From 35b6961dcb907407dcf28820eeed71038b1c2fa8 Mon Sep 17 00:00:00 2001
From: Giacomo Travaglini <giacomo.travaglini@arm.com>
Date: Wed, 28 Apr 2021 12:45:52 +0100
Subject: [PATCH] arch-arm: Fix SMM* instructions

A recent clean up patch [1] introduced some bugs in the instruction
implementation of:

SMMLA, SMMLAR, SMMLS, SMMLSR, SMMUL, SMMULR

[1]: https://gem5-review.googlesource.com/c/public/gem5/+/42387

Change-Id: I459fe99bd2711e00027e9ef0c7796af7a374a509
Signed-off-by: Giacomo Travaglini <giacomo.travaglini@arm.com>
Reviewed-on: https://gem5-review.googlesource.com/c/public/gem5/+/44945
Reviewed-by: Gabe Black <gabe.black@gmail.com>
Maintainer: Gabe Black <gabe.black@gmail.com>
Tested-by: kokoro <noreply+kokoro@google.com>
---
 src/arch/arm/isa/insts/mult.isa | 34 ++++++++++++++++++---------------
 1 file changed, 19 insertions(+), 15 deletions(-)

diff --git a/src/arch/arm/isa/insts/mult.isa b/src/arch/arm/isa/insts/mult.isa
index 528aa669ed..7f37099d60 100644
--- a/src/arch/arm/isa/insts/mult.isa
+++ b/src/arch/arm/isa/insts/mult.isa
@@ -158,43 +158,43 @@ let {{
         ''', "overflow")
     buildMult4Inst("smlal", '''
             resTemp = PInt2.sw * PInt3.sw +
-                      (int64_t)((PInt1.uw << 32) | PInt0.uw);
+                      (int64_t)(((uint64_t)PInt1.uw << 32) | PInt0.uw);
             PInt0 = (uint32_t)resTemp;
             PInt1 = (uint32_t)(resTemp >> 32);
         ''', "llbit")
     buildMult4InstUnCc("smlalbb", '''
             resTemp = PInt2.sh0 * PInt3.sh0 +
-                      (int64_t)((PInt1.uw << 32) | PInt0.uw);
+                      (int64_t)(((uint64_t)PInt1.uw << 32) | PInt0.uw);
             PInt0 = (uint32_t)resTemp;
             PInt1 = (uint32_t)(resTemp >> 32);
         ''')
     buildMult4InstUnCc("smlalbt", '''
             resTemp = PInt2.sh0 * PInt3.sh1 +
-                      (int64_t)((PInt1.uw << 32) | PInt0.uw);
+                      (int64_t)(((uint64_t)PInt1.uw << 32) | PInt0.uw);
             PInt0 = (uint32_t)resTemp;
             PInt1 = (uint32_t)(resTemp >> 32);
         ''')
     buildMult4InstUnCc("smlaltb", '''
             resTemp = PInt2.sh1 * PInt3.sh0 +
-                      (int64_t)((PInt1.uw << 32) | PInt0.uw);
+                      (int64_t)(((uint64_t)PInt1.uw << 32) | PInt0.uw);
             PInt0 = (uint32_t)resTemp;
             PInt1 = (uint32_t)(resTemp >> 32);
         ''')
     buildMult4InstUnCc("smlaltt", '''
             resTemp = PInt2.sh1 * PInt3.sh1 +
-                      (int64_t)((PInt1.uw << 32) | PInt0.uw);
+                      (int64_t)(((uint64_t)PInt1.uw << 32) | PInt0.uw);
             PInt0 = (uint32_t)resTemp;
             PInt1 = (uint32_t)(resTemp >> 32);
         ''')
     buildMult4InstUnCc("smlald", '''
             resTemp = PInt2.sh1 * PInt3.sh1 + PInt2.sh0 * PInt3.sh0 +
-                      (int64_t)((PInt1.uw << 32) | PInt0.uw);
+                      (int64_t)(((uint64_t)PInt1.uw << 32) | PInt0.uw);
             PInt0 = (uint32_t)resTemp;
             PInt1 = (uint32_t)(resTemp >> 32);
         ''')
     buildMult4InstUnCc("smlaldx", '''
             resTemp = PInt2.sh1 * PInt3.sh0 + PInt2.sh0 * PInt3.sh1 +
-                      (int64_t)((PInt1.uw << 32) | PInt0.uw);
+                      (int64_t)(((uint64_t)PInt1.uw << 32) | PInt0.uw);
             PInt0 = (uint32_t)resTemp;
             PInt1 = (uint32_t)(resTemp >> 32);
         ''')
@@ -220,33 +220,37 @@ let {{
         ''', "overflow")
     buildMult4InstUnCc("smlsld", '''
             resTemp = PInt2.sh0 * PInt3.sh0 - PInt2.sh1 * PInt3.sh1 +
-                      (int64_t)((PInt1.uw << 32) | PInt0.uw);
+                      (int64_t)(((uint64_t)PInt1.uw << 32) | PInt0.uw);
             PInt0 = (uint32_t)resTemp;
             PInt1 = (uint32_t)(resTemp >> 32);
         ''')
     buildMult4InstUnCc("smlsldx", '''
             resTemp = PInt2.sh0 * PInt3.sh1 - PInt2.sh1 * PInt3.sh0 +
-                      (int64_t)((PInt1.uw << 32) | PInt0.uw);
+                      (int64_t)(((uint64_t)PInt1.uw << 32) | PInt0.uw);
             PInt0 = (uint32_t)resTemp;
             PInt1 = (uint32_t)(resTemp >> 32);
         ''')
     buildMult4InstUnCc("smmla", '''
-            PInt0 = PInt3.sw + ((PInt1.sw * PInt2.sw) >> 32);
+            PInt0 = (((int64_t)PInt3.sw << 32) + (PInt1.sw * PInt2.sw)) >> 32;
         ''')
     buildMult4InstUnCc("smmlar", '''
-            PInt0 = PInt3.sw + ((PInt1.sw * PInt2.sw + (0x1ULL << 31)) >> 32);
+            PInt0 = (((int64_t)PInt3.sw << 32) +
+                     (PInt1.sw * PInt2.sw) +
+                     (0x1ULL << 31)) >> 32;
         ''')
     buildMult4InstUnCc("smmls", '''
-            PInt0 = PInt3.sw - ((PInt1.sw * PInt2.sw) >> 32);
+            PInt0 = (((int64_t)PInt3.sw << 32) - (PInt1.sw * PInt2.sw)) >> 32;
         ''')
     buildMult4InstUnCc("smmlsr", '''
-            PInt0 = PInt3.sw - ((PInt1.sw * PInt2.sw + (0x1ULL << 31)) >> 32);
+            PInt0 = (((int64_t)PInt3.sw << 32) -
+                     (PInt1.sw * PInt2.sw) +
+                     (0x1ULL << 31)) >> 32;
         ''')
     buildMult3InstUnCc("smmul", '''
-            PInt0 = (PInt1.sw * PInt2.sw) >> 32;
+            PInt0 = ((int64_t)PInt1.sw * PInt2.sw) >> 32;
         ''')
     buildMult3InstUnCc("smmulr", '''
-            PInt0 = (PInt1.sw * PInt2.sw + (0x1ULL << 31)) >> 32;
+            PInt0 = (((int64_t)PInt1.sw * PInt2.sw) + (0x1ULL << 31)) >> 32;
         ''')
     buildMult3InstCc("smuad", '''
             PInt0 = resTemp = PInt1.sh0 * PInt2.sh0 + PInt1.sh1 * PInt2.sh1;