freedreno/ir3: Realign disasm shader stats

To better match up with what mesa shader-db stats look like, for easier
comparision.

Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7997>

3 files changed, 110 insertions, 98 deletions

diff --git a/src/freedreno/.gitlab-ci/reference/glxgears-a420.log b/src/freedreno/.gitlab-ci/reference/glxgears-a420.log
index c5ef966f48e..576aa98a301 100644
--- a/src/freedreno/.gitlab-ci/reference/glxgears-a420.log
+++ b/src/freedreno/.gitlab-ci/reference/glxgears-a420.log
@@ -1047,7 +1047,7 @@ t3			opcode: CP_LOAD_STATE4 (30) (131 dwords)
 				- output (full): 6-13 (cnt=8, max=13)  (estimated)
 
 				- shaderdb: 74 instr, 27 nops, 47 non-nops, 7 mov, 1 cov
-				- shaderdb: 0 last-baryf, 0 half, 4 full, 14 constlen
+				- shaderdb: 0 last-baryf, 0 half, 4 full, 13 constlen
 				- shaderdb: 28 cat0, 8 cat1, 15 cat2, 22 cat3, 1 cat4, 0 cat5, 0 cat6, 0 cat7
 				- shaderdb: 10 sstall, 1 (ss), 0 (sy)
 109ce66c:			0000: c0813000 01200000 00000000 10000002 40700000 10030002 40700001 00001004
@@ -1091,7 +1091,7 @@ t3			opcode: CP_LOAD_STATE4 (30) (35 dwords)
 				- shaderdb: 11 instr, 5 nops, 6 non-nops, 0 mov, 0 cov
 				- shaderdb: 5 last-baryf, 0 half, 1 full, 0 constlen
 				- shaderdb: 6 cat0, 0 cat1, 1 cat2, 0 cat3, 0 cat4, 0 cat5, 4 cat6, 0 cat7
-				- shaderdb: 65531 sstall, 1 (ss), 0 (sy)
+				- shaderdb: 0 sstall, 1 (ss), 0 (sy)
 109ce878:			0000: c0213000 00700000 00000000 00000000 00000000 01c00000 c7c60000 01c00002
 109ce898:			0020: c7c60001 01c00004 c7c60002 01c00006 c7c60003 00002000 473090fc 00000000
 109ce8b8:			0040: 03000000 00000000 00000000 00000000 00000000 00000000 00000000 00000000
@@ -1679,7 +1679,7 @@ t3			opcode: CP_LOAD_STATE4 (30) (131 dwords)
 				- output (full): 10-17 (cnt=8, max=17)  (estimated)
 
 				- shaderdb: 67 instr, 23 nops, 44 non-nops, 4 mov, 1 cov
-				- shaderdb: 0 last-baryf, 0 half, 5 full, 14 constlen
+				- shaderdb: 0 last-baryf, 0 half, 5 full, 13 constlen
 				- shaderdb: 24 cat0, 5 cat1, 15 cat2, 22 cat3, 1 cat4, 0 cat5, 0 cat6, 0 cat7
 				- shaderdb: 10 sstall, 1 (ss), 0 (sy)
 109cee34:			0000: c0813000 01200000 00000000 10000002 40700000 10030002 40700001 00001004
@@ -1716,13 +1716,13 @@ t3			opcode: CP_LOAD_STATE4 (30) (35 dwords)
 				- used (full): 0-3 (cnt=4, max=3)
 				- input (half): (cnt=0, max=0)
 				- input (full): 0-3 (cnt=4, max=3)
-				- max const: 0
-
 				- output (half): (cnt=0, max=0)  (estimated)
 				- output (full): (cnt=0, max=0)  (estimated)
-				- shaderdb: 11 instructions, 5 nops, 6 non-nops, (11 instlen), 5 last-baryf, 0 half, 1 full
+
+				- shaderdb: 11 instr, 5 nops, 6 non-nops, 0 mov, 0 cov
+				- shaderdb: 5 last-baryf, 0 half, 1 full, 0 constlen
 				- shaderdb: 6 cat0, 0 cat1, 1 cat2, 0 cat3, 0 cat4, 0 cat5, 4 cat6, 0 cat7
-				- shaderdb: 1 (ss), 0 (sy)
+				- shaderdb: 0 sstall, 1 (ss), 0 (sy)
 109cf040:			0000: c0213000 00700000 00000000 00000000 00000000 01c00000 c7c60000 01c00002
 109cf060:			0020: c7c60001 01c00004 c7c60002 01c00006 c7c60003 00002000 473090fc 00000000
 109cf080:			0040: 03000000 00000000 00000000 00000000 00000000 00000000 00000000 00000000
@@ -2108,13 +2108,13 @@ t3			opcode: CP_LOAD_STATE4 (30) (131 dwords)
 				- used (full): 0-8 10-17 (cnt=17, max=17)
 				- input (half): (cnt=0, max=0)
 				- input (full): 2-8 (cnt=7, max=8)
-				- max const: 52
-
 				- output (half): (cnt=0, max=0)  (estimated)
 				- output (full): 10-17 (cnt=8, max=17)  (estimated)
-				- shaderdb: 67 instructions, 23 nops, 44 non-nops, (56 instlen), 0 last-baryf, 0 half, 5 full
+
+				- shaderdb: 67 instr, 23 nops, 44 non-nops, 4 mov, 1 cov
+				- shaderdb: 0 last-baryf, 0 half, 5 full, 13 constlen
 				- shaderdb: 24 cat0, 5 cat1, 15 cat2, 22 cat3, 1 cat4, 0 cat5, 0 cat6, 0 cat7
-				- shaderdb: 1 (ss), 0 (sy)
+				- shaderdb: 10 sstall, 1 (ss), 0 (sy)
 109cf40c:			0000: c0813000 01200000 00000000 10000002 40700000 10030002 40700001 00001004
 109cf42c:			0020: 63818000 00011007 63818001 00001008 63820000 0001100b 63820001 0000100c
 109cf44c:			0040: 6382800a 10010002 40700000 0001100f 6382800d 00001005 63818000 00000010
@@ -2147,13 +2147,13 @@ t3			opcode: CP_LOAD_STATE4 (30) (35 dwords)
 				- used (full): 0 2-5 (cnt=5, max=5)
 				- input (half): (cnt=0, max=0)
 				- input (full): 0 (cnt=1, max=0)
-				- max const: 0
-
 				- 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), 3 last-baryf, 0 half, 2 full
+
+				- shaderdb: 9 instr, 4 nops, 5 non-nops, 0 mov, 0 cov
+				- shaderdb: 3 last-baryf, 0 half, 2 full, 0 constlen
 				- shaderdb: 5 cat0, 0 cat1, 4 cat2, 0 cat3, 0 cat4, 0 cat5, 0 cat6, 0 cat7
-				- shaderdb: 0 (ss), 0 (sy)
+				- shaderdb: 0 sstall, 0 (ss), 0 (sy)
 109cf618:			0000: c0213000 00700000 00000000 00002000 47300002 00002001 47300003 00002002
 109cf638:			0020: 47300004 00002003 47308005 00000000 03000000 00000000 00000000 00000000
 *
@@ -2502,13 +2502,13 @@ t3			opcode: CP_LOAD_STATE4 (30) (131 dwords)
 				- used (full): 0-13 (cnt=14, max=13)
 				- input (half): (cnt=0, max=0)
 				- input (full): 2-5 (cnt=4, max=5)
-				- max const: 52
-
 				- output (half): (cnt=0, max=0)  (estimated)
 				- output (full): 6-13 (cnt=8, max=13)  (estimated)
-				- shaderdb: 74 instructions, 27 nops, 47 non-nops, (61 instlen), 0 last-baryf, 0 half, 4 full
+
+				- shaderdb: 74 instr, 27 nops, 47 non-nops, 7 mov, 1 cov
+				- shaderdb: 0 last-baryf, 0 half, 4 full, 13 constlen
 				- shaderdb: 28 cat0, 8 cat1, 15 cat2, 22 cat3, 1 cat4, 0 cat5, 0 cat6, 0 cat7
-				- shaderdb: 1 (ss), 0 (sy)
+				- shaderdb: 10 sstall, 1 (ss), 0 (sy)
 109cf96c:			0000: c0813000 01200000 00000000 10000002 40700000 10030002 40700001 00001004
 109cf98c:			0020: 63818000 00011007 63818001 00001008 63820000 0001100b 63820001 0000100c
 109cf9ac:			0040: 63828006 10010002 40700000 0001100f 63828009 00001005 63818000 00000010
@@ -2544,13 +2544,13 @@ t3			opcode: CP_LOAD_STATE4 (30) (35 dwords)
 				- used (full): 0-3 (cnt=4, max=3)
 				- input (half): (cnt=0, max=0)
 				- input (full): 0-3 (cnt=4, max=3)
-				- max const: 0
-
 				- output (half): (cnt=0, max=0)  (estimated)
 				- output (full): (cnt=0, max=0)  (estimated)
-				- shaderdb: 11 instructions, 5 nops, 6 non-nops, (11 instlen), 5 last-baryf, 0 half, 1 full
+
+				- shaderdb: 11 instr, 5 nops, 6 non-nops, 0 mov, 0 cov
+				- shaderdb: 5 last-baryf, 0 half, 1 full, 0 constlen
 				- shaderdb: 6 cat0, 0 cat1, 1 cat2, 0 cat3, 0 cat4, 0 cat5, 4 cat6, 0 cat7
-				- shaderdb: 1 (ss), 0 (sy)
+				- shaderdb: 0 sstall, 1 (ss), 0 (sy)
 109cfb78:			0000: c0213000 00700000 00000000 00000000 00000000 01c00000 c7c60000 01c00002
 109cfb98:			0020: c7c60001 01c00004 c7c60002 01c00006 c7c60003 00002000 473090fc 00000000
 109cfbb8:			0040: 03000000 00000000 00000000 00000000 00000000 00000000 00000000 00000000
@@ -3057,13 +3057,13 @@ t3			opcode: CP_LOAD_STATE4 (30) (131 dwords)
 				- used (full): 0-8 10-17 (cnt=17, max=17)
 				- input (half): (cnt=0, max=0)
 				- input (full): 2-8 (cnt=7, max=8)
-				- max const: 52
-
 				- output (half): (cnt=0, max=0)  (estimated)
 				- output (full): 10-17 (cnt=8, max=17)  (estimated)
-				- shaderdb: 67 instructions, 23 nops, 44 non-nops, (56 instlen), 0 last-baryf, 0 half, 5 full
+
+				- shaderdb: 67 instr, 23 nops, 44 non-nops, 4 mov, 1 cov
+				- shaderdb: 0 last-baryf, 0 half, 5 full, 13 constlen
 				- shaderdb: 24 cat0, 5 cat1, 15 cat2, 22 cat3, 1 cat4, 0 cat5, 0 cat6, 0 cat7
-				- shaderdb: 1 (ss), 0 (sy)
+				- shaderdb: 10 sstall, 1 (ss), 0 (sy)
 109d00b4:			0000: c0813000 01200000 00000000 10000002 40700000 10030002 40700001 00001004
 109d00d4:			0020: 63818000 00011007 63818001 00001008 63820000 0001100b 63820001 0000100c
 109d00f4:			0040: 6382800a 10010002 40700000 0001100f 6382800d 00001005 63818000 00000010
@@ -3098,13 +3098,13 @@ t3			opcode: CP_LOAD_STATE4 (30) (35 dwords)
 				- used (full): 0-3 (cnt=4, max=3)
 				- input (half): (cnt=0, max=0)
 				- input (full): 0-3 (cnt=4, max=3)
-				- max const: 0
-
 				- output (half): (cnt=0, max=0)  (estimated)
 				- output (full): (cnt=0, max=0)  (estimated)
-				- shaderdb: 11 instructions, 5 nops, 6 non-nops, (11 instlen), 5 last-baryf, 0 half, 1 full
+
+				- shaderdb: 11 instr, 5 nops, 6 non-nops, 0 mov, 0 cov
+				- shaderdb: 5 last-baryf, 0 half, 1 full, 0 constlen
 				- shaderdb: 6 cat0, 0 cat1, 1 cat2, 0 cat3, 0 cat4, 0 cat5, 4 cat6, 0 cat7
-				- shaderdb: 1 (ss), 0 (sy)
+				- shaderdb: 0 sstall, 1 (ss), 0 (sy)
 109d02c0:			0000: c0213000 00700000 00000000 00000000 00000000 01c00000 c7c60000 01c00002
 109d02e0:			0020: c7c60001 01c00004 c7c60002 01c00006 c7c60003 00002000 473090fc 00000000
 109d0300:			0040: 03000000 00000000 00000000 00000000 00000000 00000000 00000000 00000000
@@ -3490,13 +3490,13 @@ t3			opcode: CP_LOAD_STATE4 (30) (131 dwords)
 				- used (full): 0-8 10-17 (cnt=17, max=17)
 				- input (half): (cnt=0, max=0)
 				- input (full): 2-8 (cnt=7, max=8)
-				- max const: 52
-
 				- output (half): (cnt=0, max=0)  (estimated)
 				- output (full): 10-17 (cnt=8, max=17)  (estimated)
-				- shaderdb: 67 instructions, 23 nops, 44 non-nops, (56 instlen), 0 last-baryf, 0 half, 5 full
+
+				- shaderdb: 67 instr, 23 nops, 44 non-nops, 4 mov, 1 cov
+				- shaderdb: 0 last-baryf, 0 half, 5 full, 13 constlen
 				- shaderdb: 24 cat0, 5 cat1, 15 cat2, 22 cat3, 1 cat4, 0 cat5, 0 cat6, 0 cat7
-				- shaderdb: 1 (ss), 0 (sy)
+				- shaderdb: 10 sstall, 1 (ss), 0 (sy)
 109d068c:			0000: c0813000 01200000 00000000 10000002 40700000 10030002 40700001 00001004
 109d06ac:			0020: 63818000 00011007 63818001 00001008 63820000 0001100b 63820001 0000100c
 109d06cc:			0040: 6382800a 10010002 40700000 0001100f 6382800d 00001005 63818000 00000010
@@ -3529,13 +3529,13 @@ t3			opcode: CP_LOAD_STATE4 (30) (35 dwords)
 				- used (full): 0 2-5 (cnt=5, max=5)
 				- input (half): (cnt=0, max=0)
 				- input (full): 0 (cnt=1, max=0)
-				- max const: 0
-
 				- 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), 3 last-baryf, 0 half, 2 full
+
+				- shaderdb: 9 instr, 4 nops, 5 non-nops, 0 mov, 0 cov
+				- shaderdb: 3 last-baryf, 0 half, 2 full, 0 constlen
 				- shaderdb: 5 cat0, 0 cat1, 4 cat2, 0 cat3, 0 cat4, 0 cat5, 0 cat6, 0 cat7
-				- shaderdb: 0 (ss), 0 (sy)
+				- shaderdb: 0 sstall, 0 (ss), 0 (sy)
 109d0898:			0000: c0213000 00700000 00000000 00002000 47300002 00002001 47300003 00002002
 109d08b8:			0020: 47300004 00002003 47308005 00000000 03000000 00000000 00000000 00000000
 *
@@ -3884,13 +3884,13 @@ t3			opcode: CP_LOAD_STATE4 (30) (131 dwords)
 				- used (full): 0-13 (cnt=14, max=13)
 				- input (half): (cnt=0, max=0)
 				- input (full): 2-5 (cnt=4, max=5)
-				- max const: 52
-
 				- output (half): (cnt=0, max=0)  (estimated)
 				- output (full): 6-13 (cnt=8, max=13)  (estimated)
-				- shaderdb: 74 instructions, 27 nops, 47 non-nops, (61 instlen), 0 last-baryf, 0 half, 4 full
+
+				- shaderdb: 74 instr, 27 nops, 47 non-nops, 7 mov, 1 cov
+				- shaderdb: 0 last-baryf, 0 half, 4 full, 13 constlen
 				- shaderdb: 28 cat0, 8 cat1, 15 cat2, 22 cat3, 1 cat4, 0 cat5, 0 cat6, 0 cat7
-				- shaderdb: 1 (ss), 0 (sy)
+				- shaderdb: 10 sstall, 1 (ss), 0 (sy)
 109d0bec:			0000: c0813000 01200000 00000000 10000002 40700000 10030002 40700001 00001004
 109d0c0c:			0020: 63818000 00011007 63818001 00001008 63820000 0001100b 63820001 0000100c
 109d0c2c:			0040: 63828006 10010002 40700000 0001100f 63828009 00001005 63818000 00000010
@@ -3926,13 +3926,13 @@ t3			opcode: CP_LOAD_STATE4 (30) (35 dwords)
 				- used (full): 0-3 (cnt=4, max=3)
 				- input (half): (cnt=0, max=0)
 				- input (full): 0-3 (cnt=4, max=3)
-				- max const: 0
-
 				- output (half): (cnt=0, max=0)  (estimated)
 				- output (full): (cnt=0, max=0)  (estimated)
-				- shaderdb: 11 instructions, 5 nops, 6 non-nops, (11 instlen), 5 last-baryf, 0 half, 1 full
+
+				- shaderdb: 11 instr, 5 nops, 6 non-nops, 0 mov, 0 cov
+				- shaderdb: 5 last-baryf, 0 half, 1 full, 0 constlen
 				- shaderdb: 6 cat0, 0 cat1, 1 cat2, 0 cat3, 0 cat4, 0 cat5, 4 cat6, 0 cat7
-				- shaderdb: 1 (ss), 0 (sy)
+				- shaderdb: 0 sstall, 1 (ss), 0 (sy)
 109d0df8:			0000: c0213000 00700000 00000000 00000000 00000000 01c00000 c7c60000 01c00002
 109d0e18:			0020: c7c60001 01c00004 c7c60002 01c00006 c7c60003 00002000 473090fc 00000000
 109d0e38:			0040: 03000000 00000000 00000000 00000000 00000000 00000000 00000000 00000000
@@ -4439,13 +4439,13 @@ t3			opcode: CP_LOAD_STATE4 (30) (131 dwords)
 				- used (full): 0-8 10-17 (cnt=17, max=17)
 				- input (half): (cnt=0, max=0)
 				- input (full): 2-8 (cnt=7, max=8)
-				- max const: 52
-
 				- output (half): (cnt=0, max=0)  (estimated)
 				- output (full): 10-17 (cnt=8, max=17)  (estimated)
-				- shaderdb: 67 instructions, 23 nops, 44 non-nops, (56 instlen), 0 last-baryf, 0 half, 5 full
+
+				- shaderdb: 67 instr, 23 nops, 44 non-nops, 4 mov, 1 cov
+				- shaderdb: 0 last-baryf, 0 half, 5 full, 13 constlen
 				- shaderdb: 24 cat0, 5 cat1, 15 cat2, 22 cat3, 1 cat4, 0 cat5, 0 cat6, 0 cat7
-				- shaderdb: 1 (ss), 0 (sy)
+				- shaderdb: 10 sstall, 1 (ss), 0 (sy)
 109d1334:			0000: c0813000 01200000 00000000 10000002 40700000 10030002 40700001 00001004
 109d1354:			0020: 63818000 00011007 63818001 00001008 63820000 0001100b 63820001 0000100c
 109d1374:			0040: 6382800a 10010002 40700000 0001100f 6382800d 00001005 63818000 00000010
@@ -4480,13 +4480,13 @@ t3			opcode: CP_LOAD_STATE4 (30) (35 dwords)
 				- used (full): 0-3 (cnt=4, max=3)
 				- input (half): (cnt=0, max=0)
 				- input (full): 0-3 (cnt=4, max=3)
-				- max const: 0
-
 				- output (half): (cnt=0, max=0)  (estimated)
 				- output (full): (cnt=0, max=0)  (estimated)
-				- shaderdb: 11 instructions, 5 nops, 6 non-nops, (11 instlen), 5 last-baryf, 0 half, 1 full
+
+				- shaderdb: 11 instr, 5 nops, 6 non-nops, 0 mov, 0 cov
+				- shaderdb: 5 last-baryf, 0 half, 1 full, 0 constlen
 				- shaderdb: 6 cat0, 0 cat1, 1 cat2, 0 cat3, 0 cat4, 0 cat5, 4 cat6, 0 cat7
-				- shaderdb: 1 (ss), 0 (sy)
+				- shaderdb: 0 sstall, 1 (ss), 0 (sy)
 109d1540:			0000: c0213000 00700000 00000000 00000000 00000000 01c00000 c7c60000 01c00002
 109d1560:			0020: c7c60001 01c00004 c7c60002 01c00006 c7c60003 00002000 473090fc 00000000
 109d1580:			0040: 03000000 00000000 00000000 00000000 00000000 00000000 00000000 00000000
@@ -4872,13 +4872,13 @@ t3			opcode: CP_LOAD_STATE4 (30) (131 dwords)
 				- used (full): 0-8 10-17 (cnt=17, max=17)
 				- input (half): (cnt=0, max=0)
 				- input (full): 2-8 (cnt=7, max=8)
-				- max const: 52
-
 				- output (half): (cnt=0, max=0)  (estimated)
 				- output (full): 10-17 (cnt=8, max=17)  (estimated)
-				- shaderdb: 67 instructions, 23 nops, 44 non-nops, (56 instlen), 0 last-baryf, 0 half, 5 full
+
+				- shaderdb: 67 instr, 23 nops, 44 non-nops, 4 mov, 1 cov
+				- shaderdb: 0 last-baryf, 0 half, 5 full, 13 constlen
 				- shaderdb: 24 cat0, 5 cat1, 15 cat2, 22 cat3, 1 cat4, 0 cat5, 0 cat6, 0 cat7
-				- shaderdb: 1 (ss), 0 (sy)
+				- shaderdb: 10 sstall, 1 (ss), 0 (sy)
 109d190c:			0000: c0813000 01200000 00000000 10000002 40700000 10030002 40700001 00001004
 109d192c:			0020: 63818000 00011007 63818001 00001008 63820000 0001100b 63820001 0000100c
 109d194c:			0040: 6382800a 10010002 40700000 0001100f 6382800d 00001005 63818000 00000010
@@ -4911,13 +4911,13 @@ t3			opcode: CP_LOAD_STATE4 (30) (35 dwords)
 				- used (full): 0 2-5 (cnt=5, max=5)
 				- input (half): (cnt=0, max=0)
 				- input (full): 0 (cnt=1, max=0)
-				- max const: 0
-
 				- 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), 3 last-baryf, 0 half, 2 full
+
+				- shaderdb: 9 instr, 4 nops, 5 non-nops, 0 mov, 0 cov
+				- shaderdb: 3 last-baryf, 0 half, 2 full, 0 constlen
 				- shaderdb: 5 cat0, 0 cat1, 4 cat2, 0 cat3, 0 cat4, 0 cat5, 0 cat6, 0 cat7
-				- shaderdb: 0 (ss), 0 (sy)
+				- shaderdb: 0 sstall, 0 (ss), 0 (sy)
 109d1b18:			0000: c0213000 00700000 00000000 00002000 47300002 00002001 47300003 00002002
 109d1b38:			0020: 47300004 00002003 47308005 00000000 03000000 00000000 00000000 00000000
 *
@@ -5216,13 +5216,13 @@ t3		opcode: CP_LOAD_STATE4 (30) (35 dwords)
 			- used (full): (cnt=0, max=0)
 			- input (half): (cnt=0, max=0)
 			- input (full): (cnt=0, max=0)
-			- max const: 0
-
 			- 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 last-baryf, 0 half, 0 full
+
+			- shaderdb: 5 instr, 4 nops, 1 non-nops, 0 mov, 0 cov
+			- shaderdb: 0 last-baryf, 0 half, 0 full, 0 constlen
 			- shaderdb: 5 cat0, 0 cat1, 0 cat2, 0 cat3, 0 cat4, 0 cat5, 0 cat6, 0 cat7
-			- shaderdb: 0 (ss), 0 (sy)
+			- shaderdb: 0 sstall, 0 (ss), 0 (sy)
 108ce4b0:		0000: c0213000 00600000 00000000 00000000 03000000 00000000 00000000 00000000
 *
 t3		opcode: CP_LOAD_STATE4 (30) (35 dwords)
@@ -5242,13 +5242,13 @@ t3		opcode: CP_LOAD_STATE4 (30) (35 dwords)
 			- used (full): 0-3 (cnt=4, max=3)
 			- input (half): (cnt=0, max=0)
 			- input (full): (cnt=0, max=0)
-			- max const: 3
-
 			- output (half): (cnt=0, max=0)  (estimated)
 			- output (full): 0-3 (cnt=4, max=3)  (estimated)
-			- shaderdb: 9 instructions, 4 nops, 5 non-nops, (9 instlen), 0 last-baryf, 0 half, 1 full
+
+			- shaderdb: 9 instr, 4 nops, 5 non-nops, 4 mov, 0 cov
+			- shaderdb: 0 last-baryf, 0 half, 1 full, 1 constlen
 			- shaderdb: 5 cat0, 4 cat1, 0 cat2, 0 cat3, 0 cat4, 0 cat5, 0 cat6, 0 cat7
-			- shaderdb: 0 (ss), 0 (sy)
+			- shaderdb: 0 sstall, 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 00b6e7979be..82f2135fbe9 100644
--- a/src/freedreno/common/disasm.h
+++ b/src/freedreno/common/disasm.h
@@ -44,6 +44,7 @@ struct shader_stats {
 	int nops;
 	int ss, sy;
 	int constlen;
+	uint16_t sstall;
 	uint16_t mov_count;
 	uint16_t cov_count;
 	uint16_t last_baryf;
diff --git a/src/freedreno/ir3/disasm-a3xx.c b/src/freedreno/ir3/disasm-a3xx.c
index 64e5f78e6a2..184e814fb4c 100644
--- a/src/freedreno/ir3/disasm-a3xx.c
+++ b/src/freedreno/ir3/disasm-a3xx.c
@@ -91,12 +91,13 @@ struct disasm_ctx {
 	/* current instruction repeat indx/offset (for --expand): */
 	unsigned repeatidx;
 
+	int sfu_delay;
+
 	/* tracking for register usage */
 	struct {
 		regmask_t used;
 		regmask_t rbw;      /* read before write */
 		regmask_t war;      /* write after read */
-		unsigned max_const;
 	} regs;
 };
 
@@ -260,14 +261,13 @@ static void print_reg_stats(struct disasm_ctx *ctx)
 	fprintf(ctx->out, "%s- input (full):", levels[ctx->level]);
 	print_regs(ctx, &ctx->regs.rbw, true);
 	fprintf(ctx->out, "\n");
-	fprintf(ctx->out, "%s- max const: %u\n", levels[ctx->level], ctx->regs.max_const);
-	fprintf(ctx->out, "\n");
 	fprintf(ctx->out, "%s- output (half):", levels[ctx->level]);
 	print_regs(ctx, &ctx->regs.war, false);
 	fprintf(ctx->out, "  (estimated)\n");
 	fprintf(ctx->out, "%s- output (full):", levels[ctx->level]);
 	print_regs(ctx, &ctx->regs.war, true);
 	fprintf(ctx->out, "  (estimated)\n");
+	fprintf(ctx->out, "\n");
 
 	/* convert to vec4, which is the granularity that registers are
 	 * assigned to shader:
@@ -275,15 +275,21 @@ static void print_reg_stats(struct disasm_ctx *ctx)
 	fullreg = (fullreg + 3) / 4;
 	halfreg = ctx->regs.used.mergedregs ? 0 : (halfreg + 3) / 4;
 
-	// 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), %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,
-			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",
+	fprintf(ctx->out, "%s- shaderdb: %u instr, %u nops, %u non-nops, %u mov, %u cov\n",
+			levels[ctx->level],
+			ctx->stats->instructions,
+			ctx->stats->nops,
+			ctx->stats->instructions - ctx->stats->nops,
+			ctx->stats->mov_count, ctx->stats->cov_count);
+
+	fprintf(ctx->out, "%s- shaderdb: %u last-baryf, %d half, %d full, %u constlen\n",
+			levels[ctx->level],
+			ctx->stats->last_baryf,
+			halfreg,
+			fullreg,
+			ctx->stats->constlen);
+
+	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],
@@ -293,8 +299,12 @@ static void print_reg_stats(struct disasm_ctx *ctx)
 			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);
+
+	fprintf(ctx->out, "%s- shaderdb: %u sstall, %u (ss), %u (sy)\n",
+			levels[ctx->level],
+			ctx->stats->sstall,
+			ctx->stats->ss,
+			ctx->stats->sy);
 }
 
 static void process_reg_dst(struct disasm_ctx *ctx)
@@ -364,19 +374,11 @@ static void print_src(struct disasm_ctx *ctx, struct reginfo *info)
 				break;
 		}
 	} else if (info->c) {
-		int i, num = regidx(reg);
-		for (i = 0; i <= ctx->repeat; i++) {
-			unsigned src = num + i;
-
-			ctx->regs.max_const = MAX2(ctx->regs.max_const, src);
-
-			if (!info->r)
-				break;
-		}
-
-		unsigned max = (num + ctx->repeat + 1 + 3) / 4;
-		if (max > ctx->stats->constlen)
-			ctx->stats->constlen = max;
+		unsigned num = regidx(reg);
+		if (info->r)
+			num += ctx->repeat;
+		num = DIV_ROUND_UP(num, 4);
+		ctx->stats->constlen = MAX2(ctx->stats->constlen, num);
 	}
 
 	if (info->r)
@@ -1639,6 +1641,8 @@ static bool print_instr(struct disasm_ctx *ctx, uint32_t *dwords, int n)
 	if (instr->ss && ((instr->opc_cat <= 4) || (instr->opc_cat == 7))) {
 		fprintf(ctx->out, "(ss)");
 		ctx->stats->ss++;
+		ctx->stats->sstall += ctx->sfu_delay;
+		ctx->sfu_delay = 0;
 	}
 	if (instr->jmp_tgt)
 		fprintf(ctx->out, "(jp)");
@@ -1660,6 +1664,13 @@ static bool print_instr(struct disasm_ctx *ctx, uint32_t *dwords, int n)
 	if (instr->ul && ((2 <= instr->opc_cat) && (instr->opc_cat <= 4)))
 		fprintf(ctx->out, "(ul)");
 
+	if (instr->opc_cat == 4) {
+		ctx->sfu_delay = 10;
+	} else {
+		int n = MIN2(ctx->sfu_delay, 1 + ctx->repeat + nop);
+		ctx->sfu_delay -= n;
+	}
+
 	ctx->stats->instructions += nop;
 	ctx->stats->nops += nop;
 	if (opc == OPC_NOP) {