cachepc

commit 22297cd9d96fa608ff6d166fa215ff503908e539
parent 5ad426b7b5859275739616e5f62041317b5511a1
Author: Louis Burda <quent.burda@gmail.com>
Date:   Thu,  9 Feb 2023 08:41:34 -0600

Most consistent version of prime self-tests

Diffstat:
M
cachepc/asm.S
 | 
24
++++++++++++++++++------
M
cachepc/cachepc.c
 | 
6
+++---
M
cachepc/macro.S
 | 
4
++++

3 files changed, 25 insertions(+), 9 deletions(-)
diff --git a/cachepc/asm.S b/cachepc/asm.S
@@ -28,6 +28,7 @@ SYM_FUNC_END(cpc_read_pmc)
 SYM_FUNC_START(cpc_stream_hwpf_test_asm)
     push %rbx
 
+    mfence
     wbinvd
 
     readpmc $0 %r8
@@ -55,10 +56,11 @@ SYM_FUNC_START(cpc_prime_test1_asm)
 
     mov cpc_ds, %r10
 
+    mfence
     wbinvd
 
     readpmc $0 %r8
-    prime_pass prime_test1_2 %r10 %r11 %r12
+    prime_pass prime_test1 %r10 %r11 %r12
     readpmc $0 %r9
 
     mov %r9, %rax
@@ -80,10 +82,10 @@ SYM_FUNC_START(cpc_prime_test2_asm)
 
     mov cpc_ds, %r10
 
+    mfence
     wbinvd
 
-    # try to convince replacement policy to
-    # keep this line in
+    # try to convince replacement policy to keep this line in
     mov (%rdi), %rax
     mov (%rdi), %rax
     mov (%rdi), %rax
@@ -92,10 +94,14 @@ SYM_FUNC_START(cpc_prime_test2_asm)
     mov (%rdi), %rax
     mov (%rdi), %rax
 
+    # need 4 passes to evict all lines (reset cache line scores)
     prime_pass prime_test2_1 %r10 %r11 %r12
+    prime_pass prime_test2_2 %r10 %r11 %r12
+    prime_pass prime_test2_3 %r10 %r11 %r12
+    prime_pass prime_test2_4 %r10 %r11 %r12
 
     readpmc $0 %r8
-    prime_pass prime_test2_2 %r12 %r11 %r10
+    prime_pass prime_test2 %r12 %r11 %r10
     readpmc $0 %r9
 
     mov %r9, %rax
@@ -117,9 +123,10 @@ SYM_FUNC_START(cpc_prime_test3_asm)
 
     mov cpc_ds, %r10
 
+    mfence
     wbinvd
-    barrier
 
+    # try to convince replacement policy to keep this line in
     mov (%rsi), %rax
     mov (%rsi), %rax
     mov (%rsi), %rax
@@ -129,7 +136,11 @@ SYM_FUNC_START(cpc_prime_test3_asm)
     mov (%rsi), %rax
     mov (%rsi), %rax
 
-    prime_pass prime_test3 %r10 %r11 %r12
+    # need 4 passes to evict all lines (reset cache line scores)
+    prime_pass prime_test3_1 %r10 %r11 %r12
+    prime_pass prime_test3_2 %r10 %r11 %r12
+    prime_pass prime_test3_3 %r10 %r11 %r12
+    prime_pass prime_test3_4 %r10 %r11 %r12
 
     readpmc $0 %r8
 .rept L1_ASSOC
@@ -152,6 +163,7 @@ SYM_FUNC_START(cpc_eviction_prio_test_asm)
 
     mov cpc_ds, %r10
 
+    mfence
     wbinvd
 
     prime eviction_prio_test %r10 %r11 %r12
diff --git a/cachepc/cachepc.c b/cachepc/cachepc.c
@@ -161,8 +161,8 @@ cpc_save_msrmts(struct cpc_cl *head)
 	cl = head;
 	do {
 		if (cl->first) {
-			BUG_ON(cl->set >= L1_SETS);
-			BUG_ON(cl->line != 0);
+			WARN_ON(cl->set >= L1_SETS);
+			WARN_ON(cl->line != 0);
 			if (cl->count > L1_ASSOC) {
 				CPC_ERR("OOB count %llu for set %u\n",
 					cl->count, cl->set);
@@ -171,7 +171,7 @@ cpc_save_msrmts(struct cpc_cl *head)
 			}
 			cpc_msrmts[cl->set] = cl->count;
 		} else {
-			BUG_ON(cl->count != 0);
+			WARN_ON(cl->count != 0);
 		}
 		cl->count = 0;
 
diff --git a/cachepc/macro.S b/cachepc/macro.S
@@ -42,9 +42,13 @@
 
 # clobbers rax, rbx, rcx, rdx, cl_tmp, (cl_out)
 .macro prime name cl_in cl_tmp cl_out
+    mfence
+    wbinvd
+
     prime_pass pass1_\name \cl_in \cl_tmp \cl_out
     prime_pass pass2_\name \cl_in \cl_tmp \cl_out
     prime_pass pass3_\name \cl_in \cl_tmp \cl_out
+    prime_pass pass4_\name \cl_in \cl_tmp \cl_out
 .endm
 
 # clobbers rax, rbx, rcx, rdx, cl_tmp1, cl_tmp2, pmc_tmp, pmc_tmp2