fpu: carry the decoded wb through the tag store (rebase of #356)

hw/rtl/fpu/VX_fpu_unit.sv

@@ -61,16 +61,16 @@ module VX_fpu_unit import VX_gpu_pkg::*, VX_fpu_pkg::*; #(
 
         fpu_header_t fpu_hdr, fpu_hdr_wb, fpu_hdr_store;
 
-        // FPU always writes back, so header.wb is always 1 at dispatch.
+        // Carry the decoded wb through the tag store: FPU compares may target
+        // integer x0 (e.g. feq.s x0, ...), which decode does not reserve in the
+        // scoreboard, so forcing wb=1 on readback would trip the invalid
+        // writeback assertion.
         always_comb begin
-            fpu_hdr_store    = per_block_execute_if[block_idx].data.header;
-            fpu_hdr_store.wb = 1'b0;
+            fpu_hdr_store = per_block_execute_if[block_idx].data.header;
         end
 
-        // Force wb=1 on the readback path.
         always_comb begin
-            fpu_hdr_wb    = fpu_hdr;
-            fpu_hdr_wb.wb = 1'b1;
+            fpu_hdr_wb = fpu_hdr;
         end
 
         wire [TAG_WIDTH-1:0] fpu_req_tag, fpu_rsp_tag;

tests/kernel/conform/main.cpp

@@ -31,6 +31,8 @@ int main() {
 
 	errors += test_jalr();
 
+	errors += test_feq_x0_scoreboard();
+
 	if (0 == errors) {
 		PRINTF("Passed!\n");
 	} else {

tests/kernel/conform/tests.cpp

@@ -505,3 +505,25 @@ int test_jalr() {
 	}
 	return 0;
 }
+
+///////////////////////////////////////////////////////////////////////////////
+
+void __attribute__((noinline)) do_feq_x0_scoreboard() {
+	asm volatile(
+		"fsgnj.s f24, f0, f0\n"
+		"fsgnj.s f1, f0, f0\n"
+		"feq.s x0, f24, f1\n"
+		:
+		:
+		: "memory");
+}
+
+int test_feq_x0_scoreboard() {
+	PRINTF("FEQ x0 Scoreboard Test\n");
+	int num_threads = std::min(vx_num_threads(), 4);
+	int tmask = make_full_tmask(num_threads);
+	vx_tmc(tmask);
+	do_feq_x0_scoreboard();
+	vx_tmc_one();
+	return 0;
+}

tests/kernel/conform/tests.h

@@ -30,4 +30,6 @@ int test_qshfl();
 
 int test_jalr();
 
+int test_feq_x0_scoreboard();
+
 #endif