nv50/ir: fix s32 x s32 -> high s32 multiply logic

Retrieving the high 32 bits of a signed multiply is rather annoying. It
appears that the simplest way to do this is to compute the absolute
value of the arguments, and perform a u32 x u32 -> u64 operation. If the
arguments' signs differ, then negate the result. Since there is no u64
support in the cvt instruction, we have the perform the 2's complement
negation "by hand".

This logic can come into use by the IMUL_HI instruction (very unlikely
to be seen), as well as from constant folding of division by a constant.
Fixes dolphin's divisions by 255.

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Cc: "10.1 10.2" <mesa-stable@lists.freedesktop.org>
Reviewed-by: Ben Skeggs <bskeggs@redhat.com>
(cherry picked from commit d3a5cf052c38087b395871b5b46776e2a7d4a7d7)

2 files changed, 82 insertions, 11 deletions

diff --git a/src/gallium/drivers/nouveau/codegen/nv50_ir_lowering_nv50.cpp b/src/gallium/drivers/nouveau/codegen/nv50_ir_lowering_nv50.cpp
index b17d57d0bfd..0fb76663ffe 100644
--- a/src/gallium/drivers/nouveau/codegen/nv50_ir_lowering_nv50.cpp
+++ b/src/gallium/drivers/nouveau/codegen/nv50_ir_lowering_nv50.cpp
@@ -37,18 +37,25 @@ namespace nv50_ir {
 //    ah*bl 00
 //
 // fffe0001 + fffe0001
+//
+// Note that this sort of splitting doesn't work for signed values, so we
+// compute the sign on those manually and then perform an unsigned multiply.
 static bool
 expandIntegerMUL(BuildUtil *bld, Instruction *mul)
 {
    const bool highResult = mul->subOp == NV50_IR_SUBOP_MUL_HIGH;
 
-   DataType fTy = mul->sType; // full type
-   DataType hTy;
+   DataType fTy; // full type
+   switch (mul->sType) {
+   case TYPE_S32: fTy = TYPE_U32; break;
+   case TYPE_S64: fTy = TYPE_U64; break;
+   default: fTy = mul->sType; break;
+   }
+
+   DataType hTy; // half type
    switch (fTy) {
-   case TYPE_S32: hTy = TYPE_S16; break;
    case TYPE_U32: hTy = TYPE_U16; break;
    case TYPE_U64: hTy = TYPE_U32; break;
-   case TYPE_S64: hTy = TYPE_S32; break;
    default:
       return false;
    }
@@ -59,15 +66,25 @@ expandIntegerMUL(BuildUtil *bld, Instruction *mul)
 
    bld->setPosition(mul, true);
 
+   Value *s[2];
    Value *a[2], *b[2];
-   Value *c[2];
    Value *t[4];
    for (int j = 0; j < 4; ++j)
       t[j] = bld->getSSA(fullSize);
 
+   s[0] = mul->getSrc(0);
+   s[1] = mul->getSrc(1);
+
+   if (isSignedType(mul->sType)) {
+      s[0] = bld->getSSA(fullSize);
+      s[1] = bld->getSSA(fullSize);
+      bld->mkOp1(OP_ABS, mul->sType, s[0], mul->getSrc(0));
+      bld->mkOp1(OP_ABS, mul->sType, s[1], mul->getSrc(1));
+   }
+
    // split sources into halves
-   i[0] = bld->mkSplit(a, halfSize, mul->getSrc(0));
-   i[1] = bld->mkSplit(b, halfSize, mul->getSrc(1));
+   i[0] = bld->mkSplit(a, halfSize, s[0]);
+   i[1] = bld->mkSplit(b, halfSize, s[1]);
 
    i[2] = bld->mkOp2(OP_MUL, fTy, t[0], a[0], b[1]);
    i[3] = bld->mkOp3(OP_MAD, fTy, t[1], a[1], b[0], t[0]);
@@ -75,24 +92,76 @@ expandIntegerMUL(BuildUtil *bld, Instruction *mul)
    i[4] = bld->mkOp3(OP_MAD, fTy, t[3], a[0], b[0], t[2]);
 
    if (highResult) {
-      Value *r[4];
+      Value *c[2];
+      Value *r[5];
       Value *imm = bld->loadImm(NULL, 1 << (halfSize * 8));
       c[0] = bld->getSSA(1, FILE_FLAGS);
       c[1] = bld->getSSA(1, FILE_FLAGS);
-      for (int j = 0; j < 4; ++j)
+      for (int j = 0; j < 5; ++j)
          r[j] = bld->getSSA(fullSize);
 
       i[8] = bld->mkOp2(OP_SHR, fTy, r[0], t[1], bld->mkImm(halfSize * 8));
       i[6] = bld->mkOp2(OP_ADD, fTy, r[1], r[0], imm);
       bld->mkMov(r[3], r[0])->setPredicate(CC_NC, c[0]);
       bld->mkOp2(OP_UNION, TYPE_U32, r[2], r[1], r[3]);
-      i[5] = bld->mkOp3(OP_MAD, fTy, mul->getDef(0), a[1], b[1], r[2]);
+      i[5] = bld->mkOp3(OP_MAD, fTy, r[4], a[1], b[1], r[2]);
 
       // set carry defs / sources
       i[3]->setFlagsDef(1, c[0]);
-      i[4]->setFlagsDef(0, c[1]); // actual result not required, just the carry
+      // actual result required in negative case, but ignored for
+      // unsigned. for some reason the compiler ends up dropping the whole
+      // instruction if the destination is unused but the flags are.
+      if (isSignedType(mul->sType))
+         i[4]->setFlagsDef(1, c[1]);
+      else
+         i[4]->setFlagsDef(0, c[1]);
       i[6]->setPredicate(CC_C, c[0]);
       i[5]->setFlagsSrc(3, c[1]);
+
+      if (isSignedType(mul->sType)) {
+         Value *cc[2];
+         Value *rr[7];
+         Value *one = bld->getSSA(fullSize);
+         bld->loadImm(one, 1);
+         for (int j = 0; j < 7; j++)
+            rr[j] = bld->getSSA(fullSize);
+
+         // NOTE: this logic uses predicates because splitting basic blocks is
+         // ~impossible during the SSA phase. The RA relies on a correlation
+         // between edge order and phi node sources.
+
+         // Set the sign of the result based on the inputs
+         bld->mkOp2(OP_XOR, fTy, NULL, mul->getSrc(0), mul->getSrc(1))
+            ->setFlagsDef(0, (cc[0] = bld->getSSA(1, FILE_FLAGS)));
+
+         // 1s complement of 64-bit value
+         bld->mkOp1(OP_NOT, fTy, rr[0], r[4])
+            ->setPredicate(CC_S, cc[0]);
+         bld->mkOp1(OP_NOT, fTy, rr[1], t[3])
+            ->setPredicate(CC_S, cc[0]);
+
+         // add to low 32-bits, keep track of the carry
+         Instruction *n = bld->mkOp2(OP_ADD, fTy, NULL, rr[1], one);
+         n->setPredicate(CC_S, cc[0]);
+         n->setFlagsDef(0, (cc[1] = bld->getSSA(1, FILE_FLAGS)));
+
+         // If there was a carry, add 1 to the upper 32 bits
+         // XXX: These get executed even if they shouldn't be
+         bld->mkOp2(OP_ADD, fTy, rr[2], rr[0], one)
+            ->setPredicate(CC_C, cc[1]);
+         bld->mkMov(rr[3], rr[0])
+            ->setPredicate(CC_NC, cc[1]);
+         bld->mkOp2(OP_UNION, fTy, rr[4], rr[2], rr[3]);
+
+         // Merge the results from the negative and non-negative paths
+         bld->mkMov(rr[5], rr[4])
+            ->setPredicate(CC_S, cc[0]);
+         bld->mkMov(rr[6], r[4])
+            ->setPredicate(CC_NS, cc[0]);
+         bld->mkOp2(OP_UNION, mul->sType, mul->getDef(0), rr[5], rr[6]);
+      } else {
+         bld->mkMov(mul->getDef(0), r[4]);
+      }
    } else {
       bld->mkMov(mul->getDef(0), t[3]);
    }
diff --git a/src/gallium/drivers/nouveau/codegen/nv50_ir_target_nv50.cpp b/src/gallium/drivers/nouveau/codegen/nv50_ir_target_nv50.cpp
index 0b2f27ad497..c844fa4d3c1 100644
--- a/src/gallium/drivers/nouveau/codegen/nv50_ir_target_nv50.cpp
+++ b/src/gallium/drivers/nouveau/codegen/nv50_ir_target_nv50.cpp
@@ -331,6 +331,8 @@ TargetNV50::insnCanLoad(const Instruction *i, int s,
          return false;
       if (sf == FILE_IMMEDIATE)
          return false;
+      if (i->subOp == NV50_IR_SUBOP_MUL_HIGH && sf == FILE_MEMORY_CONST)
+         return false;
       ldSize = 2;
    } else {
       ldSize = typeSizeof(ld->dType);