Mesa (master): r600g/sb: handle more cases for folding in gvn pass
Vadim Girlin
vadimg at kemper.freedesktop.org
Tue May 28 01:26:11 UTC 2013
Module: Mesa
Branch: master
Commit: 08810ca9ef2adc0c094686b6cd8b565eecfa866d
URL: http://cgit.freedesktop.org/mesa/mesa/commit/?id=08810ca9ef2adc0c094686b6cd8b565eecfa866d
Author: Vadim Girlin <vadimgirlin at gmail.com>
Date: Tue May 28 05:24:53 2013 +0400
r600g/sb: handle more cases for folding in gvn pass
Signed-off-by: Vadim Girlin <vadimgirlin at gmail.com>
---
src/gallium/drivers/r600/sb/sb_expr.cpp | 143 +++++++++++++++++++++++++------
src/gallium/drivers/r600/sb/sb_expr.h | 3 +
2 files changed, 118 insertions(+), 28 deletions(-)
diff --git a/src/gallium/drivers/r600/sb/sb_expr.cpp b/src/gallium/drivers/r600/sb/sb_expr.cpp
index f904ea5..053ab2b 100644
--- a/src/gallium/drivers/r600/sb/sb_expr.cpp
+++ b/src/gallium/drivers/r600/sb/sb_expr.cpp
@@ -49,6 +49,14 @@ value* get_select_value_for_em(shader& sh, value* em) {
return d0;
}
+void convert_to_mov(alu_node &n, value *src, bool neg, bool abs) {
+ n.src.resize(1);
+ n.src[0] = src;
+ n.bc.src[0].abs = abs;
+ n.bc.src[0].neg = neg;
+ n.bc.set_op(ALU_OP1_MOV);
+}
+
expr_handler::expr_handler(shader& sh) : sh(sh), vt(sh.vt) {}
value * expr_handler::get_const(const literal &l) {
@@ -177,8 +185,8 @@ bool expr_handler::fold(container_node& n) {
bool expr_handler::fold_setcc(alu_node &n) {
- value* v0 = n.src[0];
- value* v1 = n.src[1];
+ value* v0 = n.src[0]->gvalue();
+ value* v1 = n.src[1]->gvalue();
assert(v0 && v1 && n.dst[0]);
@@ -187,6 +195,9 @@ bool expr_handler::fold_setcc(alu_node &n) {
unsigned cmp_type = flags & AF_CMP_TYPE_MASK;
unsigned dst_type = flags & AF_DST_TYPE_MASK;
+ bool cond_result;
+ bool have_result = false;
+
bool isc0 = v0->is_const();
bool isc1 = v1->is_const();
@@ -202,9 +213,6 @@ bool expr_handler::fold_setcc(alu_node &n) {
apply_alu_src_mod(n.bc, 1, cv1);
}
- bool cond_result;
- bool have_result = false;
-
if (isc0 && isc1) {
cond_result = evaluate_condition(flags, cv0, cv1);
have_result = true;
@@ -254,6 +262,13 @@ bool expr_handler::fold_setcc(alu_node &n) {
cond_result = false;
have_result = true;
}
+ } else if (v0 == v1) {
+ bc_alu_src &s0 = n.bc.src[0], &s1 = n.bc.src[1];
+ if (s0.abs == s1.abs && s0.neg == s1.neg && cmp_type != AF_FLOAT_CMP) {
+ // NOTE can't handle float comparisons here because of NaNs
+ cond_result = (cc == AF_CC_E || cc == AF_CC_GE);
+ have_result = true;
+ }
}
if (have_result) {
@@ -265,10 +280,7 @@ bool expr_handler::fold_setcc(alu_node &n) {
else
result = literal(0);
- n.src.resize(1);
- n.src[0] = sh.get_const_value(result);
- memset(&n.bc.src[0], 0, sizeof(bc_alu_src));
- n.bc.set_op(ALU_OP1_MOV);
+ convert_to_mov(n, sh.get_const_value(result));
return fold_alu_op1(n);
}
@@ -371,7 +383,7 @@ bool expr_handler::fold_alu_op1(alu_node& n) {
if (n.src.empty())
return false;
- value* v0 = n.src[0];
+ value* v0 = n.src[0]->gvalue();
assert(v0 && n.dst[0]);
@@ -425,7 +437,7 @@ bool expr_handler::fold_alu_op1(alu_node& n) {
case ALU_OP1_RECIP_FF:
case ALU_OP1_RECIP_IEEE: dv = 1.0f / cv.f; break;
// case ALU_OP1_RECIP_INT:
-// case ALU_OP1_RECIP_UINT:
+ case ALU_OP1_RECIP_UINT: dv.u = (1ull << 32) / cv.u; break;
// case ALU_OP1_RNDNE: dv = floor(cv.f + 0.5f); break;
case ALU_OP1_SIN: dv = sin(cv.f * 2.0f * M_PI); break;
case ALU_OP1_SQRT_IEEE: dv = sqrt(cv.f); break;
@@ -440,17 +452,20 @@ bool expr_handler::fold_alu_op1(alu_node& n) {
return true;
}
+
bool expr_handler::fold_alu_op2(alu_node& n) {
if (n.src.size() < 2)
return false;
- if (n.bc.op_ptr->flags & (AF_SET)) {
+ unsigned flags = n.bc.op_ptr->flags;
+
+ if (flags & AF_SET) {
return fold_setcc(n);
}
- value* v0 = n.src[0];
- value* v1 = n.src[1];
+ value* v0 = n.src[0]->gvalue();
+ value* v1 = n.src[1]->gvalue();
assert(v0 && v1 && n.dst[0]);
@@ -512,7 +527,62 @@ bool expr_handler::fold_alu_op2(alu_node& n) {
} else { // one source is const
- // TODO handle 1 * anything, 0 * anything, 0 + anything, etc
+ if (isc0 && cv0 == literal(0)) {
+ switch (n.bc.op) {
+ case ALU_OP2_ADD:
+ case ALU_OP2_ADD_INT:
+ case ALU_OP2_MAX_UINT:
+ case ALU_OP2_OR_INT:
+ case ALU_OP2_XOR_INT:
+ convert_to_mov(n, n.src[1], n.bc.src[1].neg, n.bc.src[1].abs);
+ return fold_alu_op1(n);
+ case ALU_OP2_AND_INT:
+ case ALU_OP2_ASHR_INT:
+ case ALU_OP2_LSHL_INT:
+ case ALU_OP2_LSHR_INT:
+ case ALU_OP2_MIN_UINT:
+ case ALU_OP2_MUL:
+ case ALU_OP2_MULHI_UINT:
+ case ALU_OP2_MULLO_UINT:
+ convert_to_mov(n, sh.get_const_value(literal(0)));
+ return fold_alu_op1(n);
+ }
+ } else if (isc1 && cv1 == literal(0)) {
+ switch (n.bc.op) {
+ case ALU_OP2_ADD:
+ case ALU_OP2_ADD_INT:
+ case ALU_OP2_ASHR_INT:
+ case ALU_OP2_LSHL_INT:
+ case ALU_OP2_LSHR_INT:
+ case ALU_OP2_MAX_UINT:
+ case ALU_OP2_OR_INT:
+ case ALU_OP2_SUB_INT:
+ case ALU_OP2_XOR_INT:
+ convert_to_mov(n, n.src[0], n.bc.src[0].neg, n.bc.src[0].abs);
+ return fold_alu_op1(n);
+ case ALU_OP2_AND_INT:
+ case ALU_OP2_MIN_UINT:
+ case ALU_OP2_MUL:
+ case ALU_OP2_MULHI_UINT:
+ case ALU_OP2_MULLO_UINT:
+ convert_to_mov(n, sh.get_const_value(literal(0)));
+ return fold_alu_op1(n);
+ }
+ } else if (isc0 && cv0 == literal(1.0f)) {
+ switch (n.bc.op) {
+ case ALU_OP2_MUL:
+ case ALU_OP2_MUL_IEEE:
+ convert_to_mov(n, n.src[1], n.bc.src[1].neg, n.bc.src[1].abs);
+ return fold_alu_op1(n);
+ }
+ } else if (isc1 && cv1 == literal(1.0f)) {
+ switch (n.bc.op) {
+ case ALU_OP2_MUL:
+ case ALU_OP2_MUL_IEEE:
+ convert_to_mov(n, n.src[0], n.bc.src[0].neg, n.bc.src[0].abs);
+ return fold_alu_op1(n);
+ }
+ }
return false;
}
@@ -579,9 +649,9 @@ bool expr_handler::fold_alu_op3(alu_node& n) {
if (n.src.size() < 3)
return false;
- value* v0 = n.src[0];
- value* v1 = n.src[1];
- value* v2 = n.src[2];
+ value* v0 = n.src[0]->gvalue();
+ value* v1 = n.src[1]->gvalue();
+ value* v2 = n.src[2]->gvalue();
assert(v0 && v1 && v2 && n.dst[0]);
@@ -606,11 +676,12 @@ bool expr_handler::fold_alu_op3(alu_node& n) {
apply_alu_src_mod(n.bc, 2, cv2);
}
- if (n.bc.op_ptr->flags & AF_CMOV) {
+ unsigned flags = n.bc.op_ptr->flags;
+
+ if (flags & AF_CMOV) {
int src = 0;
- if (v1->gvalue() == v2->gvalue() &&
- n.bc.src[1].neg == n.bc.src[2].neg) {
+ if (v1 == v2 && n.bc.src[1].neg == n.bc.src[2].neg) {
// result doesn't depend on condition, convert to MOV
src = 1;
} else if (isc0) {
@@ -622,10 +693,7 @@ bool expr_handler::fold_alu_op3(alu_node& n) {
if (src) {
// if src is selected, convert to MOV
- n.bc.src[0] = n.bc.src[src];
- n.src[0] = n.src[src];
- n.src.resize(1);
- n.bc.set_op(ALU_OP1_MOV);
+ convert_to_mov(n, n.src[src], n.bc.src[src].neg);
return fold_alu_op1(n);
}
}
@@ -635,6 +703,7 @@ bool expr_handler::fold_alu_op3(alu_node& n) {
if (isc0 && isc1 && isc2) {
switch (n.bc.op) {
+ case ALU_OP3_MULADD_IEEE:
case ALU_OP3_MULADD: dv = cv0.f * cv1.f + cv2.f; break;
// TODO
@@ -642,11 +711,29 @@ bool expr_handler::fold_alu_op3(alu_node& n) {
default:
return false;
}
-
} else {
+ if (isc0 && isc1) {
+ switch (n.bc.op) {
+ case ALU_OP3_MULADD:
+ case ALU_OP3_MULADD_IEEE:
+ dv = cv0.f * cv1.f;
+ n.bc.set_op(ALU_OP2_ADD);
+ n.src[0] = sh.get_const_value(dv);
+ memset(&n.bc.src[0], 0, sizeof(bc_alu_src));
+ n.src[1] = n.src[2];
+ n.bc.src[1] = n.bc.src[2];
+ n.src.resize(2);
+ return fold_alu_op2(n);
+ }
+ }
- // TODO
-
+ if ((isc0 && cv0 == literal(0)) || (isc1 && cv1 == literal(0))) {
+ switch (n.bc.op) {
+ case ALU_OP3_MULADD:
+ convert_to_mov(n, n.src[2], n.bc.src[2].neg, n.bc.src[2].abs);
+ return fold_alu_op1(n);
+ }
+ }
return false;
}
diff --git a/src/gallium/drivers/r600/sb/sb_expr.h b/src/gallium/drivers/r600/sb/sb_expr.h
index 1ee48a0..1b77c8e 100644
--- a/src/gallium/drivers/r600/sb/sb_expr.h
+++ b/src/gallium/drivers/r600/sb/sb_expr.h
@@ -42,6 +42,9 @@ unsigned get_predsetcc_op(unsigned cc, unsigned cmp_type);
unsigned get_killcc_op(unsigned cc, unsigned cmp_type);
unsigned get_cndcc_op(unsigned cc, unsigned cmp_type);
+void convert_to_mov(alu_node &n, value *src,
+ bool neg = false, bool abs = false);
+
class expr_handler {
shader &sh;
More information about the mesa-commit
mailing list