChoice;
-    logic [15:0] MisalignedHalfInstrF, MisalignedHalfInstrD;
-    logic [31:0] InstrF, AlignedInstrD;
-    logic [31:0] nop = 32'h00000013; // instruction for NOP
+    logic             DelayF, DelaySideF, FlushDLastCycle;
+    logic  [1:0]      InstrDMuxChoice;
+    logic [15:0]      MisalignedHalfInstrF, MisalignedHalfInstrD;
+    logic [31:0]      InstrF, AlignedInstrD;
+    logic [31:0]      nop = 32'h00000013; // instruction for NOP
+    logic             LastReadDataValidF;
+    logic [`XLEN-1:0] LastReadDataF, LastReadAdrF, InDataF;
 
     flopr   #(1)  flushDLastCycleFlop(clk, reset, FlushD | (FlushDLastCycle & StallF), FlushDLastCycle);
     flopenr #(1)  delayStateFlop(clk, reset, ~StallF, (DelayF & ~DelaySideF) ? 1'b1 : 1'b0 , DelaySideF);
     flopenr #(16) halfInstrFlop(clk, reset, DelayF & ~StallF, MisalignedHalfInstrF, MisalignedHalfInstrD);
 
+    // This flop is here to simulate pulling data out of the cache, which is edge-triggered
     flopenr #(32) instrFlop(clk, reset, ~StallF, InstrF, AlignedInstrD);
 
+    // These flops cache the previous read, to accelerate things
+    flopenr #(`XLEN) lastReadDataFlop(clk, reset, InstrReadF & ~StallF, InstrInF, LastReadDataF);
+    flopenr #(1)     lastReadDataVFlop(clk, reset, InstrReadF & ~StallF, 1'b1, LastReadDataValidF);
+    flopenr #(`XLEN) lastReadAdrFlop(clk, reset, InstrReadF & ~StallF, InstrPAdrF, LastReadAdrF);
+
     // Decide which address needs to be fetched and sent out over InstrPAdrF
     // If the requested address fits inside one read from memory, we fetch that
     // address, adjusted to the bit width. Otherwise, we request the lower word
@@ -64,34 +72,47 @@ module icache(
         end
     endgenerate
     // For now, we always read since the cache doesn't actually cache
-    assign InstrReadF = 1;
+
+    always_comb if (LastReadDataValidF & (InstrPAdrF == LastReadAdrF)) begin
+        assign InstrReadF = 0;
+    end else begin
+        assign InstrReadF = 1;
+    end
+
+    // Pick from the memory input or from the previous read, as appropriate
+    mux2 #(`XLEN) inDataMux(LastReadDataF, InstrInF, InstrReadF, InDataF);
 
     // If the instruction fits in one memory read, then we put the right bits
     // into InstrF. Otherwise, we activate DelayF to signal the rest of the
     // machinery to swizzle bits.
     generate
         if (`XLEN == 32) begin
-            assign InstrF = PCPF[1] ? {16'b0, InstrInF[31:16]} : InstrInF;
+            assign InstrF = PCPF[1] ? {16'b0, InDataF[31:16]} : InDataF;
             assign DelayF = PCPF[1];
-            assign MisalignedHalfInstrF = InstrInF[31:16];
+            assign MisalignedHalfInstrF = InDataF[31:16];
         end else begin
-            assign InstrF = PCPF[2] ? (PCPF[1] ? {16'b0, InstrInF[63:48]}  : InstrInF[63:32]) : (PCPF[1] ? InstrInF[47:16] : InstrInF[31:0]);
+            assign InstrF = PCPF[2] ? (PCPF[1] ? {16'b0, InDataF[63:48]}  : InDataF[63:32]) : (PCPF[1] ? InDataF[47:16] : InDataF[31:0]);
             assign DelayF = PCPF[1] && PCPF[2];
-            assign MisalignedHalfInstrF = InstrInF[63:48];
+            assign MisalignedHalfInstrF = InDataF[63:48];
         end
     endgenerate
-    assign ICacheStallF = DelayF & ~DelaySideF;
+    assign ICacheStallF = 0; //DelayF & ~DelaySideF;
 
     // Detect if the instruction is compressed
-    // TODO Low-hanging optimization, don't delay if compressed
-    assign CompressedF = DelaySideF ? (MisalignedHalfInstrD[1:0] != 2'b11) : (InstrF[1:0] != 2'b11);
+    // TODO Low-hanging optimization, don't delay if getting a compressed instruction
+    assign CompressedF = (DelaySideF & DelayF) ? (MisalignedHalfInstrD[1:0] != 2'b11) : (InstrF[1:0] != 2'b11);
 
     // Pick the correct output, depending on whether we have to assemble this
     // instruction from two reads or not.
     // Output the requested instruction (we don't need to worry if the read is
     // incomplete, since the pipeline stalls for us when it isn't), or a NOP for
     // the cycle when the first of two reads comes in.
-    always_comb
-        assign InstrDMuxChoice = FlushDLastCycle ? 2'b10 : (DelayF ? (DelaySideF ? 2'b01 : 2'b10) : 2'b00);
-    mux3 #(32) instrDMux (AlignedInstrD, {InstrInF[15:0], MisalignedHalfInstrD}, nop, InstrDMuxChoice, InstrRawD);
+    always_comb if (DelayF & (MisalignedHalfInstrF[1:0] != 2'b11)) begin
+        assign InstrDMuxChoice = 2'b11;
+    end else if (FlushDLastCycle) begin
+        assign InstrDMuxChoice = 2'b10;
+    end else begin
+        assign InstrDMuxChoice = {1'b0, DelaySideF};
+    end
+    mux4 #(32) instrDMux (AlignedInstrD, {InstrInF[15:0], MisalignedHalfInstrD}, nop, {16'b0, MisalignedHalfInstrD}, InstrDMuxChoice, InstrRawD);
 endmodule
diff --git a/wally-pipelined/testbench/testbench-imperas.sv b/wally-pipelined/testbench/testbench-imperas.sv
index 8947482ea..334e75e24 100644
--- a/wally-pipelined/testbench/testbench-imperas.sv
+++ b/wally-pipelined/testbench/testbench-imperas.sv
@@ -53,14 +53,13 @@ module testbench();
 //                    "rv64m/I-REMW-01", "3000"
   };
   string tests64ic[] = '{
-
+                     "rv64ic/I-C-BEQZ-01", "3000",
                      "rv64ic/I-C-ADD-01", "3000",
                      "rv64ic/I-C-ADDI-01", "3000",
                      "rv64ic/I-C-ADDIW-01", "3000",
                      "rv64ic/I-C-ADDW-01", "3000",
                      "rv64ic/I-C-AND-01", "3000",
                      "rv64ic/I-C-ANDI-01", "3000",
-                     "rv64ic/I-C-BEQZ-01", "3000",
                      "rv64ic/I-C-BNEZ-01", "3000",
                      "rv64ic/I-C-EBREAK-01", "2000",
                      "rv64ic/I-C-J-01", "3000",