freedreno/ir3: add more disasm stats

Add tracking for # of instructions per category, similar to the last
patch.  Also add a few other shader-db stats that were missing on the
disasm side, to make it easier to compare to shaders from cmdstream
traces.

Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6243>

5 files changed, 117 insertions, 48 deletions

diff --git a/src/freedreno/.gitlab-ci/reference/dEQP-VK.draw.indirect_draw.indexed.indirect_draw_count.triangle_list.log b/src/freedreno/.gitlab-ci/reference/dEQP-VK.draw.indirect_draw.indexed.indirect_draw_count.triangle_list.log
index 2471e5b9352..be5a809e3e8 100644
--- a/src/freedreno/.gitlab-ci/reference/dEQP-VK.draw.indirect_draw.indexed.indirect_draw_count.triangle_list.log
+++ b/src/freedreno/.gitlab-ci/reference/dEQP-VK.draw.indirect_draw.indexed.indirect_draw_count.triangle_list.log
@@ -821,7 +821,8 @@ t4					write SP_VS_OBJ_START_LO (a81c)
 
 							- output (half): (cnt=0, max=0)  (estimated)
 							- output (full): 8-11 (cnt=4, max=11)  (estimated)
-							- shaderdb: 13 instructions, 7 nops, 6 non-nops, (10 instlen), 0 half, 3 full
+							- shaderdb: 13 instructions, 7 nops, 6 non-nops, (10 instlen), 0 last-baryf, 0 half, 3 full
+							- shaderdb: 8 cat0, 0 cat1, 1 cat2, 4 cat3, 0 cat4, 0 cat5, 0 cat6, 0 cat7
 							- shaderdb: 0 (ss), 0 (sy)
 00000000010541a4:					0000: 48a81c02 01054000 00000000
 t7					opcode: CP_LOAD_STATE6_GEOM (32) (4 dwords)
@@ -848,7 +849,8 @@ t7					opcode: CP_LOAD_STATE6_GEOM (32) (4 dwords)
 
 						- output (half): (cnt=0, max=0)  (estimated)
 						- output (full): 8-11 (cnt=4, max=11)  (estimated)
-						- shaderdb: 13 instructions, 7 nops, 6 non-nops, (10 instlen), 0 half, 3 full
+						- shaderdb: 13 instructions, 7 nops, 6 non-nops, (10 instlen), 0 last-baryf, 0 half, 3 full
+						- shaderdb: 8 cat0, 0 cat1, 1 cat2, 4 cat3, 0 cat4, 0 cat5, 0 cat6, 0 cat7
 						- shaderdb: 0 (ss), 0 (sy)
 00000000010541b0:					0000: 70328003 00620000 01054000 00000000
 t7					opcode: CP_LOAD_STATE6_GEOM (32) (8 dwords)
@@ -913,7 +915,8 @@ t4					write SP_FS_OBJ_START_LO (a983)
 
 							- output (half): (cnt=0, max=0)  (estimated)
 							- output (full): 2-5 (cnt=4, max=5)  (estimated)
-							- shaderdb: 9 instructions, 4 nops, 5 non-nops, (9 instlen), 0 half, 2 full
+							- shaderdb: 9 instructions, 4 nops, 5 non-nops, (9 instlen), 3 last-baryf, 0 half, 2 full
+							- shaderdb: 5 cat0, 0 cat1, 4 cat2, 0 cat3, 0 cat4, 0 cat5, 0 cat6, 0 cat7
 							- shaderdb: 0 (ss), 0 (sy)
 000000000105422c:					0000: 40a98302 01054080 00000000
 t7					opcode: CP_LOAD_STATE6_FRAG (34) (4 dwords)
@@ -939,7 +942,8 @@ t7					opcode: CP_LOAD_STATE6_FRAG (34) (4 dwords)
 
 						- output (half): (cnt=0, max=0)  (estimated)
 						- output (full): 2-5 (cnt=4, max=5)  (estimated)
-						- shaderdb: 9 instructions, 4 nops, 5 non-nops, (9 instlen), 0 half, 2 full
+						- shaderdb: 9 instructions, 4 nops, 5 non-nops, (9 instlen), 3 last-baryf, 0 half, 2 full
+						- shaderdb: 5 cat0, 0 cat1, 4 cat2, 0 cat3, 0 cat4, 0 cat5, 0 cat6, 0 cat7
 						- shaderdb: 0 (ss), 0 (sy)
 0000000001054238:					0000: 70348003 00720000 01054080 00000000
 t4					write SP_CS_CONFIG (a9bb)
@@ -1507,7 +1511,8 @@ t7			opcode: CP_DRAW_INDIRECT_MULTI (2a) (12 dwords)
 
 				- output (half): (cnt=0, max=0)  (estimated)
 				- output (full): 8-11 (cnt=4, max=11)  (estimated)
-				- shaderdb: 13 instructions, 7 nops, 6 non-nops, (10 instlen), 0 half, 3 full
+				- shaderdb: 13 instructions, 7 nops, 6 non-nops, (10 instlen), 0 last-baryf, 0 half, 3 full
+				- shaderdb: 8 cat0, 0 cat1, 1 cat2, 4 cat3, 0 cat4, 0 cat5, 0 cat6, 0 cat7
 				- shaderdb: 0 (ss), 0 (sy)
 !+	00000100			SP_VS_CONFIG: { ENABLED | NTEX = 0 | NSAMP = 0 | NIBO = 0 }
 !+	00000001			SP_VS_INSTRLEN: 1
@@ -1542,7 +1547,8 @@ t7			opcode: CP_DRAW_INDIRECT_MULTI (2a) (12 dwords)
 
 				- output (half): (cnt=0, max=0)  (estimated)
 				- output (full): 2-5 (cnt=4, max=5)  (estimated)
-				- shaderdb: 9 instructions, 4 nops, 5 non-nops, (9 instlen), 0 half, 2 full
+				- shaderdb: 9 instructions, 4 nops, 5 non-nops, (9 instlen), 3 last-baryf, 0 half, 2 full
+				- shaderdb: 5 cat0, 0 cat1, 4 cat2, 0 cat3, 0 cat4, 0 cat5, 0 cat6, 0 cat7
 				- shaderdb: 0 (ss), 0 (sy)
 !+	00000100			SP_BLEND_CNTL: { UNK8 }
  +	00000000			SP_SRGB_CNTL: { 0 }
diff --git a/src/freedreno/.gitlab-ci/reference/fd-clouds.log b/src/freedreno/.gitlab-ci/reference/fd-clouds.log
index 949862982f0..89ae0409004 100644
--- a/src/freedreno/.gitlab-ci/reference/fd-clouds.log
+++ b/src/freedreno/.gitlab-ci/reference/fd-clouds.log
@@ -644,7 +644,8 @@ t4					write SP_VS_OBJ_START_LO (a81c)
 
 							- output (half): (cnt=0, max=0)  (estimated)
 							- output (full): (cnt=0, max=0)  (estimated)
-							- shaderdb: 5 instructions, 4 nops, 1 non-nops, (5 instlen), 0 half, 0 full
+							- shaderdb: 5 instructions, 4 nops, 1 non-nops, (5 instlen), 0 last-baryf, 0 half, 0 full
+							- shaderdb: 5 cat0, 0 cat1, 0 cat2, 0 cat3, 0 cat4, 0 cat5, 0 cat6, 0 cat7
 							- shaderdb: 0 (ss), 0 (sy)
 0000000001121038:					0000: 48a81c02 01011000 00000000
 t7					opcode: CP_LOAD_STATE6_GEOM (32) (4 dwords)
@@ -666,7 +667,8 @@ t7					opcode: CP_LOAD_STATE6_GEOM (32) (4 dwords)
 
 						- output (half): (cnt=0, max=0)  (estimated)
 						- output (full): (cnt=0, max=0)  (estimated)
-						- shaderdb: 5 instructions, 4 nops, 1 non-nops, (5 instlen), 0 half, 0 full
+						- shaderdb: 5 instructions, 4 nops, 1 non-nops, (5 instlen), 0 last-baryf, 0 half, 0 full
+						- shaderdb: 5 cat0, 0 cat1, 0 cat2, 0 cat3, 0 cat4, 0 cat5, 0 cat6, 0 cat7
 						- shaderdb: 0 (ss), 0 (sy)
 0000000001121044:					0000: 70328003 00620000 01011000 00000000
 t4					write VPC_VAR[0].DISABLE (9212)
@@ -1114,7 +1116,8 @@ t7			opcode: CP_DRAW_INDX_OFFSET (38) (4 dwords)
 
 				- output (half): (cnt=0, max=0)  (estimated)
 				- output (full): (cnt=0, max=0)  (estimated)
-				- shaderdb: 5 instructions, 4 nops, 1 non-nops, (5 instlen), 0 half, 0 full
+				- shaderdb: 5 instructions, 4 nops, 1 non-nops, (5 instlen), 0 last-baryf, 0 half, 0 full
+				- shaderdb: 5 cat0, 0 cat1, 0 cat2, 0 cat3, 0 cat4, 0 cat5, 0 cat6, 0 cat7
 				- shaderdb: 0 (ss), 0 (sy)
 !+	00000100			SP_VS_CONFIG: { ENABLED | NTEX = 0 | NSAMP = 0 | NIBO = 0 }
 !+	00000001			SP_VS_INSTRLEN: 1
@@ -1959,7 +1962,8 @@ t4					write SP_VS_OBJ_START_LO (a81c)
 
 							- output (half): (cnt=0, max=0)  (estimated)
 							- output (full): (cnt=0, max=0)  (estimated)
-							- shaderdb: 5 instructions, 4 nops, 1 non-nops, (5 instlen), 0 half, 0 full
+							- shaderdb: 5 instructions, 4 nops, 1 non-nops, (5 instlen), 0 last-baryf, 0 half, 0 full
+							- shaderdb: 5 cat0, 0 cat1, 0 cat2, 0 cat3, 0 cat4, 0 cat5, 0 cat6, 0 cat7
 							- shaderdb: 0 (ss), 0 (sy)
 0000000001120038:					0000: 48a81c02 01012000 00000000
 t7					opcode: CP_LOAD_STATE6_GEOM (32) (4 dwords)
@@ -1981,7 +1985,8 @@ t7					opcode: CP_LOAD_STATE6_GEOM (32) (4 dwords)
 
 						- output (half): (cnt=0, max=0)  (estimated)
 						- output (full): (cnt=0, max=0)  (estimated)
-						- shaderdb: 5 instructions, 4 nops, 1 non-nops, (5 instlen), 0 half, 0 full
+						- shaderdb: 5 instructions, 4 nops, 1 non-nops, (5 instlen), 0 last-baryf, 0 half, 0 full
+						- shaderdb: 5 cat0, 0 cat1, 0 cat2, 0 cat3, 0 cat4, 0 cat5, 0 cat6, 0 cat7
 						- shaderdb: 0 (ss), 0 (sy)
 0000000001120044:					0000: 70328003 00620000 01012000 00000000
 t4					write VPC_VAR[0].DISABLE (9212)
@@ -3502,7 +3507,8 @@ t4					write SP_FS_OBJ_START_LO (a983)
 
 							- output (half): (cnt=0, max=0)  (estimated)
 							- output (full): 4-7 (cnt=4, max=7)  (estimated)
-							- shaderdb: 2414 instructions, 1355 nops, 1059 non-nops, (1406 instlen), 0 half, 19 full
+							- shaderdb: 2414 instructions, 1114 nops, 1300 non-nops, (1406 instlen), 0 last-baryf, 0 half, 19 full
+							- shaderdb: 1120 cat0, 48 cat1, 551 cat2, 512 cat3, 183 cat4, 0 cat5, 0 cat6, 0 cat7
 							- shaderdb: 140 (ss), 0 (sy)
 0000000001120158:					0000: 40a98302 01013000 00000000
 t7					opcode: CP_LOAD_STATE6_FRAG (34) (4 dwords)
@@ -4925,7 +4931,8 @@ t7					opcode: CP_LOAD_STATE6_FRAG (34) (4 dwords)
 
 						- output (half): (cnt=0, max=0)  (estimated)
 						- output (full): 4-7 (cnt=4, max=7)  (estimated)
-						- shaderdb: 2414 instructions, 1355 nops, 1059 non-nops, (1406 instlen), 0 half, 19 full
+						- shaderdb: 2414 instructions, 1114 nops, 1300 non-nops, (1406 instlen), 0 last-baryf, 0 half, 19 full
+						- shaderdb: 1120 cat0, 48 cat1, 551 cat2, 512 cat3, 183 cat4, 0 cat5, 0 cat6, 0 cat7
 						- shaderdb: 140 (ss), 0 (sy)
 0000000001120164:					0000: 70348003 16320000 01013000 00000000
 t4					write VFD_CONTROL_1 (a001)
@@ -5339,7 +5346,8 @@ t7			opcode: CP_DRAW_INDX_OFFSET (38) (4 dwords)
 
 				- output (half): (cnt=0, max=0)  (estimated)
 				- output (full): (cnt=0, max=0)  (estimated)
-				- shaderdb: 5 instructions, 4 nops, 1 non-nops, (5 instlen), 0 half, 0 full
+				- shaderdb: 5 instructions, 4 nops, 1 non-nops, (5 instlen), 0 last-baryf, 0 half, 0 full
+				- shaderdb: 5 cat0, 0 cat1, 0 cat2, 0 cat3, 0 cat4, 0 cat5, 0 cat6, 0 cat7
 				- shaderdb: 0 (ss), 0 (sy)
  +	00000100			SP_VS_CONFIG: { ENABLED | NTEX = 0 | NSAMP = 0 | NIBO = 0 }
  +	00000001			SP_VS_INSTRLEN: 1
@@ -6777,7 +6785,8 @@ t7			opcode: CP_DRAW_INDX_OFFSET (38) (4 dwords)
 
 				- output (half): (cnt=0, max=0)  (estimated)
 				- output (full): 4-7 (cnt=4, max=7)  (estimated)
-				- shaderdb: 2414 instructions, 1355 nops, 1059 non-nops, (1406 instlen), 0 half, 19 full
+				- shaderdb: 2414 instructions, 1114 nops, 1300 non-nops, (1406 instlen), 0 last-baryf, 0 half, 19 full
+				- shaderdb: 1120 cat0, 48 cat1, 551 cat2, 512 cat3, 183 cat4, 0 cat5, 0 cat6, 0 cat7
 				- shaderdb: 140 (ss), 0 (sy)
 !+	00000100			SP_BLEND_CNTL: { UNK8 }
  +	fcfcfc00			SP_FS_OUTPUT_CNTL0: { DEPTH_REGID = r63.x | SAMPMASK_REGID = r63.x | STENCILREF_REGID = r63.x }
diff --git a/src/freedreno/.gitlab-ci/reference/glxgears-a420.log b/src/freedreno/.gitlab-ci/reference/glxgears-a420.log
index 2abe7e8d2c5..efed6fcdcfd 100644
--- a/src/freedreno/.gitlab-ci/reference/glxgears-a420.log
+++ b/src/freedreno/.gitlab-ci/reference/glxgears-a420.log
@@ -433,7 +433,8 @@ t3			opcode: CP_LOAD_STATE4 (30) (35 dwords)
 
 				- output (half): (cnt=0, max=0)  (estimated)
 				- output (full): (cnt=0, max=0)  (estimated)
-				- shaderdb: 5 instructions, 4 nops, 1 non-nops, (5 instlen), 0 half, 0 full
+				- shaderdb: 5 instructions, 4 nops, 1 non-nops, (5 instlen), 0 last-baryf, 0 half, 0 full
+				- shaderdb: 5 cat0, 0 cat1, 0 cat2, 0 cat3, 0 cat4, 0 cat5, 0 cat6, 0 cat7
 				- shaderdb: 0 (ss), 0 (sy)
 109ce1f0:			0000: c0213000 00600000 00000000 00000000 03000000 00000000 00000000 00000000
 *
@@ -458,7 +459,8 @@ t3			opcode: CP_LOAD_STATE4 (30) (35 dwords)
 
 				- output (half): (cnt=0, max=0)  (estimated)
 				- output (full): 0-3 (cnt=4, max=3)  (estimated)
-				- shaderdb: 9 instructions, 8 nops, 1 non-nops, (9 instlen), 0 half, 1 full
+				- shaderdb: 9 instructions, 4 nops, 5 non-nops, (9 instlen), 0 last-baryf, 0 half, 1 full
+				- shaderdb: 5 cat0, 4 cat1, 0 cat2, 0 cat3, 0 cat4, 0 cat5, 0 cat6, 0 cat7
 				- shaderdb: 0 (ss), 0 (sy)
 109ce27c:			0000: c0213000 00700000 00000000 00000000 20244000 00000001 20244001 00000002
 109ce29c:			0020: 20244002 00000003 20244003 00000000 03000000 00000000 00000000 00000000
@@ -1045,7 +1047,8 @@ t3			opcode: CP_LOAD_STATE4 (30) (131 dwords)
 
 				- output (half): (cnt=0, max=0)  (estimated)
 				- output (full): 6-13 (cnt=8, max=13)  (estimated)
-				- shaderdb: 74 instructions, 38 nops, 36 non-nops, (61 instlen), 0 half, 4 full
+				- shaderdb: 74 instructions, 27 nops, 47 non-nops, (61 instlen), 0 last-baryf, 0 half, 4 full
+				- shaderdb: 28 cat0, 8 cat1, 15 cat2, 22 cat3, 1 cat4, 0 cat5, 0 cat6, 0 cat7
 				- shaderdb: 1 (ss), 0 (sy)
 109ce66c:			0000: c0813000 01200000 00000000 10000002 40700000 10030002 40700001 00001004
 109ce68c:			0020: 63818000 00011007 63818001 00001008 63820000 0001100b 63820001 0000100c
@@ -1086,7 +1089,8 @@ t3			opcode: CP_LOAD_STATE4 (30) (35 dwords)
 
 				- output (half): (cnt=0, max=0)  (estimated)
 				- output (full): (cnt=0, max=0)  (estimated)
-				- shaderdb: 11 instructions, 5 nops, 6 non-nops, (11 instlen), 0 half, 1 full
+				- shaderdb: 11 instructions, 5 nops, 6 non-nops, (11 instlen), 5 last-baryf, 0 half, 1 full
+				- shaderdb: 6 cat0, 0 cat1, 1 cat2, 0 cat3, 0 cat4, 0 cat5, 4 cat6, 0 cat7
 				- shaderdb: 1 (ss), 0 (sy)
 109ce878:			0000: c0213000 00700000 00000000 00000000 00000000 01c00000 c7c60000 01c00002
 109ce898:			0020: c7c60001 01c00004 c7c60002 01c00006 c7c60003 00002000 473090fc 00000000
@@ -1675,7 +1679,8 @@ t3			opcode: CP_LOAD_STATE4 (30) (131 dwords)
 
 				- output (half): (cnt=0, max=0)  (estimated)
 				- output (full): 10-17 (cnt=8, max=17)  (estimated)
-				- shaderdb: 67 instructions, 31 nops, 36 non-nops, (56 instlen), 0 half, 5 full
+				- shaderdb: 67 instructions, 23 nops, 44 non-nops, (56 instlen), 0 last-baryf, 0 half, 5 full
+				- shaderdb: 24 cat0, 5 cat1, 15 cat2, 22 cat3, 1 cat4, 0 cat5, 0 cat6, 0 cat7
 				- shaderdb: 1 (ss), 0 (sy)
 109cee34:			0000: c0813000 01200000 00000000 10000002 40700000 10030002 40700001 00001004
 109cee54:			0020: 63818000 00011007 63818001 00001008 63820000 0001100b 63820001 0000100c
@@ -1715,7 +1720,8 @@ t3			opcode: CP_LOAD_STATE4 (30) (35 dwords)
 
 				- output (half): (cnt=0, max=0)  (estimated)
 				- output (full): (cnt=0, max=0)  (estimated)
-				- shaderdb: 11 instructions, 5 nops, 6 non-nops, (11 instlen), 0 half, 1 full
+				- shaderdb: 11 instructions, 5 nops, 6 non-nops, (11 instlen), 5 last-baryf, 0 half, 1 full
+				- shaderdb: 6 cat0, 0 cat1, 1 cat2, 0 cat3, 0 cat4, 0 cat5, 4 cat6, 0 cat7
 				- shaderdb: 1 (ss), 0 (sy)
 109cf040:			0000: c0213000 00700000 00000000 00000000 00000000 01c00000 c7c60000 01c00002
 109cf060:			0020: c7c60001 01c00004 c7c60002 01c00006 c7c60003 00002000 473090fc 00000000
@@ -2106,7 +2112,8 @@ t3			opcode: CP_LOAD_STATE4 (30) (131 dwords)
 
 				- output (half): (cnt=0, max=0)  (estimated)
 				- output (full): 10-17 (cnt=8, max=17)  (estimated)
-				- shaderdb: 67 instructions, 31 nops, 36 non-nops, (56 instlen), 0 half, 5 full
+				- shaderdb: 67 instructions, 23 nops, 44 non-nops, (56 instlen), 0 last-baryf, 0 half, 5 full
+				- shaderdb: 24 cat0, 5 cat1, 15 cat2, 22 cat3, 1 cat4, 0 cat5, 0 cat6, 0 cat7
 				- shaderdb: 1 (ss), 0 (sy)
 109cf40c:			0000: c0813000 01200000 00000000 10000002 40700000 10030002 40700001 00001004
 109cf42c:			0020: 63818000 00011007 63818001 00001008 63820000 0001100b 63820001 0000100c
@@ -2144,7 +2151,8 @@ t3			opcode: CP_LOAD_STATE4 (30) (35 dwords)
 
 				- output (half): (cnt=0, max=0)  (estimated)
 				- output (full): 2-5 (cnt=4, max=5)  (estimated)
-				- shaderdb: 9 instructions, 4 nops, 5 non-nops, (9 instlen), 0 half, 2 full
+				- shaderdb: 9 instructions, 4 nops, 5 non-nops, (9 instlen), 3 last-baryf, 0 half, 2 full
+				- shaderdb: 5 cat0, 0 cat1, 4 cat2, 0 cat3, 0 cat4, 0 cat5, 0 cat6, 0 cat7
 				- shaderdb: 0 (ss), 0 (sy)
 109cf618:			0000: c0213000 00700000 00000000 00002000 47300002 00002001 47300003 00002002
 109cf638:			0020: 47300004 00002003 47308005 00000000 03000000 00000000 00000000 00000000
@@ -2498,7 +2506,8 @@ t3			opcode: CP_LOAD_STATE4 (30) (131 dwords)
 
 				- output (half): (cnt=0, max=0)  (estimated)
 				- output (full): 6-13 (cnt=8, max=13)  (estimated)
-				- shaderdb: 74 instructions, 38 nops, 36 non-nops, (61 instlen), 0 half, 4 full
+				- shaderdb: 74 instructions, 27 nops, 47 non-nops, (61 instlen), 0 last-baryf, 0 half, 4 full
+				- shaderdb: 28 cat0, 8 cat1, 15 cat2, 22 cat3, 1 cat4, 0 cat5, 0 cat6, 0 cat7
 				- shaderdb: 1 (ss), 0 (sy)
 109cf96c:			0000: c0813000 01200000 00000000 10000002 40700000 10030002 40700001 00001004
 109cf98c:			0020: 63818000 00011007 63818001 00001008 63820000 0001100b 63820001 0000100c
@@ -2539,7 +2548,8 @@ t3			opcode: CP_LOAD_STATE4 (30) (35 dwords)
 
 				- output (half): (cnt=0, max=0)  (estimated)
 				- output (full): (cnt=0, max=0)  (estimated)
-				- shaderdb: 11 instructions, 5 nops, 6 non-nops, (11 instlen), 0 half, 1 full
+				- shaderdb: 11 instructions, 5 nops, 6 non-nops, (11 instlen), 5 last-baryf, 0 half, 1 full
+				- shaderdb: 6 cat0, 0 cat1, 1 cat2, 0 cat3, 0 cat4, 0 cat5, 4 cat6, 0 cat7
 				- shaderdb: 1 (ss), 0 (sy)
 109cfb78:			0000: c0213000 00700000 00000000 00000000 00000000 01c00000 c7c60000 01c00002
 109cfb98:			0020: c7c60001 01c00004 c7c60002 01c00006 c7c60003 00002000 473090fc 00000000
@@ -3051,7 +3061,8 @@ t3			opcode: CP_LOAD_STATE4 (30) (131 dwords)
 
 				- output (half): (cnt=0, max=0)  (estimated)
 				- output (full): 10-17 (cnt=8, max=17)  (estimated)
-				- shaderdb: 67 instructions, 31 nops, 36 non-nops, (56 instlen), 0 half, 5 full
+				- shaderdb: 67 instructions, 23 nops, 44 non-nops, (56 instlen), 0 last-baryf, 0 half, 5 full
+				- shaderdb: 24 cat0, 5 cat1, 15 cat2, 22 cat3, 1 cat4, 0 cat5, 0 cat6, 0 cat7
 				- shaderdb: 1 (ss), 0 (sy)
 109d00b4:			0000: c0813000 01200000 00000000 10000002 40700000 10030002 40700001 00001004
 109d00d4:			0020: 63818000 00011007 63818001 00001008 63820000 0001100b 63820001 0000100c
@@ -3091,7 +3102,8 @@ t3			opcode: CP_LOAD_STATE4 (30) (35 dwords)
 
 				- output (half): (cnt=0, max=0)  (estimated)
 				- output (full): (cnt=0, max=0)  (estimated)
-				- shaderdb: 11 instructions, 5 nops, 6 non-nops, (11 instlen), 0 half, 1 full
+				- shaderdb: 11 instructions, 5 nops, 6 non-nops, (11 instlen), 5 last-baryf, 0 half, 1 full
+				- shaderdb: 6 cat0, 0 cat1, 1 cat2, 0 cat3, 0 cat4, 0 cat5, 4 cat6, 0 cat7
 				- shaderdb: 1 (ss), 0 (sy)
 109d02c0:			0000: c0213000 00700000 00000000 00000000 00000000 01c00000 c7c60000 01c00002
 109d02e0:			0020: c7c60001 01c00004 c7c60002 01c00006 c7c60003 00002000 473090fc 00000000
@@ -3482,7 +3494,8 @@ t3			opcode: CP_LOAD_STATE4 (30) (131 dwords)
 
 				- output (half): (cnt=0, max=0)  (estimated)
 				- output (full): 10-17 (cnt=8, max=17)  (estimated)
-				- shaderdb: 67 instructions, 31 nops, 36 non-nops, (56 instlen), 0 half, 5 full
+				- shaderdb: 67 instructions, 23 nops, 44 non-nops, (56 instlen), 0 last-baryf, 0 half, 5 full
+				- shaderdb: 24 cat0, 5 cat1, 15 cat2, 22 cat3, 1 cat4, 0 cat5, 0 cat6, 0 cat7
 				- shaderdb: 1 (ss), 0 (sy)
 109d068c:			0000: c0813000 01200000 00000000 10000002 40700000 10030002 40700001 00001004
 109d06ac:			0020: 63818000 00011007 63818001 00001008 63820000 0001100b 63820001 0000100c
@@ -3520,7 +3533,8 @@ t3			opcode: CP_LOAD_STATE4 (30) (35 dwords)
 
 				- output (half): (cnt=0, max=0)  (estimated)
 				- output (full): 2-5 (cnt=4, max=5)  (estimated)
-				- shaderdb: 9 instructions, 4 nops, 5 non-nops, (9 instlen), 0 half, 2 full
+				- shaderdb: 9 instructions, 4 nops, 5 non-nops, (9 instlen), 3 last-baryf, 0 half, 2 full
+				- shaderdb: 5 cat0, 0 cat1, 4 cat2, 0 cat3, 0 cat4, 0 cat5, 0 cat6, 0 cat7
 				- shaderdb: 0 (ss), 0 (sy)
 109d0898:			0000: c0213000 00700000 00000000 00002000 47300002 00002001 47300003 00002002
 109d08b8:			0020: 47300004 00002003 47308005 00000000 03000000 00000000 00000000 00000000
@@ -3874,7 +3888,8 @@ t3			opcode: CP_LOAD_STATE4 (30) (131 dwords)
 
 				- output (half): (cnt=0, max=0)  (estimated)
 				- output (full): 6-13 (cnt=8, max=13)  (estimated)
-				- shaderdb: 74 instructions, 38 nops, 36 non-nops, (61 instlen), 0 half, 4 full
+				- shaderdb: 74 instructions, 27 nops, 47 non-nops, (61 instlen), 0 last-baryf, 0 half, 4 full
+				- shaderdb: 28 cat0, 8 cat1, 15 cat2, 22 cat3, 1 cat4, 0 cat5, 0 cat6, 0 cat7
 				- shaderdb: 1 (ss), 0 (sy)
 109d0bec:			0000: c0813000 01200000 00000000 10000002 40700000 10030002 40700001 00001004
 109d0c0c:			0020: 63818000 00011007 63818001 00001008 63820000 0001100b 63820001 0000100c
@@ -3915,7 +3930,8 @@ t3			opcode: CP_LOAD_STATE4 (30) (35 dwords)
 
 				- output (half): (cnt=0, max=0)  (estimated)
 				- output (full): (cnt=0, max=0)  (estimated)
-				- shaderdb: 11 instructions, 5 nops, 6 non-nops, (11 instlen), 0 half, 1 full
+				- shaderdb: 11 instructions, 5 nops, 6 non-nops, (11 instlen), 5 last-baryf, 0 half, 1 full
+				- shaderdb: 6 cat0, 0 cat1, 1 cat2, 0 cat3, 0 cat4, 0 cat5, 4 cat6, 0 cat7
 				- shaderdb: 1 (ss), 0 (sy)
 109d0df8:			0000: c0213000 00700000 00000000 00000000 00000000 01c00000 c7c60000 01c00002
 109d0e18:			0020: c7c60001 01c00004 c7c60002 01c00006 c7c60003 00002000 473090fc 00000000
@@ -4427,7 +4443,8 @@ t3			opcode: CP_LOAD_STATE4 (30) (131 dwords)
 
 				- output (half): (cnt=0, max=0)  (estimated)
 				- output (full): 10-17 (cnt=8, max=17)  (estimated)
-				- shaderdb: 67 instructions, 31 nops, 36 non-nops, (56 instlen), 0 half, 5 full
+				- shaderdb: 67 instructions, 23 nops, 44 non-nops, (56 instlen), 0 last-baryf, 0 half, 5 full
+				- shaderdb: 24 cat0, 5 cat1, 15 cat2, 22 cat3, 1 cat4, 0 cat5, 0 cat6, 0 cat7
 				- shaderdb: 1 (ss), 0 (sy)
 109d1334:			0000: c0813000 01200000 00000000 10000002 40700000 10030002 40700001 00001004
 109d1354:			0020: 63818000 00011007 63818001 00001008 63820000 0001100b 63820001 0000100c
@@ -4467,7 +4484,8 @@ t3			opcode: CP_LOAD_STATE4 (30) (35 dwords)
 
 				- output (half): (cnt=0, max=0)  (estimated)
 				- output (full): (cnt=0, max=0)  (estimated)
-				- shaderdb: 11 instructions, 5 nops, 6 non-nops, (11 instlen), 0 half, 1 full
+				- shaderdb: 11 instructions, 5 nops, 6 non-nops, (11 instlen), 5 last-baryf, 0 half, 1 full
+				- shaderdb: 6 cat0, 0 cat1, 1 cat2, 0 cat3, 0 cat4, 0 cat5, 4 cat6, 0 cat7
 				- shaderdb: 1 (ss), 0 (sy)
 109d1540:			0000: c0213000 00700000 00000000 00000000 00000000 01c00000 c7c60000 01c00002
 109d1560:			0020: c7c60001 01c00004 c7c60002 01c00006 c7c60003 00002000 473090fc 00000000
@@ -4858,7 +4876,8 @@ t3			opcode: CP_LOAD_STATE4 (30) (131 dwords)
 
 				- output (half): (cnt=0, max=0)  (estimated)
 				- output (full): 10-17 (cnt=8, max=17)  (estimated)
-				- shaderdb: 67 instructions, 31 nops, 36 non-nops, (56 instlen), 0 half, 5 full
+				- shaderdb: 67 instructions, 23 nops, 44 non-nops, (56 instlen), 0 last-baryf, 0 half, 5 full
+				- shaderdb: 24 cat0, 5 cat1, 15 cat2, 22 cat3, 1 cat4, 0 cat5, 0 cat6, 0 cat7
 				- shaderdb: 1 (ss), 0 (sy)
 109d190c:			0000: c0813000 01200000 00000000 10000002 40700000 10030002 40700001 00001004
 109d192c:			0020: 63818000 00011007 63818001 00001008 63820000 0001100b 63820001 0000100c
@@ -4896,7 +4915,8 @@ t3			opcode: CP_LOAD_STATE4 (30) (35 dwords)
 
 				- output (half): (cnt=0, max=0)  (estimated)
 				- output (full): 2-5 (cnt=4, max=5)  (estimated)
-				- shaderdb: 9 instructions, 4 nops, 5 non-nops, (9 instlen), 0 half, 2 full
+				- shaderdb: 9 instructions, 4 nops, 5 non-nops, (9 instlen), 3 last-baryf, 0 half, 2 full
+				- shaderdb: 5 cat0, 0 cat1, 4 cat2, 0 cat3, 0 cat4, 0 cat5, 0 cat6, 0 cat7
 				- shaderdb: 0 (ss), 0 (sy)
 109d1b18:			0000: c0213000 00700000 00000000 00002000 47300002 00002001 47300003 00002002
 109d1b38:			0020: 47300004 00002003 47308005 00000000 03000000 00000000 00000000 00000000
@@ -5200,7 +5220,8 @@ t3		opcode: CP_LOAD_STATE4 (30) (35 dwords)
 
 			- output (half): (cnt=0, max=0)  (estimated)
 			- output (full): (cnt=0, max=0)  (estimated)
-			- shaderdb: 5 instructions, 4 nops, 1 non-nops, (5 instlen), 0 half, 0 full
+			- shaderdb: 5 instructions, 4 nops, 1 non-nops, (5 instlen), 0 last-baryf, 0 half, 0 full
+			- shaderdb: 5 cat0, 0 cat1, 0 cat2, 0 cat3, 0 cat4, 0 cat5, 0 cat6, 0 cat7
 			- shaderdb: 0 (ss), 0 (sy)
 108ce4b0:		0000: c0213000 00600000 00000000 00000000 03000000 00000000 00000000 00000000
 *
@@ -5225,7 +5246,8 @@ t3		opcode: CP_LOAD_STATE4 (30) (35 dwords)
 
 			- output (half): (cnt=0, max=0)  (estimated)
 			- output (full): 0-3 (cnt=4, max=3)  (estimated)
-			- shaderdb: 9 instructions, 8 nops, 1 non-nops, (9 instlen), 0 half, 1 full
+			- shaderdb: 9 instructions, 4 nops, 5 non-nops, (9 instlen), 0 last-baryf, 0 half, 1 full
+			- shaderdb: 5 cat0, 4 cat1, 0 cat2, 0 cat3, 0 cat4, 0 cat5, 0 cat6, 0 cat7
 			- shaderdb: 0 (ss), 0 (sy)
 108ce53c:		0000: c0213000 00700000 00000000 00000000 20244000 00000001 20244001 00000002
 108ce55c:		0020: 20244002 00000003 20244003 00000000 03000000 00000000 00000000 00000000
diff --git a/src/freedreno/common/disasm.h b/src/freedreno/common/disasm.h
index 1a2993e6f42..0e1d8502ed7 100644
--- a/src/freedreno/common/disasm.h
+++ b/src/freedreno/common/disasm.h
@@ -44,6 +44,10 @@ struct shader_stats {
 	int nops;
 	int ss, sy;
 	int constlen;
+	uint16_t mov_count;
+	uint16_t cov_count;
+	uint16_t last_baryf;
+	uint16_t instrs_per_cat[8];
 };
 
 int disasm_a2xx(uint32_t *dwords, int sizedwords, int level, gl_shader_stage type);
diff --git a/src/freedreno/ir3/disasm-a3xx.c b/src/freedreno/ir3/disasm-a3xx.c
index b84163106fb..2555c739082 100644
--- a/src/freedreno/ir3/disasm-a3xx.c
+++ b/src/freedreno/ir3/disasm-a3xx.c
@@ -277,10 +277,21 @@ static void print_reg_stats(struct disasm_ctx *ctx)
 	// Note this count of instructions includes rptN, which matches
 	// up to how mesa prints this:
 	fprintf(ctx->out, "%s- shaderdb: %d instructions, %d nops, %d non-nops, "
-			"(%d instlen), %d half, %d full\n",
+			"(%d instlen), %u last-baryf, %d half, %d full\n",
 			levels[ctx->level], ctx->stats->instructions, ctx->stats->nops,
 			ctx->stats->instructions - ctx->stats->nops, ctx->stats->instlen,
-			halfreg, fullreg);
+			ctx->stats->last_baryf, halfreg, fullreg);
+	fprintf(ctx->out, "%s- shaderdb: %u cat0, %u cat1, %u cat2, %u cat3, "
+			"%u cat4, %u cat5, %u cat6, %u cat7\n",
+			levels[ctx->level],
+			ctx->stats->instrs_per_cat[0],
+			ctx->stats->instrs_per_cat[1],
+			ctx->stats->instrs_per_cat[2],
+			ctx->stats->instrs_per_cat[3],
+			ctx->stats->instrs_per_cat[4],
+			ctx->stats->instrs_per_cat[5],
+			ctx->stats->instrs_per_cat[6],
+			ctx->stats->instrs_per_cat[7]);
 	fprintf(ctx->out, "%s- shaderdb: %d (ss), %d (sy)\n", levels[ctx->level],
 			ctx->stats->ss, ctx->stats->sy);
 }
@@ -1529,7 +1540,7 @@ static void print_single_instr(struct disasm_ctx *ctx, instr_t *instr)
 static bool print_instr(struct disasm_ctx *ctx, uint32_t *dwords, int n)
 {
 	instr_t *instr = (instr_t *)dwords;
-	uint32_t opc = instr_opc(instr, ctx->gpu_id);
+	opc_t opc = _OPC(instr->opc_cat, instr_opc(instr, ctx->gpu_id));
 	unsigned nop = 0;
 	unsigned cycles = ctx->stats->instructions;
 
@@ -1538,15 +1549,18 @@ static bool print_instr(struct disasm_ctx *ctx, uint32_t *dwords, int n)
 				instr->opc_cat, n, cycles++, dwords[1], dwords[0]);
 	}
 
-	/* NOTE: order flags are printed is a bit fugly.. but for now I
-	 * try to match the order in llvm-a3xx disassembler for easy
-	 * diff'ing..
-	 */
+	if (opc == OPC_BARY_F)
+		ctx->stats->last_baryf = ctx->stats->instructions;
 
 	ctx->repeat = instr_repeat(instr);
 	ctx->stats->instructions += 1 + ctx->repeat;
 	ctx->stats->instlen++;
 
+	/* NOTE: order flags are printed is a bit fugly.. but for now I
+	 * try to match the order in llvm-a3xx disassembler for easy
+	 * diff'ing..
+	 */
+
 	if (instr->sync) {
 		fprintf(ctx->out, "(sy)");
 		ctx->stats->sy++;
@@ -1567,16 +1581,30 @@ static bool print_instr(struct disasm_ctx *ctx, uint32_t *dwords, int n)
 		nop = (instr->cat2.src2_r * 2) + instr->cat2.src1_r;
 	else if ((instr->opc_cat == 3) && (instr->cat3.src1_r || instr->cat3.src2_r))
 		nop = (instr->cat3.src2_r * 2) + instr->cat3.src1_r;
-	ctx->stats->instructions += nop;
-	ctx->stats->nops += nop;
-	if (opc == OPC_NOP)
-		ctx->stats->nops += 1 + ctx->repeat;
 	if (nop)
 		fprintf(ctx->out, "(nop%d) ", nop);
 
 	if (instr->ul && ((2 <= instr->opc_cat) && (instr->opc_cat <= 4)))
 		fprintf(ctx->out, "(ul)");
 
+	ctx->stats->instructions += nop;
+	ctx->stats->nops += nop;
+	if (opc == OPC_NOP) {
+		ctx->stats->nops += 1 + ctx->repeat;
+		ctx->stats->instrs_per_cat[0] += 1 + ctx->repeat;
+	} else {
+		ctx->stats->instrs_per_cat[instr->opc_cat] += 1 + ctx->repeat;
+		ctx->stats->instrs_per_cat[0] += nop;
+	}
+
+	if (opc == OPC_MOV) {
+		if (instr->cat1.src_type == instr->cat1.dst_type) {
+			ctx->stats->mov_count += 1 + ctx->repeat;
+		} else {
+			ctx->stats->cov_count += 1 + ctx->repeat;
+		}
+	}
+
 	print_single_instr(ctx, instr);
 	fprintf(ctx->out, "\n");