PPC970: 6954845d8 breaks build

[pkubaj]

FreeBSD 15.0-RELEASE, powerpc64, PPC970

TEST 41/123 zscal:i00_NAN_incx_2 [FAIL]
  ERR: test_zscal.c:639  should be true
TEST 42/123 zscal:i00_NAN [FAIL]
  ERR: test_zscal.c:625  should be true
TEST 53/123 cscal:i00_NAN_incx_2 [FAIL]
  ERR: test_zscal.c:466  should be true
TEST 54/123 cscal:i00_NAN [FAIL]
  ERR: test_zscal.c:452  should be true
RESULTS: 123 tests (119 ok, 4 failed, 0 skipped) ran in 172 ms

[martin-frbg]

Not sure what happened here - either I tested in the wrong environment, or I managed to commit with a different stack address value than I tested. Hopefully fixed now, but I have to rely on building for PPC970 on more recent POWER hardware

[pkubaj]

Sorry, but that breaks the tests even more. Here's the latest develop branch:

TEST 1/125 max:smax_zero [OK]
TEST 2/125 max:dmax_positive [OK]
TEST 3/125 max:smax_negative [OK]
TEST 4/125 min:smin_zero [OK]
TEST 5/125 min:dmin_positive [OK]
TEST 6/125 min:smin_negative [OK]
TEST 7/125 amax:dzamax [OK]
TEST 8/125 amax:scamax [OK]
TEST 9/125 amax:damax [OK]
TEST 10/125 amax:samax [OK]
TEST 11/125 ismax:negative_step_2 [OK]
TEST 12/125 ismax:positive_step_2 [OK]
TEST 13/125 ismin:negative_step_2 [OK]
TEST 14/125 ismin:positive_step_2 [OK]
TEST 15/125 drotmg:drotmg_D1_big_D2_big_flag_zero [OK]
TEST 16/125 drotmg:rotmg_D1eqD2_X1eqX2 [OK]
TEST 17/125 drotmg:rotmg_issue1452 [OK]
TEST 18/125 drotmg:rotmg [OK]
TEST 19/125 axpy:caxpy_incx_0 [OK]
TEST 20/125 axpy:caxpy_inc_0 [OK]
TEST 21/125 axpy:saxpy_inc_0 [OK]
TEST 22/125 axpy:zaxpy_incx_0 [OK]
TEST 23/125 axpy:zaxpy_inc_0 [OK]
TEST 24/125 axpy:daxpy_inc_0 [OK]
TEST 25/125 zdotu:zdotu_offset_1 [OK]
TEST 26/125 zdotu:zdotu_n_1 [OK]
TEST 27/125 dsdot:dsdot_n_1 [OK]
TEST 28/125 swap:cswap_inc_0 [OK]
TEST 29/125 swap:sswap_inc_0 [OK]
TEST 30/125 swap:zswap_inc_0 [OK]
TEST 31/125 swap:dswap_inc_0 [OK]
TEST 32/125 rot:csrot_inc_0 [OK]
TEST 33/125 rot:srotm_inc_1 [OK]
TEST 34/125 rot:srot_inc_1 [OK]
TEST 35/125 rot:srot_inc_0 [OK]
TEST 36/125 rot:zdrot_inc_0 [OK]
TEST 37/125 rot:drotm_inc_1 [OK]
TEST 38/125 rot:drot_inc_1 [OK]
TEST 39/125 rot:drot_inc_0 [OK]
TEST 40/125 dnrm2:dnrm2_neg_incx [OK]
TEST 41/125 dnrm2:dnrm2_tiny [OK]
TEST 42/125 dnrm2:dnrm2_inf [OK]
TEST 43/125 zscal:i00_NAN_incx_2 [FAIL]
  ERR: test_zscal.c:639  should be true
TEST 44/125 zscal:i00_NAN [FAIL]
  ERR: test_zscal.c:625  should be true
TEST 45/125 zscal:i_0inf_inc_2 [OK]
TEST 46/125 zscal:i_0inf [OK]
TEST 47/125 zscal:inf_i_inc_2 [OK]
TEST 48/125 zscal:inf_i [OK]
TEST 49/125 zscal:i_inf_inc_2 [OK]
TEST 50/125 zscal:i_inf [OK]
TEST 51/125 zscal:nan_i_inc_2 [OK]
TEST 52/125 zscal:nan_i [OK]
TEST 53/125 zscal:i_nan_inc_2 [OK]
TEST 54/125 zscal:i_nan [OK]
TEST 55/125 cscal:i00_NAN_incx_2 [FAIL]
  ERR: test_zscal.c:466  should be true
TEST 56/125 cscal:i00_NAN [FAIL]
  ERR: test_zscal.c:452  should be true
TEST 57/125 cscal:i_0inf_inc_2 [OK]
TEST 58/125 cscal:i_0inf [OK]
TEST 59/125 cscal:inf_i_inc_2 [OK]
TEST 60/125 cscal:inf_i [OK]
TEST 61/125 cscal:i_inf_inc_2 [OK]
TEST 62/125 cscal:i_inf [OK]
TEST 63/125 cscal:nan_i_inc_2 [OK]
TEST 64/125 cscal:nan_i [OK]
TEST 65/125 cscal:i_nan_inc_2 [OK]
TEST 66/125 cscal:i_nan [OK]
TEST 67/125 dscal:inf_nan_inc_2 [OK]
TEST 68/125 dscal:inf_nan [OK]
TEST 69/125 dscal:nan_inf_inc_2 [OK]
TEST 70/125 dscal:nan_inf [OK]
TEST 71/125 dscal:inf_0_inc_2 [OK]
TEST 72/125 dscal:inf_0 [OK]
TEST 73/125 dscal:0_inf_inc_2 [FAIL]
  ERR: test_zscal.c:226  should be true
TEST 74/125 dscal:0_inf [FAIL]
  ERR: test_zscal.c:214  should be true
TEST 75/125 dscal:nan_0_inc_2 [OK]
TEST 76/125 dscal:nan_0 [OK]
TEST 77/125 dscal:0_nan_inc_2 [FAIL]
  ERR: test_zscal.c:180  should be true
TEST 78/125 dscal:0_nan [FAIL]
  ERR: test_zscal.c:168  should be true
TEST 79/125 sscal:inf_nan_inc_2 [OK]
TEST 80/125 sscal:inf_nan [OK]
TEST 81/125 sscal:nan_inf_inc_2 [OK]
TEST 82/125 sscal:nan_inf [OK]
TEST 83/125 sscal:inf_0_inc_2 [OK]
TEST 84/125 sscal:inf_0 [OK]
TEST 85/125 sscal:0_inf_inc_2 [FAIL]
  ERR: test_zscal.c:77  should be true
TEST 86/125 sscal:0_inf [FAIL]
  ERR: test_zscal.c:65  should be true
TEST 87/125 sscal:nan_0_inc_2 [OK]
TEST 88/125 sscal:nan_0 [OK]
TEST 89/125 sscal:0_nan_inc_2 [FAIL]
  ERR: test_zscal.c:31  should be true
TEST 90/125 sscal:0_nan [FAIL]
  ERR: test_zscal.c:19  should be true
TEST 91/125 amin:dzamin [OK]
TEST 92/125 amin:scamin [OK]
TEST 93/125 amin:damin [OK]
TEST 94/125 amin:samin [OK]
TEST 95/125 axpby:zaxpby_inc_2 [OK]
TEST 96/125 axpby:zaxpby_inc_1 [OK]
TEST 97/125 axpby:zaxpby_inc_0 [OK]
TEST 98/125 axpby:caxpby_inc_2 [OK]
TEST 99/125 axpby:caxpby_inc_1 [OK]
TEST 100/125 axpby:caxpby_inc_0 [OK]
TEST 101/125 axpby:daxpby_inc_2 [OK]
TEST 102/125 axpby:daxpby_inc_1 [OK]
TEST 103/125 axpby:daxpby_inc_0 [OK]
TEST 104/125 axpby:saxpby_inc_2 [OK]
TEST 105/125 axpby:saxpby_inc_1 [OK]
TEST 106/125 axpby:saxpby_inc_0 [OK]
TEST 107/125 zgemv:2_0_nan_1_inf_1_incy_2 [OK]
TEST 108/125 zgemv:2_0_nan_1_inf_1 [OK]
TEST 109/125 zgemv:0_2_nan_1_inf_1_incy_2 [OK]
TEST 110/125 zgemv:0_2_nan_1_inf_1 [OK]
TEST 111/125 zgemv:0_nan_inf_incy_2 [OK]
TEST 112/125 zgemv:0_nan_inf [OK]
TEST 113/125 cgemv:2_0_nan_1_inf_1_incy_2 [OK]
TEST 114/125 cgemv:2_0_nan_1_inf_1 [OK]
TEST 115/125 cgemv:0_2_nan_1_inf_1_incy_2 [OK]
TEST 116/125 cgemv:0_2_nan_1_inf_1 [OK]
TEST 117/125 cgemv:0_nan_inf_incy_2 [OK]
TEST 118/125 cgemv:0_nan_inf [OK]
TEST 119/125 dgemv:0_nan_inf_incy_2 [OK]
TEST 120/125 dgemv:0_nan_inf [OK]
TEST 121/125 sgemv:0_nan_inf_incy_2 [OK]
TEST 122/125 sgemv:0_nan_inf [OK]
TEST 123/125 potrf:smoketest_trivial [OK]
TEST 124/125 potrf:bug_695 [OK]
TEST 125/125 kernel_regress:skx_avx [OK]
RESULTS: 125 tests (113 ok, 12 failed, 0 skipped) ran in 249 ms

[martin-frbg]

Probably an OS or endianness issue then - is your FreeBSD powerpc64 or powerpc64le ? (I built and tested on a big-endian POWER7 running Debian, the OSUOSL-hosted "cfarm110" in the GCC Compile Farm)

[pkubaj]

It's powerpc64, big-endian. It doesn't make sense to build PPC970 on little-endian, I'm not even sure whether that works. Build works fine on powerpc64le using POWER8 kernel.

[martin-frbg]

Thought as much, will probably have to try with qemu if TARGET=PPC970 on the aforementioned POWER7 under Debian/ppc64 is not sufficient. The short-term alternative would be to switch all ?SCALKERNEL entries in kernel/power/KERNEL to the respective generic C ones (../arm/scal.c and ../arm/zscal.c)

[pkubaj]

That's strange, I'm just using my POWER9 workstation. I just set up TARGET=PPC970 and get that crash.

[martin-frbg]

Hmm, weird. I'll be off to FOSDEM in a few minutes though, will get back to this monday at the latest.

[pkubaj]

I hope you don't take it as a slop, but Claude actually helped here:
Root Cause: The PowerPC assembly code in kernel/power/zscal.S was loading the FLAG parameter from an incorrect stack offset (104(SP)) for the FreeBSD PowerPC64 ELFv2 ABI. This caused a zero-multiplication optimization to run
unconditionally, violating IEEE 754 by returning 0 instead of NaN for 0 * NaN operations.

Fix: Changed the stack offset from 104(SP) to 112(SP) to correctly load the FLAG parameter, allowing the optimization to be properly controlled:

• When FLAG=0 (internal BLAS calls like gemv): Uses fast zero-store optimization per BLAS spec
• When FLAG=1 (external user calls): Uses proper multiplication to preserve IEEE 754 NaN handling

The patch is:

--- kernel/power/zscal.S.orig   2025-06-19 11:45:39.000000000 +0200
+++ kernel/power/zscal.S        2026-04-30 13:24:06.878012000 +0200
@@ -96,7 +96,7 @@
        fcmpu   cr0, FZERO, ALPHA_I
        bne-    cr0, LL(A1I1)
 
-       LDLONG  FLAG, 104(SP)
+       LDLONG  FLAG, 112(SP)
        cmpwi   cr0, FLAG, 1
        beq-    cr0, LL(A1I1)

[pkubaj]

Oh, I can see you already changed it in the newest version. In that case, can you revert #5654 ?

[martin-frbg]

Ah great - I must admit I am an AI sceptic, but I guess this case counts as an intelligent pattern matching, and practical use of linear algebra can't all be a bad thing...
Back in January I did not have access to any ppc running FreeBSD, so I would have had to entertain you with various random guesses for the stack offset to see which caused the least damage - going with the lame C kernels looked like the saner approach for this outdated hardware. Thankfully, a FreeBSD16 VM on POWER9 has since become available in the GCC Compile Farm. I'll rebuild and test there later