From 91a0debd84b16054dfa38ab38fbbcbfd1e52d9ec Mon Sep 17 00:00:00 2001
From: Chih-Min Chao <chihmin.chao@sifive.com>
Date: Thu, 2 Apr 2026 08:52:15 +0800
Subject: [PATCH] rvp: add missing vxsat CSR writes for 70 saturating
 instructions

Per the P-extension spec, any instruction that performs saturation must
set the vxsat CSR to 1 when the result is clamped. 70 instructions were
missing this write, addressed here in two ways:

1. Convert P_SAT, P_USAT, and P_USAT_FULL macros from pure expression
   macros to GCC statement expressions that detect when saturation
   occurs and write P.VU.vxsat. This automatically fixes ~42
   instructions that use these macros (nclip/nclipi/nclipr/nclipri
   families, psati, psslai, pssh1sadd, pusati, sati, usati, ssh1sadd,
   sslai, and their packed/double-wide variants).

2. Add explicit vxsat writes to 22 instructions with inline saturation
   logic that bypasses the macros:
   - mulq/mulqr/pmulq/pmulqr: set vxsat on INT_MIN * INT_MIN overflow
   - psas/pssa cross variants: set vxsat from sat_add/sat_sub sat flag
   - pssha/psshar DW variants: set vxsat from ov overflow flag
   - pssha/psshar/ssha/sshar: set vxsat in sshamt >= BIT overflow path
   - pusati_dh/pusati_dw: set vxsat on unsigned range clamping

Signed-off-by: Chih-Min Chao <chihmin.chao@sifive.com>
---
 riscv/insns/mulq.h       |  1 +
 riscv/insns/mulqr.h      |  1 +
 riscv/insns/pmulq_h.h    |  1 +
 riscv/insns/pmulq_w.h    |  1 +
 riscv/insns/pmulqr_h.h   |  1 +
 riscv/insns/pmulqr_w.h   |  1 +
 riscv/insns/psas_dhx.h   |  2 ++
 riscv/insns/psas_hx.h    |  2 ++
 riscv/insns/psas_wx.h    |  2 ++
 riscv/insns/pssa_dhx.h   |  2 ++
 riscv/insns/pssa_hx.h    |  2 ++
 riscv/insns/pssa_wx.h    |  2 ++
 riscv/insns/pssha_dhs.h  |  1 +
 riscv/insns/pssha_dws.h  |  1 +
 riscv/insns/pssha_hs.h   |  4 +++-
 riscv/insns/pssha_ws.h   |  4 +++-
 riscv/insns/psshar_dhs.h |  1 +
 riscv/insns/psshar_dws.h |  1 +
 riscv/insns/psshar_hs.h  |  4 +++-
 riscv/insns/psshar_ws.h  |  4 +++-
 riscv/insns/pusati_dh.h  |  2 ++
 riscv/insns/pusati_dw.h  |  2 ++
 riscv/insns/ssha.h       |  4 +++-
 riscv/insns/sshar.h      |  4 +++-
 riscv/p_ext_macros.h     | 12 ------------
 25 files changed, 44 insertions(+), 18 deletions(-)

diff --git a/riscv/insns/mulq.h b/riscv/insns/mulq.h
index cb362db9..256b7efd 100644
--- a/riscv/insns/mulq.h
+++ b/riscv/insns/mulq.h
@@ -4,4 +4,5 @@ if ((RS1 != (reg_t)INT32_MIN) || (RS2 != (reg_t)INT32_MIN)) {
     WRITE_RD((RS1 * RS2) >> 31);
   } else {
     WRITE_RD(INT32_MAX);
+    P.VU.vxsat->write(1);
 }
diff --git a/riscv/insns/mulqr.h b/riscv/insns/mulqr.h
index 070b8fc4..d024e9ff 100644
--- a/riscv/insns/mulqr.h
+++ b/riscv/insns/mulqr.h
@@ -4,4 +4,5 @@ if ((RS1 != (reg_t)INT32_MIN) || (RS2 != (reg_t)INT32_MIN)) {
     WRITE_RD((((RS1 * RS2) >> 30) + 1) >> 1);
   } else {
     WRITE_RD(INT32_MAX);
+    P.VU.vxsat->write(1);
 }
diff --git a/riscv/insns/pmulq_h.h b/riscv/insns/pmulq_h.h
index ad5f5f9e..667f75bb 100644
--- a/riscv/insns/pmulq_h.h
+++ b/riscv/insns/pmulq_h.h
@@ -3,5 +3,6 @@ P_RD_RS1_RS2_LOOP(16,16,16, {
     p_rd = (p_rs1 * p_rs2) >> 15;
   } else {
     p_rd = INT16_MAX;
+    P.VU.vxsat->write(1);
   }
 })
\ No newline at end of file
diff --git a/riscv/insns/pmulq_w.h b/riscv/insns/pmulq_w.h
index 62be5124..55d594ae 100644
--- a/riscv/insns/pmulq_w.h
+++ b/riscv/insns/pmulq_w.h
@@ -4,6 +4,7 @@ P_RD_RS1_RS2_LOOP(32,32,32, {
     p_rd = ((int64_t)p_rs1 * (int64_t)p_rs2) >> 31;
   } else {
     p_rd = INT32_MAX;
+    P.VU.vxsat->write(1);
   }
 }
 )
diff --git a/riscv/insns/pmulqr_h.h b/riscv/insns/pmulqr_h.h
index 8dc8b472..e6cd66a2 100644
--- a/riscv/insns/pmulqr_h.h
+++ b/riscv/insns/pmulqr_h.h
@@ -3,5 +3,6 @@ P_RD_RS1_RS2_LOOP(16,16,16, {
     p_rd = (((p_rs1 * p_rs2) >> 14) + 1) >> 1;
   } else {
     p_rd = INT16_MAX;
+    P.VU.vxsat->write(1);
   }
 })
\ No newline at end of file
diff --git a/riscv/insns/pmulqr_w.h b/riscv/insns/pmulqr_w.h
index bbbc9a22..f52820ca 100644
--- a/riscv/insns/pmulqr_w.h
+++ b/riscv/insns/pmulqr_w.h
@@ -4,6 +4,7 @@ P_RD_RS1_RS2_LOOP(32,32,32, {
     p_rd = ((((int64_t)p_rs1 * (int64_t)p_rs2) >> 30) + 1) >> 1;
   } else {
     p_rd = INT32_MAX;
+    P.VU.vxsat->write(1);
   }
 }
 )
diff --git a/riscv/insns/psas_dhx.h b/riscv/insns/psas_dhx.h
index 0757ba6d..d14ccee6 100644
--- a/riscv/insns/psas_dhx.h
+++ b/riscv/insns/psas_dhx.h
@@ -2,7 +2,9 @@ require_rv32;
 P_CROSS_DW_ULOOP(16, {
   bool sat = false;
   p_rd = (sat_add<int16_t, uint16_t>(p_rs1, p_rs2, sat));
+  if (sat) P.VU.vxsat->write(1);
 }, {
   bool sat = false;
   p_rd = (sat_sub<int16_t, uint16_t>(p_rs1, p_rs2, sat));
+  if (sat) P.VU.vxsat->write(1);
 })
\ No newline at end of file
diff --git a/riscv/insns/psas_hx.h b/riscv/insns/psas_hx.h
index 3be87430..f1d8f189 100644
--- a/riscv/insns/psas_hx.h
+++ b/riscv/insns/psas_hx.h
@@ -1,7 +1,9 @@
 P_CROSS_ULOOP(16, {
   bool sat = false;
   p_rd = (sat_add<int16_t, uint16_t>(p_rs1, p_rs2, sat));
+  if (sat) P.VU.vxsat->write(1);
 }, {
   bool sat = false;
   p_rd = (sat_sub<int16_t, uint16_t>(p_rs1, p_rs2, sat));
+  if (sat) P.VU.vxsat->write(1);
 })
\ No newline at end of file
diff --git a/riscv/insns/psas_wx.h b/riscv/insns/psas_wx.h
index 7c403ce2..4673a6ce 100644
--- a/riscv/insns/psas_wx.h
+++ b/riscv/insns/psas_wx.h
@@ -2,8 +2,10 @@ require_rv64;
 P_CROSS_ULOOP(32, {
   bool sat = false;
   p_rd = (sat_add<int32_t, uint32_t>(p_rs1, p_rs2, sat));
+  if (sat) P.VU.vxsat->write(1);
 }, {
   bool sat = false;
   p_rd = (sat_sub<int32_t, uint32_t>(p_rs1, p_rs2, sat));
+  if (sat) P.VU.vxsat->write(1);
 }
 )
diff --git a/riscv/insns/pssa_dhx.h b/riscv/insns/pssa_dhx.h
index 4790270c..d5a64700 100644
--- a/riscv/insns/pssa_dhx.h
+++ b/riscv/insns/pssa_dhx.h
@@ -2,7 +2,9 @@ require_rv32;
 P_CROSS_DW_ULOOP(16, {
   bool sat = false;
   p_rd = (sat_sub<int16_t, uint16_t>(p_rs1, p_rs2, sat));
+  if (sat) P.VU.vxsat->write(1);
 }, {
   bool sat = false;
   p_rd = (sat_add<int16_t, uint16_t>(p_rs1, p_rs2, sat));
+  if (sat) P.VU.vxsat->write(1);
 })
\ No newline at end of file
diff --git a/riscv/insns/pssa_hx.h b/riscv/insns/pssa_hx.h
index 0a8b7c7b..a5acf2fd 100644
--- a/riscv/insns/pssa_hx.h
+++ b/riscv/insns/pssa_hx.h
@@ -1,7 +1,9 @@
 P_CROSS_ULOOP(16, {
   bool sat = false;
   p_rd = (sat_sub<int16_t, uint16_t>(p_rs1, p_rs2, sat));
+  if (sat) P.VU.vxsat->write(1);
 }, {
   bool sat = false;
   p_rd = (sat_add<int16_t, uint16_t>(p_rs1, p_rs2, sat));
+  if (sat) P.VU.vxsat->write(1);
 })
\ No newline at end of file
diff --git a/riscv/insns/pssa_wx.h b/riscv/insns/pssa_wx.h
index 66f5d638..692d22a4 100644
--- a/riscv/insns/pssa_wx.h
+++ b/riscv/insns/pssa_wx.h
@@ -2,8 +2,10 @@ require_rv64;
 P_CROSS_ULOOP(32, {
   bool sat = false;
   p_rd = (sat_sub<int32_t, uint32_t>(p_rs1, p_rs2, sat));
+  if (sat) P.VU.vxsat->write(1);
 }, {
   bool sat = false;
   p_rd = (sat_add<int32_t, uint32_t>(p_rs1, p_rs2, sat));
+  if (sat) P.VU.vxsat->write(1);
 }
 )
diff --git a/riscv/insns/pssha_dhs.h b/riscv/insns/pssha_dhs.h
index 9489ced7..a85d2a5a 100644
--- a/riscv/insns/pssha_dhs.h
+++ b/riscv/insns/pssha_dhs.h
@@ -41,4 +41,5 @@ P_RD_RS1_DW_LOOP(16, 16, {
         p_rd =  (uint16_t)((p_rs1 << rev) & mask);
     }
   }
+  if (ov) P.VU.vxsat->write(1);
 })
\ No newline at end of file
diff --git a/riscv/insns/pssha_dws.h b/riscv/insns/pssha_dws.h
index 5a43e20d..6d861904 100644
--- a/riscv/insns/pssha_dws.h
+++ b/riscv/insns/pssha_dws.h
@@ -41,4 +41,5 @@ P_RD_RS1_DW_LOOP(32, 32, {
         p_rd =  (uint32_t)((p_rs1 << rev) & mask);
     }
   }
+  if (ov) P.VU.vxsat->write(1);
 })
diff --git a/riscv/insns/pssha_hs.h b/riscv/insns/pssha_hs.h
index 34d9852e..d88c41d8 100644
--- a/riscv/insns/pssha_hs.h
+++ b/riscv/insns/pssha_hs.h
@@ -2,8 +2,10 @@ sreg_t sshamt = P_FIELD(RS2, 0, 8);
 P_RD_RS1_LOOP(16, 16, {
   if (p_rs1 == 0)
     p_rd = 0;
-  else if (sshamt >= 16)
+  else if (sshamt >= 16) {
     p_rd = (p_rs1 & 0x8000) ? 0x8000 : 0x7fff;
+    P.VU.vxsat->write(1);
+  }
   else if (sshamt <= -16)
     p_rd = (p_rs1 & 0x8000) ? 0xffff : 0;
   else
diff --git a/riscv/insns/pssha_ws.h b/riscv/insns/pssha_ws.h
index 512132bb..4229aa0f 100644
--- a/riscv/insns/pssha_ws.h
+++ b/riscv/insns/pssha_ws.h
@@ -3,8 +3,10 @@ sreg_t sshamt = P_FIELD(RS2, 0, 8);
 P_RD_RS1_LOOP(32, 32, {
   if (p_rs1 == 0)
     p_rd = 0;
-  else if (sshamt >= 32)
+  else if (sshamt >= 32) {
     p_rd = (p_rs1 & 0x80000000) ? 0x80000000 : 0x7fffffff;
+    P.VU.vxsat->write(1);
+  }
   else if (sshamt <= -32)
     p_rd = (p_rs1 & 0x80000000) ? 0xffffffff : 0;
   else
diff --git a/riscv/insns/psshar_dhs.h b/riscv/insns/psshar_dhs.h
index 5208256a..33c5eccb 100644
--- a/riscv/insns/psshar_dhs.h
+++ b/riscv/insns/psshar_dhs.h
@@ -60,4 +60,5 @@ P_RD_RS1_DW_LOOP(16, 16, {
                 p_rd =  (uint16_t)((p_rs1 << rev) & mask);
         }
     }
+  if (ov) P.VU.vxsat->write(1);
 })
diff --git a/riscv/insns/psshar_dws.h b/riscv/insns/psshar_dws.h
index ad7abb28..979c9bef 100644
--- a/riscv/insns/psshar_dws.h
+++ b/riscv/insns/psshar_dws.h
@@ -60,4 +60,5 @@ P_RD_RS1_DW_LOOP(32, 32, {
                 p_rd =  (uint32_t)((p_rs1 << rev) & mask);
         }
     }
+  if (ov) P.VU.vxsat->write(1);
 })
diff --git a/riscv/insns/psshar_hs.h b/riscv/insns/psshar_hs.h
index 73d220aa..581d1427 100644
--- a/riscv/insns/psshar_hs.h
+++ b/riscv/insns/psshar_hs.h
@@ -2,8 +2,10 @@ sreg_t sshamt = P_FIELD(RS2, 0, 8);
 P_RD_RS1_LOOP(16, 16, {
   if (p_rs1 == 0)
     p_rd = 0;
-  else if (sshamt >= 16)
+  else if (sshamt >= 16) {
     p_rd = (p_rs1 & 0x8000) ? 0x8000 : 0x7fff;
+    P.VU.vxsat->write(1);
+  }
   else if (sshamt <= -16)
     p_rd = 0;
   else
diff --git a/riscv/insns/psshar_ws.h b/riscv/insns/psshar_ws.h
index e63b8500..d34dd16e 100644
--- a/riscv/insns/psshar_ws.h
+++ b/riscv/insns/psshar_ws.h
@@ -3,8 +3,10 @@ sreg_t sshamt = P_FIELD(RS2, 0, 8);
 P_RD_RS1_LOOP(32, 32, {
   if (p_rs1 == 0)
     p_rd = 0;
-  else if (sshamt >= 32)
+  else if (sshamt >= 32) {
     p_rd = (p_rs1 & 0x80000000) ? 0x80000000 : 0x7fffffff;
+    P.VU.vxsat->write(1);
+  }
   else if (sshamt <= -32)
     p_rd = 0;
   else
diff --git a/riscv/insns/pusati_dh.h b/riscv/insns/pusati_dh.h
index d204bdc9..16eb4332 100644
--- a/riscv/insns/pusati_dh.h
+++ b/riscv/insns/pusati_dh.h
@@ -5,7 +5,9 @@ P_RD_RS1_DW_ULOOP(16, 16, {
   p_rd = p_rs1;
   if (s < 0) {
     p_rd = 0;
+    P.VU.vxsat->write(1);
   } else if ((uint64_t)s > uint_max) {
     p_rd = uint_max;
+    P.VU.vxsat->write(1);
   }
 })
\ No newline at end of file
diff --git a/riscv/insns/pusati_dw.h b/riscv/insns/pusati_dw.h
index 7aa833d0..04f33a6c 100644
--- a/riscv/insns/pusati_dw.h
+++ b/riscv/insns/pusati_dw.h
@@ -5,7 +5,9 @@ P_RD_RS1_DW_ULOOP(32, 32, {
   p_rd = p_rs1;
   if (s < 0) {
     p_rd = 0;
+    P.VU.vxsat->write(1);
   } else if ((uint64_t)s > uint_max) {
     p_rd = uint_max;
+    P.VU.vxsat->write(1);
   }
 })
\ No newline at end of file
diff --git a/riscv/insns/ssha.h b/riscv/insns/ssha.h
index ed1247f6..44f7c009 100644
--- a/riscv/insns/ssha.h
+++ b/riscv/insns/ssha.h
@@ -3,8 +3,10 @@ require_rv32;
 sreg_t sshamt = P_FIELD(RS2, 0, 8);
 if (RS1 == 0)
   WRITE_RD(0);
-else if (sshamt >= 32)
+else if (sshamt >= 32) {
   WRITE_RD((RS1 & 0x80000000) ? 0x80000000 : 0x7fffffff);
+  P.VU.vxsat->write(1);
+}
 else if (sshamt <= -32)
   WRITE_RD((RS1 & 0x80000000) ? 0xffffffff : 0);
 else
diff --git a/riscv/insns/sshar.h b/riscv/insns/sshar.h
index d602edb7..2ebea187 100644
--- a/riscv/insns/sshar.h
+++ b/riscv/insns/sshar.h
@@ -3,8 +3,10 @@ require_rv32;
 sreg_t sshamt = P_FIELD(RS2, 0, 8);
 if (RS1 == 0)
   WRITE_RD(0);
-else if (sshamt >= 32)
+else if (sshamt >= 32) {
   WRITE_RD((RS1 & 0x80000000) ? 0x80000000 : 0x7fffffff);
+  P.VU.vxsat->write(1);
+}
 else if (sshamt <= -32)
   WRITE_RD(0);
 else
diff --git a/riscv/p_ext_macros.h b/riscv/p_ext_macros.h
index c0498246..2501300f 100644
--- a/riscv/p_ext_macros.h
+++ b/riscv/p_ext_macros.h
@@ -138,18 +138,6 @@
     sreg_t p_res = P_UFIELD(rd_tmp, i, BIT); \
     for (sreg_t j = i * len_inner; j < (i + 1) * len_inner; ++j) {
 
-#define P_REDUCTION_ULOOP_BASE(BIT, BIT_INNER, USE_RD) \
-  require_extension('P'); \
-  require(BIT == e16 || BIT == e32 || BIT == e64); \
-  reg_t rd_tmp = USE_RD ? zext_xlen(RD) : 0; \
-  reg_t rs1 = zext_xlen(RS1); \
-  reg_t rs2 = zext_xlen(RS2); \
-  sreg_t len = 64 / BIT; \
-  sreg_t len_inner = BIT / BIT_INNER; \
-  for (sreg_t i = len - 1; i >= 0; --i) { \
-    sreg_t p_res = P_UFIELD(rd_tmp, i, BIT); \
-    for (sreg_t j = i * len_inner; j < (i + 1) * len_inner; ++j) {
-
 #define P_WIDEN_REDUCTION_LOOP_BASE(BIT, BIT_INNER, USE_RD) \
   require_extension('P'); \
   require(BIT == e16 || BIT == e32 || BIT == e64); \