colin
/
Hazard3
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Fix incorrect tracking of predictedness of fetch data phase. Fixes performance regression in RV32IMC CoreMark (now runs faster, as it should.)

This commit is contained in:
Luke Wren 2022-06-25 11:32:56 +01:00
parent 979e80be99
commit 31efd07042
2 changed files with 3 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
hdl/hazard3_decode.v
View File

@ -131,9 +131,8 @@ assign d_pc = pc;
// Frontend should mark the whole instruction, and nothing but the
// instruction, as a predicted branch. This goes wrong when we execute the
// address containing the predicted branch twice with different 16-bit
// alignments! We don't care about performance in this case(it took BMC to
// find it), but need to issue a branch-to-self to get back on a linear path,
// otherwise PC and CIR will diverge and we will misexecute.
// alignments (!). We need to issue a branch-to-self to get back on a linear
// path, otherwise PC and CIR will diverge and we will misexecute.
wire partial_predicted_branch = !d_starved &&
|BRANCH_PREDICTOR && d_instr_is_32bit && ^fd_cir_predbranch;

2
hdl/hazard3_frontend.v
View File

@ -386,7 +386,7 @@ always @ (posedge clk or negedge rst_n) begin
};
end
mem_data_predbranch <=
|BRANCH_PREDICTOR && btb_match_current_addr ? (
|BRANCH_PREDICTOR && btb_match_word ? (
btb_src_addr[1] ? 2'b10 :
btb_src_size ? 2'b11 : 2'b01
) :